prompt
stringclasses 1
value | text
stringlengths 39
387k
|
|---|---|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Texture markings on a 3-D surface provide potentially useful cues to the 3-D shape of the surface when the surface is viewed in a 2-D image [1] – [11] . Local pattern changes across these 2-D images can be separated into modulations in orientation and spatial frequency, features for which it is well established that neurons early in the visual pathway are specialized. Examples are shown in Figure 1 . In Figure 1a, a horizontal-vertical plaid texture is overlaid onto a surface that is corrugated in depth as a function of horizontal position. In the perspective image of the surface, oriented components of the texture that are parallel to the surface slant (in this example the horizontal grating component) converge and diverge forming patterns that we refer to as orientation flows (shown in isolation in Figure 1b ), which exhibit minimal spatial frequency modulation. Although other oriented components also exhibit local changes in orientation, it is the pattern of flows formed specifically by the component running parallel to the surface slant that contains sufficient information to distinguish 3-D curvatures and that consistently conveys 3-D shape in isolation [12] , [13] . The visibility of these orientation flows plays a critical role in correct shape perception [14] – [16] . Components perpendicular to the surface slant (here the vertical grating) are instead modulated in spatial frequency ( Figure 1c ) which in isolation can lead to misinterpretations about surface shape [14] , [15] , [17] . Given that most surface texture patterns contain multiple oriented components, the goal of the current study is to examine how the visibility of orientation flows for 3-D shape is affected by the presence of other oriented components in the surface texture.
10.1371/journal.pone.0053556.g001
Figure 1
Orientation flows and frequency modulations in images of textured 3-D surfaces.
A. A vertically corrugated surface overlaid with a horizontal-vertical plaid texture. B. The horizontal component of the plaid from A shown in isolation. C. The vertical component of the plaid from A shown in isolation. D. The same corrugated surface as in A overlaid with a complex plaid pattern consisting of eight iso-frequency gratings equally spaced in increments of 22.5 degrees.
Many surface textures contain components of roughly the same spatial frequencies at different orientations. When these iso-frequency patterns, such as the plaid in Figure 1a , are mapped onto a developable surface, any slant in the surface out of the fronto-parallel plane causes mismatches in frequency across components such that those that form the critical orientation flows maintain the lowest frequency across the image, and are thus more salient. Saliency increases with increasing surface slant as the frequency mismatch increases [16] . This difference in saliency is even more pronounced in iso-frequency patterns that have multiple frequency components such as the one shown in Figure 1d which contains eight gratings equally spaced in orientation. Thus, for iso-frequency texture patterns on developable surfaces, slanting a surface leads to an increase in visibility of the orientation flows and should facilitate the 3-D shape percept. It has been suggested that a form of frequency-selective cross-orientation suppression (COS) may contribute to this facilitation [16] . At the neural level, neurons responding to the local orientations along the orientation flows may be suppressed by the additional presence of components of the same frequency at other orientations [18] – [20] . The mismatch in the frequencies between the orientation flow components and these components created by slanting the surface may cause a release of frequency-specific COS. This could in turn lead to a decrease in neural suppression of the response to the orientation flows and thus an increase in their visibility and enhancement of perceived 3-D slant and shape.
The current study aims to further investigate the factors that contribute to the visibility of orientation flows that are critical for 3-D shape perception by determining how the presence of components at different orientations affects orientation flow visibility in complex patterns. We are particularly interested in texture patterns and mappings for which slanting the surface does not result in increases in frequency of texture components, as is the case for developable surface mappings. If a COS mechanism contributes to 3-D shape perception as suggested by Li and Zaidi [16] and if it is physiologically and psychophysically broadband for orientation over a wide range of frequencies as found in previous work [18] – [23] , then one would thus expect that texture components at any orientation, and at the same frequency as that which forms the orientation flows, to affect orientation flow visibility similarly. To test this prediction, we used the complex plaid pattern shown in Figure 2 consisting of eight iso-frequency gratings, systematically subtracted oriented components contained in the pattern, and measured contrast thresholds of orientation flows for planar slanted surfaces and curved concave and convex surfaces. Results consistently suggest that the visibility of orientation flows is selectively reduced by the presence of components that are closest to them in orientation. Thus unlike COS mechanisms previously found to be broadband in spatial frequency and orientation, any contribution of the COS mechanisms isolated here to the perception of 3-D shape from orientation flows appears to be both frequency- and orientation-specific.
10.1371/journal.pone.0053556.g002
Figure 2
Complex plaid used in study consisting of 8 gratings equally spaced in orientation.
When slanted out of the fronto-parallel plane, the component parallel to the surface slant creates the critical orientation flows. For left-right slanting surfaces, the horizontal component forms the orientation flows for the texture mapping used in this study. When all 8 components are equal in contrast, the flows are not visible. For demonstration purposes, the contrast of the flows has been enhanced in left and right images.
Materials and Methods
1. Apparatus and Presentation
All stimuli were presented on a calibrated 22 in. Mitsubishi Diamond Pro 2070 flat screen CRT monitor with a 1024×768 pixel screen at a refresh rate of 100 Hz. The monitor was driven by a Cambridge Research Systems ViSaGe Visual Stimulus Generator controlled via a 3.2 GHz Pentium 4 PC. Experimental code was written using the CRS Toolbox for Matlab. A CRS CB6 infrared response box was used to record responses.
Observers’ head positions were fixed with a chin-rest situated 1 m away from the stimulus monitor. All stimuli were presented at the center of the screen and the monitor was elevated such that the center of each image was level with the observer’s eye. Viewing was monocular; each observer patched the same eye across all sessions. An audio cue was the only feedback given to indicate that the observer’s response had been recorded. The experiment took place in a dimly lit room. To minimize fatigue, observers typically ran no more than two consecutive sessions and took breaks as they felt necessary. Sessions were randomized within and across observers.
2. Stimuli and Procedure
To examine whether proximity in orientation of neighboring components affects visibility of orientation flows, we generated a complex plaid texture and mapped it onto planar and corrugated surfaces. To isolate the effects of component orientation while minimizing the presence of frequency modulations in the image, we chose a different texture mapping from the developable surface mapping used in Figure 1 . We used a volumetric carved solid texture mapping which creates identical orientation flows in the perspective image but which minimizes frequency modulations of other oriented components (see Li & Zaidi [15] for mapping details). The complex plaid textures were created by superimposing eight grating components at 2 cpd varying in orientation in 22.5 degree increments (0, ±22.5, ±45.0, ±67.5, and 90 degrees, where 0 deg is defined as horizontal). All components were subjected to the same texture mapping. When slanting the planar surface out of the fronto-parallel plane, only the component parallel to the surface slant creates the critical flows that contain sufficient information to distinguish 3-D curvatures [12] , [13] . In this mapping, frequency modulations are minimized across components. For surfaces slanted around the vertical axis (left/right slants), the critical flows arise from horizontal components ( Figure 2 ); whereas, the critical flows for surfaces slanted around the horizontal axis (floor/ceiling slants) arise from vertical components (this can be seen by rotating the images in Figure 2 by 90 deg). Visibility of the orientation flows was quantified by varying their contrast and determining the lowest contrast at which they were visible (the contrast threshold, CT). CTs were determined in the presence of the other components to evaluate their effect on orientation flow visibility. CTs of critical flows were measured for fronto-parallel surfaces and surfaces slanted ±60 degrees out of the fronto-parallel plane (examples shown in Figure 2 ). Additionally, similarly mapped sinusoidally corrugated surfaces were used to examine orientation flow visibility for more complex, curved surfaces. In the corrugated condition, each image contained 1.5 cycles of the sinusoidal depth corrugation, and the corrugations were simulated to span a depth amplitude of 14 cm.
CTs were measured in the presence of five different complex plaid textures shown in Figure 3 . This figure shows patterns used for the left/right slanted condition (i.e. vertical slant axis). The top row shows patterns in the absence of the critical horizontal component and the bottom row shows the same patterns with the horizontal component added in. In the top row, Pattern 1 consists of seven oriented components (all but the horizontal that creates the orientation flows). In Patterns 2–5, we systematically subtracted components (indicated below each pattern label) from Pattern 1. For example, Pattern 4 was created by removing the components that are +22.5 and −22.5 degrees away from the critical flow component (here the horizontal component). Thus, it is the composite of ±45.0, ±67.5, and +90 degree components, and missing components closest in orientation to that of the flow component. Video frames containing the complex plaid patterns ( Figure 3 , top row) were interleaved with frames containing the critical flow component so that the contrast of this component could be independently manipulated. As a result of the interleaving, the contrast of each video frame could only be 50%. Thus the contrast of each grating component of the complex plaid frame was set to the maximum possible, 7.1%. For demonstration purposes, the critical horizontal flow component has been added at a contrast equal to that of all other components in the bottom panels of Figure 3 .
10.1371/journal.pone.0053556.g003
Figure 3
Five complex plaid patterns against which contrast thresholds of the orientation flows were measured.
Patterns shown here were used in the left-right planar surface slant condition. Pattern 1 contains 7 gratings, each 22.5 deg from the next, excluding the horizontal (0 deg) grating. Pattern 2 is Pattern 1 less the 67.5 deg components. Pattern 3 is Pattern 1 less the 45 deg components. Pattern 4 is Pattern 1 less the 22.5 deg components. Pattern 5 is Pattern 1 less the 90 deg component.
All observers completed three experimental conditions: 1) planar surfaces slanted left/right about a vertical axis, 2) planar surfaces slanted along floor/ceiling slants about a horizontal axis, and 3) vertically corrugated surfaces varying in depth as a function of horizontal position. CTs of the flows were determined using a 3-down, 1-up staircase procedure, in which the contrast of the orientation flows was varied. In a single session, observers were presented with two randomly interleaved staircases (one ascending and one descending) for each of two complex plaid patterns ( = 4 staircases, 2 threshold estimates per pattern). Staircases were programmed to complete two reversals via 1.8% contrast steps followed by eight reversals via 0.4% contrast steps, from which contrast at the last six reversals were averaged as the estimate of threshold.
Each of the five patterns were blocked randomly into pairs for each session for each observer, and each pattern was tested in two separate sessions resulting in a total of 40 sessions across the three experimental conditions. Thus, each of the two planar conditions consisted of 15 sessions of randomly paired patterns at each of the three following slant conditions: 60 degrees positively slanted, 60 degrees negatively slanted, and fronto-parallel. The corrugated experimental condition contained 10 sessions of randomly paired patterns, in which there were five patterns for each of the two corrugations – convex and concave. As a result, four CT estimates for each pattern were averaged per observer.
Observers were seated 1 m away from the stimulus monitor and head position was fixed with a chinrest. Viewing was monocular in a dimly lit room. At the beginning of each session, a mean grey screen (58 cd/m 2 ) was presented with a central fixation cross for 1 minute, which remained onscreen throughout the rest of the session. A tone signaled the start of the trials after the initial 1 minute adaptation. Stimuli were presented in circular apertures spanning 6.5 degrees against the mean grey background. In each trial of the experiment, observers were presented with two sequential stimuli for 500 ms each, separated by a 400 ms inter-stimulus interval, which were accompanied by audible tones of different frequencies. One of the stimuli was one of the complex plaid patterns without the orientation flows (e.g. panels in top row of Figure 3 ), and the other was the same complex plaid with the orientation flows (e.g. analogous panels in bottom row of Figure 3 ). The interval containing the flows was randomized across trials. Observers were asked to judge which of the two intervals contained the orientation flow pattern.
Observers were presented with written as well as verbal instructions regarding the experimental task, which included examples of visual stimuli and the patterns of orientation flows they were asked to detect, at very high contrast in the presence of various complex plaid patterns. Practice sessions were run for each of the three conditions for one of the pattern types (randomly chosen). Each session lasted approximately 15 minutes. Breaks were given between sessions to minimize visual fatigue.
3. Observers and Ethics Statement
The two authors and six naïve individuals (3 experienced but uninformed) served as observers in the experiment. All observers had normal or corrected-to-normal visual acuity. All research followed the tenets of the World Medical Association Declaration of Helsinki and informed written consent was obtained from the observers after explanation of the nature of the study. The research was approved by the Queens College Institutional Review Board.
Results
We predict that if the orientation of neighboring components does not affect orientation flow visibility, CTs should be unaffected by the varying combinations of oriented components present in the pattern. If there is an effect of orientation proximity, then components that are closest in orientation to the critical flows should mask the flows more than components that are farther away in orientation. Therefore, subtracting those closest in orientation to the critical flow (Pattern 4) should cause a significant decrease in CTs.
An independent-measures ANOVA (F(4,315) = 28.83, p <.001) was conducted to compare CTs for each pattern collapsed across all surface conditions. Using the Bonferroni procedure for all pairwise comparisons, we found that CTs were significantly reduced when the components closest in orientation to that of the flows were subtracted out of the overall pattern (Pattern 4) reflecting the fact that the orientation flows were unmasked and thus most visible. This finding is inconsistent with the prediction that orientation flow visibility would be affected by components at all orientations equally. To further evaluate this finding, additional statistical tests were performed for each of the experimental conditions. These tests showed similar results. Per experimental condition, we used an independent-measures ANOVA to evaluate overall main effects. For each slant/corrugated surface, we used a repeated-measures ANOVA. All post-hoc pairwise comparisons were evaluated using the Bonferroni procedure. These are discussed below.
1. Planar Left/right Slanted Surfaces
Data for left/right slanted surfaces are presented in Figure 4 . In the graph on the left ( Figure 4A ), CTs of the orientation flows were averaged across fronto-parallel and ±60 deg slanted surface conditions. An independent-measures ANOVA yielded a significant difference ( F (4,115) = 11.14, p <.001) between pattern means. Post-hoc testing revealed significant differences only with respect to Pattern 4. For the graphs on the right ( Figures 4B to 4D ), subsequent analysis was conducted for each of the slant conditions separately – right slant, fronto-parallel, and left slant – using a repeated-measures ANOVA, which revealed significant differences ( F (4,28) = 12.28, p <.001; F (4,28) = 15.02, p <.001; and F (4,28) = 5.21, p = .003, respectively). Post-hoc analysis found similar significant differences for Pattern 4, except between Patterns 3 and 4 when surfaces were slanted in either direction ( Figure 4B and Figure 4D ). Overall, these data indicate that the absence of neighboring components (or orientations closest to the orientation flows) allows for greater visibility of the orientation flows across surface slants.
10.1371/journal.pone.0053556.g004
Figure 4
Mean contrast thresholds for the left-right planar slanted condition.
The leftmost panel shows orientation flow CTs for each of the 5 patterns tested (see Figure 3 ) averaged across the fronto-parallel, +60 and −60 deg slanted conditions. Pattern 4 is the pattern missing the components closest in orientation to those of the orientation flows. CTs for each of the fronto-parallel, +60 and −60 deg conditions are shown on the right. Error bars represent 95% confidence intervals. Asterisks indicate significance at α<0.05.
2. Planar Floor/ceiling Slanted Surfaces
To see if our results generalize to another slant axis, we ran an analogous condition for surfaces slanting around a vertical axis ( Figure 5 ). An independent-measures ANOVA found significance ( F (4,115) = 9.81, p <.001) for the overall main effects of the floor/ceiling slanted condition (left panel, Figure 5A ). Using post-hoc analysis, significant differences were found only with respect to Pattern 4. Repeated-measures ANOVAs were performed per slant condition and resulted in the following significant findings: floor slant ( F (4,28) = 7.99, p <.001), fronto-parallel ( F (4,28) = 8.50, p <.001), and ceiling slant ( F (2.71) = 12.31, p <.001). Post-hoc testing showed significant differences for Pattern 4, except between Patterns 3 and 4 for all surface slants ( Figure 5A ), Patterns 2 and 4 at fronto-parallel ( Figure 5C ), and Patterns 1 and 4 for the ceiling slanted surface condition ( Figure 5D ). Overall, the pattern of results for floor/ceiling slants is consistent with the pattern of findings for left/right slants.
10.1371/journal.pone.0053556.g005
Figure 5
Mean contrast thresholds for the floor-ceiling planar slanted condition.
See caption for Figure 4 . Error bars represent 95% confidence intervals. Asterisks indicate significance at α<0.05.
3. Vertically Corrugated Surfaces
Finally, we also wanted to see if these findings generalized from planar surfaces to more complex, curved surfaces. We thus used vertically corrugated surfaces (varying sinusoidally in depth as a function of horizontal position) that were either centrally convex or concave. Vertically corrugated surfaces are composed of local left/right slants along the surface. Since these results should be consistent with those found for left/right slants, we predict the analogous pattern of results would be found between horizontally corrugated surfaces and our results for floor/ceiling slanted surfaces. In this final condition, the average of all flows produced a similar significant finding ( F (4,115) = 9.89, p <.001) using an independent-measures design ( Figure 6A ). As previously found, post-hoc testing showed significant differences between patterns only with respect to Pattern 4. Moreover, individual curvature conditions ( Figure 6, right panel s) yielded significance for the pattern with the most proximal components subtracted out – convex ( F (4,28) = 11.69, p <.001) and concave ( F (4,28) = 9.00, p <.001), using a repeated-measures ANOVA. Subsequent analysis indicated significant differences only with respect to Pattern 4, except for Patterns 2 and 4 in the concave surface corrugation ( Figure 6B ).
10.1371/journal.pone.0053556.g006
Figure 6
Mean contrast thresholds for the corrugated condition.
The leftmost panel shows orientation flow CTs for each of the 5 patterns tested (see Figure 3 ) averaged across the concave and convex conditions. CTs for each of the concave and convex conditions are shown on the right. Error bars represent 95% confidence intervals. Asterisks indicate significance at α<0.05.
To summarize across surface conditions, CTs are significantly reduced when components that are closest in orientation to that of the flows are subtracted from the pattern. This trend indicates that visibility of orientation flows that convey 3-D slant and curvature is most impaired by these proximal components. It is worth noting that for some surface conditions, CTs for Pattern 3 were not significantly different from the CTs for Pattern 4. Since ±45 deg components were subtracted from Pattern 3 and these components are the next most proximal in orientation to the orientation of the component creating orientation flows, it is not surprising that in some cases the removal of these components still has an effect on orientation flow visibility. However, when all surface conditions were averaged together (A panels in Figures 4 , 5 , 6 ) it is clear that CTs were consistently and significantly lowest for Pattern 4. It is also worth noting that the number of components across the five patterns tested is different (seven in Pattern 1, five in Patterns 2–4, six in Pattern 5). Despite the varying number of components, all patterns other than Pattern 4 contained the proximal ±22.5 deg components. This demonstrates not only the selective effect of the proximal components on orientation flow visibility, but the lack of effect on visibility of the number of components contained in the pattern, at least for the number of components tested in these patterns.
Discussion
The goal of the current study was to examine the effects of orientation of non-critical texture components on the visibility of critical orientation flows for 3-D shape. We chose to examine this using complex plaid surface textures containing multiple spatial components, which might be more generally representative of naturalistic surfaces textures than simple plaids. Our results consistently show that, for planar and curved surfaces, orientation flow visibility is most affected by the presence of components oriented closest to them. Removal of these components caused significant increases in orientation flow visibility (as quantified by reductions in contrast threshold) while removal of all other less proximal components had no effect on their visibility.
For iso-frequency surface textures mapped onto developable surfaces, slanting the surface out of the fronto-parallel plane creates frequency mismatches between the texture components that create critical orientation flows and those that do not, thereby increasing the visibility of the orientation flows and increasing the strength of the slant percepts they convey. It has been suggested that the frequency mismatch increases orientation flow visibility via the release of a cross-orientation suppression (COS) mechanism that is frequency-selective [16] . Although COS has been previously implicated in visual processes such as orientation tuning [19] , [24] , [25] , contrast gain control [20] , [26] – [29] , and the reduction of redundancy in the coding of natural images [30] – [32] , this was the first suggestion of its potential contributions to 3-D shape perception.
The results of the current study, which utilizes a texture mapping that minimizes frequency modulations in texture components, compliment the results of Li and Zaidi [33] by showing that orientation flow visibility is not only affected by the frequencies of neighboring components, but also by their proximity in orientation such that components closer in orientation mask orientation flow visibility. Therefore, if a COS mechanism is contributing to our results, the process is not only frequency-selective, but also orientation-selective. However, given that many studies have found COS to be broadband in orientation [18] – [23] , an additional or alternative mechanism that may be contributing to the visibility of orientation flows is surround suppression, which has been physiologically and psychophysically characterized as much being more narrowly tuned for orientation [20] , [21] , [34] , [35] . The dimensions of our stimuli do not preclude the possibility that both types of suppression might be contributing to orientation flow visibility. The results also suggest that for complex texture patterns, the number of components present in the pattern has little effect on orientation flow visibility. Patterns 2–4 all contained five components, but only Pattern 4, for which the most proximal components were removed, showed significant decreases in contrast thresholds. Conversely, Patterns 1 and 5 respectively contained seven and six components (including the proximal components) and, as for Patterns 2 and 3 with fewer components, similarly had no effect on orientation flows visibility.
Although our study was not designed to parametrically quantify the orientation tuning of the underlying mechanisms at play, it is worth discussing our results in the context of orientation masking studies. In these studies, the tuning of perceptual orientation-selective mechanisms is quantified by measuring the visibility of a single test grating stimulus in the presence of a second overlaid masking grating stimulus of the same frequency but different orientation. Contrast thresholds of the test grating are typically greatest when the orientation of the mask matches that of the test, and falls off as this orientation difference increases. The tuning of the underlying mechanisms is characterized by the rate of this threshold fall-off. The tuning of the underlying mechanisms quantified in this way has been found to be on the order of 10–20 degrees, with several studies finding substantial decreases on test thresholds beyond about 12 degrees [36] – [38] . In light of these results, the fact that components 22.5 deg away from the orientation flows in our study had such a significant effect on their visibility may seem somewhat surprising. It is possible that the mechanisms isolated with masking studies are different from those contributing to orientation flow visibility in our study, which appear to be somewhat more broadly tuned. An important difference between the patterns used in this study and conventional masking studies is that ours contained multiple components. Thus, although the non-proximal components did not appear to have any effect on orientation flow visibility, perhaps orientation flow visibility in the presence of additional components requires the recruitment of additional mechanisms.
Another important difference between our stimuli and those used in masking studies is in the number of local orientations contained in the image. For our fronto-parallel conditions, all grating components were unmodulated in orientation and thus the difference in orientation between the orientation flows and the most proximal component was fixed at 22.5 deg across the image. For slanted and corrugated conditions, however, all components, including those that create orientation flows, contained multiple orientations across the image (e.g. see orientation flow patterns in inset panels of Figure 2 ). Thus there was in fact a range of orientation differences between the orientation flows and the most proximal components across the image. That said, our results for fronto-parallel, slanted, and corrugated conditions were qualitatively similar. Thus, it appears the effects of proximal components on orientation flow visibility are invariant to these orientation differences across the image.
Although not directly tested in this study, the importance of critical orientation flow patterns in the perception of 3-D shape has been substantiated in our previous work [12] – [15] , [17] , [39] – [41] . Together with Li and Zaidi [16] , our results have important implications for image processing algorithms used to optimize 3-D shapes of surfaces and objects. Since orientation flow visibility is critical for 3-D shape perception, maximizing visibility is important. One way to maximize visibility is to ensure that oriented components of texture patterns be as different as possible in frequency from those that create orientation flows, or use a texture mapping (such as that used for developable surfaces) where surface slant creates frequency mismatches from those of the orientation flows. A second way to maximize visibility, as shown by the results of this study, is to ensure that the non-critical oriented components are at orientations that are far from those of the orientation flows. Surface textures with these spatial restrictions should convey the strongest 3-D slant and shape in projected images.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Breeding success largely depends on the timing of breeding relative to the timing of maximum food availability [1] – [3] . Many forest passerine bird species in temperate regions feed their young with caterpillars, which occur only in a short period of time during spring (caterpillar biomass peak). Due to increased spring temperatures, the window in which food availability is high has shifted forward in time over the last 25 years [4] . As a result, the optimal timing for breeding advanced, but many species, like the great tit, have not adjusted their timing sufficiently, causing them to breed too late [4] – [6] .
The reproductive system of most seasonally breeding birds, including great tits, shows a clear seasonal pattern [7] , [8] . Gonads are regressed during winter, grow slowly during late winter and grow rapidly in spring until fully developed. Gonads are regressed again after the breeding season. The rapid growth in spring is affected by increasing day length, which causes the release of gonadotrophins [7] , [9] , [10] . Although it has been shown that temperature can affect the speed at which gonads develop in great tits breeding in Southern latitudes [11] , the speed of gonadal development in great tits breeding in more Northern latitudes is not accelerated by increasing spring temperatures (The Netherlands [12] , Scandinavia [11] ). In these latitudes, gonadal growth is driven by photoperiod [11] . Increasing spring temperatures due to global warming will therefore not advance the birds' readiness to reproduce. As climate change does not affect the seasonal change in photoperiod, a possible reason as to why great tits are not advancing their laying date adequately is that gonads are not fully developed early enough to allow early egg laying.
To test whether gonadal development is hampering early egg laying, gonadal development needs to be experimentally advanced. This could be done by manipulating the photoperiod a bird experiences, as is shown in an experiment where blue tits ( Cyanistes caeruleus ) in captivity (with ad libitum food) could be tricked into laying their eggs in winter (January) by exposing them to long days from December onwards [13] . Under a natural photoperiod, egg laying in January is not possible as the reproductive system will not be fully developed at that time.
Although the experienced photoperiod of captive birds is easily manipulated, photostimulating birds in the field has many practical problems, such as fitting a large number of nest boxes with a light, batteries and a timer. More importantly, birds do not always sleep in a particular nest box in the period before egg laying and thus it is difficult to determine which bird is photostimulated and by how much. Taking wild birds into captivity for photostimulation treatment for long periods can cause problems as breeding vacancies as a result of the removal of territorial birds are filled within a few days by unpaired birds, which may lead to fights after release of the original territory holder or its female (pers. comm. P. de Goede (NIOO-KNAW)).
Previous experiments have shown that the exposure to ‘a single long day’ can affect levels of hormones involved in reproduction. Nicholls and colleagues [14] kept Japanese quail under 8L∶16D and gave them a single long day of 20L∶4D resulting in an increase of luteinizing hormone (LH) and follicle stimulating hormone (FSH) within four hours of the end of the long day. These birds were kept in constant darkness thereafter and their LH and FSH levels decreased slowly over the next 8 to 10 days. Creighton and Follett [15] , who performed a similar experiment, kept Japanese quails under short day lengths after just one long day and report that LH remained elevated for three days after photostimulation. Follet and colleagues [16] have shown that LH levels of white-crowned sparrows ( Zonotrichia leucophrys gambelii ) transferred from short (8L∶16D) to long days (20L∶4D) increased six-fold in five days, with the largest increase after the first day (three-fold increase). Saab and colleagues [17] has shown that a single long day increased gonadotrophin-releasing hormone and LH concentrations in white-throated sparrows ( Zonotrichia albicollis ). A single long day also affects gonadal growth of both male and female Song sparrows ( Melospiza melodia ). Wingfield [18] showed that after one single long day, females' gonads grew for up to 60 days, even though the changes in LH and FSH were only present for a few days. In all of the experiments above food was available ad libitum . If a single long day would affect gonadal growth in wild female great tits, it would allow us to test the hypothesis that laying dates are restricted by photo-induced gonadal growth.
Outside the breeding season gonads are regressed, implying that there are costs connected with having and/or maintaining fully developed gonads. These costs can be present in terms of increased risk of predation due to lower aerial maneuverability and take-off ability [19] – [21] , but might also involve energetic maintenance costs. Thus, it is likely that advancing gonadal growth in spring also comes with a cost. If energetic costs restrict early gonadal development, only birds in habitats with high food availability might be growing their gonads as a reaction to a single long day.
The aim of this study was to determine (i) if a single long day induces gonadal growth in captive female great tits and (ii) if a single long day in spring affects laying dates in two field populations which differ in the availability of supplementary food in the period before and during egg laying. If gonadal growth restricts early egg laying, we expect photostimulated birds to lay earlier compared to control birds in both study areas. If gonadal growth is restricted by a combined effect of photoperiod and food availability, we expect only those birds in the population with available supplementary food during the pre-laying period to advance egg laying.
Results
Aviary experiment
Gonadal growth of female great tits in captivity could only be determined for 9 out of the 15 females due to technical difficulties (two control females, three females photostimulated once and four individuals photostimulated twice in December were measured). Females that were photostimulated once or twice (20L∶4D) showed gonadal growth one month after the treatment (One-sample Wilcoxon signed rank test: V = 28, P = 0.016), whereas both females of the control group did not show gonadal growth. Gonadal growth of females which were photostimulated once did not differ from females that were photostimulated twice (Two-sample Wilcoxon signed rank test: W = 5, P = 0.86; see Fig. 1 ).
10.1371/journal.pone.0035617.g001
Figure 1
Effect of photostimulation on follicle growth in captive great tits in winter.
Follicle length ( A ) and volume ( B ) of the three experimental groups in the aviary experiment before and one month after the photostimulation. All birds were kept in an outdoor aviary and only moved indoor for the photostimulation treatment. Of the females of which we successfully measured the largest follicle in December and January, two birds were kept under natural light conditions (solid lines), 3 individuals were given one long day (dashed lines) and 4 individuals were given two long days (seven days apart; dotted lines).
Field experiment
In total, 63 out of 103 females in the field experiment started egg laying in our study area ( Table 1 ). Females started laying on average (± SE) on April 18.0 (±1.17), 14.3 (±0.69) and 16.7 (±1.0) in the Hoge Veluwe 2009, Oosterhout 2009 and 2010, respectively. Laying dates of first eggs did not differ between photostimulated and control female great tits in either of the two years and in either of the two study areas (P>0.22 for all comparisons; Table 2 ; Fig. 2 ).
10.1371/journal.pone.0035617.g002
Figure 2
Effect of photostimulation on laying dates in free living great tits.
One third of the females encountered during a night check of all nest boxes were not taken indoors, 1/3 were taken into captivity and kept under natural light regime and 1/3 were taken into captivity and kept under a long light regime (according to figure 3 ). Note that random jitter is applied (on the X-axis only) to separate overlapping data point.
10.1371/journal.pone.0035617.t001 Table 1
Sample sizes of the field experiment to study the effect of photostimulation on laying dates in wild living great tits ( Parus major ).
Treatment
Sample size
Birds with laying date
Percentage with lay date
Oosterhout
2009
Indoor – one long day
10
6
59%
Indoor – natural day length
10
6
Not taken indoors
7
4
2010
Indoor – one long day
10
7
72%
Indoor – natural day length
10
5
Not taken indoors
16
14
Hoge Veluwe
2009
Indoor – one long day
15
7
53%
Indoor – natural day length
14
7
Not taken indoors
11
7
Total indoor
69
38
55%
Total not taken indoor
34
25
73%
Totals
103
63
We created 3 experimental groups: (i) photostimulated females (ii) females which experienced natural day length in captivity and (iii) females which experienced natural day length in the field. Females from group (i) and (ii) were kept indoors for one day and two nights after which they were released in the field at the location where they were caught.
10.1371/journal.pone.0035617.t002 Table 2
Results of the statistical analyses (ANOVAs) of the effect of photostimulation on first egg laying dates (date at which the first egg of a clutch is laid) of wild great tits ( Parus major ) in the field experiment.
Subset
Variable
df
Error df
Sum of squares
Mean squares
F
P
Oosterhout
Treatment * year
2
33
1.05
0.53
0.05
0.96
Oosterhout
Treatment
2
36
38.39
19.19
1.59
0.22
2009
Treatment * location
2
30
40.35
20.18
0.96
0.39
2009
Treatment
2
33
3.02
1.51
0.06
0.94
The experiment had three treatments: indoor photostimulated; indoor natural light regime; not taken indoors. The experiment was done in two locations in 2009 and in one location in 2010. We divided the analyses in two; i) only data from the Oosterhout population (2009 and 2010) and ii) only data from 2009, where the experiment was done in Oosterhout and Hoge Veluwe population.
Discussion
This study aimed to test the hypothesis that timing of egg laying is restricted by the timing of gonadal development, which is under photoperiodic control. We showed that, even in winter (average daily temperature in the outdoor aviary in December 2008 = 1.8°C, January 2009 = 0.1°C; February = 2.7°C), gonadal growth is initiated after exposing captive female great tits to ‘a single long day’ (20L∶4D). In a field experiment in spring, however, free living female great tits which were given a single long day did not advance egg laying, either in the study area with or without good food conditions and the availability of supplementary food in the period before egg laying. Although we did not measure gonadal development prior to egg laying in the field study and can therefore not confirm that our ‘single long day’ treatment also worked in spring, these results suggest that the seasonal timing of gonadal growth does not play a major role in restricting great tits from advancing their laying date.
Follicles of captive female great tits which were exposed to a single long day in winter started growing. As gonads are in a regressed state outside the breeding season and only start growing very slowly in early winter, follicles were still small in December (maximum 0.3 mm long). It is therefore difficult to measure their size: in 3 out of 15 cases, the follicle size in December was too small to be measured precisely. Follicle sizes in January ranged from 0.1 mm to 0.5 mm, which is still small compared to the accuracy with which we can measure them (0.1 mm). Although measuring follicles in winter is difficult, we have confidence in our measurements.
Both populations have a long history of selection for early laying: reproductive success, measured as the number of fledged offspring that recruited in the breeding population in the next year, was higher for early breeding females for 21 out of 25 years in Oosterhout, and for 23 out of 25 years in the Hoge Veluwe [4] . If gonadal growth would restrict egg laying, this would most likely occur in years with high spring temperatures since warm springs lead to early egg laying. Temperature in the period 16 March until 20 April correlates well with laying dates [4] . In 2010, mean temperature in this period was high and indeed, the first pair that started egg laying in the Oosterhout study population was the earliest recorded laying date in the last 55 years in this study area. Although laying dates are early in years with warm springs, gonadal growth is not affected by temperature in the period before egg laying [12] . Therefore we would expect an effect of photostimulation (especially in 2010) if gonadal growth was restricting egg laying dates.
Although follicle volumes increased after one day of photostimulation in captivity during winter, photostimulation in spring did not affect laying dates in the field. There are a number of potential explanations.
We did not measure gonadal growth in a subset of the photostimulated animals in March. It is therefore possible that the photostimulation in March did not result in gonadal growth, causing the lack of effect in laying dates. To our knowledge, no studies have focused on the seasonal variation in the strength of the response in hormonal change or gonadal growth to a single long day. However, Silverin [8] caught male great tits during different months of the year and exposed (a part of) them to a 20L∶4D light regime for 100 days. Testis growth of male great tits caught in December was less than one millimeter after 10 days of photostimulation, whereas male great tits caught in March grew their tested on average just over 2 mm. Thus, at both dates photostimulation leads to a reaction that was adequate for that time and developmental stage and it is therefore likely that follicle growth of the female great tits in our field experiment was stimulated by our experimental treatment in March.
Since the natural day length is shorter in winter than in spring, the photostimulation in winter was a relatively stronger stimulation compared to the photostimulation in spring. Silverin [8] measured testes growth of male great tits exposed to two different light regimes (14L∶10D and 20L∶4D) and showed that the gonadal maturation was faster in he 20L∶4D group compared to the 14L∶10D group. In our experiment in spring, the increase in day length was still 9 hours and 15 minutes for the photostimulated group, which is more than the increase in day length from 8L∶16D to 14L∶10D in the experiment from Silverin [8] which resulted in clear effects on hormones and gonadal development. We therefore believe the photostimulation treatment in spring is strong enough to evoke a response in gonadal growth.
Another possible explanation why photostimulated females did not advance egg laying is that, besides the primary predictive cue of photoperiod, supplementary cues are used to time egg laying [22] , for example the increase in spring temperature [12] , [23] . One of the supplementary cues might be food availability. During the aviary experiment, as well as in most other experimental studies [12] , [23] , food and water were given ad libitum while the females in the field encountered all kinds of stressors (predation risk, lower food availability, inter-species interactions et cetera). In addition, eggs are laid in cold weather conditions under which foraging efficiency is low [24] and energetic costs are high [25] , [26] . Therefore, adverse food conditions can restrict growth of gonads (either acting as a cue or as an energetic constraint). Perfito et al [27] showed that male zebra finches ( Taeniopygia guttata guttata ) under long photoperiod but with food restrictions did not develop their testis, similar to those under short day lengths, while birds under long photoperiod with ad libitum food did. Zebra finches, however, are opportunistic breeders that have evolved to use food availability as a cue, since they live in areas where food availability is unpredictable and do not follow a seasonal pattern [28] . Testis size of male European starlings kept in aviaries was not affected by a food restriction [29] , however, as the birds were able to maintain body weight during the treatment they were possibly not restricted enough. Also, in most physiological studies like these, males are used, while it is the females that determine the timing of reproduction [30] . It is likely that the female great tits in our field experiment were food restricted. If this would be the case, we would expect females in Oosterhout (which had a richer food supply) to advance egg laying compared to the control females, but not the Hoge Veluwe females. However, photostimulated females from the Oosterhout population also did not advance laying compared to the control groups in neither of two years. Therefore, we hypothesize that other supplementary cues, like temperature, or perhaps the availability of insects, prevented photostimulated females from laying early relative to the phenology of their environment.
While egg laying can not start before males and females have fully developed reproductive organs, we only photostimulated female great tits in the field experiment. Field observations showed that the male reproductive system is functioning well before that of the females [22] . Gonadal measurements of great tit breeding pairs in captivity confirmed this by showing that males have mature gonads sometimes weeks earlier compared to the exponential growth phase of female follicles [23] , [31] . Therefore, laying dates in wild birds are not likely to be restricted by the development of the male reproductive system.
It is important to know which factors hamper the lack of shift in laying dates because these can have different implications on how to adapt to future climate change. If a shift in laying date is hampered by gonadal development (which our result suggest is it not), birds have to adjust their rules in which day length is used as a cue. Since climate change is not affecting day lengths, gonadal development is likely to restrict a shift in laying dates in the future. At the moment there seems be other reasons why the shift in the advancement of laying date lags behind this shift in the phenology of the food, leading to an increasing phenological mismatch between food availability and food requirements over the last decades. Future temperature increase will further this mismatch. A better understanding of the causes and consequences of the (in)ability of birds to adapt their timing of reproduction to restore the synchrony with their prey is important and will provide insights into the effects of future climate change on population viability [32] .
Methods
Ethics Statement
The experiments reported here comply with the current law in The Netherlands and were carried out under licenses of the Animal Ethics Committee of the KNAW (Royal Netherlands Academy of Arts and Sciences, protocol CTE.08.10 & CTE 09.01).
Study areas – This study was carried out in two study areas, Hoge Veluwe and Oosterhout (the Netherlands), about 50 km apart. Strong natural selection for early laying females exists in both populations [6] . The study sites were chosen for this experiment because food availability in the period before egg laying differs between them. National Park ‘De Hoge Veluwe’ (52° 02′ 07″ N 5° 51′ 32″ E) is a mixed forest on poor sandy soils, while Oosterhout (51° 52′ 22″ N 5° 50′ 22″ E) is a rich deciduous forest on rich river clay. Besides finding food in the rich undergrowth, females of the Oosterhout population regularly fly to the nearby village (maximum distance 1 km) to feed on the abundant supplemented food (observed during radio tracking, unpublished data LtM & MEV). This food was available until most birds laid their eggs and consisted mainly of fat and peanuts. Over the last 25 years, natural selection favored early breeding in 21 of the past 25 years in Oosterhout and in 23 out of 25 years in the Hoge Veluwe [4] .
Aviary experiment
To test if one (or two) long days initiate gonadal growth, we caught 15 wild female great tits at the end of November 2008 around the Netherlands Institute of Ecology, Heteren (The Netherlands; 51°57′20″N–5°44′34″E). All females were housed in one large outdoor aviary under natural light and temperature conditions and ad libitum food and water. Six days before the light treatment (December 5 th 2008), length of the largest ovarian follicle of all females was measured during laparotomy without knowledge of the treatment each female would be assigned to. To measure follicle development, a small incision was made between the last two ribs on the left side. By parting the ribs slightly, length of the largest follicle was measured to the nearest 0.1 mm with an ocular scale. When the length of the largest follicle was too small to be measured (significantly smaller than 0.1 mm) we reported a length of 0.05 mm ( n = 3 in December). All laparotomies were carried out by SVS under light Isoflurane anesthesia. Follicle volume was calculated as V = 4/3 · π · a 3 where a is ½ the length of the follicle.
For the first photostimulation treatment (December 11 th 2008), all females were moved indoors after sunset into individual cages in two separate rooms (see Fig. 3 for a schematic overview of the treatments). Ten females in one room were exposed to light for 20 hours (7AM–3AM), then dark for four hours, after which the lights were turned on again at 7AM the following day. Five control females in the second room were kept under the natural light regime (light from 8:40AM to 4:30PM). All females were returned to the outdoor aviary the day after the photostimulation treatment. Seven days later (December 18 th 2008), five of the 10 photostimulated females were photostimulated again using the same protocol, while the other five photostimulated females and five control birds were also kept indoors but under natural photoperiod. Thus, the aviary experiment consisted of three treatment groups, each containing five birds: (i) natural short photoperiod (ii) one long day and (iii) two long days with an interval of one week in between. One month after the first photostimulation (January 15 th 2009) the length of the largest follicle was measured again for each female. One week after the last laparotomy all 15 females were released into the wild at the catching location ( Fig. 3 ).
10.1371/journal.pone.0035617.g003
Figure 3
Schematic overview of the photostimulation treatment for both the aviary and the field experiment.
For the experiment in captivity (December 2008), females were moved from the outdoor aviary to our indoor facilities after sunset, where a part of the females were kept under natural day length and a part were kept in a different room under a long day length (20L∶4D). Females were moved to their outdoor aviary in the morning of day three. For the field experiment (March 2009 and 2010), females were brought from the field into out indoor facilities for a similar treatment. On the morning of day three, all females were released in the field at the location where they were caught. Grey bars resemble darkness while open bars represent light periods; numbers in the bar are the hours of the day.
Field experiment
We carried out a field experiment to test if laying dates in the wild are restricted by gonadal development of the females. In the Oosterhout (∼150 nest boxes) and Hoge Veluwe study area (∼440 nest boxes) all nest boxes were checked at night for the presence of female great tits (Oosterhout 2009: February 25 th ; 2010: March 1 st ; Hoge Veluwe 2009: March 3 rd ). All females were banded with a uniquely numbered aluminum ring as well as a unique colour band combination. Of the 103 females encountered, we took 69 females into temporary captivity (Oosterhout: 2009 n = 20, 2010 n = 20, Hoge Veluwe: 2009 n = 29; Table 1 ). Thus, 34 females were not taken into captivity; these control females were put back in the nest box after being ringed and weighed with a pesola spring scale ( Table 1 ). The females taken into captivity were housed indoor in individual cages with ad libitum food and water.
Half of the females were kept indoors under a natural light regime (light from 7:30AM to 6:15PM), while the other half were photostimulated (light from 7AM to 3AM; Fig. 3 ), and thus experienced an increase in day length of 8 hours and 45 minutes. Thus, we created three experimental groups in this field experiment: (i) photostimulated females in captivity (ii) females which experienced natural day length in captivity and (iii) females which experienced natural day length in the field. We chose not to measure gonadal size of the females in the field after release since disturbance in the period just before egg laying could probably affect the timing of egg laying. All females were released at the field site of capture the day after the treatment (day 3, Fig. 3 ).
Nest boxes at both study sites were checked once a week from the beginning of April onwards to monitor nest building. Once the bottom of the box was covered with nest material, nests were checked daily to determine the exact laying date (date the first egg was laid). During incubation, females were identified by their unique colour code combination by exposing the colour rings with a pen. Laying dates of three females were excluded since they were likely to be replacement clutches (>30 days after the first egg laying dates in that year of that population) of which we had missed the first clutch (removed laying dates are: Oosterhout 2009: April 40 th ; Oosterhout 2010: April 34 th and April 61 st ). Results of the analysis did not change after including these data points.
Statistical analyses
We used a One-sample Wilcoxon signed rank test to test if follicles grew for the photostimulated females in the aviary experiment and a Two-sample Wilcoxon signed rank test to test if gonadal growth was different for females which were photostimulated either once or twice. Statistics presented here were done on the follicle volumes. Results did not change when using follicle length as dependent variable. To test for differences in laying dates in the field experiment we used ANOVAs. Because the experiment was done in Oosterhout in 2009 and 2010 and in the Hoge Veluwe population only in 2009, we divided the analyses of the field experiment in two parts; comparison between study areas in 2009 and comparison between years for the Oosterhout population. All statistics were done using R 2.9.2 [33] .
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Bioethanol produced from lignocellulosic biomass is a prospective and attractive substitute for petroleum-based liquid fuels as a sustainable and renewable source of energy [1] . Consolidated bioprocessing (CBP) is a leading option for the conversion of biomass to biofuel because this technique effectively simplifies the bioconversion process and reduces the cost of bioethanol production by integrating cellulase production, biomass saccharification, and sugar fermentation into a single step [2] . Clostridium thermocellum , a gram-positive thermophilic anaerobic bacterium, has been proposed as a potential candidate microorganism for bioethanol CBP production due to its high efficiency of cellulose degradation and direct production of ethanol [2] , [3] . However, the industrial application of C. thermocellum has been hampered because of its low hemicellulose utilisation, low ethanol productivity and titre, and low ethanol tolerance [4] – [6] . Various strategies, including co-cultivation with other bacteria and metabolic engineering, have been developed to improve the ethanol production and hemicellulose utilization of C. thermocellum [7] – [9] . However, the low ethanol tolerance of C. thermocellum is still one of the major bottlenecks of its bioethanol industrialisation.
Many efforts have been made to understand the ethanol tolerance mechanism and to improve the ethanol tolerance of C. thermocellum . Some ethanol tolerant (ET) strains have been obtained by gradual ethanol-adaption growth, genetic engineering, and chemical or UV mutagenesis [5] , [10] , [11] . Using these ET strains, the mechanism of ethanol tolerance of C. thermocellum has been studied via analysis of membrane composition and structure [12] , membrane proteomic profile [13] , and genome sequence [5] , [10] . Changes in the membrane proteins of C. thermocellum ethanol-tolerant strains, as well as fatty acid composition which have resulted in an increase of membrane rigidity, have been observed [12] , [13] . A recent work indicated that increased membrane fluidity is not the sole adverse effect caused by ethanol resulting in growth inhibition, and ethanol might also denature proteins thus affecting bacterial metabolism [14] . The genome sequencing of C. thermocellum ET strains revealed a large number of mutation sites, and a mutation in an alcohol dehydrogenase-encoding gene appears to confer most of the ET phenotypes [5] , [10] . The mutation changes the co-factor specificity of the alcohol dehydrogenase, but net ethanol oxidation does not appear to be a major detoxification mechanism [10] . Therefore, it is still not clear why a change of the co-factor specificity results in ethanol tolerance, and the contribution of other mutation sites to ethanol tolerance cannot be ruled out [10] . Some ET strains of C. thermocellum grow slower than the wild-type (WT) strain, and the yield of ethanol in ET strains is often lower than that in the WT strain [10] , [12] , [13] . These previous studies on the ET strains of C. thermocellum suggest that the metabolism in ET strains is significantly changed compared with the WT strain.
The “-omics” technologies, such as genomic, transcriptomic, proteomic, and metabolomic profiling, can provide deep insight into the molecular mechanisms of certain phenomena or responses [10] , [13] , [15] , [16] . In this work, to understand the mechanism of ethanol tolerance of C. thermocellum more deeply, we adopted systematic metabolomics to compare the intracellular and extracellular polar-/fatty-phase metabolites in C. thermocellum WT and ET strains in the absence and presence of 3% (v/v) exogenous ethanol (designated as ET 0 and ET 3 , respectively) using nuclear magnetic resonance (NMR), gas chromatography-mass spectroscopy (GC-MS), and ion chromatography (IC). Significant differences in the levels of many metabolites were observed in our analysis, which sheds new light on the mechanism of ethanol tolerance of C. thermocellum and provides new clues and strategies for metabolic and fermentation engineering to improve the ethanol tolerance and production of C. thermocellum .
Materials and Methods
Chemicals
Sodium chloride, K 2 HPO 4 ·3H 2 O, NaH 2 PO 4 ·2H 2 O and nonadecanoic acid (all analytical grade) were purchased from Guoyao Chemical Co. Ltd. (Shanghai, China) and used without further treatments. D 2 O (99.9% in D) and 4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS) were purchased from Cambridge Isotope Laboratories (Miami, USA). Amino acid standards for IC analysis were purchased from AccuStandard Inc. (New Haven, Connecticut USA). Carbohydrate and organic acid standards for IC analysis were purchased from Sigma-Aldrich (St. Louis,USA). Cation standards (including K + , Na + , Mg 2+ , Ca 2+ and NH 4 + ) for IC analysis were purchased from the Shanghai Institute of Measurement and Testing Technology (Shanghai, China).
Strain, Medium and Ethanol Stress Treatments
The wild-type and ethanol-tolerant strains of C. thermocellum ATCC 35609 were kindly provided by Prof. Jian Xu (Qingdao Institute of Bioenergy and Bioprocess Technology, Chinese Academy of Sciences). The culture media were derived from GS-2 media [17] with minor modifications (g/L, KH 2 PO 4 1.0, K 2 HPO 4 ·3H 2 O 5.0, Urea 1.0, MgCl 2 ·6H 2 O 2.5, CaCl 2 ·2H 2 O 0.05, FeSO 4 ·7H 2 O 0.00125, cysteine hydrochloride 3.0, cellobiose 10.0, MOPS 6.0, yeast extract 10.0, Na 3 C 6 HO 7 ·2H 2 O 3, and redox indicator resazurin 0.002, pH 7.6). All media were prepared in an anaerobic cabinet with an atmosphere of mixed gases (10% CO 2 , 5% H 2 and 85% N 2 ), and the cultivations were conducted in 1000 mL anaerobic bottles with 300 mL, 400 mL and 600 mL of fresh medium for the WT, ET 0 and ET 3 cultivations, respectively. All cultures were grown at 60°C. The biomass was determined by optical density (OD 600 ) and dry cell weight (DCW) in triplicate.
Sample Pretreatment and Metabolite Extraction
The cultivations were stopped at the late logarithmic phase by cooling in an ice-water bath for a half hour to quench metabolic activity, and the metabolites in the samples before and after quenching were checked by NMR to ensure there was no difference ( Figure S1 ). The cells were then harvested by centrifugation (4,000 g, 10 min, 4°C). The supernatants were frozen at −80°C as the extracellular metabolite samples for future IC analysis. The cell pellets were washed with phosphate buffer (137 mM NaCl, 2.7 mM KCl, 10 mM Na 2 HPO 4 , and 2 mM KH 2 PO 4 ) 3 times.
The intracellular polar metabolites were extracted using a boiling ethanol method [18] . The precipitates were used for further metabolite extraction in fatty phase. The polar extracts were concentrated and dried by rotary evaporation and lyophilisation, and the obtained extract powder was used for further NMR analysis. The fatty phase metabolites were extracted by chloroform from the precipitates of the previous boiling ethanol extraction. Briefly, 200 mg of precipitate was added into 2.5 mL of chloroform and shaken at 160 rpm for 2 h at 37°C. The supernatant was separated by centrifugation at 9,000 g for 10 min at 4°C, and the pellet was re-extracted by adding 2.5 mL of chloroform and 1.5 mL of methanol. The supernatants from the two extractions were combined and subjected to evaporation using a rotary evaporator. Finally, 1.0 mL of chloroform was added to the dried extracts, and then the extracts were flushed with nitrogen. The extracts were kept at −80°C before GC-MS analysis.
Metabolite Analysis
Intracellular polar metabolites determined by NMR
Approximately 40 mg of extracted powder of polar metabolites was added into 500 µL of 100% D 2 O together with 100 µL of phosphate buffer (100 mM, pH7.4) containing 10% D 2 O and 0.02 mM DSS. The mixture was centrifuged at 4°C (16,100 g, 5 min), and the obtained supernatant was transferred into a 5-mm NMR tube for NMR analysis. Two blanks were always prepared in parallel during extraction. All 1 H-NMR spectra were recorded at 298 K using a Bruker 600 NMR spectrometer (600.13 MHz for 1 H) equipped with a 5-mm triple resonance cryoprobe (Bruker Biospin, Germany). A standard one-dimensional noesypr1d pulse sequence was used to quantify the intracellular polar metabolites. Water suppression was achieved with a weak irradiation during the recycle delay (2 s) and the mixing time (100 ms). 64 transients were collected into 32768 data points for each spectrum with a spectral width of 12 kHz. An exponential window function with a line broadening factor of 0.5 Hz was applied to all free induction decays prior to Fourier transformation. For resonance assignment purposes, two-dimensional 1 H- 1 H TOCSY, 1 H- 1 H COSY, 1 H- 13 C HSQC and 1 H- 13 C HMBC spectra were acquired. In the COSY and TOCSY experiments, 48 transients were collected into 2048 data points for each of 256 increments, and the spectral widths were 6 kHz for both dimensions. Phase-insensitive mode was used with gradient selection for the COSY experiments, whereas the MLEV-17 was employed as the spin-lock scheme in the phase-sensitive TOCSY experiment with a mixing time of 100 ms. 1 H- 13 C HSQC and HMBC spectra were recorded using gradient selected sequences with 200 transients and 2048 data points for each of 128 increments. The spectral widths were 6 kHz for 1 H and 26 kHz (HSQC) or 33 kHz (HMBC) for 13 C. The data were processed into a 4096 by 2048 matrix via Fourier transformation.
The quantification of the intracellular polar metabolites was obtained from analysis of the 1 H-NMR spectra. These spectra were manually corrected for phase and baseline distortions using TOPSPIN (Bruker Biospin, Germany), and the spectral region (δ = 0.5–9.5) was uniformly integrated into 3166 buckets with a width of 0.003 ppm (1.8 Hz) using the AMIX package (Bruker Biospin, Germany). A region (δ = 4.67–4.90) was discarded to eliminate the effects of imperfect water presaturation. The spectral areas of each bucket were normalised to the internal reference (DSS). The absolute levels of the metabolites were calculated, as milligram per gram freeze-dried metabolite extracts, from the least overlapping NMR signals of the metabolites and the relevant DSS values with known concentrations. These semi-quantitative data were expressed in the form of the mean ± standard deviation and were also subjected to classical one-way ANOVA analysis using SPSS 13.0 software and a Turkey post-test ( p <0.05).
Intracellular fatty metabolites determined by GC-MS
The total lipid in the fatty phase extracts were converted to methyl esters using 2.5 mL of 2% H 2 SO 4 -methanol (v/v) with an internal standard (nonadecanoic acid, C19:0) for quantitative determination. The fatty acid methyl esters (FAMEs) were analysed with an Agilent 7890-5975 GC-MS system (Agilent Technologies Inc., Santa Clara, USA) equipped with a 30 mm×0.25 mm×0.25 µm capillary column (Agilent HP-INNOWAS). The oven temperature was initially 100°C and increased up to 240°C over 10 min (15°C min −1 ). The split ratio was 1:20, and helium was used as the carrier gas at a flow rate of 1.0 mL min −1 in constant flow mode. The ion source and quadrupole temperatures were 230°C and 150°C, respectively. The mass spectrometer was operated in electron impact mode at 70 eV with a scan range of 30–400 m/z. The injection sample volume was 1.0 µL.
Extracellular metabolite analysed via ion chromatography
The frozen extracellular metabolite samples were thawed and filtered using a filter with a 0.22-µm pore size. Four types of extracellular targeted metabolites (i.e., amino acids, carbohydrates, organic acids and cations) were identified using an ICS-3000 system and an ICS-5000 system (Dionex Corporation, Sunnyvale, USA). The standard curves of 20 amino acids, 10 carbohydrates, 12 organic acids, and 5 cations were measured and used to quantify the corresponding extracellular metabolites. The concentrations of the metabolites in an uncultured medium were also measured as a negative control. The negative control concentrations were subtracted from the concentration of extracellular metabolites in the samples. Therefore, the extracellular metabolite contents are presented as positive or negative values, which indicate released or assimilated metabolites, respectively.
Results and Discussion
The Growth Patterns of Wild-type and Ethanol-tolerant Strains of C. thermocellum were Different
The growth and dry cell weight curves of the WT, ET 0 and ET 3 C. thermocellum cultivations are shown in Figure S2 . Remarkable differences were observed among these three cultivations. Both the ET 0 and ET 3 cultivations displayed longer lag phases and lower maximal biomass yields than that of WT, whereas the maximal biomass of ET 3 was only around half of that of ET 0 . These results indicate that cell growth was inhibited by ethanol, which was proven in a previous report [10] . The noteworthy divergence of macroscopic growth patterns might imply great differences of microscopic metabolic profiles.
Differential Composition and Abundance of Intracellular Metabolites in WT, ET 0 and ET 3 C. thermocellum Cultivations were Detected
Intracellular metabolome analysis, referred to as metabolic fingerprinting, can directly provide the cellular metabolic profile and metabolic changes in specific environments [19] , [20] . We analysed both the polar and fatty phase intracellular metabolites of the three phenotypes of C. thermocellum to investigate the intracellular ethanol-tolerant strain metabolic changes.
The polar metabolites
The polar metabolites of the WT, ET 0 and ET 3 C. thermocellum cultivations were analysed by NMR ( Figure 1 ). The metabolite resonances ( Table S1 ) were assigned with 1D and 2D NMR data and were further confirmed with literature [21] – [23] and publicly available database data [24] . Using these data, 39 putative metabolites could be detected ( Figure 1 and Table S1 ), of which 16 were identified and quantified ( Table 1 ). These metabolites included amino acids, carbohydrates, organic acids/amines/alcohol, and nucleotide derivatives that are involved in various metabolic pathways ( Figure 2 ). As shown in Table 1 and Figure 1 , considerably different metabolic profiles of C. thermocellum were observed for the WT, ET 0 and ET 3 . Most WT and ET 0 metabolites displayed higher levels than those in ET 3 . This phenomenon suggested that ethanol significantly inhibited the metabolism of C. thermocellum even for the ET strain, which is in accordance with the observed changes in the C. thermocellum growth pattern ( Figure S2 ).
10.1371/journal.pone.0070631.g001 Figure 1
1 H NMR spectra of the intracellular polar metabolites of C. thermocellum .
The spectra of the wild type strain (WT) and the ethanol-tolerant strain without additional ethanol (ET 0 ) and with 3% ethanol (ET 3 ) were shown from bottom to top. The left halves (4.90–9.50 ppm) of all spectra are magnified two times in the Y-axis for clarification. Keys: 1, valine; 2, lactic acid; 3, threonine; 4, acetic acid; 5, glutamate; 6, pyruvate; 7, succinate; 8, dimethylamine; 9, norspermidine; 10, malonate; 11, ethanol; 12, cellodextrin; 13, phosphoenolpyruvate; 14, L-erythrose; 15, uridine monophosphate; 16, adenosine; 17, nicotinate; 18, thymidylic acid; 19, Adenosine monophosphate; 20, nicotinamide adenine dinucleotide; 21, nicotinamide adenine dinucleotide phosphate; 22, uracil; 23, uridine monophosphate; 24, inosine; 25, formate; 26, adenosine diphosphate; 27, adenosine triphosphate; 28, tyrosine; 29, tryptophan; 30, aspartate; 31, α-arabinose; 32, methanol; 33, fumarate; 34, guanine; 35, cytosine; 36, acetamide; 37, p-aminobenzoic acid; 38, uridine diphosphate glucose; 39, α-D-galactose-1-phosphate.
10.1371/journal.pone.0070631.g002 Figure 2
Metabolic pathway representation of intramolecular metabolite change.
Significantly changed intramolecular polar metabolites in C. thermocellum wild-type strain (WT) and ethanol-tolerant strain without additional ethanol (ET 0 ) and with 3% ethanol (ET 3 ) were mapped on the existing metabolic pathways. The red and green symbols indicate significantly increased and decreased metabolites, respectively. Metabolites: G1P, glucose 1-phosphate; G6P, glucose 6-phosphate; F6P, fructose 6-phosphate; G3P, glyceraldehyde 3-phosphate; E4P, erythrose 4-phosphate; R5P, ribose 5-phosphate; PEP, phosphoenolpyruvate; OAA, oxaloacetic acid; α-KG, α-ketoglutaric acid; Glu, glutamic acid; Gln, glutamine; Arg, arginine; Citru, citrulline; Orn, ornithine; Put, putrescine; Nors, norspermidine; FA, fatty acid. Enzymes: ACK, acetate kinase; ALDH/ADH, acetaldehyde dehydrogenase/alcohol dehydrogenase; CDPase, cellodextrin phosphorylase; CBPase, cellobiose phosphorylase; HK, hexokinase; PGM, phosphoglucomutase; LDH, lactate dehydrogenase; MDH, malate dehydrogenase; PTA, phosphotransacetylase; PFL, pyruvate:formate lyase; PFO, pyruvate:ferredoxin oxidoreductase; Fd H2ase, ferredoxin hydrogenase; NAD(P)H H2ase, NAD(P)H hydrogenase.
10.1371/journal.pone.0070631.t001 Table 1
Quantification of the intracellular polar metabolites from the WT, ET 0 and ET 3 cultivations of C. thermocellum .
Metabolites (integral range, ppm)
Relative contents (mg/g dry-weight metabolites)
WT
ET 0
ET 3
NAD (9.30, 9.33)
1.26±0.18 a
2.22±0.37 b
0.25±0.14 c
NADP (9.27, 9.30)
0.21±0.05 a
0.18±0.04 a
0.05±0.04 b
nicotinate (7.57, 7.64)
0.04±0.02 a
0.04±0.01 a
0.01±0.01 b
tyrosine (6.86, 6.89)
0.50±0.16 a
0.17±0.11 b
0.14±0.02 b
fumarate (6.50, 6.51)
0.02±0.00 a
0.02±0.01 a
0.01±0.00 b
UDPG (5.59, 5.63)
0.64±0.10 a
0.88±0.11 b
0.19±0.07 c
Ery (5.28, 5.30)
0.32±0.05 a
0.33±0.04 a
0.11±0.04 b
PEP (5.17, 5.19)
0.29±0.05 a
0.35±0.03 b
0.07±0.02 c
cellodextrin (3.90, 3.98)
55.59±3.93 a
108.35±15.26 b
40.89±16.76 c
norspermidine (2.95, 3.01)
1.61±0.12 a
1.97±0.31 b
0.55±0.11 c
DMA (2.71, 2.73)
0.78±0.27 a
0.16±0.09 b
0.11±0.01 b
α-KG (2.41, 2.45)
1.94±0.37 a
1.25±0.86 b
0.48±0.04 c
succinate (2.38, 2.39)
0.86±0.21 a
0.63±0.32 b
0.19±0.03 c
pyruvate (2.35, 2.36)
0.87±0.22 a
0.27±0.19 b
0.13±0.01 b
glutamate (2.36, 2.32)
3.06±0.77 a
1.25±0.77 b
0.65±0.10 b
acetate (1.89, 1.91)
2.96±1.04 a
0.88±0.56 b
0.42±0.05 b
Abbreviated metabolites: NAD, nicotinamide adenine dinucleotide; NADP, nicotinamide adenine dinucleotide phosphate; UDPG, uridine diphosphate glucose; Ery, L-erythrose; PEP, phosphoenolpyruvate; DMA, dimethylamine; α-KG, α-ketoglutaric acid. a,b,c Different letters indicated statistical significance (P<0.05) from one-way ANOVA with a Turkey post-test.
Based on extensive 2D NMR data and literature reports [23] , cellodextrins were confirmed as the most abundant metabolite in the quantified metabolites of the WT, ET 0 and ET 3 cultivations (55.59±3.93, 108.35±15.26 and 40.89±16.76 mg/g freeze-dried weight of extracted metabolites, respectively, assuming a degree of polymerisation of 4). The accumulation of cellodextrins in C. thermocellum was observed in previous studies. C. thermocellum has been demonstrated to be able of assimilating cellodextrins hydrolysed from extracellular cellulose [25] – [27] , but when cellobiose is the sole carbon source in culture media, cellodextrins are believed to be biosynthesised from cellobiose by the reverse activity of cellobiose phosphorylase and cellodextrins phosphorylase [28] , [29] . However, the concentrations of cellobiose in the extracellular metabolites were quite low in both the WT or ET strains (see section of the extracellular metabolites below), which indicates that the cellodextrins accumulated in the cells cannot be explained by the reverse activity of phosphorylases when the cellobiose in the media was almost completely taken up. These results indicate that the high-concentration accumulation of cellodextrins in cells was likely an active process, and the cellodextrins potentially serve as energy storage in cells. The ET 0 C. thermocellum harboured a higher concentration of cellodextrins than the WT, and the cellodextrin level in the ET 3 bacteria was similar to that in WT, although the other metabolites in the ET 3 bacteria were significantly decreased in comparison with the WT, which suggests the synthesis of cellodextrins in the ET strain was more active than that in the WT stain and perhaps indicates a new metabolic regulation mechanism responded to the ethanol. To check the effect of cellodextrins in C. thermocellum , the growths of the ET strain in absence and presence of additional cellodextrins were compared ( Figure S3 ). Cellodextrins caused the delay of growth of ET strain, but the highest biomass was ∼ 5–10% higher than that in the absence of cellodextrins. The delay of growth was similar to the response to ethanol stress ( Figure S2 ), whereas the higher biomass indicated the ethanol tolerance was indeed improved by cellodextrins. Further studies are needed to understand the internal mechanism of intracellular cellodextrins.
When comparing the metabolite concentrations in the ET 0 and ET 3 cultivations, nicotinamide adenine dinucleotide (NAD) was the most significantly changed metabolite ( Table 1 ). The level of NAD in the ET 3 cultivation decreased to approximately one-ninth of that in the ET 0 cultivation, whereas the level of nicotinamide adenine dinucleotide phosphate (NADP) in the ET 3 cultivation was approximately one-third of that in ET 0 cultivation. These results indicate that the biosynthesis of NAD, compared to the biosynthesis of NADP, was obviously more affected by ethanol. The low level of NAD could be disadvantageous for many NAD-specific enzymes that can be very important in the process of cell growth, such as alcohol dehydrogenase, although NADP is generally more important to cell growth. Thus, this result might partially explain why the co-factor specificity change from NAD to NADP of alcohol dehydrogenase could influence the ethanol tolerance as previously reported [10] . Further studies are needed to reveal the molecular mechanism resulting in the NAD level change, which may provide a new strategy to increase the ethanol tolerance of C. thermocellum .
The fatty phase metabolites
Fatty phase metabolites are generally the major components of the cell membrane, and thus the composition of fatty phase metabolites could provide important information about the membrane, such as the fluidity, which is considered to be a key factor in ethanol tolerance [12] . We determined the fatty acid composition of the fatty phase metabolites of C. thermocellum WT, ET 0 and ET 3 cultivations. In total, 17 metabolites were identified and quantified by GC-MS ( Figure 3 and Table 2 ). Similar to the previous results in literature [12] , the fatty phase metabolites were mainly fatty acids and plasmalogens, and the percentage of long chain fatty acids (> = 16:0) in the ET strain was higher than in the WT strain. We further detected this increasing tendency in the ET 3 compared with ET 0 cultivations, and the changes in fatty acid composition between the ET 3 and ET 0 cultivations were much less than those between the ET 0 and WT cultivations (maximal and minimal ratios for various fatty acids were 337.24% and 22.03% for ET 0 /WT, and 139.97% and 43.79% for ET 3 /ET 0 ). These results indicate that the adjustments of membrane composition under stress conditions were relatively slow compared with other metabolite changes, but some adjustments were kept in ET stains during long-term adaption. Furthermore, we observed that some fatty acids with odd carbon numbers (15:0, 17:0, P n-15:0, P i-17:0) displayed reversed changes when comparing ET 0 /WT and ET 3 /ET 0 . The percentage of these fatty acids among the total fatty acids was relatively low, so these changes might have little effect on the membrane but still reflect the responses of fatty acid biosynthesis to ethanol stress.
10.1371/journal.pone.0070631.g003 Figure 3
GC-MS spectrum of C. thermocellum for fatty phase metabolites assignment.
Keys: 1, cis-13-eicosenoic acid (20:1); 2, 10-methyl-undecanoic acid (i-11:0); 3, 12-methyl-tridecanoic acid (i-13:0); 4, tetradecanoic acid (14:0); 5, plasmalogen n-14:0 (P n-14:0); 6, 13-methyl-tetradecanoic acid (i-14:0); 7, plasmalogen n-15:0 (P n-15:0); 8, pentadecanoic acid (15:0); 9, plasmalogen n-16:0 (P n-16:0); 10, 14-methyl-pentadecanoic acid (i-15:0); 11, plasmalogen i-17:0 (P i-17:0); 12, n-hexadecanoic acid (16:0); 13, 15-methyl-hexadecanoic acid (i-16:0); 14, 14-methyl-hexadecanoic acid (i-16:0); 15, heptadecanoic acid (17:0); 16, 16-methyl-heptadecanoic acid (i-17:0); 17, octadecanoic acid (18:0); 18, nonadecanoic acid (C19:0).
10.1371/journal.pone.0070631.t002 Table 2
The fatty phase metabolites of the WT, ET 0 and ET 3 cultivations of C. thermocellum .
Metabolites
Phenotype (FA/total FA%)
Percent (%)
WT
ET 0
ET 3
ET 0 /WT
ET 3 /ET 0
cis -13-eicosenoic acid (20:1)
2.54±0.40
1.19±0.21
0.65±0.11
46.90 *
54.17
10-methyl-undecanoic acid (i-11:0)
0.34±0.08
0.21±0.06
0.16±0.03
61.95
77.14
12-methyl-tridecanoic acid (i-13:0)
0.38±0.08
0.21±0.05
0.16±0.04
56.55
74.86
tetradecanoic acid (14:0)
4.19±0.21
1.32±0.06
0.73±0.06
31.53 *
55.13 *
plasmalogen n-14:0 (P n-14:0)
0.74±0.07
0.52±0.19
0.33±0.08
69.86
64.97
13-methyl-tetradecanoic acid (i-14:0)
1.18±0.24
1.12±0.32
0.64±0.12
95.54
56.81
plasmalogen n-15:0 (P n-15:0)
0.70±0.17
1.39±0.16
0.61±0.18
196.47 *
43.79 *
pentadecanoic acid (15:0)
0.27±0.09
0.89±0.14
0.44±0.05
337.24 *
48.86 *
plasmalogen n-16:0 (P n-16:0)
0.49±0.17
0.52±0.22
0.49±0.19
107.19
93.64
14-methyl-pentadecanoic acid (i-15:0)
14.78±0.36
3.26±0.39
2.48±0.52
22.03 *
77.93
plasmalogen i-17:0 (P i-17:0)
5.31±0.89
1.61±0.26
1.88±0.38
29.48 *
120.32
n-hexadecanoic acid (16:0)
36.39±0.89
40.26±2.56
41.18±2.70
110.64
102.28
15-methyl-hexadecanoic acid (i-16:0)
9.65±2.87
14.92±2.94
16.23±4.33
154.62
108.75
14-methyl-hexadecanoic acid (i-16:0)
2.11±0.33
2.96±0.51
2.96±0.88
140.12
100.17
heptadecanoic acid (17:0)
0.51±0.13
1.18±0.15
0.88±0.17
229.73 *
74.79
16-methyl-heptadecanoic acid (i-17:0)
1.13±0.20
2.40±0.35
3.36±0.65
212.27 *
139.97
octadecanoic acid (18:0)
19.30±1.84
26.08±2.84
26.76±2.56
135.12
102.64
* The significant differences are derived from a one-way ANOVA analysis (p<0.05).
The Extracellular Metabolites of C. thermocellum WT, ET 0 and ET 3 Cultivations
The analysis of extracellular metabolites, known as metabolic footprinting, is critical for understanding microbial metabolism because extracellular metabolites are related to nutrient uptake, communication between cells and environmental factors, and metabolic end products, some of which may be particularly valuable in industrial biotechnology [20] . We identified and quantified 19 amino acids, 6 carbohydrates, 9 organic acids and 5 cations in the extracellular metabolites of the WT, ET 0 and ET 3 cultivations by IC ( Table 3 ), and these metabolites represented the majority of extracellular metabolites in the media. Because we used a complex media in the cultivations, control experiments were performed to ensure that the concentrations of metabolites were not influenced by secretory enzymes (Table. S2). For most extracellular metabolites, the differences between the ET 0 and WT cultivations were more significant than those between ET 3 and ET 0 with the exception of cellobiose, the major carbon source in the media (10 g/l), which was almost completely consumed in all three types of cultivations. The biomass of the ET cultivations was much less than that of the WT cultivation ( Figure S2 ), and therefore, additional carbon sources in the ET cultivations should have been converted into extracellular metabolites. Consistently in our results, the concentrations of most extracellular metabolites in the ET cultivations were higher than those in the WT cultivation.
10.1371/journal.pone.0070631.t003 Table 3
The extracellular metabolites of the WT, ET 0 and ET 3 cultivations of C. thermocellum .
Metabolites a
Phenotype (ppm)
Differences
WT
ET 0
ET 3
ET 0 vs. WT
ET 3 vs. ET 0
Amino acid
arginine
−24.06±1.68 b
14.15±1.48
23.80±2.72
* c
*
lysine
14.32±0.35
19.96±1.05
19.88±1.10
*
– d
asparagine
−0.20±0.70
1.11±0.45
1.04±0.58
*
–
glutamine
9.06±1.10
12.70±0.80
12.65±0.59
–
–
alanine
18.74±0.93
28.30±0.87
35.22±1.29
*
–
threonine
−2.53±0.20
−0.55±0.39
−0.38±0.23
*
–
glycine
8.81±0.51
7.97±0.26
7.86±0.51
–
–
valine
80.35±3.03
98.42±3.95
104.54±3.35
*
–
serine
0.93±0.03
1.69±0.11
1.84±0.11
*
–
proline
26.08±0.90
29.94±0.91
29.39±1.40
–
–
isoleucine
7.67±0.85
11.84±0.66
15.33±0.70
*
*
leucine
6.20±1.34
20.44±1.43
32.02±1.45
*
*
methionine
−0.45±0.01
−0.45±0.01
−0.21±0.01
–
*
histidine
1.59±0.12
2.52±0.17
2.24±0.17
*
–
phenylalanine
−3.57±0.14
−3.93±0.10
−3.52±0.12
–
–
glutamic acid
−33.32±1.07
−32.83±0.95
−33.08±0.79
–
–
aspartic acid
−28.29±1.50
−27.72±1.14
−25.07±1.18
–
–
cysteine
23.63±1.33
57.74±4.09
34.97±1.99
*
*
tyrosine
−2.78±0.10
7.32±0.37
5.47±0.32
*
–
Carbohydrates
trehalose
−70.30±5.74
−56.97±4.02
−16.76±2.40
–
*
glucose
−360.44±25.53
−245.23±17.97
−34.39±4.05
*
*
mannose
−9.64±0.47
−5.03±0.92
−3.27±0.61
*
*
ribose
38.97±1.51
10.95±3.04
−2.51±0.31
*
*
cellobiose
−8551.70±14.74
−8548.20±17.63
−8555.90±13.14
–
–
panose
−69.97±3.73
−66.22±4.31
−43.33±6.27
–
*
Organic acids
lactic acid
501.50±57.57
749.09±77.40
417.23±56.99
*
*
acetic acid
989.36±48.89
1306.56±42.95
1000.75±65.92
*
*
propionic acid
128.66±23.36
123.25±33.50
119.16±36.22
–
–
glyoxylic acid
391.48±34.44
360.62±27.8
130.51±25.61
–
*
pyruvic acid
−62.67±12.10
−62.67±19.87
−62.67±13.98
–
–
malic acid
18.52±2.66
40.97±5.83
52.08±5.15
*
–
fumaric acid
1.38±0.35
1.69±0.42
1.70±0.31
–
–
dihydroxyacetone phosphate
611.11±71.61
946.18±65.28
809.22±51.32
*
–
citric acid
−363.68±49.73
−293.94±46.65
−363.13±29.76
–
–
Cations
Na +
−1339.22±97.41
−1097.55±75.90
−565.13±45.79
*
*
NH 4 +
213.36±24.10
98.20±9.08
156.08±28.57
*
*
K +
−427.36±74.07
−1214.52±71.79
−763.65±85.92
*
*
Mg 2+
−164.73±12.77
−182.86±23.05
−193.36±31.88
–
–
Ca 2+
−6.94±0.23
−9.99±0.48
−10.44±0.33
*
–
a Metabolites that were not detected among all three cultivations and control medium are not listed in the table. These metabolites included tryptophan, arabinose, galactose, lactose, malonic acid, isocitric acid, and α-ketoglutarate. b Positive and negative values indicate amount of the released and absorbed metabolites, respectively. c The significant differences were derived from a one-way ANOVA analysis (p<0.05). d No significant difference was derived from a one-way ANOVA analysis (p<0.05).
It has been reported that C. thermocellum produces a high concentration of extracellular free amino acids [30] . In our experiments, the concentrations of 17 of 19 detected amino acids (tryptophan was not detectable in both the control medium and cultured media) in the ET cultivations were higher than those in the WT cultivation, with two exceptions, glycine and phenylalanine. Eleven secreted amino acids (the three most abundant were valine, proline and cysteine) were secreted in greater amounts by the ET strain than by the WT strain except glycine. Eight amino acids were consumed by the WT strain, 3 of which were secreted by the ET strain in contrast, whereas the other 4 amino acids, except phenylalanine, were less consumed by the ET strain. The most interesting amino acid was arginine because it exhibited the most significant change among the reversed amino acids (i.e., consumed to secreted), and the ET strain might contain mutations in the arginine biosynthesis pathway as previously reported [5] . Further studies are needed to confirm whether the mutation resulted in the overproduction and secretion of arginine in ET cultivations.
In addition to cellobiose, several carbohydrates including trehalose, glucose, mannose, and panose in media were also utilised by C. thermocellum ( Table 3 ). However, unlike cellobiose, the utilisation of these sugars in the ET strain was much less than those in the WT strain. The only secreted sugar in the WT strain was ribose, but much less ribose was secreted in the ET 0 cultivation and a small amount of ribose was taken up in the ET 3 cultivation. Ribose is a key metabolite involved in the pentose phosphate pathway, so these changes of the extracellular ribose concentration indicated that pentose phosphate pathway was significantly affected in the ET strain.
The organic acids in metabolites include both final products and intermediates ( Table 3 ). The detected final products included lactic acid, acetic acid, and propionic acid, and all of them were released into media. Lactic acid and acetic acid were the majority of the released final metabolites. Interestingly, the concentrations of lactic acid and acetic acid in the ET 0 cultivation were much higher than those in both the WT and ET 3 cultivations, unlike the released amino acids, which were released maximally in the ET 3 cultivation. Two consumed organic acids, citric acid and pyruvic acid, were important intermediates in central metabolism. Citric acid plays key role in the TCA cycle, whereas pyruvic acid is a pivot that links glycolysis and the TCA cycle. Their uptake might save energy in the metabolism, although the initial purpose of the addition of citric acid was to prevent the precipitation of salts [17] . However, other 4 metabolic intermediates, including glyoxylic acid, malic acid, fumaric acid, and dihydroxyacetone phosphate, were significantly released into the media in all three cultivations. The phenomenon of metabolic intermediate overflow has been observed for malic acid in a previous study [31] , and our results indicated that the overflow occurred for many metabolic intermediates. It has been reported that end-product-induced metabolic shifts in C. thermocellum [32] and the addition of ethanol changed the balance of metabolism. It is worth noticing that the metabolic overflow was more significant in the ET 0 than ET 3 cultivations, which suggests that the ET strain was adapted to grow with better metabolic balance in ethanol stress environments.
Five cations were detected in this study, 4 of which were taken up by C. thermocellum , whereas ammonium was released from cells ( Table 3 ). The two divalent cations, Mg 2+ and Ca 2+ , were consumed more in the ET strain than the WT strain, and the uptakes in the ET 3 cultivation were more than those in the ET 0 cultivation. This may be caused by the up-regulation of divalent cation transporter, as reported in a previous study [13] , which is considered to be a common phenomenon in ethanol-tolerant microorganisms, and supplementation of these divalent cations may be a strategy to enhance ethanol productivity [33] . The two major monovalent cations, K + and Na + , displayed different changes in the ET stain compared with the WT strain. In the ET strain, less Na + but more K + was taken up into cells. Because Na + was pumped out and K + was taken up into cells to maintain the membrane potential [34] , [35] , the ET strain should have higher membrane potential than the WT strain.
The extracellular metabolome of C. thermocellum provided abundant information about the nutrient uptake, metabolism, and membrane properties. Our data indicated that the major nutrient (cellobiose) was taken up completely in the ET strain but it was not converted into biomass or ethanol. Instead, greater amount of other metabolites, including amino acids and some organic acids, were produced in the ET strain, and some intermediate metabolites were more significantly overflowed. Therefore, the metabolic balance in C. thermocellum was not optimised for ethanol production, which might partially explain why the deletion of some by-product pathways only slightly affects the ethanol production in C. thermocellum [36] , [37] . Careful examination and re-design of some metabolic pathways in C. thermocellum are needed to obtain an engineered strain with a high ethanol yield. For example, a recent successful study reported that the introduction of an exogenous pyruvate kinase increased the ethanol yield over 3-fold in C. thermocellum with enhanced ethanol tolerance [38] . However, the ethanol yield in the study was up to 38.8 mM ethanol, which was still far below the ethanol tolerance level of the C. thermocellum WT strain. Therefore, more engineering studies, including both introducing exogenous pathways and the modification of endogenous pathways, are necessary for obtaining an industrial ethanol-producing strain of C. thermocellum .
Conclusion
The ethanol-tolerant strain of C. thermocellum displayed metabolic adaption in various metabolic pathways. The changes in the intracellular polar and fatty phase metabolites and extracellular-targeted metabolites indicated that the nutrient uptake, central metabolism, membrane structure, amino acid biosynthesis, and cation transport were adapted to ethanol stress. Cellodextrins were more significantly accumulated in the ethanol-tolerant strain as a potential anti-stress mechanism. The overflow of many intermediate metabolites indicated the imbalance of metabolic flow in C. thermocellum, and the ethanol-tolerant strain displayed a better adaption under the conditions of ethanol stress compared with the wild-type strain.
Supporting Information
Figure S1
1 H NMR spectra of the extracted intracellular polar metabolites from wild-type Clostridium thermocellum before (black) and after (red) quenching experiments.
(PDF)
Figure S2
Growth (black) and dry cell weight curves (grey) of the WT, ET 0 and ET 3 C. thermocellum cultivations.
(PDF)
Figure S3
The growth curves of ethanol-tolerant Clostridium thermocellum with and without cellodextrins.
(PDF)
Table S1
The NMR assignments of intracellular polar metabolites.
(PDF)
Table S2
The change ratio of metabolites concentrations in control experiments.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
There are more 10 million admissions to U.S. prisons and jails each year [ 1 , 2 ]. More than half of those admitted have mental health problems [ 3 ]. Major depressive disorder (MDD) is the most common serious mental illness among individuals who are incarcerated [ 3 , 4 ]. A national survey of state prisoners found that 23.5% met criteria for MDD within the past year, three times the national past-year prevalence [ 3 ]. In-prison consequences of MDD can include dropout from correctional treatment programs, inability to assertively protect oneself, physical victimization, aggressive acting out, and increased suicide risk [ 5 – 9 ]. Once out of prison, MDD increases risk of recidivism [ 10 ]. Therefore, it is important to know how to implement evidence-based MDD treatments within prisons, potentially unique and challenging implementation climates [ 11 , 12 ].
Unfortunately, research guiding implementation of evidence-based mental health interventions in justice settings is sparse. More is needed. This article describes implementation processes in a randomized hybrid effectiveness-implementation trial [ 13 ] of Interpersonal Psychotherapy (IPT) for MDD in two state prison systems [ 11 , 14 ]. Results of that trial, the first fully-powered (n = 181) randomized MDD treatment trial in any incarcerated population, indicated that IPT reduced depressive symptoms, hopelessness, and posttraumatic stress disorder symptoms, and increased rates of MDD remission relative to prison treatment as usual [ 14 ]. As a result of the trial [ 14 ], which evaluated the effectiveness and cost-effectiveness of IPT for MDD in prisons, IPT became the only MDD treatment for incarcerated individuals supported by evidence from a full-scale randomized trial.
IPT was chosen because of its strong evidence-base in non-incarcerated populations [ 15 , 16 ], its fit for the target population, and pilot data suggesting that it is acceptable and effective for MDD among incarcerated women [ 17 , 18 ]. IPT’s focus on addressing recent stressors, such as interpersonal conflicts, life changes, and grief, is a good fit for incarcerated populations, who are faced with many of these disruptive life events [ 19 ].
Cost is key for mental health implementation in prisons because prison systems can often afford few licensed mental health providers [ 11 ]. Task-shifting may increase access to care. Previous studies in this line of research have demonstrated that non-specialist bachelor’s-level prison counselors can conduct IPT [ 20 ]. The trial included both master’s-level mental health professionals and non-mental health specialists (such as bachelor’s-level re-entry planners) working in the prisons as study IPT counselors [ 11 ].
The ultimate goal of this line of work [ 11 , 14 ] was to determine whether psychotherapy in prison could be done in a way (i.e., group, non-specialist counselors) that is effective and cost-effective enough to be offered in prisons regularly [ 11 , 14 ]. Trial results suggest that the answer is “yes,” if training and supervision costs can be decreased without losing effectiveness. Supervision in this trial was effective in that IPT produced better outcomes than did prison treatment usual, and prison counselors learning IPT for the trial (some of whom had no prior formal psychotherapy training) were highly adherent (spending an average of 96% of time in sessions on model) and reasonably competent (averaging 5.1 on a 1 to 7 scale) [ 14 ]. However, 72% of the costs of IPT in the trial were for training and supervision (including study supervisor and counselor hours) [ 14 ]. Given extremely sparse prison mental health training budgets, this is not scalable or sustainable. On the other hand, single workshops typically have little effect on provider competence [ 21 ]. Results of the current analysis can help optimize future training and supervision to make it more efficient for wider-scale implementation.
This article reports results from a planned qualitative analysis of 460 structured implementation and supervision documents in this randomized effectiveness-implementation trial [ 14 ] to describe training and supervision processes and lessons learned. As noted above, effectiveness and cost-effectiveness results from the trial have been published [ 11 , 14 ]. Johnson et al. [ 22 ] reported the first part of implementation analyses, addressing the first four Consolidated Framework for Implementation Research (CFIR [ 23 ]) domains that describe how characteristics of the intervention, the inner setting (i.e., implementing facilities), outer setting (i.e., policies), and individuals involved (i.e., counselors, clients) affect implementation. The current report is the third planned paper in the series, analyzing items that had been top-level coded for the fifth and final CFIR dimension, implementation processes.
This article describes two sets of analyzes applied to items top-coded within the CFIR process dimension. The first analysis summarizes implementation processes using thematic analysis of activities within the CFIR process steps (planning/engaging and executing /reflecting/evaluating). Detailed implementation/training steps are rarely published and present a contribution to the psychotherapy training literature. The second analysis identifies common themes in discussions between supervisors and study counselors. The first analysis (description of implementation/training steps) provides context for the second (common themes in what went well and what was a challenge in supervision in this trial). Taken together, this information can guide decisions to use similar processes or adapt them in the future.
Results from these analyses can inform and optimize: (1) implementation of evidence-based mental health treatments in prisons and jails, an important effort that needs more evidence to guide it; (2) psychotherapy and interpersonal psychotherapy (IPT) training efforts generally, especially in low-resource settings. Analysis of training and supervision processes and lessons learned in the first major MDD trial among incarcerated individuals can inform psychotherapy training recommendations and optimization for prisons and other low-resource settings with high mental health need. Efficient training in such settings is important to mental health access and equity. To our knowledge, this is the first detailed description and analysis of implementation processes of a mental health intervention in a prison or jail setting.
Methods
The hybrid effectiveness-implementation trial randomized 181 individuals with MDD who were in prison to IPT (n = 91) or to prison treatment as usual (n = 90) [ 11 , 14 , 22 ]. As described above, results favored IPT across multiple outcomes, IPT treatment adherence was high (96% of time on model), and IPT treatment competence was acceptable (average of 5.1 on a 1 to 7 scale) [ 14 ]. The current manuscript reports prospectively collected implementation data from prison-based IPT counselors and IPT supervisors in the trial, as they worked with the 91 clients randomized to the IPT condition. IPT training and supervision took place from November 2011 to December 2014.
Ethics approval and consent to participate
This trial was approved by Brown University’s Institutional Review Board (FWA 00004460) and regulatory bodies overseeing prison research in participating states. All study procedures were carried out in accordance with relevant guidelines and regulations. Study investigators had access to counselor and patient participant identities during the study. Written informed consent was obtained from participants before beginning study procedures.
Sources of data
Data for this analysis consisted of a systematic, planned document review of 100% of the existing documentation kept by the study team throughout the study period (2012–2014), in order to create a structured description of implementation as it occurred. Prison counselors provided IPT for the study at 3 women’s facilities and in 3 men’s medium security facilities in two northeastern U.S. states as part of the randomized trial [ 14 ]. These counselors were trained and supervised in IPT by study clinical supervisors (JJ and JN). Document review data sources included structured process notes of clinical supervision sessions (n = 359) kept by study clinical supervisors, relevant email exchanges between the study and prison staff (n = 46), minutes from study team and prison meetings (n = 18), internal memos (n = 10), and other files that documented implementation processes (official letters, training manuals and intervention goals; n = 27) [ 11 , 14 , 22 ]. To aid with the planned document review, structured supervision process notes kept by the study team included the following questions: “What IPT elements went well?”, “What challenges were encountered? What was recommended?” “How did the counselor respond to the feedback?” “Specific barriers/facilitators discussed at the facility, counselor, client, or state prison system level.” [ 11 , 14 , 22 ].
Interventions
Treatment as usual
Prison treatment as usual for MDD typically consists of antidepressant medications (either selective serotonin reuptake inhibitors or tricyclics) [ 24 ]. Psychosocial interventions in prison are often psychoeducational and highly structured. IPT was not being provided in participating facilities other than through our study.
Group IPT
Group IPT used the study treatment manual [ 11 , 25 ]. IPT was offered to 91 study participants who had MDD, were incarcerated at participating prisons, and were randomized to the IPT condition in the trial [ 14 ]. It was delivered using 20 90-minute group therapy sessions over 10 weeks with 4 individual (pre-group, mid-group, post-group, and maintenance) sessions. Individual sessions focused on treatment planning, individualized treatment goals, and preparing clients to use group effectively. In the pre-group individual session, counselors reviewed what depression is and some of its causes, explored potential precipitating events for the current depressive episode, explained how those problems are typically addressed in IPT, and then worked with participants to choose target problems and IPT approaches for their situation (i.e., create a mutually agreed upon treatment contract). As in Wilfley et al. [ 26 ] we asked counselors to summarize the case conceptualization (which we called “goals summaries”) for clients’ review and revision at the first group session. Goals summaries included a simply worded description of the client’s precipitating crisis for MDD, its link to the current depressive episode, and IPT steps for addressing it. Because the prisons are mental health resource-poor, counselors almost always ran groups alone rather than in pairs. The mean and median number of members per group was 4; the largest group had 7 members [ 14 ]. Members were assigned to IPT groups based on logistical considerations (when they enrolled in the study and the facility in which they were housed, such as men’s medium security, women’s facility, etc.).
Training and supervision
The study hired prison counselors to offer IPT as moonlighters. Counselors were employed at participating prisons, with at least a bachelor’s degree and at least one year of prison work experience. Study clinical supervisors were external to the prisons. Study supervisors provided study counselors an initial 1.5-day IPT training (with all study counselors together) and then ongoing individual IPT supervision. Three refresher trainings were also offered during the 3-year study (see details in the results section). Ongoing supervision consisted of weekly review of counselors’ audiotaped IPT sessions, feedback on their written goals summaries for each client, and weekly individual phone consultation. When the secondary supervisor was supervising a counselor, the primary supervisor provided supervision of supervision. This 3-year, real-world pragmatic effectiveness trial did not have counselors practice and become certified in IPT before beginning groups. After the initial training, counselors began groups under study supervision. Therefore, notes made on the training and supervision process cover counselors’ entire IPT learning trajectories. This approach is also closer to what is realistic in in low-resource settings.
Qualitative data analysis
The approach to analysis used thematic analysis following the steps of framework analysis [ 27 ]. The Consolidated Framework for Implementation Research (CFIR) [ 23 ] was used as the a priori framework for the larger series of papers related to the trial. Johnson et al. [ 22 ] reported findings from the CFIR constructs of inner setting, outer setting, intervention, and individuals involved. The current report analyzes items that had been top-level coded for the fifth and final CFIR dimension, implementation processes.
Within the CFIR process domain, codes were analyzed two ways. First, we briefly summarized implementation processes using thematic analysis of activities within the CFIR process steps (planning/engaging, and executing/reflecting/evaluating [ 23 ]). Second, we used thematic analysis to identify themes in discussions between supervisors and study counselors within each section of the structured process and supervision notes. Supervision was guided by three prompts for supervisors to discuss with counselors: (1) What went well? (2) Where do you feel like you struggled? (3) Tell me about your case conceptualization for each client. Supervision notes (made by supervisors after supervision) followed a semi-structured template that included supervisors’ and counselors’ perspectives on what went well in the session/s, challenges the counselors experienced, and recommendations offered by the supervisor/receptiveness of the counselor to the supervisor’s recommendations. We used thematic analysis to identify themes within responses to each of these sections (i.e., what went well, challenges, and recommendations/response sections).
Prior to coding, JJ and SWS created a comprehensive codebook for the study. The team coded the documents using NVivo software. Data was coded by four study team members (MH, FR, JJ, and a 4 th coder) with previous qualitative experience and familiarity with the study topic. JJ was also the study principal investigator and one of the study supervisors. The four coders met as a team to establish consensus after coding the first 10 files independently. The coders coded the rest of the data individually, meeting in pairs to establish consensus. When discrepancies emerged, coder pairs discussed and reached agreement; final consensus files were integrated into the master file for analysis. Within responses top-coded within the CFIR “processes category,” we first briefly summarized implementation processes within the CFIR process steps (planning/engaging, and executing/reflecting/evaluating [ 23 ]). Second, thematic analysis was applied first to responses to the “what went well?” section of the process notes, then to responses to the “challenges” section of the process notes, and then to responses to the “recommendations/counselor response to feedback” section of the process notes, yielding three sets of themes. MH was the primary qualitative analyst with secondary interpretation and validation by JJ.
In the results section, supervisors are labeled “S01” and “S02,” and counselors are labeled “C01” through “C11.” Clients receiving group IPT are referred to as “group members” (i.e., members of the IPT treatment groups).
Results
A total of 460 supervision notes, meeting minutes and email exchanges between the 11 counselors and the two supervisors describing implementation and supervision were analyzed. Results are organized by analyses. The first set of analyses describes implementation processes pulled from the 460 study documents categorized and organized using the CFIR process steps. The second set of analyses identified themes within each section of the structured supervision notes: (a) counselor and supervisor perspectives on what went well, (b) counselor and supervisor perspectives on challenges, and (c) supervisor recommendations and counselor receptiveness.
Both study supervisors were white, non-Hispanic women in their 30s. Nine of the 11 study counselors were female. Eight counselors were non-Hispanic white; the others were Hispanic White (n = 1), Asian (n = 1), and Non-Hispanic Other (n = 1). Counselors had a median age of 33 (range 26–49) at the end of the 3-year trial. Individuals enrolled in the trial (n = 181) were 35% female and 19% Hispanic, with self-identified race described as White (59%), African American/Black (20%), Native American (2%), Asian (1%), more than one race (6%) and Other (12%). Their average age at study enrollment was 39 (range 20–61) [ 14 ].
Analysis 1: Implementation processes categorized using CFIR [ 23 ] process steps
Analysis of implementation documentation using the CFIR process steps yielded the following summary of implementation processes. It is important to understand these processes to contextualize what went well and what was a challenge in subsequent supervision. Together, this information can guide decisions to use similar processes or adapt them in the future.
Engaging and planning
Many prison specialist and non-specialist counselors indicated interest in being study counselors. Counselors described receiving additional training, supervision, and support, as well as getting paid by the study to work outside their extra hours, as motivators. Our initial training cohort included 7 prison counselors (4 non-specialists and 3 masters-level social workers), 5 of whom ended up leading IPT groups (one was in a small facility that did not have enough participants and another became overcommitted). We hired and trained an additional 4 counselors (all master’s prepared prison mental health clinicians) over the course of the study to cover new facilities or existing facilities as study counselors got promoted. We trained most of these new counselors at refresher trainings or abbreviated versions of the original training. We had to train one new counselor (C06) in only two 2-hour segments due to a last-minute schedule switch, and she required additional training over the course of her first group.
Initial training of study counselors took place over a day and a half, on a Friday evening and a Saturday to work around counselors’ prison workweek. An outline of training topics and methods is shown in Table 1 . Training procedures accommodating both licensed mental health counselors and non-specialist counselors had been honed in our pilot studies of IPT in prisons [ 17 , 18 ] and followed our prison IPT manual [ 11 , 20 ]. Based on training experiences in the pilot trials, training for this trial (1) emphasized that the manual should not be read word for word when delivering IPT; (2) used audio recordings of IPT to give counselors practice responding and a sense of what IPT sounds like; and (3) offered recorded mock sessions covering more complicated IPT techniques, like communication analysis, that counselors could borrow and watch. Counselors said the most helpful aspects of this training were role plays, videos, and the lunch we provided. As a result of feedback on the first training from counselors and supervisors, we added additional detail on the pre-group IPT session, finding a current (rather than past) interpersonal goal, provided additional example recordings, and planned refresher trainings to help counselors learn from each other’s experiences.
10.1371/journal.pone.0288182.t001
Table 1 Initial training as provided.
Topic
Time
Details
Friday Evening Session (4 hours)
Overview
15 min
Welcome, introductions, and overview
MDD
20 min
Definition, rates, treatment options
IPT rationale
30 min
Social support, IPT focus areas, and their relationship to MDD: grief, interpersonal conflicts, life changes, problem patterns.
Counseling essentials
70 min
The client is the expert on his/her own experience, needs, desires, future (why; role play to have trainees experience what it is like to be on the receiving end of direct/teach vs. listen/evoke helping styles) Empathy and reflective listening (what it is, what it is not, small group role plays to give everyone a chance to practice and to be on the receiving end, responding to audio recordings as a group) Emotions (don’t fix, just listen) The power of ounsellor expectancies for clients
IPT case conceptualization, pre-group session
45 min
Introduction and education about MDD Practice interpersonal circle exercise Role play practice identifying the interpersonal problem area How to create written interpersonal goals summaries
Role play entire pre-group session
50 min
Group leaders demonstrate and let trainees watch Trainees can lead small parts as they feel comfortable
Discussion and homework
10 min
Homework: practice reflective listening with friends in at least 3 conversations, record at least 1 (we provided audio recorders) Finish reading the treatment manual, if not already finished
Saturday Session (8 hours)
Welcome
15 min
Welcome and questions
Counseling essentials
45 min
Discussion of reflective listening homework (listen to trainee audio recordings as a group: where do you think you did well? Where do you feel like you struggled? Pointers)
Study logistics
45 min
Case notes, audio recorders, confidentiality, emergencies
IPT techniques
75 min
Discussion, lived and taped demonstration, role plays of: ounsellor as client advocate, empathy, validation, building a therapeutic alliance, non-judgmental approach, working with ambivalence, clarification, communication analysis [with role plays], problem-solving, helping clients role play, present relationship focus, helping group members (i.e., clients) support each other, attention to group process, strategies that are never part of IPT
Grief sessions
45 min
Big picture and details of addressing grief in IPT Role play reflective listening in grief work
Lunch and walk
30 min
We provided lunch
Conflicts sessions
75 min
Big picture and details of addressing interpersonal conflicts in IPT Watch a videotaped demonstration Role play communication analysis Role play helping clients change unhelpful communication patterns
Life changes sessions
30 min
Big picture, details, and role play
Problem patterns sessions
45 min
Big picture and details Role play reflecting listening in problem patterns work Role play working with ambivalence in problem patterns work
Strategies for challenging situations
45 min
Using the group, managing group process (e.g., monopolizers, quiet people, etc.), suicidality, conflict or hostility in the group, outside conflicts spill over into the group, trauma or dissociation, guardedness, violence or threats in relationships
Synthesis and summary
30 min
Review of big picture and important points, adherence and competence measures, IPT quizzes, question and answers Reflective listening Help clients resolve interpersonal problems, including accepting mixed feelings about grief, conflicts, life changes, problem patterns Work to create a respectful, supportive relationship with clients Emotions matter; help clients experience and express them. Help clients develop social support. If you are stuck, “What is that like for you?” and “Who else can you discuss this with?” are almost always good questions
Feedback
15 min
Final questions. Feedback on the training and the manual.
Note: Training followed the structure of the intervention manual. We sent the manual to counselors several weeks before the training and asked them to read the manual and come with questions.
Executing, reflecting, and evaluating
Training continued with ongoing supervisor audio recording review and weekly individual telephone supervision. In particular, supervisors listened to counselors’ pre-group IPT sessions and helped them to edit and refine the written IPT case conceptualizations (“goals summaries”). Supervisors also observed that both specialist and non-specialist counselors seemed to understand IPT much better after their first set of groups.
After having run [IPT] groups a few times , [C02] said it was easier to run groups this time . She indicated that she had a better understanding of what the therapy was about and what needed to be worked on . She stated that the improved understanding of IPT made it possible for her to know where she was going so she could “think less” and focus on what she needed to do . - S01 note on C02 supervision
We conducted 3 refresher trainings during the 3-year study (at 6, 15, and 21 months after the initial training; Table 2 ) that covered areas in which counselors were still struggling. Our first refresher training reviewed the steps to forming a mutually agreed-upon written case conceptualization and negotiating a treatment contract before the beginning of the group. Counselors had struggled with these ideas, so we refined the pre-group IPT session in the manual and provided additional training in these skills. Negotiating the case conceptualization and treatment contract was the hardest part of IPT to learn, but once counselors learned it, they were well on their way to achieving IPT competence. The refresher training also covered focusing versus flexibility in IPT groups. The second refresher training reviewed IPT treatment details. A final (full day) training covered all previous trainings in a condensed fashion for an additional cohort of counselors and as review for ongoing counselors. Because these final counselors were all master’s trained mental health providers (i.e., specialist counselors), we trained them in 8 hours rather than 12. Clinical training took place in person. Ongoing clinical supervision took place individually by telephone.
10.1371/journal.pone.0288182.t002
Table 2 Refresher trainings as provided.
Topic
Details
Refresher training 1 (half day; 6 months after initial training)
IPT case conceptualization
It is important to have a mutually agreed-upon written case conceptualization/ treatment contract (“goals summary”) before the beginning of the group. The goals summary is written for the client in lay terms. It provides a brief description of the problem leading to the current depressive episode and ways of using the IPT group to address it.Steps for negotiating the treatment contract/goals summary include: 1. Determine when the client’s most recent (not first) depressive episode started (or their mood worsened, if client has chronic MDD) 2. Ask open-ended questions about what was going on in the client’s life around that time. Ask about their relationships, any losses, any changes in their circumstances. 3. Listen carefully for interpersonal conflicts, life changes, grief, and isolation. With individuals in prison, there are usually many potential triggers. 4. Reflect potential triggers for the depressive episode back to the client: “So, it seems like there were a lot of stressful things going on in your life at the time,” list them, then ask, “Which of them bothered you or is currently bothering you the most.” 5. Discuss the most salient triggers for the current depressive episode, and then once one (or two) are identified, ask, “Would you like to work on that issue in group”? 6. Once a troublesome issue that the client would like to work on is identified, describe the IPT steps for working with that issue, how s/he could use the group, and ask him/her if that makes sense. 7. Negotiate and renegotiate as necessary. 8. Then finally, write up a user-friendly paragraph describing the issue and another short paragraph describing how to address it. 9. Bring the statement to the first group, run it by the client, and revise as necessary. Interpersonal (i.e., therapeutic) goals may be refined over time, but it is important to start group with an interpersonal goal in mind. Reviewed written examples.
Common IPT errors
Get the triggers of the CURRENT depressive episode, not the first one. Interpersonal goals should be present focused. If a client identifies a trigger from the far past, ask “Are there things going on now in your life that remind you of that time?” Get a treatment contract (i.e., interpersonal goals that the client wants to work on in group) before group starts
Focus vs. flexibility
Reviewed steps to address each IPT problem area How to decide what to work on in a session How to balance emergent issues on any given day with what you had planned for IPT work that day, especially when counselors are still learning IPT
IPT essentials
Review and importance of reflective listening and empathy When in doubt, use reflective listening. This can be used to slow things down and re-establish an alliance if something feels adversarial or the client feels misunderstood
Discussion, reflection
What has gone well? What has been a challenge? Lessons learned from specific cases
Refresher training 2 (half day; 15 months after initial training)
IPT finer points
Different ways of solving interpersonal conflicts (communicate, set boundaries, change expectations, end the relationship). The goal is to move clients from a stuck, ambivalent place, not necessarily to get them to communicate with someone they’ve given up on. Meeting clients where they are: respecting their reading level, readiness to discuss issues in group Appropriate level of detail to explore sensitive issues (e.g., trauma history, crime details) Importance of a structured mid-group individual session to review progress on goals and discuss work to do before the end of group Importance of a formal goodbye at the last session Ideas for handing challenging clients, especially Axis II, in group
IPT essentials
Review and importance of reflective listening and empathy
IPT case conceptualization
IPT case conceptualization is the purpose of the pre-group session. Review of steps. When did the mood change? Reflect the challenges you hear in an IPT frame and ask them what they think is most important (it is the client’s goal). Writing the goals summary.
IPT approach
The 2–3 steps for addressing each interpersonal problem area Practicing communication analysis
Discussion, reflection
Summaries of recent groups Wins, challenges, lessons learned.
Refresher training 3 (full day; 21 months after initial training)
Overall
This was a 1-day training and refresher training condensing all the topics from previous trainings review for ongoing and learning for new study counselors
Analysis 2: Themes observed within each section of process and supervision notes
Themes corresponding to each section of the structured supervision notes (i.e., what went well, challenges, and recommendations/responsiveness) are described below. Themes in each section are underlined. A summary of themes in each section is shown in Table 3 .
10.1371/journal.pone.0288182.t003
Table 3 Supervision themes.
What went well
Challenges
Recommendations and responsiveness
Therapy basics
1. Reflective listening and empathy 2. Focusing on emotions 3. Getting experience/conversation specifics 4. Open-ended questions
1. Reflective listening, focusing on emotions 2. Getting experience/conversation specifics
1. Reflection, emotions, specificity, open-ended questions
IPT case conceptualization
5. Case conceptualization, advancing work on interpersonal goals 6. Structure/limit-setting/protecting clients 7. Structure with flexibility
3. Pre-group session/treatment contract/case conceptualization 4. Counselors off-model 5. Keeping the group focused and doing IPT work 6. Managing structure vs. flexibility
2. Case conceptualization 3. What are the next IPT steps? 4. Keeping the group focused and doing IPT work 5. Limit-setting and structure 6. Structure vs. flexibility
IPT techniques
8. Role plays, communication analysis, work on interpersonal conflicts
7. Role plays, communication analysis, social support
Therapy processes
9. Alliance and alliance repair (including negotiating tasks, goals) 10. Group processes
7. Managing group process/conflicts
8. Managing group processes
Difficult situations
8. Challenging clients 9. Clients don’t want to feel/discuss feelings 10. Other forms of resistance (attendance, disinterest in social support) 11. Counselor schedules, frustrations with the prisons
9. Challenging clients
Counselor growth
11. Counselors putting the pieces together and getting more confident
10. Encouraging, supporting counselors 11. Counselor receptiveness
What went well
Themes found within the “what went well” section of supervision notes included: (1) reflective listening and empathy; (2) focusing on emotions; (3) getting specifics of client experiences and conversations; (4) asking open-ended questions; (5) case conceptualization and advancing work on clients’ interpersonal goals; (6) structure, limit-setting, and protecting clients; (7) structure with flexibility; (8) specific IPT techniques such as role plays, communication analysis, and work on interpersonal conflicts; (9) alliance and alliance repair (including negotiating tasks and goals of treatment); (10) group processes; and (11) counselors putting the pieces together and becoming more confident.
Supervision notes focused on several core IPT skills that are central to most psychotherapies (i.e., therapy basics). Reflective listening and empathy were among these core skills. Not all counselors had been trained to use these techniques prior to the study, but all used them well during the study, sometimes to great effect.
[Client] started out with some [resistance] at first , but [C03] was careful to validate and empathize and [the client] really engaged and did some good work… . [The client] sounded angry , [C03 ] said , you sound angry , he said , I’m not angry , [another client] said , I don’t know what you’re feeling man , but you sound pretty [expletive] angry , and [the first client] started to cry . -S01 note on C03 supervision
Reflective listening helped counselors elicit clients’ emotions about situations being discussed and create the safety needed for clients to experience those emotions, another core skill of IPT. Several of the counselors had been trained in alternate therapy styles, which tended to ask about thoughts or behaviors or focus on providing education, rather than on experiencing emotions. However, all the counselors (including the non-specialist counselors) became good at listening for, focusing on, and helping group members (i.e., clients) talk about emotions over the course of supervision.
[ C02] asked a group member to share her story . [C02] attempted to maintain focus on slowing the participant down and getting the details of the complete narrative as well as eliciting the emotions that the participant felt at that time . [C02] consistently does a good job eliciting emotional content from the group members . - S02 notes from C02 supervision
Helping clients slow down and provide specifics about their experiences and conversations in their personal narratives is another core IPT skill. Supervision notes described several times when counselors did a good job inviting clients to describe the specifics of a grief or life change narrative (to elicit and process emotion) or of a conversation that did not go well (to role play and problem-solve how to communicate better in the future; “ Could you go through what you said line by line ?”). Open-ended questions also helped clients experience emotions. Supervision notes detailed examples of excellent questions, including questions focused on exploration (“ How does this conflict with your brother affect you ?”), process affect (“ What was it like having that conversation , ” or “What is it like to talk about this now ? ”) , and building social support (“How can we support you ? ” and “When you’re struggling and you need help , who can you go to ? ” ).
One of the more difficult skills for counselors to master was keeping track of group members’ case conceptualizations and knowing how best to focus the group to advance work on interpersonal goals . Some counselors understood this well on their first set of groups, but most did better by their second set of groups.
[ C03] did a really nice job–making spot-on clinical judgments about what the core IPT issue is , what in-group issues to address and not , and doing some really nice reflecting and following the model .– S01 note on C03 supervision
[ C11] has a good sense of the women in group , their interpersonal goals , and where they need to go in the remaining groups .– S02 note on C11 supervision
Advancing work on interpersonal goals required clarity about the goals and methods used in group to achieve them, a clear agenda, structure, limit-setting , the ability to redirect less productive conversations, and the ability to protect group members . Supervision comments related to such structure included, “ Good review of confidentiality and expectations/logistics of group ” and “ Good group format : check-in , session topic , and check-out . ”
Group was giving [the client] a lot of advice and “you should have known” …. Counselor defended [her and said it wasn’t her fault] . -S01 note on C04 supervision
[ C07] does nice job giving overview of conflict session and the kinds of information she is interested hearing about to understand the conflict . -S02 note on C07 supervision
When counselors understood the bigger picture arc of IPT treatment and the steps to address each category of interpersonal problem, they were better able to have a structure but adapt in real time if other relevant issues surfaced (i.e., structure with flexibility ).
[ C11] has done a good job structuring groups , so the women know what the main focus of the group is , whose “turn” it is for the day , as well as rolling with things when they don’t go exactly as planned . -S02 note on C11 supervision
The next set of skills often commented on in supervision notes was specific IPT techniques, including communication analysis (a line-by-line telling of a conversation to determine how better to do it next time), role plays, and work on interpersonal conflicts . One example of successful communication homework included:
[ C07] did a nice job giving a group member some concrete interpersonal homework (try not to say the “f-word” for two days and see if people stick around longer to talk to you , try to find some new guys in the program to help)… He laughed about the first assignment but agreed to both , and tried them and by Thursday , was happier . Said he’s definitely getting different reactions from others . -S01 note on C07 supervision
Supervision notes often commented on therapy processes such as therapeutic alliance (including alliance repair) .
[ C01] has a nice way with the clients . She was client , understanding , and they seem to open up to her . It’s obvious that she wants to do a good job and that she is there to help . S01 note on C01 supervision
[ C06] did a nice job addressing the issues from last week’s groups; she was clear and direct with the group members about her not hearing what they were asking her for last week . The group members responded well . - S02 note on C06 supervision
Therapeutic alliance also includes agreement on goals and tasks of treatment. Supervision notes sometimes commented on how the mutually negotiated treatment contract and other choices provided by the counselors helped group members take ownership of their own therapeutic work.
How to give people options about their interpersonal goals and have them decide so they feel like they own them , they then take responsibility for them and it’s great. - C07 as quoted in S01 notes.
[ C02] did a nice job checking in with the group member to see if she felt comfortable talking about a topic (e . g ., her crime) and respecting the group member’s boundary about this . She also checked back in with her at the end of group and got the group member to agree that it would be helpful to talk about this with her other group members as well as use the group to talk through what she wants to communicate to her family and get feedback .– S02 note on C02 supervision
Supervision notes often described counselors’ efforts to foster helpful group processes , including getting group members talking to each other, asking about commonalities among group members, helping quiet members speak and monopolizers listen, and teaching group members how to use reflective listening and provide empathic feedback to help each other.
[ C04] sent an email about Conflicts Session 2 , saying that it was some of the best , most authentic work he’s ever seen , with excellent feedback from group members to each other , and that they gave each other honest feedback in one of the most skilled ways he’s ever seen… he was excited for me to hear it .– S01 note on C04 supervision
The final category of comments in the supervision notes about what went well related to when all the IPT skills were coming together and counselors were proficient in the model.
[ C01] is empathizing , working with feelings , building support , seems to get the model , working on communication and relationships… IPT has been a different way of working for [C01] (more focused on feelings , less on strict behavioral activation) , but she is really , really enjoying it… can’t wait to start next group . - S01 note on C01 supervision
Challenges
Supervision notes also had a section describing challenges that counselors experienced on the way to learning and mastering IPT. Themes found within the “what went well” section of supervision notes included: (1) reflective listening and focusing on emotions; (2) getting specifics of client experiences and conversations; (3) using the pre-group session well to be able to obtain a treatment contract and formulate a case conceptualization during that session; (4) counselors off-model; (5) keeping groups focused and doing IPT work; (6) managing structure vs. flexibility; (7) managing group processes and conflicts; (8) challenging clients; (9) clients not wanting to feel or discuss feelings; (10) other forms of resistance (e.g., poor attendance or disinterest in social support); and (11) difficult counselor schedules and counselor frustrations with the prisons.
In terms of therapy basics, counselors began at differing stages of experience with reflective listening and comfort focusing on emotions . Furthermore, even experienced counselors are not 100% empathic all the time.
She’s telling him what to do rather than empathizing . -S01 note on C07 supervision
He’s interpreting , rather than reflecting . “Sounds like you’re romanticizing”… [C03] “ confronted him”… instead of just asking “do you want to come back or just do individual sessions or be done ? ” - S01 note on C03 supervision
It’s hard for the counselors without mental health training to know how to follow and push a little for the emotionally intense stuff and how to do real therapy without getting caught up in “ I have to fix this ! ” or “What is the next question on the manual ? ” It’s a confidence/anxiety issue . ” -S01 note on C02 supervision
Supervision notes also included reminders to focus on specifics of group members’ experiences and conversations (rather than generalities), and to focus on the present (rather than the past).
The second set of challenges reflected learning IPT case conceptualization , which counselors typically understood better on their second set of groups. Supervision notes included documentation of some of counselors’ struggles with case conceptualization and negotiating a mutually agreed upon treatment contract in the pre-group session, especially initially. This was the most challenging part of learning IPT for counselors.
Overall, the pre-group [IPT] sessions were short (45 minutes vs. 75) and though the beginning of the sessions started well , [C07] had trouble getting information she needed about when major depressive episodes started and ended and so didn’t leave the sessions with ideas about what treatment goals should be or with a treatment plan negotiated with the participants. She said she had felt “frustrated” in the sessions and didn’t know how to write the goals up afterward.– S01 note on C07 supervision
[ C08] indicated that she did better with the more structured groups than with the less structured groups . She also indicated that she sometimes would feel stuck and not be sure of what direction to go in next . -S02 note on C08 supervision
As a result of these challenges, our refresher training focused much more on the steps needed to create an IPT case conceptualization and negotiate a treatment contract in the first session (see Table 2 ). We also expanded the description of the first session in the treatment manual. In addition to learning IPT case conceptualization and how to get it in the first session, non-specialist counselors had to learn about the general idea of a case conceptualization, and more experienced counselors had to work not to default to other models .
[ C01] said that she had to work to keep her inner behaviorist in check and focus on feelings rather than pleasant activities when one woman was so sad . -S01 note on C01 supervision
These challenges were addressed through supervision and additional training sessions, including reviewing the IPT adherence and competence scales with counselors.
Another set of challenges related to keeping groups focused and doing IPT work . It took practice to hold the big-picture arc of each client’s treatment goals (i.e., case conceptualization) in mind to help focus moment-to-moment interactions. Counselors also needed to know when to set firm limits, including when group members pushed boundaries, tried to help other group members in unhelpful ways, were distractible, or when groups were off topic:
Another challenge was that one of the group members was very vocal , very loud , and would run [the counselor] and everyone else over and clearly had his own agenda … When he went off , another group member would follow along . So , it was really hard for the first few weeks for C06 to get them doing interpersonal/therapeutic work because… there wasn’t a good therapeutic contract in place . -S01 note on C06 supervision
We trained this counselor very quickly when another counselor had a scheduling conflict. She had not fully understood how to make a collaborative therapeutic contract in the pre-group session. Once a good treatment contract was in place with each member, the counselor was able to help this group of clients do better IPT work.
One of the places the counselors struggled the most but mastered over time was managing structure versus flexibility in the IPT groups :
[ It was] hard for the therapist to hold onto the IPT frame in this session . I’ve had that with 3 therapists now I’ve trained , who have great exploratory/general skills , so they do well out of the gate , but then struggle a bit to stick to the structure of IPT , look at the manual , and then do the more active parts of IPT once they get closer to the middle of the group… In this active phase of group , I keep telling all of 3 them to focus solving the interpersonal problems and stick closer to the manual . -S01 note on C01 supervision
It took time for counselors to master what is core (e.g., IPT steps for working on grief, life changes, interpersonal conflicts) and what is flexible (e.g., session order) and when to go with what they had planned (typically have a 15-minute check-in and then conduct planned session) versus what walks in the door (when someone is in crisis or is upset about something related to work on one of the group member’s interpersonal goals). A common mistake was letting group check-ins go too long (e.g., 45 minutes) and not using more of the group time for interpersonal work.
Counselors also needed to manage challenging group processes, including conflicts among group members as well as members who monopolized group, derailed group, or gave unhelpful responses to other members (giving advice, shutting down emotions).
Managing the argument kept the counselor on her toes , but she handled it well . At one point , a deputy came in to check on them (they were getting loud ), which also helped . -S01 note on C01 supervision
Most conflicts among group members provided in-group opportunities to practice addressing conflicts. However, in two cases, we split groups because of conflicts that pre-existed the group and provided a distraction rather than a learning opportunity.
Counselors also sought supervision over clients they found difficult or frustrating . Most of the group members had personality disorders (72% had antisocial personality disorder and 38% had borderline personality disorder; Johnson et al., (2019) and almost all had difficult life histories.
Therapist feels she struggles with one member who tends to talk and talk … Just when she thinks she should cut him off because what he’s saying is irrelevant, he says something that makes it relevant . We talked about how to manage him.– S01 note on C07 supervision
The most common resistance to IPT work that counselors experienced in groups was implicit or explicit avoidance of emotion :
[ C07] keeps trying to focus them on feelings and specific things , and they keep trying to move into general philosophy .– S01 note on C07 supervision
One group member believes that talking about things doesn’t make people feel better . He resists [C02’s] attempts to get more details about his life change or to share how he felt in past or currently. [C02] sticks with emotions as way to get him to identify feeling words, but this group member struggles to provide much detail about his interpersonal goals or use feeling words .– S02 note on C02 supervision
Counselors also occasionally experienced resistance in other areas (attendance, the need for social support, role plays). Counselors addressed these challenges through re-establishing treatment contracts (“Given that, would you rather work on something else?” or “What would you like to do?”) and/or finding ways to accommodate clients’ concerns while doing other IPT work.
The final category of challenges documented in supervision notes was counselors being burned out, overcommitted, or frustrated with their other clinical work, with prison administration, or with their own life circumstances. Their prison positions were genuinely demanding.
I’m just getting home now and I wanted to let you know that I had to cancel group tonight because of a crisis at the prison . I was able to call all of [the group members’] units to let them know . - C06 email to S01
Are you free to talk right now? Call with [C06] did not go well, she feels too frustrated with the Department of Corrections and everything and wants to bail … I am super frustrated . -S02 email to S01
The circumstances of this group have been very , very hard for [C06] and this supervision took place by phone on a Friday night when she was sick after being at a funeral all week . - S01 note on above email from S02
[C05] is too busy , has rescheduled supervision [a lot] . … Supervision has been hard because we do it while she’s driving and she doesn’t take notes … This is her 3 rd job and she commutes 2 hours each way to her main job . - S01 note on C05 supervision
Supervisor recommendations and counselor receptiveness
The final section of supervision notes described recommendations that supervisors made to counselors and counselors’ receptiveness to those recommendations. Themes found within this section included: (1) reflective listening, focusing on emotions, specificity, and open-ended questions; (2) case conceptualization; (3) recommendations for the next IPT steps; (4) keeping groups focused and doing IPT work; (5) limit-setting and structure; (6) structure vs. flexibility; (7) role plays, communication analysis, and social support; (8) managing group processes; (9) challenging clients; (10) encouraging and supporting counselors; and (11) counselor receptiveness.
The first large category of recommendations provided to counselors during supervision addressed the basic IPT and general therapy skills of reflection, focusing on emotions, asking for open-ended questions, and helping clients be specific in describing events and conversations. More use of reflective listening generally, as well as instruction to empathize and reflect rather than interpret or argue, was common feedback discussed during supervision.
We have given her consistent feedback to use reflection rather than say “I understand what you ’ re saying , ” which has also helped a lot , and she’s starting to do that beautifully .– S01 note on C06 supervision
Feedback to ask for details of interpersonal communication and events was especially common for counselors without formal therapy training.
Get more confidence pushing for details, specifics, and feelings . Don’t let them go with general, vague statements. -S01 note on C02 supervision
Supervisors also provided frequent feedback to help counselors with IPT case conceptualization , especially with the process of eliciting the information needed for the case conceptualization and then developing treatment goals collaboratively with clients.
[ Be sure to get] a treatment contract : “Here’s what I hear , here are our options for working , what do you think ? ” Provide options and then get agreement . -S01 note on C07 supervision
The intervention manual contained an “IPT cheat sheet,” with one page summarizing IPT goal areas and basic steps for addressing each. In addition, supervisors reviewed case conceptualization and implications for the next IPT steps for each group member during each supervision session.
Discussed with group leaders the directions of where to go with each group member and what sessions to conduct next week . -S02 note on C06 supervision
We discussed [ C06] validating that it is okay to have feelings , okay to feel sad about losses , etc . We also discussed having [the client] focus on positive action-focused coping since he is trying to not get anxious and frustrated when his sister does not contact him or follow-through on things . He also values the relationship and most likely has not clearly communicated his feelings or needs to his sister . We discussed encouraging him to think through what is best for him and what he wants from this relationship . It is clearly a relationship that is highly valued , but also one that causes him stress and triggers his anger . -S02 note on C06 supervision
Recommendations to counselors also addressed strategies for keeping the group focused and working, as well as limit-setting and structure as needed. For example, supervisors sometimes suggested redirecting unhelpful responses among group members. Recommendations included:
Eight group sessions left–what does she want to be sure to get done therapeutically before group is over ? Talked about how this is a short period of time , it will go quickly and how to use the mid-group sessions and remaining group sessions to identify which work needs to be done and do it . -S01 note on C06 supervision
One group member tends to monopolize group time and emphasized importance of… keeping him to reasonable check-in time . -S02 notes on C06 supervision
After [he ] told his grief story , one group member said , “you can’t change it , get over it . ” Suggested cutting this kind of thing off , doing some education about grief , supporting validating feelings , and helping the group do that also . -S01 note on C07 supervision
Another theme of supervision recommendations was structure versus flexibility , and how to know when to do each. For example:
I asked the counselor to follow the prescribed sessions more (he often never gets to them in the session or only as a brief “topic” at the end ). He said part of the problem is that he’s trying to be flexible to go with whatever focus areas are hot for the group members that day , but it’s often something different than what he has prepared , and he can’t read the manual and run group at the same time , so it’s been hard to follow the manual and be maximally responsive to members’ needs when he doesn’t know the manual well yet . I told him for now , to go into group with one or two sessions prepared , to let them do their check-in and then to go with the prepared sessions . He has let them go a lot and now I think he can be more structured with specific work to get done . -S01 note on C04 supervision
We talked about how to tell if someone is done with the emotional part of grief work (the feelings have subsided some , are less conflicted , and you know all the details of the story) and how to continue to get details of the story and help the woman have space to sort out her feelings about them . I also suggested not to try to “fix it” by talking her out of her guilt , but to let the woman talk it through and come to her own conclusions . Later , I had a very concrete suggestion for a woman , and [C02 ] said , “But isn’t that fixing it ? ” which was a great question . So , we talked about when a concrete suggestion makes sense (for a very concrete information-based issue , like “how do I get to the store ? ” or “you might try methadone treatment”) and when emotion-focused work makes more sense (an issue with a lot of conflicted emotions or that is very emotionally laden) . This made sense to her .– S01 note on C02 supervision
Another common theme of supervisors’ recommendations included communication analysis, role plays, and helping group members build social support .
Completed role plays… to improve communication since group member does not want to end relationships . Group leaders had group members start (so not to make a big deal about roleplaying) , but group members had trouble at first and would have benefited from a little more direction at the beginning . Provided feedback that one of the group leaders could start playing the group while the group member plays the other person (his wife in this example) since he has best idea of how wife might respond in the situation . Once the group leaders jumped in and completed the 2 nd role play , it went a lot better , the guys got a sense of how to do it and both liked this aspect of the session . Also discussed… suggestions to make to group member about his communication style . - S02 note on C01 supervision
Supervisors also frequently provided feedback or suggestions about making group processes and interactions among group members maximally useful.
One group member shared a current stressor … One group member in particular tends to provide very directive feedback that can come across as unsupportive . I provided feedback to [C02] about how she could have slowed down this portion of group when the women were all talking at once and made sure that the group member who was sharing her story felt understood and to check in with her about how helpful the “directive” feedback is from the other group member . - S02 note on C02 supervision
As mentioned above, a majority of clients in the study had personality disorder/s in addition to major depressive disorder. Therefore, some feedback to counselors related to helping counselors work with clients they found challenging .
We discussed that… the main focus should be on the remaining work on her interpersonal goals . However , since these [personality] issues interfere with her social support , [C11] could also take opportunities to help her work on boundaries/finding healthy ways to get her needs met . This more positive frame might be something she could hear and be willing to work on and [C11 ] can more gently reflect both the things she does well and the things she struggles with around relationships . -S02 note on C11 supervision
Finally, as counselors were learning a new model (and sometimes learning psychotherapy itself as an entirely new skill), they tended to be hard on themselves. Much of the feedback supervisors provided was to encourage, support, and reassure the counselors .
[C07 ] wasn’t sure about this session , if the guys are discussing what they need to , and feels that she has a hard time keeping one guy on track . I said I thought the session sounded great and the reasons why . She said she is still struggling to know when she’s on model or not and I told her the reasons she was on model and that she was doing great . She said until this session , the guys haven’t been talking to each other much and the guy who talked today has been quiet . I said , today was great , and they were even expressing appreciation for the things they learned from each other ! -S01 note on C07 supervision
Counselor receptiveness . Prison counselors generally responded well to and appreciated supervision. There were only a few instances in which counselors were slightly defensive or subdued.
They’re so happy they’re getting paid for supervision , because people pay a lot for supervision and don’t get good supervision like this . -S01 note on C02 supervision
Regarding listening to the tape . I did ! And I totally heard what you were referring to . I’m definitely planning on being more vigilant in my efforts to be reflective when interacting with [ the client] . -C03 email to S01
In fact, there were several instances of counselors calling or emailing supervisors outside of their regularly scheduled meeting times to ask for help.
I think I would like to talk to you before the next session . Today’s session did not go well both in and out of group . A few members got into an argument afterwards when talking to me and I would like to talk about ways to manage the next group . FYI , [client’s] father died today , which also added to the distress . I uploaded the session already , but there were many things not on the recording that need to be addressed . Do you still have Tues at 11 : 00 available to talk ? - C04 email to S01
Discussion
Current findings
This manuscript reports results from a planned qualitative analysis of 460 structured implementation and supervision documents in a hybrid effectiveness-implementation trial of group IPT for major depressive disorder in 6 state prisons. The goal was to inform: (1) implementation of evidence-based mental health treatments in prisons and jails, an important effort that needs more evidence to guide it; (2) psychotherapy and interpersonal psychotherapy (IPT) training efforts generally, especially in low-resource settings. To our knowledge, this is the first detailed description and analysis of implementation processes of a mental health intervention in a prison or jail setting. Analysis of training and supervision processes and lessons learned from the first major MDD trial among individuals who are incarcerated can inform and help to optimize psychotherapy training recommendations for prisons and other low-resource settings. Some study counselors had prior psychotherapy training and experience and some did not. Limitations of the study include having only two study supervisors and eleven study counselors. Study strengths included prospective, structured data collection documenting implementation processes, a rigorous qualitative coding process, and a unique and important implementation context (prisons) in which there has not been any previous analysis of counselor training processes.
The current qualitative analysis of supervision processes in this trial provides an outline for prison IPT training, retraining, and supervision ( Table 1 ; Table 2 ). Themes identified in analysis of supervision notes (summarized in Table 3 ) also provide a detailed description of where training procedures produced hoped-for results and where additional retraining/focused attention was needed. Findings ( Table 3 ) indicated that supervision focused on: (1) work on psychotherapy basics (reflective listening, focusing on emotions, open-ended questions, specific experiences); (2) IPT case conceptualization (forming a conceptualization, what is and is not therapeutic work, structure and limit setting, structure vs. flexibility); (3) IPT techniques (enhancing social support, role plays, communication analysis); (4) psychotherapy processes (alliance repair, managing group processes); and (5) managing difficult situations (avoidance, specific clients, challenging work settings). The single most challenging task to train was to help counselors understand the steps needed to formulate an initial IPT case conceptualization and negotiate a treatment contract in the pregroup individual session. This is a critical task because it frames everything else that is done in IPT, and a clear sense of the direction of therapy is needed to manage IPT in group settings, especially with frequent comorbid personality disorders (72% antisocial personality disorder and 38% borderline personality disorder in this trial [ 14 ]). Our initial training on obtaining an initial case conceptualization and treatment contract in the pregroup session had not been detailed enough ( Table 1 ), so we revised our training procedures and focused on this IPT task at the refresher training ( Table 2 ). We recommend focusing on this important task at length in future IPT trainings, especially in prisons. Counselors were receptive to feedback overall; some relied on study supervisors for support in managing stressful prison working conditions.
Based on the current analysis of supervision processes, together with previous findings related to treatment fidelity and outcomes in the trial [ 14 , 22 ], we make several observations. First, supervisors’ observations in this analysis as well as previously published treatment fidelity and outcome results suggest that non-mental health trained providers can learn IPT [ 14 , 22 ] using the training processes described in the current analysis. Second, all providers needed ongoing training in IPT case conceptualization to strike an appropriate balance between structure and flexibility in pursuing IPT goals in a group setting. A clear understanding of the big-picture steps of IPT allowed them better mastery of what is core and what is flexible. For those without previous psychotherapy training, it was important to explain the overall concept of case conceptualization. This included how it is possible to learn to keep a bigger picture of where the client headed in one’s mind while carrying on an in-the-moment conversation. We reassured new counselors that it gets easier with practice. Building in more IPT case conceptualization simulation and practice writing goals summaries during initial training may be helpful in reducing the learning curve for all counselors when groups begin. Our training had some role play of the initial session (see Table 1 ), but having counselors conduct and record a few mock first sessions and write the goals summaries might help. Third, group treatment was effective and cost-effective in this trial and in general [ 14 , 28 ]. It is often necessary in settings with many clients and few resources. However, learning to track and manage multiple members’ therapeutic work in a group setting took time, especially for novice therapists. In other words, group is doable and likely necessary, but requires additional training support. Some of the counselors’ more challenging tasks (including the need to get a written treatment contract during a single pre-group session and managing multiple challenging group members) related to the group setting. Additional training in group skills and group processes [ 29 ] might be provided; however, the key challenge of the group setting was tracking multiple case conceptualizations simultaneously, which could be addressed by more training in the big picture steps of IPT. Fourth, given that a majority of group members had antisocial and/or borderline personality disorders in addition to MDD, it is remarkable that counselors had challenges with so few of them. This is a credit to the study counselors and may suggest generalizability/flexibility of IPT.
In the current analysis, study supervisors observed that both specialist and non-specialist study counselors seemed to understand IPT much better after their first sets of groups. This observation is consistent with previous findings that treatment effects of IPT in this trial (relative to treatment as usual) were driven by outcomes of counselors’ second and subsequent (not first) rounds of groups [ 14 ]. Therefore, training and regular supervision through at least the first 2 sets of groups seems ideal. In addition, counselors in this study needed and benefited from refresher trainings every 6–9 months over the 2 years of the trial ( Table 1 ; Table 2 ). However, if training and supervision taper after that (and fidelity stays high), they become more scalable and sustainable. For example, previous cost analyses for the trial determined that costs for prison IPT programs that are already up and running (once ongoing supervision is no longer required) dropped from $2,054 per client to $575 [ 14 ]. However, even these scaled back procedures are still prohibitively expensive for most systems. Our hope is that by sharing a detailed analysis of the training process, we can facilitate future work to optimize training and supervision processes for under-resourced settings.
Integration of current findings with previous findings from this trial
Current findings, in addition to previously published results from this trial provide reason for optimism for implementing IPT in prison settings. Effectiveness results [ 14 ] indicated that IPT reduced depressive symptoms, hopelessness, and posttraumatic stress disorder symptoms, and increased rates of MDD remission relative to prison treatment as usual alone. Previous analyses of the first four CFIR domains (intervention, inner setting, outer setting, individuals involved) outlined facilitators of IPT implementation in prisons. These included IPT being a good fit for the target population, counselors and prison group members being enthusiastic about IPT, and counselors being open to learning evidence-based practices and committed to helping their clients [ 22 ]. In fact, prison counselors, administrators, and group members were highly motivated to find ways to better address mental health problems in resource-poor prison settings [ 22 ]. Their dissatisfaction with what they could do with such limited prison resources and high perceived need to change something to better address client needs [ 22 ] was echoed in our current finding that counselors were highly receptive to feedback. This was an unusually motivated and open set of counselors to train.
However, limited prison mental health resources and high need at the center of implementation barriers observed in our previous analyses [ 22 ] were echoed in the current process analysis. The main challenges to IPT implementation identified previously were overcommitment of prison treatment staff and variable prison supervision and collegial support [ 22 ], due to prison financial limitations that limited the number of mental health staff hired and resources to support them. Findings in the current analysis underscored this point. As noted in the current analysis, prison counselors often relied on study supervisors for support in dealing with difficult working environments with limited supervision and collegial support. Study counselors leaned on study supervisors for the collegial support and guidance that was missing from their settings. High client needs (severity, comorbidity, high levels of trauma and life stressors) and challenging climates among correctional staff in some of the facilities exacerbated these challenges. Furthermore, this trial took place in two states that were in the lowest quintile of incarceration rates and among the better funded states, suggesting that conditions are likely worse elsewhere. This is concerning given that supervision is important for providing quality care and preventing burnout, compassion fatigue, and turnover [ 30 , 31 ].
A critical IPT implementation task for the future is to find scalable training and supervision models to make them feasible for resource-challenged prison systems. In prisons and other low-resource settings (such as low- and middle-income counties) where supervisors are scarce and training time is limited, helping counselors become competent in IPT (and other evidence-based mental health interventions) as quickly and efficiently as possible is key for scale-up. In the current trial, supervision was effective in that IPT produced better outcomes than did prison treatment usual, and prison counselors learning IPT for the trial (some of whom had no prior formal psychotherapy training) had high IPT treatment adherence and acceptable IPT treatment competence [ 14 ]. However, 72% of the costs of IPT in this trial were for training and supervision [ 14 ]. Given extremely sparse prison mental health training budgets, this is not sustainable and scalable. On the other hand, single workshops typically have little effect on provider competence [ 21 ]. Lessons learned from the current analysis can inform and help to optimize training procedures that are both effective and more efficient for wider-scale implementation.
Integration of current findings with findings from other low-resource settings
The need for more scalable approaches to psychotherapy training and supervision extend beyond prisons to other high-stress, low-resource settings such as refugee settlements, low- and middle-income countries, or World Health Organization Mental Health Gap Action Programs (mhGAP [ 32 ]). Similar to the U.S. prison context, implementation studies in these settings also highlight a lack of trained and qualified mental health professions and resource constraints as key barriers for scaling evidence-based mental health practices [ 33 , 34 ]. Both our prison work and global mental health efforts have used remote supervision and task-shifting to lay counselors [ 35 ]. Our remote supervision was offered individually; other efforts [ 35 – 39 ] have successfully used remote group supervision. Giving counselors online access to example sessions [ 35 ] and recordings of the original training [ 38 ] for review can save some trainer/supervisor time. One project [ 37 ] asks counselors watch pre-recorded training segments prior to the virtual training. In the virtual training, after group review and questions, the counselors independently rate fidelity of pre-recorded “better” and “worse” sessions and discuss their answers with each other and the trainer. This helps to demonstrate interactions that do and do not fit within the psychotherapy model. Recording, automating, and using technology for training and supervision to the extent possible can help [ 40 ]. Many models now post freely available training and example videos online. Finally, train-the-trainer approaches or self-organized learning collaboratives can help to scale expertise. For example, experienced lay mental health counselors can play an important role in supervising other lay mental health counselors [ 41 ] and peer-to-peer supervision can be a sustainable quality assurance method [ 35 ]. However, all or almost all of these approaches have relied on outside (often research) funds to provide training, supervision, and technical assistance. Additional capacity-building approaches are needed.
Conclusions and recommendations
We offer several recommendations for training professional and paraprofessional counselors in IPT. Based on current and previous [ 14 , 22 ] findings from this trial, we recommend an initial training and at least one refresher training 6 months later that focus on: (1) the steps of IPT case conceptualization as outlined in Table 2 (how to involve clients in treatment planning while still guiding treatment; understanding which components should be counselor contributions and which should be client contributions to the plan), (2) how the steps of addressing each problem area (e.g., tell the story, feel the feelings, explore what is gained and what is lost, hold the memory and find new supports in the present) guide when to change gears or stay on course in a planned session, and (3) psychotherapy basics (reflective listening, open-ended questions, getting specifics of experiences and communication). We recommend providing supervision for at least the first two rounds of groups, including review of written case conceptualizations and at least some audio recordings. However, future efforts to streamline these processes are needed.
Finally, on a larger scale, incarcerating fewer people with mental illness and providing money in prison mental healthcare contracts for more counselors and more ongoing training and supervision would help address the resource shortages that make mental healthcare in prison settings so challenging [ 11 ], as demonstrated in this and previous [ 14 , 22 ] analyses. More low-cost, scalable methods of training and supervision are needed to get mental health treatment to individuals who need it most, including those who are incarcerated and other mental health equity populations [ 14 , 22 ].
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Deconditioning caused by physical inactivity has aroused concern in Western societies. Physical inactivity has been associated with several chronic diseases of which non-specific low back pain (LBP) is one of the most predominant [ 1 ]. Low back pain is defined as pain in the lumbar or gluteal region with or without radicular pain to the lower extremities. The vast majority of individuals with pain affecting the lower back has no specific diagnosis and is categorized as having non-specific LBP [ 2 ], i.e. the cause of non-specific LBP is usually unknown. Non-specific LBP is common in adolescence [ 3 – 6 ], and the occurrence of LBP increases with age until adulthood. Moreover, there is strong evidence to suggest that adult LBP originates in adolescence [ 3 ]. Low back disorders are the most prevalent musculoskeletal health concerns in populations and can cause varying degrees of disability [ 7 ]. In adult populations, psychosocial difficulties [ 8 , 9 ], smoking [ 10 ], overweight [ 10 ], sleep disturbances [ 11 ], and poor self-rated health [ 9 ] have been suggested to be risk factors for LBP leading to increased disability. LBP is also a very common disorder among military personnel causing disability, lost worker productivity, and increased health care costs [ 12 ].
Low back pain is a common condition worldwide, and it has been estimated that its prevalence will increase substantially in future [ 13 , 14 ]. A systematic review of population-based studies estimated the global point prevalence of LBP to be 12%, with a 1-month prevalence of 23% [ 14 ]. The overall mean prevalence was 31%, the one-year prevalence was 38%, and the lifetime prevalence was 40%. Women had significantly higher point and annual prevalence; although no gender differences were seen for the 1-year or lifetime prevalence of LBP [ 14 ]. The prevalence of LBP did not increase in men after 40 years of age, while the prevalence peaked later among women. In addition, LBP has been reported to be as prevalent in middle-aged populations as among those aged 60 years or over, but the prevalence seems to decline among the oldest individuals [ 15 ]. Another systematic review concluded that the prevalence of severe forms of back pain continues to increase with age, whereas less severe back pain becomes less common after reaching a peak when the individual is 50 to 60 years of age [ 16 ]. Moreover, there are some reports that suggest the prevalence of non-specific LBP decreases with age [ 17 , 18 ].
The relation between physical activity and LBP has been demonstrated to be controversial because both high [ 19 – 21 ] and low amounts and / or intensities [ 22 , 23 ] of physical activity have been recognized as risk factors for LBP. In addition, it has been shown that individuals with chronic LBP have reduced physical fitness levels compared with healthy asymptomatic subjects [ 24 ]. Moreover, it has been stated that chronic LBP sufferers have lower aerobic capacity and a higher fat percentage than their healthy controls [ 25 ]. It should be pointed out, however, that these findings mainly originate from a cross-sectional study setting, and thus may be biased. The question whether deconditioning is the cause of LBP or whether LBP contributes to physical inactivity has been raised [ 26 ].
Low back pain leads to care-seeking behaviour and reduced health-related quality of life in adolescence [ 27 ]. Furthermore, LBP in late adolescence has been reported to predict LBP in adulthood [ 28 , 29 ]. LBP is believed to be prevalent throughout life and is often recurring [ 30 , 31 ]. When assessed for pain intensity, quality of life, disability, or health care utilization, patients with radiating LBP tend to have a poorer prognosis when compared with patients suffering from non-radiating LBP [ 32 ].
The relationship between physical activity levels and LBP was recently assessed in a review article by Hendrick and coworkers [ 33 ]. In the review article, twelve studies were identified, of which five were cross-sectional ones. They suggest that physical activity and LBP may not be associated. However, they state that there is a need for high-quality longitudinal studies. Thus, the aim of this present study was to examine the association between physician diagnosed un-specified LBP in healthy males during compulsory military service, self-reported LBP, and physical fitness levels measured on average four years after military service. We hypothesize that LBP during military service predicts lower levels of physical fitness and LBP in later life.
Subjects and methods
In Finland, military service lasting from 6 to 12 months is compulsory for all male citizens above 18 years of age, and approximately 80% of Finnish males complete military service. The special characteristics of military training are the intensity and the volume of physical training and activities, since one of the main goals of the training is to improve the physical performance of conscripts. Because military service in Finland is compulsory, the epidemiological figures can be generalized quite well to the young adult male population.
As described in detail previously [ 34 ], military service begins with an 8-week basic training (BT) period comprising 135 hours of various types of physical training. In addition, conscripts also perform 56 hours of programmed physical training during the BT period. During combat training, every conscript must complete long marches carrying personal combat gear weighing 25–35 kg. After the BT period, the amount of moderate and high-intensity programmed physical training is slightly reduced, but the intensity and volume of the marches carrying heavy combat gear increase.
During military service, all conscripts can use the medical services provided by military health care and hospitals. After careful clinical examination and the necessary diagnostic tests and imaging, the most accurate diagnosis is selected by a physician according to the 10th Revision of the International Classification of Diseases and Related Health Problems (ICD-10). All visits by conscripts to a military physician due to unspecified LBP with the ICD-10 diagnostic code M54.5, M54.4 or M54.9 are recorded. For the present study, a computer search using these ICD-10 diagnostic codes was conducted. The original, completed medical records were retrieved and reviewed to confirm the accuracy of the diagnoses and to systematically collect data for the present study. LBP that occurred during the conscript's leisure time or on the way from or to the garrison was included in the collected data. The length of absence from duty due to LBP was also recorded.
The subjects selected for this study were 1155 persons who had passed the entry medical examination for military service and who had completed physically demanding military training between 1997 and 2007. Of these, 920 participated in a refresher military training course. The most common reasons for nonparticipation in the refresher course were work, study, or health-related issues. From a total of 920 males, 778 volunteered (mean ± SD age 19.9±4.6 yrs., height 1.80±0.06m, body mass 80.3±13.4kg, and body mass index 24.7±3.8) to take part in the present study. The mean follow-up period, the period between the end of military service and participation in the refresher course, of the subjects was approximately four years (ranging from one to eleven years). The subjects were informed of the study protocol and written informed consent was obtained. The ethical statement of the present study was given and approved by the ethical committee of the Central Finland Health Care District, Jyväskylä, Finland (K-S shp:n Dnro 34/2007, 21.8.2007).
Questionnaire
Measurements were carried out in 8 different sessions during the refresher course in 2008 (March-November). During the refresher course, subjects completed a questionnaire that included questions on socioeconomic background, physical activity, and health behaviours. The urbanisation level of residence was determined by population density using four categories: city/large town (population over 90,000), small town, village (densely populated area in rural municipalities), and sparsely populated rural municipality (isolated homestead in rural municipalities). Three categories of achieved level of education were used: comprehensive school, vocational school and upper secondary school, or university. Marital status was categorised as single or other. Self-estimated health status was categorised as good, average, and poor. Smoking and alcohol consumption habits were assessed by questions about daily smoking and the frequency of consumption of more than 6 units of alcohol per day (less or more frequently than once a month). LBP was assessed with two questions: 1) Assess how many days all together you have had LBP that has radiated to the lower extremity below the knee during the last month? Alternatives: None, 1–7 days, 8–14 days, more than 14 days but not daily, daily. 2) Assess how many days all together you have had acute LBP (e.g. lumbago) during the last month? Alternatives: None, 1–7 days, 8–14 days, more than 14 days but not daily, daily. LBP during the last month (yes/no) and the intensity of LBP during the last week were studied using the Visual Analogue Scale (VAS).
Physical fitness tests
Maximal oxygen uptake (VO 2 max) was indirectly predicted during a bicycle ergometer test (Ergoline 800 S, Ergoselect 100 K or 200 K, Bitz, Germany) [ 35 ]. After a 5-min warm up, the test began with a power output of 75 W, which was increased by 25 W after every second minute. The pedalling rate of 60 rpm was maintained constant throughout the test. The heart rate (HR) was recorded continuously (Polar Vantage NV or S610, S710 or S810, Kempele, Finland). The test was terminated at volitional exhaustion. Predicted VO 2 max was determined from HR and power (Fitware, Mikkeli, Finland) as follows: VO 2 max (ml·kg -1 ·min -1 ) = 12.35*P max /kg + 3.5, where P max is maximal power in relation to body mass. During their military service, they had also run the 12-min running test. Conscripts were instructed to perform the test with a maximal effort but at progressively increasing running speed. The accuracy of the measurements was ±10 meters.
Muscle fitness was measured by tests of grip strength, sit-ups, push-ups, repeated squats, and maximal isometric leg and arm extensions. Isometric grip strength was measured twice in a sitting position (90° elbow angle) with a dynamometer (Saehan Corporation, Masan, South Korea). The best results for the right and left hands were averaged for the final outcome [ 36 ]. The results of the push-ups, sit-ups, and repeated squats were expressed as the number of correctly performed repetitions within 60 s. The detailed descriptions of these tests have been published earlier [ 37 ].
Maximal isometric leg and arm extension forces (leg and bench press) were measured bilaterally using dynamometers. Knee angle was set to 107°. During the maximal bench press, participants were in a supine position on a bench with their feet on the floor and their elbows positioned at an angle of 90°. A total of three maximal trials were performed with a 30 s recovery period between trials. The best performance recorded was included for further analysis. The participants were also instructed to produce maximal strength as fast as possible and to maintain it for 3 s. Maximal force and force production time were collected with an AD-converter (CED power 1401, Cambridge Electronic Design, Ltd, England) at a frequency of 1 kHz, on a computer [ 38 ]. More details on the testing procedures can be found in an article by Vaara et al. [ 38 ]
Statistical analysis
During the first stage, two-way tables were calculated for our main variable (yes/no physician visit due to low back pain during military service) and the categorized variables measured during the refresher training course after a mean four-year follow-up. A chi-square test was used with the level of significance defined as p = 0.05. Next, separate adjusted (adjusted by age, urbanisation level of residence, family composition, level of education, smoking, drinking) logistic regression models for physical fitness and the LBP variables measured at the follow-up were calculated in order to assess whether physician diagnosed LBP during military service is associated with levels of physical fitness or LBP later in life. Odds ratios (OR) were estimated with 95% confidence intervals (95% CI).
Logistic regression analyses were only performed for respondents who provided answers to every question, and thus respondents with incomplete answers were excluded from the analysis. The frequency of missing values for the independent variables varied from 2 to 8%.
Results
Of the 778 participants who had completed their military service between 1997 and 2006, 219 (28%) had visited a physician due to musculoskeletal symptoms (ICD-10 M-diagnosis) during their military service. The mean duration of absence from duty was 1.2 (range 1 to 44) days. Seventy-four (9.5%) conscripts had visited a physician due to unspecified LBP during their military service, and 41 (5.3%) had temporarily been absent from duty due to LBP. The mean duration of absence from duty due to LBP per physician visit was 2.6 days (range 1 to 6).
From the self-reported variables at the follow-up in 2008, level of education (p = 0.23), marital status (p = 0.31), alcohol consumption (p = 0.62), daily smoking (p = 0.54), or self-estimated perceived health (p = 0.33) were not associated with absence from duty due to LBP. In addition, the result of the 12-min running test during the military service was not associated to absence from duty due to LBP (p = 0.71).
When muscle strength and aerobic capacity in the follow-up and absence from duty due to LBP during military service were assessed ( Table 1 ), it was observed that absence from duty due to LBP during military service did not predict later physical performance, including aerobic capacity and muscle fitness.
10.1371/journal.pone.0173568.t001
Table 1 Odds ratios for absence from military service due to low back pain.
Characteristics
Adjusted * OR
Maximal oxygen uptake (l/min; by quartiles)
First (1.84 to 2.86)
1
Second (2.87 to 3.25)
1.2 (0.4–3.6)
Third (3.26 to 3.68)
2.5 (0.9–7.2)
Fourth (3.69 to 5.73)
1.4 (0.4–4.8)
Grip strength (kg; by quartiles)
First (31.5 to 46.5)
1
Second (46.6 to 52.0)
0.6 (0.3–1.7)
Third (52.1 to 58.5)
1.0 (0.4–2.3)
Fourth (58.6 to 85.5)
0.8 (0.3–2.0)
Sit-ups (reps/min; by quartiles)
First (2 to 31)
1
Second (32 to 39)
1.0 (0.4–2.2)
Third (40 to 45)
1.1 (0.4–2.9)
Fourth (46 to 72)
0.7 (0.2–2.2)
Push-ups (reps/min; by quartiles)
First (1 to 19)
1
Second (20 to 27)
0.9 (0.4–2.4)
Third (28 to 38)
0.9 (0.3–2.3)
Fourth (39 to 75)
0.4 (0.1–1.4)
Repeated squats (reps/min; by quartiles)
First (3 to 39)
1
Second (40 to 44)
0.6 (0.2–1.7)
Third (45 to 50)
0.9 (0.4–2.2)
Fourth (51 to 64)
0.8 (0.3–2.1)
Maximal leg extension (kg; by quartiles)
First (89 to 235)
1
Second (236 to 277)
0.8 (0.3–2.1)
Third (278 to 337)
1.1 (0.4–2.8)
Fourth (338 to 740)
0.9 (0.3–2.4)
Maximal bench press (kg; by quartiles)
First (48 to 76)
1
Second (77 to 87)
0.8 (0.3–2.0)
Third (88 to 110)
0.6 (0.2–1.6)
Fourth (111 to 163)
1.1 (0.5–2.7)
*Adjusted for age, BMI, level of education, marital status, drinking and daily smoking
Of the 788 respondents in the follow-up examination, 122 (15.7%) had reported LBP during the past month. LBP during military service was associated with self-reported LBP in the follow-up (p = 0.004). Of those who had been absent from duty due to LBP during military service, 13 (31.7%) reported LBP during the past month while the corresponding figure was 109 (14.8%) among those who had no LBP absence from duty.
In the follow-up, LBP during the month before the refresher military training course was not associated with temporary absence from duty due to LBP during one-week physically strenuous military training (p = 0.71). The mean LBP VAS during the military training week was 1.5 for those who had no absence from duty due to LBP and 1.9 for those who had absence from duty due to LBP. However, these results were not statistically significant (p = 0.17).
Discussion
The main finding of the present study was that unspecified LBP during military service predicts LBP in later life. An earlier study has shown that conscripts who suffer from chronic LBP before entering military service have a ten-fold higher risk of experiencing LBP during military service compared to the risk before military service [ 39 ]. A history of suffering from LBP seems to predict a later LBP episode. Thus, young men who suffer LBP during their military service should not be directed towards occupations that require a symptomless low back without special rehabilitation and muscle fitness training. According to the findings of Suni et al., the risks for LBP can be reduced by education and muscle fitness training [ 40 ]. They noticed that exercise and education improved the control of the lumbar neutral zone. This could have a prophylactic effect on LBP-related off-duty service days in the military environment when implemented as part of military service among young healthy men. However, more research data and knowledge are also needed to establish how physical activity and fitness associate with LBP.
The one-month occurrence of LBP in our study (16%) was lower than one-month occurrence described previously in a review article (23%) [ 14 ]. This is probably due to our healthy subjects that had passed physically demanding military service. In addition, they described the mean VAS scale to be 1.5 to 1.9 which can be considered to be relative low. Interestingly, none of our background variables, health behaviour, or the physical performance at the follow-up was associated with baseline LBP in the present study. Thus, focusing LBP preventive measures on those variables does not seem to be useful. A previous randomized controlled study on the effect of using orthotic insoles during military service concluded that orthotic insoles should not to be used with the aim of toward preventing LBP episodes in young male adults [ 41 ]. A previous longitudinal population-based study found that being overweight or obese in early adulthood as well as during later years increases the risk of radiating but not non-specific LBP among men [ 42 ]. The results of the present study also contradict the findings of Taanila et al. [ 43 ] who found that the risk for LBP increased among young men who had a lower educational background and lower levels of both aerobic and muscular performance.
The compulsory military service for all Finnish male citizens above 18 years of age differs from professional armies. In a conscription army, the volume and intensity of physical training have to be carefully adjusted to the conscripts’ fitness level. Therefore, the present results cannot be directly extrapolated to professional armies. However, due to the obligatory nature of military service in Finland, the epidemiological figures can be generalized quite well to the young adult male population because approximately 80% of men complete the 6 to 12 month military service period.
The strengths of the present study include the fact that the original cohort of conscripts comprised a significant number of individuals who were obliged to use the medical services provided by the Finnish Defence Forces for the management of LBP. Furthermore, all conscripts had passed two physician-performed medical examinations when entering military service. Thus, all patients with severe back diseases, such as significant scoliosis or other congenital back anomalies, severe post-traumatic disorders, rheumatoid arthritis, or ankylosing spondylitis were exempted from duty. Conscripts can thus be considered healthy young men. The accuracy and validity of the LBP during conscription is excellent, as the cohort included in this study had no alternative choices such as private clinics to attend due to LBP. Moreover, the authors conclude that the present results can very well be generalized to a young healthy male population owing to the compulsory nature of Finnish military service.
The present study also has some weaknesses. A limitation of the study is that although subjects were healthy men, their previous LBP episodes, level and recurrence before entering the military service was not known. Unfortunately, during the service it was not possible to identify less severe LBP episodes that did not result in a visit to a physician during conscription. Furthermore, only 778/1155 participated in the refresher course which provides a limitation to the study, since the prevalence of LBP in those not attending the course is not known. In addition, a mean follow-up period of four years was the midterm follow-up duration, but this time-frame was chosen due to the requirements of military service.
The main finding of the present study was that unspecified LBP during military service predicts LBP in later life. On the basis of previous literature, it is also known that LBP is a common symptom and thus, one cannot expect to be symptomless the entire life. Interestingly, none of the health behaviours nor the physical performance studied in the follow-up were associated with baseline LBP. It appears that individuals prone to LBP have symptoms during physically demanding military service and also later in their life.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Coronavirus disease 2019 (COVID-19) is an acute respiratory infectious disease caused by a novel coronavirus, called severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [ 1 – 3 ]. Due to the strong infectivity of the coronavirus and the lack of effective treatments and vaccines, COVID-19 has spread rapidly and become a pandemic worldwide with serious consequences. The social restriction policies implemented to control the infection in almost all countries have resulted in unemployment, financial hardship, and a reduction in quality of life in many countries [ 4 ], which have seriously affected the life and work of people. Simultaneously, the COVID-19 pandemic has significant implications for a global increase in mental health problems, and the psychological burden during the COVID-19 pandemic has been reported [ 5 , 6 ].
Social rhythm is defined as the regularity of timing of the behaviors in social activities and also described clinically as “daily routines” [ 7 ]. COVID-19-related social changes, particularly lockdown protocols, have indeed impacted the timing of daily behaviors in stabilizing circadian function, and social rhythm disruption may be correlated with negative psychological impacts [ 7 ]. Previous studies have reported that social rhythm disruption is connected with mental diseases, especially depression and bipolar disorder in the non-pandemic days, which has been explained by many theories [ 8 – 11 ]. Recently, several studies have reported social and circadian rhythm disruptions in students, office workers, the elderly, and patients with narcolepsy or autism [ 12 – 14 ]. However, so far it is not clear the characteristics of the disrupted rhythms of life, work and entertainment behaviors with different sociodemographic backgrounds and their associations with psychological impacts under the stress of the COVID-19 pandemic. Therefore, the aim of the present study was to evaluate the rhythm disruptions in life, work, and entertainment and their associations with the psychological impacts in the general Chinese population with different sociodemographic backgrounds during the initial phase of the COVID-19 pandemic.
Materials and methods
Participants and procedure
Chinese people with different sociodemographic backgrounds voluntarily participated in this online survey. An e-questionnaire was designed to investigate the rhythm activities and psychological impacts in Chinese people with different backgrounds including gender, age, educational background, annual income, health status, current occupation, and chronic disease status during the COVID-19 pandemic through EQxiu online questionnaire platform. The survey was conducted from 10–17 March 2020 in 22 provinces and five autonomous regions, four municipalities, and two special administrative regions ( i . e ., Hong Kong and Macao) of China. The study protocol was carried out following relevant guidelines and regulations and approved by the Ethics Committee of the Army Medical Center (Approval No. 2019–144). During the survey, the participants were asked to read and check the informed consent and then complete the questionnaire; those under 18 years old were asked to consent by their guardians before completing the questionnaire. All completed questionnaires were checked for validity and completeness.
Measures
Social rhythm
The Brief Social Rhythm Scale (BSRS) was developed to quickly assess social rhythmicity in large-scale multi-national samples and multiple languages [ 15 ]. The Chinese version of BSRS also included entertainment and physical practice, and was shown to have good reliability and validity in China [ 16 ]. The scale consisting of 17 items was used to investigate an individual’s rhythm in six aspects, i. e ., sleeping (going to bed), getting up, socializing (SGS), eating, physical practice, and entertainment (EPE) on weekdays and weekends. The first eight items addressing SGS rhythm constituted SGS-scale, and the last nine items addressing EPE rhythm constituted EPE-scale. Each item was assigned with a score of 1, 2, 3, 4, 5, or 6, corresponding to categories of very regularly, quite regularly, somewhat regularly, somewhat irregularly, quite irregularly, or very irregularly, respectively. The lower the score, the more regular the rhythm.
Zung’s depression and anxiety scale
Zung’s self-rating depression scale (SDS) and Zung’s self-rating anxiety scale (SAS) were used to evaluate the psychological characteristics of each participant [ 17 , 18 ]. There were 20 items in SAS and SDS, and each item was rated on a scale of 1–4 reflecting never, often, sometimes, and always. The raw score was the sum of 20 questions multiplied by 1.25 and rounded to get the standard score as described in previous studies [ 19 ]. The cut-off value was 50 for both SAS and SDS. The higher the score, the more obvious the anxiety or depression was, with the scores of 50–59, 60–69, and 70 points or more indicating mild, moderate, and severe anxiety or depressive symptom, respectively. The Chinese version of SDS and SAS, which was used in the present study, was reported to present good reliability, validity and excellent internal consistency, with the Cronbach coefficient of 0.784 for SDS in Chinese people [ 20 ], and of 0.931 for SAS in the general population [ 21 ].
Statistical analysis
Data were analyzed using SPSS version 23.0 (SPSS, Chicago, IL, USA). The data of normal distribution was expressed by mean±standard deviation. One-way analysis of variance with a post-hoc analysis, was used to make comparisons among groups. T-test or a nonparametric test ( i . e ., Mann-Whitney U test or Kruskal-Wallis test), where appropriate, was used to make comparisons between the group with the highest score and each of the other groups for each item. Pearson correlation or Spearman correlation, where appropriate, was used to determine the correlation between the BSRS scales and the Zung’s scales. In addition, a multivariate linear regression analysis was used to determine the independent factors for the correlation of anxiety or depression with social rhythm. Cohen’s d, an indication of effect size, was generated and reported to further indicate the strength of the difference. A P value of < 0.05 was considered statistically significant (corrected for multiple comparisons where appropriate).
Results
Sociodemographic characteristics of the participants
A total of 5872 people checked the e-questionnaire. Eighteen questionnaires with invalid data were excluded due to frivolous or incomplete answers, or the same option for different questions or orderly answers. The data of 5854 participants with valid questionnaires were further analyzed. The general information on the sociodemographic characteristics of the participants is listed in Table 1 .
10.1371/journal.pone.0250770.t001
Table 1 General sociodemographic characteristics of 5854 analyzed participants.
Feature
N (%)
Feature
N (%)
Gender
Female
3439(58.75%)
Health status
Poor
53(0.91%)
Male
2415(41.25%)
Normal
787(13.44%)
Age
-18
68(1.16%)
Good
3686(62.97%)
18–25
1960(33.48%)
Very good
1328(22.69%)
26–30
862(14.72%)
Current occupation
Businessman
481(8.22%)
31–40
1339(22.87%)
Officer
245(4.19%)
41–50
852(14.55%)
Teacher
231(3.95%)
51–60
682(11.65%)
Police
277(4.73%)
61-
91(1.55%)
Farmer
217(3.71%)
Marital status
Unmarried
2435(41.60%)
Employee
270(4.61%)
Married
3232(55.21%)
Doctor
1171(20.00%)
Divorced /widowed
187(3.19%)
Nurse
1017(17.37%)
Chronic disease
No chronic disease
4373(74.70%)
Medical technician
139(2.37%)
CDCPD
671(11.46%)
Retired re-employee
229(3.91%)
Chronic diseases only
810(13.84%)
Non-medical student
1172(20.02%)
Annual income
Less than 50000 CY
2482(42.40%)
Medical student
405(6.92%)
50000–100000 CY
1849(31.59%)
Education
Junior high school or lower
97(1.66%)
100000–200000 CY
1188(20.29%)
Senior high school
535(9.14%)
200000–300000 CY
225(3.84%)
Junior college or undergraduate
4465(76.27%)
More than 300000 CY
110(1.88%)
Master or above
757(12.93%)
CY, Chinese yuan; CDCPD, chronic diseases comorbid with psychosomatic diseases.
Disruptions of social rhythm in the participants
Generally, there were significant differences in SGS-scale and EPE-scale among participants with different sociodemographic backgrounds including gender, age, educational background, annual income, health status, current occupation, and chronic disease status ( P <0.05, Table 2 ). Moreover, the SGS-scale and EPE-scale scores were compared between the group with the highest score and each of other groups ( Table 3 ). Specifically, the female gender and poor health status were most closely associated with disrupted rhythms of life, work, and entertainment. The age group of 26–30 years, nurses and subjects with divorce or widow status, the education level of senior high school, annual pre-tax income of 50,000–100,000 Chinese Yuan (CY), or chronic disease comorbid with psychosomatic diseases mostly suffered from disrupted rhythms of life and work. The age group of older than 61, and subjects with education levels of master’s degree or above, annual pre-tax income over 300,000 CY, or chronic disease without psychosomatic diseases mostly suffered from disrupted rhythms of entertainment. Especially, nurses showed significantly more severe disruption of SGS rhythm, whereas the elderly reported significant irregularity in EPE rhythm ( Table 2 ).
10.1371/journal.pone.0250770.t002
Table 2 Scores of social rhythms in participants with different sociodemographic backgrounds during the COVID-19 pandemic (N = 5854).
Feature
SGS-scale
EPE-scale
Mean±SD
Z/X2
P
Cohen’s d
Mean±SD
Z/X2
P value
Cohen’s d
Gender
Female
21.64±9.03
-10.183
<0.001
0.269
7.71±6.02
-7.627
<0.001
-
Male
19.24±8.80
-
8.71±5.93
0.167
Age
-18
17.63±8.69
146.237
<0.001
-
7.82±5.58
112.103
<0.001
0.102
18–25
18.98±8.95
0.153
7.75±5.35
0.091
26–30
23.04±9.77
0.585
7.22±6.22
-
31–40
21.54±8.95
0.443
7.76±6.18
0.087
41–50
20.79±8.56
0.366
9.10±6.68
0.291
51–60
20.74±7.91
0.374
9.64±5.82
0.402
61-
21.27±8.78
0.417
9.87±6.14
0.429
Marital status
Unmarried
19.53±9.19
81.065
<0.001
-
7.81±5.52
2.499
0.294
-
Married
21.42±8.77
0.210
8.32±6.26
0.086
Divorced /widowed
22.04±9.21
0.273
8.76±7.25
0.147
Education
Junior high school or lower
18.92±10.07
8.880
0.031
-
7.81±7.18
11.172
0.011
0.009
Senior high school
21.18±8.97
0.237
7.75±5.97
-
Junior college or undergraduate
20.64±9.06
0.180
8.05±5.89
0.051
Master or above
20.59±8.61
0.178
8.85±6.46
0.177
Annual income
Less than 50000 CY
19.21±8.90
126.387
<0.001
-
7.88±5.64
48.455
<0.001
0.034
50000–100000 CY
22.18±9.26
0.327
7.68±6.06
-
100000–200000 CY
21.28±8.49
0.238
8.93±6.28
0.203
200000–300000 CY
20.66±8.41
0.167
9.31±6.03
0.270
More than 300000 CY
20.78±9.03
0.175
9.86±8.12
0.304
Health status
Poor
28.02±10.00
500.564
<0.001
1.178
9.53±8.85
26.969
<0.001
0.274
Normal
25.21±9.48
0.906
7.94±7.19
0.065
Good
20.93±8.36
0.468
8.35±5.90
0.146
Very good
16.90±8.85
-
7.53±5.30
-
Current occupation
Businessman
20.21±8.50
322.663
<0.001
0.302
8.39±5.92
177.273
<0.001
0.322
Officer
19.89±8.18
0.269
9.32±5.94
0.476
Teacher
21.06±8.41
0.404
9.74±6.77
0.510
Police
18.69±8.79
0.121
9.05±5.46
0.448
Farmer
19.81±9.55
0.239
7.47±6.17
0.165
Employee
20.49±7.99
0.345
8.44±5.81
0.333
Doctor
20.99±8.27
0.399
8.52±6.39
0.330
Nurse
24.47±9.91
0.740
6.46±6.08
-
Medical technician
21.31±8.65
0.428
7.63±6.56
0.185
Retired re-employee
21.63±8.78
0.462
9.62±5.28
0.555
Non-medical student
17.65±8.46
-
7.90±5.20
0.255
Medical student
20.64±8.80
0.346
8.66±6.08
0.362
Chronic disease
No chronic disease
19.71±8.82
202.421
<0.001
-
7.94±5.67
6.269
0.044
-
CDCPD
24.01±9.18
0.478
8.67±7.40
0.111
Chronic diseases only
22.98±8.82
0.371
8.68±6.37
0.123
SD, standard deviation; CY, Chinese yuan; SGS-scale: the rhythm of sleep, getting-up, and socializing; EPE-scale: the rhythm of eating, physical practice, and entertainment activities; CDCPD, Chronic diseases comorbid with psychosomatic diseases.
10.1371/journal.pone.0250770.t003
Table 3 Comparison of SGS-scale and EPE-scale scores between the group with the highest score and each of the other groups in terms of sociodemographic backgrounds.
Features
SGS-scale
EPE-scale
Highest-score group
Other groups
P- value
Cohen’s d
Highest-score group
Other groups
P- value
Cohen’s d
Marital status
Divorced /widowed
Unmarried
0.001
0.273
Divorced /widowed
Unmarried
0.227
0.147
Married
0.748
0.069
Married
0.802
0.065
Chronic disease
CDCPD
No chronic disease
<0.001
0.478
Chronic diseases only
No chronic disease
0.006
0.123
Chronic diseases only
0.086
0.114
CDCPD
1.000
0.001
Education
Senior high school
Junior high school or lower
0.222
0.237
Master or above
Junior high school or lower
0.695
0.152
Junior college or undergraduate
0.720
0.060
Senior high school
0.010
0.177
Master or above
0.811
0.067
Junior college or undergraduate
0.009
0.129
Health status
Poor
Normal
0.275
0.288
Poor
Normal
0.749
0.197
Good
<0.001
0.769
Good
0.918
0.157
Very good
<0.001
1.178
Very good
0.499
0.274
Annual income
50000–100000 CY
Less than 50000 CY
<0.001
0.327
More than 300000 CY
Less than 50000 CY
0.119
0.283
100000–200000 CY
0.056
0.101
50000–100000 CY
0.063
0.304
200000–300000 CY
0.113
0.172
100000–200000 CY
0.938
0.128
More than 300000 CY
0.714
0.153
200000–300000 CY
0.999
0.077
Age
26–30
-18
<0.001
0.585
61-
-18
0.473
0.349
18–25
<0.001
0.433
18–25
0.034
0.368
31–40
0.006
0.160
26–30
0.003
0.429
41–50
<0.001
0.245
31–40
0.042
0.343
51–60
<0.001
0.259
41–50
0.998
0.120
61-
0.799
0.191
51–60
1.000
0.038
Current occupation
Nurse
Businessman
<0.001
0.461
Teacher
Businessman
0.482
0.212
Officer
<0.001
0.504
Officer
1.000
0.066
Teacher
<0.001
0.371
Police
1.000
0.112
Police
<0.001
0.617
Farmer
0.015
0.350
Farmer
<0.001
0.479
Employee
0.781
0.206
Employee
<0.001
0.442
Doctor
0.551
0.185
Doctor
<0.001
0.381
Nurse
<0.001
0.510
Medical technician
0.007
0.340
Medical technician
0.199
0.317
Retired re-employee
0.001
0.303
Retired re-employee
1.000
0.020
Non-medical student
<0.001
0.740
Non-medical student
0.008
0.305
Medical student
<0.001
0.409
Medical student
0.953
0.168
CY, Chinese yuan; SGS-scale: the rhythm of sleep, getting-up, and socializing; EPE-scale: the rhythm of eating, physical practice, and entertainment activities; CDCPD, Chronic diseases comorbid with psychosomatic diseases.
Psychological impacts in the participants
There were significant differences in SAS and SDS scores among participants with different genders, ages, marital status, education levels, annual pre-tax incomes, health status, current occupation, and chronic diseases with or without psychosomatic diseases in each of the variables (all P <0.05) ( Table 4 ). Specifically, nurses and participants with female gender, divorce or widow status, education levels of junior high school or below, poor health status, and chronic diseases comorbid with psychosomatic diseases mostly suffered from depression and anxiety. Whereas the age group of 26–30 years, and participants with annual pre-tax income of 50,000–100,000 CY mostly suffered from depression, the age group of 31–40 years and participants with annual pre-tax income over 300,000 CY mostly suffered from anxiety. Additionally, SDS scores were significantly higher in medical students than in non-medical students. There were significant differences in the prevalence of depression and anxiety among participants with different sociodemographic backgrounds (all P <0.05, S1 and S2 Tables). The overall prevalence of depression and anxiety was 24.33% and 12.64%, respectively, as defined by SDS or SAS scores over 50. Among the different occupational groups, nurses had the highest rates of depression (32.94%) and anxiety (18.98%). Participants with chronic diseases combined with psychosomatic diseases had the highest rates of depression (45.90%) and anxiety (31.45%).
10.1371/journal.pone.0250770.t004
Table 4 Scores of depression and anxiety in participants with different sociodemographic backgrounds during the COVID-19 pandemic (N = 5854).
Feature
SDS
SAS
Mean±SD
Z/X2
P
Cohen’s d
Mean±SD
Z/X2
P
Cohen’s d
Gender
Female
43.34±11.66
-11.870
<0.001
0.290
38.80±9.39
8.441
<0.001
0.173
Male
39.97±11.59
-
37.19±9.23
-
Age
-18
39.34±10.93
63.393
<0.001
-
35.09±8.35
266.138
<0.001
-
18–25
40.13±11.63
0.070
35.82±8.27
0.088
26–30
43.87±12.20
0.391
39.66±9.97
0.497
31–40
43.59±11.68
0.376
39.84±9.76
0.523
41–50
42.48±11.89
0.275
39.40±9.94
0.470
51–60
41.27±10.76
0.178
38.12±8.61
0.357
61-
41.03±10.97
0.154
38.91±9.32
0.432
Marital status
Unmarried
40.65±11.79
63.778
<0.001
-
36.26±8.56
222.618
<0.001
-
Married
42.78±11.60
0.182
39.39±9.61
0.344
Divorced /widowed
44.57±12.05
0.329
40.87±10.63
0.478
Education
Junior high school or lower
45.22±13.29
12.821
<0.05
0.281
39.85±10.43
10.900
<0.001
0.196
Senior high school
42.83±11.37
0.097
38.57±9.38
0.070
Junior college or undergraduate
41.71±11.67
-
37.92±9.24
-
Master or above
42.31±12.17
0.050
38.86±9.85
0.098
Annual income
Less than 50000 CY
40.73±11.74
79.210
<0.001
0.107
36.49±8.86
204.895
<0.001
-
50000–100000 CY
43.54±11.65
0.365
39.72±9.45
0.353
100000–200000 CY
42.46±11.79
0.264
39.17±9.67
0.289
200000–300000 CY
39.56±10.12
-
36.93±8.02
0.052
More than 300000 CY
42.19±12.56
0.231
40.05±11.42
0.348
Health status
Poor
62.90±11.77
778.373
<0.001
2.342
55.94±12.3
827.454
<0.001
2.112
Normal
49.73±12.34
1.149
44.56±10.55
1.120
Good
42.00±10.57
0.526
37.95±8.48
0.466
Very good
36.36±10.88
-
34.14±7.86
-
Current occupation
Businessman
40.42±10.59
249.001
<0.001
0.187
37.29±8.37
374.507
<0.001
0.355
Officer
40.25±9.25
0.182
37.86±7.63
0.449
Teacher
41.58±11.24
0.287
38.58±9.38
0.485
Police
38.89±11.34
0.042
36.71±8.89
0.271
Farmer
43.02±11.59
0.411
37.97±9.14
0.419
Employee
42.08±10.68
0.341
38.25±8.34
0.478
Doctor
43.85±12.39
0.468
40.12±10.32
0.628
Nurse
45.17±12.35
0.583
40.41±10.14
0.668
Medical technician
44.74±11.84
0.558
40.14±9.77
0.653
Retired re-employee
42.24±10.47
0.359
38.98±8.61
0.561
Non-medical student
38.43±10.73
-
34.50±7.32
-
Medical student
41.93±12.18
0.305
37.93±9.59
0.402
Chronic disease
No chronic disease
40.27±11.12
363.830
<0.001
-
36.48±8.40
36.616
<0.05
-
CDCPD
48.86±12.70
0.720
44.88±10.97
0.860
Chronic diseases only
45.27±11.42
0.444
41.47±9.45
0.558
SD, standard deviation; CY, Chinese yuan; SDS, Zung’s self-rating depression scale; SAS, Zung’s self-rating anxiety scale; CDCPD, Chronic diseases comorbid with psychosomatic diseases.
Correlations between the social rhythm and psychological impacts
Among the 5854 participants, mean scores on SGS-scale and EPE-scale were 20.65±9.01 and 8.12±6.00, respectively, and mean SAS score and mean SDS score were 30.51±7.49 and 41.95±11.75, respectively. Spearman correlation analysis found that SGS scale was moderately correlated with depression and anxiety, with the correlation coefficients of 0.550 and 0.544, respectively, (both P <0.001, S3 Table ). In the multivariate linear regression analysis where the scores of SPS-scale or EPE-scale were taken as dependent variables, the score of SDS and SAS were taken as independent variables, and the age, gender, education, and occupation status were taken as confounding factors, the correlations between the scores of SPS-scale and EPE-scale and the scores of SDS and SAS remained unchanged (all P <0.05, Table 5 , S1 and S2 Figs). The disruption of sleep, getting-up, and socializing rhythm was positively associated with depression and anxiety ( β = 0.355, P <0.001; β = 0.186, P <0.001). The disruption of eating, physical practice, and entertainment rhythm showed opposite association to psychological impact: negatively related to depression and positively related to anxiety ( β = -0.049, P = 0.036; β = 0.082, P <0.001) ( Table 5 ).
10.1371/journal.pone.0250770.t005
Table 5 Correlations between the social rhythm disruption and psychological impacts as determined by multivariate linear regressive analysis.
Standardized coefficient (β)
t
P
EXP(B) 95% CI
Lower
Upper
SGS-scale
Model 1
SDS
0.381
20.080
<0.001
0.264
0.321
SAS
0.203
10.689
<0.001
0.160
0.231
Model 2
SDS
0.373
19.463
<0.001
0.258
0.315
SAS
0.200
10.385
<0.001
0.156
0.229
Model 3
SDS
0.355
18.370
<0.001
0.243
0.302
SAS
0.188
9.791
<0.001
0.145
0.217
Model 4
SDS
0.355
18.373
<0.001
0.244
0.302
SAS
0.186
9.645
<0.001
0.143
0.216
EPE-scale
Model 1
SDS
-0.082
-3.565
<0.001
-0.065
-0.019
SAS
0.109
4.780
<0.001
0.041
0.099
Model 2
SDS
-0.049
-2.129
0.033
-0.048
-0.002
SAS
0.082
3.582
<0.001
0.024
0.082
Model 3
SDS
-0.049
-2.101
0.036
-0.048
-0.002
SAS
0.081
3.513
<0.001
0.023
0.081
Model 4
SDS
-0.049
-2.101
0.036
-0.048
-0.002
SAS
0.082
3.518
<0.001
0.023
0.081
SGS-scale: the rhythm of sleep, getting-up, and socializing; EPE-scale: the rhythm of eating, physical practice, and entertainment activities; SDS/ SAS: Zung’s self-rating depression and anxiety scales.
Model 1: uncorrected; Model 2: corrective factors: gender, age, marital status, and education level; Model 3: corrective factors: based on the corrective factors of Model 2 and health status, annual pre-tax income, and current occupation; and Model 4: corrective factors: based on the corrective factors of Model 3 and psychosomatic diseases. The results of Model 4 were presented in this table
Discussion
The COVID-19 pandemic has brought great challenges and altered daily routines. It also undoubtedly contributes to emotional stress, fear, sadness, loneliness, anxiety and depression, and thus is associated with a global increase in psychological impacts [ 5 ]. One of the many mechanisms by which COVID-19-related changes in social rhythm impact mental health involves circadian disruption [ 7 ]. A previous study showed that the early identification and intervention of rhythm disruption are of important clinical significance [ 21 ]. Rhythm disruption is part of the pathogenic cascade and is associated with mood disorders, such as anxiety, depression, low happiness, strong loneliness, and bipolar disorder during the non-outbreak period of infectious disease [ 7 , 15 , 22 ]. Limited studies have found that sleep disruption is related to depression in the general population during the COVID-19 pandemic [ 11 ]. The present study found that the disruption of social rhythm was quite different among people with various sociodemographic backgrounds. The female gender, low-degree education level, lower or higher than average income, poor health status, age group of 26–30 years and older than 61 years, nurses and subjects with divorce mostly suffered from disruption of social rhythm. In addition, the present study observed that the participants with female gender, age group of 26–40 years, divorce or widow status, low education level, lower or higher than average income, poor health status, and chronic diseases comorbid with psychosomatic diseases, and nurses mostly suffered from depression and/or anxiety.
There are two major possible explanations for the observed gender differences in rhythm disruption, depression, and anxiety. First, the steroid receptors are expressed in almost every site that receives direct suprachiasmatic nucleus (SCN) input [ 23 ]. The endogenous circadian clocks can be reset by estrogen hormone signals in women, affecting the balance of the hypothalamus-pituitary-adrenal (HPA) axis, and making the HPA axis more responsive to stress than males [ 23 , 24 ]. The disruption of circadian rhythms within the HPA axis, and the sleep-arousal system differs between the genders, and is associated with dysfunction and disease [ 23 ]. Understanding gender differences in the circadian timing system can lead to improved treatment strategies. Second, the psychosocial contexts differ between the genders, and the impact of traditional gender roles and socio-cultural influences on stress and anxiety are greater in women than in men [ 25 ].
In this study, we found that people with low education levels or low income mostly suffer from irregular social rhythms, likely due to their weak competitiveness in the job market and inability to adapt to the new E-commerce models during the COVID-19 pandemic due to the isolation or even lockdown [ 26 ]. Moreover, they did not have sufficient judgment while facing various, overloaded and timely information on the COVID-19 pandemic, and easily felt stressed, depressive and anxious [ 27 ]. Interestingly, both the people with high incomes and those with poor health status were mostly suffered from disrupted rhythms. On one hand, the disrupted rhythm in people with high incomes may be explained as follows: during the COVID-19 pandemic, the consumption expenditures of high-income groups dropped sharply by as much as 17%, while the expenditures of low-income families dropped by only 4% [ 28 ]. At the same time, entertainment venues that provide services for high-income groups, such as cultural activity centers, clubs and gymnasiums, golf playgrounds, luxury party venues, were mostly closed. All these changes would impact severely on the social rhythms of these groups of people. On the other hand, the disrupted rhythm in people with poor health status can be explained that the regular activities, such as walking, jogging, swimming and fitness, which were significantly restricted since their poor health status with basic diseases, weak immune function, and low ability to fight pathogens or adapt to a rhythm change [ 29 ]. In addition, our finding that a high frequency of the irregular rhythm in the divorced group might be related to the disintegration of the family, loose family relations, economic problems, and the problem of raising children as a single parent. Especially, the isolation during the COVID-19 pandemic might have aggravated the fear for the unknown future, loneliness, helplessness, or irregular lifestyle [ 30 ].
The present study found that the people aged 26–40 years old and the elderly reported similarly high SGS-scale scores and were vulnerable to circadian disruption. The former is the main force of the society that promote the consumption, cultural creativity, and economic development [ 31 ]. However, when they must stay at home or fight against the COVID-19 pandemic due to the sudden outbreak of the disease, it is difficult for them to maintain a normal social rhythm. It is worth noting that the elderly reported the highest rate of rhythm disruption of eating, physical practice, and entertainment behaviors in the present study. Previously, Liu et al . [ 32 ] conducted a study on lifestyle and the health status of Chinese elderly, which included the elderly who lived in their own houses, with 20.6% of them being widows or widowers. They found that lack of physical practice, advanced age, and alcohol consumption were risk factors for their self-rate health status whereas a long sleep time was the protective factor [ 33 ]. Indeed, it has been reported that advanced age is characterized by a progressive decreasing amplitude and phase advance of circadian rhythmicity in overall biological functions, including blood pressure, hormonal circadian secretions, and immune function changes [ 33 ]. The elderly could not go out for dancing after dinner, exercise in the morning, or visit old friends or relatives due to recommended social isolation during the COVID-19 pandemic. Furthermore, they were inherently at high risk for COVID-19 infection and easily generated negative emotions [ 34 ]. These various factors were prone to form a vicious circle.
It is noteworthy to mention that nurses had the highest score of rhythm disruption in going to bed, get-up, and socializing among occupations, suggesting that they were more affected by the pandemic than other occupations such as businessmen, officers, teachers, police, farmers, doctors, and students. Night shift work has been thought to play a role through a disruption of the circadian rhythm, decreased synthesis of melatonin and sleep deprivation [ 35 ]. We assume that nurses undertake a working articulation in shifts to keep the continuity of healthcare throughout a whole day, and the COVID-19 pandemic results in heavy workload, insufficient rest, quarantine, and social isolation. Consequently, the already dysfunctional social rhythm of nurses is further exacerbated. McElroy et al . reported that night and long shifts caused fatigue and adversely affected the family and social life of hospital employees [ 36 ]. Rosa et al . indicated that shift work could cause psychological consequences, including anxiety, stress, and depression, which is an obstacle for social and family relationships and a risk factor for metabolic disorders, diabetes, cardiovascular disorders, and breast cancer [ 37 ].
The present study firstly reported the critical connection between social rhythm disruption and psychological impacts in a large-scale population under the stress of such a pandemic. Our study further found that the scores of social rhythms were positively correlated with SDS and SAS scores under the stress of the COVID-19 pandemic. The following factors are speculated to be involved in the connection. The first factor is the social zeitgeber theory. The notion that life events increase the likelihood of mood episodes via decreased zeitgeber scaffolding for a vulnerable circadian system has been known as the social zeitgeber hypothesis, and the “social rhythm” refers to the cross-day stability of the timing of behaviors that act as zeitgebers [ 7 ]. Social stimuli under the stress of COVID-19 may affect the circadian behavioral programs by regulating the phase and period of circadian clocks ( i . e . a zeitgeber action, either direct or by conditioning to photic zeitgebers) which leads to various adverse mental health outcomes [ 38 , 39 ]. In turn, depressive episodes that arise because of life events disturbing social zeitgebers subsequently affect biological rhythms, and thus derail social and biological rhythms [ 39 , 40 ]. This present study expanded the scope of application of these rhythm theories from patients with mood disorders to the population with different sociodemographic backgrounds. The second factor is the lockdown protocols that have been adopted to minimize COVID-19 pandemic transmission. The protocols may weaken the zeitgeber-setting mechanism, and impact the stability of the timing of daily behaviors, such as sleeping, getting up, socializing, eating, physical practice, and entertainment, which, in turn, lead to mental health disorders, such as anxiety and depression [ 7 ]. The third factor is the daily function of HPA activity, which is also regulated by the biological clock [ 41 ]. Anxiety disorders and major depressive disorder have been associated with increased and blunted HPA axis reactivity to social stress respectively, which engages the diverse biological pathways that bolster successful stress adaptation and promote stress resilience. Adaptation to the challenges of stress is compromised when these pathways are no longer functioning optimally, resulting in the disrupted rhythms on psychological impacts [ 42 ]. Of course, other mechanisms may also connect social rhythm dysregulation with psychological impacts, including circadian gene variations [ 43 ] and social jetlag [ 44 ].
Based on the findings of the present study, more effective measures that combine circadian rhythms with the diagnosis and treatment of psychosomatic diseases may improve psychological impacts and maximize the effectiveness of treatment. Therefore, the following measures are recommended for stabilizing daily routines, especially for those specific population who mostly suffered from disrupted rhythms: 1), self-management strategies for setting up and increasing regularity of daily activities and lifestyle habits during the COVID-19 pandemic [ 45 ]; 2), spending 1 hour/day outdoors, walking for more than one hour a day and basking in the sun, which may regulate the body clock by the light-dark cycle [ 7 ]; 3), scheduling social interactions at the same time of the day, and seeking the best candidates who may share thoughts and feelings in real time [ 7 ]; 4), eating meals at the same time every day, and adjusting the eating patterns and food categories. 5), avoiding daylight hour naps and the nocturnal blue spectrum light pollution from computers, mobile phones, electronic tablets and televisions, which suppresses the sleep-helping hormones [ 45 ]. Chronotherapy seems to be a promising approach to restore the proper circadian pattern in the elderly with an obvious sleep problem through adequate sleep hygiene, timed light exposure, and the use of a chronobiotic medication [ 12 ]; 6), taking drugs that adjust the biological clock for nurses with sleep disorders caused by nurses’ shift system when necessary [ 46 ]; and 7), encouraging online therapy program for normalizing the disrupted social rhythms in COVID-19 environment, with the support of health care professionals and persuasive systems design and user experience/interaction features [ 7 ].
The present study provides new ideas for the simultaneous intervention from the perspective of social rhythms and psychology, and would lay an epidemiological foundation for the interaction of social rhythms with the diagnosis and treatment of psychosomatic diseases and the maximization of therapeutic efficacy. However, there are a few limitations in the present study. First, this cross-sectional study cannot reveal causality, and the nature of voluntary participation may result in selection bias. Second, in this study, we did not track the 24-hour circadian rhythm. Indeed, it is necessary to track the circadian rhythms/sleep-wake cycle by a multi-lead-sleep monitor considering different psychological impacts, which will be implemented in our further research. Third, the survey response rate was not obtained since this online-survey was participated voluntarily and the lockdown and social isolation policy did not encourage a face-to-face survey, and it was difficult to get the survey response rate. Fourth, we did not consider obtaining the individual informed consent for teenagers themselves under 18 years old, in addition to their parents’ consent, which will also be implemented in our further similar research.
In conclusion, there exist rhythm disruptions among Chinese people with different sociodemographic backgrounds, which are closely associated with psychological impacts under the stress of the COVID-19 pandemic. Social rhythm disruption is independently associated with depression and anxiety. Interventions should be applied to the people vulnerable to the rhythm disruptions during the COVID-19 pandemic.
Supporting information
S1 Fig
(TIF)
S2 Fig
(TIF)
S1 Dataset
(XLSX)
S1 Table
Prevalence of depression with different grades in participants (N = 5854).
(DOCX)
S2 Table
Prevalence of anxiety with different grades in participants (N = 5854).
(DOCX)
S3 Table
Spearman’s correlation between subscale 1, subscale 2, Zung’s self-rating depression scale (SDS) and Zung’s self-rating anxiety scale (SAS).
(N = 5854).
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Community-acquired pneumonia (CAP) remains a major health concern worldwide with substantial morbidity and mortality, especially among geriatric populations [ 1 ]. Moreover, the severity of clinical manifestations of CAP varies significantly [ 2 , 3 ]. Owing to the diversity in clinical conditions and the lag in a clear definition of the causative pathogen, CAP remains a huge global challenge [ 4 ].
Lipids are the major components of alveolar surfactant [ 5 ]. In recent years, the rapid development of lipidomics has enabled us to gain new insights into the mechanism of disease development [ 6 , 7 ]. Lipids play a critical role in cellular energy storage, structure, and signaling [ 8 ]. They are not only considered as the components of membranes but also act as an indispensable factor in the immune response by organizing signaling complexes in cellular membranes [ 9 ]. Studies have confirmed that lipids act as important inflammatory mediators during the infection process [ 7 , 9 , 10 ]. Changes in lipid components of the serum or plasma can occur during acute lung injury, sepsis, bacteremia, and viral infections [ 11 – 14 ]. Bioactive lipids located further downstream in the biological system can better reflect the body’s redox balance, oxidative stress, signaling, apoptosis, and inflammation than biological and gene expression biomarkers in body fluids; therefore, lipids provide more relevant and richer features in CAP. However, none of the studies thus far have focused on changes in lipid abundance of the serum of patients with different degrees of CAP.
In the current study, untargeted lipidomic analyses using high-performance liquid chromatography-mass spectrometry (HPLC-MS) were performed to identify CAP-related lipidomic signatures. The relationship between the abundance of identified lipids and the severity of CAP was then investigated. The characterization of the lipidomic response in CAP patients can help to further understand the heterogeneity of the disease and provide new insights for diagnosis and treatment.
Materials and methods
Study population
This study was performed from January 2017 to October 2018 and included patients hospitalized at the Peking University People’s Hospital (PKUPH), Fujian Provincial Hospital, Sichuan University West China Hospital. This study was registered at ClinicalTrials.gov (NCT03093220) and was endorsed by the Institutional Review Board of the PKUPH. All CAP patients were recruited from the Respiratory Medicine Department or Intensive Care Unit. All participants provided written informed consent prior to the collection of any data. Authors had access to information that could identify individual participants during and after data collection. The study was approved by the Medical Ethics Committee of the PKUPH (No. 2016PHB202-01).
Inclusion criteria for the study were evidence of pulmonary infiltrate on the chest, a chest radiograph showing either a new patchy infiltrate, leaf or segment consolidation, ground-glass opacity, or interstitial change [ 4 , 15 ]. Additionally, the inclusion criteria included at least one of the following signs: (a) the presence of cough, purulent sputum production, and dyspnea; (b) fever (core body temperature >38.0°C); (c) auscultatory findings of abnormal breath sounds and rales; or (d) leukocytosis or leukopenia (peripheral white blood cell counts >10 × 10 9 /L or <4 × 10 9 /L), and if symptom onset began in communities. Severe CAP (SCAP) was defined as the presence of at least one major criterion or at least three minor criteria published by the American Thoracic Society in 2007 [ 2 ]. The major criteria were: invasive mechanical ventilation and septic shock with the need for vasopressors; the minor criteria included respiratory rate ≥30 breaths/minute, oxygenation index ≤ 250, multipolar infiltrates, confusion or disorientation, uremia (BUN level ≥ 20 mg/dL), leukopenia (WBC count <4000 cells/mm 3 ), thrombocytopenia (platelet count <100,000 cells/mm 3 ), hypothermia (core temperature <36°C), and hypotension requiring aggressive fluid resuscitation.
The exclusion criteria were age <18 years or the presence of any of the following: women who were pregnant or lactating, patients who had surgery within 3 months before onset, patients with evidence of nosocomial infections, and immunosuppressive conditions. Furthermore, patients with malignant tumor, chronic neurological diseases (e.g., Parkinson’s disease, multiple system atrophy), end-stage renal or liver disease, and active tuberculosis or pulmonary cystic fibrosis were excluded as well.
Demographic information was obtained using standard forms (including age, gender, smoking history, underlying diseases, complications, symptoms, signs, laboratory test results, and clinical treatment methods). Patient outcomes were evaluated at discharge and assessed through structured telephone interviews 30 days after enrolment ( S1 File ).
Blood sample collection and preparation
A total of 5 mL of fasting peripheral venous blood was drawn from adult CAP patients during the first 72 h of hospital admission, injected into the sterile pro-coagulation tubes, and allowed to stand at room temperature for 30 min. The sample was centrifuged and the serum was aliquoted into 2–3 Eppendorf tubes and stored at -80°C for processing.
At the time of the assay, the serum samples were thawed on ice and centrifuged at 14,000 g at 4°C for 20 min, and then 100 μL of the supernatant was transferred to a new Eppendorf tube. Next, 400 μL of chloroform/methanol (v/v = 2:1) was added to the supernatant liquid and the tube was vortexed for 1 min and left to stand for 20 min; this process was repeated thrice after which the tube was centrifuged at 12,000 rcf for 5 min. The lower chloroform layer was transferred using a glass syringe and dried under nitrogen. Finally, the lipid samples were stored at -80°C in the freezer as dry pellets. To monitor the repeatability and stability of the analytical system, a pooled quality control (QC) sample solution was prepared and extracted, by combining 10 μL aliquots of each sample using the process described above.
Lysophosphatidylethanolamine (LPE) (14:0), phosphatidylethanolamine (PE) (14:0_14:0), phosphatidylcholine (PC) (14:0_14:0), phosphatidylserine (PS) (14:0_14:0), and D5-triglyceride were used as internal standards in the analyses. The average coefficient of variation (CV) of the internal standards was 21% in all the samples and 10.7% in the QC samples; which is in the acceptable range for the analysis.
Untargeted UHPLC-MS/MS method for lipid analysis and lipid identification
Reverse-phase chromatography was selected for the LC separation using Cortecs C18 column (2.1×100 mm, Waters). Mobile phase A was prepared by dissolving 0.77 g of ammonium acetate in 400 mL of HPLC-grade water, followed by the addition of 600 m of HPLC-grade acetonitrile (PH ~7). Mobile phase B was prepared by mixing 100 mL of acetonitrile with 900 mL of isopropanol. The UHPLC system Ultimate 3000 (Thermo Fisher, CA) coupled with Q Exactive (Orbitrap) mass spectrometer (Thermo Fisher, CA) was used to acquire raw data. Details of parameters of the Q Exactive mass spectrometer and HPLC conditions are described in S2 File .
LipidSearch software v4.1.16 (Thermo, CA) was used to identify and quantify lipids containing more than 1,500,000 theoretical MS/MS fragment ions from 18 major lipid classes in the database, based on accurate precursor mass and characteristic fragments [ 16 ]. Lipid identification was based on a MS/MS match. The mass tolerance for the precursor and fragment was 8 ppm and 10 ppm, respectively. Only lipids with a chromatographic area >5E6 were regarded as a confident identification. A retention time shift of 0.25 min was allowed for quantitation. A retention time shift of 0.15 min was performed for “alignment.” M-score and chromatographic areas were used to reduce false positives. The criteria to eliminate the false positives were lipid species dependent. In addition, since ammonium acetate was used in the mobile phase, the adducts of + H, + NH 4 were used in the positive mode search, and—H, + CH 3 COO in the negative mode [ 17 ]. Finally, mass spectrum data composed of chemical formula, detected the ion adduct with accurate precursor mass, retention time, peak shape and distribution of integrated peaks, and MS/MS fragment match were obtained ( S3 File ).
Analysis and processing of UHPLC-MS/MS lipidomics data
It was necessary to pre-process the data to reduce the interference in the information and facilitate the mining of more meaningful biological information. The data pre-processing was performed using MetaboAnalyst 4.0 [ 18 ]. To remove the noise and improve the data quality, a lipid was included in the subsequent data analysis if it had a non-zero value for at least 80% of the samples of any group. Missing value estimations and filling in the data matrix were performed. Considering that most of the missing values were caused by low-abundance compounds (i.e., those below the detection threshold), we replaced all the missing values with small values (the half of the minimum positive values in the original data). We normalized the original abundance matrix of lipids to correct for the effects of factors such as individual differences or differences in sample collection or processing on absolute compound concentrations using constant sum, log transformed and auto-scaling (mean-centered and divided by the standard deviation of each variable) methods. The obtained normalized data were further analyzed.
Multidimensional statistical analysis
The normalized data were imported into SIMCA-P 14.0 (Umetrics, Sweden) for multivariate variable analyses (MVA). First, a principal component analysis (PCA) model was established to observe the overall distribution of samples in each group, explore possible factors affecting sample aggregation, and identify outliers. Second, according to the clinical grouping, an orthogonal partial least squares discriminant analysis (OPLS-DA) model was established for the samples to extract the different information between the groups, and the variable importance on projection (VIP) value was obtained. A VIP value greater than 1 indicated that the between-group differences were greater than the within-group differences. We used the R2X, R2Y, and Q2 parameter values to determine the quality of the OPLS-DA model. The quality of the MVA models was evaluated using a cross-validation analysis of variance (CV-ANOVA) and permutation test (500 iterations) [ 19 ].
Statistical analyses
Categorical variables are expressed as numbers (percentages) and analyzed using a chi-square test or Fisher’s exact test. The Kolmogorov–Smirnov test was used to evaluate the distribution of continuous variable data. Normally distributed continuous variables were expressed as means ± standard deviations (mean ± SD) analyzed using the Student’s t-test or analysis of variance with post-hoc Tukey HSD test. Continuous nonparametric data were presented as medians and interquartile ranges (25th and 75th percentiles) and analyzed using the Mann–Whitney U or Kruskal–Wallis H test, as appropriate. A heat-map with a Euclidean distance measure of relative intensity of metabolites (logarithmic scale) and a Pearson’s correlation heatmap were generated using MetaboAnalyst 4.0 (Wishart Research Group, University of Alberta, USA) [ 18 ]. The receiver operating characteristic (ROC) curve was analyzed using the multivariate logistic regression data, and the area under the curve (AUC), sensitivity, specificity and 95% confidence interval (95% CI) were calculated to evaluate the performance of lipids and clinical indicators. The Youden index (farthest to diagonal line) was used to determine an optimal cut-off point for test results (Classification: include cut-off value for positive classification; Test direction: larger test result indicates more positive test). The two-tailed Spearman’s rho test correlation coefficient (r) was calculated to evaluate the strength and direction of the linear relationship between the abundance of target lipids and clinical indicators. Furthermore, to adjust the multicollinearity inherent in lipidomics data, a stepwise approach was used to perform multiple linear regression (MLR) analysis. The Kaplan–Meier method was applied to establish a 30-day survival curve, and logarithmic rank tests were utilized to compare the survival rates. Cox proportional hazards regression analysis was used to analyze the effect of target lipids on 30-day survival.
Statistical tests were performed using the SPSS statistics version 19.0 (IBM, NY, USA) and MedCalc Software version 15.8 (MedCalc Software, Ostend, Belgium).
Results
Demographic and clinical characteristics of participants
The final study population consisted of 28 patients with CAP (15 non-severe CAP [NSCAP] and 13 severe CAP [SCAP]) and 20 age, sex, and underlying disease matched non-CAP controls (NC). As indicated in Table 1 , there were a range of etiologies of CAP, including bacteria, viruses, and fungi, which was typical for a heterogeneous CAP patient population. In particular, there were no significant differences in age, sex, smoking history, smoking index, and underlying disease among the three groups ( p > 0.05). However, in the laboratory tests, the inflammatory response-related indicators such as the percentages of neutrophil (NE%), lymphocyte (LY%), and monocyte (MO%) and the levels of white blood cell (WBC), neutrophil (NE), lymphocyte (LY) and monocyte (MO) were all significantly different in the SCAP group compared with that in the NSCAP group (all p < 0.05). The levels of serum C-reactive protein (CRP) and procalcitonin (PCT) were both higher in the SCAP group ( p < 0.05). In terms of physical examination, respiratory frequency of SCAP patients dramatically increased ( p < 0.05). CURB-65 (confusion, urea, respiratory rate, blood pressure, and age ≥65 years) and pneumonia severity index (PSI) were prominently higher in the SCAP group than in the NSCAP group ( p < 0.05). Comparing the detection of pathogens between the two groups, we found that the detection rate of bacteria in patients with SCAP was higher. In addition, hospitalization days and 30-day mortality were both substantially higher in patients with SCAP than in those with NSCAP ( p < 0.05) ( Table 1 ).
10.1371/journal.pone.0245770.t001
Table 1 Demographic and clinical characteristics of 48 study participants.
Characteristic
SCAP
NSCAP
NC
p value
(N = 13)
(N = 15)
(N = 20)
Male sex—no. (%)
8 (61.5)
6(40.00)
7(35.0)
0.304 a
Age—years
62.08±18.21
65.40±17.86
60(48.25–67)
0.141 c
Smoking history—no. (%)
4(30.80)
5(33.30)
3(15.00)
0.396 a
Smoking index
0(0–14.50)
0(0–10.00)
0(0–0)
0.533 a
Underlying diseases —no. (%)
COPD
0 (0)
1 (6.70)
0 (0)
0.325 a
Asthma
2 (15.40)
1 (6.70)
0 (0)
0.203 a
Bronchiectasis
0 (0)
1 (6.70)
0 (0)
0.325 a
Interstitial lung Disease
1 (7.70)
0 (0)
0 (0)
0.253 a
Hypertension
4 (30.80)
5 (33.30)
5 (25)
0.856 a
Cardiovascular disease
3 (23.10)
0 (0)
4 (20)
0.151 a
Cerebrovascular disease
1 (7.70)
0 (0)
2 (10)
0.466 a
Cardiac insufficiency
1 (7.70)
3 (13.30)
0 (0)
0.264 a
Autoimmune disease
1 (7.70)
0 (0)
0 (0)
0.253 a
Physical examination
T Max (°C)
38.55±1.45
37.75±1.22
NA
0.125 b
Respiratory frequency (times/min)
23.08±3.87
20(20–21)
NA
0.020 c
Systolic pressure (mmHg)
125.17±31.56
120(115–128)
NA
0.272 c
Diastolic blood pressure (mmHg)
74±14.09
72.60±10.03
NA
0.765 b
Mean arterial pressure (mmHg)
84.05±29.78
89.93±8.18
NA
0.469 b
Lung rales—no. (%)
10(76.92)
9(60.00)
NA
0.435 c
Disorder of consciousness rales—no. (%)
3(23.08)
0 (0%)
NA
0.087 a
Laboratory results
WBC (×10 9 /L)
13.68±8.64
6.38±2.48
NA
0.016 b
NE (×10 9 /L)
12.05±8.40
4.22±2.34
NA
0.002 b
LY (×10 9 /L)
0.95±0.50
1.46±0.56
NA
0.022 b
MO (×10 9 /L)
0.60±0.50
0.50±0.30
NA
0.547 b
NE percentages (%)
84.72±8.12
60(57.80–64.40)
NA
<0.001 c
LY percentages (%)
10.05±6.38
29.60(19.90–31.60)
NA
<0.001 c
MO percentages (%)
4.53±2.50
7.88±3.07
NA
0.005 b
NLR (%)
10.60 (4.30–22.90)
6.43±7.12
NA
0.001 c
PLR (%)
348.08±320.11
127.30(83.60–448.27)
NA
0.188 c
PLT (×10 9 /L)
228.38±110.76
218.13±66.91
NA
0.766 b
CRP (mg/L)
144.78±139.20
32.60±31.39
NA
0.009 b
PCT (μg/L)
0.54(0.32–7.48)
0.09±0.06
NA
0.002 c
PaO 2 (mmHg)
73.13±18.45
72.83±31.43
NA
0.978 b
FiO 2 (%)
44.36±17.82
21(21–25)
NA
0.003 c
PaO 2 /FiO 2
199.89±66.86
372.00±47.23
NA
<0.0001 b
PaCO 2 (mmHg)
39.82±10.32
36.06±4.51
NA
0.374 b
SaO 2 (%)
95.56±2.47
95.80±2.69
NA
0.875 b
HCO 3 (mmol/L)
25.81±6.20
24.26±4.67
NA
0.580 b
Total cholesterol (mmol/L)
3.49±1.36
4.31±0.67
NA
0.020 b
Triglyceride (mmol/L)
1.82(1.03–2.64)
0.84±0.20
NA
0.006 c
Detected pathogen —no. (%)
Bacteria
7(53.85)
2(13.33)
NA
0.042 a
Virus
2(15.38)
8(53.33)
NA
0.055 a
Fungus
4(30.77)
5(33.33)
NA
1.000 a
CURB-65
1(1–0)
2(1–2)
NA
0.006 c
PSI
93.85±28.70
67.40±30.51
NA
0.027 b
Hospitalization Days
18.62±9.49
10.07±3.43
NA
0.003 b
30-day mortality-no. (%)
4(30.77)
0(0)
0(0)
0.003 a
Note: Descriptive statistics. Variables are expressed as numbers (percentages). Normally distributed continuous variables are expressed as means ± standard deviations (mean ± SD) and continuous nonparametric data are presented as medians and interquartile ranges (25th and 75th percentiles).
a Chi-square test or Fisher’s exact test
b Student’s t-test or analysis of variance with post-hoc Tukey HSD test
c Mann-Whitney U or Kruskal-Wallis H test.
Abbreviations: BMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, NE neutrophil, LY lymphocyte, MO monocyte, NLR neutrophil/lymphocyte ratio, PLR platelet-lymphocyte ratio, PLT blood platelet, CRP C-reactive protein, PCT procalcitonin, PaO 2 partial pressure of oxygen, FiO 2 Fraction of inspiration O 2 , SaO 2 oxygen saturation, CURB-65 confusion, urea, respiratory rate, blood pressure, and age ≥65 years old, PSI pneumonia severity index, NA not applicable.
Global lipidomic profiles of human serum
Serum lipidomic profiles of the 48 participants were generated using an untargeted lipidomic profiling analysis using HPLC-MS/MS. Overall, the lipids were categorized into six classes: glycerophospholipids (GP), glycerolipids (GL), sphingolipids (SP), prenol lipids (PR), sterol lipids (ST), and fatty acid (FA). A total of 509 lipid species were detected in the electrospray ionization positive (ESI+) mode. The top three dominant classes comprised over 99% of the total lipid signal, including GP (44.78%), GL (38.19%), and SP (16.06%). The lipid profile of each of the three groups was different. Compared with the NC group, the abundance of zymosterol ester, dihexosylceramide (Hex2Cer) in the CAP group was significantly increased ( p <0.01, p <0.01, respectively,) while the abundances of lysophosphatidylcholine (LPC), diacylglycerol (DG), and PE showed a statistically significant decrease (all p <0.01). As the disease gradually worsened, the abundance of the eight lipid subclasses detected in SCAP patients decreased significantly, including LPC ( p <0.01), LPE ( p <0.01), hexosylceramide (HexCer) ( p <0.01), Hex2Cer ( p <0.05), trihexosylceramide (Hex3Cer) ( p <0.05), ganglioside GM3 (GM3) ( p <0.05), campesterol ester (CmE) ( p <0.05), and cholesteryl ester (ChE) ( p <0.0001) ( S1 Table ).
In the electrospray ionization negative (ESI-) mode, although 195 lipid species were detected, only the FAs were recognized more effectively. Only three classes of compounds were detected in the negative ion mode, namely, GP, SP, and FA. They accounted for 57.14%, 25.83%, and 17.03%, respectively. Subclasses such as PC (47.14%), sphingomyelin (SM, 22.35%), and free fatty acid (FFA, 17.03%) contributed 74.05% to the total lipid signal. Similarly, there are numerous differences in serum lipid profiles between NSCAP, SCAP, and NC. The abundance of Hex2Cer in the CAP group was significantly higher than in the NC group ( p <0.001), while the abundance of LPC, PI, and PE in this group showed the opposite trend ( p <0.001, p <0.05, p <0.05, respectively). At the same time, compared with NSCAP, the abundance of LPC, HexCer, Hex3Cer of SCAP group decreased ( p <0.01, p <0.05, p <0.05, respectively), while the abundance of PE increased ( p <0.05) ( S1 Table ).
Multivariate models established by untargeted lipidomics analysis
PCA is an effective approach for classifying data, detecting outliers, and validating the stability and reproducibility of an analytical method. All identified lipids were subjected to a PCA using MetaboAnalyst 4.0 to explore the major effects that potentially drive the differences in lipid profiles in CAP patients (NSCAP and SCAP) and NC ( Fig 1 ). The optimal PCA model contained eight components. The principal component (PC) 1(the first component) explained the largest variation in the variable (24.6%), followed by PC2 (12.2%), etc. Hence, the scores scatter plot composed of PC1 and PC2 shows the positional relationship between variables. The value of R2 (cum) and Q2 (cum) (0.674 and 0.449), respectively, represented the fit and predicted power of the model. All QC samples were tightly clustered, indicating that this instrument has good repeatability and stability. As shown in Fig 1A , no obvious outlier was detected in all serum samples. Importantly, the obvious intra-group clustering and inter-group separation between the three groups suggested that whether CAP or SCAP, the patient’s serum lipid profile has changed significantly.
10.1371/journal.pone.0245770.g001
Fig 1
PCA scores plot of lipidomic profiles in CAP group (including NSCAP and SCAP) and NC.
Blue, severe CAP (SCAP); Green, non-severe CAP (NSCAP); Red, non-CAP control (NC), Navy blue, quality control (QC) samples; Black, CAP. No sample was placed outside the ellipse that describes the 95% CI of Hotelling’s T-squared distribution. The three groups of lipid profiles can be clearly distinguished.
OPLS-DA analysis identified the biggest variation in lipid profiling using a few orthogonal latent variables. To further eliminate the interference factors of the disease and maximize the extraction of the information on the differences in lipid mass spectra between the different groups, we used a supervised clustering method to verify the OPLS-DA model of the serum samples of the patients. To prevent overfitting, we performed a permutation test (500 iterations) on those models. The OPLS-DA score plot showed obvious a discriminatory trend between both the CAP group versus the NC, NSCAP versus NC, SCAP versus NC, and NSCAP versus SCAP groups (S1 Fig in S4 File ). The CV-ANOVA p -values for all the models were less than 0.0001, indicating that all the differences between the groups were significant. The Q2Y of all models was higher than 0.9, which revealed the model had an excellent interpretation ability and superior predictive power (Q2>0.5) ( Table 2 ). After the 500-it iteration of the permutation test, the R2 and Q2 values were less than the original model ( Table 2 ), and the Q2 regression line was less than 0 in the Y-axis intercept (S2 Fig in S4 File ), which proved that the model was robust and there was no overfitting. Notably, in the OPLS-DA model, the predictability of separating the SCAP group from the NC group (Q2 = 0.848) was better than separating the NSCAP group from the NC group (Q2 = 0.703). That is, compared with the NSCAP group, more changes were observed in the SCAP group with a better isolation from the NC group.
10.1371/journal.pone.0245770.t002
Table 2 Evaluation parameters of our OPLS-DA model.
Group
Evaluation parameters of OPLS-DA original model
Permutation test
PC a
R2X
R2Y
Q2(cum)
CV-ANOVA ( p )
R2
Q2
CAP vs NC
1+4+0
0.476
0.979
0.685
0.000001
0.913
-0.566
NSCAP vs NC
1+2+0
0.329
0.947
0.703
0.00000257
0.845
-0.489
SCAP vs NC
1+2+0
0.448
0.964
0.848
0.00000000172
0.790
-0.509
NSCAP vs SCAP
1+2+0
0.366
0.964
0.536
0.00757053
0.847
-0.386
a : PC The number of principal components in this model, which represents the number of predicted components plus the number of orthogonal components.
Abbreviations: CAP, community-acquired pneumonia; NC, non-CAP control; NSCAP, non-severe CAP; SCAP, severe CAP.
Differential abundance of lipids associated with CAP disease and acute exacerbations
To identify lipids with significant abundance changes between the CAP groups versus the NC and NSCAP versus SCAP groups, we normalized the relative peak intensity data of all lipids detected and performed T-test for the univariate analysis. A total of 295 lipids were statistically different in abundance (Benjamini-Hochberg adjusted p -value (FDR) < 0.05) between the CAP and NC. The VIP score was used to quantify the contribution of each lipid level to the overall separation between the two groups in the OPLS-DA model. There were 226 lipids with VIP values greater than 1 between the CAP and NC. Compared with NSCAP and SCAP simultaneously, 297 lipids had statistical differences in abundance (FDR<0.05), and VIP of 246 lipids exceeded 1.
Interestingly, we found that not only did the relative abundance of many lipids were different between the NC and CAP groups, but the levels of these lipids also changed significantly as the condition of CAP deteriorated ( Fig 2A ). To screen out the highest potential lipid that could distinguish CAP from NC and assess the severity at the early stage, we set the comparisons between the groups (CAP vs NC and NSCAP vs SCAP) to meet VIP> 1 and FDR <0.05 as the screening criteria. A total of 83 lipids met the screening criteria both in the comparison of the CAP versus HC and NSCAP versus SCAP groups. Remarkably, the relative levels of most of the differential lipids were continuously up-regulated or down-regulated in the CAP and SCAP group ( S2 Table ).
10.1371/journal.pone.0245770.g002
Fig 2
The relative abundance of different lipids changed in the three groups of samples.
Red, green, and blue represent NC, NSCAP, and SCAP, respectively. (a) Hierarchical cluster heatmap of the relative abundance of 83 lipids in CAP (including NSCAP and SCAP) compared to NC. Distance measure: Pearson; Clustering algorithm: ward. Row represents lipids and column represents serum samples. Red, green, and blue represent NC, NSCAP, and SCAP, respectively. Light blue indicates lower relative abundance, while greater brown indicates higher intensity of lipids. (b-e) Changes in the relative abundance of PC (16:0_18:1), PC (18:2_20:4), PC (36:4) and PC (38:6) in serum between CAP group versus NC and NSCAP versus SCAP group. * p <0.05, ** p <0.01, *** p <0.001, **** p <0.0001.
Receiver Operating Characteristic (ROC) curve analysis
ROC analysis was performed to investigate whether the abundance of screened lipids could be efficiently utilized for building a sensitive biosignature of severe status in CAP. We further selected lipids with an AUC greater than 0.85 in both comparisons (CAP vs NC; SCAP vs CAP) as target lipids ( Fig 3 ). There were four lipids in total, PC (16:0_18:1), PC (18:2_20:4), PC (36:4), and PC (38:6) ( Table 3 ), which were screened out. Furthermore, the relative increase in the AUC (95% CI) for all the four lipids were superior to PSI (0.749, 0.550–0.892) and CURB-65 (0.772, 0.575–0.908) for discriminating the SCAP group from the CAP group ( S3 Table ). Simultaneously, the relative abundance of PC (18:2_20: 4), PC (36:4), and PC (38:6) was significantly lower in the CAP group than in the NC group, and their abundance continued to decrease as the disease worsened. However, the relative abundance of PC (16:0_18:1) showed the opposite trend ( Fig 2B–2E ). The multiple logistic regression analysis revealed that a combination of these four lipid had an AUC value of 0.952 (0.848–0.993), with 78.57% sensitivity and 100% specificity when distinguishing CAP patients from NC, indicating that they can serve as a lipid panel of potential biomarkers for diagnosing CAP ( S3 Table ). However, when distinguishing between NSCAP and SCAP, the AUC (0.959 [0.808–0.998]) of the combined signature does not show better performance than a single lipid PC (38:6) (AUC = 0.959) ( p > 0.05). Subsequently, the lipids with relatively lower AUC value (PC [16:0_18:1] and PC [36:4]) were combined as an indicator. It was found that the AUC of combined indicator (0.938 [0.779–0.994]) was superior to that of a single lipid (all p <0.0001), with sensitivity of 84.62% and specificity of 93.33% ( S3 Table ).
10.1371/journal.pone.0245770.g003
Fig 3
Flow chart of target differential lipid screening process.
CAP, Community-acquired pneumonia; NC, Non-CAP controls; NSCAP, Non-severe CAP; SCAP, Severe CAP; VIP, variable importance on projection; ROC, Receiver operating characteristic; AUC, area under the curve.
10.1371/journal.pone.0245770.t003
Table 3 Comparison results of four different lipids meeting the screening criteria.
Accepted Description
Ion Formula
CAP vs NC
SCAP vs NSCAP
Trend
VIP
p -value
AUC
Trend
VIP
p -value
AUC
PC(16:0_18:1)
C44 H85 O10 N1 P1
↑
1.269
<0.001
0.879
↑
1.179
<0.010
0.933
PC(18:2_20:4)
C48 H83 O10 N1 P1
↓
2.175
<0.00001
0.927
↓
1.747
<0.010
0.954
PC(36:4)
C44 H81 O8 N1 P1
↓
1.975
<0.00001
0.888
↓
1.389
<0.010
0.877
PC(38:6)
C46 H81 O8 N1 P1
↓
2.182
<0.00001
0.914
↓
1.863
<0.010
0.959
Abbreviations: CAP, community-acquired pneumonia; SCAP, severe CAP; NSCAP, non-severe CAP; NC, non-CAP control; PC, phosphatidylcholine; VIP, variable importance on projection value; ↑ Up-regulation ↓ Down-regulation.
Correlation between the levels of four target lipids and clinical indicators
Since these clinical indicators showed abnormal distributions, Spearman’s rank correlation test was applied to further explore whether the relative abundance of the four target lipids were correlated with the clinical parameters: WBC, NE, LY, LY (%), MO (%), and NE (%) in serum, CRP, PCT, FiO 2 , PaO 2 /FiO 2 , CURB-65, and PSI. Owing to the lack of laboratory test data in the NC, we calculated the correlation between clinical indicators and the relative level of four lipids in the NSCAP and SCAP groups (S3 Fig in S4 File ; S4 and S5 Tables).
According to the results of the correlation analysis, the relative abundances of the four target lipids (PC [16:0_18:1], PC [18:2_20:4], PC [36:4], and PC [38:6]) were related to indices of infection. Considering the strong co-linearity between those lipids, an MLR analysis was conducted to evaluate the biochemical indices, which were independently correlated to lipid abundance. Finally, we identified that decreasing relative levels of PC (18:2_20:4), PC (38:6), and PC (36:4) were negatively related to FiO 2 after p value adjustment. In addition, change in the relative abundance of PC (18:2_20:4) was inversely correlated with PCT, while the level of PC (16:0_18:1) had a positive linear relationship with PCT ( Fig 4 ).
10.1371/journal.pone.0245770.g004
Fig 4
The relative levels of PC (16:0_18:1), PC (18:2_20:4), PC (38:6), and PC (36:4) show significantly correlations with clinical indicators.
(a-c) The relative levels of PC (18:2_20:4) PC (38:6) and PC (36:4) were inversely correlated to FiO 2 of patients with CAP. (d) The abundance of PC (16:0_18:1) had a positive linear relationship with PCT (e) The abundance of PC (18:2_20:4) was negatively correlated with PCT of CAP patients. Solid black line, the fitted regression line. Area within the dotted line lines, the 95% confidence intervals.
The relative abundance of target lipids are associated with the prognosis of patients with SCAP
The AUCs for PC (16:0_18:1) and PC (18:2_20:4) were 0.885 and 0.954 for discriminating non-survivors from patients with CAP, respectively ( p < 0.001 for both comparisons) ( Table 4 ). The optimal normalized relative abundance threshold for predicting death was 0.392 of PC (16:0_18:1), with a sensitivity of 100% and specificity of 79.17%. While patients with PC (18:2_20:4) abundance <0.097 exhibited a noticeable increase in the risk of death, this threshold yielded sensitivity and specificity of 92.31% and 86.67%, respectively, for prediction of 30-day mortality ( Table 4 ).
10.1371/journal.pone.0245770.t004
Table 4 Area under the curve (AUC) and thresholds for discriminating non-survivors from patients with CAP.
Threshold
Sensitivity (%)
Specificity (%)
AUC
p value
95% CI
Lower limit
Higher limit
PC (16:0_18:1)
> 0.392
100
79.17
0.885
< 0.0001
0.708
0.974
PC (18:2_20:4)
≤ 0.353
100
45.83
0.708
0.1312
0.507
0.863
PC (36:4)
≤ -0.155
100
62.5
0.771
0.003
0.574
0.907
PC (38:6)
≤ 0.147
100
58.33
0.760
0.0099
0.563
0.9
Four lipids combined
--
100
79.17
0.875
<0.0001
0.695
0.969
PC (18:2_20:4)+ PC (36:4)+ PC (38:6)
--
100
66.67
0.792
0.0007
0.597
0.921
CURB-65
> 1
75
75
0.802
0.0057
0.609
0.927
PSI
> 114
50
91.67
0.625
0.5246
0.423
0.799
Abbreviations: CAP, community-acquired pneumonia; CURB-65, confusion, urea, respiratory rate, blood pressure, and age ≥65 years old score; PSI, Pneumonia Severity Index score.
The optimal threshold values of the ROC analysis were used as the cut-off to regroup to higher abundance group (normalized relative abundance was higher than cut-off value) and lower abundance group (normalized relative abundance was less than cut-off value) of CAP patients. Further, we compared the length of hospital stay of patients with higher and lower abundance groups. The results showed that patients with an elevated level of PC (16:0_18:1) (normalized relative abundance >-0.580) had significantly longer duration of hospital stays ( p <0.05). Moreover, with the gradual decrease of serum PC (18:2_20:4), PC (36:4), and PC (38:6) levels, the length of hospitalization days of patients also increased significantly (all p <0.05, Fig 5A–5D ). The Kaplan–Meier curve showed that there were statistically significant differences in mortality between the higher abundance and lower abundance groups of PC (18:2_20:4), PC (16:0_18:1), PC (36:4), and PC (38:6) ( p = 0.0035, 0.0338, 0.0340 and 0.0219, respectively) ( Fig 5E–5H ). The dynamic changes of the relative abundances of specific lipids in serum were closely related to the disease progression and prognosis.
10.1371/journal.pone.0245770.g005
Fig 5
Comparison of the days of hospitalization between the higher abundance group and lower abundance group of PC (16:0_18:1), PC (18:2_20:4), PC (36:4), PC (38:6) and Kaplan–Meier analysis of 30-day mortality in patients with SCAP.
(a-d) There were statistical differences in the days of hospitalization between the high abundance and low abundance groups of PC (16:0_18:1), PC (18:2_20:4), PC (36:4) and PC (38:6). Red dots, high abundance group; Blue squares, low abundance group. * p <0.05, ** p <0.01 (e-h) There were statistically significant differences in 30-day mortality between the high abundance group and the low abundance group of PC (16: 0_18: 1) (e), PC (18:2_20:4) (f), PC (36: 4) (g) and PC (38:6) (h).
Discussion
In this study, we used UHPLC-MS/MS technology to conduct exploratory studies and describe lipid mass spectrometry characteristics in 48 human serum samples, including the 28 CAP and 20 NC. We found that the lipid profiles of the NC, NSCAP, and SCAP groups were significantly different in both positive and negative ion modes. The untargeted lipidomic analysis showed that CAP patients could be clearly discriminated from NC patients, as well as the NSCAP group from the SCAP group, which further suggested that CAP and disease exacerbation caused significant fluctuations in lipidomic biochemical homeostasis. We further selected four lipids with AUCs greater than 0.85 in both comparisons of the CAP versus NC and NSCAP versus SCAP groups, as target lipids. The molecular levels of PC (18:2_20:4), PC (36:4) and PC (38:6) were significantly negatively correlated with FiO 2 after p value adjustment. The abundance of PC (18:2_20:4) was inversely correlated with the PCT level, while the level of PC (16:0_18:1) had a positive linear relationship with PCT.
Lipids account for 90% of the surfactants in the lungs, and disorders of surfactants during pneumonia may cause changes in lipid metabolism [ 20 ]. Korneev et al. reported that large amounts of lipid metabolites were produced by bacteria, and lipids were important for the structure and function of bacteria [ 21 ]. At the same time, studies have shown that lipids are important inflammatory mediators during infection, and changes in lipid metabolites have been observed in cases such as sepsis [ 22 ], bacteremia [ 23 ], and viral infections [ 24 , 25 ]. The differences in the lipid profile of the NC, NSCAP, and SCAP groups found in this study may be related to these reasons. Arshad et al. observed a striking decrease in the phospholipid concentrations in acute CAP, which largely normalized with clinical recovery. The greatest changes were seen in PC, followed by LPC, SM, and Cer [ 26 ]. Thus, the level of phospholipids can serve as highly accurate biomarkers for the diagnosis of CAP. This finding is consistent with our research. Moreover, decreased blood phospholipid concentrations have been documented in other invasive bacterial infections [ 27 ]. Studies by Arshad et al. also confirmed that at least part of the observed reduction in plasma concentration might be pre-programmed at the level of infected cells due to the phospholipase activity [ 26 ]. Additional studies are still needed to unravel the causes and consequences of reduced phospholipid abundance in CAP.
We further strictly limited the criteria used to select the differential abundance of lipids in this study, based on the combination of VIP > 1 and FDR < 0.05, which was consistent with the methods used in some previous metabolomics studies [ 28 , 29 ]. In combination with the ROC analysis, four lipid small molecules were finally selected as target lipids and combined with clinical indicators for further correlation analyses. Studies have shown that changes in the abundance of PC (16:0_18:1), PC (18:2_20:4), PC (36:4), and PC (38:6) are closely related to pathogen infection, membrane injury, and liver injury [ 30 – 32 ]. Fluctuations in the abundance of these lipids lead to the change of glycerophospholipid metabolism. The disorder of the glycerophospholipid metabolic network is closely related to the occurrence and development of many diseases, such as coronary heart disease [ 33 ], atherosclerosis [ 34 ], brain injury [ 35 ], pain, and inflammation [ 36 ]. Studying the dynamic changes of the molecular composition of glycerophospholipids helps explain the molecular mechanism of disease pathogenesis and progression.
Hence, PCT is currently considered an infection-related biomarker and can be used as a specific indicator of bacterial infection [ 37 ]. Remarkably, the degree of PCT elevation is closely related to bacterial load and infection severity [ 38 ]. In addition, studies have shown that the prognostic utility was substantially improved when combined phospholipid and PCT criteria were applied to 28-day mortality outcome predictions in ICU patients with severe sepsis or septic shock [ 39 ]. The abundances of PC (16:0_18:1) and PC (18:2_20:4) had significant linear correlation with PCT level in our study. The results suggested that plasma phospholipids not only have strong CAP biomarker potential but fluctuations in its levels may also be closely related to bacterial infections. Evaluation of bacterial diseases, viruses, or other pathogen categories may help decide treatments to stop or gradually upgrade the use of antibiotics. Langley et al.’s [ 40 ] research showed that profiles of specific metabolites measured on days 1 and 7 differed markedly between survivors and non-survivors of severe septic shock, and patients who did not survive within 90 days showed a marked decrease in PC species. Consistently, in this study, we found that as the relative abundance of PC (18:2_20:4), PC (36:4), and PC (38:6) decreased, and the hospitalization days significantly extended. Particularly, as the levels of PC (16:0_18:1), PC (36:4), and PC (38:6) decreased, 30-day mortality rates increased significantly.
This study had several limitations. Even though our results agree well with published evidence, this study is limited by the lack of an external validation cohort and the relatively small sample sizes. Further research is, therefore, needed before the identified biomarkers can be advanced to clinical application. In addition, only the serum lipid relative abundance was detected at the time of admission; dynamic and follow-up changes (in response to treatment) were not investigated. The effects of changes in lipid expression during the pathogenesis of CAP should be further investigated. Moreover, multi-omics data are urgently needed to be integrated into the CAP study to further support the development of precision medicine.
Conclusions
In conclusion, we demonstrated that lipidomic approaches based on HPLC-MS/MS could be used successfully to reveal changes in lipid abundance in CAP and establish a metabolite signature related to disease severity. It helps to adjust the treatment plan for the specific disordered lipid characteristics and related disease states exhibited by the patients.
Supporting information
S1 File
Questionnaire for the 30-day outcome of patients with community-acquired pneumonia (original language and English).
(DOCX)
S2 File
Supplemental methods for Ultra High-Performance Liquid Chromatography-Mass Spectrometry (UHPLC-MS/MS) analysis.
(DOCX)
S3 File
MS/MS spectra for assigning the lipid molecules.
(DOCX)
S4 File
S1 Fig. Orthogonal partial least squares discriminant analysis (OPLS-DA) of serum samples in the CAP group (including NSCAP and SCAP) and NC. (a) OPLS-DA score plot discriminates all CAP patients versus NC (b) OPLS-DA score plots of NSCAP versus NC group. (c) OPLS-DA score plots of SCAP versus NC. (e) OPLS-DA score plots of NSCAP versus SCAP. The model of OPLS-DA reflect good separation trends among SCAP, NSCAP and NC. Blue, severe CAP (SCAP); Green, non-severe CAP (NSCAP); Red, non-CAP control (NC), Black, CAP. S2 Fig. Permutation tests of all the OPLS-DA models. All OPLS-DA models have been verified using permutation tests. Permutation verification was established after 500 iterations. The Permutations Plot helps to assess the risk that the current OPLS-DA model is spurious. The plot shows, for a selected Y-variable, on the vertical axis the values of R2 and Q2 for the original model (far to the right) and of the Y-permuted models further to the left. The horizontal axis shows the correlation between the permuted Y-vectors and the original Y-vector for the selected Y. The plot above strongly indicates that the original model is valid. The criteria for validity are: all blue Q2-values to the left are lower than the original points to the right or the blue regression line of the Q2-points intersects the vertical axis (on the left) at, or below zero. The permutation test of the OPLS-DA model constructed by CAP versus HC groups (a), NSCAP versus HC (b), SCAP versus HC (c) and SCAP versus SCAP (d). S3 Fig. Correlation between the target lipids and the clinical indicators. The color transition from dark blue to red indicates the correlation from low to high.
(DOCX)
S5 File
Structural confirmations of the four target lipids.
S1 Fig. Some representative MS/MS spectra to assign the lipid molecules. S2 Fig. Representative chemical structures and fragmentation of PC in positive mode and negative mode. Red lines displays collision induced fragments generated in negative mode and blue lines are fragments in positive mode. S3 Fig. Determination of the chemical structure of PC (16:0_18:1) in the serum extract using tandem mass spectrometry. S4 Fig. Determination of chemical structures of PC (18:2_20:4) in serum extract using tandem mass spectrometry. S5 Fig. Determination of chemical structures of PC (36:4) in serum extract using tandem mass spectrometry. S6 Fig. Determination of chemical structures of PC (38:6) in serum extract using tandem mass spectrometry.
(DOCX)
S1 Table
Global serum lipidomic profiles of 48 patients generated through untargeted lipidomic profiling analysis using HPLC-MS/MS.
(XLSX)
S2 Table
The 83 lipids that discriminate the NC, NSCAP and SCAP groups.
(XLSX)
S3 Table
Areas under the curve (AUCs) and thresholds for all ROC analysis.
(DOCX)
S4 Table
Correlation analyses between the abundance of four target lipids for CAP and the clinical parameters—Spearman’s rank correlation coefficient.
(XLSX)
S5 Table
Correlation analysis between the abundance of four target lipids for CAP and the clinical parameters— p value.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Traumatic brain injury (TBI) remains a major worldwide health and socioeconomic problem. It is the most common cause of death and disability in people between 15 and 30 years of age [1] . The weighted average mortality for severe TBI is 39%, and for unfavorable outcome on the Glasgow Outcome Scale (GOS) is 60% [2] .
TBI has a dynamic pathophysiology that evolves in time, consisting of primary injury, followed by a combination of systemic disorders (hypoxia, hypotension, and hypercarbia) and local events, which together lead to secondary injury [2] . As brain is the functional regulator for metabolic activities, a complex milieu of metabolic alterations may occur in TBI, consisting of hormonal changes, aberrant cellular metabolism, and inflammatory cascade [3] . The abnormal metabolic processes, mainly including hypermetabolism, hypercatabolism, and glucose intolerance, have been recognized as incredibly essential elements of secondary injuries [3] – [5] . Not only can they complicate the initial period of hospitalization and stabilization, but also they may negatively impact rehabilitative treatments [3] . Nutritional support, in addition to providing daily calories, has been appreciated as an important adjunctive therapy for metabolic disorders following TBI [6] .
Nutritional support constitutes an important issue in intensive care for critically ill patients. However, it is generally neglected and underestimated in the subgroup of TBI population. In the recent most important trials in nutrition, Casaer et al. only included 0.6% of patients with neurological diseases [7] . Heidegger et al. incorporated 15% of neurological patients, and only included those with functioning gastrointestinal tract [8] . Interestingly, disagreement on the role of early parenteral nutrition even existed between the two trials. Currently, nutritional support for TBI patients, especially the appropriate timing, route, and formula of feeding, has not been well illustrated yet. In the earlier Cochrane Review, a trend towards better outcome with early nutritional support for TBI patients was shown, but without any statistically significant result [6] . In addition to not including updated studies, it had several flaws of its own. Although the Brain Trauma Foundation has recommended achieving full caloric replacement by day 7 following TBI, no agreement has been reached in the optimal timing or route of feeding [9] . In fact, nutritional support is frequently underestimated in the clinical management of severe TBI patients. Given insufficient previous evidence, as well as the introduction of recent randomized controlled trials (RCTs), we perform this meta-analysis and systematic review, aiming to compare the effects of different timings and routes of feeding, and to explore the effect of immune-enhancing formula on outcomes in TBI patients.
Methods
Search Strategy
The overview of this meta-analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) statement ( checklist S1 ) [10] . A computerized bibliographic search for all relevant articles from 1980 to October 2012 was performed by Ovid Medline, EMBASE, and the Cochran database. We used the following search core terms: “traumatic brain injury”, “craniocerebral injury”, “head injury”, “head trauma”, “enteral nutrition”, “enteral tube feeding”, “parenteral nutrition”, “gastrointestinal intubation”, “nutritional support”, “prospective cohort”, and “randomized controlled trial”. The language was limited to English. We also manually searched the references of selective papers to identify additional potentially eligible studies.
Selection Criteria
Studies were selected into the meta-analysis if they: i.) were RCTs or non-randomized prospective studies (NPSs); ii.) investigated TBI patients; iii.) compared the effect of different feeding routes (enteral nutrition [EN] vs. parenteral nutrition [PN], or nasogastric enteral feeding vs. non-nasogastric enteral feeding), different feeding timings (early or delayed), or different immunonutritional elements (such as probiotics, arginine, glutamine, nucleotides, and ω-3 fatty acids), on the outcome variables; iv.) reported the number of outcome events in different interventions.
Data Extraction
Two assessors (XW and YD) independently reviewed the full manuscripts of eligible studies. Data were extracted in standardized data-collection forms. Extracted data included first author’s name; year of publication; sample size; patients’ characteristics (mean age, gender); starting time of feeding; treatment arms; outcome variables and the score of quality assessment. Any discrepancy was resolved by discussion or a third author (CGH). Selected RCTs were critically appraised using the Jadad scale, which scores studies’ description of randomization (2 points), blinding (2 points) and attrition information (1 point) [11] . The Newcastle–Ottawa scale (NOS) was used to evaluate the methodological quality of prospective cohort studies, as recommended by the Cochrane Non-Randomized Studies Methods Working Group [12] . The quality of a study was judged on the selection of the study groups, the comparability of the groups, and the ascertainment of the outcome of interest.
Studied Outcomes
Primary outcomes of clinical importance included mortality and functional outcome on GOS score. Secondary outcomes include the length of stay (LOS) in hospital or ICU, and major complications. Infectious complications and feeding-related complications were assessed, respectively. We defined infectious complications as pneumonia (ventilator or non-ventilator-associated pneumonia and other lower respiratory tract infections), central nervous system (CNS) infection, bloodstream infection (laboratory-confirmed-bloodstream infections and clinical sepsis), or urinary tract infection. Feeding-related complications include feeding intolerance, aspiration, diarrhea, constipation and vomiting, and abdominal distention.
Statistical Analysis
Data relating to outcomes were combined from pertinent studies. We used risk ratios (RR) and the associated 95% confidence intervals (CIs) to pool binary outcomes, including mortality, poor outcome and complications. Mean differences (MDs) with 95% CIs were used for continuous outcomes, which included the LOS in hospital or ICU, and the length of ventilator days. Review Manager 5.1.7 (Cochrane Collaboration, 2012) was used to process the meta-analysis. The Mantel-Haenszel method was used to test the significance of treatment effect, and the random-effects model was used to estimate the overall RRs. Heterogeneity of treatment effects between studies was statistically explored by the I 2 statistic. I 2 statistic of 0%–40% indicates unimportant heterogeneity, 30%–60% indicates moderate heterogeneity, 50%–90% indicates substantial heterogeneity, and 75%–100% indicates considerable heterogeneity [13] . Besides, we performed subgroup analyses based on the following factors which may contribute to the heterogeneity: study design, sample size, publication year, staring time of early nutrition, and different route of feeding. The sensitivity analyses were carried out by excluding studies one by one, or by employing a fixed effect model. All reported P values were two-sides, and P values less than 0.05 were deemed as statistically significant. The publication bias was examined visually by inspecting the funnel plots on Review Manager 5.1.7, and statistically by using the Egger’s regression model, calculated by Stata 12.0 (Stata Corporation, College Station, TX, USA).
Results
709 articles were found in total from the initial search, of which 34 eligible articles were selected after screening of titles and abstracts. Further, 10 studies were excluded, with 5 of non-English language, 1 comparing combined EN and PN with PN [14] , 1 comparing two different fat emulsions [15] , 1 comparing essential amino acid with placebo [16] , 1 comparing intermittent EN with continuous EN [17] , and 1 compared different infusion speed of EN [18] . In the remaining 24 studies included in qualitative synthesis, 8 articles lacked sufficient data relating to our outcomes [19] – [25] . Thus 16 studies were pooled into the meta-analysis, including 13 randomized controlled trials [26] – [38] , and 3 NPSs [39] – [41] . The search flow diagram was shown in Figure 1 and protocol S1 . The characteristics of these studies were shown in Table 1 . Of the 13 RCTs, the mean Jadad score was 2, ranging from 0 to 5. All NPSs have a high NOS score of 8.
10.1371/journal.pone.0058838.g001
Figure 1
The flow diagram shows the selection of studies for the meta-analysis.
10.1371/journal.pone.0058838.t001
Table 1
Characteristics of included studies of meta-analysis.
Author (year)
Sample size
Mean age, y
Male sex, %
Treatment arms (No.)
Timing from admission/injury to starting of nutrition
Initial GCS
Primary outcomes
Nutritional outcomes
Quality score #
Rapp (1983)
38
40
NA
Early PN (20) vs Delayed EN (18)
Early: within 48 hours after admission; delayed: till the bowel sounds were present and the gastric residual volume was less than 100 ml/hour
Mean: 7.5
Mortality; GOS
Nitrogen balance; nitrogen intake; caloric intake; nitrogen excretion; serum albumin level; serum transferrin values; lymphocyte count; serum glucose levels
R1B0A1 = 2
Hadley (1986)
45
28
88.9
Early PN (24) vs Early EN (21)
Within 48 hours of admission
≤10 (mean: 5.8)
Infection rate; GCS; mortality
Nitrogen loss; nitrogen intake; nitrogen balance; nitrogen excretion; serum albumin levels; weight loss
R0B0A1 = 1
Young (1987)
51
33
82.4
Early PN (23) vs Delayed EN (28)
Early: within 48 hours postinjury; delayed: till the termination of low wall suction
4–10
GOS; mortality; complication rate
Nitrogen balance; caloric balance; nitrogen intake; lymphocyte counts; albumin levels
R1B0A1 = 2
Grahm (1989)
32
27
90.6
Early jejunal (17) vs Delayed gastric (15)
Nasojejunal: within 36 hours of admission; gastric: after day 3 when gastric function returned
≤10
Infection rate; length of ICU stay
Nitrogen balance; caloric intake; nitrogen intake
R0B0A0 = 0
Borzotta (1994)
49
27
81.6
Early PN (21) vs Early EN (28)
Within 72 hours of injury
3–8
Complication rate
Measured resting energy expenditure; nitrogen balance; nutrition excretion
R1BOA0 = 1
Minard (2000)
27
33
70.4
Early EN (12) vs Delayed EN (15)
Early: within 60 hours of injury; delayed: when the gastroparesis resolved
3–11
Mortality; LOS in ICU and hospital; ventilator days; complication rate
Caloric intake
R1B0A1 = 2
Falcan (2004)
20
27
95
Early EN (10) vs Early EN with glutamine and probiotics (10)
Within 48 hours of admission
Mean: 7
Infection rate; LOS in ICU; period of mechanical ventilation
Nitrogen balance; caloric intake; protein intake
R1B0A1 = 2
Kostadima (2005)
41
47
78
Gastrostomy (20) vs Nasogastric tube (21)
Within 24 hours of intubation
<6
Incidence of VAP; length of stay in ICU; duration of mechanical ventilation; mortality
NA
R1B0A1 = 2
Briassoulis (2006)
40
10
NA
Immune enhancing diet (20) vs Regular formula (20)
Within 12 hours of admission
Mean: 6.2
Rate of nosocomial infections; LOS; length of mechanical ventilation; mortality
Cytokines concentrations; nitrogen balance
R2B1A1 = 4
Hartl (2008)
797
≥16
NA
Early nutrition (755) vs Delayed nutrition (42)
Early: within 7 days postinjury; delayed: after 7 days postinjury
3–8
Mortality
NA
S4C202 = 8
Khorana (2009)
40
41
80
Standard EN (20) vs Immunonutrient containing EN (20)
Within the first 24 hour after operation
5–10
Complication rate; LOS in ICU
IL-6 and IL-10 levels
R2B2A1 = 5
Acosta-Escribano (2010)
104
38
86
Transpyloric feeding (50) vs Gastric feeding (54)
Within the first 24 hour after admission
Mean: 6
Incidence of VAP; GI complications; length of stay in ICU and hospital; mortality
Efficacious volume of diet
R1B0A1 = 2
Justo Meirelles (2011)
22
31
90.9
Early PN (10) vs Early EN (12)
Early after admission when hemodynamically stable
9–12
Mortality; complication rate; LOS in ICU
Calories intake; nitrogen intake; nitrogen balance; urinary nitrogen loss; serum glucose; CRP; albumin
R1B0A1 = 2
Chourdakis (2011) †
59
35
79.7
Early EN (34) vs Delayed EN (25)
Early: Within 24–48 hours after injury; delayed: when the gastroparesis resolved (48 h-5d)
≥9
Mortality rate; complication rate; LOS in ICU
TSH, FT3, FT4, cortisol, and testosterone levels
R1B0A1 = 2
Chiang (2012) †
297
0–99
72.4
EN (145) vs Non-EN (152)
Early: within 48 hours postinjury; delayed: no feeding formula added for 7 days
4–8
GOS; mortality; LOS in ICU
NA
S4C202 = 8
Dhandapani (2012) †
95
35
NA
Early EN (64) vs Delayed EN (31)
Early: within 7 days postinjury; Delayed: after 7 days postinjury
4–8
GOS
Anthropometric measurements (mid-arm circumference; triceps skin fold thickness); serum albumin levels; urinary creatinine levels; incidence of malnutrition
S4C2O2 = 8
†
Non-randomized prospective studies.
#
Jadad score for RCTs: randomization (R0-2), blinding (B0-2) and attrition information (A0-1); Newcastle-Ottawa Scale (NOS) for cohort studies: selection (S0-4), comparability (C0-2), outcome (O0-3).
Abbreviations: EN, enteral nutrition; GCS, Glasgow Coma Scale; GOS, Glasgow Outcome Score; LOS, length of stay; NA, not available; NNG, non-nasogastric feeding; NG, nasogastric feeding; PN, parenteral nutrition; vs, versus; VAP, ventilator-associated pneumonia.
Early VS Delayed
8 studies were available for the comparison of early nutrition with delayed nutrition, including 5 RCTs [25] , [26] , [29] , [31] , [37] , and 3 NPSs [39] – [41] . Research conducted by Grahm et al. compared early jejunal feeding with delayed gastric feeding without data of mortality or functional outcome, and thus was excluded from the pooled analyses [29] . The pooling data of the other seven studies indicated that early nutrition was associated with a significant reduction of the mortality rate compared with delayed nutrition, but with moderate heterogeneity (RR = 0.35; 95% CI, 0.24–0.50; P<0.05; I 2 = 44%) ( Figure 2A ).
10.1371/journal.pone.0058838.g002
Figure 2
Comparison of the effect of early feeding and delayed feeding on outcomes in patients with TBI.
(A) Forest plot illustrates the different effects on mortality. (B) Forest plot shows the different effects on poor outcome. PO, poor outcome.
The heterogeneity was explored by subgroup analyses according to the sample size, publication year, study design, starting time of early feeding, and the route of feeding. Most studies have a small sample size below 100 except for two studies ( Table 1 ). Five studies initiated the early nutrition within 72 hours after injury [26] , [28] , [31] , [37] , [40] . In two NPSs, mortality was compared in patients who started feeding within/out of the first 7 days postinjury. Three studies compared the early enteral nutrition with delayed enteral nutrition [31] , [37] , [41] . Two trials compared early PN with delayed EN [26] , [28] . The results of subgroup analyses were shown in Table 2 . Notably, the RCT subgroup, the subgroup of publication year earlier than 2005, and the subgroup that compared early PN with delayed EN were not found to be with statistical significances (RR = 0.62 with 95% CI 0.32–1.22, RR = 0.54 with 95% CI 0.23–1.27, and RR = 0.57 with 95% CI 0.19–1.76, respectively), demonstrating that these subset factors might be the sources of heterogeneity. Meta-regression was not further performed due to the limited number of available studies.
10.1371/journal.pone.0058838.t002
Table 2
Subgroup analyses for studies evaluating the effects of early nutrition and delayed nutrition on mortality.
Subgroups
N
RR (95% CI)
Heterogeneity (I 2 )
Total
7
0.35 (0.24, 0.50)
44%
Sample size
<100
5
0.49 (0.27, 0.87)
30%
>100
2
0.28 (0.22, 0.35)
0
Publication year
<2005
3
0.54 (0.23, 1.27)
35%
>2005
4
0.29 (0.23, 0.37)
6%
Study design
RCT
4
0.62 (0.32, 1.22)
14%
NPS
3
0.28 (0.23, 0.35)
0
Starting time of early nutrition
<72 h
5
0.45 (0.24, 0.87)
56%
<7 d
2
0.26 (0.18, 0.39)
0
Compared route
Early PN vs. delayed EN
2
0.57 (0.19, 1.76)
62%
Early EN vs. delayed EN
3
0.37 (0.21, 0.68)
0
Early EN vs. no feeding
1
0.29 (0.22, 0.38)
-
Unknown
1
0.24 (0.15, 0.38)
-
Abbreviations: EN, enteral nutrition; N, number of studies; NPS, non-randomized prospective study; PN, parenteral nutrition; RCT, randomized controlled trial; RR, relative risk.
Four studies reported the functional outcome on GOS. There was a significant lower risk of poor outcome in patients who received early nutrition compared with those who received delayed nutrition (RR = 0.70; 95% CI, 0.54–0.91; P<0.05) ( Figure 2B ). Four studies investigated the occurrence of infectious complications in both groups, in terms of pneumonia, CNS infection, bloodstream infection, and urinary tract infection. Generally, early nutrition was significantly associated with a lower risk of infectious complications compared with delayed nutrition (RR = 0.77; 95% CI, 0.59–0.99; P<0.05). In the subgroup analyses, however, no similar statistical significance was observed ( Figure 3 ). Three studies reported the rate of feeding complications, including diarrhea and feeding intolerance. However, no statistical significance was revealed ( figure S1 ). Two studies reported the mean value and standard deviation (SD) of length of stay (LOS) in intensive care unit (ICU) [31] , [41] , whereas no significant difference was observed between early and delayed groups (P = 0.68) ( figure S2 ).
10.1371/journal.pone.0058838.g003
Figure 3
Comparison of the effect of early feeding and delayed feeding on infectious complications in patients with TBI.
The funnel plots of data relating to mortality were found to be symmetrical, suggesting a low likelihood of having publication bias ( Figure 4A ). No publication bias was revealed by Egger test either (P = 0.239).
10.1371/journal.pone.0058838.g004
Figure 4
Funnel plots for the detection of publication bias.
(A) Funnel plot of studies evaluating the effects of feeding timings on mortality, which is approximately symmetric. (B) Funnel plot of studies evaluating the effects of feeding routes on mortality, which appears to be symmetric.
EN VS PN
Five RCTs involving a total of 215 patients evaluated delivery route of nutrition support (EN vs PN) in TBI patients ( Table 1 ). As it showed before, two trials compared early PN with delayed EN [25] , [26] . In the other three trials, both PN and EN were started early after admission [27] , [30] , [38] .
Aggregating data of the five studies demonstrated a trend toward a lower mortality rate associated with PN (RR = 0.61; 95% CI, 0.34–1.09; I 2 = 8). Nevertheless, statistical significance was not revealed (P = 0.09) ( Figure 5A ). In order to investigate the impact of different starting time on the results, the subgroup analysis was performed. However, no statistically significant result was revealed in any subgroup (P>0.05) ( Table 3 ). It is similar when subanalyzing the impact of publication year ( Table 3 ). The low heterogeneity may justify the application of a fixed-effect model. In sensitivity analyses by using this model, a marginal statistical significance was revealed (RR = 0.56; 95% CI, 0.32–0.99; P = 0.05). Further, only by excluding the study by Young et al. [28] , a statistical significance without heterogeneity was demonstrated (RR = 0.35; 95% CI 0.15–0.83; P = 0.02; I 2 = 0).
10.1371/journal.pone.0058838.g005
Figure 5
Comparison of the effect of enteral feeding and parenteral feeding on outcomes in patients with TBI.
(A) Forest plot illustrates the different effect on mortality. (B) Forest plot shows the different effect on poor outcome. PO, poor outcome.
10.1371/journal.pone.0058838.t003
Table 3
Subgroup analyses for studies evaluating the effects of parenteral nutrition and enteral nutrition on mortality.
Subgroups
N
RR (95% CI)
Heterogeneity (I 2 )
Total
5
0.61 (0.34, 1.09)
0
Publication year
<1990
3
0.60 (0.29, 1.27)
24%
>1990
2
0.47 (0.09, 2.41)
0
Compared timing
Early PN vs. delayed EN
2
0.57 (0.19, 1.76)
62%
Early PN vs. early EN
3
0.52 (0.16, 1.69)
0
Abbreviations: EN, enteral nutrition; N, number of studies; PN, parenteral nutrition; RR, relative risk.
Three of the five studies reported the functional outcome on GOS. Pooling data revealed a trend toward reducing the rate of poor outcome in PN groups (RR = 0.73; 95% CI, 0.51–1.04). Nevertheless, statistical significance was not observed in the overall analysis or in any subgroup analysis according to the timing of nutrition (P>0.05) ( Figure 5B ). The fixed effect model also failed to show any significant alteration. The reported infectious complications mainly included pneumonia, central nervous system (CNS) infection, bloodstream infection, and urinary tract infection. Pooling data suggested that PN patients may have a slight trend of lower rate of infection complications compared with EN patients (RR 0.89; 95% CI, 0.66–1.22), especially in reducing the occurrence of pneumonia, whereas statistical significance was not revealed (P = 0.48) ( Figure 6 ). The heterogeneity was at an unimportant level across our pooled analyses.
10.1371/journal.pone.0058838.g006
Figure 6
Comparison of the effect of enteral feeding and parenteral feeding on infectious complications in patients with TBI.
The funnel plots of data relating to mortality were found to be symmetrical, suggesting a low likelihood of having publication bias ( Figure 4B ). No publication bias was revealed by Egger test either (P = 0.621).
Standard Formula VS Immune-enhancing Formula
Three trials compared the immune-enhancing formula (arginine, glutamine, probiotics, and ω-3 fatty acids et al.) with the standard formula of EN in TBI patients [32] , [34] , [35] . Infection rate was reported in all trials, and the aggregating data revealed that immune-modulating formula was associated with a statistical significant reduction in infection rate in contrast with the standard formula (RR = 0.54; 95% CI, 0.35–0.82; P<0.05) ( figure S3 ).
Non-nasogastric (NNG) Feeding VS NG Feeding
Five trials compared NNG feeding with NG feeding in EN support. Taylor et al. compared a mixed group (intestinal or gastric) with a standard gastric group, and thus was excluded [18] . Minard et al. compared early nasoenteric feeding with delayed NG feeding [31] . Grahm et al. compared nasojejunal feeding with NG feeding [29] . Kostadima et al. compared percutaneous endoscopic gastrostomy with standard EN in the occurrence of ventilator-associated pneumonia [33] . Escribano et al. randomized patients to receive transpyloric feeding or gastric feeding [36] . The pooling data of 4 trials showed a significant reduction in the occurrence of pneumonia in patients receiving NNG feeding compared with those receiving NG feeding (RR = 0.62; 95% CI, 0.40–0.96; P = 0.03). The NNG group was associated with a trend to reduce mortality rate and shorten ventilator day, whereas failing to show statistical significances. Three trials additionally reported the LOS in ICU, and the aggregated results revealed no statistical significance in the two arms ( figure S4 ).
Discussion
We carried out a comprehensive literature search to detect prospective studies, including RCTs and NPSs, to compare the effects of different routes, timings and formulae of nutritional support on clinical outcomes in TBI patients. Explicit criteria were utilized for study selection and methodological quality assessment.
In comparison of different timing for nutritional support, our meta-analysis demonstrated beneficial effects of early nutrition on reducing mortality, improving functional outcome, and decreasing infectious complications. Although a trend was indicated for early nutrition in lowering the risk of stratified specific infections, no statistical significance was revealed. There was also no significant difference in feeding-related complications between the two arms. Notably, opposite to pooled data of NPSs, the data of RCTs failed to demonstrate statistical significance in mortality. The RCTs were largely carried out in early years with smaller sample sizes and limited nutritional support approaches, which might contribute to this discrepancy. Similar concern may also explain the heterogeneity from publication year. The earlier Cochrane review also suggested a trend toward using early nutrition to reduce mortality and improve functional outcome by analyzing fewer studies, whereas without statistical significance [42] . Furthermore, several queries were concerned. For study by Hadley et al., the PN and EN were both started within 48 hours after admission, and thus it may not be reliable for considering it as the comparison of early nutrition with delayed nutrition [27] , [42] . Additionally, they included one study that compared different infusion speeds of EN [18] . In light of the simultaneous initiation of EN in both groups, it is not likely to be justified to consider this trial as the comparison of nutritional timings, and thus we excluded it. The Brain Trauma Foundation has cautiously recommended achieving full caloric replacement by day 7 following TBI based on limited evidence [9] . Taken together, our results reinforced the inclination of early nutrition for TBI patients.
When compared with EN, our results showed that there might be a trend toward using PN to reduce mortality and improve functional outcome, whereas statistical significance was only marginally revealed when the fixed effect model was used. In the subgroup analyses based on timing of feeding and publication year, no statistical significant findings were revealed. Moreover, significant difference in the rate of complications was not recognized in any subgroup analysis between the two arms. In fact, EN and PN show unique advantages, respectively. The use of EN is superior to PN in patients with functioning gastrointestinal tracts [43] . Compared with PN, EN formulae may conveniently make use of more effective substrates to support cell and organ function, have a lower risk of hyperglycemia or hyperosmolarity, be administered at lower rates to avoid overfeeding, and better support the gut mass and barrier function. However, the use of enteral feeding in patients with gastrointestinal intolerance is associated with underfeeding and consequent malnutrition [43] , [44] . Less than 70% of patients receive an adequate enteral caloric intake even in the most experienced and motivated ICUs [45] . In comparison, PN patients have benefits in obtaining more dependable nutrient bioavailability, getting nutrition effects in a shorter period, requiring no functional GT tract, and staying away from satiety, abdominal distention or other enteral feeding complications [43] . However, overfeeding (the administration of excess dextrose, fat, or calories) and refeeding syndrome (rapid feeding of patients with preexisting malnutrition) may occur, and thus induce a variety of metabolic complications, including hyperglycemia, hypertriglyceridemia, thiamine deficiency, hypervolemic, and hypercapnia [44] . In our pooled studies, enrolled patients unanimously experienced moderate or severe TBI, mostly with GCS lower than eight [25] – [27] , [30] , [38] . They are always comatose, intubated or mechanically ventilated, with malfunctioning parasympathetic and sympathetic system, disturbed hypothalamic-pituitary axis, elevated intracranial pressure (ICP), increasing endogenous opioids and endorphins, and widespread prescription of narcotics. All of these unfavorable factors may contribute to impaired GT function, delayed gastric emptying, and increased risk of EN intolerance [3] . In this context, PN might be superior to EN for initial life saving nutritional support. Notably, PN was unanimously initiated early after admission in related studies. Thus, our result should not be misunderstood as opposition to the suggestion that EN is preferable whenever possible with functional gastrointestinal tracts [46] . However, data of feeding related complications, especially the data of hyperglycemia were insufficient across included studies, and further persuasive evidence is warranted.
Given the prevalence of inconveniences for routine nasogastric EN, other alternative EN routes have been attempted, including nasojejunal feeding, percutaneous endoscopic gastrotomy feeding, and transpyloric feeding [29] , [33] , [36] . Pneumonia rate was shown to be significantly reduced by NNG, which may be associated with the prevention of aspiration by NNG feeding. There was also a trend toward reducing mortality rate and decreasing ventilator days, but failed to demonstrate statistical significances. It has been evaluated that at least 20% of TBI patients don’t tolerate enteral alimentation at all in the first week [29] . By using NNG route, patients will probably tolerate enteral feeding as well as avoid the hyperalimentation brought about by PN. Although data on feeding complications were scant, in light of its well tolerance and prevention of pneumonia, we are inclined to side with ESPEN guidelines, which suggested that when jejunal feeding can be carried out easily, it should be given [47] .
Furthermore, our results showed that immune-enhanced formulae were associated with a significant reduction in infectious complications compared with standard formulae. Although a growing number of studies emphasized the importance of nutrition content of foods on post-injury recovery, studies relating to TBI are scant, especially according to our criteria of study design. In fact, the effect of immune-enhanced formulae has been widely investigated in general population, but with confusing and undefined conclusions for critically ill patients [48] . The ESPEN guideline has recommended that immune-modulating formulae, enriched with arginine, nucleotides, and ω-3 fatty acids, are superior to standard enteral formulae in trauma patients [47] . In contrast, the updated guidelines from the Canadian Practice Group in 2009 and the American Dietetic Association evidence analysis library did not recommend the routine use of immune-modulating diets in critically ill patients [48] . The most prominent controversy was the effect of immune-modulating diet on mortality and functional outcomes. For example, the use of arginine-contained formulae has shown greater mortality, and it is hypothesized that arginine may be converted to nitric oxide and thus contributed to hemodynamic instability [4] . Our results could only suggest the benefits of immune-modulating diet for TBI patients in reducing infectious complications. The effects on mortality or functional outcome could not be elucidated by the few studies, and further studies are warranted to investigate the effects of particular diets on the outcome of TBI.
We are aware of the limitations of this meta-analysis. Trials with statistically significant results may be more likely to be published and cited, and are preferentially published in English language journals [49] . We included only studies written in English language and therefore, may have missed relevant studies published in non-English language journals. Besides, several studies (e.g. Rapp et al. and Young et al.) are included in 2 or more of the reported analyses; this has a potential risk of overemphasizing positive results. Although results from Egger’s tests and funnel plots did not show evidence of publication bias, their capacities to detect bias was limited by the small number of studies [49] . The studies included were of relatively poor quality, with most of RCTs having a Jadad score of <3 (13/15). Only a few trials described the method of randomization [34] , [35] . Blinded design was only described in two trials on immune-enhanced formulae [34] , [35] . Although it seems difficult to conceal the route of nutritional support, studies with inadequate or unclear concealment of allocation may overestimate the intervention effect. The sample sizes were relatively small across included studies, especially in the RCTs. Small sample size might contribute to the failure of randomization and imbalance between clinical variables, and thus failed to detect the statistically significant effects. In fact, the compared different arms were not well-controlled. For example, the comparison of route for nutritional support would be more convinced if started simultaneously. However, it is commonly seen that EN was initiated late until the recovery of gastroparesis [26] , [27] . In fact, in subgroup analyses of timing of feeding, two studies that compared early PN with late EN were both revealed to be the potential source of heterogeneity. Question may be raised that whether the effects found were related to the parenteral nutrition or perhaps more due to the early onset of nutrition versusing delayed onset. The impossibility to differentiate here was a substantial confounding factor in the interpretation of results. In the sensitivity analyses of route of feeding, the exclusion of study by Young et al. has led to a significant change of results. Notably, only this study initiated the EN support until the termination of low wall suction, which was non-conventionally performed and might contribute to the heterogeneity [28] . Furthermore, our meta-analysis was absent of aggregating various nutrition indexes, such as caloric intake, nitrogen intake, and nitrogen balance. In previous systematic reviews, it has been revealed that the measurement methods, definitions of metabolic abnormalities, and energy expenditure following TBI varied greatly, which may restrict the incorporation [50] , [51] . Additionally, the included studies utilized different criteria for inclusion and exclusion. Especially, the differences in severity of disease across studies may explain some of the heterogeneity. For example, a number of studies specifically investigated patients with mechanical ventilation [32] – [34] , [36] – [38] . All of these limitations restrict the strength of conclusions drawn from our meta-analysis.
Though disputes of optimal nutritional support would continue, we postulate that the optimal clinical decision in nutritional support should be personalized, in terms of the individual profile, including nutritional status, severity, complications, feeding tolerance, and day-to-day changes in clinical conditions. Last but not least, greater multidisciplinary efforts from nutritionists and clinicians are required for better management of the nutritional support for TBI patients.
Conclusions
Our meta-analysis lends support to early initiation of nutritional support for TBI patients, which can decrease mortality, reduce complications and facilitate recovery. PN appears to be superior to EN in reducing mortality and improving outcome in the acute gut-intolerant phase of TBI. Immune-modulating formulae seem to be superior to standard formulae in reducing infectious complications. Small-bowel feeding was recommended if possible. However, our results should be interpreted with caution given the various limitations. Further well-designed RCTs are expected to clarify the optimal nutritional strategies for TBI patients.
Supporting Information
Figure S1
Forest plot shows the effect of early nutrition and delayed nutrition on feeding compliations. (A) Forest plot illustrates the effect on diarrhea. (B) Forest plot illustrates the effect on feeding intolerance. I, intolerance.
(TIF)
Figure S2
Forest plot shows the effect of early nutrition and delayed nutrition on length of stay in the intensive care unit.
(TIF)
Figure S3
Forest plot shows the effect of standard and immuno-modulated nutritional formulae on infectious complications.
(TIF)
Figure S4
Forest plot shows the effect of non-nasogastric and nasogastric enteral feeding on outcomes in patients with TBI. (A) Forest plot illustrates the effect on pneumonia. (B) Forest plot shows the effect on mortality. (C) Forest plot shows the effect on ventilator days. (D) Forest plot shows the effect on length of stay in the intensive care unit. Pneumo, pneumonia.
(TIF)
Checklist S1
PRISMA Checklist.
(DOC)
Protocol S1
PRISMA Flowchart.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Non-communicable diseases (NCDs) are considered a global health challenge. The World Health Statistics (2023) estimated that of the 55.4 million deaths from all causes in 2019, 74% deaths were from NCDs [ 1 ]. Global deaths projections indicate that by 2048 there will be nearly 90 million annual deaths, with a 90% increase in the deaths attributable to NCDs from 2019 [ 1 ].
The major burden of NCDs occurs in low-and middle-income countries (LMIC) [ 2 ]; which registered 77% of the approximately 41 million deaths attributed to NCDs in 2021 [ 3 ]. LMICs, including countries in Sub-Saharan Africa (SSA), are expected to register a greater increase of the NCD burden due to their population ageing and rapid urbanization [ 4 ]. This is a major problem as LMIC healthcare services are predominantly oriented towards acute conditions and infectious diseases. Globally, international donors have the notion that investment should first go to communicable diseases before they start focusing on NCDs, and locally, health systems base their prevention and control measures on infectious diseases [ 5 ].
The demands imposed by the rising NCD burden on LMIC health systems including diagnosis, management and treatment, will require a reorientation of healthcare systems to improve their responsiveness and address the new challenges accordingly [ 6 ]. Overall health system strengthening has been identified as a critical strategy in addressing the NCD burden [ 6 , 7 ]. Health system strengthening requires allocation of adequate numbers of qualified human resources (HR), one of its key building blocks. One of the most pressing challenges of LMICs’ health systems is the HR scarcity. In 2019 it was estimated that in SSA there were 2.9 physicians per 10,000 inhabitants in contrast to 6.5 physicians per 10,000 inhabitants in South Asia. Similarly, nurse and midwifery personnel ranged from 9.7 per 10,000 inhabitants in South Asia to 18.3 per 10,000 inhabitants in SSA [ 8 ]. These figures are up to 5 to 10 times less than the estimates for high-income countries’ physicians (33.4 per 10,000 inhabitants) and 6 to 11 times fewer nurses and midwifery personnel (114.9 per 10,000 inhabitants) for the same year [ 8 ]. This scarcity is further aggravated by factors such as training, mentoring and other supporting measures from health systems, which are largely shaped by health sector priorities [ 9 ].
In Mozambique, NCD have received some prominence since the early 2000s when reports started to identify a significant burden that has worsened over time. For example, in 2001, mortality from NCD ranged from 13 to 24% [ 10 ], then rose to 28% in 2012 [ 11 ]. Similarly, the estimated prevalence of hypertension and type 2 diabetes increased from 33% to 39% and 2.9% to 7.4%, respectively, from 2005 to 2015 [ 12 – 15 ]. This increase in prevalence was accompanied by a deterioration in other NCD health parameters, such as fewer people being 1) aware of their disease, 2) being in treatment, and 3) with their diseases controlled [ 14 – 16 ].
In recognition of the burden of NCD in Mozambique, particularly of hypertension and diabetes, NCD were included in national level health and priority policy documents, such as the Poverty Reduction Action Plan (PARPA II), which declared NCD as a priority in 2006, and later in 2008, the first National Strategy for Prevention and Control of NCD [ 17 ] was developed. The strategy aimed to reduce exposure to NCDs risk factors and associated morbidity and mortality by increasing awareness about NCDs, strengthening and integrating NCD-related training, improving access and quality of NCD prevention and care services, and strengthening NCD surveillance, monitoring and evaluation [ 10 ]. However, the reported HR shortage in Mozambique’s health system is severe, with less than one doctor per 10,000 people and approximately three nurses per 10,000 people [ 18 , 19 ], which poses a significant challenge to the effort to strengthen the health system.
This study’s aims were: (i) to explore self-reported training and competencies of NCD-related healthcare providers, and (ii) to assess the perceived challenges posed by the current health system structuring in the health sector context of the country. By documenting the perceived challenges to managing diabetes and hypertension, the study team expects to contribute to informing the healthcare workers’ training requirements and healthcare services structuring to provide NCD-related services tailored to the needs of Mozambique’s and similar health contexts.
Material and methods
Study design
We conducted a qualitative study between February and March 2017. This study was comprised of semi-structured interviews with government officials within the Mozambican Ministry of Health, district health authorities, health facility managers, and health providers. The interviews focused on exploring factors related to the training healthcare workers received, the competencies acquired, and the existing resources to support NCDs (diabetes and hypertension) prevention and management at the primary health care (PHC) level. An interview guide was developed based on the literature and integrated the following topics: healthcare workers, diagnostic tools and technologies, medicines and guidelines, referral system and other type of support, and NCD prioritization. The consolidated criteria for reporting qualitative research (COREQ) [ 20 ] was used for this study reporting.
Study setting
This study was conducted at the national level and in two districts in Southern Mozambique. The districts were Nlhamankulu and Moamba, with suburban and rural characteristics. In both districts, two primary health facilities were selected from the 15 existing facilities. In Nlhamankulu District the facilities selected include Xipamanine and Chamanculo health centers; and in Moamba District the faciltiies selected include Moamba and Sábiè health centers. Healthfacility selection was based on their characteristics, such as location (suburban vs. rural areas), easy access, catchment area, target population, and the number of patients assisted with NCD. Primary health facilities constitute the entry point to the national health services, offering mainly promotive, preventive, and curative services for minor diseases, such as acute respiratory illness and diarrhea without dehydration. In addition, health programs for malaria, HIV/AIDS, sexually transmiited infections (STI), and tuberculosis are primarily implemented at this level. These primary health facilities comprise 95% of the national health facility network and are estimated to cover 80% of the population with health services [ 18 ]. However, despite this high proportion, they only have 36% of the total health care providers [ 21 ].
Nlhamankulu District has seven health facilities, comprised of two general hospitals and five health centers. These seven facilities have 209 health care providers, including mid-level technicians, general practitioners, and specialist doctors. Of the 209 providers, 80% are distributed between the two selected study facilities, which combined provide health services to 62% of the districts 130,000 inhabitants. Per annual district health reports, the main reported health problems in 2018 were malaria, diarrhea, malnutrition, HIV/AIDS and tuberculosis [ 22 ]. The report did not include data on hypertension and diabetes.
Moamba District has ten health facilities, comprised of nine health centers and one health post, all primary level, with 102 health care providers and 25 community health workers. Of the 102 providers, 66% are distributed between the two study facilities, covering almost 50% of the district´s 90,000 inhabitants [ 23 ]. As in Nlhamankulu, the main health problems reported in 2018 were malaria, diarrhea, malnutrition, HIV/AIDS and tuberculosis. As of 2018, there were 963 cases of hypertension and eight cases of diabetes reported in Moamba [ 23 ].
In general, at the PHC level, NCD related services are provided by three types od cadres. Doctors [ 24 ] and health technicians [ 25 ], who have more clinical focus, including diagnosis and treatment, and nurses [ 26 ], who are more focused on general care to patients, not necessarily directed to a specific type of disease. The pre-service training for all categories have a number of hours dedicated to practice at the health facilities, where students apply the theoretical knowledge, which is more than two times the hours dedicated to learning theory at the classroom [ 24 – 26 ]. Practical learning components within health facilities are shaped by the health sector priorities [ 9 ]. Although both infectious diseases and NCDs co-exist, attention is much more directed towards responding to the former, more specifically to HIV, tuberculosis and malaria, as the most prioritized [ 5 , 6 , 10 ].
Data collection procedures and sampling
Semi-structured interviews
Semi-structured interviews were conducted by five researchers (two females and three males), each with a minimum of a bachelor’s level education in anthropology, biology, or medicine, and all of them had previous experience conducting qualitative interviews ranging from 2 to 15 years. Interviews were over two months, between February and March 2017.
The researchers had a three-day refresher training on qualitative methods and discussed the study protocol to gain familiarity. Interviews were audio recorded, and the interviewers took notes during the interview. Interviews were conducted in Portuguese because all participants could communicate fluently in Portuguese. Interviews lasted between 30 and 60 minutes and continued until all topics in the interview guide were covered and no new issue emerged (theoretical saturation). The interviews were anonymized.
In order to attain a complete understanding of the breadth and depth of factors that may influence NCDs-related health professionals’ qualifications and their capacity to provide optimal care, a variety of stakeholders were interviewed. A total of 26 MOH health workers and managers at various levels were invited for interviews, and two were unavailable due to concurrent agenda, resulting in 24 interviewees conducted and a 92% response rate. Three national level, four district level, and seventeen health facilities levels managers and providers were interviewed. We used a convenience approach to select respondents, considering a minimum of 6 months at the current position or role, deemed enough to understand the context within which they operate. At the national level, interviewed individuals were viewed as key opinion leaders from the Ministry of Health who could understand the contextual factors leading to the development of the policy documents and the development of the health workforce. At the district level, health authorities (health directors and chief medical officers) and focal points for the different NCD programs were interviewed to explore their experiences regarding the availability of staff and the type of conditions in place to provide specific health care services at that level. Lastly, at the health facility level, managers, wards-in-charge or sectors-in-charge personnel and health practitioners were interviewed to explore their daily experiences in implementing NCD-related activities.
Data management and analysis
The audio recordings were transcribed verbatim in Portuguese and NVivo software ( NVivo 12 , QSR International) was used to facilitate coding (thematic analysis).
A codebook with a predetermined list of themes (categories) and sub-themes (sub-categories) was developed and a deductive perspective used for data analysis. These themes were identified by drawing on the adapted model shown below ( Fig 1 ), which puts healthcare providers’ capacity building as the core asset for adequate provision of NCD-related care. We used a phenomenological approach to explore healthcare providers and their manager’s perceptions of how the structure of the health system influences how they are prepared to respond to the NCD service’s needs, and the additional investment on diagnostic tools, equipment, supplies, and medicines necessary to provide quality services under the policy prioritization context it functions.
10.1371/journal.pone.0297676.g001
Fig 1
Study’s data analysis model.
The lead and second authors conducted a parallel coding guided by the developed codebook. After concluding the process individually, they jointly compared the coded transcripts. They reviewed all coded transcripts to reach a consensus on which quotes to present that best captured respondents’ perspectives and responded to the analysis domains. After being sorted by theme (category), data was summarized, synthesized, and abstracted in each category. Translation of the relevant sections of the transcriptions to English language was done at the time of writing this manuscript.
Ethics approval and consent to participate
This research was conducted in accordance with the Declaration of Helsinki for research involving human participants. The Institutional Bioethics Committee of Health of the Faculty of Medicine and Maputo Central Hospital in Mozambique approved the research project with the reference: CIBS FM&HCM/55/2016. Written informed consent was obtained from each participant after being taught about the voluntary nature of the study and that they could withdraw at any time without consequences. Interviewers introduced themselves and then provided the interviewees with an information sheet about the study and gave each participant an informed consent document to read and sign. Time was provided for asking and answering questions about the study before the consent form was signed. They were also informed about the risks and benefits of participating in the study.
Results
For both diseases, hypertension, and type 2 diabetes, we describe the findings following the three domains identified on the model: providers capacity building, health system, and policy. Within each domain we identified common barriers ( Table 1 ) across the two districts and all four health facilities, which were: (i) inadequacy of NCD’s healthcare providers training and poor guidance to enhance providers skills; (ii) disrupted referral system, poor availability of diagnostic meansand medicines; and (iii) low NCDs’ priority.
10.1371/journal.pone.0297676.t001
Table 1 Structured themes and subthemes from the interviews by level of analysis.
Analysis Level
Themes
Sub-themes
Healthcare providers capacity building
Inadequacy of NCD’s healthcare providers training
Pre-service training inadequate for NCD-related needs
Lack of in-service training opportunities
Poor guidance to enhance providers skills
Lack of NCDs’ related guidelines and algorithms
Health System Structuring
Disrupted referral system
Incomplete patient follow-up through the referral system
Poor availability of diagnostic means and medicines
Shortage of functioning diagnostic tools and equipment
Lack of local laboratories’ capacity
Inconsistent availability of laboratory supplies
Limited quantities of available medicines
Policy
Low NCDs’ priority
Resources planning incompatible with the needs
Poor investment on NCDs related services
Healthcare providers capacity building
Inadequacy of NCD’s healthcare providers training
Pre-service training inadequate for NCD-related needs . The NCDs related competencies acquired during pre-service training are limited as evidenced by the documents review presented above and supported by the interviews. For example, six of the eight healthcare workers interviewed referred to the need to consolidate and broaden their NCD-related knowledge, to improve their capacity to treat their patients and be able to teach them how to use non-therapeutic measures confidently.
P1: “…one of the main problems we have is that clinicians can only provide limited information to patients…they are able to tell that they (patients) have the disease and prescribe the first line medicines but are unable to explain about appropriate diet and lifestyle and long-term measures or even why patients have to live on drugs for their entire life.
Likewise, health managers and non-physician providers recognized the providers’ deficiency to diagnose and adequately treat NCDs and expressed the need for acquiring additional skills due to limited pre-service training exposure to NCDs’ core subjects. The lack of skills delays the patients’ diagnosis and affects the appropriate NCDs management, mainly diabetes, at the PHC.
P2: “…the training (pre-service) and information I have on diabetes and hypertension is not enough to take care of patients… I feel like I need more.”
FP1 : “… hypertension is not a problem amongst us here…providers treat and follow-up patients with no complaints … but we have many difficulties with diabetes …”
M1 : “I notice that they (providers) have a basic knowledge , but not so thoughtful that it can ensure that they suspect or confirm all existing cases (of diabetes) here . ”
This has greatly contributed to the observed NCDs related healthcare providers shortage described in official documents from and surveys conducted by the MOH. This shortage was evident during the study’s data collection, observing that the four health facilities visited had one or two non-physician clinicians (providers who complete 2 to 3 years of clinical training after secondary education) [ 27 ], who partially dedicated their time to NCDs in addition to the doctors, who were available in only two health facilities but not always on duty due to others non-clinical responsibilities.
Lack of in-service training opportunities . Respondents noted that short in-service training is required to refresh healthcare providers’ skills to manage NCDs. All eight providers interviewed recognized the benefits of having short in-service refresher courses to improve their skills in managing NCDs,
P4. “In my opinion, having in-service training would help us greatly improve our approaches towards patients with diabetes and hypertension…”
and, to ensure updated approaches and the use of new available techniques and diagnostic means and tools.
P6 : “…the only time I was trained for these two diseases was during my pre-service training…I think I need more…there are new approaches, and we are back on time…”
However, they reported long periods without having the opportunities to receive such training, as opposed to the periodic sessions witnessed within the HIV, TB, and Malaria programs.
P5 : “… since I have started working at this Centre (more than 5 years ago) we never received any type of in-service training for hypertension or diabetes …”
The reported gaps in pre- and in-service training collectively contribute to low confidence of healthcare workers in managing NCDs. Respondents from two health facilities highlighted the delay in diagnosing patients resulting from their fear of the complexity of managing diabetes.
P3 : “… We would always avoid doing screening for diabetes , patients would present signs and symptoms , but we would ignore them and look for other diagnoses … when we decided to measure the first glycaemia it was around twenty-five , maybe the patients were already developing the disease and we ignored it until that it ended up reaching that extreme …”
Poor guidance to enhance providers skills
The managers interviewed at all levels suggested a degree of negligence from frontline providers by not using the diabetes and hypertension management guidelines developed to help them improve their capacity to diagnose and treat such conditions.
M4: “…we have developed specific guidelines (for diabetes and hypertension) and sent them out to all provinces, however every time we do supervision visits, we find them stored in drawers under their (providers) desks…”
On the other hand, the healthcare workers from the rural district were unanimous about not receiving any clinical guideline and, therefore, stated this resulted in either not having the confidence of managing the cases appropriately or contributing to inconsistent procedures between clinicians to manage the same condition.
P2 : “… I think one aspect that should be improved are the procedures…they are not uniform and therefore each clinician proceeds his/her own way … we need to avoid different approaches for the same problem …”
Health system organization
Disrupted referral system
Incomplete patient follow-up through the referral system . Health workers and managers at the PHC level pointed out that once patients are referred to higher levels for acute management, it becomes more challenging to follow them up because they do not receive appropriate support to monitor patients after being assisted by a specialist. Additionally, they lack resources and medicines since this level of care does not have the same resources as is observed with higher levels, which interrupts the patient follow-up and jeopardizes the trust of the service.
P8 : “… our providers have difficulties to adjust medications at this level , so we send them to higher levels , however they get lost when they come back and need different medicines than those that we have at this level …”
Poor availability of diagnostic means and medicines
Shortage of functioning diagnostic tools and equipment . National-level managers indicated that all health facilities in the country consistently received equipment for screening and testing,
however, health facilities’ managers and healthcare workers reported an equipment shortage and malfunctioning.
P3 : “… we don’t have enough working equipment…it should be normal for us to have hypertension devices for each technician or per cabinet and a stethoscope , but we don’t…we only work with two devices for this entire health facility …"
M1 : “… regarding glycaemia , we have a chronic problem , when we have strips , they are not suitable to the glucometers we do have , and vice-versa …”
P1 : “… when we talk about hypertension , the major problem we have here is related to the equipment (sphygmomanometers)…we have the digital type , which uses batteries , and often times we don’t have batteries …”
Lack of local laboratories’ capacity . According to healthcare workers, there is lack of laboratory capacity, which demoralizes both providers and the patients to seek diabetes-related care due to long waiting time, limited availability of supplementary exams, and the need to expend additional money to return to the health facility only to collect the result.
P6: “…it is difficult to follow-up diabetic patients when we only use glycemia…other means such as glycated hemoglobin, which is ideal for these cases, is never available…”
P4: “…we don’t process glycemia here, so we have to send the blood sample to Matola Provincial Hospital (referral hospital) …sometimes we have the results the following day, but others not…so we ask the patients to return after two or three days. However, some of them do not return, saying they do not have money for transport…they actually get upset when they come and don’t find their results……”
Inconsistent availability of laboratory supplies . A healthcare worker from one of the health facilities highlighted the generalized laboratory supplies stock-outs often with a long waiting time to replenishment.
P7: “…we have had difficulties, sometimes in the lab we have reagents stockouts…it’s normal for us to stay a week without blood glucose reagents, so we send the patients to Chamanculo to do it there, although sometimes they also have the same problems…”
Limited quantities of available medicines . All managers at different levels pointed out that the system was experiencing varied levels of essential NCDs medicines stockouts. District-level managers associated the medicines shortage with the absence of health partners interested in NCDs as opposed to what they observe for communicable diseases such as HIV and TB, which have enough resources and fewer reported stockouts.
M3: “…most of the times we have to work with almost no resources, not even basic (first line) medicines and that makes it difficult for us to diagnose and treat these diseases…if we had the same resources we have for HIV and TB we would do better…”
Policy
Low NCDs’ priority
Resources planning incompatible with the needs . Respondents raised this issue, mentioning that low prioritization of NCDs has negatively influenced the quality of their response, resulting in poor resource availability due to insufficient financing. For example, due to limitations in BP cuffs and blood glucose strips, participants mentioned that they only measure parameters like blood pressure and blood glucose on selected patients, such as those 35 years and older, based on their perceived higher risk.
PF2: “…with the equipment problem we have we try to make a rational use of the existent, for example for hypertension we measure blood pressure of patients (health facility users) who are old and only rarely we measure those under the age of 35 if they have any complaints that relate to the disease…for diabetes we have many more difficulties…”
The reported scenario is opposed to other health programs, such as those related to HIV and TB, which have enough resources and rarely register stockouts.
P8“…most of the times we have to work with almost no resources and that makes it difficult for us to diagnose and treat these diseases…if we had the same resources we have for HIV and TB we would do better…”
Furthermore, while it is rare to notice stockouts of HIV and TB related medicines, the same does not hold for diabetes and hypertension. These patients, who should collect their medicines every month, in the best case scenario, collect them for half that period, when they are available, to allow more people to access medication. Consequently, the cost of collecting medicines for these patients is exacerbated by the combined costs of transport and opportunity, given the uncertainty of its availability. Some patients opt to by in private pharmacies through their own means or loans.
FP2: “…it is normal to have limited stocks of anti-hypertensives in this health facility…we have to manage what we have, so instead of giving enough for the whole month, we give small portions and ask the patients to come back to see if we received more medicines later…”
P8: “…patients with conditions buy at the private pharmacy, but we only have one pharmacy here in the countryside, and sometimes they also run out of stock…”
Poor investment on NCDs related services . Some healthcare workers understand the observed level of investment in HIV and related services as a clear message about where their focus should be, disregarding the importance of other health conditions including NCDs.
P5: “…I think we have cases of diabetes and hypertension in the district, however we don’t make so many diagnosis because our main focus is on HIV patients…”
Discussion
Although the reviewed cadres’ training curriculum incorporates NCD-related content, the findings from this study suggest that nurses and non-physician clinicians rarely engage in diagnosing and managing NCDs at the PHC level due to limitations in their pre-service training. The healthcare sector of Mozambique lacks diagnostic equipment, tools and supplies, and medicines which constitutes an additional barrier to healthcare providers ability to manage NCDs appropriately; and, there is a lack of guidelines and training in their use to improve providers’ skills.
The training models adopted by health training institutions in LMIC, which have a significant practical component completed within the health facilities, are shaped by the health sector´s priorities [ 9 ]. Although both infectious diseases and NCDs co-exist, attention is much more directed towards responding to the former, more specifically to HIV, tuberculosis and malaria. This is frequently driven by donor funding, as shown by the high level of investment and the amount of existing supporting resources (guidelines, diagnostic equipment and supplies, medicines, patient registries, to name some) in health facilities [ 28 – 30 ]. Therefore, health students are much more likely to practice and learn about infectious diseases, while the NCDs lag behind. Secondly, there is a well documented investment in short in-service trainings targeted towards doctors, non-clinician health technicians, and nurses that mainly focuses on three priority diseases, HIV, tuberculosis and malaria [ 29 ]. Funders are interested in this type of training to show results within the time frame of specific grants by continuously adapting to the local needs, ensuring providers are up-to-date on the latest approaches and international guidelines and improving their performance against pre-determined indicators [ 9 , 29 ]. In Mozambique, external institutions supporting the large programs for HIV, TB and malaria will frequently provide professionals with technical skills that are available to provide mentorship and guidance to the frontline providers in case they need it [ 9 ]. However, this same level of technical support has not yet prioritized NCD care.
Apart from the healthcare workers’ limited exposure to NCD-related knowledge and practice during their pre-service training, after their integration into the healthcare system they face additional challenges. The necessary equipment and tools to support them with their activities were neither sufficient nor adequate to help them improve the quality of services they offer. For example, the diagnostic equipment required to screen for and diagnose both diabetes and hypertension was revealed to be unreliably available [ 31 ]. When available, the equipment often did not correspond to the actual needs of patients, such as those needing extra-large BP cuffs even though obesity is a common risk for diabetes [ 15 ].
Moreover, providers could benefit from using MOH produced clinical guidelines to improve their management of diabetes and hypertension. Apart from supporting healthcare providers, the guidelines serve other purposes, including harmonization of clinicians’ approaches [ 32 ], improving the confidence of clinicians who are unsure on how to proceed, and reversing clinicians’ beliefs adapted to outdated practices [ 33 ]. However, some mid-level providers perceive their use as an additional burden to the providers’ complex working environment.
Lastly, the NCDs referral system does not consider the uneven distribution of resources between levels of care, further aggravating the perceived barriers at the PHC level. The nature of the NCDs discussed here requires the same core type of resources for their management when there are no complications. Therefore, it would be needed to ensure the availability of similar types of resources, including qualified providers, laboratory equipment and consumables, and medicines, at the different levels, including the PHC where the demand is higher and there are competing priorities.
Although there are apparent efforts from the government to change the current scenario, early detection, treatment, and control of diabetes and hypertension is unsatisfactory in Mozambique’s health care system, similar to other countries in the African region, which results in frequent and severe complications and sequelae, along with premature deaths [ 14 , 16 , 34 ]. For more than a decade now, Mozambique has taken decisive steps in recognizing the importance of NCDs in the national epidemiological profile by highlighting them in the government’s guiding documents, developing NCDs specific strategic plans, and investing in medicines and laboratory consumables for NCDs to provide them to the public at no cost. The introduction of reforms in the system to tackle the NCD burden and ensuring a thorough implementation of the plans would enable an appropriate response from the health services, but this is not visible. A similar observation was made in a recent study by Heller et al. [ 35 ], who concluded that policies, strategies and effective interventions to NCD control exist, but their applicability is questionable, especially in LMICs. In fact, the National Health Service has theoretically adopted integrated management of diseases to allow a comprehensive provision of health care. However, most of the programs, including HIV, TB, and malaria, still function vertically, which further affects the availability of other less prioritized services [ 36 ], such as the NCDs. This leads to uneven access to resources and services, contributing to the system’s inefficiencies [ 36 , 37 ].
There were some limitations in this study. First, the NCD-related skills reported relied mainly on self-reported assessment, which could affect the reliability of the study findings. A strength was that we used data triangulation to enhance the study findings, validity, and credibility. Second, in Mozambique there are several independent health training institutions graduating students, that are then placed within the health system which may mean that some health professionals were not trained under any of the assessed curricula. Further, we did not conduct an in-depth evaluation of the content of the training curricula since this was not the aim of the study. We only visited a limited number of health facilities for this study, all located in a relatively resources favored provinces (Maputo Province and City). This may have biased our findings by underestimating the level of resource scarcity in the PHC-level health facilities.
The health system needs to adapt to offer effective treatment, self-management support, and regular follow-up to NCD patients [ 38 – 40 ]. Some tasks, such as counselling for self-management, are needed to manage NCDs but are only learned by practice. Unfortunately, most of the time, the workload of clinicians does not allow them sufficient time to provide adequate counselling to patients and their families. Tasks like this can be shifted to non-clinicians, such as lay counsellors, who have been successfully used in other health programs. Thus, comprehensive NCD care can be provided to patients by providing appropriate training and evidence-based skills, as well as organizing health care providers into team care models for managing chronic health problems [ 38 ]. This approach would reduce the number of trained professionals needed to achieve satisfactory results of NCDs control [ 41 – 43 ] by restructuring the few existing providers and ensuring an even resources distribution among the types of health problems facing the healthcare system.
Conclusions
Mozambique´s health strategic and policy documents reflect the concern at national governmental level about NCDs, however this has not translated into practice. There is a gap in health resources, including human, financial, and material, to respond to the needs faced by the country’s health system. This is more significant for NCDs as they need to compete with the major infectious diseases, which overall are better funded by external partners. The healthcare workers available to provide NCD-related care and management at the PHC level of Mozambique’s health system are inappropriately skilled. In addition, these professionals face a lack of diagnostic equipment and tools to adequately respond to NCD-related needs, particularly at the primary care level. Any increase in global and national responses to the NCD challenge must include investments in human resources and appropriate equipment.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Dengue, an arbovirus infection with an explosive epidemic potential, is a major public health problem in many tropical and subtropical countries today [ 1 ]. The incidence of dengue has dramatically increased over the past decades, with the number of symptomatic dengue infections reported to be doubling every 10 years between 1990 and 2013 [ 2 ]. Dengue hemorrhagic fever (DHF), a severe and potentially life-threatening form of the disease, has also reported to be increasing steadily during this period, driving the increase in hospitalization rates for dengue, particularly in children [ 3 – 7 ]. Yet, an increasing number of studies have shown that there is substantial under-reporting of dengue cases to national surveillance systems, which prevents an accurate estimation of the disease burden of dengue in endemic countries [ 8 ]. Despite its limited effectiveness and high cost, vector control is the mainstay of dengue control and outbreak response in endemic areas [ 9 – 12 ]. The disease is expected to further expand its geographical range due to favorable conditions provided by rapidly growing high density urban areas along with socio-economic changes [ 13 ], increased worldwide travel and trade [ 14 ], and climate change [ 15 ]. A growing literature shows that dengue imposes an enormous socioeconomic burden on households, health care systems, and governments in endemic countries [ 16 – 18 ], particularly during outbreaks [ 19 – 22 ].
Reported as a public health problem since the 1950s, dengue causes frequent outbreaks in Thailand and is hyperendemic with all four distinct serotypes of the dengue virus in circulation for more than five decades [ 23 ]. Although dengue has traditionally affected children, there has been a shift in the mean age of dengue cases towards older age groups in Thailand and other dengue hyperendemic countries in Southeast Asia [ 24 – 28 ]. Most dengue cases now occur in individuals aged 5–24 years [ 29 ], which account for one third of the total population in Thailand, and the disease is more common in adolescents and young adults [ 30 ]. The incidence of DHF varies widely from year to year, exhibiting as much as a tenfold difference between years [ 26 ]. During the period 2000–2011, the incidence of DHF was higher in children aged 5–14 years than those aged 15 years or older [ 29 ]. While the case fatality rate of dengue has been declining steadily over the past decade, the highest rates are seen in children aged 0–4 years [ 29 ].
Frequent and severe illness can cause considerable social and economic disruption to households by requiring one or multiple visits to health care providers and hospitalization. Dengue illness often leads to school and work absenteeism, medical and non-medical expenditures, and foregone income. Illness related costs incurred by patients and household members constitute a severe economic burden for households, particularly in developing country settings. To accurately assess the overall economic burden of dengue, cost-of-illness data at the household level are, therefore, essential. Within the context of a European Union funded research project on dengue [ 31 ], we conducted a prospective hospital-based cost-of-illness study to assess the cost and impact of hospitalized dengue cases on households in a highly endemic area in eastern Thailand. Previous cost-of-illness studies in Thailand focused primarily on pediatric dengue patients (aged under 15 years) [ 32 – 36 ]. In view of the shift in the age distribution of dengue cases, we expanded the focus to cover adult dengue patients (aged 15 years and above).
Material and methods
Study design
We conducted a prospective, hospital-based cost-of-illness study in Chacheongsao province in eastern Thailand. Chachoengsao is a highly endemic area for dengue with a population of 700,902 in 2015 [ 37 ] and a surface area of 5,351 km 2 . The province is divided into 11 districts with 93 sub-districts, and has one province-level and nine district-level hospitals in total. Historically an agriculture-based province with rice paddies, fruit plantations, and livestock, it has become industrialized in recent years, transitioning from rural to semi-rural and semi-rural to semi-urban.
One provincial-level and two district-level public hospitals participated in the study. The study population included hospitalized pediatric (aged under 15 years) and adult (aged 15 years and above) patients who were clinically diagnosed with dengue. All patients or their legal guardians were invited to participate in the study and asked to sign an informed consent form. Patients who did not give consent were excluded from the study. The recruitment period was from March to September in 2015 and overlapped with the peak season of dengue illness.
Research procedures
We adapted a patient questionnaire, which was successfully used in previous cost-of-illness studies in several dengue endemic countries [ 17 ]. It was translated into and back-translated from Thai by two researchers who were fluent in both languages, and the discrepancies were resolved through discussion. The questionnaire was piloted on 10 patients and validated before its administration. It collected information on the demographic and socio-economic characteristics of the patients and other household members, the characteristics of dengue illness episodes, work and school absenteeism, health care service utilization, household health care spending and coping strategies, care provided to the patients by household members, and household income loss due to the dengue illness episode.
Data collection and management
Patients or their legal guardians were interviewed in-person after recovery from the illness by six experienced public health officers, and each officer received a half-day one-on-one training about the study protocol. The interviews took place at the hospital, the patient’s workplace or home, or any other place convenient for the patient. Each interview lasted about 30–45 minutes. Patients or their legal guardians were compensated for their time with a stipend in the amount of 200 Thai Baht (THB) (US$5.8). We followed up with 5–10 patients by phone because there was missing data or inconsistent information in the completed questionnaires. Data were entered into a Microsoft Access Database (2015, Microsoft Corp, Redmond, WA) and analyzed using SAS 9.4 (SAS Institute Inc., Cary, NC).
Analytical framework
The unit of analysis was a dengue case, defined as a documented acute febrile illness with a clinical diagnosis of dengue at the time of hospital discharge. This study examined all hospitalized dengue cases regardless of disease severity.
Household expenditures on dengue include direct medical and non-medical costs and indirect costs incurred by the household during the dengue illness episode. Direct medical costs comprised all household out-of-pocket payments for medical services received by the patient prior to and during hospitalization. Direct non-medical costs included out-of-pocket payments for transportation, food and lodging for the patient and accompanying household members while seeking and receiving medical care for the illness episode. Indirect costs incurred by the household were assessed as the sum of lost paid work by the patient and other household members aged 15 years and above while caring for the patient during the dengue illness episode. We valued lost paid work as the higher of the reported income loss or the estimated income loss calculated conservatively as the product of the minimum daily wage (300 THB [ 38 ]) in Thailand times the number of reported workdays lost by the patient or other household members. The value of time forgone from leisure or other non-market activities was not included in the calculation of indirect costs. If reimbursements were paid to the household by health and/or income protection insurance, the amount reported was subtracted from the sum of the direct and indirect costs for that particular household to arrive at a total cost per case. We also reported on the number of school days lost by the patient and other household members due to dengue illness, as well as the total number of days household members cared for the patient during the illness episode. All costs were presented as mean (± standard deviation, SD) and expressed in 2015 US$ based on 34.2 THB to 1US$ currency exchange rate [ 39 ].
Ethics statement
The protocol for this study was reviewed and approved by the Ethical Review Boards of Mahidol University, Heidelberg University, Chachoengsao Provincial Public Health Office, and Buddhasothorn Hospital. Signed informed consent was obtained from all patients or their legal guardians. Participant information, such as gender, age, clinical diagnostic status and contact information, was obtained from the hospital records. All the data collected through the cost of illness questionnaire and the hospital records were analyzed anonymously.
Results
Characteristics of the study population and dengue illness episodes
A total of 570 hospitalized patients, who were clinically diagnosed with dengue, were eligible to participate in the study. Of these hospitalized patients, 224 were recruited into the study. The general characteristics of hospitalized dengue patients and dengue illness episodes are summarized in Tables 1 and 2 . Overall, 48% were female, and 24% were aged under 15 years of age. The age distribution of the study population is presented in Fig 1 . The mean household size was 4.4 persons (SD 1.9). Among adult hospitalized patients, 18% had primary school education or less, 36% secondary school education, and 46% vocational/high school/college education or above. Of the 224 hospitalized patients, 168 (75%) and 56 (25%) had a clinical diagnosis of Dengue Fever (DF) and Dengue Hemorrhagic Fever (DHF), respectively, at the time of hospital discharge. About 73% of DF patients and 86% of DHF patients were aged 15 years or above. There were no deaths in this cohort of hospitalized dengue patients.
10.1371/journal.pntd.0005961.g001
Fig 1
The age distribution of hospitalized dengue patients, Chacheongsao province, Thailand, 2015.
10.1371/journal.pntd.0005961.t001
Table 1 General characteristics of the hospitalized dengue patients, Chacheongsao province, Thailand, 2015.
Sex, n (%)
Male
117 (52)
Female
107 (48)
Age distribution, n (%)
<15 years
53 (24)
≥15 years
171 (76)
Highest level of education among adult patients, %
Primary school or less
18
Secondary school
34
Vocational school or more
45
Missing
3
Highest level of education in household with pediatric patients, %
Primary school or less
27
Secondary school
18
Vocational/high school, college or more
49
Missing
6
Patients currently studying, n (%)
53 patients < 15 years of age
48 (91)
171 patients ≥15 years of age
43 (25)
Patients currently working, n (%)
53 patients <15 years of age
2 (4)
171 patients ≥15 year of age
97 (57)
Household members currently studying, n (%)
602 household members
106 (17)
Household members currently working, n (%)
602 household members
312 (52)
10.1371/journal.pntd.0005961.t002
Table 2 General characteristics of the hospitalized dengue cases by patient age group and disease severity, Chacheongsao province, Thailand, 2015.
Patient age group
<15 years
≥15 years
All
Dengue diagnosis at hospital discharge, n (%)
DF
45 (85)
123 (72)
168 (75)
DHF
8 (15)
48 (28)
56 (25)
All
53 (100)
171 (100)
224 (100)
Duration of illness (days), mean±SD
DF
8.0±2.1
9.5±2.5
9.0±2.5
DHF
10.0±2.9
10.5±4.8
10.5±4.6
All
8.3±2.3
9.9±3.6
9.5±3.4
Patient feeling bad or very bad (days), mean±SD
DF
4.3±2.6
3.5±2.1
3.7±2.2
DHF
2.5±0.8
4.5±2.2
4.3±2.2
All
4.0±2.5
3.8±2.2
3.8±2.2
Time between onset of illness and seeking first treatment, n (%)
<24 hours
25 (48)
61 (37)
86 (39)
24–48 hours
8 (15)
47 (28)
55 (25)
>48 hours
19 (37)
59 (35)
78 (36)
Duration of hospital stay (nights), mean ± SD
DF
4.0±2.6
3.7± 1.5
3.8±1.9
DHF
4.6±1.3
4.2±2.1
4.3±2.0
All
4.1±2.5
3.8±1.7
3.9±1.9
Overall, hospitalized dengue patients reported 9.5 days (SD 3.4) of illness, including 3.8 days (SD 2.2) during which the patient felt bad or very bad. The mean duration of illness was 8.0 days (SD 2.1) for pediatric DF patients and 10.0 days (SD 2.9) for pediatric DHF patients, including 4.3 days (SD 2.6) and 2.5 days (SD 0.8) during which the patient felt bad or very bad, respectively. Adult DF and DHF patients, respectively, reported 9.5 days (SD 2.5) and 10.5 days (SD 4.8) of illness, including 3.5 days (SD 2.1) and 4.5 days (SD 2.2) during which the patient felt bad or very bad.
Utilization of health care services
Forty-seven percent of caretakers reported seeking care for their children within 24 hours of onset of illness, 15% reported seeking care one to two days after onset of illness, and 36% waited more than two days. Thirty-six percent adult patients sought care within 24 hours after onset of illness, 27% sought care one to two days after onset of illness, and the remaining 35% waited more than two days. Table 3 presents the type of health facility visited and the type of health provider consulted by hospitalized dengue patients during their illness episode. Fifty-four percent of the patients sought care at a hospital first and got hospitalized during their first visit, followed by 15% visiting a doctor’s office and 14% a pharmacy. Sixty-eight percent of the first visits occurred in a public health facility. About 51% and 17% of the patients reported a second and a third visit, respectively, where 64% and 84% of these visits resulted in hospitalization, and 82% and 97% of them occurred in public health facilities. Two patients had multiple hospitalizations during their illness episode.
10.1371/journal.pntd.0005961.t003
Table 3 Type of health facilities visited and health providers consulted by hospitalized dengue patients, Chacheongsao province, Thailand, 2015.
First visit
Second visit
Third visit
Fourth visit
Fifth visit
Type of health facility, n (%)
Hospital
120 (54)
73 (64)
32 (84)
2 (50)
2 (100)
Emergency care
6 (3)
7 (6)
1 (3)
0 (0)
0 (0)
Outpatient department at a hospital
13 (6)
13 (11)
4 (10)
2 (50)
0 (0)
Health center
20 (9)
3 (3)
0 (0)
0 (0)
0 (0)
Doctor’s office
34 (15)
13 (11)
0 (0)
0 (0)
0 (0)
Laboratory
0 (0)
1 (1)
1 (3)
0 (0)
0 (0)
Pharmacy
31 (14)
4 (4)
0 (0)
0 (0)
0 (0)
Type of health provider, n (%)
Public provider
152 (68)
93 (82)
37 (97)
4 (100)
2 (100)
Private provider
68 (30)
20 (18)
1 (3)
0 (0)
0 (0)
Don’t know
4 (2)
1 (0)
0 (0)
0 (0)
0 (0)
Dengue patients spent, on average, 3.9 (SD 1.9) nights in the hospital. While the mean number of hospital nights for pediatric DF and DHF patients was 4.0 (SD 2.6) and 4.6 (SD 1.3), respectively, adult DF and DHF patients spent, on average, 3.7 (SD 1.5) and 4.2 (SD 2.1) nights in the hospital, respectively. None of the hospitalized dengue patients reported receiving care in the intensive care unit.
Dengue illness affected school attendance and productive activities of the patients and other household members. Table 4 presents the mean number of school days missed and work days lost. Of the 91 hospitalized dengue patients who were studying at the time of illness, 79 reported missing school with an average of 6.8 (SD 4.0) days. Of the 53 pediatric dengue patients, 48 were in school at the time of illness, and 45 missed an average of 6.5 (SD 3.8) days of school. The mean number of school days missed was 6.6 (SD 3.9) and 6.1 (SD 3.1) days for pediatric DF and DHF patients, respectively. Of the 171 hospitalized adult patients, 43 were in school at the time of illness, and 34 reported missing school with an average of 7.2 (SD 4.2) days. Of the 99 hospitalized dengue patients who were working for pay at the time of illness, 97 were adult patients, and 94 lost an average of 6.9 (SD 3.5) days of work due to the illness episode. The mean number of work days lost for adult DF and DHF patients was 6.6 (SD 3.6) and 7.6 (SD 3.1) days, respectively. The burden of a hospitalized dengue case on household members was also considerable. Of the 602 household members, 52% and 17% reported to be working and attending school, respectively. On average, household members missed 1.2 (SD 2.9) days of school and lost 4.1 (SD 3.9) days of work. The mean total number of days cared for the patient during the illness episode was 7.2 (SD 4.9) per household.
10.1371/journal.pntd.0005961.t004
Table 4 Impact of a hospitalized dengue case on patients’ school attendance and productive activities, Chacheongsao province, Thailand, 2015.
Pediatric patients (<15 years)
Adult patients (≥15 years)
All
Days of school missed, mean±SD
DF
6.6±3.9
5.9±2.7
6.3±3.5
DHF
6.1±3.1
10.3±5.7
8.6±5.1
All
6.5±3.8
7.2±4.2
6.8±4.0
Days of work lost, mean±SD
DF
7.0±2.8
6.6±3.6
6.6±3.6
DHF
-
7.6±3.1
7.6±3.1
All
7.0±2.8
6.9±3.5
6.9±3.5
Household costs of a hospitalized dengue case
Table 5 presents the direct, indirect and total costs of hospitalized dengue cases to households by patient age category and disease severity. The mean total household cost of a hospitalized pediatric and adult dengue case was US$155.4 (SD 112.1) and US$186.8 (SD 184.7), including a mean reimbursement of US$7.7 (SD 24.1) and US$20 (SD 138.9), respectively. The direct costs for pediatric and adult patients amounted to US$81.9 (SD 76.5) and US$109.3 (SD 190.4), constituting 52% and 59% of the total household costs while the mean indirect costs were US$81.1 (SD 66.3) and US$97.5 (SD 110.3), respectively. The direct non-medical costs accounted for the majority of the direct costs to households regardless of patient age category, and were US$67.2 (SD 66.4) and US$78.6 (SD 94.7), constituting 82% and 72% of the direct costs, respectively, the rest being the direct medical costs.
10.1371/journal.pntd.0005961.t005
Table 5 Household costs of hospitalized dengue illness by patient age group and disease severity, Chacheongsao province, Thailand, 2015. All costs are reported in 2015 US$.
DF
DHF
All
Pediatric patients (<15 years), mean±SD
Direct cost
80.0±76.3
94.1±82.6
81.9±76.5
Medical cost
15 . 8 ± 25 . 7
8 . 3 ± 10 . 7
14 . 7 ±24.3
Non-medical cost
64 . 2 ± 63 . 8
85 . 9 ± 84 . 4
67 . 2 ±66.4
Indirect cost
81.7±68.7
77.7±53.5
81.1±66.3
Total cost
153.6±115.3
166.3±96.3
155.4±112.1
Adult patients (≥ 15 years), mean±SD
Direct cost
96.5±127.1
141.6±294.6
109.3±190.4
Medical cost
20.0±52.0
57.6±246.8
30 . 7 ± 138 . 8
Non-medical cost
76.5±99.8
84.0±81.3
78 . 6 ± 94 . 7
Indirect cost
84.7±80.1
129.8±160.2
97.5±110.3
Total cost
171.2±167.2
226.1±220.1
186.8±184.7
Overall, the total household cost of a hospitalized dengue case increased with disease severity. The mean total cost of a pediatric DF and DHF case to households was US$153.6 (SD 115.3) and US$166.3 (SD 96.3), respectively. Adult patients reported a mean household cost of US$171.2 (SD 167.2) and US$226.1 (SD 220.1) for a DF and DHF case, respectively. The direct non-medical costs similarly accounted for the majority of the direct costs to households regardless of dengue disease severity. Among adult hospitalized patients, the direct medical costs and the indirect costs were notably higher for DHF cases compared to DF cases.
Overall, 74% of the households reported that the patient received free medical care during their illness episode, and 62% reported that the patient was covered by the Thai Universal Coverage Scheme for health care. A great majority of the households (94%) reported not borrowing money from outside the household or selling or transfering any household assets to finance the dengue illness episode. About 27% reported using household savings, and 41% reported that other household members helped finance the dengue illness episode.
Discussion
This was a prospective cost-of-illness study, which aimed to quantify the direct, indirect and total costs of hospitalized dengue cases to patients and their households in a highly endemic, semi-rural area in eastern Thailand. Overall, wide regional variations in dengue incidence occur annually in Thailand [ 29 ]. During the period from 2014 to 2016, dengue-related morbidity and mortality rates in Eastern region where Chacheongsao province is located were notably higher than the national rates [ 40 ]. In the study year of 2015, dengue morbidity and mortality rates were 328.7 and 0.6 per 100,000 population for the Eastern region, compared to the national rates of 222.6 and 0.23 per 100,000 population, respectively [ 41 ]. While most published cost-of-illness studies on dengue have focused on pediatric patients, this study included adult patients in view of the recent shift in the mean age of dengue cases reported in Thailand and other endemic countries in Southeast Asia.
The average household cost of a hospitalized dengue case in Chachoengsao province was US$153.6–166.3 and US$171.2–226.1 for pediatric and adult patients, respectively. These costs fell within the range of those reported in other studies for Thailand, ranging from US$44 in 2001 [ 42 ] to US$118 in 1994 [ 43 ], and for adult patients from US$138 to US$162 in 1994 (unadjusted costs) [ 43 ]. Overall, the total cost of a hospitalized dengue case to the household was higher for adult patients than pediatric patients. Unsurprisingly, the severity of dengue illness was found to increase the financial burden on households due to more prolonged and complicated treatment and longer illness period.
The direct medical costs constituted only a small portion of the direct costs and the total costs, and were US$14.7 for pediatric patients and US$30.7 for adult patients. This could be explained by the fact that the majority of the study population was covered by the Thai Universal Coverage Scheme for health care and paid THB30 (US$ 0.88) per visit or admission. This co-pay was waived for children and elderly and for households with income less than THB2,800 (US$81.9) per month. The direct non-medical costs increased with the severity of dengue illness for both pediatric and adult patients. Our findings showed that the indirect costs were as significant as the direct costs, constituting about half of the total costs to the households. The indirect costs reported for pediatric patients in this study were similar to those reported by other studies, ranging from US$20 in Khamphaeng Phet in 2001 to US$42 and US$51 for clinically diagnosed and laboratory confirmed dengue, respectively, in Khon Kaen in 2005 (unadjusted costs) [ 44 ].
The average monthly household income in Chacheongsao province was THB27,555 (US$806) in 2015 [ 45 ]. Our findings showed that a hospitalized dengue episode cost approximately 19–23% of the monthly household income, which was lower than what was previously reported for Thailand as 37% [ 42 ]. However, some households had to use additional financial sources, such as household savings (27%), or sought financial assistance from other household members (41%) to finance the dengue illness episode. These coping strategies that deal with direct costs of illness can potentially undermine future income streams and threaten the economic sustainability of households. Our study potentially underestimated the total costs of dengue illness to households because dengue episodes tend to cluster at the household level, affecting multiple household members simultaneously. An earlier study in Thailand found that dengue affected an average of 1.4 family members per household per episode [ 42 ].
There were several limitations to this study. It was conducted in three hospitals in a single province with a focus on hospitalized clinically diagnosed dengue cases in the public health sector. Previous cost of illness studies did not find any significant differences in health seeking behavior and overall costs between laboratory confirmed and clinically diagnosed dengue cases, and laboratory confirmed dengue and non-dengue febrile cases [ 34 , 46 ]. The total cost of a hospitalized dengue case to a household is, however, expected to vary depending on the type of hospital where care is received and whether the hospital is public or private. Similar to other cost of illness studies, this study relied on self-reported costs, which are prone to recall bias. It has been shown that a recall timeframe of one year is appropriate for rarely used health care services, such as hospitalization, and much shorter timeframes are recommended for more frequently used services, such as doctor visits [ 47 ]. To minimize potential recall bias, patients were contacted within one to six weeks of hospital discharge to check up on patient recovery progress and set up an interview date, and were interviewed about within four to ten weeks of recovery. The issue of self-reported costs is particularly important when considering the indirect costs associated with an illness episode. The method of asking patients how many days they could not work has been shown to overestimate the productivity losses from a disease [ 48 ]. This is mainly because patients, particularly in low and middle income countries, would not have been working for all of those days in the absence of a disease. It is also a common coping strategy that other household members fill in for a sick person or for a parent caring for a sick child to sustain the household productivity [ 48 ]. We, therefore, estimated the value of lost paid work conservatively using Thailand’s minimum daily wage when the patient or household member reported number of work days lost rather than income loss. Despite these limitations, our study has the most comprehensive cost-of-illness data for hospitalized dengue cases in Thailand to date.
The costs of inpatient care services absorbed by the government through the Thai Universal Coverage scheme is beyond the scope of this study. We, however, received the hospital bills of 89 patients from the provincial hospital. The hospital bills indicated the length of hospital stay and the service fees for inpatient care provided during hospitalization for each dengue patient. The mean inpatient service fees for DF and DHF cases in the provincial hospital were US$145.7 (SD 78.9) and US$152.1 (SD 107.1) for pediatric patients and US$132.2 (SD 62.5) and US$144.9 (SD 82.3) for adult patients, respectively. The service fees are negotiated fees paid to hospitals by the public health insurance. Therefore, they do not cover the full economic cost of inpatient care provided at hospitals and are not comparable to the direct medical costs of hospitalized dengue cases reported in other costing studies. Still, these service fees are relatively high compared to the estimated direct medical costs of hospitalized DF and DHF cases at secondary care hospitals in Colombo district, Sri Lanka: US$ 51 (SD 1) and US$129 (SD 3) for pediatric patients and US$32 (SD 1) and US$91 (SD 2) for adult patients (2012 US$), respectively [ 22 ].
Conclusions
This study showed that dengue related illness imposes financial hardship on households in Thailand when hospitalization is required. Although direct medical costs were covered for a majority of hospitalized patients by the Thai Universal Coverage Scheme for health care, direct non-medical and indirect costs were of great economic significance to households. These hidden costs of dengue illness are likely to increase given the shift in the mean age and severity of dengue cases in Thailand and other dengue affected countries in the region. This begets the question of whether households can be protected from these hidden costs through innovative policy measures in dengue endemic countries. To fully understand the economic impact of dengue illness on households, it is necessary to collect cost of illness data for both hospitalized and non-hospitalized dengue cases and in both the public and private health sectors. The total cost of a hospitalized dengue case in public facilities accounted for about 19–23% the monthly household income. High household costs of dengue illness strongly justify efforts to improve the coverage of preventive and control measures against dengue. Such cost of illness data are also key to evaluating the cost-effectiveness of these measures, including dengue vaccines.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction A considerable body of evidence shows that geographic distance is a good predictor of the genetic structure of European populations. A southeast-northwest cline, possibly associated with the Pleistocene settlement of the continent and the Neolithic demic diffusion from the Fertile Crescent [ 1 , 2 ] (but see 3 ), has been initially highlighted for classic genetic markers [ 1 ] and later corroborated by the analysis of Y chromosome and autosomal polymorphisms [ 4 , 5 , 6 ]. One exception to this scenario, however, is that no clear evidence of clinal variation has been observed for mitochondrial DNA, which is supposedly a consequence of the higher female compared to male migration associated with the prevalence of patrilocality [ 7 , 8 , 9 ]. Finns, Sardinians, Basques and European Jewish provide important departures from this pattern, a finding which is currently explained by bottlenecks and/or their reduced genetic exchange with other European populations [ 10 , 11 , 12 , 13 , 14 , 15 ]. A potential but yet to be well explored source of diversity in the European genetic landscape is represented by groups that have settled in mountainous environments. In particular, great mountain ranges, such as the Alps, Pyrenees and Carpath, may have not only acted as barriers to gene flow for resident populations, but have possibly, since prehistory, also offered a place for the settlement of small, and sometimes culturally diverse, communities.
The Alps are one of the broadest mountain ranges of Europe, with a longitudinal extension of approximately 1,200 kilometers. They cover eight different countries and over 100 peaks of over 4000 m a.s.l. There is a substantial consensus among archeologists regarding the notion that many alpine areas had already been inhabited in the Paleolithic [ 16 , 17 ], with a more intense peopling starting from the Neolithic [ 18 , 19 ]. However, occupation of the upper valleys remained scattered and small in number until a more systematic process of colonization and demographic expansion began in the late Middle Ages [ 20 ]. Another key passage concerning the demographic history of the Alps is represented by the “break-up of isolates”. In fact, a dramatic decline of endogamy began in the first half of the 20th century due to an increase in individual mobility and the depopulation of the mountain areas thanks to socio-cultural changes linked to industrialization [ 21 , 22 ].
At present, Alpine populations can be considered as a mosaic of groups that are separated by physical and cultural boundaries, whose remarkable cultural diversity is clearly demonstrated by the presence of minorities that speak Franco-Provençals, Occitans, French, German, Ladin, Friulian and Sloven languages [ 23 , 24 ]. From a bio-anthropological point of view, they offer a unique opportunity to study the impact of geographical, demographic and cultural factors on genetic structure [ 25 ]. Such a target requires the simultaneous investigation of distinct linguistic groups and, ideally, the analysis of genetic systems with different modes of evolution and transmission. Unfortunately, the population genetic studies that have been carried out so far are scanty and most of them only focused on a limited number of populations or single groups [ 26 , 27 , 28 , 29 , 30 ].
In this study, we present new high resolution data on Y chromosomal variation in three distinct Alpine ethno-linguistic groups, Italian, Ladin and German. Combined with data on Y chromosome and mitochondrial variation taken from our previous research work and the literature, these results are used to answer four questions: (i) how is genetic diversity patterned in alpine ethno-linguistic groups?; (ii) what micro-evolutionary forces might have shaped their genetic structure?; (iii) how do the observed patterns compare with what has been noticed in other European groups, in particular with well known genetic outliers and other groups settled in great mountain ranges?; (iv) are there factors, other than geography and language, that should be taken into account when studying the genetic structure of European mountain populations?
The Populations under Study
Our study is primarily based on unpublished Y chromosome data (17 Short Tandem Repeats, STRs, and 50 Single Nucleotide Polymorphisms, SNPs) from 610 unrelated individuals belonging to 15 populations from the Eastern Italian Alps (Trentino-Alto Adige, Veneto and Friuli regions; see Table 1 and Figure 1 ).
10.1371/journal.pone.0081704.t001 Table 1
Populations included in the present survey.
Population (region)
Abbreviation
Sample size
Language
Census size *
Adige (Trentino)
ADI
56
Romance (Italian)
166394
Badia (South Tyrol)
BAD
44
Romance (Ladin)
10644 †
Fassa (Trentino)
FAS
47
Romance (Ladin)
9894 †
Fersina (Trentino)
FER
26
Romance (Italian)
2575
Fiemme (Trentino)
FIE
41
Romance (Italian)
18990
Gardena (South Tyrol)
GAR
51
Romance (Ladin)
10198 †
Giudicarie (Trentino)
GIU
51
Romance (Italian)
36282
Lessinia (Veneto)
LES
24
German
13455 §
Luserna (Trentino)
LUS
25
German
286
Non (Trentino)
NON
48
Romance (Italian)
37832
Primiero (Trentino)
PRI
41
Romance (Italian)
9959
Sappada (Veneto)
SAP
38
German
1307
Sauris (Friuli)
SAU
29
German
429
Sole (Trentino)
SOL
65
Romance (Italian)
15235
Timau (Friuli)
TIM
24
German
500
* ISTAT (2011) ( http://demo.istat.it )
† This value refers to Ladin speaking communities only [ 23 ]
§ This value refers to Cimbrian speaking communities only [ 23 ]
10.1371/journal.pone.0081704.g001 Figure 1
Geographic location of the populations under study (see table 1 for population acronyms).
Ten populations belong to the main Romance language [Italians (Adige, Fersina, Fiemme, Giudicarie, Non, Primiero and Sole valleys); Ladins (Fassa, Badia, and Gardena valleys)], five to the German-linguistic isolates [two Cimbri groups (from Luserna and the Lessinia area); the communities of Sappada, Sauris and Timau].
Ladins are thought to be related to pre-Indo-European speaking tribes who probably represent the most ancient settlers of the Alps [ 31 ]. The Dolomitic Ladins are the remnant of a wider group that started settling in a broader territory in 1000 AD. As for the Ladins, the other Romance speaking groups of Italians are thought to be linked to the most ancient peopling of the area [ 31 ]. Finally, the ethno-linguistic Germanic islands of the Eastern Alps are in continuity with nuclei that migrated from Bavaria, Carinthia and Tyrol in the late Middle Ages, a process driven by the landed aristocracy and the monasteries with the objective of a more intensive exploitation of marginal territories [ 20 ].
The dataset was integrated with an extensive search of literature data on unilinear transmitted markers [ 32 ] relative to populations living in the Alps or in other European mountain ranges (Pyrenees) (see Table S1 ).
Material and Methods
Sampling and ethic statements
Buccal swabs were collected in apparently healthy and unrelated donors selected according to the place of birth of the sampled individual and of their parents and grandparents. The procedure and informed consent were reviewed and approved by the “ Comitato Etico per la Sperimentazione con l’Essere Umano ” of the University of Trento (samples from Trentino), “South Tyrolean Ethics Committee” (samples from Alto Adige, POLYS project) and the institutional review board of the Istituto Italiano di Antropologia (samples from Veneto and Friuli). All participants provided written informed consent to participate in this study.
Laboratory analyses
The DNA was extracted using the ‘Nucleic Acid Isolation System’ by the QuickGene-810 instrument following the standard protocols for blood and swab samples (FUJIFILM) or using a modified “salting-out” procedure.
The 17 Y-chromosomal short tandem repeats (STRs) included in the AmpFlSTR Yfiler Amplification Kit (AB Applied Biosystems; DYS19, DYS389I, DYS389II, DYS390, DYS391, DYS392, DYS393, DYS385ab, DYS437, DYS438, DYS439, DYS448, DYS456, DYS458, DYS635 and GATA H4.1) were typed in all samples (with the exclusion of 59 samples from Non and Sole valleys belonging to the R-M269* lineage which had been previously published [ 3 ]). PCR products were analyzed by capillary electrophoresis in an ABI 3100 Genetic analyzer (Applied Biosystem, Foster City, CA). Fifty Y-specific unique-event polymorphisms were examined in hierarchical order (M17, M102, M153, M170, M172, M173, M201, M222, M223, M224, M241, M253, M26, M267, M269, M280, M282, M304, M319, M35, M410, M423, M438, M45, M47, M521, M67, M78, M89, M9, M92, P37.2, S116, S127, S139, S144, S145, S167, S21, S28, S29, SRY2627, V12, V13, V148, V19, V22, V27, V32, V65). Firstly, all samples were tested by one basal multiplex (MY1) following the protocol reported in Onofri et al. [ 33 ] with the addition of UEPs M269, M17, M201, M267, M282 and M304. Afterwards, all the samples derived for the M269 mutation (T>C), M35 (C>G), M170 (A>C), and M172 (A>C) were further analyzed using the specific multiplex for haplogroups R1b*, E*, I* and J2* , respectively ([ 3 , 34 ] and Brisighelli F and Capelli C, personal communication). The protocol includes first PCR amplification reactions by using the Qiagen Multiplex PCR kit with the conditions specified by the producer [ 35 ] and subsequent purification by enzymatic method (ExoSAP; [ 36 ]). The purified products were then used for a single-base extension reactions by the SNAPShot method (Applied Biosystems Carlsbad, CA).
Phylogenetic relationships between markers and nomenclature follow the International Society of Genetic Genealogy (April 2013, Ver 8.43), ( http://www.isogg.org/tree/ ). The population data obtained were submitted to the Anthro-Digit database ( http://www.isita-org.com/Anthro-Digit/data.htm ).
Statistical analysis
Unless otherwise stated, statistical analyses were performed using 15 STRs, having excluded the duplicated DYS385 loci. The level of intra-population genetic variation was analyzed through the calculation of haplotype diversity (HD) and the number of different haplotypes (H). Multi-Dimensional Scaling of Fst genetic distances based on Y chromosome STRs (Reynolds’ distances, [ 37 ]) and a Principal Component Analysis plot based on haplogroup frequencies were obtained using SPSS software (release 16.0.1 for windows, SPSS Inc.). We partitioned genetic variance at different hierarchical levels of population subdivision according to language groups (Italian, Ladin and German) by means of a molecular analysis of variance (AMOVA). In this analysis, we also used mitochondrial DNA literature data (HVR1, 333 bp from 16033 to 16365; see Table S2 ) [ 32 ]. All parameters of intra and inter-population genetic diversity were calculated using the Arlequin software (version 3.5.1.2, [ 38 ]).
We used a coalescent based simulation approach in order to evaluate whether the observed values of within-group genetic diversity may be attributed solely to the size of the founding group (see Tofanelli et al. [ 39 ] for a review of simulation methods for uniparental markers). We separated Italians into two sub-groups, western and eastern, according to their different current census size and previous mtDNA evidence [ 40 ]. Adige valley and Cimbrian populations were not considered to be part of the simulations because of the difficulties and uncertainties in modeling their evolutionary history. Based on current historical records, we designed two different topologies, one for the German-speaking island group and one for the two Italian sub-groups and Ladin speaking group. In both topologies (see Figure S1 ) three sub-populations split from a large source population at a certain time (T1) which were identified as Central-Western Europe but which differ in splitting times (32-40 generations for German speaking islands and 90-110 generation for all the other groups). According to Bramanti et al. [ 41 ], effective population sizes for source and sink populations were set as 1/10 of census size. Growth rate for the source population was set at 0.0018 from 1800 to 300 generations ago, and increased to 0.022 from then to the present day [ 42 ]. The growth rate for the sink populations was set as half of the highest value of the source. A symmetrical gene flow between source and sink was allowed (0.005-0.01), while admixture between sink populations was allowed to vary between 0.01-0.02 and 0.02-0.03. We simulated 10K random genealogies for the Y chromosome (15 STRs) using the mutation rate estimates of Ballantyne et al. [ 43 ] assuming a generation time of 25 years. For each scenario, we randomly sampled 50 individuals from each sink population and analyzed within-group diversity for each simulation using Arlequin 3.5 [ 38 ].
Results and Discussion
Patterns of genetic diversity in the linguistic groups of the Italian Alps The Eastern Italian Alps embrace an important portion of the ethno linguistic diversity of the alpine arch, encompassing Romance (including Ladins and Italians) and German speakers. Their genetic characterization highlights a high level of diversity not only among single populations, but also within linguistic groups, a pattern which is likely to be due to a complex interplay of demographic histories and isolation determined by environmental and cultural factors.
The extent of diversity among Alpine populations is shown by the plots based on STR and SNP data ( Figure 2A and 2B ). The spatial relationships among populations differ between the two plots, with the SNP-based patterns probably mirroring more ancient population relationships due to their slower evolutionary rate. However, with both data-type populations under study are well separated and no linguistic structure of genetic diversity is detectable. This latter feature may be appreciated in a quantitative way by an AMOVA performed among linguistic groups, which produced low values of intergroup variation (from 0.007 to 0.020; see Table S3 ).
10.1371/journal.pone.0081704.g002 Figure 2
Plots of the genetic relations among populations under study. (a) Multi-Dimensional scaling plot of Fst genetic distances (15 STRs; stress value=0.128); (b) Principal Component Analysis plot based on haplogroup frequencies. First component (x axis) and second component (y axis) explain 16.96% and 13.95% of total variance, respectively. Acronyms are given in Table 1 .
To gain further insights into the genetic diversity occurring within each linguistic group, we went one step further by focusing on their genetic structure. The Italian speaking group was found to be the most genetically homogeneous. Within group variation (0.04, p<0.05) is lower than in other Alpine groups and geographically distant European populations, but higher than observed among Northern Italian populations ( Table S4-S5 ). Furthermore, they show high haplotype diversity values, with the highest observed in the Adige valley (0.997 ± 0.004 ; see Table S6 ). R1b S28*, a haplogroup found at high frequencies in most Alpine groups, is the most frequent in all populations (from 17 to 45%), the only exception being the Primiero, where G-M201 prevails (~49%, see Table S7 ). This pattern may be explained in two, not mutually exclusive, ways. Italian speaking populations have, since historical times constituted the most numerous ethno-linguistic group in the Eastern Alps, and they did not suffer from any historically documented bottleneck. Their present census values are comparable or higher than those of the other two groups under study (see Table 1 ). Furthermore, having settled in zones which are characterized by wider valleys, lower altitudes (from 200 to 1022 m a.s.l.) and more accessible mountain passes, they have probably been less geographically isolated than other groups (e.g. Ladins). Finally, the Adige river has provided a supplementary communication route, favoring population movements and interactions [ 44 ].
The genetic differentiation among Ladin valleys noticed in the plots is supported by other analyses of STR haplotype distribution. Their intra-group variation (0.075, p<0.05) is similar to what has been found in geographically distant European populations (6 populations, distances ranging from 366 to 2520 km, 0.074 p<0.05; see Table S4-S5 ) but much higher than what has been found in a set of Northern Italian populations (6 populations, distances ranging from 57 to 396 km, 0.006 p>0.05; see Table S4-S5 ) . Another likely effect of genetic drift may be seen in the intra-population diversity values (HD), which are lower than those observed in Italian speaking communities and in most European populations ( Table S6 ). This is particularly evident for the communities from the Gardena and Badia valleys (South-Tyrol), which, correspondingly, depart more evidently from the main central group in the genetic distance plot ( Figure 2A ). Signatures of intra-group diversity are also provided by a phylogeographic approach. A further signal of the high within-group diversity is given by the finding that the prevalent haplogroup in the Fassa and Badia communities (S28*-R1b*) and Gardena valley (S-145 R1b*; Table S7 ) do not coincide. These two lineages of the main S116-R1b* haplogroup show a quite distinct continental distribution, with the former reaching its highest frequencies in south-central Europe (with frequencies peaks in France, northern Italy and the Alps), and S145-R1b found mainly in the north-Atlantic Europe [ 3 , 45 ]. On the whole, our results support the definition of Ladins as “small genetically isolated populations (subject to strong genetic drift), having a predominantly European ancestry” [ 27 ]. However, it should be noted that the inclusion of a third population (Fassa Valley) and the higher resolution of Y chromosome genotyping make our inferences more robust. The genetic signatures we observed may be an echo of the processes of fragmentation and/or assimilation of Ladin communities, first by Latin groups starting from the 15th century b.C, and then by German-speaking people (Gardena and Badia valleys) from the end of the 4 th century, and the consequent reduction in their settlement area and demographic size [ 46 ]. Moreover, the considerable altitude of the Ladin valleys (from 1120 to 1345 m a.s.l.) might have further increased a reciprocal isolation among fragmented Ladin communities [ 31 ].
The German speaking populations show the most marked signatures of genetic drift. As predicted by the outlying positions of Sappada, Timau and Luserna in the plot of genetic distances, the intra-group variation is very high (0.240, p<0.05), around two times higher than that found for geographically distant European populations. Moreover, the haplotype diversity values in these populations are the lowest of the the dataset, with the exception of Lessinia (see Table S6 ). Different haplogroups prevail in Sappada (E1b-V13 63%) and Timau (R1a-M17 56%), and different R1b subhaplogroups in Sauris (S139 34%), Lessinia (S116 17%) and Luserna (M269 84%). The considerable differentiation among German-speaking populations may be also seen as a consequence of their demographic history. In fact, they are in continuity with small founding groups [ 47 ] which settled in the present day location in Medieval times. Furthermore, as we have recently proposed [ 30 ], a relative reciprocal isolation could have occurred even among the linguistically closely related communities of Sappada, Timau, and Sauris as a result of “local ethnicity”. In this condition, the members of each community tend to identify their ancestry with their own village rather than considering themselves as part of the same ethnic group, similarly to what occurs in other alpine regions [ 48 ].
The genetic differentiation between the two Cimbri populations of Luserna and Lessinia deserves further discussion. Both these communities derive from Bavarian populations that colonized a vast territory of the Eastern Italian Alps starting from 1053 AD (Veneto; [ 49 ]) to 1216 AD (Trentino; [ 44 ]). Luserna is genetically very distant from all the other Alpine populations (average Fst=0.328; see Table S6 ) and shows a strikingly low intra-population diversity (0.483±0.119). Paternal lineages are represented mostly by the R1b-M269* (frequency of 84%), with six different STR haplotypes associated with only one founder surname. Lessinia shows different, if not opposite, features. The average genetic distances from other populations (Fst=0.097; see Table S6 ) is less than one third compared to Luserna, while HD is close to the highest values of our dataset (0.978±0.019; Table S6 ). The prevalent haplogroup, R1b-M269*, accounts for only one third of the total, the rest represented by different lineages (G-M201, I1-M253, M410-J2A and K-M9), which are associated with twenty-three different surnames. The demographic history of the Luserna and Lessinia communities may help explain their differentiation. Luserna was founded by few families which moved from Lavarone, the first known Cimbrian settlement in the territory of Trentino [ 44 ]. This could have led to a strong founder effect in this community, a hypothesis supported by a previous study of mtDNA polymorphisms [ 40 ]. Moreover, Luserna is located on a high plateau (1,333 m a.s.l.) and is quite isolated from the surrounding areas. By contrast, Lessinia, a more extensive area with reliefs of low altitude (Giazza, 758 m a.s.l.), and has been colonized since the XIII century AD through several migration waves of small groups of settlers for more than one century. From the XV century AD, this community opened to, and probably admixed with, Italian neighboring groups [ 49 ].
On the whole, our genetic characterization indicates three main genetic patterns. Italian speaking populations show slightly higher level of within-group diversity than observed among distant European populations. The strongest signals of departure from the European genetic background can be seen among German speaking populations, while the intra-group and intra-population diversity level of Ladins fall between the former two groups. These signals seem to reflect the different demographic history of the three groups and their genetic isolation due to the mountainous environment (for all groups) and use of different languages from their neighbors (Ladins and German speakers). Nonetheless, the fact that the Y chromosome is a single locus transmitted by father to sons means that our inference needs further support from other genetic systems with a diverse mode of inheritance. Therefore, we thought it would be useful to repeat the analysis of intra and inter-population diversity with maternally transmitted mitochondrial DNA polymorphisms (hypervariable region 1) [ 27 , 30 , 40 ]. Despite some minor differences regarding Sauris and Ladins from the Gardena valley (both show an outlying position in the mtDNA MDS plot and the latter a lower rank for haplotype diversity), mtDNA and Y chromosome patterns substantially match (see Figures S2 - S4 and Table S3 ).
As discussed above, the intensity of the genetic signals observed in the Alpine linguistic groups seems to comply with what is to be expected for isolated population groups characterized by a different demographic profile. Therefore, a cause effect relationship between these two conditions and the different patterns of genetic diversity is worth taking into consideration. However, it did not escape our attention that such intensity seems to be inversely correlated with the supposed size of the founding groups, reflecting present census values (see Table S8 ). We then decided to test the alternative hypothesis that our observations could be the result of differences in the long–term effective size among groups, without any substantial effect of genetic isolation. To this purpose, we carried out coalescent simulations for all of our linguistic groups, with comparable levels of gene flow to those expected for non isolated groups. The distributions obtained ( Figure 3 ) are incompatible (Ladins and Italians) or only marginally compatible (German speakers) with the observed Fst values. A scenario combining the effects of founding group size and continued genetic isolation seems, therefore, to provide the best explanation for the observed level of within-group differentiation detected in both geographic and geographic/linguistic isolates.
10.1371/journal.pone.0081704.g003 Figure 3
Posterior densities of Fst genetic distances for the micro-evolutionary scenarios. Distributions of Fst values obtained by coalescent simulations (see Materials and Methods), with vertical lines representing observed values of within group diversity: (A) German speaking islands; (B) Italians (Non, Sole and Giudicarie valleys); (C) Ladins (continuous) and Italians (dashed; Fiemme, Fersina and Primiero valleys).
The Alpine linguistic groups in the European genetic background
The genetic distinctiveness of Alpine populations can be better appreciated contextualizing our results into the body of knowledge regarding European populations. A first comparison is to be made with open populations, to see whether group under study actually depart from the continental genetic structure. As shown by Roewer et al. [ 50 ], the distribution of Y chromosome variation at the continental level complies with an isolation by distance model. By contrast, the historical stratification and complexity of the peopling processes occurred in the Eastern Alps does not predict any simple relation between genetic structure and geographic distances. Accordingly, the correlation between geographic and Y chromosomal genetic distances is statistically insignificant (Spearman’s rho correlation value, R 2 = 0.61, p=0.99). Coherently, the alpine populations are widely dispersed in the MDS plot of Y-STR genetic distances despite their geographic proximity ( Figure 4 ), with Sappada, Timau and Luserna behaving as outliers. The histogram plot ( Figure 5A ) based on 15 Y-STR loci highlights the substantial departure of Alpine populations from open continental groups (Austria, Croatia, Italy, Poland, Portugal, Serbia and Spain; see Table S4 ). In fact, average and median value of genetic distances between Alpine and open populations (0.095; 0.078) are substantially greater than between the latter (0.061; 0.061).
10.1371/journal.pone.0081704.g004 Figure 4
Multi-Dimensional Scaling plot of Fst genetic distances between Alpine and European populations. Plot based on 15 Y chromosome STRs (stress value=0.141). Population acronyms are given in Table 1 and Table S4 .
10.1371/journal.pone.0081704.g005 Figure 5
Distribution of Fst genetic distances among Alpine, Sardinian, Basque, Finn and European populations. Frame A shows the Fst genetic distances distributions among European populations and between European and Alpine populations (see Table S4 ; Italy is represented only by La Spezia). Frame B shows the Fst genetic distances between some European outliers (Sardinians, Basques and Finns), Alpine groups and European open populations.
Using the same approach ( Figure 5B ), we observed that the genetic differentiation of Ladins and German speakers from Europeans is comparable or even greater to that observed for well known continental outliers (see Table S4 ). In fact, the average value of Ladins (0.092) is higher than Sardinians (0.078), whereas their median Fst is slightly lower (0.075 vs 0.088). The signal is even stronger for the German speakers, whose average (0.144) exceeds that of Basques (0.121), whereas the two median values are rather close (0.111 vs 0.121) and 14.3% of Fst is above the upper bound of the range of genetic distances between Europeans and Basques. However, all these values are lower than those obtained for Finns (average 0.209; median 0.208) who are known to have undergone severe bottlenecks and further local episodes of drift [ 13 ].
As the final step of our study, we further extended our dataset by including other populations that have settled in great mountain range systems, from the Pyrenees (5) [ 51 ] and from South Tyrol (3) [ 27 ]. The results of the AMOVA ( Table 2 and Table S9 ) show that Y chromosome intra-group variation within human groups that have settled in mountainous environments is relatively high and statistically significant, the South Tyroleans being the only exception (see below). Not unexpectedly, this is in sharp contrast with the low and insignificant diversity observed among open populations settled on plains at comparable geographic distances (-0.003, p=0.555; see Table S9 ). Focusing on mountain populations, it turns out that Alpine groups host the greatest Y chromosome among-population diversity. Interestingly, this does not hold only for German speakers and Ladins, who are the only groups subject to both geographic and linguistic isolation, but even for Italians, who show the apparently weakest signals of genetic drift.
10.1371/journal.pone.0081704.t002 Table 2
Within group diversity among population groups under study based on 5 Y chromosome STRs (DYS19, 390, 391, 392 and 393) and the hypervariable region of mtDNA (from 16033 to 16365).
Y chromosome
mtDNA
German speakers
0.315 (0.184 - 0.380)
0.077 (0.057 - 0.091)
Italian speakers
0.053 (0.023 - 0.066)
0.008 (0.006 - 0.010)
Ladin speakers
0.077
0.035
North-Eastern Italy *
-0.003
0.001
Pyreneans
0.018 (0.002 - 0.031)
n.a.
South Tyroleans
0.004 (-0.006 - 0.011)
0.030 (0.014 - 0.046)
Values in brackets refer to minimum and maximum values obtained by jacknife procedure (see Table S9 ). Statistically insignificant values are in bold.
* This group is composed by three geographically close plain populations (Brescia, Treviso, Vicenza; see Table S4 ).
South Tyroleans provide an exception to the high and statistically significant Y-chromosome intra-group diversity of Alpine populations. A possible explanation for this finding comes from their particular social structure. In fact, since at least the early fourteenth century South Tyroleans have mostly complied with an inheritance and succession system known as Geschlossener Hof (“closed holding”), which entails an impartible transfer of the farm [ 52 , 53 ]. This system typically prescribes that only one son – generally the first born – takes over the economic unit consisting of the farmstead and the attached lands and succeeds into the position of a peasant house-father, while the other sons have the option to remain in the family farm as employees or to receive an economic compensation and relocate elsewhere [ 54 , 55 , 56 , 57 ]. Therefore, this practice may favor male dispersal, increasing the probability for sons other than the first born to marry far from the original community. Conversely, female mobility is less socially favored than in patrilocal groups. In the long term and under regimes of prevalent male mobility within the original groups, the Geschlossener Hof may lead to a pattern which is opposite to what would be expected for patrilocal groups. This is, in fact, the case of Tyrolean populations, who show only statistically significant intra-group variation for mtDNA polymorphisms (0.004, p=0.220 and 0.030 p=0 for Y chromosome and mtDNA respectively). Therefore, the Geschlossener Hof could have shaped intra-group variation of paternal lineages in the opposite way to “local ethnicity” [ 30 ]. If this was supported by further evidence, it would provide an example of divergent effects of socio-cultural factors on genetic diversity in populations who are closely related from a historical, linguistic and environmental point of view.
In conclusion, the comparison between Y chromosomal and mitochondrial patterns of variation suggests that not only geographic factors and linguistic diversity, but also socially induced sex biased gene flow should be taken into account when studying the genetic structure of Alpine populations. We believe this is an important avenue for any future research work which aims to shed light on the yet to be explored complexity of the genetic structure of European populations.
Supporting Information
Figure S1
Topology used for the simulations of evolutionary scenarios.
(TIF)
Figure S2
Multi-dimensional scaling plot of Fst genetic distances among Alpine populations based on mtDNA HVR-I sequences (stress value=0.153). Acronyms are given in Table 1 . (TIF)
Figure S3
Haplotype diversity of Alpine populations: (a) mitochondrial DNA values based on HVR-I region; (b) Y chromosome values based on 15 STRs (acronyms as in Table 1 ).
(TIF)
Figure S4
Analysis of molecular variance (AMOVA) within groups under study based on mtDNA sequences (hypervariable region 1) and 15 Y chromosome STRs.
(TIF)
File S1
17-loci Y-STR haplotypes and haplogroups of the populations under study. (XLS)
Table S1
Literature data (Y chromosome 5 STRs) on populations settled in mountain ranges.
(DOC)
Table S2
Literature data on mtDNA (sequences of the hypervariable region 1) on Alpine populations.
(DOC)
Table S3
Analysis of molecular variance (AMOVA) among linguistic groups (Fst values below the diagonal, p-values above the diagonal).
(DOC)
Table S4
Literature data (Y chromosome 15 STRs) on European and Northern Italian open populations.
(DOC)
Table S5
Analysis of the molecular variance (AMOVA) within European and Northern Italy open populations groups based on 15 Y chromosome STRs (acronyms as in Table S4 ).
(DOC)
Table S6
Y chromosome (15 STRs) genetic diversity in 15 Alpine populations.
(DOC)
Table S7
Haplogroup frequency distribution in populations under study (acronyms as in Table 1 ).
(DOC)
Table S8
Mean census size and analysis of molecular variance (AMOVA) within Alpine linguistic groups under study based on 15 Y chromosome STRs (acronyms as in Table 1 ).
(DOC)
Table S9
Analysis of molecular variance (AMOVA) within groups under study based on 5 Y chromosome STRs, including results of jacknife procedure (acronyms as in Table 1 , Table S1 and S4 ). (DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Staphylococcus aureus nasal carriage increases a patient’s risk for developing a health care-associated infection with this micro organism, at least after cardiac surgery, orthopedic surgery and in peritoneal dialysis. [1] , [2] , [3] , [4] , [5] Preoperative screening for nasal carriage and subsequent treatment of carriers with mupirocin and chlorhexidine reduces the risk for the development of hospital-acquired S. aureus infections by 79% for deep-seated infections and 55% for superficial infections. [6] Consequently, the mean duration of hospital stay is reduced in treated carriers by approximately 2 days. A cost benefit analysis shows that the strategy is cost-effective and saves lives. [7] .
In vascular surgery little is known about the relation between nasal carriage of S. aureus and surgical site infections (SSI).
For this reason we conducted a prospective analysis of S. aureus nasal carriage in patients undergoing vascular surgery and the occurrence of surgical site infections.
Methods
A prospective cohort study was performed on all patients who underwent elective vascular surgery between January 1 st 2010 and December 31th 2010 in the Amphia Hospital, Breda, The Netherlands.
Operations included were; central reconstructions for both abdominal aortic aneurysms and occlusive disease (endovascular (EVAR) and open), peripheral bypass procedures (autologous and PTFE), endarterectomies of the femoral and carotid artery, embolectomies and Artero-venous access procedures.
Patients were screened on the day that they were admitted to the vascular surgery department of the Amphia hospital in Breda. Screening was performed using a dry, sterile swab, which was rotated four times in each nostril. The swab was placed in saline and centrifuged. Part of the sample was processed for polymerase chain reaction (PCR) on the presence of S. aureus , and part was inoculated onto a blood agar plate, to allow nasal and infecting strains to be compared in case a surgical site infection did occur.
The GeneXpert MRSA/SA Assay (Cepheid, Sunnyvale, CA) is a real-time PCR-based method, which identifies S. aureus and also can differentiate whether a S. aureus is a Methicillin-susceptible (MSSA) or Methicillin-resistant (MRSA). [8] , [9] , [10] , [11] .
Patients were followed prospectively for the development of Surgical site infections (SSI) which were defined according to the criteria of the Centers for Disease Control. [12] .
10.1371/journal.pone.0038127.t001 Table 1
Baseline and surgical characteristics and surgical site infections caused by S. aureus .
Characteristics
N
224
Sex, Male/Female (%)
171/53 (76.3/23.7)
Age mean (SD)
70 (10.1)
Type of Surgery N/(%)
N° of SSI’s (%)
N° of S. aureus SSI (%)
Aortic open repair
52 (23)
6 (11)
3 (6)
Aortic endovascular repair
38 (17)
1 (3)
1 (3)
Femoral endarterectomy
34 (15)
3 (9)
3 (9)
Peripheral bypass surgery
Autologous bypass
42 (66)
2 (5)
1 (2)
PTFE bypass
22 (34)
4 (18)
1 (5)
AV access surgery
14 (6)
0 (0)
0 (0)
Peripheral embolectomy
4 (2)
1 (25)
1 (25)
Carotid endarterectomy
18 (8)
0 (0)
0 (0)
Total
224 (100)
17 (8)
10 (4)
The main criteria are the presence of: redness, heat, swelling or pain around the wound within 30 days after the initial procedure, and the presence of a positive culture, drainage of the wound, or pus after a diagnostic puncture. When prosthetic material had been used the follow up was extended up to one year. Infections were differentiated between superficial, deep seated and organ based infections.
Screening was performed as part of the infection control strategy of the Amphia hospital using non-invasive sampling. Approval of the medical ethical committee and informed consent were not applicable.
10.1371/journal.pone.0038127.t002 Table 2
Relation between S. aureus carriage and surgical site infections caused by S. aureus .
Surgery type
S. aureus SSI-rate (%)
RR
95% CI
P-value *
Central reconstructions (n = 90)
Non-carriers (n = 65)
1 (1,5)
Carriers (n = 20)
3 (15)
9.8
1.1–88.6
0.039
Inconclusive (n = 5) **
0
Other procedures (n = 134)
Non-carriers (n = 94)
5 (5)
Carriers (n = 35)
1 (3)
0.5
0.1–4.4
0.48
Inconclusive (n = 5) **
0
*
Fisher’s exact test;
**
Inconclusive screening results were not used for analyzation.
A stratified analysis was performed for patients with central vascular surgery, as we expected a possible difference for the importance of nasal carriage between patients suffering from peripheral arterial occlusive disease (PAOD) and patients suffering from central diluting vascular disease.
Patients suffering from PAOD have a gradient of lower limb ischemia, which ranges from impaired walking distance, due to inappropriate blood flow to the lower limbs (intermittent claudication), to critical limb ischemia. In those patients hypo perfusion of the lower limbs often results in ischemia or even ischemic ulcers. These ulcers may be colonized with pathogens, which may be introduced into the surgical wound. This may alter the role of nasal carriage as there is an additional source of S. aureus in the patient.
10.1371/journal.pone.0038127.t003 Table 3
Relation between S. aureus surgical site infections and risk factors for SSI.
Factor
Univariate
Multivariate
RR
95% CI
P-value
RR
95% CI
P-value ***
Duration of the surgical procedure
NA
0.62 **
1.0
0.98–1.02
0.89
ASA class 1 or 2 (compared to >2)
2.9
0.4–21.5
0.29 *
3.4
0.4–30.2
0.26
BMI
NA
0.99 **
1.0
0.88–1.34
0.89
S. Aureus screening
9.8
1.1–88.6
0.039 *
12.8
1.1–147.9
0.041
*
Fisher exact test;
**
t-test;
***
Chi-square.
Statistical analyses was performed with SPSS software v 19.0 (SPSS Inc., Chicago, IL, USA), the Fisher exact test was used to determine significance.
A multivariate analysis was performed for evaluation of several other known risk factors on the development of SSI’s. Chi-square test was used to determine significance.
A P- value <0,05 was considered significant.
Results
As shown in Table 1 , 224 patients were included. There were a total of 17 SSI’s, 13 of which were superficial, and 4 where deep seated SSI’s. The PCR of nasal swabs showed that 159 (71.0%) were negative for S. aureus , 55 (24.6%) were positive and 10 (4.5%) were inconclusive because of inhibition of the amplification reaction. In 214 patients with conclusive results, there were 16 surgical site infections, 10 of which were caused by S. aureus and 6 by other pathogens.
The incidence of surgical site infections in nasal carriers of S. aureus is 4 out of 55 (7.3%), whereas the incidence in non-carriers is 6 out of 159 (3.8%)(RR = 1.9, 95%CI 0.5–7.5).
A stratified analysis was performed for central reconstruction surgery, peripheral bypass surgery and other procedures as shown in Table 2 . In peripheral bypass surgery 2 S. aureus surgical site infections occurred in patients who did not carry S. aureus (n = 45) and no infections in Patients who carried S. aureus (n = 17) ( P> 0.05).
In the central reconstruction surgery population, there was 1 surgical site infection in patients who did not carry S. aureus (n = 65), this SSI occurred after an aortoilliac bypass procedure because of occlusive disease, and there were 3 infections with S. aureus in Patients who carried S. aureus (n = 20), 1 SSI after an EVAR procedure, 1 SSI after open aneurysm repair and 1 after an aortoilliac bypass procedure because of occlusive disease (RR = 9.8, 95%CI 1.1–88.6, P = 0.039).
A multivariate analysis including 3 other risk factors known from the literature for the development of surgical site infections, did not alter the effect of nasal carriage ( Table 3 ).
Discussion
Our study shows that surgical site infections in vascular surgery occur relatively frequent and that the majority (62%) are caused by S. aureus . Especially the central reconstructions and the peripheral bypass procedures have a relatively high incidence of surgical site infections, compared to, for example, carotid endartectomies and AV access procedures.
Overall there is no significant relation between nasal carriage of S. aureus and the occurrence of surgical site infections. However, a stratified analysis on patients who underwent abdominal aortic surgery shows a significant association. The effect in this group is comparable to what has been found previously in cardiothoracic and orthopedic surgery. [1] , [2] , [3] , [5] .
In other vascular procedures no significant effect was found. The infections in this group mainly occurred in peripheral procedures of patients with occlusive disease. Patients with occlusive vascular disease cope with insufficient blood flow to at least one of the, mostly lower, limbs. This insufficient blood flow is often associated with ischemic disease, e.g. gangrene of non healing ulcers. As this wounds can be infected or colonized prior to surgery with a large scale of different pathogens, this could limit the role of S. aureus nasal carriage. In our study no significant effect after peripheral vascular surgery was found. However, for one patient who was positive for S. aureus nasal carriage and who developed a S. aureus SSI, accidently a sample of both the nasal swab as well as a wound swab were available for typing. This showed that the two trains were identical. Considering the small number of patients and the frequent presence of wounds before surgery we consider the role of nasal carriage in peripheral vascular surgery unresolved.
All S. aureus strains were methicillin susceptible and no MRSA was found which reflects the low rate of MRSA in Dutch hospitals. Also all strains were mupirocin susceptible. Potentially administration of mupirocin could reduce the risk of nasal carriage. [6] A cost effectiveness analysis showed that treating every patient with S. aureus eradication therapy, without screening for nasal carriage is the most cost-effective way for preventing surgical site infections. [7] However, as recent studies reported mupirocin resistant MRSA strains [13] , [14] , it should only be used in proven MSSA and MRSA carriers to limit the risk for development of further resistance.
Based on the results of this study we conclude that S. aureus carriers who undergo central reconstructive surgery have a significant higher risk for the development of SSI which can be decreased by perioperative eradication of S. aureus in nasal carriers. [6] .
This is important because infection with S. aureus after aortic reconstructions is related to severe complications and a high mortality rate.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Mitochondria are ubiquitous eukaryotic organelles derived from an ancient endosymbiotic α-proteobacterium. Tracing the evolutionary history of mitochondrial proteins is important in understanding how a free-living α-proteobacterium became an integrated eukaryotic organelle. Several mitochondrial proteins are closely related to α-proteobacterial proteins; however, it has been shown that many mitochondrial proteins are derived from bacteria not related to the endosymbiont [ 1 , 2 ]. The majority of the remaining mitochondrial proteins are eukaryote novelties that have no obvious homologues in bacteria [ 1 , 2 ]. To further complicate things, specific protein complexes are not either of endosymbiotic or eukaryotic origin, but instead, several systems like the mitochondrial import complexes and the complexes of the electron transport chain have mixed origins [ 2 ]. Some mitochondrial proteins have obscure evolutionary histories. The current study investigates the evolutionary origins of one such protein, the mitochondrial outer membrane protein MAPL (mitochondria-associated protein ligase, also called MULAN/GIDE/MUL1).
MAPL has been implicated in several processes specific to metazoan (multicellular animal) cells such as NF-kB activation, innate immunity and antiviral signaling, suppression of PINK1/parkin defects, mitophagy in skeletal muscle, and caspase-dependent apoptosis [ 3 – 6 ]. This, coupled with the fact that MAPL has homologues in α-proteobacterial, other bacteria, and archaea [ 7 ], calls into question the conserved ancestral function of MAPL. If MAPL is of endosymbiotic origin, then it likely has an ancient cellular function in addition to the aforementioned metazoan-specific functions. Indeed, other studies have uncovered MAPL’s role in potentially ancient processes like mitophagy, Akt regulation, regulation of mitochondrial dynamics, and sequestration as cargo in mitochondria-derived vesicles (MDVs) targeted to peroxisomes [ 6 , 8 – 15 ].
Human MAPL has a BAM (beside a membrane)/GIDE (growth inhibition death E3 ligase) domain located between two transmembrane domains topologically oriented such that both N- and C-termini reside in the cytosol while the BAM domain resides in the mitochondrial intermembrane space (IMS) [ 10 , 11 , 14 ]. BAM domain proteins are found in every domain of life (eukaryotes, bacteria, and archaea) and its patchy distribution has been explained by several independent horizontal gene transfer (HGT) events [ 7 ]. In human MAPL, a C-terminal cytosolic RING finger domain is present following the last transmembrane domain. Most studies have focused on the function of this RING domain. In the case of its role in regulating mitochondrial dynamics, MAPL stabilizes Drp1 via SUMOylation [ 9 ] and promotes the degradation of Mfn via ubiquitylation [ 5 ]. Hence, upregulation of MAPL results in small fragmented mitochondria [ 5 , 6 , 9 ]; whereas depletion of MAPL results in elongated mitochondria [ 6 ]. The substrates of MAPL’s RING domain when involved in other cellular processes remain unknown; however, at least four different E2 conjugating enzymes as MAPL interacting partners suggesting the existence of several other substrates [ 13 ].
Compared to the RING domain, very little is known about the function of MAPL’s BAM domain. Previous work has shown that only the BAM domain is required for proper packaging of MAPL into MDVs trafficked to peroxisomes [ 8 – 10 ]. Due to their endosymbiotic origins it is commonly held that the mitochondrial membranes are distinct from those of the membrane trafficking system and do not participate in vesicular trafficking. However, recent evidence from experiments in human tissue culture cells suggests there are at least two pathways in which mitochondrial membranes are trafficked to other organelles via MDVs [ 10 , 7 , 16 , 17 ]. One pathway directs MDVs to lysosomes/multivesicular bodies in a PINK1-dependent manner [ 18 ], whereas another pathway that directs MAPL-containing MDVs to peroxisomes is retromer-dependent [ 8 ]. It is important to note that MDVs have not been reported in any organism other than humans and the extent to which MDVs can be generalized to other model systems is unknown. Unfortunately the proteins/protein complexes identified as required for MDV formation (PINK1 and retromer) have several other cellular functions and cannot be used as markers for the presence of MDVs. MAPL has been described as a specific cargo in MDVs trafficked to the peroxisome and represents the best candidate for tracking MDVs in other organisms [ 10 ]. The only other published study [ 12 ] focusing on MAPL’s BAM domain suggests that it is involved in a mitochondrial stress response pathway. This study shows that, in the presence of hydrogen peroxide, IMS-located Omi/HtrA2 protease cleaves the BAM domain of MAPL resulting in the degradation of the protein [ 12 ]. Inactivation of Omi/HtrA2 results in increased levels of mitophagy due to the increased levels of MAPL. A connection between Omi/HtrA2-dependent cleavage of MAPL and its delivery to peroxisomes has yet to be tested.
The potential homology of MDVs and bacterial OMVs (outer membrane vesicles) has not gone unnoticed [ 7 , 16 , 19 ]. Bacterial OMVs are important for nutrient acquisition, biofilm development, and pathogenesis [ 20 ]. While it is attractive to assume that these two processes are homologous (especially in light of the presence of MAPL homologues in prokaryotes), this avenue has not yet been investigated, as the machinery required for OMV production in bacteria is still unknown [ 20 ]. Only by examining MAPL function in other organisms can we begin to gain insight into the ancient conserved function of MAPL. In this study we searched for MAPL homologues in diverse eukaryotes in order to gain insight into the origins and ancient conserved function of MAPL.
Materials and Methods
Genome databases
Publicly available genomes were obtained from various online resources including the Joint Genome Institute, the Broad institute ( http://www.broadinstitute.org/ ), the National Center for Biotechnology Information (NCBI), and some private websites. A list of genomes used in this study is included in S1 Table .
Homology searching
BAM domain containing proteins were retrieved using a combination of BLAST [ 21 ] and HMMer ( http://hmmer.janelia.org/ ) searching strategies of both the NCBI non-redundant database as well as individual genomes. Each putative MAPL homologue was assessed for domain structure using the Pfam database ( http://pfam.xfam.org/ ). Only sequences that contained recognizable BAM/GIDE domains were retained for phylogenetic analysis. Some BAM/GIDE proteins were not present in predicted proteomes were reconstructed from genomic assembly sequences or EST sequence reads available from NCBI. Sequences retrieved in this study are listed in S2 Table .
Phylogenetic analysis
BAM domain-containing proteins were aligned using MUSCLE [ 22 ] and manually adjusted as needed using Mesquite ( http://mesquiteproject.org ). Model testing was performed using ProtTestv1.3 [ 23 ] with a Gamma rate distribution and accounting for invariant sites as appropriate. Phylogenetic tree reconstructions were carried out using MrBayes v3.2.2 [ 24 ] for Bayesian analysis. Maximum likelihood bootstrap values were obtained using PhyML [ 25 ] and RaxML [ 26 ] with 100 pseudoreplicates using the LG [ 27 ] model. Species-specific duplications and long branches representing highly divergent sequences were removed from subsequent analyses in order to limit the effects of long-branch attraction.
Results and Discussion
MAPL-related proteins are present in basally diverging lineages of Viridiplantae, Holozoa and Fungi
We first searched for BAM domain proteins in representatives from diverse eukaryotic groups ( Fig 1A ). We identified putative MAPL homologues in the predicted proteomes of most metazoa (multicellular animals) as well as in the choanoflagellate Salpingoeca rosetta and the ichthyosporean Capsaspora owczarzaki , close single-celled relatives of animals. The presence of MAPL in organisms outside metazoa suggests an ancient origin of MAPL. In order to determine if MAPL was present in the ancestor of Opisthokonta we searched for homologues in fungi and identified MAPL-related proteins in only the chytridiomycete Spizellomyces punctatus and the neocallmastigomycete Oprinomyces sp. ( Fig 1A ). The S . punctatus BAM domain-containing protein lacked the characteristic C-terminal RING domain present in MAPL, but the protein in Orpinomyces sp . clearly contains a C-terminal RING domain. We thus conclude that metazoan MAPL with its extant domain structure likely descended from a protein that was present in the common ancestor of animals and fungi. However, the possibility of HGT of MAPL into these two fungal species cannot be ruled out by these data alone.
10.1371/journal.pone.0128795.g001
Fig 1
Distribution of BAM domain proteins across the tree of life.
A. BAM domain distribution across the three domains of life. BAM domain proteins are present in all three domains of life, but only plants, animals, and a single fungus contain BAM proteins that are followed by a RING domain. Filled circles indicate many taxa contain at least one BAM protein. Open circles indicate only one or two species were identified with BAM proteins. B. Distribution of MAPL (BAM-RING) in holozoa (clade comprising animals and their closest single-celled relatives), with particular focus on non-vertebrates. Most species contain MAPL, but several instances of loss are recorded. C. Expansion of MAPL in the vertebrate lineage followed by loss in mammals. Multiple MAPL paralogues are present in non-mammalian vertebrates (MAPL2 and MAPL2-like). D. Expansion of MAPL in multicellular plants. Green algae contain a single MAPL whereas multicellular plants have gained a paralogue. The Capsella - Arabidopsis clade has further gained a paralogue that has lost the RING domain (MAPL-R).
Although MAPL is present in numerous opisthokonts it is noticeably absent from several lineages. We could find no evidence for MAPL in any sequenced nematode species, or any fungal lineage outside those already mentioned. Although MAPL was present in one species of filasteria ( C . owczarzaki ) and one species of choanoflagellate ( S . rosetta ) MAPL could not be identified in sequences of genomes from the other sequenced holozoan protists Sphaeroforma arctica or Monosiga brevicolis . These absences suggest that although MAPL has been retained in diverse lineages, it has been repeatedly lost over the evolutionary history of the Opisthokonta.
Previous studies have shown that the fish Salmo salar contains two different MAPL paralogues [ 28 , 29 ]. Our investigation led to the discovery of a number of vertebrate species that contain MAPL paralogues ( Fig 1C ). This prompted a phylogenetic analysis (see below).
Although Andrade-Navarro et al. (2009) [ 7 ] found BAM domain-containing proteins in multicellular plants, the extent to which this protein is found in diverse eukaryotes was not investigated. We therefore searched for MAPL homologues in the predicted proteomes of members of the Archaeplastida (glaucophytes, red and green algae, and plants) and other protistan clades. We identified candidate proteins in all multicellular plants examined as well as several unicellular/colonial green algae including Volvox carteri , Micromonas sp ., Chlorella variabilis , Coccomyxa subellipsoidea , Ostreococcus tauri and Chlamydomonas reinhardtii ( Fig 1D ). We also identified a plant-specific MAPL protein that we designated MAPL-P ( Fig 1D and see below). Andrade-Navarro et al. [ 7 ] reported that proteins that lacked the C-terminal RING domain could be found in many plants; however, we found that the majority of archaeplastid BAM domain-containing proteins retained the C-terminal RING domain. Of the genomes investigated, only Capsella rubella and Arabidopsis thaliana contained proteins lacking a C-terminal RING domain. Although 33 diverse protist genomes were searched, BAM domains were identified in only two organisms outside the Opisthokonta and Viridiplanta, the rhizarian Bigelowiella natans and the brown alga Ectocarpus siliculosis (See S2 Table for a list of organisms lacking detectable BAM domains). These two proteins lacked the characteristic C-terminal RING domain present in metazoan MAPL proteins.
Opisthokont and archaeplastid MAPL likely derive from a common ancestral protein
Andrade-Navarro et al. [ 7 ] suggested that BAM domains in various organisms are derived from several independent HGT events. If this were the case, a phylogenetic reconstruction would reveal distinct relationships between certain prokaryotic and eukaryotic proteins. In order to assess the relationships of eukaryotic BAM proteins we performed a phylogenetic analysis on all putative MAPL proteins along with BAM protein sequences from diverse bacterial and archaeal species ( Fig 2 ). It is important to note that no prokaryotic sequence contained a C-terminal RING domain ( Fig 1 ) and thus the RING domain was not included in the phylogenetic analysis.
10.1371/journal.pone.0128795.g002
Fig 2
Phylogenetic reconstruction of BAM domain-containing proteins from opisthokonts, archaeplastida, and prokaryotes.
BAM domain-containing protein sequences were aligned using MUSCLE, Sites that could not be aligned with confidence (including the eukaryote-specific RING domains) were removed manually. The resulting alignment was subjected to phylogenetic analysis (see methods section for details). In this analysis, prokaryotic BAM proteins group together to the exclusion of all eukaryote proteins. Thus, the BAM domain-containing proteins present in various eukaryotes cannot be traced to independent HGT events. In this and all following phylogenetic analyses, numerical values represent Bayesian posterior probabilities and maximum-likelihood bootstrap values (Bayesian/PhyML/RAxML). Node values are given to highlight the clades of interest, denoted by coloured boxes and annotated by protein name. All other node support is iconized as inset.
Our analysis revealed that prokaryotic BAM proteins form a robustly supported clade to the exclusion of all opisthokont and archaeplastid MAPL proteins. The plant-specific clade designated as MAPL-P formed a robust clade with no obvious affinity to the other plant sequences. The A . thaliana and C . rubella proteins that lack the C-terminal RING domain (MAPL-R) group weakly with other archaeplastid MAPL proteins but strongly group within the larger eukaryotic clade ( Fig 2 ). This suggests that HGT events are not responsible for the presence of the MAPL-R proteins in the Archaeplastida. The lack of the RING domain can instead be attributed to secondary loss.
Independent expansion of MAPL in multicellular plants and animals
When searching for MAPL homologues in diverse eukaryotes, we noticed that several genomes contained more than one BAM-RING protein. Several vertebrate genomes contained two distinct MAPL proteins while several plant genomes contained many BAM domain-containing proteins. In order to identify lineage-specific duplications we reconstructed the phylogenies of these proteins in Viridiplantae and Opisthokonta.
Within the vertebrates we found that, while mammals contain only a single MAPL protein, amphibians and reptiles contain two MAPL paralogues while the fishes Danio rerio and S . salar contained three and two paralogues, respectively (See Tacchi et al. [ 28 , 29 ]). The apparent discordance between the number of paralogues in the two fishes compelled us to search for the potential missing S . salar sequence. In our searches of the S . salar EST database we found a third MAPL transcript that encodes a protein similar to the third D . rerio sequence. Our reconstructed phylogeny demonstrates that there was likely a duplication of MAPL at the base of vertebrates producing two proteins, MAPL and MAPL-like, one of which (MAPL-like) was subsequently lost in the mammalian lineage ( Fig 3 ). These data also demonstrate that another independent duplication occurred in the fish lineage producing MAPL-like2 in fishes. It is interesting to note that MAPL and MAPL-like2 (the fish-specific paralogue) have been studied in S . salar , but the function of MAPL-like has not yet been investigated [ 28 , 29 ].
10.1371/journal.pone.0128795.g003
Fig 3
Phylogenetic analysis of Opisthokont MAPL proteins.
This analysis demonstrates that all vertebrate MAPL proteins group together to the exclusion of all other opisthokont MAPL proteins. MAPL has been retained in all major vertebrate clades. MAPL-like is an ancient vertebrate protein lost in the mammalian lineage. MAPL-like2 is specific to fishes. The RING domain was included in the alignment in this analysis as the vast majority of predicted proteins contained this domain. Node support as in Fig 2 .
Within Archaeplastida, multicellular plants contain very closely related paralogues of MAPL proteins. Most plant MAPL proteins fall into one of two clades, one rather divergent clade specific to multicellular plants (long branch MAPL-P clade in Fig 2 ) and a less divergent clade that shares features with the MAPL proteins found in green algae. While most of the proteins that we identified contained C-terminal RING domains, some of the paralogues identified in Arabidopsis thaliana and the closely related species Capsella rubella lacked the C-terminal domain. Our phylogenetic analysis of Archaplastid MAPL ( Fig 4 ) further indicates that MAPL-R proteins represent a divergent lineage-specific expansion of MAPL in the Arabidopsis / Capsella clade that has subsequently lost the C-terminal RING domain.
10.1371/journal.pone.0128795.g004
Fig 4
Phylogenetic analysis of Archaeplastid MAPL proteins.
This analysis demonstrates that A . thaliana and C . rubella BAM proteins that lack the RING domain group within a weakly supported clade comprising sequences from multicellular plants that retain a RING domain. The RING domain was excluded from this analysis. Node support as in Fig 2 .
Stramenopile and rhizarian BAM/GIDE proteins are related to bacterial proteins
In order to determine if the BAM proteins identified in representatives from the SAR (stramenopiles, alveolates, rhizaria) clade are more closely related to other eukaryote sequences or prokaryote sequences, BAM protein sequences from B . natans and E . siliculosis were added to our previous alignment used to generate Fig 2 but with all long-branches and lineage-specific duplications removed. To our surprise, we found that the SAR BAM domain proteins were excluded from the other eukaryote sequences ( Fig 5 ). This suggests that recent HGT events might be responsible for the presence of BAM domains in these two species alone.
10.1371/journal.pone.0128795.g005
Fig 5
Phylogenetic analysis of SAR BAM proteins.
In this analysis the BAM proteins from E . siliculosis and B . natans group with prokaryotic sequences suggesting that these eukaryotic proteins have a different origin than other MAPL and might be derived from recent HGT events. The RING domain was excluded from this analysis. Analysis and node support as in Fig 2 .
Functional and evolutionary implications
Presence of conserved BAM-RING domain architecture in diverse eukaryotes and the relative rarity of loss of the RING domain suggests that both domains are important for the overall function of MAPL. If the domains functioned independently it would be expected that the domains would be found separated from one another more frequently over the course of eukaryote evolution. This is interesting because studies to date have not found a link between the two domains and the different functions discovered in animals seem to be distinct.
Duplication and divergence in some lineages, like vertebrates and plants, suggests that functional specialization has occurred. The duplication at the base of vertebrates is of particular interest because the divergent paralogue is lost in mammals marking a relatively rare example of secondary loss in the mammalian lineage. Investigation of MAPL-like proteins in model fish, reptile, and amphibian species will be of great interest for understanding the conserved function of MAPL in mammalian cells.
Although eukaryotic MAPL appears to derive from a single eukaryotic protein with similar domain architecture to extant MAPL, the ancestry of MAPL remains difficult to resolve. Previous work [ 7 ] suggested that MAPL may have arisen in bilateria and was subsequently horizontally transferred to plants and prokaryotes. Our discovery of MAPL in fungi, unicellular holozoa and green algae makes this suggestion unlikely. Since only archaeplastids and opisthokonts retain MAPL it is impossible to determine if the protein was present in the LECA (last eukaryotic common ancestor) or if MAPL was horizontally transferred very early between the opisthokont and viridiplantae lineages. Although unlikely, it is conceivable that MAPL might have evolved in either the ancestor of all opisthokonts or the ancestor of all viridiplantae and a single HGT event between them could account for the current distribution of the protein. We feel that a more parsimonious explanation is that MAPL is ancient and was present in the LECA (perhaps by HGT from bacteria or even the mitochondrial endosymbiont) but was independently lost in several eukaryotic lineages while retained in Opisthokonta and Viridiplantae.
Conclusions
MAPL is involved in organelle dynamics, mitophagy, mitochondria to nucleus signaling, and is packaged as cargo in MDVs destined for peroxisomes. The presence of MAPL in diverse eukaryotes allows us to infer its ancient origins, but several questions still remain. Specifically, although MAPL is ancient we currently have no way of knowing if it was present in the LECA or if it arose later in evolution. It is also still unclear how many HGT events may have occurred between prokaryotes.
Its taxonomic distribution suggests that MAPL functions in several single-celled eukaryotes. As multicellular animals are cluttered with proteins and pathways required for obligate multicellular life, investigation into MAPL function in single-celled organisms like the choanoflagellate S . rosetta , the filasterian C . owczarzaki , and the green alga C . reinhardtii is critical to understanding MAPL’s ancestral function. We hope that our study provides the impetus for future functional work on MAPL in organisms outside metazoa.
Supporting Information
S1 Table
Genome databases used in this study and indication of BAM presence or absence.
(XLSX)
S2 Table
Proteins retrieved in this study.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Bisphosphonates (BPs) had long been used in metabolic bone disease as osteoporosis, tumor-associated hypercalcaemia and metastases-induced osteolysis due to their ability to inhibit bone resorption. BPs were able to bind divalent cations like Ca 2+ or zinc constituting the basis of their bone-targeting property and their inhibition of the proteolytic activity of matrix metalloproteinases (MMP), respectively. The nature of their side chains gave rise to a variety of possible structures and stereochemistry determining their different potencies [1] – [4] . Non-nitrogen containing BPs (non N-BPs) acted by forming non hydrolysable ATP-analogues and were less effective than nitrogen-containing BPs (N-BPs) in inhibiting bone metastasis [5] . However, Zoledronate treatment of patients were reported to induce toxic side effect characterised by osteonecrosis of the jaw while non N-BP did not produce this effect [6] , [7] . N-BPs, such as zoledronate, acted on the mevalonate pathway inhibiting the farnesyl diphosphate synthase (FPP) thereby depleting the cells of the farnesyl (FPP) or geranylgeranyl (GGPP) diphosphate isoprenoids [8] . Isoprenoids were required for translocation and anchorage of small G proteins like Rho or Ras to the plasma membrane assuring their ultimate involvement in signal transduction during several important normal and tumor cellular pathways.
However, in vivo efficacy of all BPs on extra-osseous sites or primary tumors was still debated. Only a small number of studies demonstrated their in vivo antiproliferative activity on tumors or metastasis present in soft tissues [9] . The reasons were the poor oral bioavaibility (0.3–7% in humans) due to chelation of metal ions by phosphonic acid group inside the digestive lumen, poor membrane permeability due to poor BP lipophilicity as well as strong uptake by bone tissue [10] . Previously, our laboratory developed a new strategy to overcome BP hydrophilicity by masking the phosphonic acid with organic protecting groups and introducing hydrophobic functions in the side chain [11] . We previously demonstrated that an esterified BP with methyl group displayed antitumor growth and antiangiogenic activities on A431 tumors being more effective in vivo than in vitro [12] . In order to further increase the lipophilicity of BPs (and their entering into the cells), we synthesized new aromatic 1-hydroxymethylene-1,1-bisphosphonic acids containing phenyl or halogen phenyl ring in their side chains. Interestingly, we showed that these compounds exhibited potent antiproliferative activities in vitro on human epidermoid A431 cells [13] . In parallel, recently crystallographic and computational investigation revealed that the presence of phenyl ring in the side chain permitted non N-BPs to interact with farnesyl enzyme [14] . Thus, we synthesized a class of BPs that contained bromobenzyl in their side chains (BP7033Br, Fig. 1 ). For the first time, we symmetrically esterified one of each phosphonic acids with aromatic groups (BP7033Br ALK, Fig. 1 ).
10.1371/journal.pone.0004685.g001 Figure 1
Chemical structure of BP7033Br and BP7033Br ALK.
The first step (1) was an Arbusov reaction between an activated carboxylic function and a very reactive species, the bis(trimethylsilyl) phosphite and the second one (2) was hydrolysis.
In this study we tested the effects of BP7033Br and BP7033Br ALK on MDA-MB-231 xenograft growth and metastasis. We found that the addition of hydrophobic bromobenzyl groups on non N-BPs side chain rendered them efficient in inhibiting estrogen responsive as well as non-responsive breast cancer cells like MDA-MB-231 and metastatic subpopulation (D3H2LN) cell growth as well as migration and invasion in vitro . Further, we demonstrated that BP7033Br and BP7033Br ALK differently reduced the proteolytic activity of MMP-9 and MMP-2. These two non-N BPs inhibited MDA-MB-231 xenografts associated with tumor angiogenesis reduction. Importantly, we demonstrated that only esterification of m-bromobenzyl bisphosphonate revealed to induce antimetastatic effect in nude mice.
Results
BP7033Br ALK was more potent than BP7033Br in inhibiting breast cancer cell viability
We investigated the effect of BP7033Br and its esterified analogue BP7033Br ALK on the proliferation of several breast cancer cell lines with different estrogen-receptor statuses as T47D ( Fig. 2A ), MCF-7 (B), SKBR3 (C), MDA-MB-231(D) and its D3H2LN metastatic subclone (E). The special features of D3H2LN subclone are that these cells exhibited more important tumor growth than MDA-MB-231 and that when injected in the left heart ventricle, they induced faster metastasis and more multiple tissues sites [15] . After 72 h, BP7033Br differently inhibited the several cell lines. The maximal T47D cell viability inhibition induced by BP7033Br was about 38% and the IC 50 was not reached even at the maximal concentration (1 mM). MCF-7 and SKBR3 cells were inhibited by 50% with 500 µM of BP7033Br and this concentration was the more effective for SKBR3 cells. Viability of both MDA-MB-231 and D3H2LN was inhibited by BP7033Br in a dose-dependent manner with the same efficacy (IC 50 = 500 µM, Fig 2B ). For these cells, the maximal effect (80%) was achieved at 1 mM of BP7033Br. Considering BP7033Br ALK, we showed that this esterified BP was more effective than BP7033Br on T47D, MCF-7 and SKBR3 cells with maximal effect achieved at 250 µM. BP7033Br ALK also inhibited the MDA-MB-231 and D3H2LN cell lines in a dose-dependent manner with the same efficacy (IC 50 = 170 µM, Fig. 2D and E ). In addition, BP7033Br ALK was at least 4-fold more efficient than non esterified BP7033Br since maximal inhibition of cell viability (90%) was reached at 250 µM. Breast cancer cell viability inhibitions induced by both BPs were also time-dependent and started after 24 h (data not shown).
10.1371/journal.pone.0004685.g002 Figure 2
BP7033Br and BP7033Br ALK inhibited viability of different breast cancer cells.
T47D (A), MCF-7 (B), SKBR3 (C), MDA-MB-231 (D) or D3H2LN cells (E) (1×10 5 ) were treated with BP7033Br and BP7033Br ALK at increasing concentrations for 72 h. Then, the cells were washed and incubated with 0.1 mL of MTT (2 mg/mL) for 4 h. Optical density was measured at 570 nm using a Labsystems Multiskan MS microplate reader. Data represents the mean value (±SD) of three independent experiments.
BP7033Br and BP7033Br ALK inhibited MDA-MB-231 breast cancer cell viability through different cell cycle arrests
In order to understand the effect of the BPs on breast cancer cell viability, we evaluated their effect on cell cycle progression ( Fig. 3 ). MDA-MB-231 and D3H2LN cells ( Fig. 3A and 3B , respectively) were treated with effective doses of BP7033Br and BP7033Br ALK for 72 h (500 µM and 200 µM, respectively). BP7033Br blocked 10% of MDA-MB-231 cells into the G0/G1 phase ( P = 0.0002) and diminished the cell number in the S phase ( P = 0.017, Fig. 3C ). In contrast, BP7033Br ALK increased the number of MDA-MB-231 cells into the S phase ( P = 0.005) accompanied by a reduction of the proportion of cells in G0/G1 phase ( P <0.001). Concerning the D3H2LN cells, BP7033Br increased about 20% the number of cells ( P = 0.048, Fig. 3D ) in the G0/G1 phase, as observed for MDA-MB-231 cells, and completely inhibited the proportion of cells in the G2/M phase. Also, BP7033Br ALK increased the number of D3H2LN cells into the S phase to about 15% ( P = 0.006, Fig. 3D ).
10.1371/journal.pone.0004685.g003 Figure 3
BP7033Br and BP7033Br ALK inhibited MDA-MB-231 breast cancer cell cycle progression.
Distribution of MDA-MB-231 (A) and D3H2LN (B) cells treated with BPs in different cell cycle phases was determined as described in “ Materials and Methods ”. Histograms show the percentage of MDA-MB-231 (C) and D3H2LN (D) cell repartition. Each column represents a mean (±SD) of three independent experiments. *P control versus BP7033BrALK and ** P control versus BP7033Br <0.05.
BP7033Br and BP7033Br ALK induced MDA-MB-231 breast cancer cell death
The Ann-V-positive/PI-negative population corresponds to cells in an early apoptotic phase and the Ann-V-positive/PI-positive one to cells in a late apoptosis phase ( Fig. 4 ). We evaluated the apoptotic effect of efficient doses of both BPs on breast cancer cell lines. Both BP7033Br (500 µM) and BP7033Br ALK (200 µM) induced apoptosis of the MDA-MB-231 cells inducing the same amount of cells in both early and late apoptosis ( Fig. 4A ). Total percentage of MDA-MB-231 cells in apoptosis (late plus early) induced by BP7033Br and BP7033BR ALK is about 37 and 25% as compared to control, respectively ( Fig. 4B ). The percentage of apoptotic cells was not significantly different between the two BPs treatments( P = 0.069) although concentration used for BP7033Br ALK was lower (200 µM) Concerning D3H2LN cells, the effect of the two BPs was slightly less effective (20%) as compared to the MDA-MB-231 parental cells ( Fig. 4C and 4D ).
10.1371/journal.pone.0004685.g004 Figure 4
BP7033Br and BP7033Br ALK induced MDA-MB-231 breast cancer cell apoptosis.
Preconfluent MDA-MB-231 (A) and D3H2LN (C) cells were treated with 500 µM BP7033Br or 200 µM BP7033Br ALK for 72 h in a serum-containing medium. Percentages of MDA-MB-231 (B) and D3H2LN (D) were evaluated as described in “ Materials and Methods ”. Each column represents a mean (±SD) of three independent experiments.
BP7033Br and BP7033Br ALK inhibited MDA-MB-231 cells migration, invasion and MMP-9 and MMP-2 proteolytic activation
We next explored the effect of BPs on the migration of the two MDA-MB-231 cell lines ( Fig. 5A ). MDA-MB-231 cells as well as D3H2LN subpopulation ones, migrated through the lower chamber side when 10% FCS-DMEM was used as chemo attractant. BP7033Br (125 µM) inhibited about 55% the migration of MDA-MB-231 cells but was less efficient on the D3H2LN cell migration (32%, P <0.001). The same concentration of BP7033Br ALK induced stronger motility inhibition of both cell lines (62 and 50%, respectively). Since invasion is also important in the metastasis process, we investigated the inhibition induced by BPs on the MDA-MB-231 cell invasion ( Fig. 5B ). In the presence of 10% FCS-DMEM, MDA-MB-231 cell lines invaded the inserts coated with matrigel. In contrast to the results obtained with MDA-MB-231 cell migration, we observed a better inhibition of D3H2LN invasion by BP7033Br as compared to BP7033Br ALK (84% versus 40%, P <0.001). Concerning the MDA-MB-231 cells, BP7033Br ALK inhibited cell invasion about 65% whereas BP7033Br reduced it about 45%.
10.1371/journal.pone.0004685.g005 Figure 5
BP7033Br and BP7033Br ALK inhibited MDA-MB-231 breast cancer cell migration, invasion, MMP-9 and MMP-2 activities.
BPs inhibited MDA-MB-231 breast cancer cell migration (A) and invasion (B). Cells (2.5×10 5 ) with 125 µM of BPs were added to each 8 µm-insert in the upper chamber of boyden chamber. After 24 h, cells invading the chamber were fixed, stained and counted as described in “ Materials and Methods ”. BPs inhibited MMP-9 and MMP-2 activities (C and D, respectively). Lyophilized conditioned media were normalized to the number of cells and subjected to 10% SDS-polyacrylamide gels containing 1 mg/mL gelatine. Lane 1, 2 and 3 represent the control, BP7033Br and BP7033Br ALK conditioned medium of treated cells, respectively. Each column represents a mean (±SD) of three independent experiments. * P versus MDA-MB-231 control <0.05, ** P versus D3H2LN control <0.05.
To further study the BPs effects on invasion, we evaluated the effect of BPs on MMP activity in the D3H2LN cells ( Fig. 5 ). D3H2LN cells showed activation of MMP-9 and MMP-2 (lane 1, Fig. 5C and 5D ). BP7033Br strongly inhibited the activation of MMP-9 ( Fig. 5C , lane 2, P = 0.0037). In contrast, BP7033Br ALK only reduced about 45% the MMP-9 activity ( Fig. 5C , lane 3, P = 0.0065). In addition, BP7033Br inhibited about 40% ( P = 0.001) MMP-2 activity whereas BP7033Br ALK reduced MMP-2 activation about 25% ( P = 0.01, Fig. 5D , Lane 2 and 3, respectively).
Both BP7033Br and BP7033Br ALK inhibited the D3H2LN tumor growth and angiogenesis in nude mice
All 7 mice developed tumors within one week after D3H2LN inoculation and the BPs treatments were initiated at the end of the first week ( Fig. 6A ). Both BP7033Br and BP7033BR ALK used at 11 mg/kg (286 µg/mouse) twice a week significantly inhibited D3H2LN growth after 21 days as compared to control ( P = 0.034 and P = 0.038, respectively). D3H2LN tumor growth was inhibited by about 80% with both BPs. All mice treated with the two BPs were alive at the end of the treatment and did not present significant loss of body weight ( Figure 6B ). In addition, no macroscopic differences were observed between treated and control mice liver and kidney after autopsy (data not shown).
10.1371/journal.pone.0004685.g006 Figure 6
BP7033Br and BP7033Br ALK inhibited D3H2LN tumor growth and esterified analogue completely inhibited angiogenesis.
(A) D3H2LN cells were inoculated in nude mice as described in “ Materials and Methods ”. After 1 week, mice were treated with BPs (11 mg/kg), twice a week, for 21 days. Each column represents the mean of tumor volume (mm 3 ) (±SD, n = 7). Body weight (BW) ratio was determined for each group (B). Endothelial cells in tumor sections were stained in controls (C), BP7033Br (D) and BP7033Br ALK (E) Microvessels were indicated by arrows and necrosis area by double asterisks (magnification ×100). Quantification of micro-vessel density (F). Each column represents a mean (±SD) of three independent experiments. * P BP7033Br and BP7033Br ALKversus control <0.05.
GSL-1 selectively labelled the endothelial cells and thus enabled to determine the relative density of micro-vessels in D3H2LN tumors xenografted in nude mice ( Fig. 6C–E ). BP7033Br ALK treated tumors exhibited large areas of necrosis ( Fig. 6E ). BP7033Br ALK treatment strongly reduced the micro-vessel density (80%) in viable field of tumors as compared to control and BP7033Br treatments ( Fig. 6F , P = 0.01 and P = 0.047, respectively).
Only BP7033Br ALK inhibited D3H2LN metastasis
Since D3H2LN was described to develop bone and soft tissues metastasis after intracardiac injection in nude mice [15] , we further studied the effect of both BPs on this D3H2LN metastasis model ( Fig. 7 ). Animals (n = 7) with successful intracardiac injection at day 0 ( Fig. 7A ) were treated with each BPs at 11 mg/kg twice a week during 4 weeks. The control group received sterile PBS. Injections started the first week when micrometastasis appeared. Within 2 weeks after intracardiac injection, all animals exhibited metastasis at multiple sites in the head, thorax, abdomen, legs or spine ( Fig. 7A ). Ex vivo imaging of the different tissues after the final imaging in vivo , showed bone lesions (legs) as well as lung, lymph nodes, ovary, bladder and liver ones ( Fig 7B ). Treatment with BP7033Br ALK at 11 mg/kg twice a week, induced inhibition about 80% of luminescence signalling and 50% reduction of the mean metastatic sites number per animal ( P = 0.01 and 0.08, respectively, Fig. 7C ). In contrast, BP7033 did not show any significant reduction in metastasis bioluminescence signals or number sites ( Fig. 7D ).
10.1371/journal.pone.0004685.g007 Figure 7
Only BP7033Br ALK inhibited D3H2LN metastasis.
D3H2LN cells were injected into the left ventricle of nude mice (n = 7). Day 0 showed the successful intracardiac cells injection. Within 2 weeks, when metastasis were initiated, mice were treated with BP7033Br ALK or BP7033Br (A). At the indicated days, the bioluminescence images were acquired for control (c, left panel) and BPs treated mice (BP7033BrALK and BP7033Br middle and right panel, respectively). Ex vivo data confirm soft tissue metastasis from D3H2LN cells injection (B). Quantification of the mean metastatic sites and the photons/s after BP7033Br ALK treatment (C). Quantification the photons/s after BP7033Br treatment (D). Each column represents a mean (±SD) of three independent experiments. * P versus control <0.05.
Discussion
BPs represented an emerging class of drugs for cancer therapy. In this work, we demonstrated the efficacy of a new class of non-N-BPs which exhibited higher antiproliferative activities on breast cancer cells compared to previously described non-N-BPs such as clodronate [16] . Both types of m-bromobenzyl BPs inhibited the viability of several breast cancer cell lines with different estrogen-receptor statuses. We showed that the esterified BP was the more effective on estrogen-responsive cells since the maximal inhibition was reached at 250 µM in contrast to non esterified BP that did not induce maximal inhibition even at 1 mM. In addition, at 250 µM, BP7033Br ALK is effective on cells independently from the estrogen-receptor status. Both types of our BP inhibited viability of estrogen non-responsive cells and particularly that of MDA-MB-231 and D3H2LN cells, the last cell line being the more aggressive ones. It was worth to note that we demonstrated for the first time a dramatic improvement of antiproliferative effect of non-N-BPs on breast cancer cells since clodronate at the same concentration range (200 µM) and the same time-treatment (72 h) did not reduce MDA-MB-231 cell viability [17] – [19] . In addition, clodronate demonstrated mitogenic effects via MCF7 estrogenic receptor [16] and we never observed this effect with our BPs. Based on our results, it appears also that BP7033Br ALK antiproliferative effect is estrogen-receptor-independent. The occurrence of this new effect of non-N-BPs could result from the addition of aromatic functions in the side chain. Heterocycle in the side chain was implicated in the induction of cell apoptosis by preventing the prenylation of signalling proteins such as Ras or Rho [20] . We previously demonstrated the inhibition of Ras processing using non bromo-containing BP7033 [21] . Also, the addition of phenyl function in the side chain of BPs rendered the catalytic pocket of geranyl and/or farnesyl synthase enzymes of the mevalonate pathway more accessible [14] , [22] . BP7033Br ALK reduced MDA-MB-231 and D3H2LN cell viabilities about 90% with a concentration 4-fold inferior to that of BP7033Br. These data were in agreement with our previous results on epidermoid A431 cell proliferation that showed a beneficial effect of esterification of the phosphonic groups [12] , [13] . Our hypothesis was that such esterified compounds rendered BPs more hydrophobic increasing their cell uptake and could therefore act like prodrugs releasing active BP into the cells. In accordance, characterisation of the hydrophilicity demonstrated a shift toward lipophilicity of the BP7033Br ALK compound (Log P values of −0.75 versus −031, respectively). Alternatively, one could also hypothesize that esterified BPs had their own mechanisms since they blocked the cells into the S phase while non esterified BPs inhibited the G0/G1 cell phase transition. On the other hand, both type of BPs (esterified or not) induced cell death apoptosis of both MDA-MB-231 and D3H2LN cells. These results are interesting since D3H2LN had a metastatic potential greater than MDA-MB-231 and consequently could be more resistant to apoptosis as it was described for metastatic cells [23] , [24] . Both BPs induced strong D3H2LN metastatic cell apoptosis but the concentration of the esterified analogue used was 2-fold lower. Also, the two BPs inhibited migration of MDA-MB-231 cell lines with a more important effect obtained with BP7033Br ALK. In contrast, BP7033Br ALK was less effective in inhibiting D3H2LN cell lines invasion concomitant with a less important effect on MMP-9 and MMP-2 activities. BP7033Br strongly inhibited MMP-9 and MMP-2, the major form of metalloproteinases present in extracellular matrix. Since MMPs are zinc-dependent endopeptidases, we speculated that the reduction of BP7033Br ALK effect could be due to a reduction of available phosphonic acid groups able to chelate zinc and consequently inhibit MMPs. These results were in agreement with previous studies which showed that MMPs inhibition activity by BPs was related to zinc chelation [2] . However, we hypothesised that release of BP7033Br ALK with free phosphonic group could be more important in in vivo system because of the presence of phosphodiesterases in serum. Also, we found that these BPs had no influence on MMPs expression (data not shown).
Interesting results were the BPs antitumor effects observed on D3H2LN xenografts growth and metastasis. D3H2LN cells were obtained from a MDA-MB-231 subclone isolated from a lymph node metastases and induced an increased xenograft tumor growth as compared to parental cells when injected in vivo [15] . Both BP7033Br and BP7033Br ALK inhibited about 80% the D3H2LN tumor growth after intratumoral injection of about 286 µg BPs per mouse twice a week during 21 days. We established that this new class of BPs was the more potent among the current non N-BPs since clodronate, even used at 1600 mg twice daily during several weeks (as compared to BP7033 ALK corresponding human dose of 770 mg twice a week during only 2 weeks) failed to reduce primary tumor growth [25] . Also, we found that our BPs were 10-fold more potent than the non halogenated phenyl analogues [26] . In addition, BP7033Br ALK better inhibited D3H2LN vessel density than BP7033Br. This point is also supported by the large necrosis area not detected in BP7033Br treated tumors. In addition, we can not exclude that these large necrosis areas could also be due to a greater amount of esterified BP penetrating into tumor to induce cell death. As compared to N-BPs, it was noteworthy that risedronate or ibandronate failed to inhibit MDA-MB-231 tumor growth in nude mice [27] , [28] . Furthermore, no pre-clinical works demonstrated an antiproliferative effect of zoledronate on primary breast tumor growth in nude mice. The only study demonstrating an inhibition effect of zoledronate on primary tumor growth used mesothelioma tumors which involved calcification that could uptake the drug [29] . In addition, the efficacy of zoledronate on bone metastasis seems to be supported by its affinity for osseous tissues rather than its direct antiproliferative effect on tumor cells [25] . Also, zoledronate was a compound rapidly eliminated from plasma, resulting in renal excretion, rapid bone or calcified tissues uptake and accumulation partly due to its phosphonic functions. Here, we showed that the symmetrical esterification of the phosphonic groups may improved BPs soft tissue bioavaibility limiting osseous or calcified tissue uptake. Also, as their chemical structures are close to the apomine BP which presented aliphatic ester group, we suggested that their half-life will be close to that found for this drug (156 h with micromolar plasma concentration, [30] ). Thus, esterified BP7033Br ALK abrogated angiogenesis, both soft tissue and bone metastasis whereas BP7033Br did not. Noteworthy, in BP7033Br treated mice, luminescence signalling of leg osseous metastasis was not significantly reduced because 2/7 mice did not respond to the treatment in contrast to BP7033Br ALK treatment that induced significant reduction (data not shown). Indeed, the esterified functions seem to be important for the BPs distribution within the systemic system and less for local injection (subcutaneous tumors) since the two N-BPs studied both inhibited D3H2LN xenograft growth. To note, was the toxic adverse effects of N-BPs inducing osteonecrosis of treated patients [7] and no apparent side effect reported with non N-BPs in the present study.
In conclusion, esterified m-bromobenzyl non N-BPs constituted a new class of drugs with potent direct antitumor, antiangiogenic and antimetastatic efficacy on breast tumors.
Materials and Methods
Bisphosphonates synthesis and hydrophilicity characterisation
The bisphosphonate molecule evaluated in this study corresponds to [2-(3-Bromo-phenyl)-1-hydroxy-1-phosphono-ethyl]-phosphonic acid (MW: 400 g/L, Fig. 1 ). This compound (BP7033Br) was synthesised as previously described [31] , [32] . The novel esterified analogue, {2-(3-Bromo-phenyl)-1-hydroxy-1-[hydroxy-(4-methoxy-phenoxy)-phosphoryl]-ethyl}-phosphonic acid mono-(4-methoxy-phenyl) ester named BP7033Br ALK (MW: 572 g/L, Fig. 1 ), was synthesised following the same methodology in two steps. Briefly, a very reactive species, the bis(trimethylsilyl)( p -methoxyphényl) phosphite, was reacted following an Arbusov (1) reaction with an activated carboxylic function and then the intermediate was hydrolysed (2). Log P values were determined using a shake-flask method as described previously [33] . Aqueous sodium chloride (0.9% w/v) and n-octanol phase were saturated for a week. Bisphosphonates were dissolved at a concentration of 0.02 M in the aqueous phase (2.5 mL). An equal volume of saturated n-octanol was added and the solution was mixed. The content of aqueous phases was determined by UV spectroscopy respectively at 268 and 278 nm (BP7033 Br and BP7033Br ALK). The logarithm of the ratio of BPs concentrations in the organic and aqueous phases was calculated to determine following log P values: log P (BP7033Br) = −0.75 ; log P (BP7033BrALK) = −0.31. In addition, 0at physiological pH both compounds BP7033 Br and BP7033Br ALK were readily soluble at 2 mM concentration in both distilled water and aqueous sodium chloride solution (0.9% w/v). BP7033Br ALK becomes insoluble in water at concentration higher than 20 mM and below 10 mM in aqueous sodium chloride solution. BP7033 Br was still soluble at 50 mM in both solutions.
Cell lines
The human breast adenocarcinoma MDA-MB-231, T47D, MCF-7 cell lines were obtained from the American Type Culture Collection (Manassas, VA, USA). SKBR3 were kindly obtained from Dr Cavailles (U824, CRLC, Montpellier, France). Cells were maintained in Dulbecco's minimal essential medium supplemented with 10% fetal bovine serum (FBS), 1% sodium pyruvate and antibiotics (1% penicillin sodium and 1% streptomycin) at 37°C in a humidified atmosphere containing 5% carbon dioxide. D3H2LN cell line isolated from MDA-MB-231 lymph node metastasis was obtained from Caliper Life Sciences (Alameda, CA, USA). D3H2LN cell line was a subclone selected from a MDA-MB-231 stable clone expressing firely luciferase. MDA-MB-231 cells expressing luciferase were injected into the mammary fad pad of nude mice and after 12 weeks of growth in vivo , they were harvested and re-propagated in vitro . This subclone was injected once more into the mammary fad pad of nude mice to yield a second cell line D3H2LN, harvested from a lymph node metastasis [15] . D2H2LN cells were cultured in Minimum Essential Medium with Earl's Balanced Salts Solution MEM/EBSS medium supplemented with 10% fetal bovine serum, 1% nonessential amino acids, 1% L-glutamine, and 1% sodium pyruvate and antibiotics (all from Hyclone, Logan, UT, USA) at 37°C in a humidified atmosphere containing 5% carbon dioxide.
Cell viability
MDA-MB-231 and D3H2LN cell viability was assessed using the MTT-microculture tetrazolium assay [34] . The cells were then incubated with different concentrations of BP7033Br or BP7033Br ALK, for 72 h at 37°C in a 5% CO 2 -incubator. Optical density was measured at 570 nm using a Labsystems Multiskan MS microplate reader.
Cell Cycle Analysis
Cells (2×10 5 ) were incubated with 500 µM BP7033Br or 200 µM BP7033Br ALK for 72 h. Adherent cells were harvested, washed with cold PBS, then fixed with ice-cold 70% ethanol at −20°C for 1 h. Cells were then treated with RNAse A (200 µg/mL) and stained with propidium iodide (50 mg/mL) at room temperature for 30 min in the dark. After incubation, the red fluorescent cells were analysed by flowcytometer (Becton Dickinson). DNA histograms were created using Cell Quest software (Becton Dickinson) analysing 1×10 4 events per sample. The relative distribution of cells in the phases of the cell cycle was calculated with ModFiLT software (Becton Dickinson).
Cell death detection
To reveal a phosphatidylserine translocation specific to early apoptosis stage, the cells were stained with a FITC-labelled annexin-V (Ann-V). The ultimate stage of apoptosis or the first stage of necrosis was revealed by incorporation of PI, which enters into the cells when cell membrane damage has occurred. Cells (1×10 5 ) were incubated with 500 µM BP7033Br or 200 µM BP7033Br ALK for 72 h in serum-containing medium. No organic vehicle was used since both BPs were soluble in water or culture medium until 20 mM. In addition, cells were harvested and apoptotic cells were determined using the Annexin V-FITC Apoptosis Detection kit (Beckman coulter, Fullerton CA, USA). Then, the cells were analysed by flow cytometry (Becton Dickinson, Heilderberg, Germany).
Cell migration and invasion assay
Cell migration experiments were performed using migration chambers containing 8 µm pore size (Becton Dickinson). Cells (2,5×10 5 ) with 125 µM of BPs were added to each insert (upper chamber). 10% FCS-DMEM was used as chemoattractant. For invasion assays, Boyden invasion chambers with 8 µm pore size filters coated with Matrigel (Falcon, MA, USA) were used. After 24 h incubation, non-migrated cells in the upper chambers were removed by wiping cells with a cotton swab. Then, cells on the lower filter face were fixed, stained and counted. Results were expressed as a percentage, relative to controls normalised to 100%. Experiments were performed in triplicate.
Gelatine zymography
The serum-free conditioned media of MDA-MB-231 and D3H2LN cells treated with both BPs for 24 h in 6-wells plates were lyophilized. Lyophilized conditioned media were normalized to the number of cells mixed with non-reducing LaemmLi buffer and subjected into 10% SDS-polyacrylamide gels containing 1 mg/mL gelatine in the presence or absence of BP. The gel was washed 3 times at room temperature in a solution containing 2.5% (v/v) Triton X-100 and incubated at 37°C for 24 h in 50 mM Tris/HCL, pH 7.5, 0.2 M NaCl, 5 mM CaCl2 and 0.05% Brij 35. The gel was stained for 60 min with 0.25% (w/v) R-250 Coomassie blue in 45% (v/v) methanol/10% (v/v) acetic acid and destained in 25% methanol (v/v)/10% acetic acid (v/v).
Xenografts in nude mice
All in vivo experiments were carried out with local ethical committee approval and accordingly to the UKCCCR guidelines. Animals were kept in a temperature-controlled room on a 12: 12 light-dark schedule with food and water ad libitum . D3H2LN cells (2×10 6 cells) were inoculated subcutaneously (s.c) in 4-week-old athymic nude mice (nu/nu, Janvier, France, n = 21). Then, mice were arbitrarily placed in control (n = 7) and in each BPs treated group (n = 7). The administration of BP7033Br or BP7033Br ALK (11 mg/kg) started 1 week after cell inoculation when tumors reached 30 mm 3 . The concentration of 11 mg/kg corresponds to about 250 µM for BP7033 ALK and 500 µM for BP7033 if considering the MW of both BPs and the mean mouse weight of about 30 g. Control group received 0.1 mL of PBS 1×. Treatments or PBS solutions administration were intratumoral twice a week for 4 weeks. Tumor volume was measured once a week along to major axes using callipers. Tumor volume (mm 3 ) was calculated as following: V = 4/3πR 1 2 R 2 where R 1 <R 2
Tolerability of the dose used (11 mg/kg) was evaluated according to the same protocol used for BPs injection on each tumor and tumor-free healthy mice (n = 5). Each animal was weighted once a week during the treatment.
Immunohistochemical analysis
Tumour specimens were fixed in 4% paraformaldehyde and embedded in to paraffin using standard procedure. Routinely, 5 µm-sections were stained in haematoxylin and eosin. For immunohistochemical studies the sections were deparaffinized and rehydrated. Endogenous peroxidase was inactivated with 3% H 2 O 2 and washed in TBS (Tris 0.05 M, NaCl 1.5 M, pH 7.6) followed by pre-incubation in 10% normal goat serum for 1 h at room temperature. Endothelial cells were specifically labelled with GSL-1 isolectin B4 (Vector Laboratories, Burlingame, CA, U.S.A.). The sections were incubated for 1 h with the 1∶50 diluted GSL-1 isolectin at room temperature. The sections were then incubated with goat antibody against GSL-1 isolectin B4 (1∶400 dilution, Vector Laboratories) for 30 min, washed with TBS and incubated with biotinylated rabbit anti-goat immunoglobulins (1∶400 dilution; Dako,Glostrup, Denmark) for 20 min in a moist chamber at room temperature. After three washes with TBS, the samples were incubated with streptavidin-biotin peroxidase (LSAB kit;Dako) for 10 min using diaminobenzidine tetrahydrochlorideas the chromogen. Finally, the slides were washed in water and counterstained with hematoxylin. Intratumor microvessel areas were determined as described previously [20] . For each tumour, 10 randomly selected non serial sections were studied.
Intracardiac experimental metastasis model
Female nude mice (8–10 weeks) were anasthetized by intraperitoneal injection of 120 mg/kg ketamine and 6 mg/kg xylazine on the day of injection and by exposure to 1–3% isoflurane on subsequent imaging days. On day 0, anaesthetized animals were injected with D3H2LN (1×10 5 cells) in 100 µl sterile PBS into the left ventricle of the heart by non surgical means. Anesthetized mice were placed in the IVIS™ Imaging System (Xenogen) and imaged from both dorsal and ventral views approximately 5 min after intraperitoneal injection of D-luciferin (Caliper Life Science). A successful intracardiac injection was indicated on day 0 by systemic bioluminescence distributed throughout the animal. Only mice with satisfactory injection continued the experiment and were treated by intraperitoneal injection of PBS (control) or with both BPs (11 mg/kg). Assessment of subsequent metastasis was monitored in vivo once a week by imaging for up to 3 weeks.
Bioluminescent Imaging
Bioluminescence images were acquired with the IVIS imaging system (Xenogen) at 5 min after intraperitoneal injection of D-luciferin to anesthetized animals as described above. Acquisition times of the beginning of the time course started at 5 min and were reduced in accordance with signal strength to avoid saturation. Analysis was performed using LivingImage software (Xenogen) by measurement of photon flux (photon/s/cm 2 ) with a region of interest (ROI) drawn around the bioluminescence signal.
Statistical analysis
Statistical significance was determined by the Student's t-test. P<0.05 was considered significant
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Chronic kidney disease (CKD) is a major risk factor for cardiovascular disease and all-cause mortality, placing a huge burden on the health care system [1] , [2] . CKD is a complex trait, regulated by interactions of several environmental and genetic factors [3] . CKD can arise as a consequence of diabetes, hypertension, immunologic disorders (such as lupus or primary glomerulonephritis), and a variety of other exposures. While the genetic causes of monogenic forms of renal diseases are well established, those contributing to the common forms of CKD are still largely unknown [4] . Recent meta analyses of genome wide association studies (GWAS) identified several loci for kidney function indices and CKD, collectively explaining a very small fraction of the variability of these traits, leaving most of their genetic components undetermined [5] . Earlier linkage analyses and candidate gene studies also produced inconsistent or unconfirmed results [6] – [9] . Most study populations were ascertained on the basis of disease status, resulting in enrichment for diabetes, obesity, and hypertension, thus complicating genetic studies of CKD [10] , [11] .
Genetic studies of CKD can benefit from a pathway-based targeted approach that includes testing for genetics and environmental interactions [12] for an intermediate quantitative trait [13] such as glomerular filtration rate (GFR) that is widely used to measure kidney function. GFR is a complex trait with an estimated heritability of 25–75% [14] . However, only about 1.5% of its variability has been explained by the genetic loci that have been identified so far [5] .
In the current study, we examined the genetic factors that may influence GFR in rural Chinese participants of the “Genetic Epidemiology Network of Salt Sensitivity” (GenSalt) study, who did not have clinical evidence of overt CKD. Estimated GFR (eGFR) values were calculated from serum creatinine measurements obtained during a three-day observation period preceding dietary intervention, while the participants consumed their usual diet. We tested single nucleotide polymorphisms (SNPs) in 26 genes from pathways of blood pressure regulation for association with eGFR measures. This genetic study of GFR provides insight into the genetic determinants of kidney function in a general population sample of individuals without CKD, and potentially into the initiation and progression of CKD.
Materials and Methods
Ethics statement
Institutional review boards at Tulane University Health Sciences Center, Washington University School of Medicine, University of Texas School of Public Health, Fu Wai Hospital and Chinese National Human Genome Center at Beijing, Chinese Academy of Medical Sciences approved the GenSalt study. Written informed consents for the baseline observation and for the intervention program were obtained from each participant.
Study population
The GenSalt Study was conducted in Han Chinese families living in six rural villages in Northern China. Families were recruited through 18–60 year old probands who were either prehypertensive or had stage-1 hypertension (SBP 130–160 mm Hg and/or DBP 85–100 mm Hg), but had never been treated with antihypertensive medication. Parents, spouses, siblings, and offspring were invited to participate in the study. Family members were excluded if they had stage-2 hypertension, a history of CVD, diabetes, or heavy alcohol consumption, or were pregnant, on a low sodium diet or taking anti-hypertensive medications. A total of 3,025 individuals in 631 families participated in this study. A large number of demographic, anthropomorphic, and medical variables were measured in GenSalt participants. More information regarding participants recruitment and measurements are available elsewhere [15] . Institutional Review Board approval for this study was obtained at all of the participating institutions and all study participants signed an informed consent document.
Phenotype measurements
Serum creatinine was measured during a 3-day baseline observation period while the study participants consumed their usual diet prior to a GenSalt dietary intervention. During this period, an overnight fasting blood sample was obtained from each participant by venipuncture. This was used to measure serum creatinine by the modified kinetic Jaffe reaction method. GFR was estimated using an amended formulation of the Modification of Diet in Renal Disease (MDRD) study equation, specifically designed for use in healthy individuals [16] : eGFR in mL/min per 1.73 m 2 = 216 × (serum creatinine in mg/dL) −0.490 × (age in years) −0.192 × 0.923 (if female).
Gene and SNP selection and genotyping
The GenSalt genotyping effort focused on 26 blood pressure candidate genes. The genes were selected based on their presumed role in blood pressure homeostasis and being a part of blood pressure regulation pathways such as the renin angiotensinsystem ( REN , RENBP , AGT , AT2R1 , AT2R2 , ACE ), the aldosterone system ( CYP11B1 , CYP11B2 , MLR , HSD11B1 , HSD11B2 , CYP3A5 ), the endothelial system ( EDN1 , NOS3 , SELE ), the sympathetic nervous system ( GRK4 , ADRB2 ), alternative renin angiotensinsystem pathway ( ACE2 , APLN , AGTRL1 ), as well as atrial natriuretic peptide genes ( NPR3 , NPPA ), sodium channels genes ( SCNN1B , SCNN1G ), and intracellular messengers genes ( GNB3 , ADD1 ). We selected 234 SNPs within these genes based on linkage disequilibrium (LD) structure in the Chinese population from the International HapMap project [17] . SNP genotyping was performed using the SNPlex platform (Applied Biosystems, Foster City CA) according to the manufacturer's protocol [18] . We excluded 41 SNPs in three genes from the association analysis due to low call rate (<80%), low minor allele frequency (MAF <0.05), or severe deviation from Hardy-Weinberg Equilibrium (HWE) (p<0.001). Detailed information concerning the remaining 193 SNPs within 24 genes is listed in Table S1 .
Statistical analysis
Plink and PedCheck programs were used to assess the Mendelian consistency of SNP genotype data [19] , [20] . Programs from the Affected-Sib-Pair Interval Mapping and Exclusion package (ASPEX) and the Graphical Representation of Relationships (GRR) package were used to check for potential misreported relationships within pedigrees [21] , [22] . Haploview (Broad Institute, Boston MA) was used for SNP descriptive statistics [23] . The Generalized Estimation Equation (GEE) method was used to test for associations between eGFR and the genetic variants, accounting for familial correlation [24] . Values for eGFR were adjusted for significant covariates including age, age 2 , age 3 , gender, BMI, high density lipoprotein cholesterol (HDL-C), hypertension, and field center. Neither smoking nor alcohol consumption was significantly associated with eGFR and were not included as covariates in statistical analysis. GEE analysis was performed with SAS 9.1 using proc genmod, and exchangeable working correlation matrix. False Discovery Rate (FDR) was used to correct for the multiple testing in GEE analysis [25] . For interactions among genes, we used the Generalized Multifactor Dimensionality Reduction program (GMDR) [26] to determine joint effects of each pair of SNPs. The best model identified by GMDR for eGFR was verified using GEE to account for familial correlation. Several web algorithms and data bases were used for SNP annotation and bioinformatics analysis including UCSC [27] , SNPnexus [28] , [29] , PolyPhen-2 [30] , SIFT [31] , and FastSNP [32] .
Results
Table 1 presents the basic characteristics of the 3,025 GenSalt participants who were included in this study. The study sample consisted of healthy free living members of three-generation families with an average age of 50 years, and approximately equal proportions of males and females. Their average values for BMI, HDL-C, serum creatinine, and eGFR were all within normal ranges. Table S1 shows characteristics of the 193 SNPs within 26 loci involved in blood pressure regulation that were chosen according to LD in the Han Chinese population (International HapMap project).
10.1371/journal.pone.0092468.t001 Table 1
Basic characteristics of the study subjects.
Healthy subjects
N
3025
Age
50.0± 16.6
Male %
51.3
Hypertension %
17
Smokers %
34.5
Alcohol consumers %
27
Body Mass Index
23.1 ± 3.2
HDL (mg/dL)
51.47 ± 11.3
Serum creatinine (mg/dL)
0.93 ± 0.21
eGFR (mL/min per 1.73 m 2 )
104.91 ± 14.07
HDL: High Density Lipoprotein, eGFR: estimated Glomerular Filtration Rate. Values are mean ± standard deviation for Age, Body Mass Index, HDL, serum creatinine, and eGFR.
Table 2 presents results of genetic analysis that identified 12 SNPs in six genes that showed significant associations with eGFR (p<0.05). The four significant SNPs in ACE were highly correlated with r 2 ranging from 0.78 to 0.96. The three SNPs in the hydroxysteroid 11-beta dehydrogenase 1 gene ( HSD11B1 ) were less correlated (r 2 from 0.35 to 0.43), and the two significant SNPs in the alpha-adducin gene ( ADD1 ) were not significantly correlated (r 2 = 0.17). The FDR for all significant p-values was 0.59, which means that seven of these 12 significant SNPs might have been false positives. Figure 1 shows the average eGFR values for each of the three genotypes for nine of the significant SNPs (excluding three highly correlated ACE SNPs). Three of the significant SNPs (AGT_rs4762, GRK4_rs2488815, and SCNN1G_rs4299163) can be considered as risk SNPs where the additive effect of the minor allele was associated with lower values of eGFR. The other significant SNPs can be considered as protective, where the minor allele was associated with higher eGFR values.
10.1371/journal.pone.0092468.g001 Figure 1
Mean eGFR values and standard errors for genotypes of SNPs that showed significant associations (0, homozygotes for common alleles; 1, heterozygotes; 2, homozygotes for minor alleles).
10.1371/journal.pone.0092468.t002 Table 2
SNPs that showed significant associations with eGFR in GenSalt participants.
Gene
Chr.
SNP
Region
HWpval
Call Rate
MAF
Maj/Min
P *
ACE
17
rs4316
exon
0.0917
97.2
0.353
T/C
0.0077
ACE
17
rs4343
exon
0.4012
92.7
0.354
A/G
0.0170
ACE
17
rs4353
intron
0.2622
97.6
0.393
G/A
0.0181
ACE
17
rs4331
exon
0.2927
96.2
0.353
G/A
0.0313
ADD1
4
rs3775067
intron
0.934
97.1
0.34
C/T
0.0061
ADD1
4
rs12503220
utr
0.3043
93.1
0.134
G/A
0.0231
AGT
1
rs4762
exon
0.3746
95.6
0.073
C/T
0.0051
GRK4
4
rs2488815
intron
0.0255
96.7
0.206
C/T
0.0279
HSD11B1
1
rs4844880
utr
0.6649
89.5
0.356
T/A
0.0166
HSD11B1
1
rs2235543
utr
0.0703
92.5
0.35
C/T
0.0308
HSD11B1
1
rs846908
intergenic
0.9629
92.5
0.251
G/A
0.0419
SCNN1G
16
rs4299163
intron
0.8433
96.6
0.103
G/C
0.0247
HWpval: Hardy-Weinberg p value, MAF: minor allele frequency, Maj/Min: Major/Minor allele. *p values adjusted for age, age 2 , age 3 , gender, BMI, high density lipoprotein cholesterol (HDL-C), hypertension, field center, and family structure.
Figure 2 shows the cumulative effects on mean adjusted eGFR values for carriers of the minor alleles for all nine significant SNPs ( Panel A ) and separately for the six protective alleles ( Panel B ) and three risk alleles ( Panel C ). Carriers with increasing numbers of the minor protective alleles had up to more than 4 mL/min per 1.73 m 2 higher mean eGFR values (p = 0.001) ( Figure 2 , Panel B ). Carriers with increasing numbers of the minor risk alleles had mean eGFR values that were as much as almost 3 mL/min per 1.73 m 2 lower (p = 0.006) ( Figure 2 , Panel C) .
10.1371/journal.pone.0092468.g002 Figure 2
The cumulative effect of the minor alleles in all of the 9 significant SNPs (Panel A), the 6 protective SNPs (Panel B), and the 3 risk SNPs (Panel C) on the value of the mean adjusted eGFR.
The best fitting trend line and p value from t-tests between those with no minor allele and those with the largest possible number of minor alleles in each category are shown.
We also tested the SNPs for effects of gene by gene interactions (GxG) on eGFR. We identified a joint effect on eGFR between a nonsynonymous SNP in the gene for cytochrome P450, family 11, subfamily B, polypeptide 1 (CYP11B1_rs4541, Ala386Val) and a synonymous SNP in the beta-2-adrenergic receptor gene (ADRB2_rs1042718). Figure 3 shows the joint effects on mean eGFR values of interactions between CYP11B1_rs4541 and ADRB2_rs1042718 genotypes. The mean adjusted eGFR value in homozygotes for the ADRB2_rs1042718 minor allele (AA) depended on their genotype for CYP11B1_rs4541. Homozygotes for the minor allele of ADRB2_rs1042718 (AA) who are also homozygous for the major allele of CYP11B1_rs4541 (CC) had the lowest mean eGFR values. Homozygotes for the minor allele of ADRB2_rs1042718 (AA) who are heterozygous for CYP11B1_rs4541 (CT) had the highest eGFR. The difference between these two joint genotypes was approximately 5 mL/min per 1.73 m 2 .
10.1371/journal.pone.0092468.g003 Figure 3
Mean adjusted eGFR as a result of the genotypic interaction between CYP11B1_rs4541 and ADRB2_rs1042718.
The data points represent the eGFR for the nine possible combinations of the three ADRB2 genotypes versus each of the three possible CYP11B1 genotypes. The number of individuals at each point is provided.
Discussion
The overall goal of this study was to conduct a comprehensive examination of the effects of variability in genes from pathways of blood pressure regulation on renal GFR. The GenSalt study cohort was comprised of rural Han Chinese villagers to minimize the genetic heterogeneity that is encountered in most association studies that are conducted in admixed urban populations. None of our study participants were taking antihypertensive medication, so the complexity associated with the antihypertensive drugs is absent from our study. In addition, we employed an amended version of the MDRD eGFR equation that was specifically designed for use in healthy free living individuals and eliminated the underestimation of GFR with the equation that was previously in use [16] .
In our analyses of individual SNPs, the Thr207Met polymorphism (rs4762) in the angiotensinogen gene ( AGT ) showed the strongest association with eGFR ( Table 2 ). AGT plays a role in the renin-angiotensin system (RAS), a primary pathway in blood pressure regulation with strong influences on cardiovascular and renal disease. AGT encodes preangiotensinogen in the liver, which is subsequently cleaved by renin to generate angiotensin I. Angiotensin I converting enzyme ( ACE ), converts angiontensin I to angiotensin II [33] , [34] , a potent vasoconstrictor that also affects renal hemodynamics by decreasing renal cortical blood flow, total renal plasma flow, urinary sodium excretion, and GFR [33] , [34] . Moreover, angiotensin II increases glomerular capillary pressure, potentially contributing to glomerulosclerosis [35] , [36] . AGT_rs4762 (Thr207Met) is a probably damaging SNP [30] , [31] as it substitutes a non-polar amino acid (methionine) for a polar amino acid (threonine). Furthermore, threonine at this position in AGT is highly conserved among divergent species ranging from human to zebrafish [37] . Previous studies in Asians have identified associations of AGT_rs4762 with diabetic nephropathy in Taiwanese patients [38] and hypertension in different Asian populations based on meta-analysis [37] .
We also found significant associations with eGFR for four correlated SNPs in the ACE gene, another key player in the RAS pathway of blood pressure regulation. Three of these SNPs (ACE_rs4316, ACE_rs4331, ACE_rs4343) are exonic variants, but do not cause amino acid substitutions (synonymous SNPs). ACE_rs4343, was previously reported to be significantly associated with diabetic nephropathy in an Asian Indian population [39] .
Our analysis identified a significant association of eGFR with two intronic SNPs (rs3775067 and rs12503220) in the adducin 1 gene ( ADD1 ). ADD1 encodes the alpha subunit of the cytoskeleton protein adducin, which plays an important role in hypertension and renal function via sodium homeostasis [40] . Many previous studies have reported associations of ADD1 variants with hypertension, renal functions and renal diseases, in different populations including Chinese [40] – [48] .
Another SNP that showed associations in our study was rs2488815, an intronic SNP in the G protein_coupled receptor kinase 4 ( GRK4 ) gene, a major player in sodium homeostasis and blood pressure regulation [49] . GRK4 is expressed in the renal proximal tubule, where about 70% of renal sodium reabsorption takes place. Increased GRK4 activity leads to decreased dopamine signaling and increased AngII receptor expression and function, both of which increase sodium retention and blood volume which ultimately leads to hypertension [50] , [51] . GRK4 variants have been shown to be associated with hypertension and blood pressure traits in different populations including Han Chinese [52] .
Three SNPs in the hydroxysteroid 11beta dehydrogenase1 gene ( HSD11B1 ) were associated with eGFR in the current study. HSD11B1 is a NADP dependent enzyme that functions in the proximal tubule and medullary interstitial cells of the human kidney. HSD11B1 plays a role in the metabolism of the endogenous glucocorticoids, which in turn modulate sodium homeostasis, renal blood flow, and GFR [53] . HSD11B1 enzymatic activities are thought to be involved in obesity, hypertension, and other components of the metabolic syndrome. HSD11B1 overexpression in mice has been associated with dose-dependent hypertension and AGT expression in liver [54] – [56] .
Our analyses of individual SNPs identified associations of eGFR with an intronic SNP (rs4299163) in the gene encoding the gamma subunit of the epithelial sodium channel gene ( SCNN1G ). Epithelial sodium channels ( ENaC ), are the main regulators for sodium transport in the kidney [57] , [58] , and rare variants in SCNN1G cause Liddle Syndrome, a monogenic form of hypertension [59] . Other variants in SCNN1G cause pseudohypoaldosteronism type 1, a rare inherited form of renal tubular acidosis [60] , [61] . Many studies have identified linkage of SBP with the region that contains SCNN1G on chromosome 16 [62] . A fine mapping study of this region detected associations of SBP with three SCNN1G intronic SNPs, including rs4299163 [62] .
In addition to analyses of individual SNPs, we tested for interactions among genes (GxG interactions) that influence eGFR. Our GxG analyses identified a joint effect on eGFR between a conservative nonsynonymous SNP rs4541 (Ala386Val) in CYP11B1 and a synonymous SNP rs1042718 in ADRB2 . CYP11B1 is one of the cytochrome p450 genes encoding 11β hydroxylase, a protein involved in the synthesis of cortisol in the adrenal cortex [63] . Cortisol is associated with Cushing's syndrome, hypertension of chronic renal failure, hypertension related to low birth weight, and essential hypertension [64] , [65] . Glucocorticoid-Remediable Aldosteronism, a rare form of hypertension, is caused by a gene fusion between CYP11B1 and CYP11B2 [66] . ADRB2 encodes the beta 2 adrenergic receptor, a member of the G-protein superfamily receptors, playing a role in metabolism regulation and also in blood pressure regulation by mediating vasodilation and vascular resistance [67] , [68] . ADRB2 SNPs were previously found to be associated with hypertension and blood pressure traits in different populations including Han Chinese but many of the results are inconsistent [68] – [70] . ADRB2_rs1042718, the SNP showing significant GxG interaction, is part of a haplotype recently found to be associated with weight, insulin, and homeostasis model assessment (HOMA) score in Korean adolescents [71] . The exact mechanism of interaction between these two coding region SNPs in relation to eGFR warrants further investigation.
We compared the results of our association study with a recently published meta-analysis of kidney function traits in several East Asian populations, including GenSalt [72] . The SNPs with significant associations in our study ( Table 2 ) were not significantly associated with eGFR in the meta-analysis. This discrepancy could stem from differences in the study populations since we included only Han Chinese from Northern China, while the meta-analysis included Japanese, Malay, Indian, Korean, and Chinese. GenSalt participants were relatively healthy free living individuals from three-generation families, while the meta-analysis included hospital and population based cohorts with no exclusion of diseased individuals. In addition, we employed an equation to calculate eGFR that was specifically designed for use in healthy individuals [16] , while the meta-analysis used a different equation specific for Japanese individuals [73] .
The genes included in this study were all in pathways known to be involved in regulation of blood pressure which may play an important role in regulating kidney function. However, there are established overlaps, such as the RAS pathway which is involved in renal function decline since treatment of patients with renin-angiotensin inhibitors slows the progression of kidney disease [74] – [76] . Such protective effect of RAS blocking on kidney function may be through both blood pressure and non-blood pressure dependent mechanisms. Our study was not designed to determine whether these genes impact GFR independent of their effects on blood pressure, or whether their effects on GFR are linked to the same pathophysiologic cascades as their involvement in hypertension. However, their significant association with kidney function remains after adjustment for hypertension.
In conclusion, we have identified common variants in genes from pathways of blood pressure regulation and their interactions that influence kidney function, providing new insights into the genetic determinants of kidney function. A longitudinal association between these common variants and changes in kidney function remains to be investigated.
Supporting Information
Table S1
SNP characteristics. Detailed information about the 193 SNPs used in analysis.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The morphologic assessment of renal biopsies is a well-established practice and provides important diagnostic and prognostic information. Standardized assessment and reporting for these specimens is essential [ 1 ]. Although most renal biopsy diagnoses are dependent on qualitative features, semi-quantitative metrics are routinely applied for prognostic measures and treatment decisions. Semi-quantitative approaches were first introduced in lupus nephritis and emphasized the detailed evaluation of all lesions in glomerular and tubulointerstitial compartments [ 2 – 4 ]. Subsequently, similar approaches to systematic semi-quantitative biopsy interpretation have been applied to other renal diseases [ 5 – 9 ]. Thus, accuracy in reporting quantitative and semiquantitative values is emphasized by consensus papers for routine renal pathology reporting [ 1 , 10 ]. Renal pathologists report glomerular numbers in renal biopsies utilizing various approaches, including a) attempting assessment of the total number of glomeruli by looking at all sections available, b) averaging the number of glomeruli per section either counting all glomeruli in each levels as individual glomeruli, or by using the level with the most and the least, c) providing a range using the sections with the least and the most glomeruli. It is, however, appreciated that there are technical limitations independent of the individual pathologists experience and accuracy.
The Nephrotic Syndrome Study Network (NEPTUNE) implemented a digital pathology repository (DPR) that includes whole slide images (WSI) of all kidney biopsy levels available for each case and a copy of the de-identified pathology report [ 11 ]. The NEPTUNE digital pathology protocol (NDPP), by application of glomerular annotation in discontinuous but sequential sections, sought to standardize reporting metrics by refining accuracy. The annotation and enumeration of glomeruli in all available WSI sections enables not only the estimation of the overall number of glomeruli, but also of affected glomeruli by any parameter of choice [ 12 ].
Here, we applied WSI and annotation software to evaluate the most basic quantitative metrics: total number of glomeruli and number/percentage of globally sclerotic glomeruli. Additionally, by comparing the digital quantitative analysis to the manual light microscopy (LM) analysis as routinely reported, we test its potential value in improving clinical practice.
Methods
NEPTUNE DPR and case selection
Renal biopsies are an enrollment criteria of the NEPTUNE study (ClinicalTrials.gov Identifier: NCT01240564). All patients were consented at enrollment, and the study was approved by Institutional Review Boards of all participating institutions ( http://www.rarediseasesnetwork.org/cms/neptune ).
According to the NEPTUNE protocol, all renal biopsies are digitized and together with the de-identified PDF of the pathology report, stored in the NEPTUNE DPR [ 11 , 13 ]. From a total of 392 digital renal biopsies, 317 cases with a diagnosis of minimal change disease (MCD), focal and segmental glomerulosclerosis (FSGS), membranous glomerulopathy (MN), and IgA nephropathy (IgAN) were eligible for inclusion in this study. Other glomerular diseases were excluded. The pdf of the pathology report was reviewed by two pathologists to determine whether the number of glomeruli and the number of globally sclerotic glomeruli (GS) present in the biopsy was documented. Cases where the pathology report was missing the glomerular number assessment (40/317), either as a total number of glomeruli per biopsy as determined by an assessment across all histology sections, or as an average number of glomeruli per histology level, were excluded. A total of 277/317 cases were selected, distributed across the following diagnosis: 79 MCD, 108 FSGS, 54 MN, and 36 IgAN ( Fig 1 ), all of which contain contained documentation of the the assessment of GS. No glass slides were available for re-review as part of this study.
10.1371/journal.pone.0156441.g001
Fig 1
Summary of workflow for case selection and annotation using the NEPTUNE Digital Pathology repository.
Glomerular counting
To test the value of annotation for glomerular enumeration and counting, ten pathologists annotated glomeruli in all WSI levels available by remotely accessing the 277 cases stored in the NEPTUNE DPR ( Fig 1 ). As previously described, annotation of glomeruli was achieved by visualizing and aligning up to 4 WSI simultaneously (DIH, Leica, Dublin IR) and enumerating each glomerulus with a unique number that was maintained in all levels examined [ 11 ]. ( Fig 2 ) Following initial annotation, each case underwent quality control review for accuracy of annotation by a different pathologist. Two of the ten pathologists independently retrieved the total number of annotated glomeruli in all 277 cases by remotely accessing each biopsy WSI section/level in the NEPTUNE DPR and recording the highest number used for glomerular annotation.
10.1371/journal.pone.0156441.g002
Fig 2
Digital review and annotation of whole slide images in the NEPTUNE digital pathology repository.
A) Example of multiple levels visualized at the same time. The software allows overlapping and orientation of the sections to facilitate multilevel reconstruction and representation of the biopsy levels. B) Digital annotation of whole slide images. Two levels are shown in alignment. Glomeruli in the left panel are annotated in red. In a subsequent level (right panel), two of the red-annotated glomeruli are still present (3 and 4), while one has disappeared (6). Three additional glomeruli, including one adjacent to glomerulus 3 and 6 have appeared in the deeper level (annotated in blue).
To investigate concordance between number of glomeruli counted on annotated WSI and number of glomeruli reported in the anonymized pathology reports, reported glomerular count was also recorded by two pathologists either as the absolute total number of glomeruli (139 cases) or the average of the number of glomeruli per histology level (138 cases). The number of annotated glomeruli was then compared to the number of reported glomeruli as follow:
a) for the 139 cases where the total number of glomeruli were recorded in the pathology report, the total number of reported glomeruli per biopsy was compared to the total number of annotated glomeruli counted on the corresponding WSI; b) for the 138 cases where the average number of glomeruli per histology level was reported, this average was compared to the average number of annotated glomeruli per WSI level per biopsy, calculated by dividing the total number of annotated glomeruli per number of WSI levels in each case.
To investigate the impact of annotation on assessment of globally sclerotic glomeruli, two pathologists counted the annotated GS by remotely accessing the NEPTUNE DPR and each biopsy WSI section/level, and recorded the total number or average number of reported GS by reviewing the report pdf. The number and percentage of annotated GS were compared with those recorded in the pathology report. Discrepancy for GS between the two pathologists were resolved by webinar consensus review. The percentages of annotated globally sclerotic glomeruli from WSI and from the pathology reports were calculated using annotated (total or average) and reported number of glomeruli, respectively.
Statistical methods
Descriptive statistics including mean, range, median and interquartile range (IQR). Wilcoxon signed rank tests were used to test the difference in total number and number of globally sclerotic glomeruli from annotated WSI versus given in pathology reports. The Wilcoxon test was used due to non-normality of the paired differences of annotated versus reported glomeruli. The analysis was performed separately for cases with pathology reports giving total versus average numbers of glomeruli, and presented both overall and with stratification by diagnosis (FSGS, MCD, MN, and IgANP).
To estimate the differences between annotated and reported counts and percentages with global sclerosis by categories of number of glomeruli or percent globally sclerotic glomeruli, linear regression models were used. The outcomes were the differences between annotated and reported values, and glomeruli categories were the covariates. All analyses were conducted in SAS software V9.4 (SAS Institute Inc., Cary, NC, USA).
Results
Assessment of annotated glomeruli on WSI
A total of 9,379 annotated glomeruli were counted, of which 1,322 were classified as globally sclerotic. Across all four diagnostic categories, the average number of glomeruli per biopsy was 33 with a range of 3 to 164. Percent globally sclerotic averaged 16.7%, with a range of 0 to 100.
Comparison of total number of annotated versus reported glomeruli
For the 139 cases where the total number of discrete glomeruli was given on the pathology report, the total number of annotated glomeruli per case was higher on average than the total number of glomeruli reported (Annotated: mean glomeruli per case = 30.7; range 3–149; median = 23; IQR 13–43. Reported: mean glomeruli per case = 18.3; range 2–71; median = 16; IQR 10–25; p<0.01) ( Table 1 ). The difference between annotated and reported numbers of glomeruli increased as the number of glomeruli increased, with a ratio approaching 2:1 ( Fig 3A ). This relationship is also shown in Table 2 , where the increase of annotated over reported glomeruli is proportional to the number of reported glomeruli.
10.1371/journal.pone.0156441.t001
Table 1 Number of annotated versus reported glomeruli stratified by diagnosis.
Total
Average
Reported
Annotated
Significance *
Reported
Annotated
Significance *
Overall N = 139
Range
2–71
3–149
<0.01
Overall N = 138
Range
4–125
4–164
<0.01
Mean
18.3
30.7
Mean
21.4
36.0
FSGS N = 40
Range
5–71
5–149
<0.01
FSGS N = 68
Range
4–83
4–164
<0.01
Mean
18.7
29.9
Mean
20.1
35.2
MCD N = 33
Range
4–51
5–106
<0.01
MCD N = 46
Range
4–125
7–163
<0.01
Mean
20.2
38.8
Mean
25.8
42.9
MN N = 37
Range
2–45
3–74
<0.01
MN N = 17
Range
8–34
8–55
<0.01
Mean
17.1
26.8
Mean
19.3
28.2
IGA N = 29
Range
4–38
3–65
<0.01
IGA N = 7
Range
5–21
7–44
<0.01
Mean
16.9
27.6
Mean
13.4
24.0
Total: n = 139 cases and Average: n = 138 cases.
*Wilcoxon signed rank test
10.1371/journal.pone.0156441.t002
Table 2 Difference in glomerular count of number of annotated versus reported glomeruli stratified by number of glomeruli.
Total
Average
# glom
# cases
Mean Δ (annot—report)
Significance
# glom
# cases
Mean Δ (annot—report)
Significance
1–10
37
3.8[0.1,7.5]
0.05
1–10
27
5.7[0.6,10.9]
0.03
11–20
55
9.0[5.9, 12.0]
<0.01
11–20
51
13.5[9.8,17.3]
<0.01
21–30
29
20.1[15.9, 24.3]
<0.01
21–30
35
15.3[10.7,19.9]
<0.01
31–40
12
23.3[16.8,29.9]
<0.01
31–40
17
15.8[9.3,22.3]
<0.01
>40
6
38.8[29.6,48.1]
<0.01
>40
8
46.5[37.0,56.0]
<0.01
Total: n = 139 cases and Average: n = 138 cases.
10.1371/journal.pone.0156441.g003
Fig 3
Digital annotation of whole slide images improves accuracy of glomerular enumeration.
(A,B) Scatter plot of annotated glomerular number versus reported glomerular number by diagnosis in cases reporting (A) total number of glomeruli or (B) average number of glomeruli per level. Dotted diagonal line shows where annotated = reported, and highlights the increase in number of glomeruli found by annotation. (A: n = 139; B: n = 138). (C,D) Box plot of annotated glomerular number and reported total glomerular number stratified by diagnosis in cases reporting (C) total number of glomeruli (n = 139) or (D) average number of glomeruli (n = 139).
Comparison of average number of annotated versus reported glomeruli
Similarly, for the 138 cases where the average number of glomeruli per level was given on the pathology report, the average numbers of annotated glomeruli per level were higher than the average numbers given on pathology reports (Annotated: mean = 36; range 4–164; median = 30; IQR 19–46. Reported: mean = 21.4; range 4–125; median = 16; IQR 12–28; p<0.01)( Table 1 ). Again, the approximate 2:1 ratio is seen in the scatterplot between annotated and reported average glomeruli ( Fig 3B ). The increase in the difference with increasing categories of reported glomeruli ( Table 2 ) mirrors the effect seen with total glomeruli above.
Stratification by center and disease category
Differences in total and average annotated and reported number of glomeruli were maintained across the 25 centers (data not shown) and all disease categories. ( Fig 3C and 3D ; Table 1 )
Comparison of annotated versus reported globally sclerotic glomeruli
In the 139 cases where the percentage of globally sclerotic glomeruli was obtained from the reported total number of glomeruli, there was no significant difference in the mean globally sclerotic glomeruli percentage determined on annotated WSI (14.4%) and the pathology reports (13.7%) ( Table 3 , Fig 4A ). In contrast, in cases reporting an average glomerular number (138 cases), annotation yielded a greater mean globally sclerotic glomeruli percentage compared with the corresponding pathology reports (17.9% vs 15.8%, p<0.01) ( Table 3 , Fig 4B ). When cases were stratified by disease ( Fig 4C and 4D ; Table 3 ), this discrepancy was statistically significant only in FSGS (29.8% vs 25.8%, p = <0.01). While the overall discrepancy was small in the percentage of globally sclerotic glomeruli between annotated WSI and pathology reports, in cases with > 40% globally sclerotic glomeruli there was a significant underestimation in the biopsy reports by both total and average methodologies (p < 0.01) ( Fig 4A and 4B ; Table 4 ); this effect was evident in FSGS and IGA, particularly for the biopsy reports using average glomeruli ( Fig 4C and 4D ).
10.1371/journal.pone.0156441.t003
Table 3 Number of annotated versus reported percent globally sclerotic (%GS) stratified by diagnosis.
Total
Average
Reported %GS
Annotated %GS
Significance *
Reported %GS
Annotated %GS
Significance *
Overall N = 139
Range
0–80
0–100
0.71
Overall N = 138
Range
0–100
0–100
<0.01
Mean
13.7
14.4
Mean
15.8
17.9
FSGS N = 40
Range
0–80
0–65.6
0.99
FSGS N = 68
Range
0–100
0–100
<0.01
Mean
16.6
16.0
Mean
25.8
29.8
MCD N = 33
Range
0–25
0–21.7
<0.01
MCD N = 46
Range
0–15.3
0–14.3
0.81
Mean
3.3
2.1
Mean
1.2
1.3
MN N = 37
Range
0–66.7
0–75
0.65
MN N = 17
Range
0–28.6
0–24.1
0.91
Mean
9.1
8.8
Mean
5.8
6.1
IGA N = 29
Range
0–69.2
0–100
0.02
IGA N = 7
Range
0–68.8
0–72.7
0.69
Mean
27.4
33.3
Mean
37.2
40.6
Total: n = 139 cases and Average: n = 138 cases.
*Wilcoxon signed rank test
10.1371/journal.pone.0156441.t004
Table 4 Difference in %GS glomeruli between annotated versus reported stratified by number of glomeruli.
Total
Average
%GS
# cases
Mean Δ (annot—report)
Significance
%GS
# cases
Mean Δ (annot—report)
Significance
0–19%
103
-0.80[-22.5,0.1]
0.30
0–19%
92
-0.20 [-1.4, 1.1]
0.77
20–39%
17
-0.24[-38.4, 3.4]
0.90
20–39%
15
3.0 [-0.2, 6.2]
0.07
40–59%
10
6.5[1.8,11.2]
<0.01
40–59%
17
7.5 [4.5, 10.6]
<0.01
60–100%
9
12.8[7.7,17.9]
<0.01
60–100%
14
10.3 [7.0, 13.7]
<0.01
Total: n = 139 cases and Average: n = 138 cases.
10.1371/journal.pone.0156441.g004
Fig 4
Digital annotation of whole slide images modestly improves accuracy of global glomerular sclerosis enumeration.
(A,B) Box plot of annotated percent (%) globally sclerotic glomeruli and reported %globally sclerotic glomeruli in cases reporting (A) total number of glomeruli (n = 139) or (B) average number of glomeruli (n = 138). (C,D) Box plot of difference between number of annotated and reported % globally sclerotic glomeruli stratified by % globally sclerotic glomeruli in cases reporting (C) total number of glomeruli (n = 139) or (D) average number of glomeruli (n = 138).
Discussion
Glomerular counting is a fundamental metric for renal biopsy providing the denominator for glomerular involvement in primary and secondary kidney diseases. In this study we evaluated the NEPTUNE protocol for glomerular annotation to provide a standardized and reliable estimate of number of glomeruli. Although Whole Slide Imaging is not approved for primary diagnosis in routine pathologic practice in the US [ 14 ] it is an enabling technology for the development and validation of evaluation of information gained from light microscopy. WSI with and without annotation has been demonstrated to enhance reproducibility of pathology assessment compared to glass slide review [ 15 – 17 ].
The application of digital annotation facilitates accuracy for renal pathology metric that cannot be achieved by conventional LM. The accuracy of determining the number of glomeruli in a given biopsy by LM is limited by visual-spatial memory of the pathologist across multiple histologic levels. This issue is resolved by the digital software allowing visualization and alignment of multiple levels at the same time, facilitating identification of the same or different glomeruli thru the biopsy. Thus, we found significant differences between the number of glomeruli reported for routine patient care and obtained by LM evaluation, and the number of glomeruli identified by digital annotation. We also demonstrated that discrepancy in glomerular number increases proportionally to the number of glomeruli present within a biopsy, demonstrating the limitation of manual light microscopy to accurately determine the number of glomeruli present. In contrast, the annotated and reported percentages of GS were generally more similar. However, discrepancies were found within disease subgroups and particularly for cases with percent GS over 40%, with annotated values showing higher percent sclerosis in all cases. These data suggest that up to a certain threshold, the discrepancy between annotated and conventional approaches are probably minimally critical for prognostic factors, but may become relevant for biopsies rich in glomeruli and with significant sclerosis. Glomerular density and percentage of GS have been discussed as a risk factor in disease progression [ 18 ], IgA [ 19 ], risk of progression in membranous glomerulonephritis [ 20 ] and obesity-related glomerulopathy [ 21 ]. Thus, accurate assessment with improved numerators and denominators is essential in improving prognostication.
Accurate glomerular counting is fundamental to determining biopsy adequacy and improved prognostication. It is anticipated that with digital enumeration of glomeruli, the binomial distribution of abnormal glomeruli (i.e., with segmental lesions), and thus the number of glomeruli needed for accurate sampling, will be greater than previously suggested [ 22 ].
Although the ultimate proof of the increased accuracy of digital annotation versus conventional light microscopy would be provided by direct comparison of glomerular number estimate on glass slide and annotated WSI, glass slides were not available and we had to rely on values extracted from the pathology report.
Our study demonstrates the limits of the conventional approach to microscopic examination of tissue in tasks that require enumeration and integration across multiple tissue sections. We have carried out this study in the context of glomerular disease that is a part of the NEPTUNE protocol. The data demonstrates that low frequency events can be accounted for with relative accuracy, as demonstrated by both the overall glomerular enumeration as well as percent global sclerosis data. However, as the number of events increases, the inaccuracy of the estimation increases. With the glomerular number, this reached a level of 2:1 when WSI and manual microscopy are compared. For percent glomerular sclerosis, inaccurate determinations reached statistical significance at 40%GS. These findings suggest that quantization of the histopathology of glomerular disease by an intensive approach, utilizing WSI and piecemeal evaluation of individual glomerular profiles will result in new insights to disease classification, prognostic markers, and potential markers of response to therapy.
These findings suggest the accuracy of detection is object specific, and that findings for one object may not generalize to other objects. Although this study was performed within the limited context of primary glomerular disorders, the findings have broad implications for many observer-based task where integration across image planes is required. The evaluation of renal biopsies for other disorders, including diabetes, SLE and transplant are obvious fields of inquiry, as are renal biopsies of medical conditions of the liver, where portal triads and other features are routinely enumerated. The overall implications are that there are some conventional approaches that may be enhanced by the adoption of computer-aided diagnostic tools. Manual review and enumeration of renal biopsies with current tools is time consuming and complex, however the development of tools to aid the reviewer in a value-added strategy are underway.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Stereotactic ablative radiotherapy (SABR), or stereotactic body radiotherapy (SBRT), has been shown to be an excellent treatment option for early-stage non-small cell lung cancer (NSCLC) and lung metastases when a biologically effective dose (BED) of ≥100 Gy 10 is delivered [1] – [4] . However, treatment related death from severe pulmonary toxicities, hemoptysis, or esophagitis, has been reported when various dose fractionation schedules were delivered to treat centrally located lesions [5] – [9] . This happened mainly when a large fractional dose has been delivered, leading to the overdosing of the organs at risk (OARs) adjacent to the tumor target. Therefore, respecting the OAR dose constraints is essential when treating central lesions close to the mediastinal structures to avoid potentially catastrophic consequences [10] .
Excellent OAR sparing has been routinely achieved through intensity modulated radiotherapy (IMRT), which generates highly conformal dose avoidance of structures immediately adjacent to the tumor target in various sites [11] , [12] . More recently, advanced techniques of IMRT delivery under image guidance, helical tomotherapy (HT) [13] and volumetric modulated arc therapy (VMAT) [14] , [15] , have been shown to produce more conformal dose distribution, and better OAR sparing when compared to IMRT, or three dimensional conformal radiotherapy (3D-CRT) at various sites [16] – [24] . Thus, HT & VMAT may be more suitable when treating centrally located lung lesions with SABR.
Previously, we have demonstrated the feasibility of HT-based SABR for centrally located lung lesions which are very close to critical OARs in the thorax; while VMAT has been shown to be superior to IMRT or 3D-CRT for lung SABR in OAR sparing [22] – [25] . In this dosimetric study multi-arc VMAT and HT and compared directly in their ability to maximally spare immediately adjacent OARs when SABR is delivered to the treat centrally located lung lesions. In addition, potential benefits of increasing the number of arcs for VMAT-based SABR in this setting are explored. In this study, 7 Gy × 10 fractions was investigated because it was associated with an excellent toxicity profile when bulky tumors were treated, and the clinically sound BED achieved with this dose fractionation schedule (119 Gy 10 ) [26] .
Methods
Patient and Tumor Characteristics
This study has been approved by the institutional review board (IRB) at the University of Arizona. Since no actual human subjects were involved, no informed consent was needed per IRB. A total of 12 patients with centrally located lesions have been randomly identified. These patients had undergone 3D or intensity-modulated SBRT for stage I-II NSCLC or metastasis to the lung in the department of radiation oncology at the University of Arizona. Central location is defined as the area within 2 cm of the proximal bronchial tree, which includes the lower trachea, carina, mainstem bronchi, and the lobar bronchi. The critical structures are the esophagus, the heart, the spinal cord, major vessels in the mediastinum, and the major airway (lower trachea, carina, mainstem bronchi, and lobar bronchi). The tumor location, size, and its immediately adjacent OARs in each case are listed in Table 1 .
10.1371/journal.pone.0059729.t001
Table 1
Patient tumor characteristics.
Patients
Location
PTV volume (cc)
Immediately adjacent structures
PTV to structure distance (cm)
1
RLL
153.68
Heart
0.15
2
LUL
70.54
Aortic arch
1.06
L pulmonary artery
0.11
3
RUL
69.22
Heart
0.50
SVC
0.23
4
RML
14.04
Heart
1.22
R middle lobar bronchus
0.10
5
RUL
56.52
R mainstem bronchus
0.20
R pulmonary artery
0.23
6
RUL
34.91
R brachiocephalic artery
0.19
7
LUL
22.79
Aorta
0.11
8
RUL
133.64
Esophagus
1.30
SVC
0.26
Trachea
1.06
9
RUL
147.22
Heart
1.54
R pulmonary artery
0.39
10
RLL
65.76
Esophagus
0.53
Heart
0.13
R pulmonary artery
0.22
11
LUL
24.11
Aortic arch
0.10
12
RML
22.69
Heart
0.23
R middle lobar bronchus
0.14
R pulmonary vein
0.97
Target Volume Delineation
The gross tumor volume (GTV) was delineated at the lung window level on the treatment planning CT. The clinical target volume (CTV) was defined as the GTV and its immediately adjacent areas which were felt to be at a high risk for microscopic disease extension. The planning target volume (PTV) was the CTV with a 5 mm expansion to account for set up errors and residual tumor motion. Particular attention was paid to avoid overlapping any target volumes with the OARs. In cases for which 4D CT was available, internal target motion was accounted for by 4D CT simulation. The lungs, esophagus, spinal cord, and the heart were contoured for each patient. The major vessels and major airway were contoured only when they are adjacent to the GTV. All the target delineation was performed in the clinical Pinnacle treatment planning system, version 9.0 (Philips Medical Systems, Bothell, WA).
Treatment Planning
Tomotherapy plans were generated in the Tomotherapy Hi-Art planning system using 6 MV photons delivered without a flattening filter. Longitudinal aperture size of 1.05 cm or 2.5 cm, a pitch of 0.3, and a modulation factor of 3 were used. Please refer to our previous study for details [25] . VMAT plans are generated with Smart Arc (SA) using the clinical version 9.0 of Pinnacle to be delivered with 6 MV photons. The machine specification of a Varian linear accelerator with 120 leaf interdigitating MLC is used. VMAT plans were generated with coplanar partial arcs to spare as much contralateral lung as possible. The arc length varied from 150° to 240°. 2-arc and 8-arc plans were created for each case. For 2- arc plans, the delivery time was constrained to 3 minutes. Delivery time was not limited for 8-arc plans. Continuous gantry motion, dose-rate variation, and MLC motion were approximated by optimizing individual beams at 4° gantry angle increments. The machine configuration was based on the “Recommended Smart Arc Physics Parameters” from Philips (Andover, MA. Recommended Smart Arc Physics Parameters. Philips Application Note 2009-03 Rev. A). Except that the “Max MU” limitation, which is 999 by default, has been changed to 5999.
All SABR plans prescribed 70 Gy delivered in 10 daily fractions to the PTV with heterogeneity corrections. They were optimized to have at least 95% of the PTV receiving 100% of the prescription dose with collapsed-cone convolution (CCC) algorithm for both HT and SA. Please refer to our previous study for details on the dose constraints used [27] . PTV coverage took precedence over OAR sparing in all plans. All treatment plans were designed under the same set of planning guidelines agreed upon among the authors with similar levels of emphasis placed on the PTV and the OARs. HT planning was conducted at the University of Arizona, and VMAT planning was conducted at the Cancer Hospital & Institute at Peking Union Medical College. In addition, all the plans were designed to deliver a dose that is used in daily clinical practice.
Plan Comparison
Various lung dose parameters and the maximum dose (D max ) to specific OARs generated in HT and VMAT plans were compared to assess their ability for OAR sparing. For the PTV, the dose covering 95% of the PTV (D 95 ), the % PTV receiving ≥70 Gy (V 70 Gy ), the mean dose (D mean ) & D max , the homogeneity index (HI), and the conformation number (CN) were generated and compared between different techniques. The HI and CN are previously defined [28] , and are described below: (1) (2) D 2 and D 98 represent the doses to 2% and 98% of the PTV, D p is the prescription dose. For the CN, the first portion (1 st parentheses) is an assessment of target volume coverage by 95% of the prescription dose; and the second part (2 nd parentheses) is an assessment of normal tissue sparing (the volume of normal tissue receiving ≥95% of the prescribed dose). The CN values between 0 and 1 with 1 representing the ideal conformity ( Fig. 1a ). In other situations that may be encountered in SABR delivery, a less-than-ideal CN is achieved when the target is partially covered by the desired dose with proportionately increased irradiation of the healthy tissue ( Fig. 1b ); or increased volume of healthy tissue may be irradiated within the limit of the allowed dose constraints due to the need to adequately cover the target volume and to spare a critical structure that is in its proximity ( Fig. 1c ).
10.1371/journal.pone.0059729.g001
Figure 1
Illustration of the possible scenarios of dose conformity described by the conformation number (CN).
The shade represents the target volume, the dotted line represents the desired isodose, the small solid in c) represent a critical structure that is immediately adjacent to the target. a). the ideal dose conformation with CN = 1. b). Less than optimal coverage of the target volume. c). In situations where the target is next to a critical structure, both adequate dose coverage of the target and the sparing of the critical structure are desired. As a result, more healthy tissue is irradiated in the context of the healthy tissue dose constraint as shown. The CN will be <1 is both b) and c).
The differences in tumor characteristics, such as tumor size and the distance between the PTV & its immediately adjacent structures, were sought between the group of patients for whom optimal PTV coverage and OAR sparing was achieved (Group 1) and those whose plans were suboptimal (Group 2). Group 1 included both HT and VMAT plans (2 and/or 8 arc plans). Group 2 included patients for whom either the HT or both VMAT plans could not successfully spare ≥1 immediately adjacent OAR if adequate PTV coverage was maintained.
Statistical Analysis
Dosimetric parameters generated in the HT, 2-arc, and 8-arc plans were compared through a randomized complete block ANOVA. After the dose parameters for the OARs and the target volumes were obtained from the HT and the VMAT plans, they were normalized to the OAR dose constraints listed in Table 2 . Selected OAR dose parameters from the HT, and VMAT plans were compared using multifactorial ANOVA while controlling for differences between patients and various OARs. In the assessment of each treatment technique’s influence on OAR sparing, multiple logistic regression was then performed on these selected normalized parameters. In analyzing the differences in tumor characteristics between groups 1 & 2, one-way ANOVA was used. Statistical significance was defined by a p value <0.05. All analyses were performed using JMP-Pro/v9.0.2 (SAS Institute Inc., Cary, NC).
10.1371/journal.pone.0059729.t002
Table 2
Comparison of PTV dose coverage parameters generated through helical tomotherapy, VMAT with 2 arcs, and 8 arcs with absolute doses illustrated in mean ± standard deviation.
VMAT
P value
HT
2 Arcs
8 Arcs
HT vs. 2 Arcs
HT vs. 8 Arcs
2 Arcs vs. 8 Arcs
PTV dose coverage parameters
D 95 (Gy)
70.61±0.61
70.00±0.00
70.00±0.00
0.0003
0.0003
0.99
V 70 Gy (%)
96.15±1.22
95.00±0.00
95.00±0.00
0.0006
0.0006
0.99
D mean (Gy)
74.03±1.74
76.12±1.53
74.99±1.69
0.002
0.11
0.06
D max (Gy)
81.98±3.66
82.76±2.77
80.76±2.39
0.46
0.26
0.07
CN
0.64±0.06
0.61±0.11
0.63±0.11
0.17
0.73
0.29
HI
19.23±10.95
21.45±6.66
17.83±5.75
0.24
0.45
0.06
Results
HT and VMAT SABR plans were generated for all 12 patients to meet the PTV dose coverage criteria. The PTV, lung, and other OARs’ dose parameters generated with each treatment approach are summarized in Tables 2 , 3 , and -4, respectively, with any two different techniques compared directly.
10.1371/journal.pone.0059729.t003
Table 3
Comparison of lung dosimetric parameters generated through helical tomotherapy, VMAT with 2 arcs, and 8 arcs with absolute doses illustrated in mean ± standard deviation.
VMAT
P value
HT
2 Arcs
8 Arcs
HT vs. 2 Arcs
HT vs. 8 Arcs
2 Arcs vs. 8 Arcs
Total lung
MLD (Gy)
6.48±2.33
7.23±3.47
6.50±2.53
0.24
0.97
0.25
V 5
21.16±8.03
22.89±10.15
22.31±9.25
0.23
0.42
0.68
V 10
15.88±5.48
15.16±6.18
14.62±5.78
0.38
0.13
0.51
V 20
10.49±4.16
10.27±4.12
9.94±3.86
0.82
0.57
0.73
Ipsilateral lung
MLD (Gy)
10.72±3.67
12.70±4.98
12.03±4.50
0.03
0.15
0.45
V 5
34.10±13.29
38.22±14.09
37.64±13.66
0.03
0.06
0.75
V 10
28.35±11.11
30.19±10.90
27.93±10.36
0.42
0.15
0.32
V 20
19.50±8.81
21.42±8.06
20.78±7.68
0.36
0.54
0.76
Contralateral lung
MLD (Gy)
1.62±0.75
1.64±0.76
1.60±0.67
0.90
0.95
0.85
V 5
7.04±6.27
8.76±7.10
8.10±5.70
0.36
0.57
0.72
V 10
2.11±2.88
1.11±1.11
0.98±1.16
0.06
0.04
0.79
V 20
0.27±0.42
0.04±0.09
0.05±0.07
0.04
0.04
0.98
HT vs. 2-arc VMAT
For the PTV, D 95 and V 70 Gy were significantly higher in the HT plans ( Table 2 ). HT also generated significantly lower D mean ( p = 0.002). No significant difference in the D max , CN, and HI was found. For the total lung (volume of both lungs – GTV), no significant differences in the mean lung dose (MLD), V 5 , V 10 , and V 20 was observed ( Table 3 ). The ipsilateral MLD and V 5 were significantly lower in HT plans ( p = 0.03, 0.03, respectively). On the contrary, 2-arc VMAT achieved lower V 20 in the contralateral lung ( p = 0.04). HT generated significantly lower D max for the heart and the major vessels ( p = 0.02, 0.01, respectively), while a trend toward lower Dmax for the major airway was observed ( p = 0.07) ( Table 4 ). However, lower D max to the spinal cord was found in VMAT plans ( p = 0.03).
10.1371/journal.pone.0059729.t004
Table 4
Comparison of the maximum dose other organs at risk (OARs) generated through helical tomotherapy, VMAT with 2 arcs, and 8 arcs with absolute doses illustrated in mean ± standard deviation.
VMAT
P value
HT
2 Arcs
8 Arcs
HT vs. 2 Arcs
HT vs. 8 Arcs
2 Arcs vs. 8 Arcs
D max for other OARs (Gy)
Spinal cord
18.84±7.44
14.18±8.57
13.65±8.20
0.03
0.02
0.79
Esophagus
22.72±13.06
22.71±13.52
22.28±11.84
0.99
0.62
0.62
Heart
23.08±22.83
29.77±26.70
29.66±26.54
0.02
0.03
0.97
Major airway
34.10±15.02
37.97±17.37
37.30±17.09
0.07
0.12
0.74
Major vessels
46.30±3.20
50.91±6.51
49.03±5.35
0.01
0.09
0.24
HT vs. 8-arc VMAT
Significantly higher D 95 and V 70 Gy for the PTV for HT was the only observed difference in the target dose indices ( Table 2 ). Total-lung dose parameters were equivalent, while the contralateral V 10 & V 20 was significantly lower in VMAT plans ( p = 0.04, 0.04, respectively) ( Table 3 ). For other OARs, HT achieved lower D max to the heart, while VMAT achieved lower D max to the spinal cord ( p = 0.03, 0.02, respectively) ( Table 4 ).
2-arc vs. 8-arc VMAT
For the PTV, a trend toward significance for lower D mean , D max , and HI were observed ( p = 0.06, 0.07, 0.06, respectively) ( Table 2 ). No significant difference between the 2 & 8-arc plans was found in any of the OAR parameters ( Tables 3 & 4 ).
D max for Adjacent OARs and MLD, V20 for the Total Lung ((MLD total , V 20, total ) after Normalizing to the Dose Constraints Used
No statistically significant difference between the three different techniques was observed in the normalized dose parameters when patient and OAR differences were controlled ( Fig. 2 ). However, the treatment technique was found to be a statistically significant factor influencing OAR sparing (meeting dose constraints), favoring HT for all OARs as an aggregate, when compared with VMAT techniques ( p = 0.0004); specifically affecting MLD total ( p = 0.0219), and D max to the heart ( p = 0.0219) & the major vessels ( p = 0.0033). OAR sparing was successfully achieved in all HT plans. However, OAR overdosing was found in 2 &/or 8-arc plans in patients 3, 9, 10, 11, 12 ( Fig. 2 ). Increasing from 2 to 8 arcs decreased the esophageal D max for patient 10, and the MLD total for patient 3 to below the dose threshold ( Fig. 2a, f ). The same did not occur for the heart, the major airway, the major vessels, and the MLD total for patients 9, 10, 11, and 12 ( Fig. 2b, c, d, and f ). However, increasing from 2 to 8 arcs decreased the D max to the major vessels and the spinal cord in many cases ( Fig. 2d, e ).
10.1371/journal.pone.0059729.g002
Figure 2
Comparison of dose parameters to the organs at risk.
Organs at risk: a) Esophagus, b) heart, c) major airway, d) major vessels, e) spinal cord, f) and g) mean lung dose (MLD) and V 20 for the total lung, after normalized to the absolute dose constraints between helical tomotherapy (Tomo), 2-arc, and 8-arc VMAT plans.
Tumor Characteristics for Patients for Whom Neither VMAT Plans Achieved Optimal OAR Sparing
The tumor characteristics for group 2 (patients 9–12) were compared with those for group 1 (patients 1–8). Given the small sample size for each group, no significant difference was found in PTV diameter, volume, and distances to the closest and furthest immediately adjacent OARs ( Table 5 ). 2/4 cases in group 2 had 3 immediately adjacent structures (50%), while only 2 such cases were found in group 1 (25%). The median PTV to its closest OAR distance was also slightly shorter in group 2 when compared with group 1 (0.13 cm vs. 0.17 cm).
10.1371/journal.pone.0059729.t005
Table 5
Tumor characteristics for groups of patients for whom optimal OAR sparing and tumor volume dose coverage can be achieved with HT or any one form of VMAT (Group 1) and those patients among whom optimal OAR sparing cannot be achieved if optimal tumor volume coverage is desired with VMAT (Group 2).
PTV diameter (cm)
PTV volume (cc)
Number of immediately adjacent normal structures
Shortest distanceto the PTV (cm)
Longest distance to thePTV (cm)
Group 1
6.05±1.90
69.42±50.46
1.88±0.84
0.17±0.06
0.61±0.50
Group 2
5.94±1.74
65.07±58.61
2.25±0.96
0.19±0.13
0.79±0.62
Discussion
Although higher D 95 and V 70 Gy were found in HT plans, this is mostly due to the differences in how target coverage parameters were executed in the treatment planning systems (TPS) under comparison. No significant difference in dose conformality was found between HT and VMAT plans. However, increasing from 2 to 8 arcs led to a trend toward lower PTV D mean , D max , and HI ( Table 2 ). Thus, suggesting a potential for improving dose homogeneity by increasing the number of arcs when treating targets in areas of complex geometry with VMAT. This finding is consistent with what has been previously observed by Guckenberger et al [29] . Poor CN has been found with all three different techniques ( Table 2 ). The CN achieved in our VMAT plans was lower than what has been reported in the literature [22] . This may be partially due to the degree of complexity in OAR sparing in close vicinity to the target; which is especially true when multiple OARs are immediately adjacent to the PTV, making it extremely difficult to conform the dose to the PTV in all directions. In these cases, less conformity is observed due to increased dose to the healthy tissue that has the least demanding dose constraint in the context of each specific case as previously shown in Fig. 1c . This is also illustrated in Figure 3 , where the desired isodose can be seen to include additional normal lung tissue due to the sparing of immediately adjacent normal structures. In the same illustration, increased dose homogeneity from the 8-Arc plan is also shown as the 77 Gy isodose is significantly diminished when compared with the 2-Arc plan. Thus, supporting that dose homogeneity may be improved by increasing the number of arcs.
10.1371/journal.pone.0059729.g003
Figure 3
Illustration of a comparison of the 2 and 8 Arc plans demonstrating that the shape of the isodose covering the PTV is largely dependent on the immediately adjacent critical structures (yellow and blue) that need to be spared in one patient.
As a result, slightly increased volume of the normal lung tissue is included in the high dose volume lateral to the PTV (blue shade) away from the central structures. Also shown here is that when comparing to the 2 Arc plan, the high dose region included by the 77 Gy isodose in the 8 Arc plan is greatly diminished, demonstrating increased homogeneity.
Lung dose parameters in HT and VMAT plans are shown in Table 3 . VMAT plans demonstrated significantly lower contralateral lung dose parameters when compared to HT plans ( Table 3 ). This can possibly be explained by the difference in the degree of the arc generated in 2 different TPS, which is partial arc for VMAT and full arc for HT. Due to the already very low values of the contralateral lung dose parameters, VMAT’s potentially improved contralateral lung sparing may not be of any clinical significance. For the ipsilateral lung, potential factors of clinical relevance, the V 5 , and MLD, were significantly lower in HT plans when compared to 2-arc plans [30] . However, this significance was lost when HT and 8-arc plans were compared. Similar to what was observed for the ipsilateral lung, HT has demonstrated significantly lower D max to the major vessels than 2-arc VMAT. However, this significance was lost when HT was compared to 8-arc VMAT. Although no significant difference was observed in any OAR dose parameters between 2 & 8 arcs VMAT plans, these observations again suggest that increasing from 2 to 8 arcs for VMAT-based SABR may have a potential for improving conformal dose avoidance in areas of complex geometry.
With patient and OAR differences controlled, no difference in a series of normalized dosimetric parameters was found between the three techniques ( Fig. 2 ). However, the technique used was found to be a significant factor influencing the ability to meet OAR dose constraints favoring HT over VMAT. Optimal OAR sparing was achieved in all cases by HT, but only the first 8 cases for VMAT ( Fig. 2 ). Increasing from 2 to 8 arcs helped to meet the dose constraints for certain structures for patients 3 & 10 ( Fig. 2 a, f), and decreased D max for the spinal cord and the major vessels in many cases ( Fig. 2 d, e). This did not lead to meeting the dose constraints for all immediately adjacent OARs for patients 9–12, among whom increased number of immediately adjacent structures and short distance between PTV and the closest OAR were common. No statistically significant difference in tumor characteristics was found between patients 9–12 and patients 1–8 due to the small number of patients studied ( Table 5 ). However, our findings suggest that HT may be more appropriate in cases which demand conformal dose avoidance of multiple structures (≥2) that are very close to the PTV in the delivery of SABR for centrally located lung lesions, even though increasing the number of arcs in VMAT may improve OAR sparing in certain situations. Our findings are corroborated in a study comparing HT and VMAT in delivering conventionally fractionated radiotherapy in several body sites by Rong et al, which demonstrated improved target dose homogeneity and lower doses to more critical structures in the HT plans [28] . This is mainly due to the increased freedom of intensity modulation created by delivering image-guided IMRT under synchronous gantry rotation and couch motion with HT [13] , [25] .
Although shown to be more capable of OAR sparing in setting of lung SABR for central lesions, HT is associated with much longer fractional treatment delivery time of >40 minutes for each case, mainly attributing to the complexity of intensity modulation required. Due to this fact, the exact treatment time for the HT plans was not recorded. On the other hand, 2 & 8 arc VMAT had average fractional treatment delivery times of only 180 & 331 seconds, respectively. Thus, VMAT remains to be more desirable for targets in areas of relatively less complex geometry. The prolonged treatment time associated with HT can be potentially improved by implementing the dynamic jaw and dynamic couch feature [31] . At the current time, this remains a problem for HT-based SABR mainly because of its associated increase in intrafractional motion, which can be critical when treating central lesions with SABR. As a result, proper respiratory motion management and careful body immobilization are essential [32] , [33] . In our experience, 4D CT simulation to account for tumor motion in various locations remains the most straight forward approach for respiratory motion management for HT clinically. Furthermore, treatment efficiency can be improved by dividing the fractional dose into two consecutive fractions (7 Gy delivered in 2 consecutive fractions, 3.5 Gy/fraction) [34] .
The dose delivery & calculation accuracy have been commented elsewhere, which were found to be adequate for both VMAT and HT [28] . In a study by Takahashi et al, the CCC algorithm closely approximated the Monte Carlo algorithm in the dose calculation specific for lung SBRT [35] . This warrants the validity of SABR dose calculation for both HT and VMAT, which is also critical in the setting of centrally located lesions closely surrounded by multiple critical OARs. Early clinical reports on treating mostly peripheral lesions with HT-based and VMAT-based SABR have been promising [34] , [36] . A prospective clinical study investigating how to best apply these advanced techniques in the treatment of central lung lesions with SABR will be conducted in the near future.
Due the virtual nature of this dosimetric study, no further quality assurance is conducted. However, treatment planning accuracy for VMAT and HT are implied from studies conducted in the past (25, 37). But it will be done as part of a prospective study in the future.
Conclusion
In delivering SABR or SBRT for centrally located lung lesions, HT appears to be superior to VMAT in OAR sparing mainly for targets with multiple immediately adjacent structures. Although increasing arc number cannot achieve the aim of sparing all the OARS in cases associated with complex geometry, it may help lowering the doses to them. However, VMAT may be preferred over HT in cases of simpler geometry due to much shorter treatment delivery time.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
“ There is proportional value in our attention to each action—so you will not lose heart if you devote no more time than they warrant to matters of less importance .”
– Marcus Aurelius, Meditations [ 1 ]
When confronted with several choices, we need to have an evaluation of how good each option is. Each choice has some immediate consequences, but also takes us into a new state where new choices emerge, and so on. Think of chess as an example. One intuitive way to solve a sequential decision-making problem like chess is to prospectively think into the future. This idea, known as model-based planning in the reinforcement learning literature [ 2 ], expands a mental decision-tree by simulating a number of future action sequences. Although this method is accurate (in terms of statistical efficiency), evaluating deep trees is computationally expensive (in terms of time, working memory, metabolic energy, etc.). In chess, for example, it is impossible even for the best supercomputers to expand the tree of all possible strategies up to the end of the game. Therefore, several solutions have been provided in the artificial intelligence literature for how to approximate the values of choices without expanding a search tree to its fullest extent [ 3 ] or how to make the best use of limited computational resources to plan better [ 4 ].
To avoid the costs of planning altogether, a drastic alternative is to rely on heuristic methods that evaluate choices without any tree expansion. For example, a chess player can evaluate a chess position, without investigating the possibility of that position leading to a win or lose, by simply counting up the values of their pieces—a common heuristic utilized by novice players. Another example of approximate evaluation techniques, widely used in both natural and artificial intelligence. is using habits. This method, known as model-free reinforcement learning [ 2 , 5 ], simply “caches” the average of previously realized rewards ensued by performing each action, and uses the cached values for evaluating those choices should they come up again in the future. Although using such heuristics frees cognitive resources from model-based planning, the downside is their inaccuracy. Habits, for example, take many trials to form, and they are always unreliable in changing environments.
Rather than clinging to one of these extreme solutions (i.e., full planning vs. heuristics/habits), an intelligent agent can instead combine the two in order to harvest the relative advantages (i.e., accuracy vs. affordability) of both techniques [ 6 – 9 ]. This, in theory, is achievable by forward planning up to some depth and then exploiting heuristic values as proxies for consequences that may arise in the further future. That is, when the depth of planning is say d , the agent computes the value of a choice by adding the first d rewards predicted by explicit simulation, to the value of the remaining actions estimated by the heuristic/habitual values. For example, a chess player could think three steps ahead, and then estimate, heuristically, the strength of the position he could achieve after those three moves. This integrative approach has been used in artificial intelligence for example for obtaining super-human Go performance [ 10 ]). Furthermore, it was shown recently that humans also use this scheme, named plan-until-habit, for integrating planning and habitual processes in a normative way, and that their depth of planning depends on the time-pressure imposed on them [ 11 ].
The plan-until-habit (or plan-until-heuristic, in general) scheme aims at mitigating the computational costs of planning by appealing to the habitual system after the planning system has sufficiently expanded the decision-tree. Obviously, the first questions to be asked in this framework are “in which directions the decision-tree should be expanded?”, and “when should the expansion stop?”. In this paper, we present, for the first time, a principled algorithm for optimal tree-expansion in the plan-until-habit framework. The algorithm is based on a speed/accuracy tradeoff: deeper planning leads to more accurate evaluations, but at the cost of slower decision-making. As a proof of concept, we show through simulations how this algorithm expands the decision-tree effectively and efficiently in a simulated grid-world environment. We further show that our algorithm can explain several behavioral patterns in animals and humans, namely the effect of time-pressure on the depth of planning, the effect of reward magnitudes on the direction of planning, and the gradual shift from goal-directed to habitual behavior during training. The algorithms also provide several predictions testable in animal/human experiments.
Results
Theory sketch
From an external-observer viewpoint, the questions to be answered by an agent are of the type “what action should be taken?”. From a metacognitive perspective, however, the agent should first think about how to think (e.g., how deep she should plan). In fact, the question she could ask at each step of the planning process is “Should I expand the decision-tree one step further?”, and if yes, “In what direction?”.
To answer these, assume that the agent has already expanded a tree to a certain extent ( Fig 1A ). This means that the agent knows, possibly with some uncertainties, a few next states to be visited upon taking each action, and the immediate rewards associated with each of those transitions. She can, therefore, sum up the predicted rewards along each trajectory (i.e., action-sequence) and have an estimate of the total rewards to be achieved. On the top of this “total immediate rewards”, each trajectory ends in a frontier state which represents the edge of the current planning horizon along that trajectory. The habitual (or any other heuristic) values on this frontier state supposedly reflect the total (discounted) rewards to be expected from that point on. Therefore, the sum of “total immediate rewards” and the habitual value of the frontier node provides an estimate of the total expected reward of each trajectory ( Fig 1B ).
10.1371/journal.pcbi.1006827.g001
Fig 1
Overview of the pruning scheme, illustrated via an example.
(A) A snapshot of the search tree. Nodes of the tree represent states, and each state has a number of available actions, denoted with circles, that lead to next states. Blue graphs show value distributions for the leaves of the tree, estimated by the model-free (MF) or any other heuristic system. Green graphs show the immediate rewards for previously expanded state-actions, estimated via the model-based (MB) system. (B) Each path from the root to a leave forms a strategy, A i , with a corresponding value distribution. These distributions are obtained by summing up the value distributions of the leaves with the immediate reward distributions accumulated along the way. (C) To compute the value of uncertainty resolution ( vur ), say for A 3 , the agents assumes that one further expansion would result in a sharper value distribution (one of the black/grey distributions). The location (i.e., the mean) of the new distribution cannot be known in advance, but it can be treated as a random variable, whose distribution can be analytically obtained ( Eq 14 ). The vur for A 3 is therefore the expected value, over all possible sharper distributions (grey curves), of the additional rewards that can be obtained by a policy improvement in the light of that potential new information (i.e., the sharper distribution). (D) After computing vur for all strategies A i , the highest vur (in this case, for A 3 ) is compared to the cost of expansion. If it is bigger than the cost, the tree expands along the direction of that strategy. This corresponds to loading a new node, which is the successor state of the leaf of A 3 , from the MB system and adding it to the tree.
Habitual values, however, can be highly unreliable due to the inflexible nature of habit formation. For each given trajectory, therefore, the dependence of its estimated total rewards on uncertain habitual values renders the whole estimation uncertain. If expanding the tree along that trajectory would make value estimation less dependent on habitual values and thus reduce uncertainty, that expansion is worth considering. In this sense, the critical value to be computed for each trajectory is the “value of uncertainty reduction” ( vur ). vur computation for a trajectory should examine whether a new piece of information, possibly providable by a further expansion of the tree along that trajectory, could change agent’s decision about what action to be taken, and how much extra value is expected to be gained by that policy improvement. vur is, in fact, the expected value of policy improvement-induced rewards, computed over all possible new pieces of information that could be provided by expanding the trajectory one step further ( Fig 1C ). Although the agent readily possesses those new pieces of information in her memory (because she has a model of the environment), loading them into working memory and taking them into the value-estimation account is worth doing only if the value of uncertainty reduction is more than its cost.
Here is the general scheme of our algorithm: at each stage of planning, vur is computed for each trajectory on the search tree (we discuss later that previously-computed vur -values can be reused later under certain conditions). The trajectory with the highest vur is expanded if its vur is bigger than the cost of expansion. Otherwise, the expansion process is terminated and the agent chooses an action (e.g., using soft-max rule) according to the estimated values derived from the tree.
In this paper, we assume that the cost of expansion simply reflects the opportunity cost of time. That is, assuming that each expansion takes ϵ time units, the total cost of one expansion is R ¯ ϵ , where R ¯ is the average reward the agent receives in the given environment.
As explained before, the main motivation for expanding the tree is reducing value-estimation uncertainties. There could be several reasons for why expansion reduces uncertainty. In many cases, like chess, heuristic estimations become more precise as the game advances. In general, proximity to goal sometimes makes it easier to evaluate the states. Another way that expansion reduces uncertainty, which is the focus of our formal model, is through temporal discounting. By each level of expanding a trajectory, the dependence of its estimated value on the less-reliable habitual system is shifted one step further into the future.
As a simplified example, imagine you are in a maze and you have already thought two steps ahead along a certain trajectory, T 1 , of actions, and those two steps will take you to the state s ′. You can use the MF value, V MF ( s ′) of that state to compute the total value of the trajectory: V ( T 1 ) = r 1 + γ . r 2 + γ 2 . V MF ( s ′), where r 1 and r 2 are the immediate rewards expected to be received by performing the first and the second actions on the trajectory T 1 . Assuming that the estimates of the immediate rewards have zero uncertainty, and that the MF estimates always have variance σ 2 (i.e., uncertainty)), the total uncertainty of V ( T 1 ) will be ( γ 2 . σ ) 2 = γ 4 . σ 2 . Now, if you think one step deeper and expect to land in state s ′′ after taking the first three steps of trajectory T 2 , then V ( T 2 ) = r 1 + γ . r 2 + γ 2 . r 3 + γ 3 . V MF ( s ′′ ). Therefore, its variance will be ( γ 3 . σ ) 2 = γ 6 . σ 2 . This toy example shows that as a natural consequence of temporal discounting, by increasing the depth of planning, the total uncertainty of trajectories decreases, due to the reduced reliance on uncertain MF values. Therefore, the discount factor is the critical variable that determines the extent of uncertainty reduction by each expansion.
In this paper, we only consider environments where the transition between states via actions are deterministic (i.e., deterministic transition function for the Markov decision process; See Methods for how this assumption can be relaxed). Therefore, the expanded tree, at each point, is a deterministic tree. In order to compute vur , let’s define a strategy in a tree as a combination of actions that an agent can take to reach a leaf in the tree (see Fig 1 ), and define a frontier search as the set of all strategies that agent can take in a given tree (e.g., the search frontier in Fig 1 is { A 1 , A 2 , A 3 , A 4 , A 5 }). Based on this definitions, as shows in the Methods section, the value of uncertainty reduction for strategy A i , given the search frontier F , can be written as:
VUR ( A i | F ) = E μ i * [ max ( μ i * , max A ∈ F - A i E [ V ( A ) ] ) ] ︸ with expansion - max A ∈ F E [ V ( A ) ] ︸ without expansion , (1)
where F − A i is the set F excluding A i . According to this equation, computing vur ( A i | F ) requires μ i * , which is the expected mean of strategy A i after the potential expansion. However, this variable can be computed before expansion, by μ i * ∼ N ( μ i , ( 1 - γ 2 ) σ i 2 ) (see Methods section), in which γ is the discount factor, and μ i and σ i 2 are respectively the mean and the variance of the MF-value distribution for the last action on A i . In other words, vur is computable based on μ i * , the expectation with respect to the predicted value of A i after expansion, instead of its realized value which is not available before the expansion (a more general form of the above equation without reliance on the discount factor is presented in the Methods section).
The right-hand side of Eq 1 is composed of two parts: the amount of future rewards that are expected to be gained with the expansion of strategy A i , and the amount expected to be gained without the expansion of A i . vur is the difference between these two quantities. The without-expansion term is simply the value of the best strategy that is currently available to the agent. In the with-expansion term, the outer ‘max’ operator implies that if after expanding, A i turns out to be worse than the other available strategies ( F − A i ), then the best strategy among the other ones will be taken. Otherwise, A i will be taken.
The agent, however, needs to calculate this term before the expansion of A i and therefore the term is calculated based on the expectation with respect to the predicted value of A i after expansion (denoted by μ i * ) instead of its realized value which is not available before the expansion.
It can be shown that in the case of normally distributed MF value functions, Eq 1 has a closed-form solution (see S1 Text for details):
VUR ( A i | F ) = { σ i [ ϕ ( μ i - μ β σ i ) - μ i - μ β σ i Φ ( - μ i - μ β σ i ) ] + μ β - μ α if A i is the best strategy σ i [ ϕ ( μ i - μ α σ i ) - μ i - μ α σ i Φ ( - μ i - μ α σ i ) ] otherwise (2)
where μ i and σ i are, respectively, the mean and the standard deviation of strategy A i . Furthermore, μ α and μ β are the means of the, respectively, first-best and second-best strategies in the currently-expanded tree. First-best and second-best strategies are the strategies that have the highest and the second-highest mean values. Finally, ϕ and Φ are, respectively, the probability density and cumulative distribution functions of a standard normal distribution.
A central principle for any meta-control algorithm is that the cost of meta-reasoning (here, the cost of computing arg max A VUR ( A | F )) should be lower than the cost of expensive reasoning (here, one-step expansion of the decision-tree). In terms of memory cost, tree-expansion would require loading information about the expanding nodes from the long-term to the working memory. Furthermore, it would require engaging an additional working memory slot to store such information. Meta-reasoning, however, has minimal memory cost, since all the variables for computing arg max A VUR ( A | F ) already exist in the working memory (i.e., are in the already-expanded tree).
In terms of computational-time cost, we should stress that even though we want to find the strategy with the maximum vur value, this does not necessarily require computing vur ’s of all strategies at each time step. vur ( A i | F ) only depends on μ i , σ i and μ α (or μ β ). Therefore, vur values can be cached, and reused as long as the aforementioned parameters have not changed (i.e., the newly-added strategies are not first- nor second-best strategies). From an algorithmic point of view, computing vur of a given A i can be viewed as a constant time operation. Therefore computing arg max A vur ( A | F ) is in the order of O ( | F | ) in the worst case, where | F | is the cardinality of F (i.e., number of items in the search frontier). However, as shown in the appendix, as the tree expands, the expected cost becomes constant (i.e., O ( 1 ) ) asymptotically, given that the agent caches previously computed vur values. This is intuitively becuase as the depth of the tree grows, the uncertainty around the value of the to-be-expanded strategy shrinks (becuase of the discounting factor), which makes it less likely that the strategy (which is not currently the best strategy) becomes the best one after expnasion (or second best strategy). As such, the chances that a new expansion affects previusly computed vur values becomes smaller and smaller as the tree gets deeper. This rate of decrement is faster than the rate at which new potential strategies are added to the tree as it gets deeper, and therefore overall the number of vur values that need re-computation remains constant as in the limit.
Pruning in a grid world environment
Just as a proof of concept, we would like to see whether our method can be beneficial in a setting in which an agent is combining both MF and MB information for efficient planning. For this, we first trained an agent in an episodic grid-world environment where she obtains imperfect estimates of state-values by the model-free system. After training, she utilizes both the MF and the MB systems to use the plan-until-habit scheme, where the MB system is used to construct the tree, and the MF systems is used for estimating the values of state-actions that lie on the frontier of the tree. We predict that the increased accuracy in model-free estimates, as a result of training, would bias the direction of expanding the tree towards better states.
The agent starts each episode in the center of a 7 × 7 grid and can choose to go up, down, left, or right at each state. All the transitions are deterministic and are associated with a unit cost. The bottom right cell is the goal state that concludes the episode. This state is not associated with any reward, but is implicitly rewarding since it terminates the costly walk in the grid world. Evidently, the optimal policies are combinations of three right moves and three down moves. Given the structure of the task, for easier geometric interpretation and without loss of generality, the MF system learns state values, rather than state-action values.
To apply our plan-until-habit pruning algorithm, we require an MF system that learns not just the mean, but also the variance (i.e., uncertainty) over the state values. In our implementation, the agent estimates the value of a state by generating a number of trajectory samples from the state, similar to the first-visit Monte Carlo method described in [ 2 ], and utilizing the trajectories’ return statistics. However, instead of estimating the Q -values with Monte Carlo averages, we use independent conjugate normal priors and obtain posterior estimates of Q ’s, which are conditioned on the trajectory returns (see S1 Text ). We obtain N trajectory samples starting from each state, such that each sample consists of a trajectory resulting from a fixed uniform random policy that assigns 1 4 probability to each direction {UP, DOWN, LEFT, RIGHT}.
We test our planning model in two different settings. First, we assume the agent has no experience interacting with the environment (i.e., N = 0). This condition results in the posterior Q -values having large and equal variances. We compare this with the case where the agent has collected some samples (i.e., N = 10), resulting in more accurate estimates of state values. In both cases, we employ the same pruning mechanism, with a variable number of possible tree expansions (capturing working-memory limitations; see Discussion section) selected uniformly from [5, 25] and γ = 0.95.
As displayed in Fig 2A , in the no-experience condition, the search tree is explored in all directions almost uniformly. In the second condition, however, the search is directed more towards the goal state as illustrated in Fig 2B . These results are in line with our intuition that the agent prunes more aggressively as she gathers more experience and thus, is better able to judge what the promising states or actions are.
10.1371/journal.pcbi.1006827.g002
Fig 2
Grid-world pruning simulation results.
Reaching the bottom-right corner of the map with minimum moves is rewarding. The heatmaps show the frequencies of state-visits during the tree expansion when the agent starts from the middle of the map, and (A) the agent has had no prior exposure to the environment, or (B) after some exposure (i.e., 10 trajectory samples from each state) resulting in more accurate estimates of model-free values.
Human-like pruning
Behavioral evidence suggests that humans, when planning, curtail any further evaluation of a sequence of actions as soon as they encounter a large punishment on the sequence [ 12 ]. In a behavioral task [ 12 ], subjects were required to plan ahead in order to maximize their income gain. The environment in the task is composed of six states. Each state affords two actions, each of which transitions the subject to another state deterministically. Subjects see their current state on a display and press the ‘U’ or ‘I’ buttons on the keyboard to transition to a different state.
In the first phase of the experiment, subjects learn the deterministic transition structure of the environment. In the second phase, transitions are associated with specific gains or losses, which are visually cued to make it easier to remember. At each trial in this stage, subjects are told to take a certain number of actions, varying between 2 and 8, and collect all the rewards and punishments along their chosen trajectory. This forces them to think ahead and plan in order to find a relatively profitable trajectory among 2 2 = 4 to 2 8 = 256 options. For example, in the setting described in Fig 3A , 8 possible trajectories resulting from 3 consecutive actions are displayed.
10.1371/journal.pcbi.1006827.g003
Fig 3
Example search trees from [ 12 ].
A : Starting at state 3, subjects make three consecutive decisions (pressing ‘U’ or ‘I’), each of which are associated with a gain or loss. Two trajectories maximize the cumulative rewards in this example and achieve −20. B and C : State transition frequencies of subjects. Higher frequencies are illustrated with thicker lines. If a transition is not taken by any of the subjects, then it is illustrated with a dashed line. Yellow backgrounds show the optimal trajectories. Colors red, black, green, and blue denote the transition rewards of P , −20, + 20 and + 140 respectively. B : P = −140 condition. It can be seen that the subjects avoid the action associated with the large punishment. C : P = −70 condition. Subjects are eager to take transitions with large losses when such transitions lead to large gains (i.e., + 140), which in fact is the optimal strategy. Reprinted with permission from [ 12 ].
Out of all 12 transitions, 3 of them are associated with a large loss. The magnitude of this loss is manipulated across trials (from {−140, −100, −70}) such that for certain losses (i.e., −100 and −70), Pavlovian pruning results in suboptimal strategies. In other words, pruning a strategy that starts with a −100 or −70 loss would result in discarding the most profitable course of actions, since such actions will eventually lead to highly rewarding states. The results of this experiment show that humans prune infrequently if pruning results in prematurely discarding optimal trajectories. Conversely, they tend to prune liberally when pruning does not eliminate the optimal trajectories. That is, they prune more when the loss on a trajectory is so large (i.e., −140) that cannot be compensated for by future rewards.
We aimed to replicate this task in our simulations. Because in the first part of the experiments subjects learn the transition and the immediate rewards through repetitive exposure, we assume that the agent (i.e., our simulation of a subject) knows the transition and reward structures. Since the immediate state-action rewards are visually cued, subjects, after observing their starting state s and their available actions a 1 and a 2 , presumably incorporate the immediate rewards of those actions into their planning at no cost. Therefore, we assume that the agent starts the decision tree with two already-expanded actions, with values Q ( a i ) = R ( s , a i ) + γV ( T ( s , a i )), where i ∈ 1, 2, and R ( s , a ) and T ( s , a ) are the immediate reward and successor states resulting from taking action a at state s .
As in the previous experiment, we obtain the posterior Q -value distributions of the agent through a training stage. Similar to the training phase of the original study, we have the simulated agent interact with the environment for 100 episodes, during which she observes transitions and collects reinforcements. At each trial, the agent is located in a random state and is allowed to make a certain number of moves, which is sampled uniformly from {2, 3, 4}. She selects actions following uniform random policy, and stores the mean cumulative reinforcements collected after taking action a at state s , similar to the first-visit Monte Carlo algorithm [ 2 ]. Those mean values are then used for obtaining the posterior Q -distributions assuming a conjugate normal distribution as in the previous experiment (see S1 Text ). The prior is a normal distribution with mean and standard deviation of 0 and 1000, respectively. After the training stage, the agent moves on to the pruning state, where she starts at state s and is asked to mentally expand the planning tree for n ∈ {2, 4, 6, 8, 10, 12 s } steps. We record the frequency with which the agent expands the early branch with the large punishment, which we very between −40 and −140. Finally, we set γ to 0.95 as before.
One critical observation in [ 12 ] is that subjects prune more frequently as the magnitude of the punishment increases. As shown in Fig 4 , our simulation results account for this pattern. Intuitively, observing a punishment on a trajectory reduces the expected value of the trajectory and thus, reduces the overlap between the value-distribution of that trajectory and that of the best trajectory. When the punishment is large enough, the overlap becomes very small even if the trajectories have highly uncertain value estimates. Small overlap is equivalent to low “value of uncertainty resolution” expected from expanding the unpromising trajectory, because there is a very small chance that the new pieces of information will render the unpromising trajectory better than the currently best strategy.
10.1371/journal.pcbi.1006827.g004
Fig 4
The frequency of pruning the branch with the large punishment.
The black area on the right is the region where the agent does not prune (i.e., expands) the punishment branch. Each condition is averaged over 300 simulations.
In the simulations, we also vary the maximum number of branches allowed to be expanded, reflecting constraints on the working memory capacity (see Discussion section). Not surprisingly, as the memory capacity is increased, pruning frequency decreases ( Fig 4 ).
Another important aspect of the study is that the likelihood of selecting the optimal sequence of actions by the subjects was affected by three factors: (i) subjects were less likely to choose the “Optimal Lookahead” sequence when it contained a large loss, (ii) this effect became larger as the size of the loss increased, and (iii) the optimal sequence was more likely to be chosen when the tree was shallow (i.e., when the subjects were supposed to choose a smaller number of actions). These three effects are shown in the top panel of Fig 5 for the data reported in Huys et. al. [ 12 ]. The bottom panel displays the prediction of our method based on the simulations in the same task. It can be seen that similar to the actual data, we predict that the subjects will be more successful in picking the optimal sequence when it does not contain a large loss, the tree is shallow and the loss is small (i.e., the effect is strongest in the −140 group and the weakest in the −70 group). One notable qualitative mismatch between the top and bottom panels is that, our model assigns a higher probability of choosing optimal sequences for smaller depths than what is shown for the actual data on the top panel. This is because, in our setting, the agent is very likely to make enough expansions to find the optimal sequence for a tree of depth 2, as there are only 2 2 = 4 possible sequences—which can be spanned with a small number of expansions. The number of expansions are sampled from round (Gamma(4, 2)) + 1, where + 1 ensures positivity. Given this distribution, it is often the case that the agent performs enough expansions to find the optimal. However, if we look at the top left plot in Fig 5 , we see that the probability of choosing the optimal sequence is low if it contains a large loss—even for depth of 2. This might suggest that the subjects do not fully use their “expansion budgets”, if performing expansions do not seem advantageous. The same could be done in our scheme by stopping expansions altogether if the maximum vur is below a threshold. However, we refrained from doing so, and instead used a random number of expansions for simplicity, and for limiting the flexibility of the model to prevent overfitting. Other than this, all other parameters are kept the same as the ones used for generating Fig 3 .
10.1371/journal.pcbi.1006827.g005
Fig 5
The top panels show the effect of different factors on choosing the optimal sequence of action.
The panels are adapted from [ 12 ]. The x-axis denotes the number of actions the subjects were supposed to take, which determines the maximum depth of the search tree. The y-axis denotes the probability of choosing the Optimal Lookahead sequence. The blue lines represent the condition that the optimal sequences of actions included a big loss, and the green lines represent the condition that the optimal sequence of actions did not include a big loss. The amount of big loss is varied among the panels, and is mentioned by Group X on top of the panels, in which X denotes the amount of big loss (X = -140, -100, -70). The bottom panels are similar to the top panels but using the data obtained from the simulations of the model in the same settings.
Previously, the punishment-induced pruning discussed here was explained assuming that a Pavlovian system, reflexively evoked by large losses, curtails further evaluation of the corresponding sub-tree [ 12 , 13 ]. In our computational framework, however, this pruning pattern emerges naturally, rather than devising new mechanisms, from a speed-accuracy tradeoff. Furthermore, the normative nature of our explanation depicts punishment-induced pruning as an adaptive mechanism in the face of cognitive limitations, rather than depicting it an a “maladaptive” Pavlovian response [ 12 ].
The effects of training and decision-making times on depth of planning
Several lines of research have shown a transfer of control over behavior from goal-directed to habitual decision-making during the course of learning [ 14 – 17 ]. Previous accounts of interaction between MB and MF algorithms [ 18 , 19 ] explained this behavior by showing that the MF value estimates become more and more accurate along the course of experiencing a task. As a result, they eventually become more accurate than MB estimates [ 18 ], or become accurate enough that the extra information that MB planning can provide is not worth the cost of planning [ 19 ]. Therefore, a binary transition from goal-directed to habitual responding occurs in behavior.
Our model also explains the transition, but also suggests that it is gradual, rather than binary. As MF estimates become more accurate, the variance in strategy values decrease and thus, vur values also decrease monotonically (see S1 Text for an analytical proof of this effect). This implies that an experienced agent would construct a shallower search tree and hence, spends less time planning compared to an inexperienced agent. Furthermore, in contrast to the previous accounts that propose ad-hoc [ 18 ] or optimal, but with very strong assumptions (i.e., MB tree-expansion has an infinite depth), [ 19 ] models for MB-MF arbitration mechanisms, our proposed model’s optimality is based on more reasonable assumptions.
Our algorithm further predicts that in a plan-until-habit scheme, time-limitation would reduce the depth of planning. That is, time pressure would monotonically limit the total number of branches to be expanded, pressing the agent to switch to habitual/heuristic values at a shallower depth. This is due to the fact that every tree-expansion step is assumed to take a certain amount of time, ϵ . Therefore, our model, for the first time, accounts for recent evidence showing that humans use a plan-until-habit scheme and that time pressure reduces their depth of MB planning [ 11 ], resulting to a relying on habitual responses at a shallower level.
In this experimental study [ 11 ], participants first learned the stationary transition structure of the environment in a three-step task. They then navigated through the decision tree, in each trial, to reach their desired terminal state. The rewarding value of the terminal states was non-stationary and changed along the trials, allowing to measure, from participants’ choices, whether or not they use a plan-to-habit scheme; and if they do, what depth of planning they adopt. The experiment imposed a decision time-limit of either 2000 or 700 milliseconds to two different groups of participants. While both groups showed a significant behavioral signature of plan-to-habit responding, participants that experienced a shorter time-limitation showed pruning the tree and switching to MF values at shallower levels.
Plan-to-habit pruning in comparison
In this section, we qualitatively compare our plan-to-habit pruning algorithm to other methods, such as Monte Carlo tree search.
Mean-based pruning, variance-based pruning
Let us consider a simple pruning algorithm that expands the tree only according to the mean value of the strategies, and ignores their variances (e.g., the algorithm always—or stochastically- expands the strategy with the highest mean value, arg max A E [ V ( A ) ] ). The critical drawback of such algorithm is that it does not expand uncertain trajectories that have relatively smaller mean values. The true value of a strategy with a low estimated mean but high estimated uncertainty might be even higher than the strategy known to have the highest estimated mean. Therefore, uncertain strategies should be given the chance to prove their worth. In this sense, our algorithm proposes an optimal weighting of mean and variance in order to prioritize expansions.
Furthermore, note that an algorithm that only takes into account the mean values cannot explain the canonical experimental evidence of the gradual transition from goal-directed to habitual behavior over time [ 14 – 17 ]. Explaining such a transition, at least in all the existing accounts, requires keeping track of the MB and MF uncertainties, and taking them into account when arbitrating between the two systems [ 18 , 19 ].
Similarly, an algorithm that expands the tree only on the basis of the uncertainty of trajectories’ values, would only favor mental exploration of uncertain trajectories, even when their low mean value renders them totally unpromising.
Monte Carlo tree search
Monte Carlo tree search (MCTS) is a family of algorithms that incrementally and stochastically builds a search tree to approximate state-action values. This incremental growth, as in our algorithm, prioritizes the promising regions of the search space by directing the growth of the tree towards high-value states.
A so-called tree policy is used to traverse the search tree and select a node which is not fully expanded, i.e., it has immediate successors that are not included in the tree. The node is then expanded by adding one of its unexplored children to the tree, from which a trajectory will be simulated for a fixed number of steps or until a terminal state is reached. Such trajectories are generated using a rollout policy which is typically fast to compute—for instance at each step of the trajectory actions are selected randomly and uniformly. The outcome of this trajectory (i.e., cumulative discounted rewards along the trajectory) is used to update the value estimates of the nodes in the tree that lie along the path from the root to the expanded node.
MCTS algorithms diverges from our approach mainly in how the value of states and actions are computed. The former relies on simulated experiences, called rollouts, whereas the latter relies on summaries of past experiences in terms of “cached” values (or model-free values). As such, the latter is much cheaper to compute, but is dependent on the policy with which those experiences are collected. In MCTS, however, values depend mostly on the tree policy, which is adaptive. Consequently, relying on past experiences, as in vur model, is cheaper but less flexible.
Our plan-to-habit pruning algorithm can be compared to MCTS methods on another level by focusing on tree policies. The most popular MCTS tree policy is “UCT” (Upper Confidence Bound 1 applied to trees) [ 20 ], which is based on a successful multi-armed bandit algorithm called “UCB1” (Upper Confidence Bound 1). UCB1 assigns scores to actions as a combination of their (empirical) mean returns and their exploration coefficients, which reflects how many times an action is sampled in comparison to other actions. UCT adapts this UCB1 rule to MCTS by recursively applying this rule to select actions down the tree starting from the root node.
UCT is simple and has successfully been utilized for many applications. However, it has also been noted [ 21 , 22 ] that UCT’s goal is different from that of approximate planning. UCT attempts to ensure a high net simulated worth for the actions that are taken during the Monte Carlo simulations that comprise planning. However, all that actually matters is the real worth of the single action that is ultimately taken in the world after all the simulations have terminated. To put it in another way, in planning, simulations and expansions are valuable, only because they help select the best action. However, UCT actually aims to maximize the sum of rewards obtained in simulations, rather than paying direct attention to the quality of actual (i.e., not simulated) actions. Consequently, it tries to avoid simulations with potentially low rewards, even though they might help select better actions. In other words, even though UCT explicitly computes an “exploration bonus” that favors infrequently visited nodes, it still underestimates how valuable exploration is. In fact, it has been shown that modifying UCT to explore (asymptotically) more when selecting root actions increases its performance [ 21 , 22 ]. Our model does not suffer from this problem of underexploration as it explictly quantifies the expected gain of expanding a node.
Discussion
Finding optimal or near optimal actions requires comparing the expected value of all possible plans that can be taken in the future. This can be achieved by explicitly expanding a model that represents the underlying structure of the environment, followed by calculating the expected value of each plan. However, the computational complexity of this process grows exponentially with the depth of search for optimal plans, which makes it infeasible to implement in all but the smallest environments. Indeed, evidence shows that humans and other animals use alternative ways that have lower computational complexities than explicit search. Examples are using ‘cached’ values of actions instead of recalculating them at each decision point [ 18 ], or using ‘action chunking’, in which actions span over multiple future states [ 23 ]. Here, we suggest that such decision-making strategies are not operating independent of the planning processes, but they interact in order to provide a planning process that adapts its extent according to time and cognitive resource and therefore, scales to complex environments. In particular, the model that we suggest is built upon two bases: (i) the planning process is directed toward the parts of the environment’s model that are most likely to benefit from further deliberation, and (ii) the planning process uses ‘cached’ action values for the unexpanded (i.e., pruned) parts of the tree. Simulation results showed that the model prunes effectively in a synthetic grid world, and that it explains several patterns reported in humans/animals.
Namely, a sequential decision-making task has demonstrated that humans use strategies such as ‘fragmentation’ and ‘hoarding’, in addition to pruning, for efficient planning. The pruning process, however, was shown to play a significant role on the top of those strategies [ 13 ]. Indeed, the data shows that humans stop expanding a branch of the model once they encounter a large punishment. This effect was previously accounted for, in the model-based planning framework, by adding a new parameter that encodes the probability of stopping the search after encountering a large punishment. The model here does not explicitly contain such a parameter, but the pruning effect emerges naturally based on the fact that the value of uncertainty resolution is lower for the branches of the model that start with large punishments and therefore, they are more likely to be pruned.
Another component of the model here is using the cached values for unexpanded parts of the model, which is in line with previous works [ 11 , 12 ]. The psychological nature of such cached values can be related to either Pavlovian (as used in [ 12 ]) or instrumental (as used in [ 11 ]) processes in the brain, depending on whether cached values are coded for state or for state-action pairs, respectively. In the former case, our algorithm represents a collaborative interaction between instrumental model-based and Pavlovian processes [ 24 ]. In the latter case, it represents interaction between instrumental model-based and instrumental model-free processes. The theoretical framework we presented here is readily compatible with either case.
As discussed in the previous sections, temporal discounting of future rewards (and punishments) is a necessary component in the current framework. Reduction of uncertainty is a variable that changes monotonically with the discount factor: the smaller the γ , the less dependence of the value of each strategy on uncertain cached values on the leaves and therefore, the more reduction of uncertainty by deepening the tree. However, when a new piece of information on a leaf at depth d is achieved, its policy-improvement impact on the root-level actions is measured at the root of the tree, thus discounted by a factor γ d . Therefore, the smaller the γ is, the less valuable a given uncertainty reduction is. This effect counteracts the above-mentioned effect of γ on the degree of uncertainty reduction. As a result, discount factor has a non-monotonic effect on vur and thus, on the depth of planning. vur is equal to zero for γ -values of zero and one, and reaches a maximum for an intermediate value of γ (its exact value depends on other parameters).
In sum, we proposed a principled algorithm for pruning in a plan-until-heuristic scheme. While we showed the ability of the model in accounting for several behavioral patterns in humans/animals, whether or not people use such algorithm requires further direct experiments. Such experiments could test the effect of variables like the mean and the variance of cached values on the probability of expanding a node. On the theoretical front, our algorithm can benefit from several improvements, most notably, from relaxing the assumption that the environment has a deterministic transition structure. In that case, the algorithm could increase the efficiency of the state-of-the-art algorithms that use a plan-until-heuristic scheme in complex games [ 10 ]. Furthermore, whereas we simply assume here that planning and action execution cannot be performed in parallel, it is reasonable to assume that agents deliberate over upcoming choices while performing previously chosen actions.
Methods
We focus on deterministic Markov decision processes (MDPs). The environment is composed of a finite set of states S ; a finite set of actions A ; a (potentially partial) transition function T : S × A ⇸ S ; and a reward function f R : S × A × S → R . The agent interacts with the environment via a (potentially stochastic) policy π : S × A ⇸ [ 0 , 1 ] s.t. ∑ a π ( s , a ) = 1 for all s , with the goal of maximizing the expected value of the cumulative discounted rewards E [ R t | s t = s ] , where R t = ∑ i = 0 ∞ γ i r t + i , s is the start state, and γ is the discount factor. The state-action values of a policy π are defined as Q π ( s , a ) = E π [ R t | s t = s , a t = a ] . Finally, the optimal state-action values are defined as Q *( s , a ) = max π Q π ( s , a ).
We assume for now that the model-based (MB) system has perfect knowledge of the environment (i.e., the reward and transition functions) (we will relax this assumption later). The agent uses some of this information to build a search tree representation, which relates the current state s t to other states that can potentially be occupied in the future. The root of the tree is s t , and its immediate children include the one-step-reachable states.
Let us illustrate the formation of a search tree. The agent creates a tree node, containing information about her current state s t , which becomes the root of the tree, meaning all other nodes will stem directly or indirectly from it. The agent picks an action a available at s t to expand, which in turn adds s ′ ≔ T ( s t , a ) to the tree as a child node of s t . Now, if the agent continues planning, she can either expand an action from s t , assuming there are more than one action available at s t , or she can choose to expand from s ′. The planning process is composed of iteratively selecting an action to expand from the set of unexplored node-action pairs and adding the resulting new state to the tree as a new node.
Let us consider the state of a tree at a given time, containing a total number of n unexpanded node-action pairs. This means, there are n trajectories that start from s t and terminate at one of the unexpanded state-action pairs. We call each trajectory a “strategy”, denoted by A i , which is a tuple of state-action pairs, and introduce the search frontier F = { A 1 , A 2 , …, A n } as the set of all strategies for a given tree. We define expanding a strategy A by adding s ′, the immediate successor state of the unexplored state-action pair at the end of A , to the tree and adding the resulting new strategies to the frontier. These new strategies have the form A + 〈 s ′, a ′〉, where a ′ denotes any action available at s ′, and + is a tuple-concatenation operator. Note that after the expansion, if A is no longer unexplored—that is, has no unexpanded actions—then A is removed from F . This process of tree expansion goes on until an action is taken or the frontier is empty. The latter condition means the tree captures all possible trajectories in the MDP, which can only happen in an episodic MDP where no matter what actions the agent takes, she ends up in a terminal state (i.e., the state that ends the episode) after a finite number of actions.
We also assume that the agent has an estimation of the expected cumulative discounted rewards of each state-action pair 〈 s , a 〉, encoded by a random variable Q ( s , a ). A model-free (MF) system, for example, can represent such Q -values as random normal variables by tracking the first order statistics (i.e., mean) and second order statistics (i.e., variance) of the values [ 25 , 26 ]. Given that state-action values are the expected longterm discounted rewards, any stochastic estimation of it will be normally distributed given the Central Limit Theorem assuming a fixed sampling policy and a reasonable ( f R has finite variance for all 〈 s , a , s ′〉) reward structure. Thus, it is reasonable to represent Q ’s as random normal variables. With these settings, and in keeping with the plan-until-habit scheme, the value of a strategy A i that ends with an 〈 s M , a M 〉 at depth M with Q ( s M , a M ) ∼ N ( μ s M , a M , σ s M , a M 2 ) can be estimated by
V ( A i ) = r 1 + γ r 2 + γ 2 r 3 + ⋯ + γ M - 1 r M + γ M Q ( s M , a M ) , (3)
where each r i corresponds to the MB estimation of reward after taking the i th action in the strategy. Assuming that there is no uncertainty in estimating the immediate rewards (As discussed later, it is straightforward to relax the assumption of zero uncertainty for immediate rewards), r 1 , r 2 , .., r M , the total variance of V ( A i ) ∼ N ( μ i , σ i 2 ) is σ i 2 = γ 2 M σ s M , a M 2 . It can be seen that as a strategy gets deeper, MF value distributions (i.e., Q ’s) get discounted more, which will form the basis of our method.
We seek to compute the value of expanding the tree along A i . The agent knows that expanding A i will lead to a new, yet unknown state, s M +1 , where an action a M +1 with the highest Q -value, Q ( s M +1 , a M +1 ), among other actions of that state exists. This potential expansion will lead to a new strategy, A i * , with its value estimated by:
V ( A i * ) = r 1 + γ r 2 + γ 2 r 3 + ⋯ + γ M - 1 r M + γ M r M + 1 + γ M + 1 Q ( s M + 1 , a M + 1 ) . (4)
Note that r M +1 , s M +1 , a M +1 , and Q ( s M +1 , a M +1 ) are unknown prior to expansion. To reflect this, we use the notation V ¯ ( . ) to denote an unknown value estimation:
V ¯ ( A i * ) = r 1 + γ r 2 + γ 2 r 3 + ⋯ + γ M - 1 r M + γ M r ¯ M + 1 + γ M + 1 Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) , (5)
where r ¯ M + 1 and Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) denote, respectively, the immediate reward and the value distribution of the successor state-action pair, both unknown prior to expansion and thus, denoted with a bar (¯). Intuitively, E [ V ( A i ) ] should be equal to E [ V ¯ ( A i * ) ] , because they result from the same information prior to an expansion. Only with the extra information obtained from an expansion, namely after observing r ¯ M + 1 and Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) , the agent hopes to gain precision. In fact, we assume the agent’s probability estimates are coherent in the sense that her expectations of r ¯ M + 1 and Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) are in line with E [ V ( A i ) ] . Therefore, we have:
E [ V ( A i ) ] = E Q ¯ , r ¯ [ E [ V ¯ ( A i * ) | Q ¯ , r ¯ ] ] , (6)
where we drop the subscript M + 1 of r and arguments s ¯ M + 1 , a ¯ M + 1 of Q ¯ for brevity. This equality is also known as the law of total expectation, and here it suggests that an expansion may change the expected value of V ( A i * ) but not in expectation . We should emphasize that an agent does not necessarily need to obey this, but not doing so might result in inefficiencies. Particularly, if Eq 6 is not obeyed, then a Dutch book may be formed such that the agent would expect to lose value by performing tree expansions.
Also, note that,
Var Q ¯ , r ¯ [ E [ V ¯ ( A i * ) | Q ¯ , r ¯ ] ] ≥ Var [ E [ V ( A i ) ] ] = 0 , (7)
which means that while the agent knows the exact mean of A i ’s value ( Var [ E [ V ( A i ) ] ] = 0 ), the mean of the new strategy’s value is unknown prior to expansion. This variability in the expected value of the new strategy creates the possibility that the true (i.e., after expansion) expected value of A i * is even higher than the mean value of the best currently-expanded strategy. In fact, prior to expansion, the agent believes that acting on the basis of its currently-expanded tree will pay her max A ∈ F E [ V ( A ) ] , which is the mean value of the best strategy. However, if the true expected value of A i * is even higher than max A ∈ F E [ V ( A ) ] , then the agent can change her policy and “gain” extra reward. The expectation of this “gain”, given the distribution over the expected value of A i * , computes the value of expanding a strategy. In other words, expanding a strategy will yield a net expected increase (assuming the expanded strategy has variance in its value) in the expected value of the best strategy, which we refer to as the value of uncertainty resolution ( vur ). The vur along the strategy A i is equal to the expected value of policy improvement-induced reward resulting from observing r ¯ M + 1 and Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) Formally, given the current state of the search frontier F , vur ( A i | F ) is simply the difference between the expected value of best strategy after expanding A i (i.e., observing r ¯ M + 1 and Q ¯ ( s ¯ M + 1 , a ¯ M + 1 ) ) and before expanding A i :
VUR ( A i | F ) = E Q ¯ , r ¯ , [ max ( E [ V ¯ ( A i ) | Q ¯ , r ¯ ] , max A ∈ F - A i E [ V ( A ) ] ) ] - max A ∈ F E [ V ( A ) ] (8) ≥ 0 . (9)
where F − A i is the set F excluding A i assuming A i will be fully explored after expansion, and thus be removed from F . Otherwise, the max should run over F . The second (with minus) term in Eq 8 is the expected value of the best strategy in the frontier. The first term is the expected value of the best strategy after expansion. The vur is always non-negative because of Jensen’s inequality: max is convex and thus, the expectation of the max of random variables has to be larger than or equal to the maximum of expectations.
In order to progress further analytically, we make an assumption and assert that Var[ Q ( s M , a M )] = Var[ Q ( s M +1 , a M +1 )]. That is, we assume that MF value distributions for 〈 s , a 〉 and its immediate successor state-action pairs have the same uncertainty, possibly because the habitual system has had a similar number of experiences (i.e., samples) of neighboring actions and they are possibly of similar values. We can see in Eq 4 that only Q ( s M +1 , a M +1 ) contributes to the uncertainty in V ( A i * ) . Therefore we have,
V ( A i * ) ∼ N ( μ i * , γ 2 M + 2 σ s M + 1 , a M + 1 2 ) (10) = N ( μ i * , γ 2 ( γ 2 M σ s M , a M 2 ) ) (11) = N ( μ i * , γ 2 σ i 2 ) , (12)
where μ i * ∈ R is the mean, which we will obtain shortly, and σ i 2 = γ 2 M σ s M , a M 2 is the variance of V ( A i ). However, both V ( A i ) (magenta curve in Fig 1C ) and V ( A i * ) (black/grey curves in Fig 1C ) are estimating the value for the same action at the root state, s t . Therefore, the value distributions V ( A i ) ∼ N ( μ i , σ i 2 ) and V ( A i * ) ∼ N ( μ i * , γ 2 σ i 2 ) should be consistent as in Eq 6 , implying
V ( A i ) = E μ i * [ V ( A i * ) ] , (13)
which can only be satisfied if
μ i * ∼ N ( μ i , ( 1 - γ 2 ) σ i 2 ) . (14)
The distribution over μ i * represents the probability distribution of the expected value of a strategy after expansion. This variability comes from the fact that we will have additional pieces of information, namely r M +1 and Q ( s M +1 , a M +1 ).
Note that in equation Eq 10 , the only source of variance in A i * is assumed to be the variance in Q ( s M +1 , a M +1 ). In other words, the agent is assumed to have no uncertainty in estimating r 1 , r 2 , .., and r M . It is straightforward to relax this assumption by keeping track of the variance of r 1 , r 2 , .., and r M , denoted by σ r 1 2 , σ r 2 2 , . . , σ r M 2 . In that case, Eq 10 will be replaced by
V ( A i * ) ∼ N ( μ i * , σ r 1 2 + γ 2 σ r 2 2 + . . + γ 2 M σ r M + 1 2 + γ 2 M + 2 σ s M + 1 , a M + 1 2 ) (15) = N ( μ i * , σ i 2 - γ 2 M ( ( γ 2 - 1 ) σ s M , a M 2 + σ r M 2 ) ) , (16)
which gives
μ i * ∼ N ( μ i , γ 2 M ( ( 1 - γ 2 ) σ s M , a M 2 - σ r M 2 ) ) ) , (17)
where σ s M , a M 2 = γ - 2 M σ i 2 again.
This will take MB imperfection information about the reward function into account. Eq 10 also assumes that the agent has perfect information regarding the transition function. Given that our algorithm is only developed for MDPs with deterministic transition function, this assumption is feasible. Relaxing these assumptions (i.e., deterministic, and perfect knowledge of, transition function) are left for future work.
Relaxing the assumption on deterministic transition function would result the estimated value, V ( A i * ) , of the strategy A i * to become a mixture of Gaussians, rather than a simple Gaussian distribution. Computing μ i * and vur for such cases would significantly increase the computational cost of meta-cognition and hence, developing approximation methods would be required. For example, one could resort to Monte Carlo methods, where a set of transitions are sampled from the stochastic transition function, over which the vur is averaged.
Given we now know the distribution of μ i * , we can rewrite the vur definition given in Eq 8 :
VUR ( A i | F ) = E μ i * [ max ( μ i * , max A ∈ F - A i E [ V ( A ) ] ) ] - max A ∈ F E [ V ( A ) ] , (18)
where μ i * is distributed according to Eq 14 and F − A i is the set F excluding A i . We show in S1 Text that there is a closed-form solution for vur ( A i | F ) defined above.
Utilizing this uncertainty resolution mechanism, the agent can simply find the most promising strategy to expand, via arg max A i ∈ F VUR ( A i | F ) . The agent can continue expanding the search tree by reducing the uncertainties of the most promising branches until the value gained by expansion is less than the opportunity cost of expanding (as in [ 19 ]), or the search can continue until the working memory is full. The latter termination condition could be implemented based on the assumption that the working memory has a limited number of slots [ 27 , 28 ] (e.g., for storing states of the expanded tree). Alternatively, one could assume that the working memory is inherently corrupted by noise, and that the level of this noise increases with the number of items in memory [ 29 ]. It is straightforward to incorporate this mechanism into our algorithm: expansion results in the variance of V ( A i * ) to decrease by a factor γ 2 , but also increases by an additive factor that is proportional to the number of items (e.g. states) currently stored in the working memory. Thus, one can compute when the noise overwhelms the resolved uncertainty.
It is noteworthy that in this paper, computing vur is based on the assumption that when the value of expansion is bigger than its cost and thus an expansion should occur, an action will be executed immediately after that expansion. In fact, our model does not compute the value of further expansions following the next potential expansion. Relaxing this assumption would require computing the value of expanding all subsets of available and potentially-emerging strategies. In this case, for a certain subset like T 1 , T 2 , one needs to compute vur ( T 1 , T 2 | F ) and compare it with B . C , where B = 2 is the number of expansions being considered, and C is the cost of one single expansion. We show in S1 Text (section “on considering vur values independently”) that the value of expanding several strategies before performing an action is not necessarily equal to the sum of the value of expanding each of those strategies independently. In general, computing the optimal sequence of expansions for a budget of B would be NP-complete in B, as it reduces to stochastic knapsack problem [ 30 ].
Another interesting outcome of this model is that the relationship between vur and γ roughly follows an inverse U-shaped curve. If γ = 0, then V ( A ) as given in Eq 3 will be a scalar; as such, vur will be 0. If γ = 1, then the variance of E [ V ( A * ) ] as given Eq 14 will be zero, which too will result in vur being 0. The interpretation of these conditions is easy: if you do not care about the future, then no need to plan; and in the latter condition, the agent cannot gain precision by discounting the model-free estimates.
Supporting information
S1 Text
Proofs and derivations.
We provide vur -related proofs and derivations.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Speech conveys more than the linguistic message intended by a speaker. It provides information about the speaker such as physical characteristics, regional accent, and emotional state. Since multiple sources of information often converge on the same acoustic parameters, the two fundamental questions are how the different sources of information contribute to speech acoustics, and how listeners disentangle these sources of information during speech perception. In this study, we investigated the relationship between emotional tones of voice (emotions hereafter) and lexical tones by examining how four common emotions shape the acoustic characteristics of Mandarin tones, and how the emotions affect the perception of Mandarin tones.
Emotional tone is defined as the vocal expression of emotion, which conveys a speaker’s affective states, motivational states, or intended emotions [ 1 – 5 ]. The primary acoustic correlates of vocal emotions include fundamental frequency (F0), mean amplitude, and duration [ 1 , 2 , 6 ]. Previous research showed that F0 is the primary acoustic correlate of emotions [ 7 – 10 ], whereas amplitude and duration serve as secondary cues [ 11 , 12 ]. Importantly, F0 and amplitude are highly correlated with each other [ 8 ].
Lexical tones are used to distinguish words in tonal languages. In Mandarin, segmentally identical words can be distinguished on the basis of F0 height or contour. For example, the syllable /ba/ means “eight”, “uproot”, “grip”, or “father” with Tone 1 (a high-flat tone), Tone 2 (mid-rising), Tone 3 (mid-falling-rising), or Tone 4 (high-falling), respectively. The primary acoustic correlate of lexical tones is F0 [ 13 ]. Amplitude and duration also vary systematically among Mandarin tones [ 13 – 15 ], and both contribute to Mandarin tone perception as secondary cues [ 16 – 19 ]. However, F0 remains the most powerful cue for the perception of Mandarin tones [ 19 – 21 ].
Since the acoustic characteristics most relevant for lexical tones coincide with those for emotions, the convergence raises the question of how emotions affect the acoustics and perception of lexical tones. A tonal language like Mandarin offers a unique opportunity to examine this question.
Theories of emotion
The two approaches to the analysis of emotion are the dimensional theory of emotion and the theory of basic emotions [ 22 ]. The difference between these two approaches is that emotions are either described as independent dimensions [ 23 ] or discrete entities [ 24 ]. In the dimensional approach, Russell (1980) [ 23 ] proposed a circlex model of emotion, which showed that each emotion could be arranged in a circle controlled by two orthogonal dimensions in space: valence and arousal [ 25 – 28 ]. The position of each emotion on the quadrant reflects different amounts of valence and arousal traits [ 27 , 29 ]. The valence dimension is associated with a person’s subjective feeling, ranging from displeasure to pleasure. The arousal dimension is associated with the energy of a person’s subjective feeling, ranging from sleep to excitement [ 28 ].
The theory of basic emotions suggests that human emotions are composed of a limited number of basic emotions [ 30 ]. Each basic emotion has its proprietary neural circuits which are structurally different [ 24 , 25 , 31 , 32 ]. Although the idea of basic emotions is commonly accepted, there is no consensus on the exact number of basic emotions. Plutchik (1962) [ 33 ] proposed eight primary emotions (anger, fear, sadness, disgust, surprise, anticipation, trust, and joy). Ekman (1992) [ 24 , 34 ] proposed seven basic emotions (fear, anger, joy, sad, contempt, disgust, and surprise), but later changed to six (happiness, anger, sadness, fear, disgust, and surprise). Izard [ 35 ] proposed seven basic emotions (fear, anger, happiness, sadness, disgust, interest, and contempt). Recent studies examining facial expressions, neural mechanisms, and brain imaging suggest that the number of basic emotions could be further reduced to four (fear, anger, joy, and sadness) [ 36 – 40 ]. As an exploratory study of the tone-emotion relationship, we adopt the framework of four basic emotions (anger, fear, happiness, and sadness) in the current study.
How emotions affect speech acoustics
There is ample evidence that different emotions result in distinct acoustic characteristics [ 2 , 5 – 7 , 41 – 54 ]. Physiologically, the sympathetic nervous system is aroused by emotions such as anger, fear, or happiness, resulting in a higher heart rate and blood pressure, a dry mouth, and occasionally muscle tremors [ 55 , 56 ]. Consequently, speech is loud, fast, and has intense high-frequency energy. On the other hand, sadness arouses the parasympathetic nervous system. Heart rate and blood pressure decrease, salivation increases, and speech is produced slowly and with little high-frequency energy. These physiological changes are reflected in amplitude, energy distribution across the frequency spectrum, frequency of pauses, and duration. For example, higher arousal associated with excitement, fear, and anger have been shown to generate higher mean F0 [ 7 , 9 , 57 ], higher mean amplitude [ 58 – 61 ], and shorter duration [ 56 ].
The influence of specific emotions on speech acoustics varies across studies. When compared with a neutral tone of voice, an angry voice has a higher mean F0, a wider or similar F0 range, a higher mean amplitude, and a shorter duration [ 2 , 5 – 7 , 41 – 51 , 54 ]. A fearful voice shows a higher mean F0, a narrower, wider, or similar F0 range, a higher or lower mean amplitude, and a shorter duration [ 2 , 5 – 7 , 41 – 51 , 54 ]. A happy voice has a higher mean F0, a wider or similar F0 range, a higher or equal mean amplitude, and a shorter or longer duration [ 2 , 5 – 7 , 41 – 54 ]. A sad voice has a lower or similar mean F0, a narrower or wider or similar F0 range, a lower mean amplitude, and a longer duration [ 2 , 5 – 7 , 41 – 54 ]. In sum, some emotions have fairly consistent acoustic features, whereas other emotions are more variable. The variability, however, is consistent with the idea that emotion is sociocultural in nature, i.e., there are cross-linguistic and cross-cultural differences in the acoustic manifestation of emotions [ 62 ]. The variability is also consistent with the observation that features of emotions vary across speakers, sexes, and contexts [ 63 ].
Several studies compared the acoustic characteristics of emotions between tonal and non-tonal languages. Ross, Edmondson, and Seibert (1986) [ 64 ] examined acoustic characteristics of neutral, happy, sad, angry, and surprising emotions using Thai, Taiwanese, Mandarin (all tonal languages), and English (a non-tonal language). They found greater F0 variations in English compared to the tonal languages, suggesting that non-tonal languages have a greater degree of freedom in using F0 to convey emotions. In contrast, no significant difference was found in duration or amplitude between the tonal and non-tonal languages.
Anolli, Wang, Mantovani, and De Toni (2008) [ 65 ] investigated acoustic differences among happy, sad, angry, fear, scornful, prideful, guilty, and shameful emotions in Mandarin and Italian (a non-tonal language). They found that emotions were characterized by significant variations in F0 and amplitude for Italian but not for Mandarin. In contrast, duration varied significantly among the emotions for Mandarin but not for Italian. Since Italian is a syllable-timed language, it is also likely that the less variation of syllable duration in Italian reflects the language-specific prosodic structure.
Wang, Lee, and Ma (2016, 2018) [ 46 , 66 ] examined acoustic correlates of angry, fear, happy, sad, and neutral emotions in Mandarin and English. Semantically-neutral declarative sentences were embedded in different contexts to elicit angry, fear, happy, sad, and neutral emotions. Comparable English sentences were constructed with a direct translation of the Mandarin sentences. Acoustic analysis showed that F0 variations among the emotions were significantly greater in English than in Mandarin. In contrast, duration variations were significantly greater in Mandarin than in English.
Studies using other tonal languages also show more restricted F0 variations for emotions in a tonal language (Chong, Kim, and Davis, 2015 [ 67 ] for Cantonese; Luksaneeyanawin, 1998 [ 68 ] for Thai). To our knowledge, the only exception to this pattern is Li, Jia, Fang, and Dang’s (2013) [ 69 ], who showed greater F0 variations associated with emotions in Mandarin compared to Japanese, which is a non-tonal language that uses lexical pitch accent extensively.
Regarding the effect of emotions on specific lexical tones, Chao (1933) [ 70 ] noted that Mandarin uses successive additional tones and edge tones to implemet the intonation for emotions (see Liang and Chen, 2019 [ 71 ], for an illustration). Li, Fang, and Dang (2011) [ 44 ] examined how emotions affect the F0 and duration of Mandarin utterances ranging from one to fourteen syllables. The results showed that anger and disgust were associated with an additional falling tone, and happiness and surprise were associated with an additional rising tone. Non-neutral emotions resulted in a different F0 range, register, contour, or duration. For example, happiness and surprise were associated with a higher F0 range and higher register, whereas sadness and disgust were associated with a reduced F0 range and lower register.
In sum, most studies comparing tonal and non-tonal languages show that F0 variations associated with emotions are greater in non-tonal languages. This suggests that lexical tones constrain the availability of F0 for emotions in tonal languages. In contrast, amplitude or duration variations associated with emotions appear to be greater in tonal languages [ 14 , 64 , 65 ], suggesting that amplitude or duration may be used to compensate for the restricted use of F0 in conveying emotions. Studies on Mandarin further showed that emotions shape F0 and duration characteristics of Mandarin tones.
How emotions affect speech perception
The speech perception literature shows that emotions affect speech perception at various levels of processing. Mullennix, Bihon, Bricklemyer, Gaston, and Keener (2002) [ 1 ] examined how variations in emotions and talker voice affect spoken word recognition in English. They presented pairs of names (e.g., Todd-Tom ) produced by either the same or different talkers, and with the same or different emotions. The participants’ task was to judge whether the names in a pair were the same or different. The results showed that variations in emotion slowed down judgments of both the names and talker voices, indicating emotion affected perception of consonants and talker characteristics. Kitayama and Ishii (2002) [ 72 ] and Ishii et al. (2003) [ 73 ] presented words spoken in a pleasant or unpleasant tone of voice. While ignoring the emotional tone, listeners were asked to judge whether the word meaning was pleasant (e.g., grateful , satisfaction ) or unpleasant (e.g., complaint , dislike ). The results showed that emotion variations slowed down judgments of word meaning. Nygaard and Lunders (2002) [ 52 ] examined how emotions (happy, neutral, and sad) affect the perception of homophonic words (e.g., die/dye ). They found that selection of word meaning was compromised by the emotion of the words. Nygaard and Queen (2008) [ 53 ] presented happy (e.g., cheer ), sad (e.g., upset ), or neutral (e.g., chair ) words spoken with consistent, inconsistent, or neutral emotions. Listeners were asked to repeat the words they heard. The results showed that listeners responded more quickly when the meaning of the words matched the emotions.
Similarly, research on tonal languages show the impact of emotions on speech perception, and the effect is further modulated by tonal language experience. Singh, Lee, and Goh (2016) [ 74 ] examined how changes in emotion and Mandarin tone affect consonant recognition, and how consonant changes affect emotion recognition and Mandarin tone identification. For Mandarin-speaking listeners, variations in Mandarin tone and emotion made consonant recognition less accurate. Consonant variations also made Mandarin tone identification and emotion recognition less accurate. Consonants and prosody (Mandarin tone and emotion) affect each other to the same extent. In contrast, for English-speaking listeners, consonant recognition was affected by prosodic variation to a different degree than prosody recognition was affected by consonant variations. That is, the effects of emotions and lexical tones on segmental perception depend on tonal language experience.
Liang and Chen (2019) [ 71 ] examined how emotions and tonal language experience affect Mandarin tone perception. Four Chinese pseudo words (i.e., mong , ging , ra , bü ) were created, and each had four lexical tones variations. Each syllable-tone combination was embedded in the middle of carrier phrases. The syllable immediately before the pseudo words was manipulated to create four tonal contexts (i.e., chu1 , du2 , xie3 , lian4 ). For instance, mong1 was embedded in carrier phrases (1) zhi3 chu1 [mong1] zhe4ge0 zi4 “Please point out the word [ mong1 ]”; (2) wo3 hui4 du2 [mong1] zhe4 ge4 zi4” I can read the word [mong1]” ); (3) wo3 hui4 xie3 [mong1] zhe4 ge4 zi4 ” I can write the word [mong1] ”; (4) wo3 xiang3 lian4 [mong1] zhe4 ge4 zi4 ” I can practice the word [mong1] .” All stimuli were produced with an angry, happy, sad, or neutral emotion. Mandarin listeners and Dutch-speaking learners of Mandarin were asked to identify the Mandarin tone of the pseudo words. The results showed that stimuli produced with the neutral emotion resulted in higher accuracy than those produced with non-neutral emotions. However, only the angry voice resulted in significantly lower accuracy relative to the neutral voice for both groups of listeners. In addition, Tone 4 was identified more accurately than Tone 1 in the angry voice.
In sum, research on speech perception shows that emotions affect the perception of segmental phonemes, talker voices, and lexical tones. The effect of emotion on lexical tone perception depends on both stimulus characteristics and tonal language experience. Particularly relevant to the current study, Liang and Chen’s (2019) [ 71 ] findings further demonstrated that emotions affect lexical tone perception to different degrees depending on specific Mandarin tones.
How emotions are perceived in speech
In addition to understanding emotion’s effect on speech perception, we also explore how emotions themselves are perceived in speech. A common approach to studying emotion recognition is to recruit professional actors to produce speech materials with different emotions. Listeners are then asked to identify the emotions of the stimuli [ 2 , 4 , 7 , 50 , 75 – 77 ]. Studies using non-tonal languages showed that sad and angry voices are easier to identify than fearful and happy voices [ 2 , 4 , 7 , 50 , 75 – 77 ]. Studies using Mandarin have reported similar findings: negative emotions such as sadness [ 41 , 46 , 66 ] and fear [ 43 ] are easier to identify than positive emotions such as happiness [ 41 , 43 ]. It has been suggested that negative emotions are prioritized in vocal communication because they convey warnings in situations of attack, loss, and danger. Consequently, negative emotions need to be communicated more effectively to ensure human survival [ 43 , 78 , 79 ]. In contrast, positive emotions such as happiness are usually expressed through additional communication channels (e.g., facial expression), which may explain why a happy voice is identified with lower accuracy when only the vocal channel is used [ 7 , 43 , 50 ].
There is limited evidence on how lexical tones affect the perception of emotions. Wang, Ding, & Gu (2012) [ 80 ] investigated emotion recognition from Mandarin sentences by native and non-native speakers. Mandarin words with various tones ( qi4che1 “car”, zhao4pian4 “picture”, xin1fang2 “new house”, dian4nao3 “computer”, and xue2xiao3 “school”) were embedded in a semantically neutral carrier phrase ( zhe4 shi4 ta1 de0 [target word] “This is his [target word]”). The sentences were recorded with six emotions (happiness, fear, anger, sadness, boredom, & neutral) and presented to listeners for emotion recognition. The results showed that native listeners had an overall higher accuracy than non-native listeners, but both groups recognized sadness with the highest accuracy and boredom with the lowest accuracy. Since tones were not systematically manipulated, it is not clear whether the results could inform the effect of lexical tone on emotion recognition.
Wang and Lee (2015) [ 41 ] and Wang and Qian (2018) [ 47 ] constructed sentences composed exclusively of a particular Mandarin tone (e.g., wang1 bin1 xing1qi1tian1 xiu1 fei1ji1 “Wang Bin fixed the airplane on Sunday” or with a mixture of different tones (e.g., wo3 bu4gan3 xiang1xin4 zhe4 shi4 zhen1de0 “I cannot believe this is true”). The sentences were recorded with various emotions. It was hypothesized that emotions would be recognized less accurately in the Tone 1-only sentences because of the restricted F0 variation imposed by the (level) tone. An alternative hypothesis was that the restricted F0 variation in the Tone 1 sentences would have allowed emotions to surface more easily, thus facilitating emotion recognition. However, the results showed that emotions were recognized equally well regardless of tonal composition; that is, the restricted F0 variation imposed by the level tone did not compromise or facilitate emotion recognition. Emotional recognition seems quite robust irrespective of specific lexical tones.
Benefit of context
The presence of context can alter a listener’s interpretation of a speech sound [ 81 ]. Such perceptual adaptation forms the basis of speaker normalization [ 82 , 83 ]. In speech audiometry, a carrier phrase is typically included in a word recognition task to provide a cue for the listeners to focus their attention on the target words [ 84 – 86 ]. Previous studies showed that word recognition accuracy is typically higher when embedded in a carrier phrase [ 84 , 85 , 87 , 88 ]. The presence of a carrier phrase is particularly helpful under challenging listening conditions. For example, Lynn and Brotman (1981) [ 85 ] found that word identification in a carrier phrase was 10% more accurate than in isolation in the presence of speech-shaped noise. Since lexical tones produced with emotions are likely to deviate from the citation form, the presence of a context is likely to help listeners retrieve the intended tones more effectively.
In the current study, we examine Mandarin tone identification and emotion recognition in two contexts: when the target syllables are embedded in a carrier phrase (in context), and when the target syllables are extracted from the carrier phrase (in isolation). Note that the syllables in the “isolation” condition were not produced in isolation; rather, they were excised from the carrier phrase. We predict that Mandarin tone identification would be less accurate when the target syllables were extracted from the carrier phrase. This is because the citation form of a tone is likely to be altered due to the influence of neighboring tones [ 89 ]. Without the carrier phrase, it would be challenging for listeners to recover the tone. Furthermore, the carrier phrase provides information about talker characteristics such as speaking F0 range, which has been shown to facilitate tone identification from the multi-talker input [ 90 ]. We also predict that emotion recognition would be less accurate when the target syllables are presented in isolation. Since the talkers who recorded the stimuli were instructed to produce emotions for the entire utterance, the presence of emotions when preceded by a carrier phrase should facilitate emotion recognition from the target syllables.
The present study
The above review shows that both acoustics and perception of speech are shaped by emotions. Emotions affect Mandarin tone identification to different degrees depending on specific tones [ 71 ], but specific Mandarin tones do not seem to affect emotion recognition differently [ 41 , 47 ]. To further clarify the interaction between lexical tones and emotions in speech, this study examines the acoustics and perception of Mandarin tones produced with various emotions, and the perception of emotions embedded in Mandarin tones. Following Liang and Chen (2019) [ 71 ], we use syllables produced with four Mandarin tones in a sentence-medial position. Extending Liang and Chen (2019) [ 71 ], we use multiple speakers of both sexes to record the stimuli. We also examine both the acoustics and perception of Mandarin tones and emotions. Finally, we examine the perception of Mandarin tones and emotions in two contexts: when the target syllables are presented with the carrier phrase, and when the target syllables are extracted from the carrier phrase.
Based on prior research, we predict that the acoustics of emotions would be affected by specific Mandarin tones. Liang and Chen’s (2019) [ 71 ] findings lead us to predict that the accuracy of Mandarin tone identification would be affected by emotions to different degrees depending on specific emotions and Mandarin tones in the stimuli. Following findings from Wang and Lee (2015) [ 41 ] and Wang and Qian (2018) [ 47 ], we predict that emotion recognition would remain robust irrespective of the specific Mandarin tones in the stimuli. Finally, we predict that Mandarin tone identification and emotion recognition would be less accurate when the target words are extracted from the carrier phrase.
Experiment 1
In this experiment, we investigated the acoustic characteristics of Mandarin tones produced with emotions by multiple talkers of both sexes. Anger (ANGRY hereafter), fear (FEAR hereafter), happiness (HAPPY hereafter), and sadness (SAD hereafter) were selected because they are considered four basic emotions [ 36 – 40 ] (see introduction for opposing views). The acoustic effects of these four emotions were evaluated relative to the neutral tone of voice (NEUTRAL hereafter). Four acoustic measures including mean F0, F0 range, amplitude, and duration were chosen because they are most relevant to both lexical tone and emotional tone distinctions.
Based on the literature reviewed, we expect that ANGRY would result in a higher mean F0, a wider or similar F0 range, a higher mean amplitude, and a shorter duration when compared to the NEUTRAL baseline. FEAR would result in a higher mean F0, a narrower, wider, or similar F0 range, a higher or lower mean amplitude, and a shorter duration. HAPPY would result in a higher mean F0, a wider or similar F0 range, a higher or equal mean amplitude, and a shorter or longer duration. SAD would result in a lower or similar mean F0, a narrower or wider or similar F0 range, a lower mean amplitude, and a longer duration. These predictions are summarized in Table 1 . We also expect that the acoustic differences among the emotions would be modulated by specific Mandarin tones.
10.1371/journal.pone.0283635.t001
Table 1 Summary of predicted acoustic characteristics of the four emotions relative to the neutral emotion in Experiment 1. The symbols >, <, and = indicate an emotion is associated with a higher, lower, or comparable value compared to the neutral emotion.
Mean F0
F0 range
Mean amplitude
Duration
Angry
>
> or =
>
<
Fear
>
> or = or <
> or <
<
Happy
>
> or =
> or =
> or <
Sad
< or =
> or = or <
<
>
Method
Talkers
The use of human subjects in this study was reviewed and approved by the Institutional Review Board of National Yang Ming Chiao Tung University (IRB No. 1000063). Written informed consent was obtained from all talkers. No minors participated in this study. No medical records or archived samples were used in this study.
Eight professional actors (4 women and 4 men; mean age of 32.4 ± 7.1 years) were recruited to record the speech materials. All were native speakers of Taiwan Mandarin with no reported history of speech, hearing, or language disorders. Each talker was compensated $1,600 New Taiwan Dollars ($54 USD) per hour for their participation.
Speech materials
Three syllables /fa/, /ꓙi/, and /pΗu/ with the four Mandarin tones were selected as target syllables, resulting in 12 syllable-tone combinations [發/fa1/], [筏/fa2/], [髮/fa3]/, [法/fa4/], [西/ꓙi1/], [錫/ꓙi2/], [洗/ꓙi3]/, [夕/ꓙi4/], [鋪/pΗu1/], [葡/pΗu2/], [譜/pΗu3/], and [瀑/pΗu4/]. These syllables were chosen because: (1) they included the three most common vowels in Taiwan Mandarin [ 91 , 92 ], (2) all began with a voiceless or aspirated consonant to facilitate identification of syllable onset, (3) all syllable-tone combinations were real words in Mandarin, and (4) the meanings of all syllable-tone combinations were emotionally neutral.
The 12 syllable-tone combinations were paired with five emotions (ANGRY, FEAR, HAPPY, SAD, and NEUTRAL) and embedded in a semantically neutral carrier phrase / ni3 ʂuo1 [target word] ts5 / “You say the word [target word]”. In the carrier phrase, the syllable following the target syllable began with a voiceless consonant / ts5 / to facilitate identification of syllable boundaries. The 60 word-emotion combinations were produced twice by eight talkers for a total of 960 stimuli.
Procedure
Speech recordings took place in a sound-treated booth in the Department of Biomedical Engineering of National Yang Ming Chiao Tung University with a GRAS Type 40AC microphone at 0-degree azimuth. The microphone was placed 30 centimeters from the participant’s mouth. The sampling rate was 44,100 Hz with 16-bit quantization. Before the recording, the first author discussed with the actors the emotional tones that they should aim for. The actors then completed the recording in the booth while being monitored by the first author.
The 120 stimuli were recorded in five blocks separated by emotions. The order of the blocks and the order of the stimuli within a block were randomized for each participant. Before the recording started, the participants were given 10 minutes to familiarize themselves with the stimuli. Breaks were given between blocks for the participants to adjust their emotions. Our goal was to elicit a broad-focus analysis, i.e., distributing the prosodic change over the whole sentence instead of focusing narrowly on the target syllable. To that end, the participants were instructed to avoid pausing before and after the target syllables, and to avoid placing excessive emphasis on the target syllables.
Acoustic and statistical analysis
The recordings (except for the NEUTRAL stimuli) were rated by 30 native speakers of Taiwan Mandarin to evaluate how well the intended emotions were present in the speech materials. The raters included 19 women and 11 men with ages ranging from 21 to 49 years (mean age 33.5 ± 7.8 years). For each stimulus, the raters were asked in a four-alternative forced-choice task to choose an emotion (ANGRY, FEAR, HAPPY, or SAD) that best represented the speech sample they heard. The raters were also asked to provide a score on a Likert’s 5-point scale indicating the degree of the match, with 1 being the worst match and 5 being the best match. Stimuli were chosen for acoustic analyses only if they were correctly identified by all 30 participants and if they received an average rating of 3.0 and above. All stimuli met both criteria and were included in the acoustic analyses.
The acoustic analyses were performed with the Praat program [ 93 ]. Two landmarks were identified from the waveform: (1) the last glottal pulse of the syllable immediately before the target syllable, and (2) the last glottal pulse of the target syllable. The target syllable was then extracted based on these two landmarks. The acoustics measures were taken from the target syllables including mean F0, F0 range, mean amplitude, and duration.
Results
Fig 1 shows the F0 contours of the four Mandarin tones produced with the five emotions. The F0 contours were time-normalized and averaged over speakers of the same sex. Specifically, for each token, F0 was measured every 10% from the beginning (0%) to the end (100%) to obtain 11 data points. Each of these 11 points was then averaged over all speakers of the same sex.
10.1371/journal.pone.0283635.g001
Fig 1
F0 contours of target syllables as a function of emotion, Mandarin tone, and talker sex.
The F0 contours of the Mandarin tones appear to be consistent with traditional descriptions of the Mandarin tones in citation form: Tone 1 is flat, Tone 2 is falling then rising, Tone 3 (which is in a non-final position in the carrier phrase) is falling, and Tone 4 is falling but in a higher register. The F0 plot is meant to show that the Mandarin tones were produced as intended. For quantitative analysis of the F0 contours, a more rigorous approach such as Functional Data Analysis should be taken [ 94 , 95 ].
To evaluate the acoustic difference between the emotions, for each of the four acoustic measures (mean F0, F0 range, mean amplitude, and duration), a linear mixed-effects model was built for each of the four measures separately using R 3.6.3 (R Core Team, 2021) [ 96 ]. Mandarin tone (T1, T2, T3, and T4), emotion (ANGRY, FEAR, HAPPY, SAD, and NEUTRAL), and the tone-emotion interaction were entered as fixed effects. Talker, talker sex, syllable type, and repetition were entered as random effects.
Mean F0
Fig 2 shows the mean F0 of the target syllables as a function of emotion and Mandarin tone. Fig 1 suggests that the overall F0 contours of the Mandarin tones produced with the four emotions are similar to those of NEUTRAL; therefore, we calculated mean F0 as a summary measure for quantitative comparisons. The linear mixed-effects model revealed significant main effects of Mandarin tone, χ 2 (3, N = 8) = 973.95, p < .001, emotion, χ 2 (4, N = 8) = 1749.66, p < .001, and tone-emotion interaction, χ 2 (12, N = 8) = 70.18, p < .001. Post hoc pairwise comparisons (Tukey adjusted) were conducted to disentangle the interaction ( Fig 2 ). It was found that ANGRY had the highest mean F0 and NEUTRAL had the lowest. The ranking of FEAR, HAPPY, and SAD varies on specific Mandarin tones. Full output of the model is available in S1 Table .
10.1371/journal.pone.0283635.g002
Fig 2
Boxplot showing the mean F0 of target syllables as a function of Mandarin tone and emotion.
(* p < .05).
F0 range
Fig 3 shows the F0 range of the target syllables as a function of emotion and Mandarin tone. The linear mixed-effects model revealed significant main effects of Mandarin tone, χ 2 (3, N = 8) = 779.23, p < .001, emotion, χ 2 (4, N = 8) = 123.02, p < .001, and their two-way interaction, χ 2 (12, N = 8) = 114.64, p < .001. The results of post hoc pairwise comparisons are also shown on the figure. There does not appear to be a consistent pattern in the ranking of the emotions. Full output of the model is available in S1 Table .
10.1371/journal.pone.0283635.g003
Fig 3
Boxplot showing the F0 range of target syllables as a function of Mandarin tone and emotion.
(* p < .05).
Mean amplitude
Fig 4 shows the mean amplitude of the target syllables as a function of emotion and Mandarin tone. The main effects of Mandarin tone, χ 2 (3, N = 8) = 121.1, p < .001, and emotion, χ 2 (4, N = 8) = 1801.44, p < .001 were observed, but not their interaction χ 2 (12, N = 8) = 3.92, p = .98. Post hoc pairwise comparisons showed that ANGRY has the highest mean amplitude and NEUTRAL has the lowest. No difference was observed across SAD, HAPPY, and FEAR. Full output of the model is available in S1 Table .
10.1371/journal.pone.0283635.g004
Fig 4
Boxplot showing the mean amplitude of target syllables as a function of Mandarin tone and emotion.
Duration
Fig 5 shows the duration of the target syllables as a function of emotion and Mandarin tone. Similar to the mean amplitude, the main effects of Mandarin tone, χ 2 (3, N = 8) = 32.5, p < .001, and emotion, χ 2 (4, N = 8) = 639.74, p < .001 were significant but their interaction χ 2 (12, N = 8) = 14.2, p = .29 was not. Post hoc pairwise comparisons indicate SAD has the longest duration than all other emotions. Full output of the model is available in S1 Table .
10.1371/journal.pone.0283635.g005
Fig 5
Boxplot showing the duration of the target syllables as a function of Mandarin tone and emotion.
Summary and discussion
Emotions leave a mark on the acoustic characteristics of Mandarin tones. Consistent with previous studies [ 2 , 5 – 7 , 41 – 53 ], findings from our acoustic analyses support the observation that emotions shape the acoustic characteristics of speech for both tonal and non-tonal languages [ 63 – 65 , 97 ]. Our inclusion of all four Mandarin tones produced by talkers of both sexes further revealed that the impact of emotions varies depending on specific Mandarin tones. ANGRY has the highest mean F0 and mean amplitude. SAD has the longest duration. In contrast, we did not observe a systematic difference in F0 range.
Table 2 summarizes our findings compared to previous research. There are similarities but also discrepancies. Methodological differences such as talkers (amateurs in previous studies vs. professional actors in the current study) and materials (sentences in previous studies vs. syllables extracted from a carrier phrase in the current study) are likely to have contributed to the discrepancies. We used professional actors in the current study because they are typically more proficient in producing the desired emotions [ 7 , 63 ]. The presence of the intended emotions was verified in the current study with an independent emotion judgment task. As for materials, since the syllable is the tone-bearing unit in Mandarin, our choice to analyze syllables instead of sentences allowed us to examine the effect of emotions on specific Mandarin tones systematically.
10.1371/journal.pone.0283635.t002
Table 2 Summary of findings from Experiment 1 regarding the effect of emotion relative to the neutral emotion.
Mean F0
F0 range
Mean amplitude
Duration
Predicted
Actual
Predicted
Actual
Predicted
Actual
Predicted
Actual
ANGRY
>
>
> or =
> or =
>
>
<
=
FEAR
>
>
> or = or <
> or = or <
> or <
>
<
=
HAPPY
>
> or =
> or =
> or =
> or =
>
> or <
> or =
SAD
< or =
>
> or = or <
> or =
<
>
>
>
Among the acoustic measures, mean F0 and F0 range appear to yield the most consistent results between previous research and the current study. ANGRY, FEAR, and HAPPY consistently resulted in a higher F0. However, the utility of this measure is difficult to evaluate because of the lack of specificity in the predictions. For example, previous research showed that FEAR and SAD could result in a wider, comparable, or narrower F0 range in Mandarin tones compared to the neutral emotion. Although the current study showed the same results, no consistent patterns could be extracted without taking into consideration specific Mandarin tones.
Among the emotions, ANGRY consistently results in a higher mean F0, a greater F0 range, and a higher mean amplitude. This finding appears to support the proposal that negative emotions are conveyed more effectively in vocal communication to ensure human survival [ 43 , 78 , 79 ]. If an emotion results in relatively stable acoustic changes in lexical tones, recognition of that emotion is likely to be more robust. However, the other two negative emotions FEAR and SAD did not result in a similar pattern of consistent acoustic changes. It has yet to be determined if ANGRY has a special status among the negative emotions.
Experiment 2
Findings from Experiment 1 show that emotions shape the acoustic characteristics of Mandarin tones. The extent of the effect also depends on specific Mandarin tones. In Experiment 2, we ask how the acoustic changes induced by emotions would affect the perception of Mandarin tones. Liang and Chen (2019) [ 71 ] found that ANGRY resulted in lower accuracy of Mandarin tone identification relative to the neutral emotion. This difference was driven by more accurate identification of Tone 4 compared to Tone 1. We expect to find a similar interaction between Mandarin tones and emotions.
More generally, we evaluate two possibilities regarding how listeners interpret the acoustic signal in terms of Mandarin tones and emotions. On the one hand, Mandarin tone identification may be compromised by emotions because emotions make tones more variable acoustically, and thus more challenging to identify. In this scenario, greater acoustic changes (e.g., those associated with ANGRY) should lead to a less accurate tone identification. On the other hand, Mandarin tone identification may not be compromised by emotions if listeners are able to attribute the acoustic changes into the Mandarin tones and emotions, respectively. Assuming more predictable acoustic changes facilitate identification, more consistent acoustic changes (e.g., those associated with ANGRY) should lead to more accurate tone identification. It should be noted that neither scenarios assume that lexical tones are defined by absolute pitch, amplitude, or duration. Rather, they are two possible ways listeners parse the acoustic signal into distinct sources of acoustic variability.
In addition to Mandarin tone identification, we examine how well the emotions themselves could be recognized from the stimuli. Previous research has shown that negative emotions tend to be identified with higher accuracy than positive emotions. Furthermore, specific Mandarin tones do not seem to affect emotional tone recognition disproportionately [ 41 , 47 ]. The inclusion of both positive and negative emotions and all four Mandarin tones in the current study would allow us to evaluate these observations.
Method
Participants
This study was reviewed and approved by the Institutional Review Board of National Yang Ming Chiao Tung University (IRB No. 1000063). All participants signed a written informed consent. No minors participated in this study. No medical records or archived samples were used in this study.
Thirty-six adults (23 women and 13 men) with ages ranging from 19 to 23 years (mean age 20.08 ± 0.91 years) participated in Experiment 2. All participants were native speakers of Taiwan Mandarin and reported no known history of speech and hearing disorders. Each participant was paid $1,000 NTD ($34 USD) for their participation. All participants passed a screening of Mandarin tone identification in the neutral emotion presented in isolation and in context. Identification accuracy of Mandarin Tone 1, Tone 2, Tone 3, and Tone 4 was 97%, 93%, 93%, and 98% in isolation, and 100%, 100%, 99%, and 100% in context.
Stimuli
The stimuli used in this experiment were selected from one female and one male talker among the eight talkers who recorded the stimuli for Experiment 1. The talkers who received the highest rating in the emotion judgment task (reported in Experiment 1) in their respective sex group were chosen. For tone identification, all five emotions were included, resulting in 240 stimuli (3 syllables, 4 tones, 5 emotions, 2 repetitions, and 2 talkers). The 240 stimuli were presented in isolation or embedded in the carrier phrase, resulting in a total of 480 trials. For emotion recognition, the NEUTRAL stimuli and response option were excluded to prevent participants from using NEUTRAL as a default response for stimuli that they were uncertain about. As a result, there were 192 stimuli (3 syllables, 4 tones, 4 emotions, 2 repetitions, and 2 talkers). The 192 stimuli were presented in isolation or embedded in the carrier phrase, resulting in a total of 384 trials.
Procedure
This experiment took place in a sound-treated booth in the Department of Biomedical Engineering at National Yang Ming Chiao Tung University. The LabVIEW program (National Instruments) on a Windows 10 laptop computer was used for stimulus delivery and response acquisition. Stimuli were presented at each participant’s preferred hearing level over a pair of Beyerdynamic DT 990 PRO headphones.
The participant’s task was to listen to each stimulus and identify the Mandarin tone and the emotion of the target syllable. The stimuli were presented in four blocks in the following order: (1) isolated syllables for Mandarin tone identification; (2) isolated syllables for emotion recognition; (3) target syllables embedded in the carrier phrase for Mandarin tone identification; and (4) target syllables embedded in the carrier phrase for emotion recognition. The order of stimuli within each block was randomized for each participant. The Random Number Generator in LabVIEW was then used to generate a unique presentation order for each participant. Brief breaks were provided between the blocks.
For Mandarin tone identification, four response buttons marked with “一聲 (Tone 1)”, “二聲 (Tone 2)”, “三聲 (Tone 3)”, and “四聲 (Tone 4)” were displayed at the four corners of the computer screen and equidistant from the center of the screen. For emotion recognition, four response buttons marked with “生氣 (ANGRY)”, “害怕 (FEAR)”, “快樂 (HAPPY)”, and “傷心 (SAD)” were displayed at the four corners of the computer screen and equidistant from the center of the screen. The NEUTRAL response option was not included to avoid listener using NEUTRAL as a default for stimuli that they were not sure about. At the beginning of each trial, a cursor appeared briefly at the center of the screen, followed by the auditory stimulus. Listeners responded by clicking one of the four buttons on the screen using a computer mouse. The next trial was then presented 500 ms after the response. If participants were not sure about the tone identity, they were told to make their best guess. Each experimental session took approximately 60–90 minutes to complete.
Results
Mandarin tone identification
Fig 6 shows the accuracy of Mandarin tone identification as a function of emotion, Mandarin tone, and context. Overall, Mandarin tones were identified more accurately in context, and the effect of emotion appeared to be much more variable when the target syllable was presented in isolation. For example, all four Mandarin tones were identified quite well in NEUTRAL regardless of context, but the presence of other emotions compromised the identification of isolated Mandarin tones disproportionately compared to those presented in context.
10.1371/journal.pone.0283635.g006
Fig 6
Boxplot showing the Mandarin tone identification accuracy as a function of Mandarin tone, emotion, and context.
To evaluate these observations statistically, a mixed-effect logistic regression model was fitted to the Mandarin tone identification data. Mandarin tone (T1, T2, T3, and T4), emotion (ANGRY, FEAR, HAPPY, SAD, and NEUTRAL), context (in isolation and in context), and the tone-emotion interaction were entered into the model as fixed effects. Talker sex, syllable type, and repetition were entered as random effect. The dependent variable was the binary Mandarin tone identification (i.e., correct and incorrect). Full output of the model is available in S2 Table .
All main effects were significant: Mandarin tone, χ 2 (3, N = 36) = 83.12, p < .001; emotion, χ 2 (4, N = 36) = 486.38, p < .001; and context, χ 2 (1, N = 36) = 1534.71, p < .001. The tone-emotion interaction was also significant: Mandarin tone-emotion, χ 2 (12, N = 36) = 492.89, p < .001. Post hoc pairwise comparisons indicate that tone identification accuracy varied across different emotions ( Table 3 ).
10.1371/journal.pone.0283635.t003
Table 3 Summary of pairwise comparisons for Mandarin tone identification accuracy as a function of emotion. The symbol > indicates a significant difference ( p < .05) and = indicates no significant difference.
Emotion
Mandarin tone
ANGRY
T4 > T2 > T1 = T3
FEAR
T1 > T4 > T2 > T3
HAPPY
T2 > T3 > T4 = T1
NEUTRAL
T4 = T1, T4 > T3, T4 > T2, T1 > T2
SAD
T1 > T2 = T3 = T4
To examine the specific types of Mandarin tone identification errors, Tables 4 and 5 show confusion matrices of Mandarin tone identification responses for syllables presented in isolation ( Table 4 ) and in context ( Table 5 ). To facilitate interpretation of the confusion patterns, the most common error for each Mandarin tone that exceeds 20% is highlighted in gray . For syllables presented in isolation ( Table 4 ), Mandarin tones produced with NEUTRAL were rarely misidentified as other tones. There appears to be a response bias where Mandarin tones, particularly Tones 1 and 3, were most misidentified as Tone 4 when presented in an ANGRY tone of voice. Similarly, there appears to be a response bias for tones to be identified as Tone 1 when presented in a FEAR tone of voice. In contrast, the errors for Mandarin tones produced with HAPPY or SAD did not show distinct bias patterns.
10.1371/journal.pone.0283635.t004
Table 4 Confusion matrices of Mandarin tone identification responses for syllables presented in isolation across emotions. The most common error that exceeds 20% for each emotion is highlighted in gray.
Mandarin tone response
Emotion
Mandarin tone stimulus
Tone 1
Tone 2
Tone 3
Tone 4
Neutral
Tone 1
97%
2%
0%
1%
Tone 2
0%
93%
7%
0%
Tone 3
2%
2%
93%
3%
Tone 4
1%
1%
0%
98%
Angry
Tone 1
68%
12%
2%
18%
Tone 2
11%
71%
16%
2%
Tone 3
18%
6%
40%
36 %
Tone 4
9%
1%
3%
87%
Fear
Tone 1
86%
10%
2%
2%
Tone 2
26 %
58%
15%
1%
Tone 3
25 %
13%
46%
16%
Tone 4
20 %
2%
8%
70%
Happy
Tone 1
52%
41 %
3%
4%
Tone 2
5%
83%
11%
1%
Tone 3
4%
5%
66%
25 %
Tone 4
32 %
7%
5%
56%
Sad
Tone 1
83%
13%
3%
1%
Tone 2
13%
61%
25 %
1%
Tone 3
13%
13%
60%
14%
Tone 4
24 %
4%
17%
55%
10.1371/journal.pone.0283635.t005
Table 5 Confusion matrices of Mandarin tone identification responses for syllables presented in context across emotions.
Mandarin tone response
Emotion
Mandarin tone stimulus
Tone 1
Tone 2
Tone 3
Tone 4
Neutral
Tone 1
100%
0%
0%
0%
Tone 2
0%
100%
0%
0%
Tone 3
0%
1%
99%
0%
Tone 4
0%
0%
0%
100%
Angry
Tone 1
78%
1%
2%
19%
Tone 2
0%
98%
2%
0%
Tone 3
0%
1%
98%
1%
Tone 4
3%
0%
0%
97%
Fear
Tone 1
96%
2%
1%
1%
Tone 2
4%
93%
2%
1%
Tone 3
3%
8%
88%
1%
Tone 4
7%
1%
1%
91%
Happy
Tone 1
92%
1%
1%
6%
Tone 2
2%
96%
1%
1%
Tone 3
1%
2%
97%
1%
Tone 4
11%
2%
0%
87%
Sad
Tone 1
95%
3%
0%
2%
Tone 2
1%
97%
2%
0%
Tone 3
1%
4%
95%
0%
Tone 4
1%
2%
1%
96%
For syllables presented in context ( Table 5 ), Mandarin tone identification was remarkably accurate. There were only three instances where the accuracy fell below 90%, and only one of those was below 80%. In terms of confusion patterns, there was only one error that approached 20%, where an ANGRY Tone 1 was misidentified as Tone 4. There were no other dominant errors. In sum, the presence of the carrier phrase effectively neutralized the negative impact of emotions on Mandarin tones identification from isolated syllables.
Emotion recognition
Fig 7 shows the accuracy of emotion recognition as a function of emotion, Mandarin tone, and context. Overall, accuracy appears higher for target syllables presented in context than in isolation. Importantly, the four emotions appear to be affected by context to different degrees. For example, FEAR appears to be identified disproportionately worse in isolation.
10.1371/journal.pone.0283635.g007
Fig 7
Boxplot showing the emotion recognition accuracy as a function of Mandarin tone and context.
A mixed-effect logistic regression model with Mandarin tone (T1, T2, T3, and T4), emotion (ANGRY, FEAR, HAPPY, and SAD), context (in isolation and in context), and the tone-emotion interaction as fixed effects, talker sex, syllable type, and repetition as random effects were constructed. The dependent variable was the binary emotion recognition (correct and incorrect). Full output of the model is available in S3 Table .
All main effects were significant: Mandarin tone, χ 2 (3, N = 36) = 92.92, p < .001; emotion, χ 2 (1, N = 36) = 775.08, p < .001; and context, χ 2 (1, N = 36) = 1843.42, p < .001. The tone-emotion interaction was also significant: χ 2 (9, N = 36) = 327.95, p < .001. Post hoc pairwise comparisons indicate that emotion recognition accuracy varied across different Mandarin tones ( Table 6 ). For example, ANGRY was recognized most accurately in all but Mandarin Tone 3, whereas FEAR was recognized least accurately in all Mandarin tones.
10.1371/journal.pone.0283635.t006
Table 6 Summary of pairwise comparisons for emotion recognition accuracy as a function of Mandarin tone. The symbol > indicates a significant difference ( p < .05) and = indicates no significant difference.
Mandarin tone
Emotion
Tone 1
ANGRY = HAPPY > SAD > FEAR
Tone 2
ANGRY = HAPPY > SAD > FEAR
Tone 3
SAD > ANGRY = HAPPY > FEAR
Tone 4
ANGRY > HAPPY > SAD > FEAR
To examine the types of emotion recognition errors made by listeners, Tables 7 and 8 show confusion matrices of emotion recognition responses for syllables presented in isolation ( Table 7 ) and in context ( Table 8 ). To facilitate the interpretation of the confusion patterns, the most common error for each emotion that exceeds 20% is highlighted in gray .
10.1371/journal.pone.0283635.t007
Table 7 Confusion matrices of emotion recognition responses for syllables presented in isolation. The most common error that exceeds 20% for each emotion is highlighted in gray.
Emotion response
Mandarin tone
Emotional tone stimulus
ANGRY
FEAR
HAPPY
SAD
Tone 1
ANGRY
77%
9%
13%
1%
FEAR
11%
21%
56 %
12%
HAPPY
8%
6%
82%
4%
SAD
4%
29 %
25%
42%
Tone 2
ANGRY
76%
5%
17%
2%
FEAR
14%
36%
29 %
21%
HAPPY
5%
8%
76%
11%
SAD
2%
21 %
5%
72%
Tone 3
ANGRY
53%
15%
9%
23 %
FEAR
7%
34%
13%
46 %
HAPPY
12%
12%
49%
27 %
SAD
2%
22 %
3%
73%
Tone 4
ANGRY
93%
1%
3%
3%
FEAR
11%
37%
24%
28 %
HAPPY
8%
11%
71%
10%
SAD
6%
29 %
4%
61%
10.1371/journal.pone.0283635.t008
Table 8 Confusion matrices of emotion recognition responses for syllables presented in context.
Emotional tone response
Mandarin tone
Emotional tone stimulus
ANGRY
FEAR
HAPPY
SAD
Tone 1
ANGRY
97%
1%
1%
1%
FEAR
1%
87%
6%
6%
HAPPY
0%
2%
96%
2%
SAD
0%
15%
0%
85%
Tone 2
ANGRY
97%
2%
0%
1%
FEAR
1%
88%
4%
7%
HAPPY
0%
1%
97%
2%
SAD
0%
8%
1%
91%
Tone 3
ANGRY
96%
2%
0%
1%
FEAR
1%
85%
7%
7%
HAPPY
1%
2%
96%
1%
SAD
0%
6%
0%
94%
Tone 4
ANGRY
99%
1%
0%
0%
FEAR
1%
87%
3%
9%
HAPPY
1%
2%
95%
2%
SAD
1%
7%
0%
92%
For isolated syllables ( Table 7 ), emotion recognition accuracy ranged from 21% to 93%. There appears to be a response bias where SAD stimuli were most commonly identified as FEAR in Tone 1 and 4. Similarly, there appears to be a response bias where FEAR stimuli were commonly identified as HAPPY in Tones 1 and 2, but as SAD in Tones 3 and 4. In contrast, ANGRY and HAPPY stimuli did not result in any dominant bias patterns.
For syllables presented in context ( Table 8 ), emotions were recognized more accurately, ranging from 85% to 99%. In terms of confusion patterns, the only error that exceeded 10% was the SAD response to the FEAR stimulus in Tone 1 (15%). Because of the high accuracy, none of the emotions resulted in any dominant bias patterns.
Summary and discussion
As predicted, Mandarin tone identification accuracy varied across emotions, i.e., not all tones were identified equally well for different emotions. We replicated Liang and Chen’s (2019) [ 71 ] finding that Tone 4 is identified more accurately than Tone 1 in ANGER. Our results further revealed additional contingencies on emotion ( Table 3 ). Tone identification accuracy also varied across contexts. All tones were identified more accurately when the target syllables were embedded in the carrier phrase, but not all tones were identified equally well across the two contexts ( Table 3 ). For syllables presented in isolation, confusion analyses showed that Tone 4 was the most common error for the ANGRY stimuli, and Tone 1 was the most common error for the FEAR stimuli ( Table 4 ). For syllables presented in the carrier phrase, confusion analyses did not reveal any notable patterns except that Tone 4 was a common error for ANGRY Tone 1 stimuli ( Table 5 ).
The hypothesis that ANGRY stimuli would result in less or more accurate tone identification depending on how listeners interpret the acoustic signal was not supported by our data. Although ANGRY resulted in the greatest acoustic changes (as shown in Experiment 1), Mandarin tones produced with ANGRY were not identified disproportionately worse. Similarly, although ANGRY was associated with the most consistent acoustic changes (as shown in Experiment 1), Mandarin tones produced with ANGRY were not identified disproportionately better. With the extensive interactions between tone, emotion, and context, the current data is inconclusive as to how listeners parsed the acoustic signal into distinct sources.
Regarding emotion recognition, accuracy of emotion recognition varied across Mandarin tones, i.e., not all emotions were recognized equally well in different Mandarin tones ( Table 6 ). However, ANGRY was recognized most accurately in three of the four Mandarin tones. Like Mandarin tones, emotions were recognized more accurately when the target syllables were embedded in the carrier phrase. Unlike Mandarin tones, the order of emotion identification accuracy was more consistent between in isolation and in context ( Table 6 ), i.e., ANGRY was recognized better than HAPPY, followed by SAD, and FEAR was recognized least accurately. For syllables presented in isolation, confusion analyses showed that the most common error for SAD stimulus was FEAR response ( Table 7 ). For syllables presented in context, confusion analyses did not reveal any notable error patterns ( Table 8 ).
Taken together, data from Mandarin tone identification and emotion recognition revealed several similarities. Mandarin tones identification accuracy varied across emotions, just as emotions recognition accuracy varied across Mandarin tones. In addition, Mandarin tone identification and emotion recognition were both more accurate with syllables presented in the carrier phrase. The presence of the carrier phrase facilitated both Mandarin tone identification and emotion recognition, especially for those tones and emotions that were identified poorly in isolated syllables. Finally, for both Mandarin tone identification and emotion recognition, there were distinct error patterns in isolated syllables, but barely any notable error patterns in context.
There are also differences between Mandarin tone identification and emotion recognition. First, the accuracy of identifying specific Mandarin tones varied greatly across different emotions ( Table 3 ), but the accuracy of recognizing specific emotions was more consistent across different Mandarin tones. For example, ANGRY was consistently identified better than most other emotions, and FEAR was consistently identified worse than all other emotions ( Table 6 ). In other words, how well a Mandarin tone is identified depends heavily on specific emotions, but how well an emotion is identified does not depend as much on specific Mandarin tones. At first sight, this finding appears to suggest that emotion recognition is more robust than Mandarin tone identification. However, the accuracy of emotion recognition (ranging from 21% to 93%) was not higher than Mandarin tone identification (40% to 98%) in isolated syllables. The accuracy in the carrier phrase also seemed comparable: emotion recognition (85% to 99%); Mandarin tone identification (78% to 100%). Emotions do not seem to be inherently easier to identify compared to Mandarin tones. Rather, our data suggest that their mutual influence is asymmetrical.
Furthermore, Mandarin tone identification and emotion recognition differed in their interaction with context. The accuracy of identifying specific Mandarin tones varied substantially between the two contexts ( Table 3 ), but the accuracy of recognizing specific emotions was quite consistent between the two contexts ( Table 6 ).
Conclusions, limitations, and future directions
Three conclusions can be drawn from our findings. First, acoustic characteristics of Mandarin tones are shaped by emotions in complex but systematic ways, depending on specific Mandarin tones and specific emotions. Second, emotions affect Mandarin tone identification to a greater extent than Mandarin tones affect emotion recognition. Finally, the presence of carrier phrase facilitates both Mandarin tone identification and emotion recognition.
There are a number of limitations of the study. First, this study was conceptualized with four-emotion models; therefore the experimental materials were recorded with only four basic emotions. As noted in the introduction, there are different opinions as to the number of basic emotions, and basic-emotion models do not fully capture valence and arousal in the dimensional approach to emotion. For example, of the four emotions used in this study, only happiness is a positive emotion, and only sadness is considered low in arousal. Inclusion of a wider range of emotions would have allowed a more nuanced examination of the interaction between lexical and emotional tones. Second, only one carrier phrase was used; i.e., we did not systematically manipulate the tone that precedes the target syllable like Liang and Chen (2019) [ 71 ] did. It is not known whether our results would generalize to other tonal contexts. Third, the stimuli were produced by speakers of Taiwan Mandarin, and the participants in the perception experiment were also Taiwan Mandarin speakers. It is not known whether our results would generalize to other variants of Mandarin. Fourth, we did not include NEUTRAL emotion in the stimuli for the emotion recognition task. Since NEUTRAL emotion was included in the acoustic analysis, it would have been informative to include it in the perception experiment. Fifth, the syllables presented in isolation in the current study were excised from a carrier phrase; therefore, it is not known whether the findings would generalize to syllables produced in isolation. Six, since syllables in isolation were presented before syllables in context, familiarization from the isolation could have boosted response accuracy in the context blocks. Finally, we used the neutral condition as a baseline for emotion-related comparisons. Comparisons among the other emotions could generate further insights.
In addition to addressing these limitations, future studies could consider the following extensions. First, although the acoustic measures used in this study were commonly used in research on lexical tones and emotions, additional acoustic measures on voice quality and quantitative measures of F0 contour (e.g., Functional Data Analysis [ 94 , 95 ]) would be useful. Second, although our use of naturally produced stimuli in the perception experiment preserved all the acoustic cues available to listeners, it was not possible to isolate the contributions of specific acoustic cues. Future studies could systematically manipulate those cues to identify their individual contributions to lexical tone identification and emotion recognition. Third, although our data revealed how the acoustics and perception of lexical tones depend on factors including emotions and context, it is not clear why these factors affect lexical tone perception in such complex ways. Finally, the asymmetry in the mutual influence between lexical tones and emotions highlights the need for further research on how tonal language experience shapes lexical tone identification [ 71 ] and emotion recognition.
Supporting information
S1 Table
Full output of a linear mixed-effect logistic regression model for acoustic analysis of four Mandarin tones with five emotions.
(DOCX)
S2 Table
Full output of a linear mixed-effect logistic regression model for Mandarin tone identification.
(DOCX)
S3 Table
Full output of a linear mixed-effect logistic regression model for emotion recognition.
(DOCX)
S4 Table
All data of two experiments.
(RTF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Dosage compensation in eutherian mammals is normally achieved by random inactivation of one of the X chromosomes in females early in development, resulting in silencing of the majority, but not all, of the genes on the inactive X [1] – [3] . This remarkable epigenetic phenomenon is initiated at a locus on the X chromosome called the X-inactivation center (Xic). It involves multiple cis - and trans -acting factors that work through a number of steps which include counting of the number of X chromosomes, choice of which X chromosome will be inactivated, spreading of the inactivation signal from the Xic along the length of the chromosome, and maintenance of the inactivated state [4] , [5] . The Xist gene, which is located in the Xic and produces a non-coding RNA larger than 15 kb [6] – [10] , is an early, critical player in the initiation and propagation of the inactivation signal. An early step in X chromosome inactivation (XCI) is the cis binding of Xist RNA to the chromosome chosen for inactivation by a process which extends the length of that chromosome. The mechanism by which Xist RNA coats the inactive X chromosome and spreads this signal is unknown, but Xist binding precedes other changes in the chromosome including alteration of the methylation and acetylation states of histones, late replication, and methylation of CpG islands for those genes which undergo X inactivation [11] – [13] . Chromosomal localization studies have now shown that Xist coating of the inactive X chromosome leads to formation of a nuclear compartment which excludes the transcription machinery [14] – [16] , yet it remains unclear what molecular components are involved in this process.
Gartler and Riggs hypothesized that spreading of the inactivation signal involves some sort of booster elements that facilitate transmission along the chromosome [17] . Studies on the limited spreading of X inactivation in X;autosome translocations [18] , [19] and in Xist transgenes inserted into autosomes [20] led to suggestions that these booster elements, or way stations, may not be exclusive to the X, but enriched there. In 1998 Lyon hypothesized that LINE-1 (L1) retrotransposons may be the way stations [21] , partially because these elements are the most common dispersed repetitive sequences in mammalian genomes, are ubiquitous throughout the mammals, and are at a higher density on X chromosomes than on autosomes [22] – [29] .
L1 elements are the most highly represented retroelements in mammalian genomes [22] . Furthermore, in most species that have been examined to date, active L1s from the same species belong to a single lineage and active copies have very high sequence similarity [22] . Even where multiple lineages have been documented, one lineage produces the bulk of new insertions [29] – [31] . Although the high similarity among dispersed L1 copies was not part of the original evidence Lyon considered when she proposed a role for L1s in XCI, such similarity could be an important prerequisite for direct or indirect interaction with Xist.
In the time since Lyon's initial proposal that L1s may function as way stations for propagation of the Xist-mediated inactivation signal, extensive additional information has come to light [32] . Additional species have been shown to preferentially accumulate L1s on the X chromosome, and it has been shown that regions of the X that undergo inactivation tend to have a higher L1 density than those escaping inactivation [24] , [26] , [33] – [36] . L1s on the inactive X have been shown to be methylated later, and by a different methylase, than those on the active X [37] . Further studies of X;autosome translocations showing incomplete spreading or reduced maintenance of inactivation on autosomal segments have reinforced the idea that there is an inherent difference between the autosomes and the X [38] – [45] . Two groups have shown that L1 densities in X;autosome translocations appear to support the Lyon hypothesis [46] , [47] , and another group has shown that even though L1 elements appear to be a major factor correlated with X chromosome inactivation, a number of other DNA sequence features may also influence X inactivation [48] , [49] . Even on the autosomes, higher L1 densities are correlated with monoallelically expressed genes [50] .
Still other results have been interpreted as either not supporting the involvement of L1s in XCI, or suggesting a less critical role [34] , [51] , [52] . The identification of different types of heterochromatin on the X [53] and of important boundary elements [54] has highlighted suggestions that control of inactivation may occur at both the gene-specific level and the level of chromosomal domains [12] , [55] .
Yet the recent studies showing that accumulation of Xist on the inactive X chromosome leads to formation of a transcriptionally repressive nuclear compartment, continue to implicate involvement of repetitive sequences in XCI [14] , [15] . It has been shown that the compartment initially contains silenced repetitive DNA, with genic DNA only shown to enter the compartment during later silencing of those sequences. This raises the question once again of whether specific types of repeated sequences may be involved.
Even though the process now appears likely to involve cooperative interactions among multiple molecules and multiple regions of the Xist RNA, it is unclear which interactions are direct and which are mediated by other molecules. Similarly, despite large amounts of correlational evidence, the potential role of L1s in this process remains unclear. Their density patterns on chromosomes may reflect their function in X chromosome inactivation or may alternatively be a consequence of their biology and evolutionary history [36] . Even if L1s do play a role in XCI, that role could involve direct interaction with other components of the process, such as Xist, or could be more generally related to their repeated nature, nucleotide composition or other features. The prevalence of L1s in all mammalian genomes makes it difficult to discern whether their presence is a necessary part of the X inactivation process.
If L1 elements are indeed major players in this process, then loss of L1 activity in a species, followed by the mutational decay seen in previously deposited L1 sequences as they mutate over time, should have repercussions for X inactivation. The gradual loss of putative way stations by mutational decay would eventually lead to escape from inactivation for X-linked genes unless there is sufficient compensatory evolution. If interaction between L1s and Xist is direct, a higher rate of evolution of the Xist gene would be expected in order to ensure recognition of either way stations that are mutating to greater degeneracy or co-option of different repetitive sequences to serve as way stations. Even if recognition is indirect, increased evolution of intermediate factors might drive compensatory evolution of Xist .
We previously identified the first known L1 extinction event in a group of sigmodontine rodents in which L1 activity appears to have ceased about 8.8 million years ago [56] , [57] . These rodents have served as a model system to study the effects of L1 elements on genome evolution [58] . In this paper, we have used them to examine the repercussions of loss of L1 activity on X inactivation and Xist evolution.
We show here that the L1-inactive species, Oryzomys palustris , still exhibits X chromosome inactivation in spite of extinction of L1 activity. We have also isolated the majority of the Xist RNA from O. palustris and an L1-active sister species, Sigmodon hispidus . We compare these using published Xist sequences from the mouse, Mus musculus , and the vole, Microtus arvalis , as outgroups; we find no statistical support for more rapid evolution of the Xist RNA in O. palustris either by analysis of the overall dataset or by sliding window analysis of smaller regions. These results provide no support for the involvement of L1s in X chromosome inactivation, but additional interpretations remain possible.
Previously known tandem repeats within this region of Xist , some of which are known to be essential for Xist function, are found here to show no unusual changes, but O. palustris has gained a new repeat with potential to form RNA secondary structures. A number of regions of Xist are also found to be more highly conserved in all four species examined, implicating specific Xist regions for importance in X inactivation.
Materials and Methods
Tissues and DNA
Female (accession numbers TK47552 and TK72607) and male (accession no. TK72547) S. hispidus tissues were obtained from The Museum at Texas Tech University. Female (KE03 and F2) and male (KE02) O. palustris tissues were obtained from Dr. Kent Edmonds (Indiana University, New Albany, IN). Genomic DNA was extracted as previously described [59] .
Isolation of CpG islands and methylation analysis
CpG islands at the 5′ ends of the Zfx genes from O. palustris and S. hispidus were cloned after PCR amplification using previously described primers and amplification conditions [60] .
A methylation assay that does not require fresh tissue was employed due to the problems of working with non-laboratory species. This methylation assay used differential cutting of male and female genomic DNA by the restriction enzymes Hpa II and Msp I, followed by PCR and agarose gel analysis of the CpG islands associated with those genes [61] , [62] .
Isolation and sequencing of transcribed regions of the Xist gene
The majority of the region transcribed in the Xist gene was isolated from both O. palustris and S. hispidus by a combination of PCR and RT-PCR in three steps, walking from a 5′ region into overlapping 3′ regions. Degenerate primers were based on previous design considerations [63] and on alignments of previously isolated Xist genes. The most 5′ primer was also based on alignment of multiple copies of the A repeat from Mus musculus and three Microtus species [7] , [10] . Sequence obtained from the 3′ part of region 1 for each species was used to design nondegenerate 5′ PCR primers for isolation of region 2, and the same approach was used for region 3. Primer sequences and PCR conditions are in Methods S1 . Total RNA was isolated with RNeasy midi kits (Qiagen Corp., Valencia, CA), and cDNA synthesis was done using the Superscript 1 st Strand Synthesis System (Invitrogen Corp., Carlsbad, CA). The amplified region 3 sequence from S. hispidus is 1,940 bp shorter than the O. palustris sequence. This appears to be due to hybridization of the 3′ XT9R PCR primer to an interstitial area showing sequence similarity in both species.
Sequencing was done with an Applied Biosystems 3730 DNA Analyzer. Unless otherwise specified, contig analyses and initial sequence analyses were done using the LASERGENE (DNASTAR, Madison, WI) and Vector NTI (Informax, Bethesda, MD) analysis packages. The cloned Xist sequences present between the most 5′ and 3′ primers have GenBank accession numbers GQ201417 ( O. palustris ) and GQ201418 ( S. hispidus) .
Alignments, tandem repeat analysis, and phylogenetic analysis
Initial Xist alignments done using ClustalW included the following transcribed Xist regions from each of these species: bp 1–10,039 of the O. palustris sequence; bp 1–10,085 of the S. hispidus sequence; bp 621–10,541 from the exons of Microtus arvalis (GenBank accession no. AJ310129); and bp 638–12,170 of Mus musculus (GenBank accession no. NR_001463). The C repeat from the M. musculus sequence is present in only one copy in the other species, so all but one copy of this repeat was removed from the alignment by eliminating bases 3,173–4,678 of the Mus sequence. The final alignment of 11,615 characters was finished by manual adjustment (see Alignment S1 ). For phylogenetic analyses, gapped sites and regions where alignment ambiguity precluded determination of homology were removed, giving rise to a 7,284 character alignment. Repeat analysis was done using the LASERGENE package (DNASTAR, Madison, WI) and Tandem Repeats Finder [64] .
To conduct the likelihood-ratio test (LRT) of the molecular clock, a model of sequence evolution was selected using DT-M od S el [65] , and this model was verified to fit the data with an absolute goodness-of-fit test [66] . With maximum likelihood as an optimality criterion, P aup * 4.0 b10 [67] was used to estimate the best phylogeny constrained to fit the molecular clock and the best unconstrained phylogeny. The difference in log-Likelihood scores between these phylogenies (lnL clock −lnL unconstrained ) formed the test statistic that was evaluated against the χ 2 distribution.
Sliding window analyses were done by constructing likelihood trees for either non-overlapping adjacent windows or overlapping windows. The model of sequence evolution selected for the entire gene was used within each window, but parameter values were re-optimized for each of the windows. The command files used for non-overlapping sliding window analyses are in Analyses S1 .
Results
The Zfx gene shows X chromosome inactivation in females from both Oryzomys palustris and Sigmodon hispidus
To address the question of whether X chromosome inactivation can occur in a mammalian species that no longer has active L1 elements, we looked at the methylation status of the CpG island near the 5′ end of the X-linked gene, Zfx , in O. palustris and a sister sigmodontine species that has retained L1 activity, S. hispidus . A portion of the CpG island of the Zfx gene from each species was initially cloned and sequenced to confirm its identity, using PCR primers previously shown to amplify this island in other mammalian species. The methylation assay shown in Figure 1 was performed by digesting genomic DNA with either the methylation-sensitive enzyme, Hpa II, or the methylation-insensitive isoschizomer, Msp I, then PCR amplifying the CpG island. After Hpa II digestion and PCR, presence of the predicted band in females (but not in males) from both species shows that Zfx undergoes X inactivation in both species.
10.1371/journal.pone.0006252.g001 Figure 1
X-inactivation of the Zfx gene from S. hispidus (Shis) and O. palustris (Opal).
Genomic DNA from female and male individuals was digested with either Hpa II (H, methylation sensitive), Msp I (M, methylation insensitive isoschizomer of Hpa II), or Bam HI (C, control with no Bam HI sites in the regions amplified). Subsequent PCR for a portion of the Zfx CpG island, followed by gel electrophoresis, resulted in the band seen above only if there was no restriction digestion within the amplified CpG island. The amplification of a CpG island PCR product in females after digestion with Hpa II shows that the island contains methylated CpGs, and therefore an inactivated gene. Absence of a product after Hpa II digestion in males shows that their CpG island is nonmethylated and therefore capable of activity. Msp I cuts the CpG island, serving as a negative control, while Bam HI does not cut the CpG island, serving as a positive control.
This result, showing that XCI can still occur more than 8 million years after loss of L1 activity, suggests that highly similar L1 elements are not necessary for propagation of the X-inactivation signal in this species. On the other hand we reasoned that if L1s play a role as ‘way stations’ in XCI by direct interaction with Xist RNA, adaptive evolution of the Xist gene in O. palustris may have allowed L1s to retain their role in XCI despite the divergence of the extinct L1s in this species. Similarly, if another repetitive sequence replaced that function, the Xist gene might also be expected to undergo substantial evolution. This led to examination of the rate of evolution of the majority of the Xist RNA in L1-active and L1-inactive species.
O. palustris Xist RNA appears to be evolving at the same rate as S. hispidus Xist RNA
Transcribed Xist sequences were isolated from O. palustris and S. hispidus by a combination of PCR and RT-PCR. This resulted in isolation of the majority of the Xist RNA from both species and covered all of the regions shown by Wutz and coworkers [68] to mediate silencing and Xist localization to the inactive X chromosome in Mus musculus (top bar in Figure 2 ). The regions isolated from both species also include a portion of the 5′ tandem repeat, A, plus all of the other previously known major repeats: F, B, C, D, and E ( Figure 2 ). An alignment that covers approximately 10,050 bp of Xist was then modified by removal of gaps and regions showing alignment ambiguity and was used as the basis for phylogenetic analyses. Four species were included in these analyses: the L1-inactive O. palustris and its L1-active sister species, S. hispidus , plus two L1-active outgroups, Microtus arvalis and Mus musculus .
10.1371/journal.pone.0006252.g002 Figure 2
Maps of Xist sequences from four rodent species.
Xist RNAs from Mus musculus , Microtus arvalis , Oryzomys palustris , and Sigmodon hispidus are aligned, showing the six major previously described tandem repeat regions, A, F, B, C, D, and E, plus the G repeat unique to O. palustris (cross hatched boxes or vertical bars). Dotted lines indicate gaps inserted in species which have only a single copy of the C and G repeats. Mus and Microtus maps show full length RNAs, while the Oryzomys and Sigmodon maps show the regions isolated in this study. The shaded bar at the top is a functional map summarizing the previously shown relative importance of different regions in Mus for Xist silencing and chromosomal localization [68] . The amount of shading in each region of the top bar is proportional to its importance for activity in Mus : regions in black showed greatest importance; white regions were found not to be necessary for activity. An asterisk is added above the white bar covering the Mus C repeat because other work has suggested it may have an important function for Mus Xist [78] . The three regions amplified by PCR are delineated at the bottom. The 5′ primers for amplification of Xist from O. palustris and S. hispidus were located within the A repeat, so only two copies of that repeat from these species were recovered 3′ of the primer bindings sites.
In order to compare rates of Xist evolution in the above lineages, we used a likelihood ratio test (LRT) to evaluate the hypothesis that these genes were evolving according to a molecular clock [69] . The null hypothesis was that the rates of evolution did not differ among the four taxa. The HKY+I model of sequence evolution (see Methods Supplement file) was selected and found to have an acceptable absolute fit to the data ( p = 0.13). When the data were constrained to match the assumptions of the molecular clock, a single optimal phylogeny with −lnL = 21468.81106 was found. The unconstrained phylogeny had a −lnL = 21470.2427 and is shown in Figure 3 . As can be seen, the terminal branches for the O. palustris and S. hispidus sequences are of identical lengths. The difference between the likelihood values for the constrained and unconstrained phylogenies (δ = 2.8632) could not reject the molecular clock at the P = 0.95 level of significance (χ 2 distribution P 0.05 (2 df ) = 5.99147). This examination of the dataset in toto shows that all branches of the tree appear to be evolving at approximately the same rate, so there is no statistical support for more rapid evolution in O. palustris than in S. hispidus in spite of the loss of L1 activity.
10.1371/journal.pone.0006252.g003 Figure 3
Phylogenetic analysis of Xist sequences from four rodent species.
This maximum-likelihood tree was constructed using all ungapped characters that could be unambiguously aligned. Branch lengths are listed. The column on the right indicates whether each species shows L1 activity.
Sliding window analysis of contiguous Xist regions
It is possible that specific, small regions of Xist are evolving at different rates in each species but that this signal is lost in the total data set, so we also conducted the LRT of the molecular clock using a sliding window approach. Sliding window analyses can be useful in searches for molecular regions important for common functions across a number of species. These would be regions showing reduced evolutionary rates, and thus shorter terminal branches, in all the species examined. Alternatively, a relatively long terminal branch for a particular species, in a window showing short branches in the other species, might be a fingerprint of selection in that species for altered function or loss of function in that region.
Six analyses were done with non-overlapping window sizes of 50, 100, 200, 300, 400, and 500 bp. For each window, we conducted the test both by constraining the topology estimated from the window to the overall species phylogeny, and by search without this constraint. While this approach allowed us to identify particular regions of the Xist gene that violated the assumptions of the molecular clock, there was no a priori expectation of how many of these regions would result from stochastic processes associated with nucleotide substitution. We generated these expectations by simulating data under the null model (e.g., the phylogeny estimated with the molecular clock enforced) using the model of sequence evolution appropriate for our empirical data. For each window size, we repeated the analysis described above for 1000 simulated data sets.
Table 1 summarizes the results of the above analyses. The top row shows the results obtained when LRTs of the molecular clock were conducted in 14 windows by starting at positions 1–500, then moving in 500 bp steps (e.g., positions 1–500, 501–1000, etc.), until the end of the alignment was reached. No regions were identified where the clock hypothesis was rejected in either the analysis using topologies constrained to the species phylogeny or the analysis with that constraint relaxed. When 400 bp windows were used with constrained tree topologies, one region was found to reject the clock hypothesis, but parametric bootstrapping showed that this number of significant windows is within the expected range ( p = 0.626). All six window sizes were used in a similar fashion yet none of the 12 analyses showed statistical support for rejection of the molecular clock. The 34 trees arising from the windows showing non-clocklike evolution in Table 1 were visually examined to compare branch lengths, and none of those trees appeared to contain a long O. palustris branch. Thus, at these levels of resolution, there is not only no statistical support for an increased rate of evolution within the O. palustris sequences, but there is also no suggestive region identified for additional investigation.
10.1371/journal.pone.0006252.t001 Table 1
Sliding Window Test of the Molecular Clock.
Window size
Number windows
Number significant windows a , constrained topology b
P value
Number significant windows a , unconstrained topology c
P value
500 bp
14
0
1.0
0
1.0
400 bp
18
1
0.626
0
1.0
300 bp
24
1
0.717
1
0.622
200 bp
36
4
0.121
2
0.499
100 bp
72
5
0.701
3
0.435
50 bp
145
10
0.659
7
0.545
a Rejecting the molecular clock at p = 0.05 level. b Constrained to the species topology. c Not constrained to the species topology.
Sliding window analysis of overlapping Xist regions
To explore the evolution of Xist in more detail, we carried out a sliding window analysis similar to the one described above, except with overlapping sliding windows rather than contiguous windows. The results are shown in Figure 4 . This approach has increased resolution but precludes statistical analysis because overlapping windows are not independent. For this figure, maximum likelihood trees, unconstrained for the clock but constrained to the species phylogeny, were constructed as described in the previous section from 100 bp windows, but with only a 10 bp slide before moving to the next, partially overlapping window. The relative terminal branch length in each window for each species was then plotted in the figure. We find it striking that there appear to be so few sequence regions with either consistently low or high branch lengths across all four species. This may reflect the lack of sequence conservation which has been reported for Xist [9] , [10] . Two areas of relatively short branch lengths in all four species are the region at positions 1–200 and the region around position 6200 (grey areas in Figure 4 ). The 1–200 bp region includes the A repeat, which has been shown to be essential for Xist silencing [68] , [70] . The region around 6200 (bp 7806 in the O. palustris sequence) is slightly 5′ of the E repeat. It is centered on the exon IV sequence shown to be highly conserved throughout eutherian mammals [71] , [72] and predicted to form a stable RNA stem-loop structure involving more than 100 bases [73] . That putative RNA structure appears to be conserved in the O. palustris and S. hispidus sequences. Interestingly, we see no region that shows a relatively long branch in O. palustris coincident with short branches in the other species, and thus no evidence for accelerated evolution of Xist in the L1-inactive species.
10.1371/journal.pone.0006252.g004 Figure 4
Sliding window analysis of terminal branch lengths of Xist in four rodent species.
Likelihood trees were determined for 100 bp windows with a 10 bp slide between each window. The terminal branch length in each window for each species was then plotted. The upper numbers on the X axis indicate the position in the data matrix used for this analysis (see Supplementary Sliding Window file). The lower numbers in parentheses indicate the position in the original alignment before gaps and areas with ambiguous alignment were removed (see Supplementary Alignment file). The shaded areas show two selected regions with short branch lengths in all four species.
Tandem repeats, indels and potential secondary structures in Xist
One of the unusual features of Xist RNA from all species analyzed at this point has been the presence of a surprising number of tandem repeats, which have been proposed to have unknown functions in XCI [7] – [10] . In order to address the question of whether loss of L1 activity is associated with any major change in the amount of tandemly repeated Xist RNA in O. palustris , Tandem Repeats Finder was used under multiple stringency settings to determine the total number of bases of tandem repeats in the roughly 10 kb analyzed from each of the four species. Although there were differences between species, O. palustris did not show either an unusually high or unusually low amount of tandemly repeated DNA. Similarly, there were no major differences in the number and sizes of indels in the O. palustris sequence when compared to the S. hispidus sequence, and the overall length of this region ( O. palustris , 10,039 bases; S. hispidus , 10,085 bases) was very similar in each species.
The region analyzed in this study was chosen to include the six major types of tandem repeats that have been previously described in Xist . Sequence comparison shows that all six of the repeat regions, the A, B, C, D, E, and F repeat regions found in all other mammals examined to date, are also present in both S. hispidus and O. palustris ( Figure 2 ). The two copies of the A repeat isolated from both species here retain the same stem and loop sizes shown to be important for Xist silencing in Mus [68] , suggesting continued functional selection. The C repeat is tandemly duplicated 14 times in Mus musculus [7] but present in only one copy in all other species examined.
A new repeat, G, was found in O. palustris , arising from sequence within the D repeat region. The G repeat consists of four copies of an 84 bp unit. The most 5′ copy of this repeat is shown in Figure 5 aligned with the single copies of this region from each of the three other species analyzed in this study. In O. palustris this repeat was found to contain a highly conserved 7 base sequence (underlined in Figure 5 ) separated by one base in each unit from an inverted copy (also underlined). This could give rise to a stem-loop in each of the 4 repeats, or alternatively, pairing between the repeats could produce RNA pseudoknots. It is unclear whether the acquisition of this G repeat in a species which has lost L1 activity has any significance, but is interesting in light of prior suggestions of functional roles for both tandem repeats and secondary structures within Xist. No evidence was found for unusual changes in the other repeats.
10.1371/journal.pone.0006252.g005 Figure 5
Alignment of the O. palustris G repeat region.
The first monomer of the O. palustris G repeat, which is repeated four times, is shown in bold. It is aligned with the corresponding single copies in the other three species. Dashes indicate gaps. Periods indicate identity. The first group of 7 underlined bases is a highly conserved sequence separated by one base in O. palustris from an inverted copy (also underlined).
Discussion
If L1s play a major role in propagation of Xist along the inactivated X chromosome [21] , loss of L1 activity would certainly have ramifications for XCI. It would seem that XCI must occur in the L1-inactive species, O. palustris , since it shows normal XX-XY sex determination, but it was not known until now that this is indeed the case. Our finding of CpG island methylation in a gene on the X in females but not males indicates that XCI does indeed continue to occur in the absence of new L1 deposition indicates that L1 activity per se is not required for XCI, and calls into question any direct role for L1s in X inactivation.
If L1s do play a direct role in XCI, how might XCI continue in the absence of deposition of new L1s? Those elements already present would diverge by accrual of new mutations, and in the absence of retrotransposition of new L1copies, the sequence divergence among L1 elements would increase. What would be the ramifications of such a scenario? There appears to have been no deposition of new L1 elements in O. palustris for roughly 8.8 million years [57] . During that time the most recently deposited L1 remnants have diverged from their last active ancestor by 8.6% and from each other by 15.2%. A weak signal in genomic Southern blot analysis and a lack of evidence for X chromosome accumulation in fluorescent in situ hybridization provide supporting evidence for the substantial divergence among the fossil L1s remaining in the O. palustris genome [56] , [57] . The idea that L1s serve as way stations for Xist spread presupposes that at least some part of the L1 sequence is conserved within species. It is possible that the L1s still present in the O. palustris genome may not have deteriorated sufficiently by mutational decay for potential way station signals to completely lose function, or this species may have evolved a route for Xist spread that is independent of L1s. If either of these two scenarios is true, then Xist sequences might be expected to show a higher rate of evolution in O. palustris than in species that have retained L1 activity.
We found that rates of evolution in the Xist sequences analyzed here do not differ significantly between O. palustris and the three species with L1 activity ( S. hispidus , M. arvalis , M. musculus ). This result is consistent when the data are analyzed in toto as well as when various regions of the gene are analyzed separately with a statistically robust, sliding-window approach. Parametric bootstrapping demonstrated that the numbers of windows where the molecular clock hypothesis was rejected in the empirical data were within the expected range of stochastic variation, and since roughly 1/3 of the sites in this data set are variable, lineages evolving at different rates should be detectable with even the smallest windows used. It is worth noting that the most recently inserted L1s in O. palustris and S. hispidus are subject to very different evolutionary forces. Neutral substitution and increasing variation among previously deposited L1s predominate in O. palustris , while natural selection to maintain an active lineage and very closely related new L1 insertions prevail in S. hispidus . The equal rates of Xist evolution in the face of these very different evolutionary processes acting on elements in the L1-inactive and the L1-active species strongly suggests that Xist evolution is independent of L1 evolution.
Even though we found no statistical support for an elevated rate of evolution for O. palustris Xist RNA, we cannot unequivocally say that L1s are not involved in XCI. Many models for the interaction of Xist with other molecular components remain possible [74] , [75] . The nature of the molecular interactions potentially involving Xist and L1s could greatly affect the expected level of Xist evolution. Direct interactions between the two nucleic acids would be likely to elicit a larger evolutionary signal in Xist upon loss of L1 activity than a scenario in which the two nucleic acids interact indirectly within a complex. Likewise, it is reasonable to assume that the large functional changes required as a molecular complex evolves to recognize a non-L1 signal might lead to compensatory changes within Xist RNA, even if the two nucleic acids do not interact directly. Our data do not support either of these scenarios. But if the structure of an L1-interacting protein domain is relatively independent of the other molecules involved in the hypothetical complex, then compensatory Xist mutation would be less. Under that hypothesis, evolutionary effects on Xist might not be observed. If the role for L1s in XCI was not based on direct interaction, but rather on some general feature such as the A-richness of L1s, such role could be maintained in the absence of new L1 deposition for much longer than a role involving direct interactions, e.g., the AT-richness of the O. palustris L1s has been maintained since the extinction of L1 activity. An additional possibility is that a region of L1 elements outside of those we have analyzed is important for Xist function, leading to selection for conservation of that L1 region even while loss of L1 activity has lead to neutral mutational decay of the majority of each element.
It also remains possible that the approach used here would not detect small Xist regions altered in a molecular complex that includes L1s. High resolution analyses of these data, such as the one shown in Figure 4 , with Xist sequences from additional species may allow researchers to pinpoint other slowly evolving regions that may be important for general Xist function, and perhaps to identify candidate regions for L1 involvement in XCI.
It is interesting that the sliding window analyses done here detected the previously known conservation of both the A repeat and the exon IV likely stem structure region. This result suggests continued functional importance for these regions in all four species, yet in spite of this sensitivity, no candidate regions were identified for possible L1 mediated function.
An unusual aspect of all published Xist RNAs is their high level of tandem repeats, some of which have been shown to be functionally important, possibly working in a cooperative fashion [9] . The six major repeat regions previously identified were found to be present in the sequences we have analyzed in this study. Their conservation, plus the retention of likely secondary structure in the A repeat, suggests that no major changes have occurred with respect to these repeats. On the other hand, the acquisition of the new G repeat in O. palustris is noteworthy. Its potential for the formation of additional RNA secondary structures should be considered in terms of both general Xist function and a potential role in a species that has lost L1 activity.
We have explored the possibility that a new family of repetitive sequences may have arisen to high copy numbers in O. palustris to either replace putative functions of L1 elements or to fill the niche vacated by loss of L1 activity. We found a family of endogenous retroviruses, the mysTR family, which are present at unprecedented copy numbers of approximately 10,000 relatively closely related elements in this species [58] . However, it remains unclear at this point whether the mysTR family has any relevance to L1 extinction. No obvious difference in densities of mysTR elements on the X versus the autosomes in O. palustris was observed, and there is no evidence from the current study that Xist underwent rapid evolution to recognize mysTR as an alternate way station. Any correlation of mysTR amplification with loss of L1 activity may be clarified by our ongoing study of this family in a number of rodent species related to the ones described here.
We have also addressed the question of whether L1s might be involved in XCI by looking at L1 activity in a mammal which has lost the need for dosage compensation, the Ryukyu spiny rat, Tokudaia osimensis , which has an XO karyotype in both males and females. We reasoned that this karyotype might have evolved as an alternate pathway for dosage compensation after loss of L1 activity. However, we found that L1s are still active in this species and have continued to accumulate at a higher density on the X than on the autosomes [36] . These results support the idea that the higher density of L1s on the X may be due to reduced recombinational removal of L1s on the X, relative to the autosomes [28] , [76] .
The most parsimonious interpretation for the results presented in this study and in the Tokudaia study described above is that L1 elements are not involved in propagation of X inactivation. It remains possible, however, that the general rate of Xist evolution is so high [10] , or the region of way station recognition is so small, that adaptation in O. palustris cannot be detected by these methods. A critically altered region in Xist might also be outside of the area we have analyzed, even though this area includes all the regions that have been shown to be important for silencing and localization [68] , [77] . It is also possible that there has not yet been sufficient mutational decay of previously deposited L1s in O. palustris to elicit obvious compensatory evolution from homologous Xist sequences or that the new G repeat is the result of compensatory evolution in the presence of deteriorating L1 elements.
It is now clear that X inactivation can continue to occur in the absence of new L1 deposition for millions of years. The current study, like so many others aimed at testing the role of L1s in XCI, does not show unequivocally whether they are involved. However, in light of this study and recent work of others, perhaps it is time to consider that even if L1s facilitate XCI in some way; their role might be more complex than that of way stations for the spread of the inactivation signal. Greater progress might be made by focusing on X inactivation models that incorporate the increasing evidence for cooperation between molecules involved in the process, as well as by considering alternative explanations for accumulation of L1s on the X.
Supporting Information
Methods S1
PCR primers and methods for characterization of the Xist gene
(0.03 MB DOC)
Alignment S1
XIST alignment
(0.08 MB TXT)
Analysis S1
Commands for sliding window analysis of Xist. Areas of ambiguous alignment and gapped sites are excluded from 4the alignment. See Alignment S1 for the complete alignment.
(0.06 MB TXT)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The widespread occurrence of sex is usually attributed to the fact that recombination generates new gene combinations, thereby increasing the rate of adaptive evolution whilst negating Muller’s ratchet and the associated irreversible accumulation of deleterious mutations [1] . However, if sexual reproduction prevails in populations with small effective population sizes, consanguineous mating over several generations will result in severe inbreeding, potentially leading to inbreeding depression [2] . The deleterious effects of inbreeding are predicted to be least noticeable in populations with a long history of inbreeding, as this tends to purge deleterious mutations in the population [3] – [6] . Conversely, the beneficial effects of outcrossing are expected to be most pronounced in the most inbred populations [2] . This prediction is based on the assumption that in highly-inbred populations at least some loci will have been fixed for recessive deleterious alleles by drift, and that outcrossing will restore these loci to a heterozygous state [7] . Hybridization may result in heterosis (or hybrid vigour), an increase in fitness in the F1 offspring [2] . However, the benefits of outcrossing and hybrid vigour tend to decline in subsequent generations due to the breakdown of co-adapted gene complexes and epistatic gene interactions [2] , [8] .
Among micro- and macroparasites, hybridization has been observed at an intraspecific [9] – [11] as well as interspecific level [12] – [14] . It can lead to the emergence of new diseases [15] , [16] , with hybrid origins and/or current interbreeding events shown for viruses (e.g. Spanish flu, human rotavirus) [17] , [18] , bacteria (e.g. Haemophilus influenza ) [19] , fungi and oomycetes [20] , protozoa (e.g. Leishmania infantum , Trypanosoma cruzi ) [11] , [21] and various schistosome species [22] , [23] . Immediate consequences of hybridization may involve increased pathogen fecundity, infectivity, virulence and transmission rates, wider host spectra, a shorter maturation time and phenotypic changes [15] , [16] . For instance, hybrid vigour in F1 offspring of Leishmania infantum / L. major crosses led to increased resistance to immunity in an atypical vector [24] . However, the fitness advantages of hybridization are typically short-lived; most laboratory hybrids that show hybrid vigour lose their fitness in subsequent generations [8] , [25] , [26] .
For gyrodactylids (a specious group of monogeneans), there is phylogenetic evidence that co-infecting species may hybridize before and after host switches [14] , [27] – [29] . These ectoparasitic monogeneans are ubiquitous on teleost fish [30] and are largely transmitted between fish by direct host-host contact with a single parasite being sufficient to seed a population [31] . Most give birth to live young, which are already pregnant when born, a phenomenon termed hyperviviparity [32] . The first born daughter is derived asexually by mitotic division when the mother is still an embryo; the second daughter can be produced by automictic parthenogenesis, and all other daughters (up to five in total recorded to date) are either produced via parthenogenesis or potentially sexual reproduction [31] . Hyperviviparity in combination with extreme progenesis allows these parasites to produce offspring in as little as 24 h (e.g. G. turnbulli ) [33] resulting in explosive population growth. It has been hypothesized that during epidemic population growth, sexual reproduction is more common than parthenogenesis due to crowding effects [34] , [35] . However for all monogeneans, the occurrence of sexual recombination has been assumed (e.g. [28] ), but never actually proven.
In this study, we used a microsatellite marker to confirm sexual reproduction in monogeneans. This methodology was then applied to assess the effects of hybridizing three G. turnbulli strains that have been isolated and inbred in the laboratory for 1, 8 and 12 years (circa 2×10 2 to 2×10 3 generations; assuming 2 days/generation). We then (conservatively) estimate the proportion of parasites that have a hybrid origin. By comparing the infection trajectories, maximum parasite burdens and duration of infections in parental and mixed populations, we provide evidence for hybrid vigour and/or inter-strain competition. By analysing time-to-infection for parasites transmitting to fish infected with same-strain or different-strain individuals, we show inbreeding avoidance behaviour within strains.
Materials and Methods
This work was conducted using the guppy ( Poecilia reticulata ) – Gyrodactylus turnbulli model system with all procedures conducted under UK Home Office licence (PPL 30/2357) regulations and approved by the Cardiff University Ethics Committee.
Source of Animals and Infection Procedures
Guppies ( Poecilia reticulata , Peters) were originally isolated from a wild population from the River Tunapuna (20P, 678513E, 1177415N), Trinidad, and thereafter maintained in laboratory cultures for ca. 8 years. The fish were maintained at 25±1°C with a 12 h light : 12 h dark cycle at a density of approximately 120 fish per 40L aquarium, and fed at least twice daily on Aquarian® fish flakes plus Artemia and/or Daphnia spp. in addition to bloodworm once a week.
Three different Gyrodactylus turnbulli strains were used: Gt3 (isolated from a pet shop guppy in Nottingham in 1997), Gt1 (isolated from wild guppies in the Lower Aripo River, Trinidad, in November 2001) and Gt8 (isolated from a pet shop guppy in Cardiff in March 2008), which are routinely maintained at Cardiff University at 25±0.5°C on ornamental guppies. In this study, crosses are made between these lines and we use the term ‘hybrid’ or ‘hybrid genotype’ sensu stricto : “(1) a progeny of a cross between parents of different genotype; (2) heterozygote” [36] .
Experimental infections utilized donor guppies carrying >20 parasites per fish from either of the three laboratory strains ( Gt1, Gt3 and Gt8 ) and which were anaesthetized with 0.02% Tricaine Methanesulfonate (MS222, PharmaQ, UK). Guppies that were naïve to these parasites were also anaesthetized, sexed, measured and infected by bringing the infected donor guppy into contact with the recipient until the gyrodactylid moved from one fish to the other. To ensure infection by one parasite strain was not biased by prior exposure to another strain, half of the fish were first infected with one strain (i.e. first infection), then the other (i.e. second infection) so avoiding any ‘priority effects’. Extreme care was taken to prevent cross-infection of individual fish with the wrong strain of parasite by changing anaesthetic and glassware after each infection procedure. At various stages during Experiments 1 and 2, parasites were removed from anesthetized hosts with fine watchmaker’s forceps and stored in 100% ethanol for subsequent molecular analysis.
Experiment 1: Sexual Recombination in Gyrodactylids
To maximise our chances of detecting hybrid genotypes and to achieve high parasite infrapopulations (assuming sex may only occur at high parasite densities) [34] we began the infection experiment with 10 parasites on each fish. In total, 24 guppies were infected with ten monogeneans each, five from two different parasite strains, resulting in 12 fish infected with Gt3 and Gt1 (parasite line Gt3 × Gt1 ) and 12 fish infected with Gt8 and Gt1 (parasite line Gt8 × Gt1 ). Fish were maintained in pairs in 1L plastic pots to which two naïve fish were added the subsequent day (D1). Thereafter, water was changed every day and any deceased guppies were replaced with naïve fish. On D14, the parasite cultures were screened and the most heavily infected fish (1–2 fish per culture, infected with over 100 parasites) were euthanized with an overdose of MS222 and fixed in ethanol. A sample of parasites from these highly infected, fixed fish were removed with entomological pins and stored individually in fresh 100% ethanol.
Experiment 2: Hybrid Fitness and/or Inter-strain Competition
In this experiment, recipient fish (n = 101) were infected with either two parasites of the same strain (paternal parasite lines: Gt1 × Gt1 , Gt3 × Gt3 and Gt8 × Gt8 ; hereafter the “pure parasite population”) or two parasites from different strains (hybrid parasite lines: Gt3 × Gt1 and Gt8 × Gt1 , hereafter the “mixed parasite population”). The time-to-infection success (i.e. the time a parasite took to transfer from the donor fish to the recipient fish during infections) [37] was recorded for both parasites. The infected recipient fish ( Gt3 : n = 10; Gt1 : n = 11; Gt8 : n = 11; Gt3 × Gt1 : n = 38; Gt8 × Gt1 : n = 31) were maintained individually in 1L pots.
On day 1 (D1) all fish were screened following procedures by Schelkle et al. [38] to check whether the infection had established (i.e. ≥2 parasites on a fish were regarded as successful establishment). From D1 onwards, all fish were screened every other day to follow the population dynamics of the parasites until the fish were screened clear for at least three times [38] . Additionally, on D7 half of the parasites (median: 6, range: 1–26) were removed with watchmaker’s fine forceps from the anaesthetized host and any monogeneans from fish containing mixed parasite population were fixed for subsequent molecular analysis to assess the presence of hybrid genotypes on individual fish. Halving the parasite load at this stage ensured we had samples to genotype while still allowing the fitness experiment to continue on a like-for-like basis (i.e. all parasite infrapopulations were halved).
Microsatellite Genotyping
The three parasite strains Gt1 , Gt3 and Gt8 were previously genotyped at four microsatellite loci [39] . However, only a single locus (TurB02) unambiguously discriminates between two of the three lines (i.e. was fixed for different alleles at two of the three lines). Hence, we use this locus to estimate the rate of sexual reproduction. Gt3 and Gt8 have different alleles to Gt1 at the TurB02 locus, which allows us to detect potential hybrid genotypes between Gt3 and Gt1 , and Gt8 and Gt1 , but not Gt3 and Gt8 . Therefore, a mixed population of both latter strains was not included in this study.
In Experiment 1, a random sample of 48 gyrodactylids from 10 different mixed parasite cultures were screened to examine whether any hybrid (i.e. heterozygous) individuals were present, which would unambiguously establish sexual reproduction. In Experiment 2, 240 parasites (139 gyrodactylids isolated from 23 Gt3 × Gt1 infected fish, and 101 from 23 Gt8 × Gt1 infected fish) were successfully genotyped. Extraction was performed using the following lysis protocol: 50 µl of lysis buffer containing 1xTE, 0.45% Tween 20 and 9 µg of Proteinase K was added to a whole gyrodactylid and incubated at 65°C for 30 min and then denatured at 95°C for 10 min. PCR amplification was performed as described in [39] using 5 µl Master Mix (Qiagen multiplex PCR kit), 2 pmol of each microsatellite primer (TurB02: forward ACGAGTGACAATAAAGCTGG , reverse ATCAATAGTTGAATGG ) and 2 µl of DNA following manufacturer’s instructions. The PCR cycling profile consisted of 15 min at 95°C, followed by 40 cycles of 40 s at 95°C, 90s at 52°C and 90 s at 72°C and a final extension for 30 min at 72°C. PCR products were loaded on an ABI3130XL using size standard ROX-350 (Applied Biosystems) and chromatograms were analyzed using PeakScanner version 1.0 (Applied Biosystems).
Parasites were identified as being of hybrid or parental strains based on their genotypes. All parental strains were homozygous for the locus used: Gt3 and Gt8 parental genotypes presented only allele 244 bp, whereas Gt1 presented the allele 246 bp. This enabled us to distinguish hybrid genotypes in crosses with Gt8 × Gt1 and Gt3 × Gt1 . Since just one microsatellite locus was used to differentiate between heterozygous or homozygous individuals, non-detection of hybrid genotype monogeneans could have occurred for five reasons: (1) One of the parasite strains did not establish after infection, but the parasite infection was still recorded as being established. (2) No sexual reproduction occurred. (3) No sexually produced parasites were sampled. (4) Sexually reproduced parasites were homozygous for the microsatellite loci due to Mendelian segregation of alleles over several generations. (5) Sampled parasites were backcrossed individuals. Hence, hybridization rates quoted for both experiments below are significant underestimates and should be interpreted as a conservative estimate of sexual reproduction in our study system.
These five factors that potentially inflate the false-negative rate (i.e. the conclusion that no hybridisation occurred) also have implications for Experiment 2. Due to this potential non-detection of hybrid genotypes, analysis of the infection trajectories in Experiment 2 was only performed on replicates for which molecular analysis had revealed the presence of hybrid genotype parasites. This data was directly compared with the control treatments (parental strains only). Fish on which only parental genotypes were found, despite infection with two different strains, were discarded from the analysis. The rationale was that homozygous F2, F3, etc. and back-crossed individuals, and parasites that did not engage in sex, were indistinguishable from asexually produced individuals or non-establishment of one of the strains.
Statistical Analysis
For Experiment 1, the occurrence of sex was assessed by scoring the presence of homozygous and heterozygous parasites on infected fish. Differences in cross-breeding success between parasite populations were assessed using a χ 2 contingency table test in Minitab v15. A direct comparison of hybrid genotype frequency with Experiment 2 could not be conducted due to the differences in timing of gyrodactylid collection in the experiments.
We investigated whether there was evidence for hybrid vigour and/or inter-strain competition in mixed parasite population by comparing the infection intensity trajectories, maximum parasite burdens and duration of infections in parental and mixed populations in Experiment 2. We used a combination of general linear models (GLMs) and a general linear mixed model (GLMM) based on restricted maximum likelihood (REML) analyses in R and ASReml-R (R v2.9.2) [40] . All models used a Gaussian error distribution and an identity link function. For mixed parasite populations only data from fish with known hybrids (confirmed by molecular analysis; n = 12) were considered. These data were compared and contrasted to the data of the pure lines (n = 32). For infection intensity trajectories and maximum parasite burden data were normalised by a natural logarithmic, ln(x), transformation within the model; data for the day of peak parasite burden and parasite loss did not need transformation. For the GLMM, animal ID (to account for repeated parasite counts on the same fish) and days post-infection (with a spline to account for non-linear effects) were added to the random model. Standardized residual distributions were assessed visually with histograms and normality plots plus Shapiro-Wilk normality tests for all models. Initial model terms are presented in Table 1 . All model terms were assessed as significant with α = 0.05 as critical value, and models were reduced with stepwise deletions (using the Log-Likelihood method for random terms in the GLMM). Finally, the minimal model for the GLMM was used to predict the infection trajectory on individual, isolated hosts.
10.1371/journal.pone.0039506.t001 Table 1
Model terms and interactions in the GLM and GLMM analyses used for Experiment 2.
Infection trajectory measure (model)
Dependent variable
Independent continuous (cont.)/categorical variable (cat.)
Infection intensity
Parasite intensity:
Parasite population (cat.)
trajectory (GLMM)
Ln (x) transformed
Fish length (cont.)
Fish sex (cat.)
Days post-infection (cont.)
Animal ID (cat.) *
Spline (days post-infection) (cont.) *
Maximum parasite
Maximum parasite burden:
Parasite population (cat.)
burden (GLM)
Ln (x) transformed
Fish length (cont.)
Fish sex (cat.)
Fish length X sex interaction
Day of maximum
Day of maximum parasite
Parasite population (cat.)
parasite burden (GLM)
burden
Fish length (cont.)
Fish sex (cat.)
Fish length X sex interaction
Parasite loss (GLM)
Day of parasite loss
Parasite population (cat.)
Fish length (cont.)
Fish sex (cat.)
Fish length X sex interaction
*
random terms in GLMM.
We also examined inbreeding avoidance behaviour within strains by evaluating time-to-infection for parasites transmitting to fish infected with same-strain or different-strain individuals. Parasite time-to-infection was analysed with a survival analysis in R v2.9.2 using Cox’s Proportional Hazard and time-to-infection data for the second parasite to infect a fish. The initial model included two parameters: whether the second parasite infecting the recipient fish was a same or different strain individual compared to the first monogenean, as well as the strain origin of the second parasite. Parasite establishment was assessed with a binary logistic regression in Minitab using fish length, sex and parasite population as independent variables.
A χ 2 contingency table test in Minitab or Fisher’s Exact Tests (available at http://www.physics.csbsju.edu/stats/contingency.html ) [41] were used to detect (1) differences in recovery from parasite infection (i.e. screened clear from parasites on three subsequent screens) [38] ; (2) percentage of fish carrying hybrid genotypes or not; and (3) percentage of hybrid genotypes in the two mixed populations on fish with hybrids.
Proportion of hybrid genotype gyrodactylids among all genotyped parasites and fish mortality were analysed using a binary logistic regression in Minitab including intended parasite population (pure or mixed) in the former and intended parasite population, fish length and sex in the latter.
Results
Sexual Reproduction and Hybridization
Gyrodactylid strains interbred freely, with 37.5% and 34.6% hybrid genotypes recovered at D14 from the Gt3 × Gt1 and Gt8 × Gt1 parasite lines, respectively, in Experiment 1 (high density starting infection). There was no apparent difference in hybrid genotype frequency between crosses (Contingency Table Test: χ 2 = 0.201, d.f. = 1, P = 0.654). In Experiment 2 (low density starting infection), twelve fish were confirmed to carry hybrid genotype parasites: five (out of 32) initially infected with Gt3 × Gt1 (15.6%) and seven fish (out of 31) initially infected with Gt8 × Gt1 (22.5%). In this experiment significantly more hybrid genotype parasites were detected for the Gt8 × Gt1 population (10.9%) than in the Gt3 × Gt1 population (3.7%; Binary Logistic Regression: G = 3.841, d.f. = 1, P = 0.05; Fig. 1 ).
10.1371/journal.pone.0039506.g001
Figure 1
Frequencey of sexual reproduction in gyrodactylids.
Proportion of hybrid parasites (±95% CI) for the Gt3 × Gt1 and Gt8 × Gt1 parasite populations in Experiment 2. A significantly higher proportion of hybrids were detected in the Gt8 × Gt1 than in the Gt3 × Gt1 cross.
Fitness Effects of Mixed Strain Infections
The GLM and GLMM minimal models for each dependent variable analysed in Experiment 2 are presented in Table 2 . Mixed gyrodactylid populations exhibited hybrid vigour and/or inter-strain competition at a population level (GLMM: Wald = 5.805, d.f. = 1, p = 0.021). Fig. 2A shows the initial growth phase after which there is no clear pattern between parasite populations due to the onset of fish immunity. Hence, in Fig. 2B we use predicted data from the GLMM to visualise different infection trajectories.
10.1371/journal.pone.0039506.g002
Figure 2
Infection trajectory of pure and mixed Gyrodactylus turnbulli populations on individual guppies ( Poecilia reticulata ).
(A) Mean parasite intensity in pure (blue) and mixed parasite populations (green) (± Standard error of the mean; Experiment 2) in the initial phases of infection showing increased fitness in the mixed parasite population. Increased variation after day 7 is due to the onset of the host’s immune response and is not displayed in this graph. (B) Predicted mean parasite burden over time (controlled for fish standard length, ±95% Confidence Intervals) showing that increased parasite fitness in mixed parasite populations leads to faster population growth and higher maximum parasite burdens.
10.1371/journal.pone.0039506.t002 Table 2
Minimal models for the GLM and GLMM analyses in Experiment 2.
a. Infection trajectory. Model type: GLMM (random terms = fish ID, days post-infection)
Model term
Wald statistic (F)
d.f.
P
Parasite population
5.805
1
0.021
Fish length
12.7
1
0.001
Days post-infection
8.669
1
0.004
b. Maximum parasite burden. Model type: GLM
Model term
Wald statistic (F)
d.f.
P
Parasite population
7.024
1, 40
0.012
Fish length
12.639
1, 40
>0.001
c. Day of peak parasite burden. Model type: GLM
Model term
F statistics
d.f.
P
Parasite population
4.963
1, 32
0.033
Fish length
16.327
1, 32
>0.001
d. Day of parasite loss. Model type: GLM
Model term
F statistics
d.f.
P
Fish length
13.343
1, 33
>0.001
For GLMs the term, then the residual degree of freedom are given.
The apparent superiority of mixed infections was reflected in a higher parasite burden over time and increased maximum parasite burden, but not a longer duration of infection. Mixed infrapopulations reached their maximum parasite burden of 48.6±15.2 parasites per fish, which was significantly higher than the 44.2±18.3 parasites in pure parental infrapopulations (GLM: F 1,40 = 7.024, p = 0.012; Fig. 3a ). However, this peak of infection occurred later in mixed populations at 8.3±0.5 days compared to 6.7±0.6 days in pure populations (GLM: F 1,32 = 4.963, p = 0.033; Fig. 3b ). Parasite infrapopulations were lost at 14.6±1.2 and 15.5±1.5 days post-infection in mixed and pure populations, respectively (GLM: F 1,34 = 1.103, p = 0.302).
10.1371/journal.pone.0039506.g003
Figure 3
Maximum parasite burden and day of maximum parasite burden in pure and mixed parasite populations.
(A) Median maximum parasite burden (log 10 transformed) and (B) median day of maximum parasite burden. Outliers are represented by dots; the bars are the lower and upper limits; and the box represents the 1 st and 3 rd quartile with the median. Significant differences between mixed and pure parasite populations for both A (P = 0.012) and B (P = 0.033).
The infection trajectory itself was also influenced by days post-infection (GLMM: Wald = 8.669, d.f. = 1, p = 0.004) and host length (GLMM: F 1 = 12.7, p = 0.001). Larger hosts showed a higher maximum parasite burden (GLM: F 1,40 = 12.639, p<0.001) which was reached slightly later (GLM: F 1,32 = 16.327, p<0.001). Similarly, parasites were lost slightly later in large guppies (GLM: F 1,33 = 13.343, p<0.001). There was no effect of host sex on maximum parasite burden, day of peak parasite burden, day of parasite loss or the infection trajectory.
Parasite Infection, Establishment Success and Host Survival
In Experiment 2, infection success for parasites used to infect each fish was 100% for all replicates in all treatments. Time-to-infection of the second monogenean was significantly slower if the parasite on the fish (from the first infection) was from the same strain (25.3±2.2 s) than from a different strain (11.8±5.6 s; Survival analysis: z = −2.809, P = 0.005, n = 100; Fig. 4 ). Establishment success, measured on D1, increased with fish size (Binary Logistic Regression: G = 10.57, d.f. = 3, p = 0.014 for overall test; Z = −2.07, p = 0.039 for fish length, odds ratio 0.62 (95% CI: 0.39–0.98)), but was not affected by parasite population (pure: 96.9%, mixed: 89.9%; Z = 0.92, p = 0.359 for parasite population, odds ratio 2.85 (95% CI: 0.30–26.76)). There was no significant difference in fish mortalities between the parasite populations (Binary Logistic Regression: G = 5.125, d.f. = 3, p = 0.163): 16.1% mortality in pure and 25% in mixed populations.
10.1371/journal.pone.0039506.g004
Figure 4
Time-to-infection for inter- and intrastrain infections.
Time-to-infection (s; natural log transformed) was significantly lower for secondary parasites infecting a fish already infected with a different (inter-)strain parasite rather than a same (intra-)strain parasite individual. The bars are the lower and upper limits; and the box represents the 1 st and 3 rd quartile with the median.
Discussion
This study unambiguously demonstrates the occurrence of sexual recombination in a monogenean parasite, Gyrodactylus turnbulli , by making experimental crosses and using a microsatellite marker to identify sexually derived parasites (i.e. outbred or hybrid genotype offspring). Although it has long been assumed that sex does occur in monogenean hermaphrodite parasites based on their observed inter- and intra-specific mating behaviour and phylogenetic studies [28] , [31] , [35] , there was no direct evidence for the production of viable sexually produced offspring. Importantly, we also show that outcrossing (hybridisation) between monogeneans of previously inbred populations and/or inter-strain competition significantly increases various fitness components of the parasites, resulting in a higher parasite burden over time and an increased maximum parasite burden, but not a longer duration of infection.
Previously, cytogenetic observations of decondensation of the sperm nucleus after the fusion of sperm and egg cell had suggested that sexual recombination occurs in gyrodactylids [31] . However, release of male genetic material into the egg cytoplasm does not necessarily equate to sexual recombination since any sperm material can be expelled from the egg. Sperm dependent parthenogenesis, for example, is displayed by the hermaphrodite flatworm Schmidtea polychroa [42] . Further evidence of genetic exchange in the evolutionary history of gyrodactylids through sexual reproduction comes from previous phylogenetic studies (e.g. G. salaris, Macrogyrodactylus spp.) [14] , [29] . However, the current study is the first to conclusively demonstrate sexual recombination in monogeneans in contemporary populations and establishes that, at least in the laboratory conditions of the present experiment, between 3.7% and 10.9% of all genotypes are recent hybrids with ancestors from two different inbred lines.
Confirmation of recombination in viviparous gyrodactylids has important consequences for the evolutionary history of this specious group of flatworms. Over 400 species of gyrodactylids have been described and host-switching rather than co-evolution with the fish host appears to be the main mechanism of speciation [43] , [44] . Sexual reproduction between diverging populations can potentially facilitate the evolution of new species [7] , [45] , particularly in combination with a switch to a host not previously populated by either of the parental species [29] , [46] . In more recent times, host switches and hybridisation events are further facilitated by changes in range distribution associated with climate change and global fish transport. The resulting co-occurrence of previously separated species in the same habitat [47] , [48] and the new opportunities provided in these novel environments could augment the rate of hybridisation in the wild.
The occurrence and frequency of sexual reproduction in gyrodactylids may be species-specific and condition-dependent. Based on changes in haptor morphology over 20 generations of individually maintained parasites, Harris [49] speculated that G. gasterostei reproduction is largely clonal. As explained in the Material and Methods , our estimates of sexual reproduction for G. turnbulli are significantly downward biased given that we used only a single microsatellite locus to identify sexually reproduced offspring. Thereby, we will have missed hybrid genotype gyrodactylids that are homozygous for our marker locus due to, for instance, Mendelian segregation or back-crosses. The rates of sexual recombination varied by nearly a factor of three between our crosses. These differences may be due to the limitations of our methodology, or they could reflect natural variation in the rate of outcrossing.
Even a low frequency of sex in hermaphrodites is sufficient to avoid the negative fitness consequences associated with inbreeding [2] . Empirical data from the facultative sexual single-celled chlorophyte Chlamydomonas reinhardtii show that particularly in large hermaphrodite populations, sex will be maintained as a reproductive mode to overcome inbreeding depression [50] – [52] . Mate preference for unrelated individuals is likely to evolve as this reduces the fitness costs of sexual reproduction with a related individual. We previously speculated about inbreeding avoidance occurring in Gyrodactylus species [44] and empirical data from the current study is consistent with this argument, showing that the time-to-infection differs according to the relatedness of the novel gyrodactylid compared to the already existing infection. This suggests that these parasites may recognize conspecifics with similar genotypes. The mechanism for this remains unknown, but could be related to a chemical cue released through the parasite’s glands or excretory products into fish mucus [44] . Humans, mice and fish are known to use chemical communication (i.e. olfaction) to make self-referential decisions of attractiveness based on genetic variation at the Major Histocompatibility Complex (MHC) [53] . Many vertebrates show preferences for partners with dissimilar genotypes, which are identified via MHC-related chemical cues [54] . In invertebrates, behavioural mechanisms to avoid inbreeding are more commonly reported than the utilization of olfactory cues (e.g. insects: Gryllus bimaculatus and the subsocial spider, Anelosimus cf. jucundus ) [55] – [57] , but there is also evidence that the innate immune system may be involved in mate choice [54] . In the only other parasitic platyhelminth in which inbreeding avoidance has been examined, individuals of the cestode Schistocephalus solidus preferred siblings to unrelated parasites as mates [58] . Hence, to our knowledge, the current study provides the first evidence of inbreeding avoidance behaviour in platyhelminths.
Once sexual reproduction occurs, the main fitness benefit in hybridising gyrodactylid populations appears to be the higher maximum parasite burdens and extended time of population growth before the host’s immune response appears to cause an infra-population decline. Pure parasite populations started declining approximately seven days after initial infection, whereas mixed parasite infra-populations continued to grow on average for a further two days. These results are consistent with the hypothesis that: (i) hybrid genotypes are more tolerant (parasite damage limitation) and/or resistant (limitation of parasite burden) [59] to the fish immune response allowing them to maintain a reproducing population on the host for longer than parental parasites; (ii) hybrid genotypes are better at evading the host’s localised immune response (see van Oosterhout et al. for the effects of parasite mobility and distribution on parasite population dynamics) [60] ; (iii) hybrid genotypes do not activate the immune response as rapidly as parental parasites, leading to delay in the onset of the host response; and/or (iv) they benefit from hybrid vigour (heterosis) and have increased fecundity relative to the inbred parental lines (see Cable & van Oosterhout for evidence of inbreeding depression in the Gt3 line) [61] . Alternatively, the high fitness of mixed strain parasite populations could be caused by inter-strain competition, possibly resulting in increased virulence [62] . However, mixed parasite populations did not cause higher host mortalities in the current study, indicating that despite increased growth rates in parasite populations, virulence of hybrid genotypes was not increased compared to parental populations. Further evidence against inter-strain competition is the lack of evidence for competitive exclusion among other monogeneans co-infect the same host [63] , [64] .
The exact reasons for the increased fitness of mixed parasite populations cannot be disentangled in this experiment, but warrant further study. Given the reduction in fitness of the Gt3 line relative to recently wild-caught G. turnbulli ( Gt1 ) [61] , we favour the hypothesis that hybridisation ameliorated the detrimental effects of inbreeding during the up to ∼2×10 3 generations in captivity, and that this resulted in hybrid vigour. There are implications of this study for the fitness of hosts and parasites in wild populations. In natural environments, a high parasite burden is associated with increased extrinsic mortality of this host, for example due to increased risk of downstream displacement during flash-flooding [65] . This suggests that the increased parasite load in hybrid genotype infections could increase host mortality in the wild. Similarly, secondary (bacterial, fungal, etc.) infections may be more common with higher parasite burdens. In addition, other factors such as infectivity, transmission abilities and host specificity, which have not been considered for this experiment, could also be affected by hybrid vigour, and the effects of hybridisation on these life history traits and infection dynamics warrant further investigations.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The uncemented Symax hip stem was developed as an optimization of the uncemented Omnifit hip stem [ 1 ]. The design considers the geometry of the stem, surface texture, and type and extent of the osseointegrative coating [ 2 , 3 ]. Previous studies have proven early ingrowth with histological and histomorphometric analyses on retrieved implanted Symax hip stems, exclusively into the proximal part of the stem, as a result of the BONIT-hydroxyapatite (HA) coating [ 2 ]. Furthermore, a 2-year follow-up dual-energy X-ray absorptiometry (DEXA) study showed improved bone remodelling compared to the Omnifit hip stem [ 4 ]. The stem showed early stabilization in 2 independent RSA studies, and excellent clinical outcomes in a 5-year clinical and radiographic follow-up study [ 5 – 7 ]. In a Danish implant registry study that included 1,055 total hip arthroplasties (THAs) in a single centre, the estimated 6.5-year survival rate of the Symax hip stem with all-cause revision as the endpoint was 97.5% (CI 96.6%-98.3%) [ 8 ]. Most common reasons for revision surgery in that study were periprosthetic fractures (n = 11) and recurrent dislocations (n = 10) [ 8 ].
The current study illustrates the ‘phased introduction’ of the uncemented Symax hip stem [ 9 , 10 ]. The idea of a stepwise clinical introduction of a new orthopaedic implant is to ensure quality of orthopaedic implants, and thus patient safety [ 10 ]. The Dutch Arthroplasty Registry (LROI) is a nationwide population-based registry covering all hospitals in the Netherlands, which was initiated by the Dutch Orthopaedic Association in 2007 [ 11 ]. The LROI contains prospectively collected data on primary and revision arthroplasty. Patient characteristics are recorded at the time of the primary procedure. In 2017, the registry completed its first 10 years of data collection, prompting us to evaluate the cumulative revision rates and the reasons for revision of the uncemented Symax hip stem in this first decade of registering.
Primary aim of this study was to evaluate the cumulative revision rates and the reasons for revision of the uncemented Symax hip stem in total hip arthroplasties 2007–2017 in the Netherlands. Secondary aim was to determine the associations between patient characteristics and reasons for revision. We hypothesized that the Symax hip stem meets the benchmark criteria for best prostheses following the National Institute for Health and Care Excellence (NICE) and Orthopaedic Data Evaluation Panel (ODEP) guidelines. The NICE guidance states that the best protheses should demonstrate a ‘benchmark’ revision rate of 5% or less at 10 years, or, as a minimum, a 3-year revision rate consistent with this benchmark [ 12 ]. The ODEP-ratings consists of a number and a letter, and a star (optional) [ 13 ]. The number represents the number of years for which the product’s performance has been evidenced. The letter represents the strength of the evidence (data). Letter ‘A’ represents strong evidence, and letter ‘B’ represents acceptable evidence. A star will be added when a benchmark replacement rate is defined of less than 1 in 20 (5%) at 10 years.
Patients and methods
Registry
The Dutch nationwide LROI database contains 99% of all primary total hip arthroplasties and 98% of revision hip arthroplasties [ 14 ]. It contains information on patient characteristics such as age, gender, and general health (ASA score). Since 2014, body mass index (BMI), smoking behaviour, orthopaedic vitality (i.e. Charnley score), and postal code were added to the database. Furthermore, hospital of surgery, anonymized (encrypted) surgeon, type and date of surgery, indication for surgery, surgical approach, fixation, and prosthesis characteristics (as specified below) were also registered. Implant information was retrieved from the LROI implant library, is based on the article number and includes among others name and type of the prosthesis, material, and femoral head size [ 11 ]. Finally, data from the LROI were matched with the encrypted citizen service number of the national insurance database on healthcare (Vektis 2017) in order to obtain information on the vital status (or date of death) of registered patients [ 11 , 14 ].
Data collection
All THAs in the Netherlands with an uncemented Symax hip stem registered in the LROI in the period between 2007 and 2017 were included. It is possible for a patient to be registered twice, if both hips were operated (bilateral THA). A primary THA was defined as the first time a total hip prosthesis is implanted to replace a hip joint. Revision arthroplasty was defined as any change (replacement, removal, or addition) of one or several components of the joint prosthesis [ 14 ]. The encrypted personal citizen number allowed linkage of revision arthroplasty procedure to the primary procedure. Reasons for revision surgery were categorized as infection, cup/liner wear, periprosthetic fracture, dislocation, loosening femoral component, loosening acetabular component, periarticular ossifications, and other. It was possible to register more than one reason for revision.
Prosthesis
The Symax hip stem is an uncemented design forged from Ti6Al4V alloy (CE 545074). Primary mechanical stability is provided by anatomical metaphyseal geometry ( Fig 1 ). The hip stem features a size-dependent anteversion, neck length and offset, with a CCD angle of 128°. Secondary biological stability is accomplished by fast osseous integration due to the BONIT-HA coating on the metaphyseal part of the stem. The BONIT-HA is an electrochemically deposited, biomimetic hydroxyapatite (HA) coating on top of a commercially pure titanium plasma spray (TPS) layer. It is deposited by low-temperature precipitation, is thin (10–20μm), and has a 3D surface with high porosity (60%) and pore interconnectivity [ 2 , 4 ]. The coating is fully resorbable and is substituted by bone for about 99% [ 15 ]. The anodization surface treatment, DOTIZE, applied on the distal part of the stem, is an electrolytical conversion of the native oxide film on titanium surfaces into a thicker and denser titanium oxide. It shows anti-galling properties, reduces protein adsorption with 19% and bone apposition compared to an untreated titanium alloy [ 2 , 4 ].
10.1371/journal.pone.0248483.g001
Fig 1
Design features of Symax hip stem.
Illustrating the anatomically anteverted proximal geometry, with the BONIT-HA coating; and the straight distal part with the DOTIZE surface treatment and a posterior chamfer.
Statistical analysis
Descriptive statistics were used for the presentation of baseline characteristics of the cohort at the time of primary procedure. Results were reported as absolute value and percentage. Follow-up started on the date of the primary surgery, and ended at revision, death, or end of the study period (1 st January 2017), whichever came first. Kaplan-Meier survival tables with 95% confidence intervals (CI) were used to estimate the cumulative 1, 5, and 7-year revision percentages. In order to assign the revision rates according to the NICE and ODEP, and to be able to compare the revision rates to other prostheses, the cumulative revision percentages for both revision of any component (e.g. cup, insert, stem, and/or head), and for revision of the uncemented Symax stem in particular, were estimated. Cox proportional hazards regression analysis was performed to assess the association between patient and procedure characteristics, and the need for revision arthroplasty. We used univariable and multivariable Cox regression analysis to examine the association between potential predictors and the outcome. All potential predictors were entered into the multivariable model. Using backward elimination on the significance of hazard ratios, we deleted non-significant variables from the model. As the number of events was lower than suggested by generic rules of thumb for multivariable modelling (i.e.: at least 10 events-per-variable), the analyses were considered exploratory. Since the amount of clustering in the data due to bilaterality was considered to be very low (<10%) we chose not to use models designed for clustered data such as the frailty model [ 16 ]. P-values of 0.05 and lower will be considered to indicate statistical significance.
Potential conflicts of interests
No conflict of interest was declared and no personal funding was received. No research grant was received.
Results
Between January 1 st 2007 and January 1 st 2017, a total of 5,013 THAs were implanted in 4,593 patients (420 bilateral) in the Netherlands. The mean age at surgery was 67.4 years (range 14–97 years), 62% were female patients, and in 83% of the patients the primary diagnosis was osteoarthritis ( Table 1 ). The median follow-up was 5.2 years (range 0–10 years).
10.1371/journal.pone.0248483.t001
Table 1 Patient and procedure-related baseline characteristics of the study cohort.
Total
n = 5013
% of subgroup
Gender
Female
3072
62%
Male
1918
38%
Missing
23
Age groups (years)
< 49
319
6%
50–59
687
14%
60–69
1735
35%
70–79
1647
33%
> 80
619
12%
Missing
6
ASA classification
ASA I
1123
23%
ASA II
2903
60%
ASA III-IV
844
17%
Missing
143
Diagnosis
Osteoarthritis
4143
83%
Acute fracture
443
9%
Osteonecrosis
157
3%
Other
270
5%
Fixation
Uncemented
4756
95%
Reversed hybrid
257
5%
Previous surgery on affected hip
Yes
280
6%
No
4523
91%
Unknown
144
3%
Missing
66
Approach
Posterolateral
1985
40%
Direct lateral
2159
43%
Anterolateral
804
16%
Other
65
1%
Revision rates
The cumulative 1, 5, and 7-year overall revision rates (with 95% CI) for revision of any component were 1.5% (1.2%-1.8%), 3.2% (2.7%-3.7%), and 3.8% (3.1%-4.4%) respectively ( Fig 2 ). The cumulative 1, 5 and 7-year stem revision rates (with 95% CI) of the Symax hip stem were 0.9% (0.6%-1.1%), 1.5% (1.1%-1.9%), and 1.7% (1.3%-2.1%) respectively ( Fig 2 ).
10.1371/journal.pone.0248483.g002
Fig 2
Kaplan-Meier cumulative revision percentages for revision of stem and THA.
Reasons for revision of the stem
In total, 76 patients underwent a revision of the stem. Of this subgroup, the mean age at surgery was 67.4 years (range 36–90 years), 66% were female patients, and in 76% of the patients the primary diagnosis was osteoarthritis ( Table 2 ). Periprosthetic fractures (n = 35) and loosening of femoral component (n = 30) were the most common reasons for revision ( Table 3 ). Since multiple reasons for revision were allowed, 12 patients were registered for both periprosthetic fractures as for loosening of the femoral component.
10.1371/journal.pone.0248483.t002
Table 2 Patient and procedure-related baseline characteristics of the revision cohort.
Total
n = 76
% of subgroup
Gender
Female
50
67%
Male
25
33%
Missing
1
Age groups (years)
< 49
9
12%
50–59
9
12%
60–69
23
30%
70–79
22
29%
> 80
13
17%
Missing
6
ASA classification
ASA I
11
15%
ASA II
48
66%
ASA III-IV
14
19%
Missing
3
Diagnosis
Osteoarthritis
58
76%
Acute fracture
13
17%
Osteonecrosis
2
3%
Other
3
4%
Fixation
Uncemented
72
95%
Reversed hybrid
4
5%
Previous surgery on affected hip
Yes
10
13%
No
63
84%
Unknown
2
3%
Missing
1
Approach
Posterolateral
33
43%
Direct lateral
29
38%
Anterolateral
13
17%
Other
1
1%
10.1371/journal.pone.0248483.t003
Table 3 Reasons for revision of the Symax hip stem (n = 76).
n
% of revisions
% of THAs
Infection
9
12%
0.2%
Cup / liner wear
1
1%
0.0%
Periprosthetic fracture
35
46%
0.7%
Dislocation
7
9%
0.1%
Loosening femoral component
30
40%
0.6%
Loosening acetabular component
2
3%
0.0%
Periarticular ossification
2
3%
0.0%
Other
15
20%
0.3%
Values represent the numbers of revision of stems, percentages (%) of the total number of revisions of stems (n = 76), and percentages (%) of the total number of THAs. One patient may have more than one reason for revision. As such, the total proportion is over 100%. Note: there are 25 revision procedures with more than one reason for revision.
Risk factors for revision of the stem
An acute fracture as primary diagnosis or a previous operation of the affected hip were risk factors for revision of the stem ( Table 4 ) in both the unadjusted and multivariable model. The proportional hazard assumptions were not violated for the evaluated risk factors. Stratified analyses for revision of the stem according to primary diagnosis and previous surgery on the affected hip showed a 5-year revision rate of the stem of 3.1% (1.4%-4.9%) for THAs for acute fractures (n = 443), compared to 1.4% (1.0%-1.8%) for THAs for osteoarthritis ( Fig 3 ). THAs in patients with a previous surgery (n = 280) on the affected hip showed a revision rate of the stem of 3.5% (1.3%-5.4%) compared to 1.4% (1.0%-1.7%) for patients without a previous surgery in the affected hip ( Fig 4 ).
10.1371/journal.pone.0248483.g003
Fig 3
Kaplan-Meier cumulative revision percentages for revision of the stem according to primary diagnosis.
10.1371/journal.pone.0248483.g004
Fig 4
Kaplan-Meier cumulative revision percentages for revision of the stem according to previous surgery to the affected hip.
10.1371/journal.pone.0248483.t004
Table 4 Cox proportional Hazard Ratios (HR, with 95% Confidence Intervals (CI)) to assess the association between patient and procedure characteristics for revision of the Symax hip stem.
HR (95% CI)
Gender
* Male
Reference
* Female
1.2 (0.8–2.0)
Age (in years)
1.0 (1.0–1.0)
ASA classification
* I-II
Reference
* III-IV
1.2 (0.7–2.2)
Diagnosis
* Osteoarthritis
Reference
* Acute fracture
2.4 (1.3–4.3)
* Osteonecrosis
1.0 (0.2–4.1)
* Other
0.8 (0.3–2.7)
Fixation
* Uncemented
reference
* Reversed hybrid
1.0 (0.4–2.8)
Previous surgery
* No
reference
* Yes
2.7 (1.4–5.2)
Approach
* Posterolateral
reference
* Direct lateral
0.7 (0.4–1.2)
* Anterolateral
0.9 (0.5–1.6)
* Other
0.8 (0.1–6.1)
Discussion
In this population-based National Dutch Implant Registry study, we evaluated the cumulative revision rates and the reasons for revision of the Symax hip stem. The cumulative 1, 5, and 7-year revision rates of the Symax THA were comparable to the cumulative revision rates for uncemented THAs in the Netherlands in 2007–2016, as these were 1.6% (1.5%-1.6%), 3.4% (3.3%-3.6%), and 4.3% (4.1%-4.4%) respectively [ 14 ]. Our study population was comparable to the total group of patients with uncemented THAs in the Netherlands for age, gender distribution, and diagnosis. The cumulative revision rates are also in line with the median 6.5-year survival rates of the Symax hip stem with all-cause revision as the endpoint of 97.5% (96.6%-98.3%) in Vejle, Denmark [ 8 ]. The NICE recommends only those hip prostheses which have (projected) revision rates of 5% or less at 10-year clinical follow-up [ 12 ]. The ODEP was initiated in the UK in 2002 and independently evaluates hip prostheses according to the NICE guidelines [ 13 ]. The ODEP assigns each prosthesis design a benchmark rating, so that they can be compared to other prostheses that meet the NICE guidelines. In conclusion, the Symax hip stem meets the benchmark criteria of the NICE guidance, and can be classified as a 7A according to the ODEP criteria [ 12 , 13 ].
The most common reasons for revision of the Symax hip stem were loosening of femoral component and periprosthetic fractures. (Aseptic) loosening of the Symax hip stem occurred in 0.4% of the study population. On the contrary, no revisions caused by aseptic loosening of the Symax hip stem were reported in a Danish single-centre study [ 8 ]. Furthermore, early stabilization of the Symax hip stem was observed already after 4 weeks in a 2-year RSA study, and minimal subsidence of the Symax stems was measured in a 3-year EBRA-FCA study [ 7 , 17 ]. Another RSA study of the Symax hip stem also showed predominantly Y-translation and Y-rotation at 3 months, and zoledronic acid had no significant effect in this femoral stem migration [ 6 ]. In this randomized controlled trial zoledronic acid maintained periprosthetic bone mineral density (BMD) during the first 12 months, while in the control group the expected loss of BMD occurred. Thereafter, periprosthetic BMD decreased even in the zoledronic acid group in Gruen zone 7, but remained 14.6% higher than in the placebo group at 4-years follow-up [ 6 ]. A study comparing BMDs around the uncemented Symax and Omnifit hip stems showed values that were statistically significant in favour of the Symax hip stems [ 4 ]. All these studies indicate a minimal risk of aseptic loosening for the Symax hip stem.
The prevalence of revision due to periprosthetic fracture for the Symax stem at 10.2-years was 1.0%, which is slightly more than in our current study (0.7%) [ 8 ]. Thien et al. found in their 2-year Nordic Arthroplasty Register Association database that the incidence of revision due to periprosthetic fracture was 0.47% for cementless THAs (Bi-Metric, CLS Spotorno, Corail, ABG I and II) [ 18 ]. Nearly all fractures occurred during the first 6 months. The exact reason for the relatively high prevalence of periprosthetic fractures of the Symax hip stem compared to other uncemented stems is unclear. 20% of the periprosthetic fractures of the stem were in patients who had a primary diagnosis of an acute fracture, and 6% because of late post-traumatic osteoarthritis. 60% of the periprosthetic fractures of the stem occurred during the first 2 months, which is in line with the findings of Thien et al. [ 18 ]. These early fractures could be initiated during the primary surgery as minimal fissures and progress to significant fractures during rehabilitation [ 18 ]. This again confirms that results of uncemented stems, implanted for proximal femoral fractures are worse compared to those implanted for osteoarthritis and avascular necrosis. Bergschmidt et al. discontinued using the Symax hip stem because of subsidence of more than 10mm in 2 patients and 3 intraoperative periprosthetic fractures outside their study population [ 19 ]. In a clinical DEXA study comparing the Symax to the Omnifit hip stem, improved stress transfer from the bone to the implant in the important posterior and medial areas was proven for the Symax hip stem. This led to improved preservation of periprosthetic bone compared to other proximally, and entirely porous or HA-coated stems, which might lead to a decrease in periprosthetic fractures and aseptic loosening [ 3 ].
The secondary aim of this current study was to estimate the associations between patient characteristics and reasons for revision. As mentioned above, a primary diagnosis of acute fracture was associated with a statistically significant increased risk for revision of the stem. This is in line with the results of a Danish implant register study in which uncemented femoral components were associated with a statistically significant increased risk of early periprosthetic femoral fractures (relative risk (RR) 4.1, CI 2.3–7.2), especially in elderly (RR 1.4 per 10 years, CI 1.2–1.6), female (RR 0.6, CI 1.1–2.2) and osteoporotic patients (RR 2.8, CI 1.6–4.8) [ 20 ]. Furthermore, Thien et al. concluded in their study on 439,629 THAs in the Nordic registry that cementless stems should be avoided when advanced age, female gender and a femoral neck fracture are present [ 18 ]. Besides this, several other studies have proven that there are more complications with uncemented than cemented femoral stems in both THA and hemiarthroplasty for displaced femoral neck fractures [ 21 , 22 ]. However, there are several uncemented stem designs. Carli et al. identified a substantial variability in the performance of uncemented stem designs, with an increased risk for periprosthetic fractures in both type 1 (‘Single-wedge’ or ‘blade-type’) stems and type 2 (‘double-wedge’ or ‘fit-and-fill’) stems [ 23 ]. Statistically significant, and clinically relevant, lower rates of periprosthetic fractures were shown in type 6 (anatomical) stems, and a group consisting of type 3 (tapered round, spline, rectangle) and type 4 (cylindrical, fully coated) stems [ 23 ]. As the uncemented Symax hip stem can be classified as a type 2 stem, it is advised not be used for acute femoral fractures.
A previous operation to the affected hip was also a statistically significant, and clinically relevant, risk factor for revision of the femoral stem. This is in line with the results of a systematic review and meta-analysis in which THA after failed osteosynthesis (salvage or conversion THA) was associated with more complications compared to primary THA for intracapsular femoral neck fractures [ 24 ]. Although the optimal treatment for intracapsular femoral neck fractures remains debatable in independently mobile patients, fixation failure occurs in about 30% of these patients [ 25 ]. So, the outcome of conversion THA must be considered thoroughly. Other studies about conversion THA after prior proximal femoral trauma have also shown heterogeneous results depending on the initial fracture and fixation type. Conversion of prior intertrochanteric fracture fixation has been associated with poorer outcomes compared to prior femoral neck fractures [ 26 ]. Conversion from prior intramedullary fixation is more complex and challenging than from prior sliding hip screw, due to the increased damage to the medullary canal and the hip abductor mechanism [ 27 ]. Therefore, one has to take into account that for the Symax hip stem, as well as for other uncemented designs, the use is associated with a higher complication risk in case of previous surgery to the hip.
The strength of this study is that it is the first nationwide study about the Symax hip stem using data from the LROI Dutch Arthroplasty Register. The data is prospectively collected, the sample size is large, and the follow-up is complete. As it is a nationwide register study the data is generalizable. A limitation of this study is that not all implanted Symax hip prostheses are incorporated in the registry, because the registry started in 2007, while the Symax hip stem was introduced in 2004. The number of registered patient characteristics was limited in this registry. For example, alcohol and drug use, fluid and electrolyte disorders, and peripheral vascular disorders are not included, while these are known to influence revision rates [ 28 ]. Smoking behaviour, BMI, Charnley score, and postal code were added since 2014 to the LROI, so these were limited available for the current population. Furthermore, not all of the complications that have occurred are registered, since only revision procedures were included in the LROI. This probably underestimates the total number of complications. Nevertheless, this is the largest cohort of patients available in the Netherlands, providing the best nationwide evidence on the survival and revision rate of the Symax hip stem.
Conclusions
In summary, overall cumulative revision rates of the Symax hip stem are comparable to the overall cumulative revision rates for best performing uncemented THAs in the Netherlands at every follow-up up to 9 years follow-up. The Symax hip stem meets the benchmark criteria of the NICE guidance, and it can be classified as a 7A according to the ODEP criteria. Periprosthetic fractures and loosening of the femoral component were the most commonly registered reasons for revision of the femoral component. An acute fracture as the primary diagnosis and a history of previous surgery on the affected hip were statistically significant, and clinically relevant, associated risk factors for revision of the Symax hip stem, and we discourage the use of the Symax hip stem in these patients.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Proximal humerus fractures account for approximately 5% of all fractures [1] – [3] . Surgical intervention is generally accepted for unstable fractures, including displaced fractures and fractures associated with osteoporosis. Recent research noted a high failure rate(8.6%–22.0%) after open reduction and internal fixation (ORIF) of proximal humerus fractures [4] – [7] . Fixation without the reconstruction of medial support is considered one of the risk factors of implant failure [5] . Gardner et al. demonstrated the direct association between medial support and subsequent loss of reduction [8] . Zhang et al. found that insertion of a medial support screw (MSS) into the medio-inferior region increased the stability of fixation of complex fractures and reduced the risk of implant failure [9] . When medial comminution and malreduction are present at the proximal humerus, the insertion of a MSS precisely to the medio-inferior region of the humeral head is one method to reconstruct the medial column support.
Investigation into the value of MSSs in the treatment of proximal humerus fractures has been limited to clinical outcome studies. No biomechanical studies have been reported. The purpose of this study was to evaluate the biomechanical advantages of MSSs in the locking plate for the treatment of proximal humerus fractures and to use this information to guide clinical practice.
Methods
A total of 30 adult synthetic left humeri (HI-C type, Orthobone, Hangzhou, China) and 30 sets of proximal humerus locking plates and screws (Double medical, Xiamen, China) ( Fig. 1 ) were prepared. All the mechanical tests were performed on an Axial-Torsional Biomechanical Testing System (301.6 Shore Western Manufacturing, California,USA) with the following load cell characteristics: a maximum force rating of 3,200 pounds (14.2 kN) and a stroke of four inches (100 mm). The torsional rating on this unit is 1,000 inch-pounds (120 Nm) with a rotary motion capability of ±140°.
10.1371/journal.pone.0103297.g001 Figure 1
The proximal screw distribution and the medial support screws (MSS) for the locking proximal humerus plate.
The 30 synthetic humeri were randomly divided into three groups ( ie , Groups A, B, and C with 10 specimens/group). A two-part surgical neck fracture was created in each proximal humerus, and all fractures were fixed with a locking proximal humerus plate.
Fracture groups
Group A ( Fig. 2a and Fig. 2d ). Ten proximal humerus fractures were fixed with medial cortical support; however, MSSs were not used. A horizontal line (line A) was made 1 cm distal to the humeral head, and transverse osteotomies were created along this line using an industrial bandsaw to simulate a two-part surgical neck fracture of the proximal humerus. Fractures were anatomically reduced and fixed with a locking proximal humerus plate. Six locking screws were inserted into the proximal holes of the plate from No. 1 to No. 6. ( Fig 1 ).
10.1371/journal.pone.0103297.g002 Figure 2
Division of the proximal humerus fracture models.
In group A, proximal humerus fractures were fixed without MSSs (Fig.2a); transverse osteotomies were created along a horizontal line (line A) (Fig. 2d). In group B, proximal humerus fractures were fixed with 3MSSs(Fig.2b); wedge osteotomies were created along a horizontal line(line A) and an oblique line (line B) to simulate medial comminution of the proximal humerus(Fig. 2e). In group C, proximal humerus fractures were fixed without medial cortical support or MSSs(Fig.2c); wedge osteotomies were created identical to group B (Fig. 2f).
Group B ( Fig. 2b and Fig. 2e ). Ten proximal humerus fractures were fixed with 3 MSSs; however, no medial cortical support was provided. A horizontal line (line A) was made 1 cm distal to the humeral head. The intersection of line A and the lateral cortex of the proximal humeral metaphysis was identified, and a line (line B) was drawn from this point to the most medio-inferior point of the humeral head. Wedge osteotomies were created along lines A and B to simulate medial comminution of the proximal humerus. Three locking screws were inserted through holes No. 7–9( Fig 1 ) into the medio-inferior region of the humeral head at the proximal part of the plate. Another three screws were randomly inserted into the other six holes (No. 1–6) of the proximal part of the plate.
Group C ( Fig. 2c and Fig. 2f ). Ten proximal humerus fractures were fixed without medial cortical support or MSSs. The fracture model was constructed identical to Group B. However, six locking screws were inserted into holes No. 1 to 6 of the proximal part of the plate.
The proximal screws in all three groups were inserted 5 to 8 mm below the subchondral bone to simulate clinical practice [10] . Humeral shafts were fixed with one cortical screw and three locking screws. Plate-bone gap was not present in all experimental models. The distal part of the humeri were removed 15 cm distal to line A. Specimens were 20 cm long, and the distal portion of each specimen was fixed by a square steel chamber filled with a commercially-available anchoring cement to a depth of 12 cm.
Biomechanical testing
Axial stiffness
Each humerus was oriented vertically in the coronal and sagittal planes. Using a plate attached to the mechanical tester, axial compression was applied with a vertical load at the apex of the humeral head using displacement control (max deflection = 0.5 mm; load rate = 5 mm/min [11] ; preload = 50 N). The maximum load was recorded, and the slope of the load-deflection curve was used to compute the axial baseline stiffness. Each test was repeated three times, and the average maximum load and average stiffness were calculated ( Fig. 3a ). Every specimen was kept within the linear elastic region to prevent any permanent damage (average linearity coefficient R 2 >0.99). No visual evidence of damage was noted.
10.1371/journal.pone.0103297.g003 Figure 3
Mechanical test modes used to assess the (A) axial stiffness; (B) torsional stiffness; and (C) shear stiffness and load-to-failure of the plated humeral constructs.
Torsional stiffness
Each humerus was positioned inside the cup of a cylindrical stainless steel block and secured using three 4.0 mm screws that were inserted into the humeral head. A torque was applied using displacement control (maximum angulation = 5°; rate = 12°/min; pre-torque = 0 Nm) to simulate rotation of the humeral head. The maximum torque was recorded, and the slope of the torque-angulation graph was used to determine torsional stiffness. The application of torque was repeated three times, and an average maximum torque and average torsional stiffness were calculated ( Fig. 3b ). All specimens were kept within the linear elastic region to avoid permanent specimen damage (average linearity coefficient R 2 >0.99). No visual evidence of damage was noted.
Shear stiffness
Each humerus was mounted distally in a vice with the shaft axis in 20° of abduction( Fig.3c ) as recommended by Koval et al. in order to simulate the shear loading across a proximal fracture site experienced when rising out of a chair or crutch weight bearing [11] .The same test procedure as that with axial stiffness testing was performed, except that maximum displacement was 1.0 mm. Load levels were chosen to prevent permanent damage to the specimens. All specimens remained within the linear elastic region to avoid permanent damage (average linearity coefficient R 2 >0.99). No visual damage was noted.
Strength testing
Load-to–failure in shear for each of the three groups was determined using displacement control (load rate = 5 mm/min; preload = 50 N) to apply a vertical force to generate shear compression on each specimen oriented 20° in abduction ( Fig.3c ). The force was applied until catastrophic failure of the implant occurred. The highest peak load was determined to indicate significant structural collapse [11] . Catastrophic fracture patterns of the bone and implant were examined and recorded.
Statistical analysis
For the statistical analysis, the SAS 11.0 (SAS Institute Inc, North Carolina, USA) was used. Statistical analyses were performed by an independent statistician blinded to surgical outcomes. The Student's t-test was employed with the one-way analysis of variance (ANOVA) test used for continuous variables. The Turkey post hoc test was used to differentiate groups for statistical differences. The level of statistical significance was set to P <0.05. The chi-square test or the Fisher's exact test was employed for binary variables with the level of statistical significance set to P <0.05.
Results
There was no evidence of plate or screw loosening or breakage. No fractures occurred during axial, torsional and shear stiffness tests. The results of four different load tests for the three groups are shown in Table 1 .
10.1371/journal.pone.0103297.t001 Table 1
The results under four different load tests for the three groups ( ±sd, n = 10).
Group
Torsional stiffness test (max angulation = 5°)
Axial stiffness test (max displacement = 0.5 mm)
Shear stiffness test (max displacement = 1 mm)
Shear failure test
Maxtorque (NM)
Torsional stiffness (NM/deg)
Max load (N)
Axial stiffness (N/mm)
Max load (N)
Shear stiffness (N/mm)
Shear failure load (N)
Group A
8.92±0.25
1.80±0.07
240.9±19.1
424.4±101.2
444.7±20.9
470.0±54.4
2949.8±355.1
Group B
9.09±0.31
1.86±0.07
169.0±19.3 a
230.7±40.54 a
228.8±29.0 a
183.9±29.6 a
2448.1±402.4 a
Group C
7.57±0.53 a , b
1.53±0.10 a , b
128.6±17.5 a , b
147.0±29.2 a , b
188.7±26.2 a , b
140.2±32.1 a
2222.6±336.4 a
F value
47.06
45.14
92.94
47.67
290.53
198.05
10.36
P value
<.0001
<.0001
<.0001
<.0001
<.0001
<.0001
0.0005
Comparison
A vs B: 0.6086
A vs B: 0.2738
A vs B: <0.001
A vs B: <0.001
A vs B: <0.001
A vs B: <0.001
A vs B: 0.0131
of three
A vs C: <0.001
A vs C: <0.001
A vs C: <0.001
A vs C: <0.001
A vs C: <0.001
A vs C: <0.001
A vs C: 0.0004
group
B vs C: <0.001
B vs C: <0.001
B vs C: <0.001
B vs C: 0.0207
B vs C: 0.0044
B vs C: 0.0561
B vs C: 0.3655
a compared with group A, P<0.05; b compared with group B, P <0.05.
Axial stiffness
The maximum load (max displacement = 0.5 mm) and the axial stiffness showed statistical differences between the groups (Group A> Group B> Group C; P≤0.0207) ( Fig. 4a , Fig 4b ).
10.1371/journal.pone.0103297.g004 Figure 4
Comparison of stiffness tests among three groups under four load steps (a,b) axial stiffness test, (c, d) torsional stiffness test, (e, f) shear stiffness test, (g) failure test.
Fig(Group A> Group B> Group C; P ≤0.0207) Fig 4 c Torsional stiffness data for all subgroups. The maximum load of Group C was statistically different from both Group A and Group B, (Group A> Group C, Group B> Group C; *,#, P <0.0001). The comparisons of Groups A and B for maximum load were not significantly different ( P = 0.6086). Fig 4 d Torsional stiffness data for all subgroups. The torsional stiffness of Group C was statistically different from both Group A and Group B (Group A> Group C, Group B> Group C; *,#, P <0.0001). The comparisons of Groups A and B for torsional stiffness were not significantly different ( P = 0.2738). Fig 4 e Shear stiffness data for all subgroups. The maximum load showed statistical differences between the groups (Group A> Group B> Group C; P ≤0.0044). Fig4 f Shear stiffness data for all subgroups. The shear stiffness of Group A was statistically different from both Group B and Group C (Group A> Group B, Group A> Group C; *,#, P <0.0001). The comparisons of Groups B and C for shear stiffness were not significantly different ( P = 0.0561). Fig 4 g The load-to-failure of Group A was statistically different from both Group B and Group C (Group A> Group B, Group A> Group C; *,#, P ≤0.0131); however, no statistical differences were noted between Group B and Group C ( P = 0.3655).
The maximum loads were 240.88 N±19.13 in Group A, 169.04 N±19.26 in Group B, and 128.58 N±17.53 in Group C. The axial stiffness was 424.4 N/mm ±101.2 for Group A, 230.7 N/mm ±40.54 for Group B, and 147.0 N/mm ±29.2 for Group C.
Torsional stiffness
The maximum load of Group C (7.57 Nm ±0.53) was statistically different from both Group A (8.92 Nm ±0.25) and Group B (9.09 Nm ±0.31) (Group A> Group C, Group B> Group C; P <0.0001). The torsional stiffness of Group C (1.53 Nm/deg ±0.10) was statistically different from both Group A (1.80 Nm/deg ±0.07) and Group B (1.86 Nm/deg ±0.07) (Group A> Group C, Group B> Group C; P <0.0001); however, the comparisons of Groups A and B for maximum load ( P = 0.6086) and torsional stiffness ( P = 0.2738) were not significantly different ( Fig 4c , Fig 4d ).
Shear stiffness
The maximum load between subgroups showed statistical differences (Group A> Group B> Group C; P ≤0.0044)( Fig 4e ). The values were 444.7 N ±20.9 for Group A, 228.8 N±29.0 for Group B, and 188.7 N±26.2 for Group C.
The shear stiffness of Group A(470.0 N/mm ±54.4) was statistically different from both Group B(183.9 N/mm ±29.6) and Group C(140.2 N/mm ±32.1)(Group A>Group B, Group A>Group C; P<0.001); however, the comparisons of Group B and Group C for shear stiffness were not significantly different(P = 0.056). ( Fig 4f ).
Shear failure
The load-to-failure of Group A (2949.76 N±355.08) was statistically different from both Group B (2448.13 N±402.39) and Group C (2222.55 N±336.41) (Group A> Group B, Group A > Group C; P ≤0.0131); however, no statistical differences were noted between Group B and Group C ( P = 0.3655). ( Fig. 4g ).
Shear failure mode
In group A, seven specimens failed by humeral shaft fracture ( Fig. 5A ), and three specimens failed by humeral head fracture. No plate bending occurred in Group A. In group B, five specimens failed by plate bending with humeral head collapse, and five specimens failed by humeral head fracture ( Fig. 5B ). In group C, six specimens failed by significant plate bending with humeral head collapse ( Fig. 5C ), and four specimens failed by humeral head fracture.
10.1371/journal.pone.0103297.g005 Figure 5
Shear Failure Mode: humeral shaft fracture (A), humeral head fracture (B) and specimens failed by significant plate bending (C).
Discussion
Biomechanical tests are currently being used in studies of fracture fixation [11] – [14] . In the current study, biomechanical tests showed that medial support from cortical bone in the proximal humerus provided the best stability when locking plates were used to treat proximal humerus fractures. When the proximal humerus was not supported by medial cortical bone, locking plating with MSSs exhibited higher biomechanical performance than locking plating without MSSs. There were differences among the three groups in the mode of failure. When the proximal humerus was supported by medial cortical bone, specimens showed no evidence of plate bending and failed primarily by fracture of the humeral shaft or humeral head. When the proximal humerus was not supported by medial cortical bone, specimens failed primarily by significant plate bending at the fracture site during axial loading. The humeral head was then crushed by the medial cortical bone of the humeral shaft, which caused a humeral head fracture.
A recent biomechanical study by Lescheid et al. also demonstrated that locking plating with medial cortical support was more resistant to axial compression and shear force compared with locking plating without medial cortical support [12] . However, there was no significant difference between specimens with medial cortical contact and other subgroups, which was different to our results. This difference may be resulted from the small number of specimens included in their research, only 6–7 specimens were included in each subgroup. In addition, the supporting role of MSS was not studied in their research. The effect of the MSS was limited to the clinical literature with no evidence of biomechanical research. Our study was conducted to verify the biomechanical benefits of the MSSs in proximal humerus locking plates and to use this information to guide clinical practice.
Based on our results, strong medial support should be reconstructed in every case. When treating two-part proximal humerus fractures, anatomical reduction of the medial cortical bone should be obtained. When the medial cortical bone is comminuted or cannot be anatomiacally reduced, 2 or 3 MSSs can be inserted in order to help the reconstruction of the medial support. If medial support is not reconstructed, the resistance to axial compression and torsional force decreases, and early failure of the fracture fixation can occur.
Compared with traditional fixation methods, the fixed-angle devices can provide a greater ability to resist angular and rotational forces, especially in osteoporotic patients [15] , [16] . Friess et al. also found that locking plates showed higher performance on both the functional range of motion and American Shoulder and Elbow Surgeons (ASES) scores after a mean 45-month follow-up [16] . However, locking plating still has early failures, especially in patients with comminuted and osteoportic fractures or fixation without reconstruction of medial support [5] , [6] . Common complications include varus deformity (16%), humeral head necrosis (10%), and screw cut out (8%) [17] . Fixation without reconstruction of medial support is one of the risk factors for implant failure [5] . Reconstruction of the medial support increases the stability of fixation by providing effective support to the humeral head [8] , sharing the varus deforming force, and decreasing the cutting force between the screws and the bone. Therefore, it is advisable to reconstruct the medial support of proximal humerus during the operation.
Anatomical reduction of the medial cortical bone of the proximal humerus is one of the methods to reconstruct the medial column support of the proximal humerus [8] . The current study confirmed that the medial cortical bone support of the proximal humerus has the best biomechanical stability. When the medial cortical bone of the proximal humerus is anatomically reduced, the cortical bone contacts, and the supporting forces are created to increase axial, torsional, and shear stiffness. Therefore, for a simple fracture of the proximal humerus, attempts should be made to achieve anatomical reduction of the medial cortical bone to gain medial column support and help avoid implant failure. When the medial metaphysis of the proximal humerus is comminuted, fractured with a bony defect, or the fracture is malreduced, one or two additional locking screws can be inserted obliquely into the medio-inferior region of the humeral head to reconstruct medial column support of the proximal humerus [10] . A cadaveric biomechanical study by Liew et al. found that the grasping force of a screw placed under the subchondral bone of the medial and inferior region was comparably stronger than that of a screw placed either in the middle of the humeral head or in the lateral and superior region [18] . Another histomorphometric study by Hepp et al. showed the highest bone strength to be in the medial and dorsal aspects of the proximal humeral head [19] . As a result, the optimal fixation of a screw is in the posterior-medial-inferior aspect of the humeral head to prevent screw cut out and implant loosening. Furthermore, MSSs in hole No. 7∼9 were comparably closer to the fracture line than that of a screw placed in hole No. 1∼6, and MSSs were relatively upwards inserted. These screws would support the proximal fragment directly from the medio-inferior part of the humeral head and might be conducive to the resistance to varus deforming force. Zhang et al. and Hardeman et al. found MSSs increase the stability of proximal humerus fractures [9] , [20] . However, their studies were limited to clinical outcomes with no evidence of biomechanical study. Through biomechanical tests, our study proved when the medial cortical bone of the proximal humerus is fractured with comminution, MSSs could help to increase the stability of fixation and should be strongly considered.
Several methods exist for the reconstruction of medial column support in comminuted proximal humerus fractures with bony defects. Egol et al. performed fracture site augmentation with calcium phosphate cement to increase stability after ORIF of the proximal humerus fractures [21] . Micic et al. achieved anatomical reduction with autogenous bone graft into the area of medial comminution [22] , while Hettrich et al. augmented the fixation with a fibular allograft inserted medially [23] . However, there is still no gold standard treatment for reconstruction of medial column support in comminuted proximal humerus fractures. And to some extent, the bone graft methods mentioned above need more stripping of soft tissue at the medial side of the proximal humerus, which is important for the blood supply of the humeral head. In our opinion, inserting two to three MSSs will help the reconstruction of the medial support if the supporting screws can be inserted precisely into the medio-inferior region of the humeral head. To achieve this, we need to adjust the locking plate to the appropriate height and confirm the position of the tips of the MSSs by intraoperative fluoroscopy. This procedure does not require stripping a wide range of soft tissue or excessive manipulation of the fracture fragments. For comminuted and osteoporotic fractures, tension band suture is routinely used as supplemental fixation to improve the stability of the humeral head [24] , [25] .
There are a number of limitations in this study. First, the effect of the surrounding soft tissues on the mechanical stability of the construct was not evaluated. Second, synthetic bone may more closely simulate normal, rather than osteoporotic, bone. The biomechanical characteristics were likely different from the complicated proximal humerus fractures seen in osteoporotic patients, especially the failure mode. More specimens might have failed by humeral head fracture or humeral head collapse if osteoporotic specimen were used. However, under the same objective conditions, the same experimental procedure was used to assess the mechanical stiffness and strength of each proximal humerus fracture fixation for three subgroups (medial cortical support, MSSs, and no medial column support). Third,the effect of cyclic loading was not investigated, which may be more predictive of the long-term performance of the construct than static load. Fourth, in order to make the study closer to clinical practice, we did not specify which three holes to be used for screw insertion in No. 1∼6 in group B. Clinically, the number of holes selected depends on the specific condition in the operation such as fracture configuration. While screw position of the aforementioned three holes might have impacted on our results, we are unable to quantify its contribution. Future studies using finite element method will be undertaken to further investigate this.
Conclusions
Proximal humerus fracture fixation with medial cortical contact demonstrated the best biomechanical characteristics. Every effort should be made to achieve anatomical reduction of the medial cortical bone of the proximal humerus. Constructs with three MSSs showed statistically higher axial, torsional, and shear stiffness than constructs without medial support. Therefore, it is recommended that if the medial comminution is present with bony defect or the medial cortical bone is malreduced, three MSSs should be inserted to reconstruct medial column support of the proximal humerus to prevent postoperative implant failure.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Molybdenum and tungsten enzymes are involved in key reactions in the global cycles of carbon, nitrogen and sulphur [1] . Most members of the family of molybdenum and tungsten enzymes catalyse oxygen transfer reactions [1] – [3] . While tungsten-containing enzymes are restricted to bacteria and archaea with most of them being found in thermophilic and hyperthermophilic archaea, molybdenum-containing enzymes are found in all kingdoms of life [4] . Within their active site, molybdenum or tungsten are chelated by a metal-binding pterin (MPT, also known as molybdopterin) forming the so-called molybdenum (Moco) or tungsten cofactors (Wco), respectively [1] . Wco and Moco biosyntheses represent multistep pathways that start from GTP and involve three isolatable intermediates: cyclic pyranopterin monophosphate (cPMP) [5] , MPT and adenylylated MPT (MPT-AMP) [6] . MPT-adenylyl-transferases (EC 2.7.7.75) convert MPT to MPT-AMP ( Figure 1 ), prior to the final metal insertion step, which requires either tungstate or molybdate to produce Wco and Moco, respectively [7] , [8] .
10.1371/journal.pone.0086030.g001 Figure 1
Adenylylation of the cofactor intermediate MPT catalyzed by hexameric MoaB ( P. furiosus ) and trimeric MogA ( E. coli ).
Biochemical and functional characterization of MPT-adenylyl-transferases from different organisms revealed that they all fold in a similar tertiary structure while presenting two different oligomeric states: trimers (MogA sub-family) and hexamers (MoaB sub-family). Structural analysis of MoaB proteins demonstrated that MoaB hexamers are formed by dimerization of trimers [9] , [10] . Catalytically active, trimeric MPT-adenylyl-transferases were identified in mesophilic organisms, e.g. E. coli MogA protein, G domain of the plant Cnx1 protein (Cnx1G) and G domain of the human gephyrin protein (GephG) [7] , [11] , [12] . In contrast, the only characterized hexameric MPT-adenylyl-transferase, MoaB, was purified from the hyperthermophilic archaeon P. furiosus and showed maximal activity at 80°C [8] , thus being close to the optimal growth temperature of P. furiosus of 100°C [13] . Therefore, enzymes of the MPT-adenylyl-transferase family from hyperthermophilic organisms represent a suitable target for the elucidation of the impact of oligomerization on thermal stability.
In this study we investigated the influence of hexamerization on the structural stability and enzymatic activity of P. furiosus MPT-adenylyl-transferase MoaB (PfuMoaB). We determined the crystal structure of PfuMoaB at 2.5 Å resolution and engineered a hexamerization-deficient variant of PfuMoaB using structure-guided mutagenesis. The resulting variant was significantly impaired in thermal stability and MPT-adenylylase activity at high temperatures. However, we found a gain of function at low temperatures suggesting that hexamer formation in PfuMoaB is a key factor contributing to thermal stability in PfuMoaB by retaining its structural integrity at high temperatures.
Materials and Methods
Protein Expression and Purification
PfumoaB-WT and EcomogA were cloned into Bam HI and Hind III restriction sites of pQE80L (Qiaqen) using pET15b_ PfumoaB-WT and pET22b_ EcomogA [8] as templates for PCR, resulting in the fusion of an N-terminal His-tag. PfumoaB-H3 was in vitro synthesized by GenScript and cloned into pQE80L in the same way as PfumoaB-WT . E. coli MPT synthase subunits MoaD and MoaE [14] , E. coli MoaB, E. coli MogA [8] and A. thaliana Cnx1G [15] were expressed and purified as previously described. For crystallization, PfuMoaB-WT was purified as described [8] . For all other procedures PfuMoaB-WT and PfuMoaB-H3 were expressed in E. coli BL21(DE3) for 16 h at 18°C. Expression was induced with 250 µM IPTG at a cell density of OD 600 of 0.5. Proteins were purified using ion metal affinity chromatography on Ni 2+ -nitrilotriacetic (Ni-NTA) matrix (Ni-NTA Superflow Cartridge, Qiagen) attached to Äkta Purifier (GE Healthcare) following the manufacturer’s instructions. Purified proteins were exchanged into buffer containing 0.1 M Tris pH 8.0 and 200 mM NaCl using PD10 columns (GE Healthcare), flash frozen in liquid nitrogen and stored at –80°C until further use.
Protein Crystallization
For crystallization, purified PfuMoaB-WT was exchanged into a buffer containing 20 mM Tris pH 8.0, 100 mM NaCl using PD10 columns (GE Healthcare). Crystallization conditions were tested using screening solutions purchased from Qiagen (The JCSG Core I, II, III and IV Suite). Routinely, 300 nL of 5 mg/ml PfuMoaB protein solution were mixed with 300 nL of screening solution applying a vapour diffusion technique in 96-well Intelli plates (sitting drop). Screens were set up using Hydra II-eDrop robot and ControlMate software (Matrix Technology). Plates were sealed with a clear seal film, centrifuged for 1 min at 1000×g and incubated at 20°C. After one week, protein crystals were observed in the condition containing 0.1 MES pH 6.0, 40% PEG 400 and 5% PEG 3000 (JCSG Core screen III, condition F1). Crystals were extracted directly from the screen condition, mounted into nylon loops (Hampton Research) and flash frozen in liquid nitrogen without addition of cryo-protectant.
Data Collection and Structure Determination
Diffraction data were collected at the BESSY II beamline BL 14.1 (Berlin, Germany) at a wavelength of 0.918 Å, with 0.5° oscillation range and a crystal to detector distance of 225 mm. Diffraction data were processed with MOSFLM [16] and SCALA of the CCP4 program suite [17] . Data collection statistics are summarized in Table S1 .
The crystal structure of PfuMoaB was determined by molecular replacement with PHASER [18] using the coordinates of Bacillus cereus MoaB (1Y5E) as a search model. Refinement was performed with REFMAC5 [19] with 5% randomly chosen reflections that were set aside to calculate R -free. The model was manually refined using COOT [20] . Water molecules were added with COOT, ligand molecules were included manually. The structure was refined to 2.5 Å resolution with final R- work and R- free of 0.18 and 0.24, respectively and was validated using COOT validation tools. Structural coordinates were deposited in the RCSB protein data bank (4LHB). Refinement statistics are shown in Table S1 .
Circular Dichroism (CD) Spectroscopy
CD spectra were recorded from 195 to 260 nm in a 0.1 cm light path quartz cuvette at 20°C with a scanning speed of 10 nm/min using a Jasco J-715 CD spectropolarimeter (Jasco). The concentration of the protein samples was adjusted to 0.2 mg/ml in 10 mM sodium phosphate buffer, pH 8.0. Each spectrum was recorded five times and averaged. Averaged CD spectra were corrected by buffer baseline by subtraction. Measured elllipticity θ in milli-degrees was converted into mean residues ellipticity [θ] in deg * cm 2 * dmol –1 using following formula: [θ] = θ * M r /(10 * (n−1)*c * L), where M r is the molecular weight of a protein in Da, n is the number of residues per monomer of protein, c is the sample concentration in mg/ml and L is the path length in cm.
Size Exclusion Chromatography (SEC)
For size exclusion chromatography (SEC) a preparative Superdex 200 16/60 prep grade and an analytical Superdex 200 10/300 column were used (GE Healthcare). Routinely, a SEC buffer containing 0.1 M Tris pH 8.0 and 0.2 M NaCl, was applied. Protein separation was conducted with a flow rate of 0.5 ml/min at 4°C. Absorbance was monitored at 280 nm. Protein-containing fractions derived from the preparative SEC column were collected, concentrated by ultrafiltration (Amicon, Millipore), flash-frozen in liquid nitrogen and stored at −80°C until further use. Molecular masses of proteins were determined using standard proteins purchased from GE Healthcare (Gel Filtration Calibration Kit, HMW).
Determination of Molecular Mass of PfuMoaB-WT and PfuMoaB-H3 by Mass Spectrometry
Protein samples (100 µl) with a concentration of 2 mg/ml were diluted to a final volume of 450 µl with 20% acetonitrile in water containing 1% acetic acid. Samples were concentrated 10-fold by centrifugation using 3 kDa Amicon ultrafiltration units (Millipore). This buffer exchange was repeated three times. Protein concentration was determined by infrared spectrometry (Direct Detect, Millipore) and adjusted to 2 mg/ml. Electrospray ionization mass spectrometry was performed with a LTQ Orbitrap mass spectrometer (LTQ Orbitrap Discovery, Thermo Scientific). The instrument’s built-in syringe pump delivered the sample at a flow rate of 1 µl/min. In the Orbitrap 100 full scan MS spectra were acquired in the m/z range 150–2,000 at a resolution of 30,000 and averaged. Data acquisition was carried out for 1 min. For charge state deconvolution spectra were exported to the MagTran software (version 1.02, [21] ). Protonated (MH+) molecular masses were calculated in the mass range 10,000–30,000 Da.
Cross-linking
Bis-sulfosuccinimidyl suberate (BS 3 ) and 1-Ethyl-3-[3-dimethylaminopropyl]-carbodiimide hydrochloride (EDC) were purchased from Thermo Scientific. 0.5 mg/ml protein solution was incubated with 1.5 mM BS 3 in a buffer containing 20 mM sodium phosphate pH 8.0 and 0.2 M NaCl for 30 min or with 2 mM EDC in a buffer containing 0.1 M MES pH 6.0 and 0.5 M NaCl for 3 h. Cross-linking reactions were quenched with 50 mM Tris pH 7.5 for 15 min and further subjected to SDS-PAGE. Bands with molecular masses corresponding to PfuMoaB-WT and PfuMoaB-H3 trimers were cut from the SDS polyacrylamide gels, digested with trypsin and analysed by LC-MS/MS in the proteomics facility of the Cologne Cluster of Excellence in Cellular Stress Responses in Aging-associated Diseases (CECAD).
Tryptic in-gel Digestion and Nano-LC ESI-MS/MS
SDS-PAGE bands of interest were subjected to tryptic in-gel digestion according to the reference [22] with minor modifications. Prior to nano-LC-MS/MS analysis the peptides were desalted using STAGE Tip C18 spin columns (Proxeon/Thermo Scientific) as described elsewhere [23] . Eluted peptides were concentrated in vacuo and then re-suspended in 0.5% acetic acid in water. Analyses using reversed phase liquid chromatography coupled to nano-flow electrospray tandem mass spectrometry were carried out using an EASY nLC II nano-LC system (Proxeon/Thermo Scientific) with a 150 mm C18 column (internal diameter 75 µm, Dr. Maisch GmbH) coupled to a LTQ/Orbitrap mass spectrometer (LTQ Orbitrap Discovery, Thermo Scientific). Peptide separation was performed at a flow rate of 250 nl/min. over 79 minutes (5 to 10% acetonitrile in 2 min., 10 to 40% in 60 min., 40 to 100% in 2 min., wash at 100%; buffer A: 0.1% formic acid in H 2 O; buffer B: 0.1% formic acid in acetonitrile). Survey full scan MS spectra (m/z 350 to 2000) of intact peptides were acquired in the Orbitrap at a resolution of 30000 using m/z 445.12003 as a lock mass. The mass spectrometer acquired spectra in „data dependent mode“ and automatically switched between MS and MS/MS acquisition. Signals with unknown charge state and +1 were excluded from fragmentation. The ten most intense peaks were isolated and automatically fragmented in the linear ion trap using collision-induced dissociation (CID). Databases containing cross-linked peptides were generated for PfuMoaB-WT and the H3 variant using the publicly available xComb software (version 1.3) with the following settings: enzyme specificity trypsin, two missed cleavages, intra- and inter-molecular cross-links, type of cross-linker used EDC, minimum peptide length 4 amino acid residues. Sequest as implemented in the Proteome Discoverer 1.3 software (Thermo Scientific) was used for identification by searching the databases generated by the xComb software. Oxidation at methionine residues was used as a variable modification and carbamidomethylation at cysteine residues as fixed modification. No modification was used for the cross-linker EDC. The xComb software computes every possible cross-link combination by linearizing each pair of tryptic peptides. Because of this linearization process, each pair of peptides is assembled in two permutations, i.e. peptide A followed by peptide B and vice versa . The resulting xComb database was created from digested peptides, and not from proteins. Therefore, an enzyme mode was generated that cleaves after a non-existing residue (specified as “J” which is not cleaved), to simulate a “do not cleave” mode. Sequest then searches the xComb database without performing any enzymatic digestion and takes each entry in its whole. Mass tolerance for intact peptide masses was 10 ppm for Orbitrap data and 0.8 Da for fragment ions detected the linear trap. Search results were filtered to contain only high confident peptides (false discovery rate ≤1%) with a mass accuracy of ≤5 ppm, and a peptide length of ≥6 amino acid residues. Only peptides for which at least two fragment spectra (peptide spectral match, PSM) were detected were further analyzed. The minimum protein score was set to ≥4.0.
Differential Scanning Calorimetry (DSC)
DSC was performed using a VP-DSC MicroCalorimeter (MicroCal, GE Healthcare). Protein samples were diluted to the concentration of 50 µM in a buffer containing 50 mM sodium phosphate pH 8.0 and 100 mM NaCl, degased and heated from 20°C to 130°C with a scan rate of 90°C per hour. Melting temperatures ( T m ) were determined using ORIGIN 7 software (OriginLab Corporation).
MPT and MPT-AMP Synthesis in vitro
cPMP was purified as previously described [5] . All reactions were conducted in a buffer containing 0.1 M Tris pH 7.2. For MPT synthesis 100 pmol MoaE, 3500 pmol MoaD and 1000 pmol cPMP were incubated at room temperature for 60 min in the presence of 1000 pmol PfuMoaB-WT or PfuMoaB-H3 variant. Next, the reaction mix was treated for 1 min at the respective temperature prior to the initiation of adenylylation with 5 mM ATP and 1 mM MgCl 2 . MPT and MPT-AMP were oxidised to their stable, fluorescent derivates, FormA and FormA-AMP [8] , respectively, and quantified by high-pressure liquid chromatography (HPLC).
Determination of MPT and MPT-AMP Content
Synthesised MPT and MPT-AMP were converted with iodine into their fluorescent derivates FormA and FormA-AMP, respectively, further treated with calf alkaline phosphatase (Roche) to yield FormA-dephospho [8] . FormA-dephospho and FormA-AMP were separated on a HPLC column C4 ReproSil 100, 250 mm x 4.6 mm, 5 µm particle size (Dr. Maisch) attached to an Agilent 1200 HPLC System (GE Healthcare) in 10 mM sodium phosphate buffer pH 3.0 and methanol gradient from 10% to 23% at 2 ml/min flow rate. FormA-dephospho and FormA-AMP were quantified as described [7] , [15] .
Bioinformatic Methods
Coordinates of crystal structures of A. aeolicus MogA (3MCI) and T. thermophilus MogA (3MCH) were extracted from the Protein Data Bank of Research Collaboratory for Structural Bioinformatics (PDB RSCB, http://www.rcsb.org/pdb/home/home.do ); protein sequences were from the protein database of the National Center for Biotechnology Information (NCBI, http://www.ncbi.nlm.nih.gov/protein ), reference number NP_213032.1 and WP_011174010.1, respectively. Interface analysis of protein crystal structures was performed using the PISA server of the European Bioinformatics Institute (EBI) ( http://www.ebi.ac.uk/ ) [24] . Salt bridge content was determined using the “Protein Interaction Calculator” (PIC) server [25] provided by Indian Institute of Science, Bangalore ( http://pic.mbu.iisc.ernet.in/ ). Hydrogen bond content was determined using the “Hydrogen Bond Calculator” server of the Center for Informational Biology of the University Ochanomizu, Japan ( http://cib.cf.ocha.ac.jp/bitool/HBOND/ ).
Results
Determination of the Crystal Structure of PfuMoaB
PfuMoaB was crystallized in the space group P3 1 21 with three molecules (chains A, B and C) in the asymmetric unit. The crystallographic model was refined to a R -factor of 0.18 ( R -free 0.24). The stereo-chemical statistics derived from the final coordinates are summarized in Table S1 . The N-terminal His-tag as well as residues 1–10 in chain A, residues 1–11 in chain B and residues 1–9 in chain C were disordered, and are therefore not present in the model. Side chains of Lys12, Lys24, Arg30, Arg87, Glu134 and Glu169 had poor electron density suggesting their high flexibility.
Each PfuMoaB subunit consists of a slightly bent central β-sheet, surrounded by six α-helices and two 3 10 helices. The β-sheet comprises six strands: five parallel (β1– β4,) and one antiparallel (β5) embedded between the inner β4 and the flanking β6 strands ( Figure 2A ). The molecular architecture is consistent with structures of other MPT-adenylyl-transferases. The functional oligomeric state of PfuMoaB was shown to be a hexamer [8] . In the crystal structure a hexamer can be generated by rotation of the asymmetric unit by 180° around the two-fold crystallographic symmetry axis. Therefore, similar to the other proteins of the MoaB family, PfuMoaB represents a dimer of trimers ( Figure 2B–C ).
10.1371/journal.pone.0086030.g002 Figure 2
Crystal structure of PfuMoaB.
(A) Ribbon representation of PfuMoaB monomer, secondary structure elements, N- and C-termini are labelled; α-helices are coloured in cyan, β-sheets in magenta, 3 10 -helices in green, loops in pink. (B) and (C) top and side view of the PfuMoaB hexamer, respectively. Subunits are shown in different colours. Zoom-in represent ionic interactions at the trimerization interface of PfuMoaB-WT. Residues mediating the contacts between subunits are shown in stick representation and are labelled. (D) Sulfate ion at the active site of PfuMoaB. The residues of the conserved Gly-Gly-Thr-Gly motif and the sulfate ion are shown superimposed with the experimentally phased electron density, contoured at 1 σ.
The active site of MPT-adenylyl-transferases encompasses a highly conserved Gly-Gly-Thr-Gly motif, which is involved in ATP hydrolysis and the coordination of the resulting MPT-AMP diphosphate [6] . In the crystal structure of PfuMoaB residual electron density was observed in close proximity to the Gly-Gly-Thr-Gly motif. As no co-purified MPT or MPT-AMP was detected in PfuMoaB, the observed electron density was attributed to components of the crystallization solution. In crystal structures of PfuMoaB homologues, E. coli MogA and E. coli MoaB, a sulfate anion was detected in the same position close to the Gly-Gly-Thr-Gly motif, most probably mimicking the phosphate of MPT [26] , [27] . PfuMoaB crystals were grown in a solution containing 2-(N-morpholino)ethanesulfonic acid (MES), which with its negatively charged sulfonate part resembles sulfate or phosphate. This suggests a binding of the MES molecule into the active site of PfuMoaB via the sulfonate. The MES molecule was manually modelled into the additional density, resulting in a good fit for the sulfonate part, but high B-factors for the MES morpholino ring, implying a high flexibility of the latter. Therefore, only a sulfate molecule was manually positioned into the respective electron density patch ( Figure 2D ) leading to an improved overall model quality.
The PfuMoaB Hexamer
The surface of each PfuMoaB monomer is involved in the organization of two interfaces: the hexameric and trimeric contact sides, respectively. At the interface between the subunits of the trimer the γ2-helix of one subunit forms a prominent bulge that intrudes into the surface between α5- and α7-helices of the adjacent monomer. Furthermore, the loop connecting β3 with α4 provides a significant contact area with C-terminal residues of α7-helix of the neighbouring subunit within the trimer. Assembly of subunits into trimers is additionally strengthened through several ionic interactions, including salt bridges between Arg109 of one monomer and Glu106 of another monomer ( Figure 2B ), and a cluster of ionic interactions involving Lys100 of one subunit and Arg124 and Glu92 of the adjacent polypeptide chain ( Figure 2C ).
Interactions between two trimers within the PfuMoaB hexamer are mainly facilitated by residues of α3- and α4-helices ( Figure 3A ). Residues Ile59, Leu62, Ile63, Phe66 and Ile69 in α3-helix and Leu97 in α4-helix form the central hydrophobic core of the trimer-trimer interface, which is reinforced by two symmetrical salt bridge pairs: i) Lys70 and Glu67; ii) Lys58 and Asp99 ( Figure 3A ). Thus, hydrophobic residues of α3-helix were identified as key residues for the association of two trimers within the hexamer. These residues are not conserved between homologous proteins from different organisms and positioned distantly from the active site ( Figure 3A , Figure 4 ). Thus, they represent a suitable target for mutagenesis to interfere with the oligomerization of PfuMoaB.
10.1371/journal.pone.0086030.g003 Figure 3
Hexamerization interface and structure-guided mutagenesis of PfuMoaB.
(A) Side view of PfuMoaB hexamer. Two trimers of PfuMoaB are depicted in grey and green. The conserved Gly-Gly-Thr-Gly motif within the active site is shown in the upper PfuMoaB monomer in red. The trimer-trimer interaction interface is boxed-in. α3-helix of the upper monomer is depicted in orange. The zoom-in shows labelled residues at the interface in stick representation. As the interface is build up by identical surfaces of each subunit, the hydrophobic residues of the bottom subunit (Ile59, Leu62, Ile63, Phe66, Ile69, Leu97) are not shown for the sake of clarity. (B) Structure-guided mutagenesis of PfuMoaB. Sequences of α3-helix of PfuMoaB-WT and EcoMogA are coloured in black, flanking residues of the helices are represented in grey. Residues are numbered accordingly to the PfuMoaB sequence. Sequences of the variants PfuMoaB I59R/I63R/F66R and PfuMoaB-H3 are shown. Arginine residues of PfuMoaB I59R/I63R/F66R are depicted in bold. Residues of PfuMoaB-H3, which were not exchanged accordingly to the EcoMogA sequence, are underlined.
10.1371/journal.pone.0086030.g004 Figure 4
Multiple sequence alignment of MPT-adenylyl-transferases from different organisms.
Corresponding MPT-adenylyl-transferases are abbreviated as follows: PfuMoaB, Pyrococcus furious ; StoMoaB, Sulfolobus tokodaii ; BceMoaB, Bacillus cereus; EcoMogA and EcoMoaB, Escherichia coli ; TthMogA, Thermus thermophilus ; AaeMogA, Aquifex aeolicus ; Arabidopsis thaliana ; HsaGephG, Homo sapiens . Secondary structure elements of PfuMoaB are shown. The conserved MPT-binding motif GGTG is highlighted with a red box, the conserved aspartate residue coordinating Mg 2+ - ion with a green box [6] , residues of PfuMoaB α3-helix with a blue box. Highly conserved residues are depicted in white letters and black background; semi-conserved residues are shadowed in grey. Consensus threshold was set to 0.8. Sequences were aligned with Clustal Omega [62] ,and modified with BoxShade server (Swiss Institute of Bioinformatics).
Structure-guided Mutagenesis of PfuMoaB
In order to disrupt the interaction between two trimers in the PfuMoaB hexamer, first, three central hydrophobic residues of α3-helix - Ile59, Ile63 and Phe66 - were exchanged against positively charged arginines ( Figure 3B ). The resulting variant PfuMoaB I59R/I63R/F66R was expressed in E. coli as N-terminally His-tagged protein and affinity purified with comparable yields to PfuMoaB wild-type (PfuMoaB-WT), however an increased tendency to precipitation was observed. Subsequent size exclusion chromatography showed that the I59R/I63R/F66R variant was purified as a heterogeneous mixture of diverse oligomeric states ( Figure S1 ). This finding suggests that mutagenesis of the selected residues was either not sufficient to disrupt the respective hexamer interface interactions, or that the exposure of hydrophobic residues derived from α3-helix caused non-specific protein-protein interactions.
Next, the entire α3-helix of PfuMoaB-WT was exchanged against the corresponding helix of the trimeric E. coli MogA (EcoMogA) homologue with two exceptions. In the PfuMoaB-WT structure, Ala64 and Ala68 are positioned at the opposite side of the hexamer interface. Their replacement against the corresponding EcoMogA Thr and Leu residues, respectively, would result in the introduction of two bulky residues and, consequently, would lead to steric clashes. Considering the fact that alanine has high helix-forming propensities, Ala64 and Ala68 of PfuMoaB were not subjected to mutagenesis. The generated variant – PfuMoaB-H3– contained an 11-residues substitution ( Figure 3B ), which resulted in the introduction of charged residues at the hydrophobic interface between trimers and disruption of two hexamer-stabilizing salt bridge pairs: Lys58-Asp99 and Glu63-Lys70.
Biochemical Characterization of the PfuMoaB-H3 Variant
The PfuMoaB-H3 variant was purified to homogeneity in the same way as PfuMoaB-WT ( Figure 5A ). Interestingly, PfuMoaB-H3 was observed to migrate slower in the SDS-PAGE than expected with a corresponding band of 25 kDa in size, while its predicted mass of 19,887.68 Da is similar to that of PfuMoaB-WT (19,971.95 Da). In order to confirm the accurate mass of PfuMoaB-H3, purified PfuMoaB-WT and PfuMoaB-H3 were subjected to mass spectrometry analysis, which detected masses of 19,872 Da and 19,888 Da, respectively (data not shown). Therefore, the observed altered gel mobility of PfuMoaB-H3 could be a result of an altered mass-to-charge ratio within the SDS-PAGE due to reduced SDS binding to the protein as demonstrated for other proteins [28] , [29] .
10.1371/journal.pone.0086030.g005 Figure 5
Biochemical characterization of the PfuMoaB-H3 variant in comparison to PfuMoaB-WT.
(A) 15% Coomassie-Blue-stained SDS polyacrylamide gel showing 200 pmol of Ni-NTA-purified PfuMoaB-WT and PfuMoaB-H3. (B) Far-UV CD spectra of Ni-NTA purified PfuMoaB-WT (solid line) and PfuMoaB-H3 (dotted line). (C) Size exclusion chromatography of Ni-NTA purified PfuMoaB-WT and PfuMoaB-H3. 5 nmol of WT and 10 nm of PfuMoaB-H3 were applied on a Superdex 200 10/300 column. Peaks referring to the different oligomerization states of both proteins are labelled. Molecular masses were determined using protein standards. Elution of PfuMoaB-WT is shown as solid line, the PfuMoaB-H3 variant as dotted line. (D–E) SDS-PAGE of cross-linked PfuMoaB-WT and PfuMoaB-H3 with BS 3 (D) and EDC (E). Samples without addition of cross-linkers were used as control (“–”). Observed oligomeric forms of both proteins are labelled. The cross-linked protein bands with a size corresponding to the trimers (designated with *) were further subjected to mass spectrometry analysis. (F) Differential scanning calorimetry of MPT-adenylyl-transferases. Melting curves of PfuMoaB-WT, PfuMoaB-H3, EcoMogA, EcoMoaB and AthCnx1G recorded by DSC. The maximum of each peak represents the respective T m value. Average T m values for each protein are summarized in the Table 2 . Measurements were performed in duplicate for each experiment.
In order to investigate protein folding, purified PfuMoaB-H3 and PfuMoaB-WT were subjected to circular dichroism (CD) spectroscopy. The far-UV CD spectrum of the H3 variant resembled that of PfuMoaB-WT in shape with a slight signal increase in the range of 210 to 225 nm ( Figure 5B ), where both α-helical and β-sheet elements have their negative maxima [30] . Possibly, the introduced changes caused partial transformation of some loop-containing regions of the protein variant into folded elements.
Next, the oligomerization state of the PfuMoaB-H3 variant was determined by size exclusion chromatography. Three major peaks with molecular masses of 74, 41 and 20 kDa were observed for PfuMoaB-H3, which correspond to trimers, dimers and monomers, respectively ( Figure 5C ). This finding confirmed on one hand the successful dissociation of the hexamer in PfuMoaB-H3 (as compared to PfuMoaB-WT), but on the other hand suggested weakening of interactions within the PfuMoaB-H3 trimer, giving rise to the presence of monomers and dimers in addition to trimers.
In order to demonstrate that the observed trimeric form of the PfuMoaB-H3 variant was not a random assembly of monomers, but rather the result of monomer-monomer interactions similar to that at the trimer interface of PfuMoaB-WT ( Figure 2B–C ), we performed chemical cross-linking of both proteins using two different reagents: (i) bis(sulfosuccinimidyl)suberate (BS 3 ), which enables formation of bonds between primary amines and (ii) ethyl-3-[3-dimethylaminopropyl]carbodiimide hydrochloride (EDC), which couples carboxyl groups to primary amines. Cross-linked PfuMoaB-H3 and PfuMoaB-WT were analysed by SDS-PAGE. Thereby, PfuMoaB-WT was observed mainly as hexamer following BS 3 -cross-linking ( Figure 5D ), or - as trimer and dimer upon EDC treatment ( Figure 5E ). In contrast, the PfuMoaB-H3 variant was found to form only dimers and, to a lesser extent, trimers, when cross-linked with BS 3 or EDC, thus confirming the results of size exclusion chromatography.
To identify the respective contact sites within the EDC-cross-linked trimers, protein bands of PfuMoaB-WT and the PfuMoaB-H3 variant were isolated from the SDS gel ( Figure 5E ) and subjected to peptide mass finger printing. The majority of the identified peptide sequences represented cross-links within monomers ( Table 1 ). Similar quantities of peptides corresponding to surface-exposed intra-molecular salt-bridges were detected for WT and the PfuMoaB-H3 variant demonstrating that monomers of both proteins exhibit similar folds as already suggested by CD spectroscopy. In contrast, peptides embracing Arg109 and Glu106, which form a salt-bridge at the trimer interface, were absent for both PfuMoaB-WT and PfuMoaB-H3 ( Table 1 ), probably due to the deeply buried position of Arg109 ( Figure 2B ). Peptides comprising residues of another trimer interface salt-bridge between Glu92 and Lys100 were found only in PfuMoaB-H3 ( Table 1 ). In the crystal structure of PfuMoaB-WT both Glu92 and Lys100 are buried within the hexamer. Therefore, our results suggest a solvent-exposed nature of the inter-trimer cluster Glu92-Arg124-Lys100 in the PfuMoaB-H3 variant, which is in agreement with a dissociation of the hexamer along the two-fold symmetry axis. The identification of the characteristic inter-subunit interactions of the trimer in the cross-linked PfuMoaB-H3 supports its ability to oligomerize using similar contact sites as WT PfuMoaB and thus confirms the existence of WT-like trimers in solutions. However, in PfuMoaB-H3 also additional cross-linked peptides were observed, which were not detected in the PfuMoaB-WT ( Table S2 ). This suggests an increased accessibility of corresponding surface residues, due to a higher dissociation of the trimer as detected in size exclusion chromatography.
10.1371/journal.pone.0086030.t001 Table 1
Analysis of the EDC cross-linked trimers of PfuMoaB-WT and the PfuMoaB-H3 variant by peptide mass fingerprinting.
Cross-link type
Cross-linked peptides 2
Cross-linked residues 2
N° PSMs 1
WT
H3
intra-subunit
SY E EVGYATVLTR - A K SYEEVGYATVLTR
E114-K111
6
6
SGPLIIE E LSK - TGLEII K SEVFHILK
E41-K157
7
9
TGL E IIK - A K SYEEVGYATVLTR
E154-K111
3
2
TF K FGVITVSDK - LG E HVYYK
K15-E47
4
0
DITI E SIKPLFDK - AKSYEEVGYATVLT R
E92-R124
0
8
DITI E SIKPLFDK - SYEEVGYATVLT R
E92-R124
0
3
DITI E SIKPLFDKELSFGEVFR - SYEEVGYATVLT R
E92-R124
0
2
inter-subunit
DITI E SIKPLFDK - DITIESIKPLFD K
E92-K100
0
6
1 trypsinized cross-linked peptides were identified by LC-MS/MS; their sequences and number of peptide-spectrum matches (PSMs) are shown. 2 residues, which are suggested to be cross-linked by EDC are shown in a separate column and underlined in the respective peptide sequences.
Thermal Stability of the PfuMoaB-H3 Variant
To probe the impact of oligomerization on thermal stability of PfuMoaB, the melting temperatures ( T m ) of both PfuMoaB-WT and the PfuMoaB-H3 variant were determined by differential scanning calorimetry (DSC). Thermal melting of PfuMoaB-WT and PfuMoaB-H3 followed single endothermic transitions, with a T m of 108.6°C and 93.3°C, respectively ( Figure 5F ), suggesting a simultaneous dissociation of the oligomers and denaturation of the protomers. The reduction of the apparent melting temperature in PfuMoaB-H3 by 15°C as compared to PfuMoaB-WT demonstrates the significant contribution of the trimer-trimer interface in thermal stability of PfuMoaB.
For comparison, T m values of mesophilic homologues of PfuMoaB were determined: trimeric E. coli MogA (EcoMogA) and the G-domain of A. thaliana Cnx1 (AthCnx1G), as well as hexameric E. coli MoaB (EcoMoaB). They were expressed as His-tagged fusions in E. coli and purified to homogeneity ( Figure S2A ). Following the determination of their physiological oligomeric state by size exclusion chromatography ( Figure S2B ), DSC analysis revealed T m values of 76.1, 69.6 and 71°C for EcoMogA, EcoMoaB and AthCnx1G, respectively ( Figure 5F ). Thus, the observed melting points of the PfuMoaB mesophilic homologues were relatively close to each other and significantly lower than that of the PfuMoaB-H3 variant ( Table 2 ). Interestingly, hexameric EcoMoaB showed an even lower melting temperature than trimeric EcoMogA. Given the remaining difference in the melting temperatures between PfuMoaB-H3 and its mesophilic homologues, we conclude that additional factors contribute to the thermal stability of PfuMoaB.
10.1371/journal.pone.0086030.t002 Table 2
Melting temperatures ( T m ) of MPT adenylyl-transferases from different organism measured by differential scanning calorimetry (DSC).
Protein 1
T m [°C] 2
PfuMoaB-WT
108.6±0.5
PfuMoaB-H3
93.3±0.5
EcoMoaB
69.6±0.8
EcoMogA
76.1±0.3
AthCnx1G
71.0±0.3
RnoGephG [61]
72.4
1 proteins from following organisms were used: PfuMoaB, Pyrococcus furiosus ; EcoMoaB and EcoMogA, Escherichia coli ; AthCnx1G, Arabidopsis thaliana ; RnoGeph, Rattus norvegicus . 2 average T m values derived from two independent DSC experiments.
Functional Characterization of the PfuMoaB-H3 Variant
To understand the contribution of hexamer formation to the PfuMoaB function, the activity of the purified PfuMoaB-H3 variant and PfuMoaB-WT was analysed by monitoring the conversion of MPT into MPT-AMP at different temperatures. MPT was first synthesised in vitro using purified E. coli cPMP and MPT synthase subunits MoaD and MoaE in the presence of PfuMoaB-WT or PfuMoaB-H3 and was subsequently adenylylated through addition of ATP and MgCl 2.
PfuMoaB-WT showed a temperature-dependent increase in MPT-AMP synthesis with a minor decrease at longer incubation times at 80°C ( Figure S3E ). While at 50°C both WT and H3 exhibited similar activities ( Figure S3C ), at high temperatures (65°C and 80°C) a significant activity loss of PfuMoaB-H3 was observed ( Figure S3D –E). Interestingly, at moderate temperatures of 25°C and 35°C, the H3 variant was 16-fold and 7-fold, respectively, more active than PfuMoaB-WT ( Figure 6 , Figure S3A –B). First, these results demonstrate that PfuMoaB-H3 remained catalytically potent, which supports our previous observation that the overall folding of the protomers has not been dramatically changed. Second, the decreased number of interactions in PfuMoaB-H3 following the dissociation of the hexamer, have probably enhanced the degree of conformational freedom within PfuMoaB-H3, thus leading to the improved catalytic velocity at mild temperatures.
10.1371/journal.pone.0086030.g006 Figure 6
In vitro adenylylation of MPT by PfuMoaB-WT and PfuMoaB-H3 variant.
Adenylylation rates were determined for both proteins at 25, 35, 50, 65 and 80°C by monitoring formation of MTP-AMP over time. Initial velocities of PfuMoaB-WT and the H3 variant at different temperatures are depicted as solid and dotted lines, respectively. Error bars represent the standard deviation of data obtained in at least two independent experiments.
Discussion
In the current study the impact of hexamerization on the thermostability of the tungsten cofactor synthesising protein PfuMoaB has been investigated. Based on the crystal structure of PfuMoaB, which we determined at 2.5 Å resolution, residues of the α3-helix were shown to form the central core of the hexamerization interface. Exchange of α3-helix against the corresponding helix of the trimeric PfuMoaB homologue EcoMogA resulted in a dissociation of the PfuMoaB hexamer and a decrease of the apparent melting temperature by more than 15°C. Therefore, hexamerization of PfuMoaB is a key determinant in thermal stability of PfuMoaB. Our finding supports earlier studies that have identified increased oligomerization states in multi-subunit proteins from thermophilic organisms [31] – [33] . At this point, we cannot exclude that, in addition to the changed oligomerization, the introduction of the new a3-helix on its own has change the stability of the PfuMoaB-H3 monomer. However, given the fact that the DSC profile of PfuMoaB-H3 again shows only a single transition from the folded to the unfolded state, suggests that similar to WT, PfuMoaB-H3 denatures in a single step. Together with our cross-linking studies we conclude that the introduced a3-helix unfolds in concert with the remaining protein and does not impact the overall fold of the PfuMoaB protomer significantly. Unfortunately, we were not able to produce crystals for structural analysis of PfMoaB-H3, which in turn could have provided ultimate proof of the structural alterations in PfMoaB-H3.
The fact, that the melting temperature of PfuMoaB-H3 was still 17–23°C higher than that of its mesophilic homologues suggests besides hexamerization additional factors contributing to the thermal stability of PfuMoaB. Various studies identified structural determinants such as increased number of hydrogen bonds [34] , [35] and ionic interactions [36] , [37] , shortening of loop regions [38] , [39] , more extensive hydrophobic interactions [40] , [41] , as well as decreased content of thermolabile residues [42] , [43] and increased proline content [44] , [45] as mechanisms to increase stability of proteins. Furthermore, it has been shown that thermostable proteins belonging to the same family can develop different strategies to adapt to high temperatures [46] – [48] . The latter will explain the presence of non-hexameric MPT-adenylyl-transferases of the MogA-subfamily in some thermophilic organisms, e.g. bacteria Thermus thermophilus (TthMogA) and Aquiefex aeolicus (AaeMogA). As an increase in the number of stabilizing inter-molecular interactions via hexamerization is not available for those enzymes, other mechanisms must contribute to thermal stability. We compared crystal structures and polypeptide sequences of trimeric TthMogA and AaeMogA with hexameric PfuMoaB and observed an increased number of intramolecular salt bridges and an increased content of proline residues in both TthMogA and AaeMogA (table S3). Furthermore, a remarkably high number of salt bridges at the trimerization interface of AaeMogA was found (table S3). These findings suggest that stabilization of the individual subunits and the trimer are the key mechanisms in MogA-type of MPT-adenylyl-transferases to adapt to elevated temperatures.
The impact of PfuMoaB mutagenesis and the resulting change of the oligomerization state on catalytic activity of the PfuMoaB-H3 variant was investigated at different temperatures. Significant reduction of the PfuMoaB-H3 activity was observed at temperatures much lower than its apparent melting point, indicating that the loss of activity was initiated before irreversible denaturation of the protein. Similar findings were reported for 3-phosphoglycerate kinase from thermophilic Thermoanaerobacter sp. Rt8.G4 and mesophilic Zymomonas mobilis [49] , 2-keto-3- deoxygluconate aldolase from S. solfataricus [50] and citrate synthase from Thermoplasma acidophilum [51] . One can assume that increased conformational flexibility of the active site enhances its susceptibility to unfolding upon increase of temperature and therefore, during the process of thermal denaturation, the active site is one of the first regions to adapt a partially unfolded conformation. Consequently, the loss of enzyme activity is seen at lower temperatures than irreversible unfolding of the entire protein [49] .
Remarkably, at moderate temperatures of 25°C and 35°C PfuMoaB-H3 was found to be more active in comparison to the WT protein. To our knowledge, this is the first report of increased catalytic activity for a given enzyme, following the alteration of its oligomerization state. Activity and conformational flexibility of enzymes are tightly interconnected thus reflecting their dynamic nature. During catalysis, enzymes undergo sequential conformational changes to enable substrate binding and stabilization of the transition state, followed by product formation and release [52] – [54] . Increased stability of proteins from thermophilic organisms inevitably leads to an increase in rigidity, which in turn affects their conformational flexibility. For 3-isopropylmalate dehydrogenase from T. thermophilus , a lower activity of the enzyme at room temperature in comparison to its mesophilic homologue from E. coli was attributed to restricted conformational movements at sub-optimal temperatures of the thermophilic protein [55] . In contrast, at temperatures close to the optimal activity ( T opt ) the conformational flexibility of the enzyme significantly increased [55] . Similar findings were reported for thermostable glyceraldehyde-3-phosphate dehydrogenase from T. maritima [56] .
Disruption of major stabilizing interactions at the interface of the PfuMoaB hexamer might have provided an additional degree of freedom to residues in individual subunits and thus, increased their conformation flexibility and resulted in much higher activity at moderate temperatures. It is important to note that a3-helix extends into a loop hosting the highly conserved Asp56 ( Figure 7 ), which was previously found to be crucial for the catalytic function of the plant homologue Cnx1G [57] . Later we could demonstrate that the corresponding Asp515 of Cnx1G is involved in the coordination of Mg 2+ -ion during ATP-dependent MPT adenylylation [6] , [7] . Consequently, once can assume that Asp56 in PfuMoaB undergoes conformational changes during catalysis. Therefore, increased conformational flexibility, introduced by the exchange of a3-helix, has probably released the structural constrain present in the hexameric PfuMoaB thus explaining the gain in catalytic velocity at lower temperature.
10.1371/journal.pone.0086030.g007 Figure 7
Active site of PfuMoaB-WT.
Two PfuMoaB subunits at the hexamerization interface are shown as ribbon in green and grey, respectively. The conserved Asp56 residue coordinating Mg 2+ (pink) is shown in sticks. MPT-AMP in the active site is derived from a superimposition with the structure of the PfuMoaB homologue A. thaliana Cnx1G (1UUY). The Mg 2+ -ion derived from a superimposition with the homologues sub-domain 3 of E. coli MoeA (1FC5) [6] , [63] .
Remarkably, Kananujia et al. showed by molecular-docking simulations that MogA proteins bind MPT stronger than MoaB proteins [58] . Together with the observed increased activity of PfuMoaB-H3, these results might suggest that lower oligomeric forms of MPT-adenylyl-transferases are more efficient at moderate temperatures. This might explain why only representatives of the MogA sub-family, but not of the MoaB-sub-family, are found today in eukaryotes ( Figure S4 ).
Interestingly, MogA and MoaB homologues are not equally distributed in prokaryotes. In Bacteria, species with only MogA ( e. g. A. aeolicus , Clostridium botulinum ), only MoaB ( e. g. B. cereus , Streptomyces avermitilis , Vibrio vulnificus ), or both homologues ( e. g. E. coli, Azorhizobium caulinodans, Bordetella bronchiseptica ) are detected. In contrast, search of the non-redundant protein sequence database in Archaea revealed only three organisms containing MogA proteins, which belong to the phylogenetically close orders of Desulforococcales ( Pyrolobus fumarii and Aeropyrum pernix ) and Acidolobales ( Acidilobus saccharovorans ). Domination of MoaB in Archaea might indicate that MogA is unable to fulfil yet unknown functions of MoaB in archaeal organisms. In E. coli , enzymes catalysing the last steps of Moco biosynthesis (MogA, MoeA, MobA and MobB) were found to form a transient multi-protein complex in vivo , most probably to protect unstable reaction intermediates, and to secure targeting of the cofactor to the corresponding apo-enzymes [59] . Considering the fact that MoaB forms hexamers, this oligomerization state might provide an additional surface for interactions with enzymes upstream and downstream of MPT adenylylation.
Noteworthy, EcoMoaB was found to be inactive for MPT adenylylation both in vivo and in vitro [8] , [60] demonstrating that its function in Moco biosynthesis was completely taken over by the MogA homologue. Future functional characterisation of MoaB homologues from bacterial organisms without MogA might clarify whether MoaB proteins are Archaea-specific proteins unable to function in Bacteria, or whether the MoaB-type MPT-adenylyl-transferases are common for all prokaryotes, while MogA is a Bacteria-specific representative.
Supporting Information
Figure S1
Size exclusion chromatography (SEC) of the PfuMoaB I59R/I63R/F66R variant. SEC elution profile of Ni-NTA purified PfuMoaB I59R/I63R/F66R variant using Superdex 200 16/60 pg column (GE Healthcare). Collected fractions (highlighted with arrows) were subsequently analysed by SDS-PAGE (shown in the SEC-chromatogram).
(DOCX)
Figure S2
Purification of His-tagged fusion proteins EcoMogA, EcoMoaB and AthCnx1G expressed in E. coli. (A) Coomassie-Blue-stained 15% SDS polyacrylamide gel of Ni-NTA purified proteins. (B) Size exclusion profiles of the Ni-NTA purified EcoMogA, EcoMoaB and AthCnx1G using Superdex 200 10/300 column. 10 nmol of each protein was applied. Observed molecular masses of the peaks correspond to EcoMoaB hexamers and EcoMogA and AthCnx1G trimers.
(DOCX)
Figure S3
In vitro adenylylation of MPT by PfuMoaB-WT and the PfuMoaB-H3 variant. Adenylylation rates were determined for both proteins at 25°C (A), 35°C (B), 50°C (C), 65°C (D) and 80°C (E) by monitoring formation of MTP-AMP in time. Adenylylation kinetics of PfuMoaB-WT and the PfuMoaB-H3 variant are depicted as solid and dotted lines, respectively. Measurements were performed in duplicate for each experiment, error bar represent the standard deviation of data obtained in two separate experiments.
(DOCX)
Figure S4
Phylogenetic tree of MPT-adenylyl-transferases. MPT adenylyl-transferases from following organisms are shown: A. thaliana (AthCnx1G), H. sapiens (HsaGephG), T. thermophilus (TthMogA), A. aeolicus (AaeMogA), E. coli (EcoMoaB and EcoMogA), B. cereus (BceMoaB), S. tokodaii (StoMoaB) and P. furiosus (PfuMoaB). Branch lengths are shown in per cent. Red branches - Archaea, violet - Bacteria, blue - Eukaryotes. Phylogenetic tree was prepared using web server of the Le Laboratoire d’Informatique, de Robotique et de Microélectronique of University Montpellier, France www.phylogeny.fr [64] .
(DOCX)
Table S1
Collection of diffraction data and refinement statistics.
(DOCX)
Table S2
Non-specific cross-links in trimers of PfuMoaB-WT and the PfuMoaB-H3 variant identified by peptide mass fingerprinting.
(DOCX)
Table S3
Comparative analysis of P. furiosus MoaB, A. aeolicus MogA and T. thermophilus MogA.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Idiopathic pulmonary fibrosis (IPF) is a devastating disease; the median survival of IPF patients is two to three years [ 1 ]. However, the natural course of IPF is variable; some patients die within a year after diagnosis, whereas some live much longer. Predicting disease progression and survival in IPF patients is important in deciding on treatment and for prompt consideration of lung transplantation. Several pulmonary functional tests—a six-month decrease of more than 10% in FVC or 15% in D L, CO —have been shown to be useful predictors of mortality in IPF patients [ 2 , 3 ]. Moreover, several serum biomarkers—surfactant proteins A and D (SP-A/SP-D, Refs. [ 4 , 5 ]), KL-6 [ 6 ], and matrix metalloproteinase-7 [ 7 ]—have also been reported to be correlated with prognosis. However, if we apply the six-month decrease of FVC or D L, CO as a biomarker to arrive at a prognosis, it takes too long when there is generally a life expectancy of only two to three years. Since most studies of serum biomarkers were performed in a single facility, a wider validation, in separate cohorts or in a multicenter analysis, has been needed. Therefore, a simple, practical, and accurate prognostic indicator in IPF is required.
Periostin is an extracellular matrix (ECM) protein belonging to the fasciclin family [ 8 , 9 ]. Periostin is induced in fibroblasts by several stimuli, including TGF-β and IL-4/IL-13. We and others previously demonstrated that periostin was highly expressed in the lungs of bleomycin-administered mice or IPF patients [ 10 – 12 ]. Periostin expression is observed in the fibroblastic foci and adjacent to α-smooth muscle actin—positive myofibroblasts, suggesting that these are the main periostin-producing cells in pulmonary fibrosis. Periostin acts on fibroblasts together with inflammatory cytokines such as TNFα or IL-1α activating NF-κB, followed by production of various inflammatory cytokines and chemokines, leading to generation of fibrosis in the lungs [ 10 , 13 ]. Thus, periostin is a key player in the pathogenesis of pulmonary fibrosis. We and others then found that serum periostin was significantly up-regulated in IPF patients [ 11 , 12 , 14 ]. Serum periostin was associated with six-month decreases of VC or D L, CO [ 11 ], clinical progression [ 14 ], and overall survival and time-to-event [ 14 ]. Taken together, these results suggest the potential of periostin as a prognostic biomarker in IPF.
However, several problems have remained unresolved. First, the previous study was performed in a single facility, and the investigated number was limited. Second, it is known that serum periostin is up-regulated in patients with various inflammatory diseases other than IPF [ 15 ]. Establishment of a detection system for serum periostin more specific to IPF is needed. Although it was reported that splicing out of exon 21 is high in IPF patients, no kit to detect splicing out of exon 21 has been developed [ 16 ]. In this study, we have developed a new periostin kit specifically detecting monomeric periostin, which showed a high ratio to total periostin in IPF compared to other periostin-high diseases. We then utilized monomeric periostin in a multicenter study to evaluate its usefulness for diagnosing IPF and for predicting IPF progression.
Materials and methods
Antibodies
Mouse monoclonal anti-periostin antibodies SS19C, SS19D, and SS20A were established as follows. Animal studies were undertaken following the guidelines for care and use of experimental animals of the Japanese Association for Laboratory Animals Science (1987) and were approved by the Saga University Animal Care and Use Committee (Saga, Japan). Mice were intraperitoneally immunized at 2- to 4-week intervals with drosophila S2 cell-derived recombinant periostin protein emulsified in TiterMax Gold adjuvant (TiterMax USA, Norcross, GA) [ 17 ]. At least two months later, splenocytes were prepared from the mice and fused with Sp2/O myeloma cells under a standard fusion protocol using polyethylene glycol. Hybridomas producing anti-periostin IgG were selected by enzyme-linked immunosorbent assay (ELISA) with S2 recombinant periostin as an immobilized antigen. IgG was purified from culture supernatant of the hybridomas using protein G affinity chromatography. Some of the antibodies were biotinylated using EZ-Link Sulfo-NHS-LC-Biotin (Thermo Fisher Scientific, Waltham, MA), or were conjugated with horseradish peroxidase (TOYOBO, Osaka, Japan) using Sulfo-EMCS (Dojindo Laboratories, Kumamoto, Japan), according to the manufacturers’ instructions. The epitopes of SS19D and SS20A, newly generated antibodies, are the R3 domain and EMI domain, respectively, whereas the epitopes of SS18A and SS17B used for the conventional periostin ELISA are the R1 domain and the R4 domain, respectively.
ELISA
The conventional periostin ELISA detecting total periostin (SS18A (the capture antibody)×SS17B (the detection antibody)) was previously described [ 11 ]. The new periostin ELISA was made of SS20A (the capture antibody) and SS19D (the detection antibody). Sensitivity (the limit of blank: LOB, the limit of detection: LOD, and the limit of quantification: LOQ) was evaluated according to CLSI EP-17A (National Committee for Clinical Laboratory Standards). LOB was calculated using the reproducibility of blank sample (0 ng/mL). Blank serum samples were prepared by using immunoprecipitation technique. LOD was evaluated using samples with low serum levels prepared by using immunoprecipitation. LOQ was determined with CV values obtained by intra-assay reproducibility. LOQ was defined as the lowest concentration of monomeric periostin quantifiable with CV of 10% or less.
Preparation of monomeric periostin
Recombinant periostin proteins derived from Drosophila S2 cell were prepared as previously described [ 17 ], and monomeric periostin was purified with NHS-activated Sepharose beads (GE Healthcare, Little Chalfont, UK) conjugated with SS19D that was reactive to monomeric periostin ( S1 Fig ). Protein assay and immunoassay were used to quantify total monomeric periostin concentration. The Bradford analysis (Bio-Rad, Hercules, CA) was used for protein assay: conventional periostin ELISA (SS18A×SS17B) was used for the ELISA assay. Standard curve of the conventional periostin ELISA kit and the new periostin ELISA kit for the purified periostin protein (oligomer and monomer) and the purified monomeric periostin protein is depicted in S2 Fig .
Immunoprecipitation and immunoblotting
Monoclonal anti-periostin antibodies were immobilized onto NHS-activated Sepharose beads at the rate of 1 mg antibody per 1 mL beads. Periostin in 0.5 mL sera was immunoprecipitated by 10 μL of SS17B-, SS18A-, SS19C-, SS19D-, or SS20A-beads, and was eluted by sample buffer containing 4% sodium dodecyl sulfate, 20% glycerol and 0.1M Tris (pH 6.8), with or without 12% 2-mercaptoethanol, followed by SDS-PAGE. Immunoblotting was essentially performed as described previously [ 18 ]. For serial immunoprecipitation, firstly periostin in 5 mL or 15 mL sera was immunoprecipitated by 10 μL of SS18A- or SS20A-beads, respectively, and the bound periostin was eluted with 0.25 M glycine (pH 2.5), followed by buffer exchange to PBS. Secondly, periostin in the eluates was immunoprecipitated by 10 μL of SS17B- or SS19D-beads, respectively, and the bound periostin was then eluted by sample buffer without 2-mercaptoethanol.
Subjects of Idiopathic Interstitial Pneumonias (IIPs)
This study was conducted by the Consortium for Development of Diagnostics for Pulmonary Fibrosis Patients (CoDD-PF, the CoDD-PF study) composed of seven hospitals (Saga Medical School, Kurume University, Oita University, Nagasaki University, Shinshu University, National Hospital Organization Kyusyu Medical Center, and Dokkyo Medical School Koshigaya Hospital) after approval by the local institutional review boards.
The patients were enrolled from 2011 to 2014. Eligible patients with IPF and fibrotic non-specific interstitial pneumonia (fNSIP) were selected according to multidisciplinary diagnosis (MDD) following global criteria [ 19 , 20 ] by two each of board-certified clinical, radiological, or histological investigators who had considerable experience in thoracic diagnosis. Concomitant therapy with up to 10 mg or the equivalent of prednisone per day was tolerated. Patients receiving other therapies for IPF, including high-dose prednisone, immunosuppressant, pirfenidone, N-acetylcysteine, and any investigational treatments for IPF, were excluded. Patients with IPF and fNSIP were eligible to participate in the present study if they were between 20 and 80 years of age and had been clinically stable with no disease exacerbation for more than 3 months prior to the first observation day. Pregnant women were excluded. Diagnosis of acute exacerbation of IPF was defined in accordance with the criteria detailed in a previous report [ 21 ].
We observed the study subjects for up to one year after the study start date (day 0). Lung functions, thin-section computed tomography (CT) images, and biomarkers were evaluated on day 0 and when the patients visited the hospitals 6–12 months after the starting date (the difference between these two times was defined as short-term change). The median (25th percentile to 75th percentile of interquartile range) of short-term duration was 189 (175–260) days. The same criteria were applied to the patients recruited for the study of biomarkers of interstitial lung disease (ILD) in the Kurume study presented in previous papers [ 11 , 14 ], and the combined data from both the CoDD-PF and the Kurume studies were analyzed. The characteristics of the selected patients are described in S1 Table .
Thin-section CT image and score interpretation
CT images were independently analyzed by two board-certified chest radiologists who were blinded to clinical information. The radiological patterns of interstitial pneumonias were as follows—definite usual interstitial pneumonia (UIP), possible UIP, or inconsistent with UIP pattern—all according to global criteria [ 22 ]. The extent of radiologic abnormalities shown in CT imaging (CT scores) were evaluated as reported previously [ 22 , 23 ].
The protocol consisted of end-inspiration in the supine position, with 0.5- to 1.5-mm collimation sections reconstructed with a high-spatial-frequency algorithm at 1-cm or 2-cm intervals. Images were interpreted at a window setting appropriate for viewing the lung parenchyma (window level, -600 to -700 Hounsfield units [HU]; window width, 1200 to 1500 HU).
The radiologists evaluated the extent of the thin-section CT features, which included the presence of ground-glass attenuation, reticulation, honeycombing, emphysema, and traction bronchiectasis. The lungs were divided into six zones (upper, middle, and lower on both sides), as reported previously [ 22 , 23 ]. The extents of all radiologic abnormalities were expressed as the percentage of lung parenchyma affected in each of the six zones, to the nearest 5%, and were averaged. We summed both the scores for honeycombing and reticulation into a reticular score. The scores for traction bronchiectasis were calculated by using the following formula: traction bronchiectasis score = (6 –most distal generation score) × number of segments: where the most distal generation score was obtained by assessing the generations of the most distal bronchial branches that were dilated as follows: 3, third-generation bronchi (segmental bronchi) as the main bronchus is the first generation; 4, fourth-generation bronchi (subsegmental bronchi); 5, distal to the fifth-generation bronchi; and where the number of segments was obtained by summing the number of pulmonary segments with traction bronchiectasis. The main bronchi and lobar bronchi were not counted. The range of traction bronchiectasis score that could be taken was 0 to 54.
Disagreements with respect to CT findings between the two radiologists were resolved by consensus after assessing the inter-observer agreement. Inter-observer agreements in evaluation of chest CT findings were analyzed as reported previously [ 11 , 14 ]. Inter-observer agreement with regard to classification of the CT pattern of IIP was excellent (kappa, 0.87, p <0.001, data not shown). Assessments of the extent of CT abnormalities and traction bronchiectasis score showed a significantly high correlation between the two independent observers (Spearman r , 0.74–0.93, p <0.001, data not shown) evaluated by using Spearman’s rank correlation coefficient.
Histological interpretation
Sixteen (10 from the CoDD-PF study and six from the Kurume study) lung specimens obtained by surgical lung biopsy (SLB) were made available for diagnosis. The histological patterns of ILD were classified as definite, probable, possible, or not UIP patterns, in accordance with global criteria [ 19 , 20 ] independently by two pathologists who were blinded to clinical information. Disagreement between the two pathologists was resolved by reaching a consensus.
Subjects of Atopic Dermatitis (AD), Systemic Scleroderma (SSc), and asthma
Serum samples were obtained from AD, SSc, and asthma patients as described previously [ 24 – 26 ]. SSc patients were confirmed to be free from ILD.
Statistical analysis
Data were expressed as the mean ± standard deviation (SD). Correlations between the two parameters were evaluated using Spearman’s rank correlation coefficient. Multi-coupled comparisons were analyzed with Bonferroni correction after a Mann-Whitney U test. The cut-off levels for various parameters were defined as the optimal value for distinguishing between two groups, using a receiver operating characteristic (ROC) curve generated by logistic regression [ 23 ]. Linear regressions between variables were calculated by the least-squares method. Differences between linear regression lines were tested using analysis of covariance. All statistical analyses were performed using SPSS software (IBM SPSS Statistics, Chicago, IL) except for the ratios of monomeric periostin, for which Prism software (GraphPad Software, La Jolla, CA) was used.
Results
Characterization of the new periostin ELISA kit
We generated a new periostin ELISA kit composed of SS20A and SS19D among the monoclonal anti-periostin antibodies that we established. Serum periostin mostly existed in the oligomeric form, with only small amounts in the monomeric form ( Fig 1A ). The oligomeric form in the non-reducing condition was dissociated into the monomeric form in the reducing condition, suggesting that the oligomeric form of periostin is assembled by intramolecular disulfide bonds. We found that the newly generated monoclonal anti-periostin antibodies, SS20A and SS19D, which comprised the new periostin kit, recognized the monomeric form, but not or only faintly the oligomeric form ( Fig 1B ). In contrast, SS18A and SS17B, comprising the conventional periostin ELISA kit, recognized both the monomeric and oligomeric forms. Consequently, serial immunoprecipitation by SS20A, followed by SS19D mimicking the new periostin kit, precipitated only the monomeric form. But serial immunoprecipitation by SS18A, followed by SS17B mimicking the conventional periostin kit, precipitated both the monomeric and oligomeric forms ( Fig 1C ). The standard curves of the conventional periostin ELISA kit and the new periostin kit for the purified periostin protein (oligomer and monomer) and the purified monomeric periostin protein are shown in S2 Fig . LOB, LOD, and LOQ were estimated to be 2.7 pg/mL, 3.7 pg/mL, and 6.0 pg/mL, respectively. These results demonstrate that the new periostin kit specifically recognizes the monomeric periostin, whereas the conventional periostin kit recognizes both the monomeric and oligomeric periostin.
10.1371/journal.pone.0174547.g001
Fig 1
Characterization of the new periostin ELISA kit.
(A) Periostin in serum was immunoprecipitated by SS18A or SS19C and detected by SS19C in non-reduced (left panel) or reduced (right panel) conditions. (B) Periostin in serum was immunoprecipitated by SS18A, SS17B, SS20A, or SS19D, respectively, and detected by SS19C in non-reduced conditions. (C) Serial immunoprecipitation of periostin by SS20A followed by SS19D (left lane) or by SS18A followed by SS17B (right lane), respectively, in non-reducing conditions.
Selection of patients with IPF or fNSIP
IPF and fNSIP patients who applied to participate in this study were selected as shown in Fig 2 . In the CoDD-PF study, 107 patients were registered in each collaborating hospital as IIP patients, and 45 cases were selected as eligible IIP patients, comprising 40 IPF patients and five fNSIP patients, respectively. We excluded the remaining 61 cases as non-IPF patients, including one patient who developed microscopic polyangiitis during the study period. Among 40 IPF cases, five were diagnosed with UIP based on pathological examinations. Thirty-five were concordant with the clinical criteria based on the MDD by each of two experienced respiratory physicians, two thoracic radiologists, or two pathologists. All five cases of fNSIP were diagnosed by pathological examinations.
10.1371/journal.pone.0174547.g002
Fig 2
Selection of patients with IPF or fNSIP.
The chart shows selection of patients with IPF or fNSIP. In the CoDD-PF study, 40 IPF patients and five fNSIP patients were selected from 107 enrolled patients. In the Kurume study, 20 IPF patients and two fNSIP patients were selected. In addition, 137 healthy donors were enrolled.
We analyzed whether pulmonary functions were correlated with the baseline or the short-term changes in the CT scores ( S1 and S2 Tables). The short-term change of % D L, CO was associated with the short-term changes of honeycombing, reticular score, or traction bronchiectasis score. These results coincide with previous reports that honeycombing, reticulation, and traction bronchiectasis score are associated with histological fibrotic process in IPF [ 21 ]. We then analyzed the correlations between %VC or % D L, CO at the base line and at various parameters ( S3 Table ). Baseline % VC or % D L, CO was weakly correlated with monomeric periostin or total periostin, respectively.
In the Kurume study, 20 IPF and two fNSIP patients fit the inclusion criteria of the CoDD-PF study. We then combined the data from both the CoDD-PF and the Kurume studies, evaluating the changes of lung functions (%VC and % D L, CO ) and analyzing various biomarkers in all of the eligible patients. We also analyzed the biomarkers present in 137 healthy donors. There was no gender-based difference in monomeric or total periostin in healthy donors (monomeric periostin: male 8.6 ± 1.97 ng/mL, female 8.7 ± 2.11 ng/mL, total periostin: male 63.5 ± 18.67 ng/mL, female 67.5 ± 18.61 ng/mL).
The greatest ability of monomeric periostin in diagnosis of IPF
We first measured monomeric periostin detected by the new periostin kit, total periostin detected by the conventional periostin kit, and conventional biomarkers for IPF—KL-6, SP-D, and LDH—in IPF and fNSIP patients. All of the investigated biomarkers—monomeric periostin (mean: 18.5 ± 9.5 ng/mL vs. 8.6 ± 2.0 ng/mL), total periostin (mean: 101.5 ± 36.2 ng/mL vs. 64.8 ± 18.7 ng/mL), KL-6 (mean: 932.7 ± 557.1 IU/mL vs. 289.3 ± 83.2 IU/mL), SP-D (mean: 230.2 ± 167.2 ng/mL vs. 45.8 ± 39.4 ng/mL), and LDH (mean: 225.6 ± 102.4 IU/mL vs. 150.7 ± 28.8 IU/mL)—were elevated in IPF patients (n = 60) compared to healthy donors (n = 137, p <0.001, Fig 3A ). These biomarkers, with the exception of total periostin, were also high in fNSIP patients (n = 7, monomeric periostin: 14.3 ± 4.7 ng/mL, p <0.01, total periostin: 88.4 ± 31.4 ng/mL, not significant, KL-6: 1042.0 ± 454.5 IU/mL, p <0.001, SP-D: 170.9 ± 89.7 ng/mL, p <0.01, LDH: 220.7 ± 58.6 IU/mL, p <0.01). The distribution of monomeric periostin in healthy donors (n = 137, mean: 8.6 ± 2.0 ng/mL) is shown in S3 Fig . In the analyses with CT scores, monomeric periostin was associated with the short-term change in reticular score ( S4 Table ). Total periostin level was associated with the baselines of honeycombing or reticular score and the short-term change in honeycombing or reticular score. KL-6 was associated with only the baseline of reticulation. SP-D was associated with the baselines of reticulation, reticular score, or traction bronchiectasis score, and the short-term change of honeycombing.
10.1371/journal.pone.0174547.g003
Fig 3
Abilities of each biomarker to diagnose IPF.
(A) Serum levels of each biomarker in IPF patients, fNSIP patients, and healthy donors. Serum levels of monomeric periostin, total periostin, KL-6, SP-D, and LDH in IPF patients (n = 60), fNSIP patients (n = 7) and control donors (n = 137). (B) ROC curve analysis of each biomarker between IPF patients and healthy donors. Monomeric periostin (red), total periostin (orange), KL-6 (black), SP-D (green), and LDH (blue) between IPF patients (n = 60) and healthy donors (n = 137). ***: p <0.001, **: p <0.01.
We next performed ROC curve analyses for IPF patients (n = 60) vs. control donors (n = 137, Fig 3B ). Monomeric periostin had the highest area under the curve (AUC, 0.958) among the investigated biomarkers (total periostin: 0.843, KL-6: 0.948, SP-D: 0.953, LDH: 0.898). When we set the cut-off values of monomeric periostin at 11.2 ng/mL, total periostin at 77 ng/mL, KL-6 at 398 IU/mL, SP-D at 96 ng/mL, and LDH at 166 IU/mL as the optimal points, respectively, the sensitivities and specificities were evaluated as 90.0% and 91.2% for monomeric periostin, 73.3% and 79.6% for total periostin, 88.3% and 92.0% for KL-6, 91.7% and 92.0% for SP-D, and 88.3% and 78.8% for LDH, respectively. These results demonstrate that monomeric periostin has the greatest ability to discriminate IPF patients from healthy donors among the investigated biomarkers comparable with KL-6 and SP-D.
The abilities of monomeric and total periostin to predict short-term progression of IPF
We next examined the correlation of each biomarker with the parameters reflecting short-term progression—short-term changes in %VC and % D L, CO —in IPF patients. Both the changes in %VC and % D L, CO were inversely associated with monomeric periostin ( r = −0.492, p <0.01 for %VC, r = −0.587, p <0.001 for % D L, CO ) and total periostin ( r = −0.428, p <0.01 for %VC, r = −0.460, p <0.01 for % D L, CO , Fig 4 ). SP-D showed a correlation with decline of % D L, CO ( r = −0.319, p <0.05), but not with %VC. Neither KL-6 nor LDH showed a significant correlation with the change of %VC or % D L, CO . It is of note that the correlations between monomeric periostin and decline of %VC were observed independently in both the CoDD-PF ( r = −0.372, p <0.05) and Kurume ( r = −0.686, p <0.01) studies ( S4 and S5 Figs). These results show that both monomeric and total periostin are good biomarkers to predict the short-term progression of IPF.
10.1371/journal.pone.0174547.g004
Fig 4
Ability of each biomarker to predict the short-term progression of IPF.
Correlations between monomeric periostin, total periostin, KL-6, SP-D or LDH and short-term change of %VC (A) or % D L, CO (B) in IPF patients (n = 44 for %VC and 39 for % D L, CO ).
Usefulness of clustering IPF patients into high and low periostin groups to predict short-term progression
Given that both monomeric and total periostin were well correlated with the short-term progression of IPF, we then investigated whether clustering IPF patients into high and low periostin groups would be useful to predict short-term progression. We searched for the optimal cut-off values to maximize the difference in lung function, setting the cut-off values to cluster IPF patients into high and low groups at 15.0 ng/mL for monomeric periostin, 100 ng/mL for total periostin, 1,000 IU/mL for KL-6, 220 ng/mL for SP-D, and 240 IU/mL for LDH. Then we compared the short-term changes of %VC and % D L, CO in the high and low groups. The differences of monomeric and total periostin between the two groups were significant in both %VC (monomeric periostin: −12.5 ± 2.7% vs. −2.0 ± 1.4%, p <0.05, total periostin: −12.5 ± 2.7% vs. −2.5 ± 1.6%, p <0.01) and % D L, CO (monomeric periostin: −22.1 ± 4.8% vs. 3.0 ± 3.6%, p <0.01, total periostin: −17.5 ± 5.5% vs. −0.9 ± 3.8%, p <0·01) changes ( Fig 5 ). Clustering of other biomarkers did not show any significant difference between the high and low groups, even using other cut-off values. These results suggest that clustering IPF patients into high and low periostin groups is useful in predicting short-term progression.
10.1371/journal.pone.0174547.g005
Fig 5
Effects of clustering IPF patients into high and low groups for each biomarker to predict the short-term progression of IPF.
IPF patients were clustered into high and low groups by the cut-off values of monomeric periostin (15.0 ng/mL), total periostin (100 ng/mL), KL-6 (1,000 IU/mL), SP-D (220 ng/mL) or LDH (240 IU/L). Short-term change of %VC (A, n = 44) or % D L, CO (B, n = 39) (upper) and proportion (down) in each high or low group is depicted.
High ratio of monomeric periostin in IPF compared to other high-periostin diseases
We have already shown that total periostin is up-regulated in various diseases other than IPF [ 24 – 26 ]. Given that measuring monomeric periostin is useful in treating IPF, we compared the ratios of monomeric periostin in IPF and three other high-periostin diseases: AD, SSc, and asthma. We confirmed that the SSc patients reviewed for this study were free from ILD. The index of total periostin/monomeric periostin was significantly lower in IPF patients (2.1, n = 40) than in either AD (14.2, n = 224), SSc (11.7, n = 37), or bronchial asthma (7.3, n = 143) patients ( Fig 6 ). These results suggest that a high ratio of monomeric periostin to total periostin is a characteristic of IPF patients.
10.1371/journal.pone.0174547.g006
Fig 6
Comparison of the ratios of monomeric periostin in IPF and other high-periostin diseases.
The comparison between monomeric periostin and total periostin in IPF patients (n = 60, open triangle) and AD (upper, n = 224, dot), SSc (lower left, n = 37, dot), or asthma (lower right, n = 143, dot) patients is shown. All patients or low-range patients in AD are shown in upper left and right panels, respectively. The regression lines are inserted.
Discussion
We previously found that serum periostin is up-regulated in IPF patients, correlated with their decline of %VC and % D L, CO [ 11 ]. The present study is the second prospective, independent validation cohort analysis. We have added or altered the following three points compared to the previous study: (1) we applied the new periostin ELISA kit specifically to detect the monomeric form of periostin in addition to the conventional periostin ELISA kit detecting both monomeric and oligomeric periostin (total periostin); (2) we performed this study in a consortium consisting of seven independent facilities; and (3) we compared the abilities of periostin with other conventional biomarkers for IPF—KL-6, SP-D, and LDH—in diagnosis of IPF and prediction of short-term progression of IPF. Consequently, we have revealed the following three major characteristics of monomeric periostin: (1) it has the greatest ability to diagnose IPF comparable with KL-6 and SP-D compared to total periostin; (2) it is able to predict the short-term progression of IPF comparable with total periostin; and (3) it has a high ratio to total periostin in IPF compared to other periostin-high diseases. We have confirmed these characteristics of monomeric periostin in a multi-center analysis. These results strongly support the usefulness of monomeric periostin as a diagnostic and prognostic biomarker for IPF.
We have shown that the new periostin ELISA kit, composed of SS20A×SS19D, recognizes only monomeric, but not oligomeric periostin. It is known that recombinant periostin protein prefers to form oligomers by intermolecular disulfide-bonds [ 27 , 28 ]. We established that most periostin exists in oligomeric form in serum and that monomeric periostin exists as only a minor fraction. Monomeric periostin may be formed by intramolecular disulfide bonds instead of intermolecular disulfide bonds. The present study shows that production of monomeric periostin would be significantly up-regulated in IPF compared to other periostin-high diseases such as AD, SSc, and asthma. Thus far, its underlying mechanism remains unclear. It is well known that oxidative stress actively contributes to the pathogenesis of pulmonary fibrosis [ 29 , 30 ]. The aberrant redox status in IPF may affect intermolecular or intramolecular disulfide-bond formation in periostin protein. It is known that periostin has eleven cysteine residues and that these cysteine residues are located at the EMI domain (Cys44, Cys60, Cys69, Cys79, Cys80, Cys92), the R1 domain (Cys208), the R2 domain (Cys311, Cys333), and the R3 domain (Cys467, Cys472), respectively. Thus far, it is unclear which cysteine residues are involved in formation of intermolecular or intramolecular disulfide-bond(s). It was previously reported that splicing out of exon 21 is high in IPF patients [ 16 ]. The correlation between production of monomeric periostin and splicing out of exon 21 is unclear, because the C-terminal domain correlated with splicing does not have any cysteine residue. However, we cannot exclude the possibility that the lack of exon 21 may affect the conformation of periostin followed by the formation of aberrant disulfide bond(s). Moreover, we did not observe different formation of oligomeric and monomeric periostin using different stimuli or cells (data not shown).
We consistently showed a good correlation of serum periostin with short-term declines of %VC and % D L, CO , which can predict mortality in IPF patients [ 2 , 3 ] as reliably as reported in a previous study [ 11 ], whereas conventional biomarkers for IPF showed no or weaker correlations. Moreover, morphologic fibroproliferative findings on thin-section CT, such as reticulation and honeycombing, were also associated with monomeric or total periostin. The producing cells and the mechanism of the release of periostin or SP-D/KL-6 are likely different. Both SP-D and KL-6 are produced in regenerating alveolar type II cells; upon epithelial damage, these proteins may leak into the bloodstream [ 31 , 32 ]. In contrast, myofibroblasts are likely to be the main periostin-producing cells in pulmonary fibrosis [ 10 , 11 ]. Periostin expression is observed in fibroblastic foci, the sites where fibrosis is progressing [ 10 ]. These findings suggest that periostin expression would more directly reflect excess proliferation of fibroblasts or active status of fibrosis, which would lead to a decline in pulmonary functions.
Thus far, it is unknown whether oligomeric and monomeric forms of periostin have different functions. It has been reported that some specific organs such as bones or periodontal ligament [ 33 , 34 ], or some cancer cells [ 35 , 36 ], show different expression profiles of periostin compared to other tissues or to normal cells. Moreover, the splicing variant deleting exons 17, 18, and 21, called periodontal ligament-specific periostin or the lung type of periostin, showed enhanced binding activity to α V β 3 integrin [ 34 ]. It would be possible that partner proteins in the oligomeric form of periostin may either enhance or inhibit the periostin functions. It would be of great interest to investigate whether oligomeric and monomeric forms of periostin have different functions.
There are three limitations in this study. First, in this study, we analyzed the correlation between the biomarkers and short-term changes in lung function. Although it has been shown that short-term changes in lung functions predict survival well in IPF patients [ 2 , 3 ], whether monomeric periostin can predict survival as well as short-term changes in lung function needs to be addressed. Second, although in this study we increased the number of investigated subjects to 60 compared to our previous study [ 11 ], we still need to analyze a larger number of subjects in order to obtain results that are more statistically robust. Third, we did not assess sufficient numbers of other ILD patients to conclude whether the diagnostic and prognostic properties of monomeric periostin are specific for IPF. To do so, we are extending and expanding the CoDD-PF study and the Kurume study. Moreover, it is of interest to analyze whether monomeric periostin can predict the efficacy of two recently approved anti-IPF agents, pirfenidone and nintedanib [ 37 , 38 ]. Although it is known that these agents are effective for some IPF patients, thus far there is no suitable biomarker to predict of efficacy of these agents. A so-called companion diagnostic for these agents would enable us to develop stratified medicine for IPF.
Conclusions
We have developed a new periostin kit specifically to detect monomeric periostin. In a multicenter study, monomeric periostin was shown to be superior in diagnosing IPF compared to total periostin. Both monomeric and total periostin can predict short-term IPF progression better than conventional biomarkers such as KL-6, SP-D, and LDH. Moreover, the ratio of monomeric periostin to total periostin was higher in IPF compared to other periostin-high diseases. These results underscore the usefulness of measuring monomeric periostin for managing IPF.
Supporting information
S1 Fig
Preparation of total and monomeric periostin.
Recombinant total periostin purified by SS18A (A) and monomeric periostin purified with SS19D (B) in non-reduced (left panel) and reduced (right panel) conditions shown in protein staining are depicted. BSA are loaded on C.
(EPS)
S2 Fig
Standard curve of the conventional periostin ELISA kit and the new periostin ELISA kit for the purified periostin protein (oligomer and monomer) and the purified monomeric periostin protein.
Closed and open circles represent total and monomeric periostin, respectively.
(EPS)
S3 Fig
Distribution of monomeric periostin in healthy controls.
Distribution of monomeric periostin detected by the new periostin ELISA kit in healthy controls (n = 137, black boxes) and IPF patients (n = 60, white boxes) is depicted.
(EPS)
S4 Fig
Abilities of each biomarker to predict the short-term progression of IPF in the CoDD-PF study.
Correlations between monomeric periostin, total periostin, KL-6, SP-D or LDH and the short-term change of %VC (A) or % D L, CO (B) in IPF patients (n = 29 for %VC and 15 for % D L, CO ). NS : not significant.
(EPS)
S5 Fig
Abilities of each biomarker to predict the short-term progression of IPF in the Kurume study.
Correlations between monomeric periostin, total periostin, KL-6, SP-D or LDH and the short-term change of %VC (A) or % D L, CO (B) in IPF patients (n = 25 for %VC and 14 for % D L, CO ). NS : not significant.
(EPS)
S1 Table
Characteristics of the subjects.
(DOC)
S2 Table
CT findings.
(DOCX)
S3 Table
Correlation between baseline %VC or % D L, CO and various parameters.
(DOCX)
S4 Table
Correlations between the biomarkers and the CT scores.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
With a concentration as high as 15–25 mM, Mg 2+ ions take part in a diverse biological functions within living cells [1] . In prokaryotes, Mg 2+ ion has been linked to the virulence as an essential regulatory signal. In eukaryotes, Mg 2+ ion has also been shown to influence the DNA and protein synthesis [2] , [3] . Three types of transporters, MgtE, MgtA/B and CorA, have been identified to own the ability to mediate transport of Mg ions across bacterial membrane [4] . Among them, CorA has been studied the most. CorA was first identified from Escherichia coli genome by Silver and colleagues in 1969 [5] and first cloned from Salmonella typhimuriom by Hmiel and colleagues in 1986 [6] . Nevertheless, the crystal structure of CorA remained unsolved until 2006, when three individual groups published the crystal structure of divalent ions bound Thermotoga maritima CorA [7] , [8] , [9] . All the three structures clearly show that the functional form of CorA protein is a funnel-like homopentamer. For each monomer, both N- and C-terminals are located at the cytosol side. The N-terminal cytosolic domain forms a “sandwiched structure” in which 7 beta-sheets locate in between 2 series of alpha-helices (helices 1–3 and helices 4–6). Helix 7, the longest helix in CorA, starts from the cytosolic sandwiched structure, includes the first transmembrane domain (TM1), and ends at the periplasmic side. Helix 8 forms the second transmembrane domain (TM2), and brings the C-terminal end back into cytosol. In pentamer, the channel is surrounded by the five TM1s, and the five TM2s form a ring encircling the channel.
The structural information of the short interhelical loop linking TM1 and TM2 was missing in all the three solved structures, probably due to its high flexibility [7] , [8] , [9] . The interhelical loop contains the signature motif “GMN” of CorA, and another highly conserved motif “MPEL” in most members of CorA family. Besides, several charged residues are present in the loop [10] . As the loop is exposed to periplasm, it was believed to be essential in initial binding of ions, and possibly substrate selection [7] . Moomaw and Maguire recently applied mutational study on the loop region of Salmonella enterica serovar Typhimurium CorA and predicted that the interhelical loops provided initial binding site for hydrated Mg 2+ ion rather than the dehydrated one [11] . They also proposed that the electrostatic interactions between ions and the negatively charged residues were not essential. On the contrary, Hu et al . compared binding of Mg 2+ ions with a well-known CorA inhibitor Co(III)Hexamine (HexCo) to Mycobacterium tuberculosis CorA and concluded that the negatively charged residues in the loop region play important roles in cations recognition [12] . Dalmas et al . combined ERP and molecular modelling tools to generate a model for the interhelical loop of T. maritima CorA, and found that the negatively charged E316 formed a “negatively charged nest” which fits nicely to a hydrated Mg 2+ ion [13] . The explicit interactions between the loops and ions, as well as the roles of negatively charged residues in ions binding are, however, still not clear yet.
In this study, we have made efforts to sample the configuration of the interhelical loops of T. maritima CorA using extensive replica-exchange molecular dynamics (REMD) simulations. Meanwhile the binding interactions of loops and Mg 2+ ions as well as HexCo ions have been explicitly explored. With the theoretical methods applied, we try to explain the roles of residues in ions binding. Moreover, by comparing the HexCo and Mg 2+ ions, we also aim to explain the inhibition mechanism of HexCo on CorA theoretically.
Model and Methods
Loop model
The monomer model was built based on homolog modelling method by Modeller software (version: 9v6) [14] . The reference structure used is the Chain A of the crystal structure of CorA solved by Eshaghi et al. (PDB code: 2IUB) with the missing residues from Y311 to G326 (sequence: YGMNFEYMPELRWKWG). Modeller constructed the coordinates of the missing residues according to de novo loop modelling technique. The resultant monomer model with twenty-eight residues, F306 to V333, was denoted as chain A in loop model. Chain A was then rotated and translated symmetrically to get Chain B, C, D and E which together make up the homo-pentameric loop model. The modelling process is shown in Figure 1 schematically. The additional first 6 and last 6 residues, a part of transmembrane helix 1 (TM1) and 2 (TM2), respectively (10), were also included in the model.
10.1371/journal.pone.0043872.g001
Figure 1
Demonstration of modelling process.
The sequence of monomer model, from F306 to V334 is shown. The residues which are missing in 2IUB are shown in red.
Simulation methods
All molecular dynamics simulations were performed with GROMACS software package version 4.0.3 [15] using the AMBER99SB force field [16] . To enhance sampling, replica exchange molecular dynamics simulations were implemented [17] , [18] . Three systems were built up with x-axis parallel to the axis of channel. System 1 is the loop model explicitly solved with 4074 SPC water molecules [19] in a dodecahedron box with an initial volume of 149.746 nm 3 (abbreviated as LOOP system for short). System 2 is the loop model and 10 HexCo ions explicitly solved with 3992 SPC water molecules in a dodecahedron box with an initial volume of 152.97 nm 3 (abbreviated as COH system for short). System 3 is the loop model and 10 hydrated Mg 2+ ions explicitly solvated with 4093 SPC water in a dodecahedron box with an initial volume of 156.66 nm 3 (abbreviated as MG system for short). 5, 35 and 25 chloride ions were added in LOOP, COH system and MG system respectively to make the systems neutral. The initial topologies of the HexCo were generated in Spartan and energy-minimized [20] . The conformations were further optimized at the HF/6-31G* level using Gaussian09 [21] , and the atomic partial charges were derived using R.E.D III package [22] . The parameters for HexCo used in this study, shown in Table S1 , is the same with a published work by Cheatham and Kollman in 1997 [23] . To validate the parameters assigned, the RDF of water-HexCo/Mg 2+ ion and the PMF of the distances between HexCo/Mg 2+ and glutamic side chain were calculated. The detail of the calculation is provided in Supporting Information. ( Section S1 and S2 , Figure S1 and S2 ). For Mg 2+ ion and chloride ions the CHARMM parameters optimized by Roux and co-workers were used [24] . The combination rule of CHARMM and AMBER force fields for the non-bonded interaction is the same. All 3 systems containing 48 replicas were simulated for 200 ns with the temperature range from 315.0 K to 455.8 K. The average replica exchange ratios are ∼39.1%, ∼37.9% and ∼36.4% for the LOOP, COH and MG systems, respectively. The heavy atoms from residues of TM1 and TM2 were constrained to mimic the effect of membrane at all simulation stages. The rest residues in the loop domain, G312 to Y327, were left to move freely. LINCS protocol [25] was applied to constrain the length of all bonds including hydrogen atoms. The integration step in the simulation is 0.002 ps, and the trajectory was output every 500 integration steps. The replica exchange was attempted every 1000 integration steps, 2ps. Coulomb interactions were treated with the particle mesh Ewald method [26] with the cutoff of 0.9 nm. The van der Waals interactions were treated with the cutoff of 1.2 nm.
Analysis methods
Trajectories from all simulations were analyzed using GROMACS software package. The commands g_mindist and g_dist were used to calculate minimum distances and inter-centre of mass distances respectively. The command g_gyrate was used to calculate the radius of gyration. The command g_traj was used to output the coordinates of certain groups of atoms. The command g_sas was used to calculate the solvent accessible surface areas (SASA). The command trjorder was used to calculate the number of water molecules around a certain group. Pymol program was used for visualization. Potential of mean force (pmf) was applied to construct the free energy surface in both 1-dimension and 2-dimension. The free energy is given by (1) Where, w(r) is the Helmholtz free energy, k B is the Boltzmann constant, T is the temperature in Kelvin, g(r) is the sample counts at distance r, and gmax is the maximum sample count.
The binding energies were analyzed by the generalized Born model with surface area modification (GBSA) using sander module in the AMBER9 package [27] . The force field used were the same as the one used for the REMD simulations. The GB model developed by Onufriev et al , [28] were applied here, and the α, β, and γ were 1.0, 0.8, and 4.85 respectively. Each structure of the REMD simulations produces three copies with all water molecules discarded (the water molecules bound to Mg 2+ ions were kept): one is the whole system, one with only ions and one with only protein. All of them were inputted into sander module and related energies were output. The energy differences between the system and the sum of separated ions and protein were taken as the binding energy.
Results
Convergence check
REMD simulations enhance the sampling through overcoming potential energy barrier at high temperatures. It is critical to check whether each replica has visited all the temperature ladders several times during the whole trajectory. Figure 2A shows the temperature evolution of one representative replica trajectory over the 200 ns simulation. It is clear that the replica visited a wide range of temperatures. The similar behaviour was also found on other replicas, of which the data are not shown. Another reason to apply REMD method in our simulation is to avoid trapping of bound ions to the protein due to the strong electrostatic interactions. To monitor it, the minimum distances between ions and protein were traced within each single replica. Figures 2B and 2C show the changes of minimum distances between one of the 10 HexCo ions and the protein, and between one Mg 2+ ion and the protein within the 200 ns trajectories. One observes that the particular Mg 2+ /HexCo ion keeps moving rather than being trapped in binding to the protein. Similar behaviour was also observed on the other ions, of which the results are not shown.
10.1371/journal.pone.0043872.g002
Figure 2
The convergence check.
Trace in temperature space of representative replicas (A); Evolution of distances between HexCo ions and protein heavy atoms for representative replicas (B); Evolution of distances between Mg 2+ ions and protein heavy atoms for representative replicas (C); Distribution of radius of gyration of protein with time interval 50–100 ns (black square), 50–150 ns (red circle), 50–180 ns (violet triangle), and 50–200 ns (green inverse triangle) at 315 K for LOOP (D), COH (E) and MG (F). Distribution of distances between ions and protein heavy atoms with time interval 50–100 ns (black square), 50–150 ns (red circle), 50–180 ns (violet triangle), and 50–200 ns (green inverse triangle) at 315 K for COH (G) and MG (H) system.
In our systems, the transmembrane domains of the protein were frozen and only the loop domains were allowed to move freely. To depict the flexibility of the loops, a global structural character, the radius of gyration (Rg), of the loop domains was employed. Thus, the convergence of the protein structures was checked through the evolution of distributions of Rg of the protein at the lowest temperature (315 K). Four Rg distribution functions sampled from LOOP, COH and MG systems over different time intervals were plotted in Figure 2D, 2E and 2F , respectively. For the LOOP system, the distribution curves of Rg over different time intervals, 50–100 ns, 50–150 ns, 50–180 ns, 50–200 ns, are plotted in Figure 2D . Among them, the curves for 50–180 ns, 50–200 ns are well overlapped, in terms of the heights and positions of peaks. For the COH system, the distributions of Rg over different time intervals, 50–100 ns, 50–150 ns, 50–180 ns and 50–200 ns, are sharing the similar pattern, among which the curves for 50–180 ns and 50–200 ns are very well overlapped ( Figure 2E ). For the MG system, the distribution curve for the sampling of 50–100 ns is slightly different from those of 50–150 ns, 50–180 ns and 50–200 ns, with an additional peak at Rg = 1.7 nm. All the three curves sampled over longer time are nearly identical, particularly for the position of peaks ( Figure 2F ). The evolution of distributions of Rg indicates that the simulations have converged through the 200 ns simulation time in the context of structure of the loops.
One of the major goals in our study is to interrogate the interactions between ions and the protein. Therefore, the convergence was also tested by the distribution of the minimum distance between ions and protein heavy atoms in different time interval from the trajectory at the lowest temperature (315 K). Four distribution curves of the minimum distance between ions and protein heavy atoms of the trajectories at 315 K over different time intervals were plotted in Figures 2G and 2H . For the COH system, the distributions of the minimum distances over 50–100 ns, 50–150 ns, 50–180 ns, and 50–200 ns are showing similar patterns ( Figure 2G ). Particularly, the curves for 50–150 ns, 50–180 ns and 50–200 ns are nearly identical, which indicates the simulation is well converged within 200 ns. For the MG system, the distributions of the minimum distances within50–100 ns, 50–150 ns, 50–180 ns and 50–200 ns, are plotted in Figure 2H . Same as the COH system, high similarity among the four curves are observed. Again, this shows the MG system is well converged. One phenomenon should be noted that in contrary to the MG system, the curves of 50–150 ns, 50–180 ns and 50–200 ns of the COH system are almost coincident with each other, indicating that COH system converged faster than MG system. Additionally, the long tails appeared in the MG system are lost in the COH system. This indicates that HexCo ions are closer to protein than Mg 2+ ions. In brief, the simulation systems are well converged within 200 ns which enables further analysis.
Distribution of the radius of gyration of the loops
Comparing the Rg distribution curves for 50–200 ns of LOOP, COH and MG systems plotted in Figures 2D, 2E, and 2F , one observes that the loop domains in LOOP system have 3 peaks at 1.52 nm, 1.57 nm and 1.64 nm. While in COH system and MG system, there is only one dominant peak in each system. Moreover, the loop domains in the COH system have a narrower distribution with a smaller Rg than that of the MG system. Therefore, the observation indicates that HexCo ions may be more efficient to limit the flexibility of the loops than Mg 2+ ions due to different binding affinities or binding patterns between HexCo and Mg ions, which will be discussed in later sections.
In Figures 3A, 3B, and 3C , three snapshots of the loop structure with Rg of 1.52 nm, 1.57 ns, and 1.64 nm respectively are shown. In Figure 3D , a snapshot of the loop structure with Rg of 1.56 nm (the peak position of the Rg distribution) in the COH system is shown. Five tightly bound HexCo ions are found there. In Figure 3E , a snapshot of loop structure with Rg of 1.6 nm is displayed, together with two closely contacted Mg 2+ ions. With smaller values of Rg, the loops seem be glued together by the ions (see Figure 3D ) and with large values of Rg, the loops tend to tilt away from the channel axis (see Figure 3E ). Generally speaking, there are very few regular secondary structures, such as α-helices and β-strands, formed in the loops (the results of secondary structure assignment using DSSP method are shown in Table S2 ).
10.1371/journal.pone.0043872.g003
Figure 3
Representative loop/loop-ions structures with distinct Rg values (A, B, C).
Conformational space of the loop structure sampled in different systems
To further investigate the conformational variability of the loop structures, the free energy surfaces constructed by PMF methods using two reaction coordinates (RC) are shown in Figures 4A (LOOP system), 4C (OCH system) and 4E (MG system). One RC (RC1) is the Rg of heavy atoms from five residues E320. The other RC (RC2) is the inter-centre of mass distances between residues F306 and E320 in X direction. The Rg of residues E320 provides a measure of the opening level of the channel entrance, since residues E320 are located right in the middle of the loop domain. The inter-centre of mass distance between residues F306 and E320 in X direction reflects the height the pentameric loop model from bottom (F306) to top (E320), because the systems were build with x-axis parallel to the axis of channel. One observes the conformational space in LOOP system ( Figure 4A ) is broader than that of COH ( Figure 4C ) and MG ( Figure 4E ) systems, which indicating the loops in LOOP system are more flexible, and HexCo/Mg 2+ ions function to limit the flexibility of the loops.
10.1371/journal.pone.0043872.g004
Figure 4
Free energy landscapes at 315 K for LOOP (A), COH (C) and MG (E) systems.
The reaction coordinates are Radius of gyration of heavy atoms from E320, and the inter-centre of mass distance between E320 and F306 in X-axis. Sketches of representative structures for LOOP (B), COH (D) and MG (F). Rounded rectangles represent the TM1 and TM2 domains. Black solid lines represent the loop domains. Black dotted lines represent the channel axis.
The conformations in COH system are mostly populated in the area where RC1 roughly equals to 1.75 nm and RC2 roughly equals to 2.0 nm. While, the conformations in MG system are populated in the area where RC1 roughly equals to 2.0 nm, and RC2 roughly equals to 1.6 nm. The results indicate in MG system, the structures were wider in channel entrance, and shorter in height, comparing with the structures in COH system ( Figures 4D and 4F ). In LOOP system, both types of structures were present, with different populations though ( Figure 4B ).
Distance between ions and the amine groups on loops
In a NMR study by Hu in 2009 [12] , the inter-helical loops of Mycobacterium tuberculosis CorA with the presence or absence of Co 2+ was studied using H/D exchange method. The results revealed the chemical shifts of amide hydrogen atoms of some residues are more sensitive to ions probably due to the nearby ions perturbing the local chemical environment. To check whether present simulations fit their results, the minimum distances between the Mg 2+ and Co atoms in HexCo to the Hydrogen atom of amine group from residues of the loops were calculated and plotted in Figure 5 . In our calculations, Mg 2+ and HexCo showed similar patterns. From Figure 5 , the residues from F315 to L321 have shorter distances than other residues in the interhelical loop, in both COH and MG systems. The results from Hu et al indicated that the Mg2+, Co2+ and HexCo have same binding sites on CorA loops and the most sensitive residues to Co2+ ions are F332 and M333, which corresponding to Y317 and M318 in Thermotoga maritima CorA used in the present study. Besides F332 and M333, other residues, from F330 to D337 (F315 to R322 in Thermotoga maritima CorA), are also sensitive to Co2+. The R322 here is not near to the cations because of its positive charge, while in the experiment the corresponding residue D337 is close to Co 2+ since it is negatively charged. The residue E316 and corresponding residue H331 in the experiment also show different behaviours, probably resulting from different charge states. Except R322 and E316, the distance profile of other residues from F315 to L321 agrees well with the experimental results of Hu et al .
10.1371/journal.pone.0043872.g005
Figure 5
The averaged minimum distances from hydrogen atoms form amine group of interhelical loop residues and the HexCo (black solid line) and Mg 2+ (black dotted line) ions.
Direction of E315 side chains
In their study Dalmas et al . developed an extracellular loop model of CorA by combining electron paramagnetic resonance (EPR) and theoretical modelling method [13] . One of the results was that the residues E316 was located towards the centre of the pentameric ring and together they form a “negatively charged nest” for hydrated Mg 2+ ions to bind. To check if our model fits Dalmas' results, we defined 2 vectors: One is the vector starting from the alpha carbon ending at the delta carbon of one residue E316; the other one is starting from the alpha carbon of the same E316 pointing to channel axis and is perpendicular to the channel axis. A schematic drawing of the 2 vectors is shown in Figure 6A . The angle α between the 2 vectors was used to define the orientation of the side chain of E316. If the cosine value of the angle α is larger than 0, the angle of the 2 vectors is an acute angle, indicating the side chain of E316 is pointing towards the channel axis. The distribution of the cosine values is shown in Figure 6B . The frames used in the plot are those in which there are Mg 2+ ions located less than 0.5 nm away from protein surface. The distribution shows a bimodal feature: the higher peak sits around 0.50 relating to an angle of 60 degree between the two vectors; the other peak is around the value of 0.15 corresponding to a side chain configuration with 81 degree of the side chain with respective to the channel axis. Clearly in the simulations the side chains of E316 are very flexible. The distribution highly populated around 0.5 indicates that the side chain of E316 is tilted to the axis of the channel instead of pointing away from the symmetry axis which is consistent with Dalmas' postulation.
10.1371/journal.pone.0043872.g006
Figure 6
Direction of E316 side chains.
Schematic drawing of the vectors from Cα to Cδ within GLU residue and from Cα of GLU to channel axis (A). The distribution of the value of cosine of the angle between the two vectors (B).
Dalmas et al . have also evaluated the solvent accessibility of CorA using fast relaxing agents, NiEdda and O 2 , soluble in water and lipid respectively. Their NiEdda accessibility curve shows two regions where high NiEdda accessibility is observed. The first region is from E316 to Y317, and the second is from L321 to W325, with an exception of a low value at R322. To compare with their results, we calculated the solvent accessible surface area (SASA) and the average number of nearby water molecules (NW) for each residue from G312 to Y327, and the results are shown in Figure 7A and 7B respectively. From the SASA results in Figure 7A one observes two distinct regions with high solvent accessible surface area: first is from F315 to Y317, and second is from E320 to W325. A similar trend can be found in the statistics of NW in Figure 7B . Our results are reasonable in line with Dalmas' experiment, considering the regions with high accessibility. However, the 3 charged residues, E316, E320 and R322, have higher values of SASA and NW value, but low NiEdda accessibility found in the experiment. This is possibly because they used the cysteine scanning technique, in which each residue was replaced by a cysteine and an underestimation of electrostatic interactions between the charged residues and water molecules may happen. Another reason could be that the large size of NiEdda may cause underestimation of the actual water accessibility near sterically hindered region. Generally speaking, our SASA and NW results reflect the trends of the accessibility profiles for the loop and are consistent with experiment done by Dalmas.
10.1371/journal.pone.0043872.g007
Figure 7
Solvent accessibility.
Averaged solvent accessible surface area for residues from G312 to Y327 (A). Ensemble averaged number of water molecules located within 0.5 nm shell from every residue in between G213 to Y327 (B). Both of the two sets of data were calculated from ensemble trajectory at 315 K.
Local binding pattern of ions
In order to gain insight into the binding pattern of ions on the loops, the averaged minimum distance from Co or Mg atoms to each residue at 315 K was calculated and plotted in Figure 8A . It is natural to think that the residues which were involved in binding with ions would have short distances to the ions. As shown in this figure, HexCo ions have an averaged minimum distance of 0.336 nm to E316 and 0.339 nm to E320, while for Mg 2+ ions the distances are 0.392 nm and 0.490 nm, respectively. The lowest average minimum distances for MG system, 0.392 nm and 0.490 nm here, are larger than the radius of equivalent hydrated sphere of Mg 2+ ions which is 0.3 nm [29] , indicating that in our system, the Mg 2+ ions bind to protein in the form of hydrated ions, which correlates well with Moomaw and Maguire's work in 2010 [11] . Additionally, the results also show E316 and E320 play important roles in ion binding. To further study the role of the two glutamic acids, the potential mean forces (PMF) as a function of distances from the Mg or Co atoms to the heavy atoms of the whole protein, of only E316 and of only E320 were calculated based on the equation 1 , and the results were plotted in Figure 8B, 8C and 8D respectively. The free energies calculated were then normalized at distance equals 2 nm, where the interactions in between could be ignored. For the COH system, the global minimum is located at 0.37 nm. The binding free energy of HexCo ions to the whole protein, only E316 and only E320 is estimated to be −16.8 kJ/mol, −8.2 kJ/mol and −9.4 kJ/mol, respectively. Clearly the binding of HexCo ions to the loop is dominantly contributed from the two residues, E316 and E320. For MG system, the global minimum is located at 0.39 nm. The binding free energy of Mg 2+ ions to the whole protein, only E316 and only E320 is estimated to be −9.5 kJ/mol, −4.7 kJ/mol and −4.4 kJ/mol, respectively. The binding of HexCo to the loops is stronger than Mg 2+ ions, with the binding free energy −7.3 kJ/mol lower.
10.1371/journal.pone.0043872.g008
Figure 8
Local binding pattern of the ions.
Averaged distances between ions and heavy atoms from residues of loop domain (M313 to Y327) for COH system (black square) and MG system (red circle) calculated from ensemble trajectory at 315 K (A). Potential of mean force as function of distance from ions to protein heavy atoms (B), to E316 heavy atoms (C) and to E320 heavy atoms (D) for COH system (black line) and MG system (red line).
The ability of HexCo ions to inhibit transportation of Mg 2+ ions by CorA has been studied by Kucharski et al . [30] . They found that the IC 50 values for Mg ions in the presence of 200 µM Ni 2+ ions are 1 µM for archeon M. jannaschii CorA, and 0.5 µM for S. typhimuriom CorA, while the values of HexCo are 200 µM for M.jannaschii CorA and 10 µM for S. typhimuriom CorA. Based on the Cheng-Prusoff equation [31] , , where K i is the binding affinity of the inhibitor, [s] is the substrate concentration, K m is the concentration of substrate at which enzyme activity is at half maximal, and , where ΔG is the binding free energy, R is the gas constant and T is the temperature, we can get the binding free energy difference, ΔΔG, between HexCo-CorA and Mg 2+ -CorA, is around −5.9 to −7.7 kJ/mol at T = 310 K (the experimental temperature). Suppose that the binding sites of HexCo and Mg 2+ ions are both on the extracellular loops, our simulation PMF data, ΔΔG = −7.3 kJ/mol, is consistent very well with the experimental measurements.
Moreover, the binding energy of ions and the loops were estimated with generalized-Born surface area (GBSA) model. The trajectory from 90 ns to 100 ns was split into 10 separated 10 ns-long trajectories where only 1 out of the 10 cations was present. The ten 10 ns-long trajectories were then concatenated into a new 100 ns trajectory, denoted as single-ion trajectory. Every frame of the new 100 ns single-ion trajectory was used for the GBSA analysis. The bound ions were then clustered according to the distances between them and E316 or E320. Each cluster is represented by two numbers, such as 1–2. The first number indicates the number of E316 residues in close contact with an ion, and the second number denotes the number of E320 residues which are close to the same ion. The clustering results were shown in Table 1 . From the table, one can observe that the binding energy is mostly contributed by the electrostatic interactions. The numerical results also show that HexCo has a lower binding energy than Mg 2+ when interacting with a same number of glutamic acid residues, which is in consistent with the previous PMF results. Further looking into the table, the 1-0 and 2-0 clusters have lower binding energy than 0–1 and 0–2 clusters in both COH and MG systems. This indicates that the E316 has a higher binding affinity than E320 for both HexCo and Mg 2+ ions. The last column of Table 1 shows the population of each cluster. The first three most populated clusters for COH system is 0-1, 1-1 and 2-1, while for MG system is 1-0, 0-1 and 2-0. Snapshots to demonstrate the six different binding patterns were extracted and listed in Figure 9 . Those structures were selected through two steps. First, the frames that belong to a particular cluster were extracted and their structures were fitted using protein heavy atoms. Second, a single linkage clustering for the pre-fitted structures was performed using the RMSD of the ion and the cluster centre was taken as the representative structure. E316 and E320 residues which were located within 0.5 nm from the ion are shown in sticks.
10.1371/journal.pone.0043872.g009
Figure 9
Snapshots taken from last 10 ns of ensemble trajectory at 315 K.
The snapshots demonstrate the different patterns of interactions between cation and protein which are labelled in the top left corner of each snapshot. Cations and their interacting residues are highlighted. Upper panel shows the snapshots taken from COH system, and lower panel shows snapshots taken from MG system.
10.1371/journal.pone.0043872.t001 Table 1
Binding energy calculated based on implicit GBSA model.
Cluster
a
G binding
b
E GB
E vdw
E ele
Population c (%)
COH system
0–1
−83.48±74.10
99.68
1.30
−184.46
35.5
0–2
−158.50±56.79
170.27
3.93
−332.70
3.3
1–0
−193.40±77.09
250.30
1.23
−444.93
7.8
1–1
−237.54±65.33
274.47
2.39
−514.40
16.7
1–2
−327.63±28.01
351.76
3.98
−683.37
0.4
2–0
−259.10±77.63
303.92
3.28
−566.30
9.6
2–1
−335.51±82.06
355.49
4.35
−695.35
13.4
MG system
0–1
−33.44±29.78
40.86
0.67
−74.97
29.0
0–2
−68.15±49.95
165.29
−0.93
−232.51
0.5
1–0
−84.23±35.83
194.63
−3.11
−275.75
36.4
1–1
−127.13±36.39
263.33
−2.18
−388.28
11.4
1–2
−160.61±29.81
311.09
−4.30
−467.40
0.1
2–0
−108.07±38.07
236.18
−1.73
−342.52
15.2
2–1
−157.20±33.99
305.04
−0.71
−461.53
4.4
a
Cluster x – y means the number of E316 residues which were located less than 0.5 nm from ions is x, and the number of E320 residues which were located less than 0.5 nm from ions is y.
b
the unit for energy terms is kJ/mol.
c
Population is the percentage of a cluster in the population where at least 1 E316 or E320 was located less than 0.5 nm from ions.
Global binding pattern of ions
It is natural to ask how many ions can bind to the protein simultaneously since the protein is pentameric. To check it, the minimum distances from each Co atom of HexCo or Mg 2+ atom to every residue of the protein were calculated from the ensemble trajectories at 315 K per frame. A residue was defined as an interacting-residue when the distance between the residue and the ion was less than 0.5 nm. And bound state was defined when there was at least one interacting-residue at the particular frame. The results are shown in Table 2 column 2. Surprisingly, there are almost 5 HexCo ions binding with protein in each frame, while only 3 for Mg 2+ ions. The results indicate that ions can bind to the loops through several different binding sites.
10.1371/journal.pone.0043872.t002 Table 2
Bound states clustered based on distances between every ion to every residue.
System
Average Number of Ions Bound
Populated Bound States
Residues involved in Binding
a
b
COH
5.15±1.007
5 (42.1%)
E320(A), E320(B), E316(B)/E320(C), E316(C)/E320(E), E316(A)/E320(D)/E316(E)
6 (26.4%)
E316(A), E320(B), E320(E), E320(A)/E320(D), E316(D)/E316(E), E316(A)/E316(D)/E320(D)/E316(E)
4 (21.7%)
E316(A), E320(B), E320(A)/E320(D), E316(B)/E316(C)/E320(C)
MG
2.89±1.155
3 (39.3%)
E316(A), E316(C), E316(A)/F315(D)/E316(D)
2 (27.2%)
E320(B), E316(D)
4 (20.8%)
E320(B), E320(C), E320(E), E316(A)/E316(C)
a
The letter inside brackets indicates the chain name for the residue.
b
The comma “,” separates different bound states, the slash “/” separates different residues in a particular bound state.
The residues involved in the multiple binding sites were then investigated based on the distances calculated. First, the bound states were further clustered according to the number of ions bound per frame. Then the first 3 most populated bound states were extracted to find out the most common interacting-residues. The results are shown in Table 2 columns 3, and 4. Consistent with the results of average minimum distances from ion to residues, the most common interacting residues are E316 and E320 for both COH and MG systems. In COH system, other residues were barely seen to be interacting residues. In MG system, F315 was seen when number of bound ions was 3, suggesting its role in the binding as a cooperative binding site.
To check where the multiple binding sites were located, we investigated the Cartesian coordinates of the ions. The coordinates of every frame were extracted from ensemble trajectory at 315 K. Since the system was built in the way that the molecular channel-axis was parallel to the X-axis of system, the Y and Z coordinates were chosen first to show the general distributions of ions. The Cartesian coordinates of the Cα atoms from Y311 and P328, 10 atoms in total, were also extracted. These atoms are constrained during the whole simulations and their coordinates were plotted here to directly show the position of the channel ( Figures 10A and 10B ). The 2-dimensional PMF methods were applied, and the free energy surface was shown in Figure 10 . The standard errors for Figure 10A and 10B are both in the scale of 10 −5 kJ/mol. And for Figure 10C and 10D , they are 0.03421 kJ/mol and 0.02101 kJ/mol respectively. From the Figure 10A , one observes the channel of COH system is mainly located in the region where y = 4.0–5.0 and z = 1.8–2.8. As to the HexCo ions, in Figure 10C , one finds the most highly sampled regions is located mainly in the channel region. As to the MG system, the channel region is around y = 3.0–4.0 and z = 2.4–3.4 ( Figure 10B ), where Mg 2+ ions are most populated. From the above results, one can see that the HexCo and Mg 2+ ions globally prefer binding to the loops located on the top of the inner circle formed by the TM1 domains, rather than those outside the outer circle formed by TM2, or in between the two domains. In comparison with the HexCo ions, Mg 2+ ions sample much broader space on the Y-Z dimensions.
10.1371/journal.pone.0043872.g010
Figure 10
The 2-dimensional PMF on Y and Z Cartesian coordinates of different atoms.
The free energy of the alpha carbon atoms of Y311 and P328 from COH system (A) and from MG system (B); The free energy of Cobalt atoms from COH system (C) and Mg atoms from MG system(D). The units of X-axis and Y axis are nm. The unit for free energy is kJ/mol.
In addition to the 2D-PMF on Y and Z dimensions, the X-coordinates were also investigated. The X-coordinates of the alpha carbon of G312 from both systems, Co atoms from the COH system and Mg 2+ atoms from the MG system were extracted from the ensemble trajectory at 315 K. The distribution of these coordinates were calculated and plotted in Figure 11 . The alpha carbon atoms of G312 residues were used to indicate the position of the channel mouth along the X-axis. The HexCo ions and Mg 2+ ions show different distribution patterns along the X-axis as shown clearly in Figure 11 . Most of HexCo ions were located closer to channel mouth than Mg ions. Conclusively, both HexCo and Mg 2+ ions bind to loops on the top of the channel, and HexCo ions bind to loops closer to the channel mouth than Mg ions.
10.1371/journal.pone.0043872.g011
Figure 11
Distribution of X-coordinates of different atoms extracted from ensemble trajectory at 315 K.
Black line represents the distribution of X-coordinates of alpha carbons of G312 from both systems. Blue line represents the distribution of X-coordinates of Co atoms from COH system. Red line represents the distribution of X-coordinates of Mg atoms from MG system.
Discussion and Summary
In this study, we explored the configuration of the flexible inter-helical loops of T. maritima CorA, and investigated the binding patterns of Mg 2+ ions and HexCo ions to the loops which are characterized by the averaged minimum distances, the PMFs and the average number of bound ions.
Based on our study both E316 and E320 are the main players for Mg 2+ and HexCo ions binding and on the whole loop there are multiple binding sites available. The side chains of E316 residues are pointing toward the central axis forming a “negatively charged nest” to ions, as suggested by Dalmas et al. (13). The E320 residues also form high amount of contacts with Mg 2+ ions.
The inhibition mechanism of HexCo ion on the transportation of Mg 2+ ions can be soundly explained. First of all, HexCo ions own a higher binding affinity than Mg 2+ ions, which gives them a priority to bind the periplasmic loops over Mg 2+ ions. Secondly, the number of HexCo ions binding on the loops simultaneously is as many as 5 in our model. The 5 ions could expel the Mg 2+ ions away from the loops through electrostatic repulsion, resulting in incapable binding of Mg 2+ ions. In addition, both ions prefer binding to the loops on the top of the channel with HexCo ions preferring binding closer to the mouth of the channel than Mg 2+ ions. We realize that the absence of explicit lipid bilayer environment in our simulations neglects the lipid-water interface and the low-dielectric medium inside the lipid bilayer which may influence the ions distributions nearby. Fortunately the loop domains which are the focus of the current study are well above the lipid-water interface. The ions distribution within the loop domains may experience little perturbations from the lipid-water interface.
In a recent experiment, Moonmaw and Maguire mutated E281 and E285 on Salmonella enterica serovar Typhimurium CorA separately (relating to E316 and E320 of T. maritima CorA) and found the overall transportation efficiency was not affected too much [11] . Our simulation study predicted that mutated negatively charged residues will reduce the ions binding affinity. Thus the above experimental data provide a hint that the bottleneck of the overall transportation process may not come from the initial binding events but may resort to the downstream ions translocation through the channel.
Unfortunately, the simulation of the whole process of Mg 2+ ions transportation through CorA is beyond the scope of the current study. We catch here the structural and energetic properties relating the initial binding of ions on the flexible loops. The one-dimensional free energy landscapes of ion binding ( Figure 8C and 8D ) are clearly funnel-like. Statistically ions move toward the mouth of the channel under a favourable environment which is constructed by the flexible loops and the aqueous medium.
A very recent work by Xia et al. [32] found that T. maritima CorA was not able to regulate the homeostasis of Mg ions. On the contrary, it could transport divalent cobalt ions with a high selectivity. Moreover they found that divalent cobalt ions could induce significant conformational changes of the protein during the transportation process, while Mg 2+ cannot. Considering the same charge state and similar ionic radii (65 pm for Co 2+ and 72 pm for Mg 2+ ) [33] , the initial binding patterns of Co 2+ on the loops should be similar to that of Mg 2+ . However, the delicate energetic differences of hydration structures between Co 2+ and Mg 2+ ions may bring about the transportation selectivity of the CorA channel.
In summary we studied the initial binding of ions to the inter-helical loops of CorA. As to other possible functions of the loops, such as the dehydration of ions, transportation of ions from the binding sites into the channel and the induction of the open/close state transitions during transportation, further experimental and theoretical studies are needed.
Supporting Information
Section S1
The radial distribution function of HexCo/Mg2+ ion - Oxygen in Water molecules.
(DOC)
Section S2
The PMF of the distances between HexCo/Mg2+ ion and the Glutamic acid side chain.
(DOC)
Figure S1
The RDF of Co atom - water Oxygen atoms (black) and Mg 2+ atom– water Oxygen atoms (red).
(TIF)
Figure S2
The PMF of the distance between HexCo (black) or Mg 2+ (red) and the glutamic acid side chain.
(TIF)
Table S1
The parameters for HexCo.
(DOC)
Table S2
Secondary structures composition calculated by DSSP at 315 K.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Antibiotics and antiviral drugs have achieved great success in recent history [ 1 ]. However, therapeutic failure may occur due to low adherence and the emergence of drug resistance [ 2 , 3 ]. The increasing amount of drug resistant pathogens is a global threat to public health [ 4 – 11 ]. The genetic barrier to drug resistance, defined as the number of mutations needed to acquire resistance, is a major determining factor of treatment outcomes [ 12 – 14 ]. Another important but often overlooked aspect of drug resistance is the fitness barrier [ 15 – 17 ]. Resistance associated mutations (RAMs) in pathogen proteins may decrease enzymatic activities, interfere with molecular interactions, or destabilize the protein structure [ 18 – 22 ]. Because of the impaired replication capacity without drug selection, drug-resistant mutants cannot normally outcompete wild-type or establish in the population [ 23 – 25 ]. However, drug-resistant mutants can sometimes reach substantial frequency in the population. Fluctuating drug concentrations may create time windows when drug-resistant mutants replicate better than wild-type virus [ 26 ]. Moreover, compensatory mutations can rescue the impaired replication capacity of mutants and stabilize drug resistance [ 22 , 27 , 27 – 29 ]. Thus, comprehensive quantification of the fitness landscape is needed to predict the evolution of drug resistance [ 30 , 31 ].
Epistasis, i.e. genetic interactions between mutations, is prevalent in molecular evolution [ 30 – 34 ]. Negative epistasis decreases fitness of the double mutant, posing constraints on gaining multiple mutations [ 35 , 36 ]. It plays an important role in shaping the local fitness landscape [ 37 ]. Positive epistasis increases replication capacity of the double mutant, facilitating pathogens to acquire and maintain drug resistance [ 38 – 40 ]. Positive epistasis may create a fitness valley that prevents drug resistant mutations from reversal [ 41 ]. Collectively, positive and negative epistasis determine the topography of the fitness landscape [ 42 ] and the course of drug resistance evolution [ 32 ]. Empirical studies on the genetic interactions between RAMs, especially in high-order mutants, are still rare [ 43 , 44 ].
HIV-1 protease inhibitors are important components of combination antiretroviral therapy [ 45 ] that target HIV-1 protease enzymatic activity [ 46 , 47 ]. Second-generation protease inhibitors have extremely high binding affinity to viral protein [ 48 ]. Resistance to them typically requires more mutations than resistance to first-generation protease inhibitors and other antiretroviral drugs [ 49 , 50 ]. For example, mutation K103N on reverse transcriptase is sufficient to confer HIV-1 nevirapine (NVP) resistance [ 51 ], while more than 4 de novo mutations are needed for protease inhibitor Darunavir (DRV) resistance [ 52 ]. Protease inhibitor-resistant viruses with multiple RAMs also have significantly reduced fitness [ 53 , 54 ]. HIV-1 gained RAMs on protease during sub-optimal protease inhibitor therapy [ 55 ]. Most resistance mutations directly affect the binding affinity between HIV-1 protease and the inhibitor, but they are likely to be deleterious because they also reduce binding to the native substrate of HIV-1 protease. To compensate the deleterious effect, some other RAMs stabilize HIV-1 protease, allowing drug-resistant virus to replicate as efficiently as its parental wild-type virus [ 27 , 56 ]. The compensatory effects between pairs of RAMs have been studied in several studies and are available on the Stanford HIV drug resistance database [ 22 , 57 – 61 ]. Meanwhile, reversals of protease inhibitor resistance-associated mutations were rarely seen clinically, even when therapy was interrupted [ 62 ] or when mutant virus infected drug-naïve patients [ 63 , 64 ]. These observations indicate that epistasis may be important for the evolution of protease inhibitor resistance. Recent analyses of sequence co-variation in drug-targeted HIV Pol proteins (protease, reverse transcriptase and integrase) and co-evolutionary Potts model provide evidence that epistasis plays an important role in drug resistance. Despite being disfavored in the wild-type background, primary resistance mutations can become entrenched by the complex mutation patterns which arise in response to drug therapy [ 65 , 66 ].
Here, we present a quantitative high-throughput genetics approach [ 67 , 68 ] to study the fitness distribution and epistasis of HIV-1 protease inhibitor RAMs. Combining these data with clinical data and fitness models, we found that positive epistasis was predominant and especially enriched among RAMs, and prevalent along drug resistance evolutionary paths. Our results suggest that fitness hills created by epistasis result in barriers that entrench RAMs, and thus drug-resistant viruses are unlikely to revert after transmission to drug-naïve patients or discontinuation of anti-retroviral drug treatment.
Results
Fitness profiling of RAMs in HIV protease
To study the interactions among RAMs in HIV protease, we constructed a library of virus mutants that covers combinations of amino acid substitutions at 11 resistance-associated sites in HIV protease ( Fig 1A , Table 1 , 2 9 × 3 2 = 4608 genotypes). To ensure sufficient coverage, we harvested more than 30000 colonies after transforming E. coli . These sites have been annotated as major drug resistance sites in Stanford Drug Resistance Database [ 61 , 69 ], and all have been shown to be strongly associated with drug resistance [ 3 ]. In our mutant library, 9 sites have one amino acid substitution and the other 2 sites have 2 amino acid substitutions ( Fig 1A , Table 1 ). 2736 out of 4608 possible genotypes (59.38%) were covered in the plasmid library.
10.1371/journal.pgen.1009009.g001
Fig 1
High-throughput fitness profiling of combinatorial HIV-1 protease mutant library.
(A) The structure of protease dimer (PDB: 4LL3). The side chains of selected resistance associated residues are shown. (B) Workflow of the fitness profiling. Protease mutations were introduced into NL4-3 background. T cells were infected by the mutant virus library. The frequency of mutants before (input library) and after (output library) selection were deep sequenced. (C) The correlation of relative fitness between two biological replicates. Pearson correlation coefficient ( R ) is 0.80. (D) Two independent validation experiments were performed. We constructed 7 protease single mutant plasmids and recovered viruses independently. We mixed each mutant virus with wild-type virus (validation 1, black dots) and passaged in T cells for 6 days. We also mixed all 7 mutant viruses together with wild-type (validation 2, red dots) and infected T cells for 6 days. The relative fitness of each mutant was quantified by the same means as that in the library. Pearson correlation coefficients ( R ) for validation 1 and validation 2 are both 0.84. Error bar is standard deviation ( n = 3). (E) The correlation of relative fitness in this study with the experimental selection coefficients in [ 71 ]. Pearson correlation coefficients ( R ) is 0.79.
10.1371/journal.pgen.1009009.t001
Table 1
List of protease inhibitor resistance associated mutations covered in the library.
a From 148840 subtype B protease sequences in Los Alamos Database [ 70 ]. b From 1951 isolates tested in PhenoSense assay [ 61 ].
Residue number
Consensus
Mutation
Prevalence in clinical dataset a
Occurrence in in vitro dataset b
10
L
F
1.54%
10.20%
32
V
I
1.37%
7.53%
46
M
I
4.32%
22.19%
47
I
V
0.88%
4.36%
50
I
V
0.30%
1.85%
54
I
L
0.68%
4.92%
54
I
M
0.48%
3.02%
74
T
P
0.37%
2.15%
76
L
V
0.46%
2.92%
82
V
T
0.64%
4.05%
82
V
F
0.33%
1.54%
84
I
V
3.00%
17.12%
90
L
M
7.71%
31.78%
We quantified the relative fitness of mutants using high-throughput fitness profiling ( Fig 1B , See Material and methods for details). We performed 3 independent transfection experiments to validate the reproducibility of fitness profiling. 20 million 293T cells were transfected and 50 million T cells were infected in each experiment. For each biological replicate, relative fitness was calculated independently. The Pearson’s correlation coefficients of single, double and triple mutations between replicates range from 0.80 to 0.82 ( Fig 1C and S1 Fig ). After filtering out mutants with low frequency or low reproducibility among replicates of input virus libraries (see Material and methods for details), we were able to estimate the relative fitness of 1219 genotypes. The fitnesses of all single mutants, and more than 70% of double and triple mutants, were quantified ( S2 Fig ).
To validate the quantification of relative fitness, we conducted competition experiments with individually constructed protease mutants. We performed two sets of validation experiments. For the first set, we packaged the mutant virus and wild-type virus independently and mixed them in pairs for head-to-head competition. The frequency of the mutant virus and wild-type virus were quantified by deep sequencing and the relative fitness was calculated in the same way as we did in library screening. A total of 7 mutants were constructed and validated. For the second set of experiments, we mixed all 7 single mutants with wild-type virus in competition experiments. The relative fitness was defined in the same way. The fitness measured in validation experiments was highly correlated with the fitness in library screening ( Fig 1D , R = 0.84 for each independent validation, Pearson’s correlation test). In addition, we compared the selection coefficients of HIV-1 protease mutants measured in an independent study by Boucher et al. [ 71 ] and the relative fitness values in our experiment ( Fig 1E , S2 Table ). The experimental results from two studies show a good correlation (Pearson’s correlation coefficient is 0.79), supporting the reliability of our experimental methods.
Positive epistasis rescues the mutational load of RAMs
We first looked at fitness effect of RAMs. In our definition, a mutant virus of relative fitness −1 means that the relative frequency of this mutant drops 10 fold after infection in cell culture. All single mutations were deleterious to virus replication ( Fig 2A ). The relative fitness of single mutants ranged from -2.33 (V82F) to -0.19 (L90M). This is consistent with previous reports that randomly introduced mutations were mostly deleterious to protease enzymatic activity or HIV-1 replication capacity [ 34 , 72 – 74 ]. Random mutagenesis in other viruses also revealed a lack of beneficial mutations in well-adapted systems [ 73 , 75 – 77 ]. RAMs in particular were also reported to be deleterious to virus replication [ 31 , 44 ]. They may destabilize viral protein, affect enzymatic activities or impact other protein-protein interactions [ 21 , 78 ].
10.1371/journal.pgen.1009009.g002
Fig 2
Positive epistasis is enriched among RAMs.
(A) Relative fitness of single mutants. Error bar is standard deviation ( n = 3). (B) The predicted relative fitness and observed relative fitness of double mutants. The predicted relative fitness was the sum of that of the two single mutants. Inset, the distribution of epistasis between double mutants. Error bar is standard deviation ( n = 3). (C)The predicted and observed fraction of viable mutants. A mutant was defined as viable if its relative fitness is higher than −4(dashed line) or −2(solid line).
We then analyzed epistasis between all pairs of RAMs. Previous studies have shown the prevalence of epistasis among pairs of random mutations [ 34 , 37 , 75 ] or spontaneously accumulated mutations [ 79 ]. However, studies focused on the epistasis among drug resistance mutations are still limited [ 30 , 39 , 72 , 75 , 80 ]. Based on the fitness effect of single RAMs, we predicted the relative fitness of double mutants with the assumption that no epistasis existed among any two single mutations (i.e., the predicted relative fitness of a double mutant was the sum of those of two single mutants)( Fig 2B ). Surprisingly, the observed relative fitness of most double mutants were significantly higher than the predicted values ( p = 2.2 × 10 −6 , two-sided Wilcoxon rank sum test), suggesting that positive epistasis is prevalent among RAMs ( Fig 2B inset). Pairwise epistasis between two RAMs is quantified as ε i , j = f i , j − f i − f j , f i represents the relative fitness of mutants i . The distribution of epistasis ranged from -0.69 (M46I and L90M) to 2.34 (L76V and V82F) and 86.6% of pairwise interactions between RAMs are positive.
We also analyzed the extent of epistasis among high-order mutants. We observed a trend that relative fitness decreased as the order of mutants increased ( S3 Fig ). This is consistent with previous reports that mutational load restricted virus replication capacity [ 30 , 81 , 82 ]. To better quantify the fitness cost of multiple mutations, we calculated the frequency of viable mutants by different thresholds, f > −2 or f > −4. The frequency of viable mutant virus decreased as the number of mutations increased ( Fig 2C ), consistent with previous observations in HIV-1 and other RNA viruses [ 83 – 86 ]. We then predicted the relative fitness of high-order mutants by summing the relative fitness of corresponding single mutants. We observed more viable mutants than would be predicted without epistasis ( Fig 2C ). This indicated pervasive positive epistasis rescued high-order mutants from lethal relative fitness, which is consistent with other clinical observations in protease inhibitor resistant virus [ 30 , 44 ]. As a result, positive epistasis partially relieved HIV-1 mutational load and allowed viruses to explore more sequence space.
Enrichment of positive epistasis among RAMs
There are two possible explanations for the observed positive epistasis among RAMs of HIV protease. The first hypothesis is that all mutations in HIV protease tend to interact positively. The second hypothesis is that epistasis among random mutations in HIV protease is on average zero, but positive epistasis is enriched among RAMs. We introduced the Potts model to test our hypotheses, while simultaneously testing whether our finding of prevalent positive epistasis among RAMs carries over to the clinical setting. Potts models, originally developed in statistical physics, have been employed previously to use the population-level frequencies and correlations between different mutations to estimate their fitness effects [ 87 – 90 ]. In the Potts model, the probability of observing a genotype A → = { A 1 , A 2 , … , A 99 } is given by equations in Fig 3A . Here the A i , i ∈ {1, 2, …, 99} are variables that represent the amino acid at site i on each of the 99 sites of protease. Two sets of Potts parameters, fields h i ( A i ) and couplings J ij ( A i , A j ), give the statistical energy E ( A → ) , which is negatively correlated with fitness. These parameters are estimated in order to reproduce the frequencies and correlations between mutations that are observed in the data. The fields h i ( A i ) represent the fitness effect of amino acids A i at sites i alone, while the couplings J ij ( A i , A j ) describe epistatic interactions between amino acids A i at site i and A j at site j . For both the couplings and the fields, positive parameter values correspond to beneficial effects on fitness, while negative values correspond to deleterious fitness effects. We applied a maximum entropy method [ 91 ] to an alignment of 20911 HIV-1 clade B protease sequences from drug naïve patients, obtained from the Los Alamos National Laboratory HIV sequence database (hiv.lanl.gov, accessed 24 March 2017) to calculate these two sets of Potts parameters.
10.1371/journal.pgen.1009009.g003
Fig 3
Positive epistasis rescues the mutational load of RAMs.
(A) The conceptual graph of Potts model. Potts model uses the probability of mutations occurring with other mutations to estimate the statistical energy. h i is the field parameter while J ij is the coupling parameter. (B) The correlation of Potts energy(Δ E = E mut − E WT ) and relative fitness of mutants with lower than 4 RAMs. Spearman correlation coefficient ( ρ ) is −0.46. (C)The cumulative density function of coupling parameters of RAMs and all other mutations. Coupling parameters between RAMs are more positive positive than those between RAMs and others ( D = 0.22, p = 2.1 × 10 −7 , two-sided K-S test) and those between other residues( D = 0.22, p = 5.1 × 10 −7 , two-sided K-S test). (D) The cumulative density function of field parameters of RAMs and all other mutations. Field parameters of RAMs and other residues are not significantly different( D = 0.25, p = 0.20, two-sided K-S test).
Then we calculated E ( A → ) for all mutants in our protease library. We found that the Potts energy for single, double or triple mutants (Δ E = E mut − E WT ) is significantly correlated with the relative fitness we measured in our screening ( ρ = −0.46, p = 1.2 × 10 −14 , Spearman’s correlation test, Fig 3B ). The correlation was lower than previous analysis in HIV-1 Gag and Env region [ 88 , 90 ]. This may be due in part to strong phylogenetic bias on the inferred Potts parameters, because protease is highly conserved. It is also possible that epistatic interactions with cleavage sites on other parts of the HIV-1 genome and complicated anti-innate immunity functions of protease [ 57 , 59 , 92 ] obscure the effects of individual mutations on replicative fitness in vitro .
The Potts couplings J ij ( A i , A j ) give the contribution of pairwise epistatic interactions between amino acids A i and A j at sites i and j , respectively. We compared the couplings among RAMs and among all other possible mutations on protease ( Fig 3C ). Couplings of other protease mutations clustered near 0, while those of RAMs are significantly more positive than that of other mutations (D = 0.22, p = 2.1 × 10 −07 , two-sided K-S test). Moreover, J ij ( A i , A j ) among RAMs were also more positive than those between RAMs and other residues (D = 0.22, p = 5.1 × 10 −07 , two-sided K-S test). Although the fields h i ( A i ) of RAMs are more negative than other mutations, the difference is not significant ( Fig 3D , D = 0.25, p = 0.20, two-sided K-S test). We note that the magnitude and the variation of field parameters is much larger than that of coupling parameters ( Fig 3C and 3D ). The Interquartile Range (IQR, i.e. the middle 50%) of field parameters is 3.55, while the IQR of coupling parameters is 0.15. The standard deviation of field parameters is 2.29, while the standard deviation of coupling parameters is 0.37. Overall, analysis based on the Potts model is consistent with our experimental results that positive epistasis is enriched among RAMs, and lends support to our second hypothesis that epistasis among random mutations in HIV protease is on average zero.
Implications of positive epistasis in evolution
To study the role of epistasis in evolution, we analyzed the evolutionary pathways covering all genotypes with up to 4 amino acid substitutions from the wild-type virus (13 single mutants, 67 double mutants, 176 triple mutants and 290 quadruple mutants) ( Fig 4A ). Mutants are linked if they differ by one amino acid substitution.
10.1371/journal.pgen.1009009.g004
Fig 4
Ruggedness in fitness landscapes prevents RAMs from reversion to wild-type.
(A) Fitness with possible evolutionary trajectories. Mutants are linked if they only have one residue difference. Red line represents an accessible path that a quadruple mutant can take and reverse to wild-type. Blue line represents an inaccessible reversal path to wild-type for that mutant. (B) Trajectory-based epistasis is calculated for each amino acid substitution and averaged over genetic backgrounds with a certain Hamming distance to the wild-type. The fitness effect of a single mutation becomes less deleterious on genetic backgrounds where other RAMs have been fixed. (C) The distribution of accessible paths for all genotypes with a certain hamming distance to wild type.
We have found that all 13 RAMs are deleterious on the wild-type background ( Fig 2A ). However, the fitness effect of a single RAM becomes less deleterious on genetic backgrounds where other RAMs have been fixed ( S4 Fig ). Following the generalized definition of epistasis proposed by Shah et al. [ 93 ], we define trajectory-based epistasis ε M , j that measures the deviation of the fitness effect if the order of mutations were reversed. ε M , j = f M , j − f M − f j , where f M and f j represent the relative fitness of background M and single mutant j [ 94 ]. For example, mutation j can be deleterious on the wild-type background but beneficial on another genetic background that mutation i has been fixed. Trajectory-based epistasis is calculated for each amino acid substitution and averaged over genetic backgrounds with a certain Hamming distance to the wild-type ( Fig 4B ). For all RAMs profiled in this study, we find that trajectory-based epistasis is overall positive and increases steadily with the number of substitutions, i.e. the fitness contribution of a specific amino acid substitution becomes more positive if more RAMs have been fixed. Our results are consistent with previous analyses of sequence co-variation in HIV-1 protease [ 65 , 66 ], where inferred epistastic interactions among mutations at PI resistance associated sites lead to entrenchment of primary drug resistance mutations. In this study, we combine the analyses of co-variation (Potts model) with comprehensive experimental fitness data of HIV-1 protease mutants (including a large number of higher-order mutants) to provide direct evidence of positive epistasis among RAMs of second-generation PIs.
We tested the hypothesis that positive epistasis prevented resistance associated genotypes from reverting to wild-type [ 41 , 95 , 96 ]. Although RAMs incurred significant fitness cost, some drug resistant mutants would not revert to wild-type after transmitting to a drug naïve patient. We quantified the frequency of accessible evolutionary pathways between mutants and wild-type in our experimentally measured fitness landscape of HIV protease RAMs. A reversal path is defined to be accessible if and only if the virus fitness increases monotonically along the path. For example, quadruple mutant V32I_M46I_I54L_V82F has many paths to revert to wild-type ( Fig 4A ). Among them, reversing V32I, I54L, V82F and M46I in order is an accessible path ( Fig 4A , red line). On the contrary, reversing I54L, V82F, M46I and V32I is not an accessible path because there are 2 steps with decreasing fitness ( Fig 4A , blue line). We found that among double mutants, 44 have two accessible reversal paths to the wild type, 20 have only one accessible reversal path, and interestingly 3 of them have none. These 3 mutants (I50V_T74P, M46I_I54M and L76V_V82F) represent local fitness peaks and the reversal to wild-type is blocked by a fitness valley. We found that the number of accessible reversal paths decreased with the accumulation of RAMs ( Fig 4C ). This indicates that protease mutants become less likely to revert to wild-type as the number of RAMs increases. Our results are consistent with clinical observations that protease inhibitor resistance associated mutations seldom reverted even when therapies were interrupted [ 25 , 62 ] or drug-naïve patients were infected [ 63 , 64 ]. The difficulty of reversal also explains the rising frequency of drug resistant HIV-1 viruses in acute phase patients [ 41 , 96 ].
Discussion
In this study, we systematically quantified the fitness effect of RAMs of HIV-1 protease. While all RAMs reduced the virus replication fitness, pervasive positive epistasis among RAMs alleviated the fitness cost substantially. Moreover, we analyzed the HIV sequence data from patients by the Potts model. We found the statistical energy inferred from HIV sequences in vivo correlated well with the replication fitness measured in vitro . Based on our fitness data and the mutational couplings inferred by the Potts model, we showed that positive epistasis is enriched among RAMs of HIV-1 protease, in both local fitness landscape and evolutionary paths. Finally, we studied the role of epistasis in evolutionary pathways. We found that positive epistasis among RAMs entrenches drug resistance and blocks the reversal paths to wild-type virus, which has important implications for the design of anti-retroviral therapies. Through this project, we also established a high-throughput platform to quantify the genetic interactions among a group of mutations. Another independent study profiled the fitness effect of all single amino acid change on HIV protease [ 71 ]. The data showed significant correlation with our study ( Fig 1E , Pearson’s correlation coefficient ( R ) is 0.79).
There are a few limitations of this study. Firstly, we only measured the fitness effect of RAMs in the absence of protease inhibitors. We are not able to quantify drug resistance of RAMs because protease inhibitors block multiple rounds of virus infection and prevent us from accurate examination of mutant frequency under drug selection. Also, we did not sequence other genes of HIV-1. HIV-1 mutates rapidly due to low fidelity of reverse transcriptase [ 97 , 98 ]. There might be compensatory mutations occurring on other proteins that rescued the protease RAMs. Secondly, the correlation between our validation experiments and high-throughput screening experiments was less than the correlation observed in similar experiments in bacteria and yeast [ 99 , 100 ]. The correlation between Potts energy and experimental fitness is also lower than previous reports on Gag and Env regions [ 88 , 90 ]. Mechanistic difference between logistic growth and viral growth may complicate the quantification of viral fitness [ 101 ]. Direct measurement of viral frequency may not linearly correlate to the probability of replication [ 102 ]. Moreover, we tested a large number of higher-order mutants (i.e. multiple mutations from the wild-type virus). Our experimental dataset not only contains clinically observed genotypes but also combinations of mutations that was not observed in patients, which are highly deleterious and may suffer from higher experimental errors. If we exclude higher-order mutants and very deleterious genotypes ( S5 Fig ), the Spearman’s correlation between fitness and Potts energy is higher ( ρ = −0.54, compared to ρ = −0.46 in Fig 3B ). Thirdly, we did not cover all clinically observed polymorphism, given the bottlenecks in virus library screening. We chose to prioritize for RAMs of second-generation protease inhibitors Darunavir (DRV) and Tipranavir (TPV), which are considered to have high genetic barriers (i.e. multiple RAMs are involved in the emergence and reversal of drug resistance) [ 52 ]. According to Stanford Drug Resistance Database [ 61 , 69 ], the RAMs that we chose contribute to the resistance to DRV and TPV ( S2 Table ). The only exception is L90M, which is frequently found in drug resistant viruses. The RAMs and the combinatorial genotypes in our library are prevalent in patients and documented in Stanford Drug Resistance Databases ( Table 1 ). Future work could be extended to cover more clinically observed polymorphism in HIV-1 protease and other drug-targeted proteins. Finally, the correlation between Potts energy and experimental fitness is confounded by many factors, like different selection pressures in vivo and in vitro , or phylogenetic bias. Nonetheless, we observe moderate but statistically significant correlation between the coupling parameters in the Potts model and the experimental epistasis ( S6 Fig , Spearman’s correlation test, p = 6.8 × 10 −3 ). We note that the coupling parameters in the Potts model and the experimental measure of epistasis (calculated for WT genetic background) are conceptually different, representing Fourier coefficients and Taylor coefficients of the fitness landscape [ 103 ]. Our findings are consistent with the literature that Potts model couplings are strongly associated with contact residues in the three-dimensional structure of protein families [ 104 , 105 ]. We tested a series of different statistical models, including the binary (Ising) model inferred via ACE, the Potts model inferred via pseudo-likelihood maximization (a popular approach to analyzing sequence data from protein families), and the Potts model inferred via ACE, to examine the epistatic effects among drug resistance mutations ( S7 Fig ). We found that the Potts model inferred via ACE is the best choice to analyze epistasis in our study.
Statistical models suggest a pervasive negative distribution of fitness effect for single mutations on HIV-1 [ 31 , 88 , 106 ]. Previous models also predicted the entrenchment of deleterious RAMs by positive epistasis [ 65 , 66 ]. This dataset provides a unique chance to experimentally test these statistical hypotheses. The predominance of positive epistasis is also observed in HIV-1 [ 30 ] and in other organisms [ 39 , 107 ]. However, they either relied on naturally-occurring resistant clones or indirectly activating gene functions. This report is the first dataset to systematically quantify the epistasis among functional residues in HIV-1 drug resistance evolution, without the bias of drug selection and in vivo evolution. Overall, our results are important for understanding drug resistance evolution. We found positive epistasis plays a critical role in HIV-1 gaining and maintaining drug resistance. Epistasis makes the fitness landscape rugged, preventing RAMs from reversion to wild-type, even when antiviral therapy is interrupted or virus transmits to a healthy individual [ 95 , 108 ].
Positive epistasis involves many kinds of molecular mechanisms. We find that the relative fitness of single mutants is not a significant factor of positive epistasis. We compared h i in the Potts model for all RAMs and other single mutants. They were not significantly different ( p = 0.20, K-S test). Physical distance between residues is a significant factor contributing to positive epistasis. The physical distances between these residues were significantly less than those between any two random residues on HIV-1 protease (D = 0.32, p = 3.9 × 10 −10 , two-sided K-S test, S8 Fig ), suggesting that physical contact among RAMs might contribute to the observed positive epistasis. Notably, their average distance was more than 10 Å, indicating most of them did not have direct contact. Some mutations may have structurally stabilizing effect to other residues. We used FoldX and Rosetta to predict the folding free energy (ΔΔ G ) as a quantification of protein stability [ 109 , 110 ] for all mutants in our library ( S8 Fig ). We notice that mutation V82F contributed to the positive epistasis on many genetic backgrounds ( Fig 4B ), but it did not contribute much to the stabilizing effect. Thus, structurally stabilizing effects cannot fully explain the predominance of positive epistasis observed in this study. Future studies on the structure and function of HIV-1 protease mutants will help elucidate the molecular mechanisms underlying the interactions among RAMs.
Material and methods
Plasmid library construction
HIV-1 RAMs were picked according to their prevalence in protease inhibitor treated patients [ 3 ]. We chose 11 residues with 13 mutations to construct a combination of HIV-1 protease mutant library ( Table 1 ).
We used a ligation-PCR method to construct the library on NL4-3 backbone, which is an infectious subtype B strain. All possible combinations of these 13 mutations are 2 9 × 3 2 = 4608 genotypes. The mutagenesis region spanned 243 nucleotides on HIV-1 genome. We split the region into 5 oligonucleotides and ligate them in order by T4 ligase (from New England BioLabs). The sequence of oligonucleotides are shown in S3 Table . After each ligation, we recovered the product by PCR and used restriction enzyme BsaI-HF (from New England BioLabs) to generate a sticky end for the next step ligation.
After making the 243-nucleotide mutagenesis fragment, we PCR amplified the upstream and downstream regions near this fragment and used overlap extension PCR to ligate them together. We then cloned it into full length HIV-1 NL4-3 background. We harvested more than 30,000 E. coli colonies to ensure sufficient coverage of the library complexity.
Virus production
The plasmid DNA was purified by HiPure Plasmid Midi Prep Kit (from Thermo Fisher Scientific). To produce virus, we used 16 μg plasmid DNA and 40 μL lipofectamine 2000 (from Thermo Fisher Scientific) to transfect 2 × 10 7 293T cells, in 3 independent biological replicates. We changed media 12 hours post transfection. The supernatant was harvested 48 hours post transfection, labeled as input virus and frozen at -80°C. We harvested 40mL viruses from each transfection. Virus was quantified by p24 antigen ELISA kit (from PerkinElmer).
Library screening
CEM cells were cultured in RMPI 1640 (from Corning) with 10% FBS (from Corning). To passage library in T cells, we added 25 mL viruses and 120 μg polybrene to 50 million CEM cells. We achieved 10 ng p24 (10 8 physical viral particles) for every million CEM cells during infection. We washed cells and completely changed media 6 hours post infection. We supplemented the cells with fresh media 3 days post infection and harvested supernatant 6 days post infection. We centrifuged supernatant at 500 × g for 3 minutes to remove the cells and cell debris. The rest of supernatant was frozen at -80°C.
In summary, we carefully controlled the experiment scales to ensure the library complexity was maintained in every step. Briefly, we harvested >3 × 10 4 E. coli colonies during bacteria transformation, which ensured ∼6-fold coverage of the expected complexity (4608 genotypes). We then transfected 2 × 10 7 HEK 293T cells with 16 μg plasmid library to package infectious viruses. We used 25 mL viruses (500 ng p24, ∼ 5 × 10 9 viral particles) to infect 2 × 10 7 million CEM cells for each biological replicate.
Sequencing library preparation
We used QIAamp viral RNA mini kit (from QIAGEN) to extract virus RNA from supernatant. We then used DNase I (from Thermo Fisher Scientific) to remove the residual DNA. We used random hexamer and SuperScript III (from Thermo Fisher Scientific) to synthesize cDNA. The virus genome copy number was quantified by qPCR. The qPCR primers are 5’ -CCTTGTTGGTCCAAAATGCGAAC-3’ and 5’ -ATGGCCGGGTCCCCCCACTCCCT-3’.
At least 2 × 10 5 copies of viral genome were used to make sequencing libraries. We PCR amplified the mutagenesis regions using the following primers: 5’ -CTAATCCTGGAGTCTTTGGCAGCGACCC-3’ and 5’ -GAAGACCTGGAGTGCAGCCAATCTGAGT-3’. We then used BpmI (from New England BioLabs) to cleave the primers and ligate the sequencing adapter to the amplicon. We used PE250 program on Illumina MiSeq platform to sequence the amplicon.
Calculation of fitness and epistasis
We used custom python codes to map the sequencing reads to reference NL4-3 genome. Mutations were called if both forward and reverse reads have the same mutation and phred quality scores are both above 30. All codes are available on https://github.com/Tian-hao/protease-inhibitor . All data were deposited in SRA (short read archive) database under accession PRJNA546460 . For each replicate of the virus library from the transfected 293T cells, we reached 4.45 × 10 5 to 6.05 × 10 5 sequencing depth. We filtered out the genotypes with frequency fewer than 5 × 10 −5 in any biological replicate and the genotypes whose frequency differ more than 10 folds between any two biological replicates.
Relative fitness f m , r of mutant m in experiment r (biological replicates) was defined as Eq 1 .
f m , r = log 10 ( F m , r , o u t p u t F m , r , i n p u t / F W T , r , o u t p u t F W T , r , i n p u t ) (1)
F m , r , input is the frequency of mutant m before screening. F m , r , output is the frequency of mutant m after passaging. F WT , r , input is the frequency of wild-type virus before screening. F WT , r , output is the frequency of wild-type virus after passaging.
The relative fitness f m was defined as the average of 3 biological replicates ( Eq 2 ). However, if relative fitness was missing in one replicate, we only average the other two replicates. The relative fitness value of all mutants was shown in S1 Table .
f m = ∑ t = 1 R f m , r / R , (2)
where R is the number of biological replicates.
Pairwise epistasis ε i , j between mutant i and mutant j was defined as:
ε i , j = f i , j - f i - f j , (3)
where f i , j refers to the relative fitness of double mutant i and j .
Trajectory-based epistasis ε M , j between a multi-mutation genotype M and another genotype differ by one mutation j was defined as:
ε M , j = f M , j - f M - f j (4)
Potts model
Data used to infer parameters for the Potts model were downloaded from the Los Alamos National Laboratory HIV sequence database, as described in the main text. Sequences were processed as previously described [ 111 ]. Briefly, we first removed insertions relative to the HXB2 reference sequence. We also excluded sequences labeled as “problematic” in the database, and sequences with gaps or ambiguous amino acids present at >5% of residues were removed. Remaining ambiguous amino acids were imputed using simple mean imputation.
Each sequence in the multiple sequence alignment (MSA) is represented as a vector of variables A → = { A 1 , A 2 , … , A N } , where N = 99 is the length of the sequence. Each of the A i represents a (set of) amino acid(s) present at residue i in the protein sequence. To choose the amino acids at each site that would be explicitly represented in the model, we first computed the frequency p i * ( A ) of each amino acid A at each site i in the MSA. To compute these frequencies, we weighted the sequences such that the weight of all sequences from each unique patient was equal to one, thereby avoiding overcounting in cases where many sequences were isolated from a single individual. We then explicitly modeled the q i most frequently observed amino acids at each site that collectively capture at least 90% of the Shannon entropy of the distribution of amino acids at that site [ 111 ]. All remaining, rarely observed amino acids were grouped together into a single aggregate state. For these data, this choice resulted in an average of three explicitly modeled states at each site (minimum of 2, maximum of 6).
The Potts model is a probabilistic model for the ‘compressed’ sequences A → , where the probability of observing a sequence A → is
P ( A → ) = 1 Z e - E ( A → ) , (5)
E ( A → ) = - ∑ i = 1 m h i ( A i ) - ∑ i = 1 m ∑ j = i + 1 m J i j ( A i , A j ) . (6)
Here the normalizing factor
Z = ∑ A → e - E ( A → ) (7)
ensures that the probability distribution is normalized. We used ACE [ 91 ] to infer the set of Potts fields h i ( A i ) and couplings J ij ( A i , A j ) that result in average frequencies and correlations between amino acids in the model ( 5 ) that match the frequencies p i * ( A i ) and correlations p i j * ( A i , A j ) observed in the data. We used a regularization strength of γ = 7 × 10 −5 in the inference, which is roughly equal to one divided by the number of unique patients from which the sequence data were obtained. We used “consensus gauge,” where the fields and couplings for the most frequent residue at each site in the protein are set to zero. We confirmed that the parameters inferred by ACE resulted in a Potts model that accurately recovered the correlations present in the data.
Validation experiments
We constructed 7 single mutants by site-directed mutagenesis. The primers used this experiment are listed in S3 Table . We used overlap-extension PCR to amplify the fragment with mutated nucleotides. We ligated the fragment with NL4-3 backbone using ApaI and SbfI. We transformed competent E.coli and picked single colonies. We sequenced the protease region of plasmids to make sure there is only desired mutant in this region. 7 mutants were L10F, I47V, T74P, L76V, V82F, V82T, L90M.
We produced mutant viruses in 293T cells, mixed them with wild-type and infected CEM cells. The frequencies of mutant virus before and after infection were quantified by deep sequencing. We did 2 biological replicates with each validation method. For validation 1, we pairwisely mixed the mutant and wild-type virus oor competition. For validation 2, we mixed all 7 mutants and wild-type virus.
Protein stability prediction
Mutants stability was predicted using either FoldX or Rosetta. For FoldX, we used the protease structure (PDB: 3S85) as reference and repaired the structure using the RepairPDB function. The free energy of the mutants was computed by using the BuildModel function under default parameters. For Rosetta analysis, we used the protease crystal structure (PDB: 6DGX) as reference and score function ddg_monomer to evaluate the effect of mutations. Each mutants were evaluated 10 times and the average score was used as ΔΔG.
Ethics statement
Reagents were acquired from the NIH AIDS Reagent program. The work is approved by UCLA IRB.
Supporting information
S1 Fig
The correlation of relative fitness among biological replicates.
All single mutants, double mutants and triple mutants are shown. R stands for Pearson correlation coefficient.
(TIFF)
S2 Fig
Coverage of protease mutant library.
(A) Fraction of expected protease mutants in each transfection virus library. (B) Number of mutant in each transfection virus library. Dashed line represents the number of all possible combinations of mutations.
(TIFF)
S3 Fig
Relative fitness of different order of mutations.
(TIFF)
S4 Fig
Relative fitness of single RAMs on different genetic backgrounds.
(TIFF)
S5 Fig
Correlation between Potts energy and relative fitness for low order mutants.
Mutants with relative fitness higher than −2.5 and numbers of mutations lower than 4 is shown. The Pearson’s correlation coefficient is −0.57. The Spearman’s correlation coefficient is −0.54.
(TIFF)
S6 Fig
The correlation between Potts’ coupling parameters with experimental epistasis.
The pairwise epistasis between all RAMs in our library was compared with Potts’ coupling parameters. The Spearman’s correlation coefficient is −0.33. The p value for the Spearman’s correlation coefficient is 6.8 × 10 −3 .
(TIFF)
S7 Fig
Correlation between relative fitness and different statistical models.
(A, B & C)The correlation between relative fitness with (A, bin) binary (Ising) model inferred via ACE, (B, plm) the Potts model inferred via pseudo-likelihood maximization, or (C, potts) the Potts model inferred via ACE. (D) Spearman’s correlation coefficients for different models. Mutants were classified according to their HD to wild-type. HD, hamming distance.
(TIFF)
S8 Fig
Structure insights on resistance associated mutations.
(A) Distribution of pairwise distance among resistance associated residues and other residues. The distance between the C- α of two residues was shown. (B & C) Correlation between mutants’ relative fitness and protein stability (ΔΔ G ). ΔΔ G is predicted by FoldX (B) or Rosetta (C). The correlation coefficients were calculated for mutants with lower than 5 mutations. ρ stands for Spearman’s correlation coefficient.
(TIFF)
S1 Table
Relative fitness of all mutants in this research.
(PDF)
S2 Table
Information of protease inhibitor resistance associated mutations covered in the library.
(PDF)
S3 Table
Sequence of oligonucleotides used in this research.
(PDF)
S4 Table
Protein stability simulated by Rosetta or FoldX.
(TSV)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The binding of transcription factors (TF) to cis elements is a key component of most biological processes. Being able to detect TF binding sites (TFBS) by inspecting genome sequences helps understanding how organisms work and how they evolved. Methods based on Chromatin Immunoprecipitation (ChIP) such as ChIP-Chip [ 1 ], ChIP-Seq [ 2 ] or ChIP-exo [ 3 ] allow the identification of all genomic regions bound by a given TF in one experimental condition and suites as Bedtools [ 4 , 5 ] offer many tools to manipulate them. To precisely identify the TFBS present in these regions, estimate their affinity, predict binding sites that might be bound in other experimental conditions, or study organism where ChIP experiments are more challenging, TF DNA binding models are extremely useful. There are multiple ways to model TFBS. The most common is the Position Weight Matrix (PWM) that, for each sequence, computes a score directly related to the TF/DNA affinity ([ 6 ] for a review). This method, however, assumes that each base of the BS contributes independently to the affinity of the TF for DNA [ 7 ] and there is evidence that interdependencies between positions exist [ 8 , 9 ] and that taking into account dinucleotide dependencies between two adjacent positions already improves predictions [ 10 ]. Several alternatives with specific advantages have been proposed using nucleotide subsequences (K-mers) rather than mononucleotide positions [ 8 , 11 , 12 ] or hidden Markov models (HMM) [ 13 ]. However, in most cases, PWMs provide simple and reliable estimation of binding affinity [ 14 ]. We propose to adapt the convenient PWM model by adding dependency information at specific positions of the matrix. As documented for several TFs [ 10 , 15 – 17 ], this will improve the prediction power of some PWM models.
Although several tools such as RSAT [ 18 ], PROMO [ 19 ], MatInspector [ 20 ] or LASAGNA [ 21 ] are available to identify overrepresented motifs in a set of sequence and build binding models, none of them allows using PWM with dependencies nor to calculate occupancy of DNA regions using biophysical models [ 22 ]. We have developed a new algorithm that uses PWM with any combination of dependent and independent positions. We incorporated it in a user friendly set of tools called MORPHEUS, which offers several specific advantages over existing tools: 1) it is a web tool that does not require any programming skill and can thus be widely used by the biologist community, 2) users can import their own matrices, not only those found in databases, 3) position interdependencies can be included between any positions of the matrix and in combination with independent positions, a possibility currently offered by none of the existing web tools, 4) a global “predicted occupancy” value can be computed for whole DNA regions using a biophysical model [ 22 ] that integrates the presence of individual binding sites.
Results and Discussion
Morpheus Matrix Format and m PWM algorithm
The Morpheus PWM format ( m PWM) allows the introduction of information on di- or tri-nucleotide dependencies between any indicated positions (not just adjacent ones) within a binding site. Unlike other models that increase model complexity for all positions, m PWM conserves the simplicity of a PWM except for interdependent positions. Using m PWM, interdependencies are defined as additional 4 (d) matrices (d = 2 for dinucleotide dependency, d = 3 for triplets) for any position combination (Example matrix files are provided as S1 Text and S2 Text).
mPWM Format Conversion Tool
We have provided a tool to generate m PWM from an alignment of transcription factor binding sites. Positions with dependency have to be detected using programs such ENOLOGOS [ 23 ] and provided as a list of dependent positions for the conversion tool to automatically generate the corresponding 4 (d) matrices. Depending on the TF structural features, the possibility is offered to generate symmetric matrices.
Morpheus tools
The Morpheus suite allows the calculation of relative affinity of TFBS from m PWM. Based on the scores of individual binding sites present in a large DNA region, Morpheus also computes the predicted occupancy using a biophysical model as previously described [ 22 , 24 , 25 ]. This possibility, not offered by any other web-tool, is particularly important as individual cis elements can vary within a regulatory region even though the occupancy and overall regulation are conserved [ 16 , 26 ]. The predicted occupancy thus offers a global measure that allows comparing regions independently of the individual binding site variations.
Morpheus webtool it is composed of three tools:
- Morpheus ‘Score’ tool scans DNA regions and computes the scores of individual TFBS. The user can choose to display only the TFBS of highest score of each region, the TFBS with a score higher than a given threshold score or all TFBS. For an easy graphical representation, this tool also generates score profiles for each sequence as well as an histogram with all scores ( Fig 1 ).
10.1371/journal.pone.0135586.g001
Fig 1
Morpheus flowchart and example of result representation.
The tool Score scans DNA regions and computes the scores of TFBS. The bottom left graph shows TFBS locations and scores; such score profile is generated for each sequence submitted. The Occupancy tool computes the TF predicted occupancy of each DNA region taking as input sequence files (in fasta format) and a binding model information ( m PWM format). Complete results are written in text files and also displayed as graphical outputs for quick results overview. Bottom right panel shows a occupancy comparison between different DNA regions. The bottom central panel illustrates the ROC-AUC curve and value obtained with the ROC-AUC tool.
- Morpheus ‘Occupancy’ tool computes the TF predicted occupancy of each DNA region using formalism described above [ 22 , 24 ] with the option of using only the scores exceeding a given threshold. Occupancy calculation is based on the correlation between predicted score and relative dissociation constant which can be obtained from in vitro measurements of relative affinities [ 27 ]. If this data is not available a relative occupancy can be calculated using default values for the parameters.
Both score and occupancy options take two files as input: a file with sequences in fasta format and a m PWM.
- Morpheus ‘ROC’ allows assessing the quality of a TF binding model by performing a Receiver Operating Characteristics (ROC) analysis [ 28 ]. This analysis measures the discriminative power of a TF matrix by comparing a set of bound regions (obtained for example from ChIP-Seq experiments) to a negative control set generated by the user. The comparison uses either the best score TFBS of each sequence or its occupancy and the Area Under the Curve (AUC) value is computed as a measure of the model predictive power.
All three programs display graphic output ( Fig 1 ) for quick results overview and text files with complete results for further analysis by the user.
For illustration of how Morpheus suite works, we present here a set of analyses performed with the LEAFY (LFY) protein, a plant TF with a central role in the evolution and development of flowers [ 29 , 30 ]. According to in vitro affinity measurement a PWM has been proposed for this factor (LFY-Trip) that includes three dependency triplets in a symmetric motif of 19 positions [ 16 ], in accordance with the information obtained from the LFY-DNA crystal structures [ 31 , 32 ].
LEAFY Binding model evaluation using ROC
The availability of ChIP-Seq data allows performing ROC analysis using the described set of bound genomic regions [ 16 ] as well as a negative set of non bound regions (see Methods for description of negative set generation) to compare the predictive power of this matrix against the previously described consensus motifs [ 31 , 33 , 34 ]. To do this, Scores or Occupancies are computed for both the positive and negative sets with each binding model (using Score or Occupancy tools) and the result serves as input for the ROC program. A histogram is generated that represents the distribution of scores or occupancy values for each data set ( Fig 2A ). This tool also generates an image file with the ROC curve and a text file with all the data. In Fig 2B , we use ROC-AUC results to illustrate the increased prediction power of the LFY-Trip PWM as compared to previously used consensus sequences. This tool can be used to compare the various matrices identified by various motifs finding algorithms in order to select the one with the best predictive power. Next, we illustrate how the LFY m PWM can be used for functional or evolutionary analysis of a regulatory relationship using the Score and Occupancy tools.
10.1371/journal.pone.0135586.g002
Fig 2
The performance of a TF model can be evaluated by its capability to discriminate between bound and non-bound regions as determined from a ChIP-Seq experiment.
The Morpheus ‘ROC’ tools computes the ROC-AUC value as a measure of the model predictive power. A) The histogram graphical output displays the distribution of score values for the best binding sites present on each DNA sequence. B) ROC data output for three binding models: two consensus motifs and LFY-trip (input data has been generated using the Morpheus ‘Occupancy’ tool). The LFY-trip model largely outperforms the two consensus models.
Prediction of LEAFY binding sites on APETALA1 promoter
We focused on the link between LFY and its direct target APETALA1 ( AP1 ) involved in the development of flowers [ 29 , 30 ]. The MADS box TF gene AP1 arose from duplication of the FRUITFUL ( FUL ) gene and this event was proposed to be important in the fixation of flower structure in eudicot plants [ 35 ]. AP1 have also experienced a more recent Brassicaceae-specific duplication [ 36 , 37 ] generating the CAULIFLOWER gene. While AP1 is a direct target of LFY, there is no evidence for direct regulation of FUL or CAL [ 16 , 38 ]. We illustrate here how Morpheus can be used to explore LFY-AP1 link through eudicot plants evolution.
The functional analysis of AP1 promoter and its regulation by LFY binding has been performed in the model plant Arabidopsis thaliana . A few promoter versions have been tested in vivo [ 39 ] including different promoter lengths (2.2, 1.7, 0.9 and 0.6 kb), and mutations in three candidate LFYBS (bs1, bs2 and bs3) displaying consensus motifs [ 31 , 33 , 34 ]. The score profile of AP1 promoter generated with Morpheus Score tool (option "limit = -25") illustrates the position of the best LFYBS in AP1 promoter ( Fig 3A ). We computed Occupancy values for all promoter versions and compared these values to the in vivo activity of the corresponding promoter fragment ( Fig 3B ). In vivo , mutations in bs2 and bs3 had weak effect while mutation in bs1 had a strongest effect, which is in accordance with their computed scores and not with the presence of the consensus sequence. The good correlation between the two types of data illustrates the power of the biophysical model to predict the impact of TFBS changes on gene expression by integrating all possible TFBS present in a regulatory region.
10.1371/journal.pone.0135586.g003
Fig 3
Comparison of LEAFY binding analysis in A . thaliana AP1 promoter using the Morpheus suite with in vivo promoter expression study [ 39 ].
A) Score profile graphic output of the Morpheus ‘Score’ tool (option limit = -25) using 2.2 kb upstream of AP1 start codon. Red dotted lines show the different promoter sizes and arrows mark the mutated BS, accordingly with promoters set described in [ 39 ]. B) Predicted occupancy (option All) shows a good correspondence with relative expression of each promoter version as determined experimentally in a published study and summarized: expression levels: +++ (high), ++ (medium), + (low),—(not detectable). The number indicates the size of the promoter (2.2, 1.7, 0.9 or 0.6 kb), m2 and m3 indicate mutations in bs2 and bs3 respectively, ∆1 indicates a deletion of bs1. In vivo , mutations of bs2 and bs3 (promoter 0.6 mutbs2bs3) has only a weak effect while elimination of bs1 drastically affects AP1 expression.
Transcriptional regulation and evolution
Next, we use the Morpheus tools to study the evolution of the link between LFY and genes of the FUL clade ( AP1 , CAL and FUL ). Scanning of 2 kb of promoter sequences in various species illustrates well the diversity of LFYBS landscapes ( Fig 4B ). As it is difficult to draw clear conclusions directly from these TFBS profiles, we computed the Occupancy (Occ) for these different promoters ( Fig 4A ). We found a higher occupancy of the AP1 ortholog promoters as compared to those of CAL and FUL in the Brassicales clade, a result in good accordance with experimental data available in Arabidopsis [ 16 , 33 , 38 , 40 ]. This analysis suggests that the link between LFY and AP1 originated before the divergence of B . rapa .
10.1371/journal.pone.0135586.g004
Fig 4
Evolutionary analysis of LFY binding on AP1 promoters.
Genomic sequences were obtained from the Phytozome database and 2 kb promoter upstream the ATG were used. Only well annotated genes were used. A) Predicted occupancy (Option limit = -23) for each promoter. Phylogenetic relationships between species are represented. B) Score profile (limit = -23) of some representative promoters. The higher occupancy for AP1 promoters in Brassicales (red) suggests that the regulatory link between LFY and AP1 in A . thaliana arose before the divergence of this clade. Interestingly, C . papaya with low occupancy in AP1 promoter has a candidate BS of very good score downstream the start codon likely to be responsible for a regulation by LFY. In Fabids, the low occupancy values suggest that LFY does not regulate AP1 , though some promoters have intermediate occupancy values (green) what will need further analysis.
Interestingly, the promoter of the AP1 gene from the Brassicale Carica papaya displays a low occupancy though evidence suggests a regulation by LFY. We wondered whether this reflected an absence of regulatory link between LFY and AP1 in this species or whether the LFY binding sites could be located outside of the promoter. We thus scanned the region downstream of the start codon and we found a predicted BS with a very good score that could be responsible for the AP1 regulation by LFY in this species despite the absence of high affinity LFYBS in the promoter. None of the other species with low promoter occupancy displayed this 3’ binding site (data not shown). This data supports the hypothesis that LFY- AP1 link originated before Brassicales divergence. The low occupancy values for AP1 promoters in Fabids suggest LFY does not regulate AP1 in these species. However, because there are intermediate Occupancy values in the Fabaceae clade, a more detailed experimental characterization would be required in these species to assay the possible existence of a regulatory relationship between LFY and AP1 .
These analyses illustrate how genomic sequences can be analysed with the Morpheus tool to generate hypotheses regarding gene regulation and regulatory network evolution. More examples can be found in three additional studies [ 16 , 41 , 42 ] that used Morpheus while under development. As more TF binding models become available, such tools will become increasingly important to exploit the genomic data, answer evolutionary questions and bringing up new working hypotheses.
Conclusions
Morpheus web allows a user-friendly suite of tools for the calculation of TFBS relative affinity on DNA sequences. It incorporates unique features such as dependency between specific positions, occupancy calculation and ROC-AUC estimation that do not exist in any currently available webtool. We have illustrated how it can be used to infer hypothesis about TFBS functional significance or about evolution of regulatory links. Experienced users can download Morpheus scripts code for specific purpose, however no programming skills are needed to use Morpheus web-tools. With all its unique characteristics and with the possibility of using any own-modified m PWM, we believe Morpheus should have strong acceptance among biologists. Morpheus web-tools, complete user guide and downloading versions are available at Morpheus website: http://biodev.cea.fr/morpheus/ .
Methods
Morpheus
All scripts for Morpheus tools are written in Python programming language (ver 2.6.7). The graphic output requires two modules: Numpy ( http://numpy.scipy.org/ ) and Matplotlib ( http://matplotlib.sourceforge.net/ ). Morpheus tools are available in the Morpheus web ( http://biodev.cea.fr/morpheus/ ), as well as downloading versions with or without graphic output, user guide and complete descriptions. The web is hosted and maintained by the GIPSI team (CEA Saclay).
When the score matrix is not directly provided, Morpheus computes it based on ‘Count’ or ‘Frequency’ matrices using W n , i = Ln(f n , i /f max , i ) where W n , i is the weight at position i for nucleotide n , f n , i is the frequency of nucleotide n at position i and f max , i is the maximal frequency observed at position i [ 43 ]. Each 4 (d) dependency matrix is preceded by a line indicating the positions involved ( S2 Text ). For score calculation the m PWM algorithm first get the value for each independent position from the independent matrix and then for all the dependency combinations from the 4 (d) matrices. If in vitro affinity data is available to correlate score with relative dissociation constant, the correlation values can be indicated in m PWM file or, if they are not indicated, the program will use default parameters (corresponding to a line curve with scope equal to one). From matrix file, m PWM algorithm first identifies the list of independent positions ( i ) and the list of dependent positions groups ( j ; each one associated with a 4 (d) matrix), then the score of each DNA sequence is calculated as:
S e q u e n c e s c o r e = ∑ p ∈ i s c o r e p n t + ∑ q ∈ j s c o r e q d e p
where
s c o r e p n t
is the score in the position p for the nucleotide nt (A,C,G or T) in the independent matrix and
s c o r e q d e p
is the score in the 4 (d) matrix of group q for the sequence combination dep (dinucleotide or triplet). Example of matrix files in Morpheus format with or without dependencies are provided as S1 Text and S2 Text, respectively.
Occupancy calculation (default parameters)
Occupancy calculation is based on the relation between predicted score and relative dissociation constant [ 16 , 24 ], score = -a * ln(Kd) + b. A and b values can be provided in the m PWM file when they have been determined. If not, Morpheus will use default parameters (a = 1.0 and b = 0.0). Occupancy calculation formalism also requires the TF concentration [X], which can optionally be indicated if available. Since this value is rarely available, as default, Morpheus uses [X] = e (b/a) corresponding to a TF concentration at which the best possible binding site (maximal score) is bound with a probability of 0.5.
Sequences sets for ROC-AUC calculation
Positive bound sequences was taken from [ 16 ] ( S3 Text ). To generate a negative set of sequences for ROC-AUC analysis, we have randomly selected in the A . thaliana genome a set of sequences that do not overlap with the positive set and with the same size distribution ( S4 Text ).
FUL clade genomic sequences
Genomic sequences were obtained from Phytozome database [ 44 ] by Blast search using the protein sequence of AtAP1 (At1g69120), AtCAL (At1g26310) and AtFUL (At5g60910). Hits without transcripts information or with incomplete gene prediction were discarded. A region of 2 kb upstream the ATG were used for relative binding score calculation. All sequences used in this study can be found in S5 Text .
Supporting Information
S1 Text
Example of mPWM format without dependency.
(TXT)
S2 Text
Example of mPWM format with dependency.
(TXT)
S3 Text
Positive Sequences Set.
(TXT)
S4 Text
Negative Sequences Set.
(TXT)
S5 Text
FUL clade genomic sequences.
(TXT)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Leprosy, also known as Hansen's disease, is a chronic infectious disease caused by Mycobacterium leprae . Leprosy remains a significant public health problem in several parts of the world. According to official reports submitted to World Health Organization (WHO) by 105 countries, the number of new cases detected during the year 2011 was 219,075. Indeed, in 2012, 33,955 new cases were detected in Brazil alone [1] . In this context; several epidemiological studies have documented the main regions of Brazil with a high prevalence of leprosy [2] – [4] .
The current strategy for leprosy control recommended by the WHO is based on multidrug therapy (MDT) that consists of the combination of rifampicin, clofazimine and dapsone (DDS) for multibacillary (MB) leprosy patients and rifampicin and dapsone for paucibacillary (PB) leprosy patients [1] . DDS (4,4′-diaminodiphenylsulfone) has bacteriostatic action against Mycobacterium leprae and is an essential component of MDT. The action of DDS is due to inhibition of dihydrofolic acid synthesis by competition with para-aminobenzoic acid (PABA) [5] . DDS distributes in all body organs including skin, liver, kidneys, and erythrocytes, and crosses the blood-brain barrier and the placenta, as well as being found in breast milk [6] . DDS was initially used as an antibiotic in humans at doses equivalent to sulfonamides, which led to severe hemolytic anemia and methemoglobinemia [7] , [8] .
Recently, our studies on the molecular structure/activity properties of DDS showed that its biological properties are strongly influenced by redox mechanisms associated with its sulphone group as well as its nucleophilic aniline rings. Hence, during the oxidative clearance of dapsone in man, hepatic CYPs exploit the propensity of the molecule to undergo electron transfer or oxidation to N-hydroxylated metabolites such as DDS-NHOH and monoacetyl-hydroxylamine MADDS-NHOH [7] – [9] . Thus, through its metabolically formed hydroxylated derivatives, DDS is able to exert local oxidative stress conditions which impacts macromolecules, such as proteins, lipids, carbohydrates and nucleic acid, ultimately leading to cellular necrosis in patients [7] , [10] . The primary manifestation of the oxidative capacity of dapsone-related hydroxylamines, is their induction of methemoglobinemia in patients which may also lead to hemolysis [7] . Indeed, methemoglobin formation is caused by the co-oxidation of the hydroxylamine metabolites, with oxyhemoglobin in erythrocytes [7] , [10] . In this study, we investigated the contribution of multidrug therapy, which includes dapsone, towards the generation of oxidative stress and cell damage through the analysis of antioxidant status (total antioxidant capacity, superoxide dismutase and catalase activities), oxidative markers (nitric oxide levels, lipoperoxidation, methemoglobin formation) and DDS levels in patients with leprosy. The results were then associated with the known redox mechanisms DDS and DDS-NHOH, through molecular modeling studies. Whilst the role of the hydroxylamine metabolites in dapsone toxicity is well established, the CYP isoforms primarily responsible for their formation have been the subject of considerable study in a variety of clinical and experimental models over past decades; indeed, CYP3A4, CYP2E1 and CYP2C9 [11] – [13] have each been postulated as the major contributor to the oxidation of this drug. Latterly, a role for CYP2C19, has been outlined in a study with recombinant isoforms [14] and in our report we also explore the potential interactions between dapsone and CYP2C19 using molecular docking analysis.
Methods
Ethics statement
The Ethics Committee of the Federal University of Pará, Brazil, approved the study protocol (protocol 079/09). It was also approved by the State Reference Unit for Leprosy Treatment Dr. Marcelo Candia, Marituba and Health unit Guama, Brazil and so gave permission to start collecting data. All participants were informed about the aims and methods of study and they also wrote and signed the informed consent before the start of the experiment and sample collection.
Population and experimental design
In this study, a total of forty-three subjects comprising twenty-three patients diagnosed with leprosy, receiving care in the Department of the State Reference Unit for Leprosy Treatment under Dr. Marcello Candia- Marituba, and in the Health Unit of Guamá, Belém, Para, Brazil. These patients were selected for the study before starting multidrug therapy (called MDT 0) and they were followed until the third month of multidrug therapy (MDT 3). Leprosy patients (age range, 20–45 years) were classified into paucibacillary (PB; 9 cases) and multibacillary (MB; 14 cases), based on WHO clinical criteria (testing positive for 2 of the 3 clinical criteria—skin lesions (≤5, PB and >5, MB), anesthesia, and nerve enlargement) and Bacteriological Index (BI) [1] . Two slit-skin smears, one from a representative lesion and the other from an earlobe, were obtained, and then stained using modified Ziehl–Neelsen technique. A minimum of 100 oil-immersion fields of the smear were examined for the presence of acid-fast bacilli, and the BI was calculated.
Leprosy patients with reactions, ulceration, a history of smoking, or those under the influence of alcohol, over 45 years of age, with co-infections or diabetes mellitus or other systemic diseases or health problems, and history of drug and nutraceutical use, including vitamins, ascorbic acid, and tocopherol, were excluded to rule out their possible influence on the study parameters. In this study, patients were divided into two groups: samples collected before administering the first dose MDT-supervised (MDT 0) and after two months in the third dose MDT-supervised (MDT 3).
Healthy adults were selected voluntarily to serve as controls (n = 20). This group did not have signs and symptoms of leprosy, other diseases, or health problems, and the volunteers were nonsmokers and free from drug use. The control group consisted of sex-matched individuals, aged 20–45 years, who were living in the same settings as those of the patients.
Blood samples (10 mL) were obtained from all participants by venipuncture in tubes containing ethylenediaminetetraacetic acid (EDTA). The whole blood samples were divided into two aliquots; the first was used immediately for determining of MetHb and GSH levels, activities of SOD and CAT and presence of Heinz Bodies, whilst the second aliquot was centrifuged at 2000× g for 6 min, to separate plasma for the analyses of MDA and total antioxidant status (TAS) levels as well as dapsone concentration analysis. In addition, the serum was collected for NO measurement ( Figure 1 ).
10.1371/journal.pone.0085712.g001 Figure 1
Sampling procedure.
DDS determination
The plasma concentrations of DDS were determined in accordance with the procedures of extraction, quantification and standardized by Kwadijk and Toraño [5] , using High Performance Liquid Chromatography (HPLC). The model used in the chromatograph was ProStarVARIAN work with ODS RP18 column, 25 cm×4 mm, UV detection at a wavelength of 295 nm and flow 1 mL/min. As mobile phase we used aqueous solution of 30% acetonitrile and as internal standard, phenacetin solution of 100 µg/mL.
Determination of Methemoglobin (MetHb) Content
MetHb content was evaluated from the change in absorbance at a wavelength of 632 nm; this change was caused by addition of potassium cyanide (KCN) to the buffered hemolysate. A dilution of the hemolysate, in which potassium ferricyanide (K 3 Fe(CN) 6 ) was used to convert all possible forms of hemoglobin (Hb) to MetHb, was used as a reference solution. The MetHb content was measured in duplicate, and values less than 2% were considered normal [15] .
Determination of Lipid Peroxidation
Lipid peroxidation was measured by quantifying MDA in blood samples of patients using the thiobarbituric acid-reactive substances (TBARS) assay. This method is a very useful, economical, and easy-to-use assay for evaluating oxidative stress [16] . Briefly, lipoproteins were precipitated by addition of samples to 0.05 M trichloroacetic acid (TCA) and 0.67% TBA in 2 M sodium sulfate. The union of lipid peroxide and TBA was performed by heating in a water bath for 90 min. The chromogen formed was extracted in n-butanol, which was measured at a wavelength of 535 nm. Lipid peroxidation was expressed as nanomoles of MDA per liter.
Measurement of Total Antioxidant Status
The total antioxidant status (TAS) is a sensitive and reliable marker to detect in vivo oxidative stress changes that may not be detectable through the measurement of a single, specific antioxidant. In this study, the TAS was evaluated by Trolox equivalent antioxidant capacity (TEAC). In this assay, 2, 2-azino-bis (3-ethylbenzothiazoline, 6-sulfonate) (ABTS2) is incubated with persulfate to produce ABTS+. This species is blue-green. Antioxidants present in the sample cause a reduction in absorption proportional to their concentration. The antioxidant capacities of the samples are expressed as TEAC using a calibration curve plotted with different amounts of Trolox and their absorbance measured at 740 nm [17] .
Determination of Nitric Oxide (NO)
The nitrate (NO 3 − ) present in the serum samples obtained from the patients was reduced to nitrite using nitrate reductase, and the nitrite concentration was determined by the Griess method [18] . Briefly, 100 µL of the serum supernatant was incubated with an equal volume of the Griess reagent for 10 min at room temperature. The absorbance was then measured on a plate scanner (Spectra Max 250; Molecular Devices, Menlo Park, CA, USA) at 550 nm. The nitrite (NO 2 − ) concentration was determined using a standard curve generated using sodium nitrate (NaNO 2 ). Nitrite production is expressed per µM.
Glutathione (GSH) Levels
Determination of the intracellular GSH levels was based on the ability of GSH to reduce 5,5-dithiobis-2-nitrobenzoic acid (DTNB) to nitrobenzoic acid (TNB), which was quantified by spectrophotometry at 412 nm. The methodology described by Ellman [19] was adapted for this determination, and GSH concentrations were expressed in µmol/mL. This assay was adapted for use in a microtitre plate using a microplate spectrophotometer system, Spectra MAX 250 (Molecular Devices, Union City, CA, USA) [20] .
Superoxide Dismutase (SOD) Activity
Determination of SOD activity was performed according to the procedure recommended by McCord and Fridowich [21] . This method evaluated the ability of SOD to catalyze the conversion of O 2 − to H 2 O 2 and O 2 . SOD activity was measured using UV spectrophotometry at a wavelength of 550 nm and was expressed in nmol/mL.
Catalase (CAT) Activity
Determination of CAT activity was performed according to the procedure recommended by Bleuter [22] . We evaluated the ability of the enzyme present in the sample to convert H 2 O 2 to H 2 O and O 2 . The decay of H 2 O 2 was measured using ultraviolet spectrophotometry at 240 nm, and CAT values were expressed as units per gram of hemoglobin (U/g protein).
Presence of Heinz Bodies
The precipitation of protein aggregates from the cytoplasmic membrane produces so-called Heinz bodies. Such inclusions can be observed under an optical microscope by staining with crystal violet or brilliant cresyl blue and visualized as globular inclusions of approximately 3 µm. Laboratory analysis for the quantification of Heinz bodies was based on the methodology employed by Rimiolli and Godoy [15] .
Statistical Analysis
Data are reported as the mean ± SD values. Statistically significant differences between groups were determined using Analysis of Variance (ANOVA) followed by Tukey multiple comparison tests. P<0.05 was considered statistically significant.
Molecular Docking
For studies of the interactions involving dapsone with the biological macromolecule human microsomal cytochrome P450 (CYP) 2C19, we conducted a computer assisted virtual molecular docking [23] , which is one of the techniques of molecular modeling, which consists of: predicting the bioactive conformation of a micromolecule (ligand) at the site of a biological macromolecule, followed by the evaluation and classification of the proposed binding mode [24] . The structure of the human microsomal CYP2C19 (4GQS, resolution 2.87 Å) obtained from the crystallographic protein data bank (PDB) and used in the study of docking was resolved experimentally by X-ray crystallography [25] . Prior to the study of virtual molecular docking, molecular mechanics calculations were performed with the three-dimensional structure developed for dapsone, based on the force-field MMFF94 [26] , so that there would be a better accommodation of the free atoms in the respective structure of dapsone. Computational tools for the performing of the virtual molecular docking differ in the manner of dealing with the flexibility of the enzyme, the ligand, its algorithm and scoring function [27] . In the present study, the molecular docking algorithm chosen was MolDock [23] . The study was performed in the flexible docking mode, using the scoring function of the docking algorithm, which is an extension of the linear potential by parts (PLP), the spatial adjustment of a wide variety of conformations of dapsone were tested at the catalytic interaction site of CYP2C19 and several docking solutions were generated automatically from the input structure of dapsone. The accuracy of the docking was further improved by using the algorithm rescoring function, which identified the most promising docking solution among the solutions obtained by the docking algorithm. We examined the complementarity of the interactions of the CYP2C19-DDS complex formed and identified the results that provided the best correlation between the poses and scores obtained.
Results and Discussion
Demographic and laboratorial characteristics of patients with leprosy
The final study cohort comprised 23 patients and 20 healthy individuals (control). Of these patients, 39% were classified as PB and 61% as MB. The twenty-three patients who initiated this research before MDT (called MDT0) were followed up until the third supervised dose of MDT (called MDT3). Most patients were female (54%) with age varying from 20–30 years (55%), 29% had a positive bacteriological index (BI) varying from degree 0.1–6, 39% presented disability (1 or 2 scores), and 28% had nerve damage (1 to 3 scores). All healthy participants did not show BI, degree of disability, or nerve damage ( Table 1 ).
10.1371/journal.pone.0085712.t001 Table 1
Demographic and laboratorial characteristics of leprosy patients and healthy individuals.
Variables
Leprosy Cases n = 23 (%)
Control n = 20 (%)
Gender (%)
Female
54
40
Male
46
60
Age
20–30
55
50
31–40
28
35
41+
17
15
Bacteriological Index
0
44
100
0.1–2.0
06
00
2.1–4.0
17
00
4.1–6.0
06
00
unknown
27
00
Degree of disability
0
44
100
1
17
00
2
22
00
unknown
17
00
Nerves affected
0
33
100
1
06
00
2
11
00
3
11
00
unknown
39
00
Dosage of DDS
After implementation of MDT (dapsone in combination with rifampicin and clofazimine) by WHO for the treatment of leprosy in order to monitor and achieve global elimination of leprosy, some adverse reactions, mainly caused by DDS, were observed in these patients [7] . The use of dapsone may cause oxidative stress leading to an imbalance between pro-oxidant and antioxidant agents [7] .
DDS mediated adverse reactions appear to be mainly due to its N-hydroxylated metabolite, DDS-NHOH and the toxicity is dose-dependent [7] , [28] . Thus, the therapeutic monitoring of drugs or their metabolites is essential to promote dose adjustment in order to control the therapeutic or toxic effects [28] . In this study, data showed that MB patients after the third dose supervised (MDT3) had a plasma concentration of dapsone of 0.518±0.029 µg/mL, while PB patients had 0.662±0.123 µg/mL. There was no significant difference between the dapsone concentrations in patients with different clinical forms (p>0.05, Table 2 ).
10.1371/journal.pone.0085712.t002 Table 2
Dapsone plasma concentrations determined in samples of leprosy patients in multi-and paucibacillary clinical forms according to the treatment time in months.
Treatment Time (months)
DAPSONE CONCENTRATIONS (µg/mL)
Multibacillary
Paucibacillary
MDT 0
ND
ND
MDT 3
0,518±0.029
0,662±0.123
ND – Not detected; Samples collected before administering the first dose MDT-supervised (MDT 0) and after two months in the third dose MDT-supervised (MDT 3).
These data were similar to that reported by Vieira et al. [28] , where 90% of patients presented therapeutic levels of dapsone of 0.5 to 5.0 µg/mL. Whilst some studies showed that concentrations of this drug in the plasma of leprosy patients are variable; they generally, remain within the accepted therapeutic range [7] , [28] . Moreover, the values obtained in this work do not correspond to toxic concentrations. According to Carraza et al. [29] , patients who had taken between 4 and 7.5 tablets of dapsone (100 mg each) had moderate to severe intoxication. In this regards, DDS levels for mild intoxication were found in patients who had average plasma concentrations up to 10 times the therapeutic level (1 µg/mL) of this drug, while 10 to 21 times (10–21 µg/mL) were moderate and over 21 times, constituted severe intoxication [30] . Although the liver is the major site of dapsone metabolism, hepatotoxicity has been observed only when the dose exceeded 300 mg/day [31] .
Oral DDS is absorbed readily from the gastrointestinal tract with bioavailability of more than 86% [32] . After absorption, the drug is transported through the portal circulation to the liver, where it is metabolized via N-hydroxylation, acetylation or glucuronidation. Moreover, the peak plasma concentration after 100 mg of oral dapsone is attained between 2 to 8 hours, and 85% of it is excreted in the urine, primarily as glucuronides, and 10% in the bile [32] , [33] . The long elimination half-life of dapsone averaging between 24 and 30 hours is thought to be due to several factors, such as significant enterohepatic recirculation, relatively high plasma protein binding (70–90%) of the drug and its acetylated metabolite (99%); indeed, the interconvertibility of the acetylated and parent forms also extends drug residence time [7] , [34] .
Biomarkers of oxidative stress
To determine oxidative stress in leprosy patients under treatment with MDT, we evaluated the systemic levels of nitric oxide, lipid peroxidation and MetHb as indicative of damage, and also the antioxidant status through the activities of SOD and CAT, GSH levels and capacity total antioxidant by TEAC.
Hydroxylamines can be formed from the parent and the acetylated derivative and they are potent oxidants which cause the hematologic adverse effects associated with dapsone, including methemoglobinemia and hemolytic anemia [7] , [10] , [34] , [35] , [36] . In this regard, leprosy patients presented with significantly enhanced MetHb percentage after the third dose supervised (MDT3), with values above the normal range (<2%). Basal levels were equivalent in untreated patients (MDT0) and healthy individuals ( Figure 2b ). In relation to the presence of Heinz bodies, Table 3 shows that only one of the untreated leprosy patients (MDT0 group) presented with Heinz bodies, with 1 body in each 500 cells counted, while an increased detection of the bodies was seen in nine (39%) of the MDT-treated leprosy patients (MDT3 group) (two or more bodies for each 500 cells analyzed).
10.1371/journal.pone.0085712.g002 Figure 2
Determination of nitric oxide (NO) in serum (a), percentage of blood methemoglobin (b) levels of malondialdehyde (MDA) in plasma (c) of patients with untreated leprosy (MDT 0) and after the third dose supervised treatment (MDT 3).
Figures in the chart are expressed as mean ± SD. *p<0.05 compared to MDT0 (ANOVA).
10.1371/journal.pone.0085712.t003 Table 3
Presence of Heinz bodies in blood smears of individuals in the study.
HEINZ BODIES
GROUP
Negative
Positive
Total
n (%)
n (%)
n (%)
Control
20(100)
00(00)
20(100)
MDT 0
22 (96)
01(04)
23(100)
MDT 3
14 (61)
09(39)
23(100)
Samples collected before administering the first dose MDT-supervised (MDT 0) and after two months in the third dose MDT-supervised (MDT 3).
Methemoglobin is generated from the oxidation of the heme group of hemoglobin to the ferric state, (Fe 2+ to Fe 3+ ), and thus cannot bind oxygen, levels which exceed 70% can result in patient fatality [37] . Oxidants such as DDS metabolites as well as forming methemoglobin, also generate other reactive species, such as superoxide and hydrogen peroxide. Indeed, the intracellular oxidative stress leads to the formation of soluble and insoluble denaturation products of hemoglobin. The most apparent insoluble products include Heinz bodies [38] , [39] reported that the long-term treatment only with DDS (100 mg daily) or MDT (dapsone in combination clofazimine and rifampicin) resulted in clinically significant hemolysis and MetHb, although the use of rifampicin and clofazimine, or clofazimine alone does not appear to increase the incidence of MetHb during treatment.
In agreement with our report, Dalpino et al. [40] showed that the percentage of MetHb of leprosy patients who were under treatment of dapsone (100 mg/day) was increased significantly compared to control. Regarding patient tolerance of DDS therapy, generally, patients with MetHb rates below 10% do not show any significant symptoms. However, a certain degree of hemolysis during DDS therapy appears inevitable, as alongside our present report, several previous studies have shown the presence of Heinz bodies during therapy with this drug [15] , [41] . In addition, Mahmud et al. [42] reported that the dapsone (100 microM) was the most potent former of methaemoglobin in human erythrocytes, and the substitution of the sulphone group with sulphur, oxygen, nitrogen, carbon or a keto group in drug resulted in a decrease in methaemoglobin formation.
Regarding nitric oxide production, we observed that their levels were significantly higher in patients of both groups MDT0 and MDT3 when compared to healthy individuals. However, there was no difference between MDT-treated and untreated patients ( Figure 2a ). These data showed that the disease process appears to be responsible for the detected NO increase in the body, as shown by our previous report [43] and other studies [44] , [45] .
Free radicals can cause cellular damage and undesired lipid peroxidation, which leads to the degradation of membrane lipid by free radicals produced by MDA. Serum/plasma levels of MDA provide some indication of the extent of f oxidative stress-related lipid peroxidation and cellular damage [46] . MDA was measured as an index of lipid peroxidation in the plasma of patients and healthy individuals. In this respect, our data showed that treatment with MDT did not alter the MDA levels in leprosy patients, as values were similar to those of healthy individuals ( Figure 2c ). In our report, we also observed that the MDA values in untreated or MDT-treated leprosy patients were similar to the control group ( Figure 2c ). These data suggest that MDT did not promote lipid peroxidation and cellular damage in the blood samples of leprosy patients. The elevated GSH contents observed in these patients may account for the normal levels of MDA in these patients after treatment with MDT ( Figure 3b ). In this regard, we observed similar levels between patients without treatment (MDT0) and healthy individuals. However, after treatment (MDT3 group), these patients showed significant increases in GSH values compared to the MDT0 group ( Figure 3b ). This protective effect of the thiol has also been demonstrated in animal models, where rats treated with 40 mg/kg of DDS yielded GSH and MetHb values significantly higher compared to control [47] .
10.1371/journal.pone.0085712.g003 Figure 3
Antioxidant capacity of blood samples from untreated patients with leprosy (MDT 0) and after the third dose of supervised treatment (MDT 3).
Concentration of superoxide dimutase (SOD) ( a ), reduced glutathione (GSH) ( b ), catalase (CAT) ( c ), Trolox equivalent antioxidant capacity (TEAC) ( d ). Figures in the chart are expressed as mean ± SD. *p<0.05 compared to MDT 0 (ANOVA).
GSH is a powerful antioxidant and a cofactor in many antioxidant enzyme reactions, so an increase in GSH levels may be relevant to the regulation of the activities of antioxidant enzymes that use GSH as a cofactor and thus the maintenance of oxidative balance [48] . In addition, some studies have shown that GSH is considered a potent inhibitor of lipid peroxidation process, and thus regulates the MDA content. One of the mechanisms by which GSH performs this protective function is by regulating the activity of GSH-dependent enzymes, such as peroxidases and peroxiredoxins, which result in reduction of intracellular oxidative stress followed by inhibition of the mitochondrial pathway of apoptosis induced by ROS [49] .
Concerning antioxidant enzymes, SOD activity was significantly reduced in samples from untreated leprosy patients (MDT0) and after the third dose supervised (MDT3) when compared to control group ( Figure 3a ). However, the basal CAT activity in untreated leprosy patients (MDT0) was similar to control group; its activity was significantly decreased by the treatment in leprosy patients ( Figure 3c ). The enzyme CAT is a tetrameric heme protein and mainly responsible for hydrogen peroxide degradation in aerobic and anaerobic organisms [50] . In summary, treatment with MDT led to a significant decrease in CAT activity in leprosy patients, but did not alter the SOD activity compared to untreated patients. These data were similar to other studies that reported that untreated leprosy patients have decreased levels of SOD compared to healthy individuals [51] , and that even after the use of MDT, the SOD levels remained low. These findings indicate that oxidative stress related to the reduction of antioxidants and free radical increase observed in these patients may also be caused by Mycobacterium leprae , as reported in a previous study [43] , [51] . Furthermore, studies reported that reduction of CAT activity may be associated with factors related to individuals, such as enzymatic deficiencies due to genetic mutations or a reduced synthesis of this enzyme by changes in their gene expression [52] . Many factors have been reported that can affect the gene expression of CAT, such as the presence of certain ions, cytokines and drugs [7] . In the case of infection by M. leprae , this agent requires ions and/or metals present in the host to regulate the expression of some of their resistance factors negatively affecting the supply of these compounds for the synthesis of metalloproteins such as CAT in the host organism [53] . On the other hand, the use of MDT has contributed to maintaining a framework of protection from oxidative stress due to host response to the infection process by M. leprae . In this sense, DDS metabolites possess oxidizing properties; amplifying the generation of reactive species, reducing of CAT activity and hemoglobin oxidation leading to formation of methemoglobin and Heinz bodies.
The ability of multidrug therapy to induce the production of free radicals can be compensated by antioxidant defense in leprosy patients. Antioxidants are best supplied by a balanced diet, but unfortunately leprosy patients are often of deprived socioeconomic status [54] . The analysis of the antioxidant capacity (AC) can provide some insight into the general biological antioxidant health the body, since it detects the presence of enzymatic and non-enzymatic antioxidants, instead of determining the concentrations of these antioxidants individually. In this sense, using the TEAC method for evaluation of AC has been recommended to evaluate various samples such as food, extracts and biological samples such as plasma [55] . In our study, we observed that the treatment with MDT did not alter the TEAC level in plasma of leprosy patients, which presented high levels similar to that of untreated patients ( Figure 3d ). Many proteins such as ceruloplasmin, transferrin, and small molecule antioxidants such as non-protein thiols, vitamins C and E, and uric acid contribute to the TEAC concentrations in plasma [56] .
Dapsone is oxidized by a number of CYPs [57] and it is likely that several are involved in its oxidation. Although only a single report has outlined the impact of CYP2C19 on dapsone oxidation, we explored the ability of the drug to bind to this CYP isoform, which is polymorphic and has significant clinical relevance in drug clearance [58] . Analysis of the internal cavities of CYP 2C19 showed two main contiguous cavities, one acting as the catalytic site ( Figure 4a ), which is in agreement with those described by Reynald et al. [25] . DDS (depicted in red) binds to amino acid residues and to the heme prosthetic group in the cavity of the catalytic site, just below the heme prosthetic group, which is depicted in green ( Figure 4b ). These two internal cavities are associated the following characteristics: volume 1271.04 Å 3 , surface 112.44 Å 2 , lipophilic surface 801.61 Å 2 , depth 28.16 Å 2 , ellipsoid main axis c/a ratio 0.22, ellipsoid main axis b/a ratio 0.80 ( Figures 4b and 5a ). The characteristics of the functional groups present in the cavities in relation to their relevant forces to molecular interactions were: 19 hydrogen bond donors, 68 hydrogen bond acceptors, 1 metal, 50 hydrophobic interactions and a hydrophobicity ratio in the order of 0.36. The amino acid composition of the cavities in relation to the polarity and electrical charge was distributed as follows: nonpolar amino acids ratio 0.57, polar amino acid ratio 0.31, positive amino acids ratio 0.03 and negative amino acid ratio 0.07.
10.1371/journal.pone.0085712.g004 Figure 4
Representation of the CYP 2C19.
Area covered (depicted as light blue transparent solid surface ) by the two internal cavities (catalytic site cavity and adjacent cavity) of human microsomal cytochrome P450 (CYP) 2C19 ( brown ). The heme prosthetic group is represented in green and the molecule of dapsone in the color red ( a ). Delimitation of the area occupied by the two internal cavities (catalytic site cavity and adjacent cavity) of the CYP2C19 ( b ).
10.1371/journal.pone.0085712.g005 Figure 5
Dapsone-CYP 2C19 complex.
Dapsone is shown in spheres and the target enzyme is colored from blue , starting from terminal N, through the red , up until terminal C ( a ). Dapsone-catalytic site CYP2C19 complex: Dapsone, the interacting amino acid residues and heme prosthetic group present in the catalytic site of CYP2C19 are displayed as sticks . The heme group and the dapsone are highlighted in a green contour ( b ). Coordination and interaction of dapsone when faced with the heme prosthetic group of the catalytic site of the CYP2C19 ( c ).
The results of the virtual molecular docking study conducted between dapsone and human microsomal CYP2C19, reveals the reactive groups present in dapsone as well as the forces and types of interactions existing between these reactive groups and the amino acid residues present in chain A of the cytochrome studied. Such data facilitates the understanding of the types of interactions and binding forces existing between them. The analysis of the interactions between dapsone and CYP 2C19 showed that DDS was located in the cavity of the reactive site, positioned near the iron atom of the heme prosthetic group ( Figures 5b–c ). The interactions which exist between DDS, the amino acid residues which contact it and the heme prosthetic group present in the catalytic site of CYP2C19 as well as their respective alpha-carbon positions in chain A are illustrated in Figure 5b .
The results of molecular docking analysis showed that several amino acid residues in the active site cavity interact with DDS. The main intermolecular interactions observed were the hydrophobic interactions, coexisting between dapsone and the respective amino acid residues Val113A, Phe114A, Ala297A, Thr301A and Leu366A which participate in the interactions, and also between dapsone and the heme prosthetic group, and the interaction by means of the hydrogen bond which exists between the hydrogen bond donating amino group (-NH2) in DDS and carbonyl group (CO) in amino acid residue Asp293A as the respecti8ve hydrogen bond acceptor ( Figure 6 ). It was evident also that the amino group attached to one of the aromatic rings and positioned nearest to the heme prosthetic group was able to interact with the existing iron atom in the heme group ( Figures 5b–c ). Our results provide some preliminary observations on the possible DDS oxidation mechanism by CYP 2C19 for the production of DDS-NHOH.
10.1371/journal.pone.0085712.g006 Figure 6
Interaction between dapsone, ligand-binding amino acid residues and the heme prosthetic group present in the catalytic site of CYP2C19 - The arrow represents the interaction between the hydrogen bond donating amino group (-NH2) in dapsone and the corresponding hydrogen bond acceptor (carbonyl group (CO) in residue Asp293A), dashed orange circular lines followed by dotted lines represent sites of hydrophobic interactions coexisting between dapsone and the respective binding residues (113A Val, Phe 114A, 297A Ala, Thr301A, Leu366A) and also with the heme prosthetic group.
Conclusions
In summary, our data showed that the treatment with MDT in leprosy patients led to a decrease in enzymatic antioxidant systems as CAT, alongside an increase in GSH, as well as a rise in MetHb levels and Heinz bodies' formations. This situation may occur in consequence to an oxidant/antioxidant imbalance systemically, which is caused by a combination of the infection process and multidrug therapy in leprosy patients. The analysis of the interactions between DDS and CYP2C19 has provided some perspective for future investigations which might explore the role of this CYP in dapsone oxidation in vivo, although it is likely that the drug is cleared by multiple CYP isoforms in vivo .
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Bioinformaticians have by now enjoyed almost two decades of publicly available protein comparison software and servers. In Cube, we shift somewhat the emphasis, and in addition to presenting our work in a way accessible to a bioinformatician, we address the needs of researchers who have no particular bioinformatics inclination, and for whom the sequence comparison is one of many steps in designing a biochemical or molecular biology experiment. In particular, Cube is structured to highlight the notion that conservation and specialization are two complementary pieces of information. Cube offers them for inspection side-by-side.
To place Cube on the map of the field, we first look at the biology involved, then discuss briefly how bioinformaticians detect and describe it, and how they disseminate their work.
Evolutionary behavior of biological sequences and the practical value of its analysis
Comparative analysis of DNA or protein sequences relies on an intuitively appealing mechanistic model of their evolution. It starts as a random process in which every region has an equal a priori chance of mutating. However, mutations that negatively impact a functionally important region get cleared out of the population.
Evolution will thus reduce the number of residue types observable at each position to the set which is allowable by the function. A thorough and illuminating analysis of the evolutionary process at work on the molecular level can be found in the body of work lead by J.H. Miller [1] , [2] . Nowadays, we can reproduce and trace the process in the lab [3] . Conversely, when we analyze conservation of residues or nucleotides, we are reverse engineering the nature-devised system, and looking for plausible functional explanation for why particular residues are conserved [4] .
Furthermore, noting that a prominent mechanism of genome evolution is gene duplication, we may enquire which of the copies (termed paralogues) changes to acquire new function [5] . We can look for residues that distinguish therwise similar groups of genes or proteins. These may, but do not need to be conserved in both paralogous groups [6] . After the gene duplication, the rate of evolution may stay the same in the two newly-founded branches (homotachy, in the fanciful terminology of [7] , or type II divergence [6] ), but is in general free to proceed at different rates (heterotachy, type I divergence). As a limiting case of the former, a position may be conserved as a different residue type in each of the branches (constant-but-different [8] , discriminant [9] ), or even, as a further extreme, conserved across two groups of related proteins. In any case, locating positions with markedly different evolutionary behavior in different paralogues can be used to understand and inform redesign of protein function [10] .
There are several practical problems to solve, though, to get meaningful results out of sequence comparison. Focusing on the word “conserved” one might note that it carries a hidden catch: it makes sense only when coupled with the definition of the set of sequences to which it applies. (Conserved in all protein kinases or conserved in CK1 group? Conserved in all vertebrates, or in mammals only?) The problem is twofold: we have to decide what defines the class of sequences within which we want to look for the conservation, and, then, we need to find those and only those sequences that belong to the class that we want to study.
While the patterns of conservation or specialization are not hard to appreciate once they are pointed out, they might be difficult to analyze systematically by a human observer - the alignment of one hundred vertebrate genes can easily approach a megabyte of data. Therefore, we would like to have ways to detect and classify of evolutionary behavior computationally.
Methods and their implementations, servers and databases of pre-calculated results
When bioinformaticians develop methods for detecting any particular type of evolutionary behavior, the fundamental way in which they present their work is by publishing the method - the scoring function or the algorithm. This is a compact way, usually involving some algebra, for explaining what the method does. At this point the methods may remain nameless. The names get attached later in the process - to the implementations, and even more often, to the servers. Implementation - the realization of the algorithm as a program is sometimes offered for download. If well written, this is the ultimate documentation for a method.
However, using an implementation directly is a task for aficionados. Servers provide shortcuts for a broader audience - they hide the implementational details from the user, and sometimes combine several sources of information. They differ widely in the way they present the output - from plain text tables that appear in the browser, to automatically generated printable reports and embedded visualization tools. It is notable however that the value expected to be added by the server increases as the field matures.
Sometimes the involved pipeline is so complicated, prone to breaking down, difficult to completely automate, or just time-consuming to complete, that the authors decide to present their results in the form of a database of pre-calculated results. The drawback of a database is that its content is fixed, and it does not allow the interested user to inquire how a change in the input data affects the offered conclusions.
Table 1 compiles (in an admittedly non-exhaustive way) method/server/database references for several notable takes on the protein sequence comparison. It also places Cube in its broader context.
10.1371/journal.pone.0079480.t001 Table 1
Comparison of several applications for comparison of protein sequences.
Name
Evolutionary behavior
Algorithm or method
Database
Server
Valdar
(degree of) variability
[11] , [12]
ScoreCons 1
rate4site, ConSurf
variability
[13]
[14]
[15] , [16]
AMAS, integer- and real-valued ET
variability
[17] – [19]
[20]
[21] , [22]
INTREPID
variability; type II div
[23]
[24]
FunShift
type I div
[25]
[26]
Diverge
type I and type II div
[6] , [27] , [28]
SDP
type II div
[29]
[30]
Treedet
type II div
[31]
[32]
SPEER
type II div
[13] , [33]
[34]
Multi-RELIEF
type II div
[35]
[35]
Capra & Singh
type II div
[36]
Cube
variability; type I and II div
[9] , [19]
[37]
this work
The table compiles the name under which a method is most often referred to, the type of evolutionary method it captures, and the references for the original (method) publication, as well as for the accompanying database and/or server publications, where applicable. “Variability” stands for the “degree of variability.” The table is not an exhaustive overview of the field, but, rather, illustrates the following. (i) Bioinformatics applications are usually presented as an algorithm and its application (third column), sometimes as a database of pre-calculated results, and sometimes as a server. Cube, described in this work, is a server. (ii) Furthermore, as of this writing, Cube is unique in that it provides a heuristic scoring both for the overall degree of variability, and for the type I and type II divergence. (iii) Type I divergence does seem to have the thinnest coverage in the literature, and is tackled by Cube. 1 http://www.ebi.ac.uk/thornton-srv/databases/cgi-bin/valdar/scorecons_server.pl .
Why Cube
It should be noted in the light of the above discussion that Cube is neither a method, nor a database. It is a server, using several methods to calculate on the spot conservation and specialization scores for the provided input. The drawback of this fact is that the users need to provide their own set of sequences for the analysis, which shifts part of the work on the users themselves. At the same time, this offers a possible advantage, because the users can provide the input from any kingdom of life, and group it according to any rule that may as well be unknown to the server. For the users working on vertebrate proteins, it might be of interest that Cube has a sister database of pre-calculated results, Cube-DB [37] , with the comparison limited to vertebrate sequences available in ENSEMBL [38] .
Behind the server are two pieces of code (available from the server's homepage) implementing several conservation detection methods [12] , [19] , [39] and one specialization detection method [9] . The specialization method implemented in Cube allows description of both divergence type I and type II events. Cube is a lightweight application with the aim of presenting our work in several formats that we have found to be practical in development and planning of experiments (mutagenesis experiments in particular): tabulation, mapping on the structure, and the sequence (by creating an image that can further be annotated). It leaves the user fully in control over the sequences that the analysis is based on. It is currently unique in that it places side-by-side and invites the contemplation of three types of evolutionary behavior: conservation and type I and type II specialization, conserved vs. determinant and discriminating residues.
We devote the following sections to more detailed description of methods and presentation of results in Cube.
Methods
Cube provides an interface to two scoring programs, one focusing on the conservation within a set of sequences, and the other on the specialization across several families. Rather than attempting to compound all the data - such as mutational propensity, spatial location, and biochemical properties of a residue - in a single score, we present them side by side, and let the user decide on their synergistic importance.
The scores implemented in Cube are all heuristics (to be distinguished from the algorithms that probabilistically model the underlying evolutionary process [6] , [13] ). They assign a single score to each position in the alignment, and assume the positions to be independent. They are “frequentist,” in that the inference is based on distribution of frequencies with which the amino acid type appears in the alignment column . In Cube, all scores are turned into ranks, which are in turn expressed as the top fraction they represent.
Conservation scoring
The user can choose between several heuristic, time-proven methods: real-valued ET [19] , and integer-valued ET [39] , majority fraction [40] , Shannon's information entropy, and Valdar's score, the last three described in [12] . All of these scores have the same common structure, where to the alignment position a value is assigned, such that . That is, the value of the score is a function of the frequency distribution of the amino acid types seen at this position. For example the majority fraction takes , the largest fraction seen at the position , and Shannon entropy takes to be .
Biochemical similarity of residues can be taken into account by using a reduced alphabet of amino acids, or by using BLOSUM [41] similarity in the case of Valdar's method. In these cases the function is parametrized in a way that depends on type similarity. This parametrization is independent of the position . Valdar's score is also the only one that attempts to correct for the uneven taxonomical sampling in the provided sequence set. rvET and ivET scores take the underlying similarity tree structure into account.
Specialization scoring
The specialization scoring is provided in two flavors. In the simpler approach, with the score termed “cube” and described in [9] , the positions are highlighted for which the overlap in distribution of amino acid types differs between the provided groups. This score is unaware of the possible relevance of biochemical similarity of some residues types. Alternatively, thus, the score that corrects for the effect is provided (“cube with similarity”). As in the case of conservation, the scoring function can be written as , the difference being that is now the function of distributions in protein groups, . The similarity is incorporated in the score by comparing the overlap with the expected overlap for (hypothetical) freely evolving residue distributions in the two groups. The scoring function does not use BLOSUM directly, but derives an evolutionary law for the distribution , such that after very long hypothetical time, every initial distribution converges to an equilibrium distribution which reproduces BLOSUM [42] . The overlap in residue type distribution between all group pairs is turned into two related but different pieces of information - discriminant and determinant score. The former rewards positions that are unique in one of the groups, while the latter seeks rarer cases in which a position is unique for each of the groups.
The scope and the limitations
The purpose of the methods implemented in Cube is to highlight residues exhibiting certain evolutionary behavior. The scores it uses are qualitative, and their absolute values carry no intrinsic meaning. Furthermore, the relative ranking of residues depends much more strongly on taxonomical sampling and the quality of the alignment, then on the precise choice of the method. In addition, when scoring the alignment positions the question of homology/orthology/paralogy arises. Faulty classification, again, may have more impact on the output than the method choice.
Implementation
The server is a mid-sized processing pipeline implemented in Perl/CGI/JavaScript, and was tested on all of the most popular web browsers. The scoring methods are implemented in C, and the code is available on the server's webpage.
Dependencies
Cube server uses MUSCLE [43] and MAFFT [44] to align sequences, and DSSP [45] to estimate the surface accessibility of individual residues. It also produces visualization for download, as a PyMOL session. [DeLano, W. (2002). The PyMOL Molecular Graphics System. ( http://www.pymol.org ). See also http://www.pymolwiki.org/index.php/Practical_Pymol_for_Beginners#Sessions .]
Results and Discussion
User's perspective
In designing Cube, we tried deliberately to keep it's interface lean. It has two main entry points. Starting from the dashboard page, the user can choose to do conservation or specialization analysis.
Conservation module
The only required input is a set of sequences in fasta format. Optionally, the sequences can be pre-aligned (the server accepts fasta and msf formats), and the reference sequence specified. In addition, the structure can be provided, and the default scoring method changed.
The server produces a 1D conservation map (the conservation score color coded and mapped on the sequence) in the png format, the tabulated information (in xls format), and the conservation mapped onto the structure (as a PyMol session, see the ‘Dependencies’ subsection in ‘ Methods ,’ above), Fig. 1 . A consistent color coding is used in all three forms of the output. The users are invited to provide any information that they already may have about the protein residues (such as transmembrane regions, post-translational modifications sites, catalytic residues and similar), numbered according to any sequence in the alignment. This information is added to the downloadable table, alongside the conservation score, residue type, and surface accessibility information.
10.1371/journal.pone.0079480.g001 Figure 1
Visualization in Cube.
Clockwise from top left: one dimensional map in png format, spreadsheet tabulation of conservation, specialization and annotation provided by the user, specialization mapped on the structure, and conservation mapped on the structure. The example shown: specialization between lysozyme C and -lactalbumin. (See http://eopsf.org/cube/help/worked_examples/spec_examples.html .).
When the structure (in PDB [46] format) is provided, the conservation score is mapped onto either the first chain or the user-specified chain in the provided PDB file. The server generates a PyMol session file in which the remaining peptide chains and ligands are indicated using a cartoon representation. From within the session, the poorly scoring residues can be hidden to emphasize the clusters of the most conserved residues.
Specialization module
The user is required to upload sequences already divided into meaningful groups. The groups can be arbitrary, but typically they are expected to represent paralogous families of proteins in comparable taxonomical samples, or protein orthologues divided into clearly distinct taxonomical groups.
In the output ( Fig. 1 ), the specialization scores are shown side-by-side with the conservation values (Shannon entropy) for each residue, both in the tabulated output (xls spreadsheet) as well as mapped on the structure (Pymol session). In the spreadsheet the results are laid out literally side-by-side in the adjacent columns. In the Pymol session, the menu on the right allows switching between the two views. The scores are also immediately shown in the browser, and available as a downloadable 1D map in the png format, and as an html version of the output table.
Documentation
The server comes with extensive help pages, worked examples, and on-the-spot help in the form of “mouseover” events, provided in the hope that it will find its place in biochemists', and molecular biologists' toolbox.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
There is an error in the first sentence of the Results section in the Abstract. The correct sentence is: Completion rates were significantly higher within families with higher education level (82% in tertiary educated families vs. 67% and 56% in secondary and primary educated families respectively) and were strongly correlated within families (ICC = 39.6) and neighbourhoods (ICC = 5.7).
There are multiple errors in Table 2 . For the “Neighbourhood variance (95% CI)” and “Family variance (95% CI) levels”, the proportional change in variance (PCV) and median odds ratio (MOR) values in Model 6 are incorrectly omitted. For the “Neighbourhood variance (95% CI)”, “Family variance (95% CI)”, and “Family variance >1 child*” levels, the median odds ratio (MOR) values for Models 1–6 are incorrect. Please see the corrected Table 2 here.
10.1371/journal.pone.0184231.t001
Table 2 The effects of parental education level, family structure and neighbourhood of residence, and its interactions on the probability of completing secondary education at age 21 among individuals born in the period 1983–1989.
Model 1
Model 2
Model 3
Model 4
Model 5
Model 6
OR
95% CI
OR
95% CI
OR
95% CI
OR
95% CI
OR
95%CI
OR
95%CI
Fixed effects
Individual level
Female
1.99
1.91–2.07
1.98
1.90–2.06
1.97
1.89–2.06
1.98
1.90–2.06
1.98
1.90–2.06
1.97
1.89–2.06
Teenage parent
0.08
0.07–0.10
0.09
0.08–0.10
0.09
0.08–0.10
0.09
0.08–0.10
0.09
0.08–0.10
0.09
0.08–0.10
Family level
Family education level
Primary
Ref
Ref
Ref
Ref
Secondary
1.77
1.68–1.86
1.58
1.50–1.66
1.60
1.51–1.70
1.51
1.37–1.66
1.48
1.36–1.61
1.44
1.27–1.62
Tertiary
4.22
3.95–4.51
3.44
3.23–3.66
3.69
3.44–3.97
3.23
2.94–3.55
3.03
2.27–3.35
2.97
2.63–3.37
Siblings
Only child
Ref
Ref
Ref
Ref
Ref
Ref
2–3
1.08
1.02–1.13
1.09
1.03–1.15
1.08
1.03–1.14
1.09
1.03–1.15
1.09
1.03–1.15
1.08
1.03–1.14
4+
0.83
0.77–0.90
1.00
0.92–1.08
0.99
0.92–1.07
1.00
0.93–1.08
1.00
0.93–1.08
1.00
0.92–1.08
Family living situation
Two parents at age 9 and 16
Ref
Ref
Ref
Ref
Ref
Ref
Both parent at age 9, one at age 16
0.42
0.40–0.45
0.52
0.49–0.55
0.59
0.53–0.66
0.52
0.49–0.55
0.52
0.49–0.55
0.59
0.53–0.67
One parent at age 9 and at age 16
0.32
0.31–0.34
0.50
0.47–0.53
0.53
0.49–0.58
0.50
0.47–0.53
0.50
0.47–0.53
0.53
0.49–0.58
Not living with parents at age 16
0.13
0.11–0.16
0.27
0.23–0.33
0.32
0.25–0.41
0.27
0.23–0.33
0.27
0.23–0.33
0.32
0.25–0.41
Maternal age at birth
<20
0.42
0.39–0.47
0.53
0.48–0.58
0.53
0.48–0.58
0.61
0.52–0.72
0.53
0.48–0.58
0.60
0.51–0.71
20–30
0.83
0.79–0.87
0.86
0.82–0.90
0.86
0.83–0.90
0.80
0.74–0.86
0.86
0.82–0.90
0.80
0.73–0.86
30+
Ref
Ref
Ref
Ref
Ref
Ref
Only one parent registered
0.93
0.79–1.08
0.85
0.72–0.99
0.86
0.73–1.01
0.85
0.73–1.00
0.85
0.72–0.99
0.86
0.82–0.96
Neighbourhood level
Urban settlement
0.94
0.90–0.99
0.97
0.93–1.02
0.97
0.93–1.01
0.97
0.93–1.02
0.89
0.83–0.96
0.89
0.82–0.96
Socioeconomic controls
Parental employment
Both parents in work
Ref
Ref
Ref
Ref
Ref
One parent in work
0.77
0.74–0.80
0.77
0.74–0.81
0.77
0.74–0.80
0.77
0.74–0.80
0.77
0.74–0.81
None parents in work
0.65
0.60–0.70
0.65
0.60–0.70
0.65
0.60–0.70
0.65
0.60–0.70
0.65
0.60–0.70
Poverty
0.42
0.39–0.44
0.42
0.39–0.44
0.42
0.39–0.44
0.42
0.39–0.44
0.42
0.39–0.44
Interactions with parental education level
Family education level * living situation
2 * Two parents at age 9, one at age16
0.88
0.76–1.01
0.87
0.76–1.01
2 * One parent at age 9 and at age 16
1.02
0.92–1.14
1.02
0.91–1.14
2 * Not living with parents at age 16
0.78
0.53–1.16
0.79
0.53–1.16
3 * Both parent at age 9, one at age 16
0.81
0.70–0.94
0.80
0.69–0.93
3 * One parent at age 9 and 16
0.82
0.73–0.91
0.81
0.72–0.91
3 * Not living with parents at age 16
0.66
0.42–1.06
0.67
0.42–1.07
Family education level * maternal age
2 * <20 years
1.09
0.98–1.21
1.09
0.97–1.21
2 * 20–30
0.83
0.68–1.03
0.83
0.67–1.03
3 * <20 years
1.12
1.02–1.24
1.15
1.04–1.27
3 * 20–30
0.71
0.56–0.91
0.78
0.61–1.00
Family education level * Urban
2 * Urban
1.10
0.99–1.21
1.09
0.99–1.21
3 * Urban
1.19
1.06–1.32
1.21
1.08–1.34
Random effects
Neighbourhood variance (95% CI)
0.13
0.11–0.16
0.11
0.09–0.13
0.11
0.08–0.13
0.11
0.08–0.13
0.11
0.08–0.13
0.10
0.08–0.13
PCV
-60.6%
-66.7%
-66.7%
-66.7%
-66.7%
-69.7
ICC(%)
2.7
2.23
2.23
2.23
2.21
2.21
MOR
1.41
1.37
1.37
1.37
1.37
1.35
Family variance (95% CI)
1.43
1.18–1.74
1.34
1.10–1.63
1.31
1.07–1.61
1.34
1.10–1.63
1.34
1.10–1.63
1.31
1.07–1.61
PCV a
- 35.9%
- 39.9%
- 41.3%
-39.9%
- 39.9%
- 41.3%
ICC (%)
29.5
28.3
27.9
28.3
28.3
27.8
MOR
3.13
3.02
2.98
3.02
3.02
2.98
Family variance >1 child *
Family variance (95% CI)
1.49
1.21–1.84
1.37
1.10–1.70
1.36
1.09–1.69
1.37
1.10–1.70
1.37
1.10–1.70
1.36
1.09–1.69
PCV
-34.7%
- 39.9%
- 40.4%
- 39.9%
- 39.9%
- 40.4%
ICC (%)
30.6
29.0
28.8
29.0
29.0
28.8
MOR
3.20
3.05
3.04
3.05
3.05
3.04
* Family level variance from the secondary analysis containing only families with more than one child ( S1 Table )
a The proportional change in variance expresses the change in variance at the particular level from the empty model
There is an error in S1 Table . For the “Neighbourhood variance (95% CI)” and “Family variance (95% CI)” levels, the median odds ratio (MOR) values for Models 1–6 are incorrect. Please see the corrected S1 Table here.
There is an error in S3 Table . For the “Family variance (95% CI)” level, the median odds ratio (MOR) values for Models 1–6 are incorrect. Please see the corrected S3 Table here.
Supporting information
S1 Table
The effects of parental education level, family structure and neighbourhood of residence, and its interactions on the probability of completing secondary education at age 21 among individuals in family groups of more than one child (N = 16,170)–three level logistic regression models estimated by mixed effects method, STATA/MP software.
(DOCX)
S3 Table
The effects of parental education level, family structure and neighbourhood of residence, and its interactions on the probability of completing secondary education at age 21 among individuals in family groups of more than one child (N = 16,170)–two level logistic regression models estimated by mixed effects method, STATA/MP software.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Tympanoplasty and mastoidectomy aided by microscopic techniques are accepted and routine methods for operating on middle ear structures in patients with chronic otitis media (COM) [ 1 ]. Despite its common application, using a microscope for these procedures is typically accompanied with limited visibility of various middle ear components including the hypotympanum, sinus tympani, and epitympanum, as well as the posterior part of the mesotympanum [ 2 , 3 ]. The ability to use both hands is one of the main advantages of this procedure, and the technique has improved the diagnostic performance of the procedure [ 4 – 7 ].
In addition to microscopy techniques, the application of flexible and rigid endoscopy has become usual for clinical evaluation of the structures of the middle ear. The assessment of these structures has been facilitated by endoscopy [ 7 ]. Comparing microscopic and endoscopic diagnostic approaches has revealed the superiority and feasibility of the latter method in evaluating middle ear pathological changes and structural abnormalities [ 8 , 9 ]. Some studies have reported successful surgeries of the middle ear including myringoplasty [ 10 ], surgery of the retraction pocket [ 11 ], stapedotomy [ 12 ], and removal of dermoid tumors of the Eustachian tubes[ 13 ] using the endoscopic approach. Because of some potential limitations of the endoscopic procedure including iatrogenic trauma, induced hyperthermia, and one-handed application, the use of this procedure is relatively uncommon in clinical settings [ 14 ].
Although the diagnostic performances of these two procedures have been widely assessed in various studies in different settings [ 15 – 17 ], to the best of our knowledge, there is insufficient evidence to compare the diagnostic performance of the procedures in COM patients. This study was conducted to compare the diagnostic performance of endoscopic and microscopic procedures in identifying the middle ear structures and associated diseases in patients with COM.
Materials and Methods
This prospective cohort study was conducted in the Otorhinolaryngology Department of Besat Hospital, Hamadan, Iran, from October 2011 to September 2012 with 58 consecutive COM patients, who were candidates for various types of tympanoplasty with or without a mastoidectomy. The Research Council for Research of Hamadan University of Medical Sciences approved the study. Written consent was obtained from the patients or their parents.
The baseline data on the demographic features, patients’ chief complaints (purulent discharge, hearing loss, or vestibular symptoms), duration of the disease, findings of the physical examination (Figs 1 and 2 ), results of tuning fork tests, and type of hearing loss (sensorineural, conductive or mixed) were collected. The candidates for second-stage hearing reconstruction or revision tympanoplasty with or without a mastoidectomy and the patients undergoing surgery by a non-post auricular incision were excluded from the study.
10.1371/journal.pone.0132890.g001
Fig 1
Preoperative transcanal microscopic view: (a) Cholesteatoma debris; (b) External auditory; canal bulging (obscures some part of the visual field); (c) Malleus.
10.1371/journal.pone.0132890.g002
Fig 2
Preoperative transcanal endoscopic view: (a) Malleus; (b) Tympanic membrane perforation (evident after the cholesteatoma removal).
Under general anesthesia, the middle ear was entered through a postauricular incision, and the tympanomeatal flap was elevated. Before the surgical intervention, the middle ear was examined with an operating microscope (Karl Storz, Germany with Sony 3CCD Color Video Camera, Japan) in different positions and in different bed positions. The visible anatomical areas were evaluated and recorded by performing gentle maneuvers on the patient`s head. Middle ear pathologies were explored with the identical technique, and the status of the ossicular chain was assessed as well. The middle ear was evaluated using a zero and 30-degree rigid endoscope (Karl Storz Image 1 HD H3 3-chip Camera Head and Diameter 4 mm, Work Length 18 cm, Karl Storz Image lens, Germany), and all of the components of the middle ear were assessed (Figs 3 and 4 ). The evaluations of the middle ear required approximately five minutes for each patient. Conventional middle ear surgery, using a microscope, was performed, and before the insertion of the tympanic membrane graft, the ear was reevaluated by endoscopy to detect any remaining disease. The exact anatomical sites were recorded, and a specimen was obtained for further pathological assessment, if any remaining disease were detected. Appropriate complementary surgical interventions were adapted for the effective eradication of detected pathologies. This second evaluation required approximately five minutes for each patient. The surgery was completed in a conventional method. All of the surgeries and microscopic and endoscopic evaluations were performed in real time by the same surgeon (a senior author, otologist and neuro-otologist, with 19 years of experience).
10.1371/journal.pone.0132890.g003
Fig 3
Preoperative middle ear microscopic view: (a) Scutum erosion; (b) Cholesteatoma and granulation tissue (around the stapes and facial recess); (d) Hypotympanum; (e) Tympanic membrane perforation; (f) Malleus.
10.1371/journal.pone.0132890.g004
Fig 4
Preoperative middle ear endoscopic view: (a) Cholesteatoma and granulation tissue (around the stapes and facial recess); (b) Eustachian tube opening; (c) Hypotympanum; (d) Tympanic membrane perforation; (e) Malleus; (f) Scutum erosion.
Real time
After collecting the data, the statistical analysis was performed using SPSS software (version 15.0, SPSS, Inc., Chicago, Illinois). The categorical variables were compared using a chi-square test. P values of 0.006 were considered to be statistically significant based on the Bonferroni correction.
Results
Included in the study were 58 consecutive patients, with an average age of 37.3 ±12.1 years (ranging from 15 to 63 years), 35 of whom were female. In 27 patients, the left ear was involved, and the right ear was involved in the others. Hearing loss, purulent discharge, and dizziness were the most common chief complaints of the patients. The average sensorineural hearing loss was 14.0 ±5.8 dB (ranging from 10 to 40), and the average conductive hearing loss was 31.9 ±9.3 dB (ranging from 15 to 50). The Rinne test was positive in 16 patients, and the Weber test was lateralized in 53 subjects.
The characteristics of the anatomical parts of the middle ear based on the microscopic and endoscopic findings are presented in Table 1 . According to the results of this table, the epitympanum (P = 0.015) and posterior mesotympanum (P<0.001) structures as well as most parts of the mesotympanum were significantly more visible through the endoscope than through the microscope.
10.1371/journal.pone.0132890.t001
Table 1 The characteristics of the anatomical parts of the middle ear based on microscopic and endoscopic findings.
Structure
Microscope, n (%)
Endoscope, n (%)
P value a
Epitympanum
18 (31.0)
31 (53.5)
0.015
Mesotympanum
Malleus
47 (81.0)
47 (81.0)
1.000
Incus
39 (67.2)
40 (69.0)
0.842
Stapes
38 (65.5)
47 (81.0)
0.059
Oval window
33 (56.9)
46 (79.3)
0.010
Round window
39 (67.2)
52 (89.7)
0.003
Promontory
58 (100.0)
58 (100.0)
1.000
Eustachian tube
30 (51.7)
53 (91.4)
0.001
Facial nerve
3 (5.2)
3 (5.2)
1.000
Posterior mesotympanum
Tympanic sinus
3 (5.2)
23 (39.7)
0.001
Hypotympanum
14 (24.1)
32 (55.1)
0.001
a The P value of the Bonferroni correction for eight multiple testings = 0.006
As shown in Table 2 , there was no significant difference in the evaluation of the ossicular chain mobility and the reflex of the round window between the microscopic and endoscopic approaches. In addition, according to the results presented in Table 3 , the diagnostic performance of both procedures was similar in identifying ossicular chain erosions.
10.1371/journal.pone.0132890.t002
Table 2 The ossicular chain mobility and reflexes of the round window in the microscopic and endoscopic views.
Structure
Microscope
Endoscope
P value
Good, n (%)
Reduced, n (%)
Good, n (%)
Reduced, n (%)
Malleus mobility
42 (89.4)
5 (10.6)
42 (89.4)
5 (10.6)
1.000
Incus mobility
32 (82.0)
7 (18.0)
33 (82.5)
7 (17.5)
0.958
Stapes mobility
32 (82.2)
6 (17.8)
40 (85.1)
7 (14.9)
0.909
Round window reflex
31 (79.5)
8 (20.5)
41 (78.8)
9 (20.2)
0.765
10.1371/journal.pone.0132890.t003
Table 3 Ossicular chain erosion in the microscopic and endoscopic views.
Structure
Microscope
Endoscope
P value
Healthy bone, n (%)
Erosive bone, n (%)
Healthy bone, n (%)
Erosive bone, n (%)
Malleus erosion
47 (81.0)
11 (19.0)
47 (81.0)
11 (19.0)
1.000
Incus erosion
31 (51.4)
26 (44.8)
32 (55.1)
26 (44.8)
0.976
Stapes erosion
37 (63.8)
14 (24.1)
44 (75.9)
14 (24.1)
0.705
At the end of the microscopic surgery and before the insertion of the tympanic membrane graft, the ear was reexamined with the endoscope to detect any remaining pathology. In 4 of the 13 patients with a cholesteatoma, the cholesteatoma had remained. This pathology was hidden in the sinus tympani in three patients and in the sinus tympani and epitympanum in another patient (Figs 5 and 6 ). In these patients, the surgery was continued for the effective eradication of the detected pathology. Granulation tissue was found in five patients. In one patient, it was detected in the medial surface of the scutum and could not be observed via microscope. Hypertrophic mucosa was found in 23 patients (most of them were in the hypotympanum), seven of which were detected by endoscopy. Tympanosclerotic plaque was found in 12 middle ears and could be seen through the microscope, except for one that was detected in the epitympanum. In two patients, there were polyps in the middle ear, both of which were observed via microscope. Other pathologies such as cholesterol granulomas were not detected.
10.1371/journal.pone.0132890.g005
Fig 5
Postoperative middle ear microscopic view after the atticotomy and cholestetoma removal.
(a) Scutum after the atticotomy; (b) Stapes capitalum; (c) Hypotympanic air cells; (d) Tympanic; (e) Membrane perforations.
10.1371/journal.pone.0132890.g006
Fig 6
Postoperative middle ear endoscopic view (evidence of a minimal remaining cholesteatoma in the tympanic sinus).
(a) Stapes capitalum; (b) The remaining cholesteatoma in the tympanic sinus; (c) Round window; niche; (d) Hypotympanic air cells; (e) Eustachian tube opening; (f) Tympanic membrane perforations; (g) Malleus.
Discussion
To compare the diagnostic performance of microscopic and endoscopic approaches in COM patients, we examined these approaches to identify the pathological or structural abnormalities in different parts of the middle ear. The results of this study showed that both methods were comparable in viewing ossicular chain mobility and reflex of the round window as well as viewing ossicular chain erosions. Various anatomical parts of the middle ear, particularly the epitympanum, posterior mesotympanum, and hypotympanum, were more visible via endoscope than microscope. Some pathologies, such as cholesteatomas, are potentially recrudescent if they remain in the middle ear. A cholesteatoma had remained in four of 13 patients. These abnormalities were hidden in the sinus tympani more frequently than in other areas. In these patients, the surgery was continued to eradicate the detected cholesteatomas. Although several studies have assessed the diagnostic accuracy of microscopic and endoscopic procedures separately, to the best of our knowledge, this study was the first in which the visibility of the two approaches were compared for diagnosing abnormalities of the middle ear.
Most of the failures in the post auricular surgical approach have been associated with difficulties in viewing different pathologies in more hidden pits of the middle ear such as the epitympanum, posterior mesotympanum, and hypotympanum [ 18 ]. One of the most appropriate approaches to locating these pathologies is a stepwise trans canal assessment of the tympanic membrane and the middle ear cavity, followed by eradication of the probable pathologies [ 3 ]. Some studies have shown that the areas that are more visible by endoscopy are those whose pathologies are hidden, such as cholesteatomas [ 4 , 19 ]. These areas, in which the pathologies might be hidden, are the epitympanum, sinus tympani, and hypotympanum [ 1 , 3 , 7 ]. Because of these limitations, for many years, surgeons have sought better tools to improve the visibility of the middle ear [ 20 – 22 ]. Accordingly, the endoscopic approach to exploration of the middle ear was suggested. Because of its limitations in middle ear surgery, it has not been widely accepted, and microscopic surgery remains the first choice method for surgical interventions in middle ear diseases [ 4 , 5 ].
Using endoscopy in middle ear surgeries has some limitations including the necessity using one instead of two hands [ 2 ], the creation of significant heat in the middle ear [ 2 ], and trauma to the middle ear because of undesirable hand movements[ 7 ]. To avoid damaging the middle ear structures and increasing morbidity, it is recommended that surgeons should not use an endoscope instead of a microscope in every ear surgery. Endoscopy could be used efficiently in particular situations such as in cases in which remaining pathologies (e.g., cholesteatomas) are suspected or if the posterior canal wall limits visibility in the confirmation of ossicular chain integrity. Another limitation of this study was that the raw data were not reevaluated by independent observer through a video review for checking reliability of the results.
Conclusion
In cases in which visibility by microscopy is disturbed and the surgeon suspects that pathologies remain in the middle ear, endoscopy could be utilized efficiently to improve the visibility and assessment of additional hidden middle ear pits and structures, particularly if there were a potentially recrudescent pathology.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Periodical organisms are very famous for their mass and synchronous reproduction, e.g., bamboos [1] – [3] , periodical cicadas [4] – [6] . The characteristic features of these organisms are the periodicity and synchronicity of reproductive events, e.g., periodical mass-flowering in bamboos [1] – [3] . Even though periodicity and synchronicity should be considered essentially two separate phenotypic traits, periodical organisms are usually associated with both periodicity and synchronicity [5] , [6] . Therefore, it is particularly important to investigate the evolutionary backgrounds of periodicity and synchronicity in these periodical organisms. Specifically, in periodical mass-flowering monocarpic (semelparous) plants, synchronous flowering is the key factor for pollination and seed predation avoidance [1] , [7] , [8] .
However, it is highly difficult, if not impossible, to investigate the relationship between periodicity and synchronicity in many periodical plants, because most these plants have extremely long life cycles (e.g., 15–120 years in bamboos). Based on the extensive studies, the confirmation of periodicity in periodical cicadas has been widely accepted [4] – [6] . However, the confirmation of periodicity in bamboos is still highly difficult [1] . It is also further difficult to evaluate whether these organisms are perfectly synchronized or not even from extensive field surveys. In bamboo species, particularly, the synchronicity of individual plants is almost impossible because we cannot identify the vegetative grown genets from other individuals [2] .
Instead of studying these long-cycle periodical organisms, we here investigated a periodical plant with relatively shorter life cycles (much less than 10 years). In the genus Strobilanthes (Acanthaceae), many species are reported to have periodicity with shorter life cycles, i.e., mass-flowering with various synchronous cycles (3–16 years), while other species flower every year [1] , [9] , [10] . We investigated the periodicity and synchronicity in supposedly periodical S. flexicaulis and its closely related S. tashiroi , both occurring on Okinawa Island, in the Ryukyu Islands, Japan. We also investigate the pollinator activities and predator satiation in S. flexicaulis to examine the selective advantages of mass-flowering.
Results
Field observations were carried out at Mt. Katsuu and Mt. Yae in the Motobu Peninsula, Okinawa Island from 1980. Mass-flowering of Strobilanthes flexicaulis was recorded at a six-year cycle in 1980, 1986, 1992, 1998, 2004, and 2010, while the rest 12 years were either few flowers or no flowers ( Fig. 1 ). The seedling experiments starting in 1998, 2004 and 2005 showed that 17 out of 20 seedlings flowered (and withered) at six years, and the rest three plants at seven years ( Table 1 ). Therefore, the interval of periodical mass-flowering had been estimated six years.
10.1371/journal.pone.0028140.g001
Figure 1
Mass-flowering years of Strobilanthes flexicaulis .
Flowering records of Strobilanthes flexicaulis at Mt. Yae and Mt. Katsuu are shown as mass flowering (large red circles), few flowers (small orange circles), no flowers (blue crosses), and non-mass-flowering (green triangles), when we have no records of flower counts. *This mass-flowering was observed by K. Nakajima (personal communication).
10.1371/journal.pone.0028140.t001 Table 1
Seedling experiments.
sowing year
individuals
6th year
7th year
1998
9
9
0
2004
8
5
3
2005
3
3
0
The seedling experiments starting in 1998, 2004 and 2005 were performed in the greenhouse in Medicinal Botanical Garden, Faculty of Pharmaceutical Sciences, Setsunan University. The numbers of flowering individuals in each year were shown.
Detailed examinations of the life history for S. flexicaulis were performed from 2008 to 2011 ( Table 2 ). We surveyed all flowering individuals in populations at Mt. Iyu, Mt. Oppa, and Awa. In all six populations, the number of flowering individuals was highest in 2010, indicating mass-flowering. Only few flowering plants are seen in 2009 or/and 2011 in four out of six populations (Mt. Iyu, Mt. Katsuu, Mt. Yae and Mt. Oppa). The populations at Mt. Yae and Mt. Katsuu show high synchronicity, more than 99% individuals flowered in 2010 with only a few one-year displacements ( χ 2 = 1.30, df = 1, P = 0.25; Text S1 and Table S1 ). Note that we exclude Mt. Iyu and Mt. Oppa from the current comparisons, since the numbers of plants at Mt. Iyu and Mt. Oppa are too small for statistical inference. In the other two populations (Mt. Nago and Awa), relatively large numbers of plants were flowered in before/after 2010. Thus the synchronicity is much lower at Mt. Nago and Awa compared with the other four localities. The chi-square test shows that the synchronicity of Mt. Nago is significantly lower than that of Mt. Katsuu and Mt. Yae ( χ 2 = 234.48, df = 2, P <0.01; Text S1 , Table S2 ). Furthermore, in Awa population, the number of flowering plants in 2011 (31 individuals) is about the same with that in 2010 (34 individuals) and 25 individuals are still not flowering at 2011 that are expected to flower next year if survived. The chi-square test shows that the synchronicity of Awa is significantly lower than that of Mt. Katsuu and Mt. Yae ( χ 2 = 827.58, df = 2, P <0.01; Text S1 , Table S3 ). Thus the flowering years are spread almost three consecutive years at Mt. Nago and Awa, showing low levels of synchronicity.
10.1371/journal.pone.0028140.t002 Table 2
Numbers of flowering individuals and survived individuals.
population
2008
2009
2010
2011
Mt. Iyu
flowering
–
0
15
0
survived/labeled
–
–
0/15
–
Mt. Katsuu
flowering *
0
11
1882
0
survived/labeled
–
0/11
0/21
–
Mt. Yae
flowering *
0
9
1043
1
survived/labeled
–
4/9
0/102
–
Mt. Oppa
flowering
–
1
>50
0
survived/labeled
–
–
0/18
–
Mt. Nago
flowering *
13
6
484
45
survived/labeled
0/13
0/6
0/30
–
Awa
flowering *
–
0
34
31
survived/labeled
–
–
0/16
–
The numbers of flowering individuals for Strobilanthes flexicaulis were enumerated in six populations from 2008 to 2011. We surveyed all flowering individuals at Mt. Iyu, Mt. Oppa, and Awa. The number of survived individuals/labeled individuals is shown in the second row of each locality, where some flowering plants were labeled and their survivals were examined in the following year.
*The chi-square test shows that the synchronicity of Mt. Nago and Awa was significantly lower than Mt. Katsuu and Mt. Yae ( Text S1 ).
In the total of six populations, we keep track of 241 plants by individual labeling ( Table S4 ). Out of the 241 individuals, 237 died after flowering ( Table 2 ). The remaining four individuals, in Mt. Yae population, had flowered partially on some branches. These branches had withered after flowering. These all four plants then had flowered in the following year and died after. Thus the death after flowering is a stable phenotypic character in S. flexicaulis even at the level of branches.
Comparing with S. flexicaulis , mass-flowering has never been observed in S. tashiroi . In the total of three populations, we labeled 108 individuals of S. tashiroi ( Table S4 ). In 2010, 76 labeled individuals were alive and the other 32 individuals were dead ( Fig. 2 ). Out of the 76 individuals, 31 flowered again. In 2011, 36 individuals survived and the other 40 individuals died. Six individuals bloomed for three consecutive years.
10.1371/journal.pone.0028140.g002
Figure 2
Life history of Strobilanthes tashiroi .
One hundred eight flowering individuals of Strobilanthes tashiroi were labeled in 2009. These individuals were pursued until 2011 and examined conditions (flowering, non flowering or dead) in each year. Three year consecutive flowering in S. tashiori clearly shows it is a polycarpic perennial plant.
We also compare pollinator activities with S. flexicaulis and S. tashiroi between the mass-flowering year and the previous/following years ( Table 3 ). Even though these species flowered in the winter season, in the mass-flowering year, three insect pollinators are seen (1.71 individuals/hour): humming-bird hawk moths ( Macroglossum corythus platyxanthum , 0.46 individuals/hour); honeybees ( Apis mellifela , 1.18 individuals/hour); and butterflies ( Byasa alcinous loochooana , 0.07 individuals/hour). In contrast, in the year before/after mass-flowering year, the same humming-bird hawk moth was observed to make a single visit during the total observation of 13.25 hours (0.09 individuals/hour). Thus the level of insect pollinations is extremely low.
10.1371/journal.pone.0028140.t003 Table 3
Pollinator observations.
year
time (hour)
Macroglossum corythus platyxanthum
Apis mellifera
Byasa alcinous loochooana
2009
10.75
0.09
0.00
0.00
2010
15.25
0.46
1.18
0.07
2011
2.5
0.00
0.00
0.00
Pollinator activities (observed individuals/hour) in Mt. Yae population before/during/after mass-flowering year are shown.
In order to evaluate the effects of predator satiation on the selective advantages of mass-flowering, we compared the percentage of fruits predated in S. flexicaulis and S. tashiroi between the mass-flowering year and the previous/following years ( Fig. 3 ). Fruit predation by the larvae of Pterophoridae sp. was observed in Mt. Yae population. Fruit predation rates in the mass-flowering year (2010) were significantly low in both S. flexicaulis ( χ 2 = 13.5, df = 1, P <0.01; Text S3 , Table S5 ) and S. tashiroi ( χ 2 = 39.4, df = 1, P <0.01; Text S3 , Table S6 ) with the chi-square test.
10.1371/journal.pone.0028140.g003
Figure 3
Surveys of fruit predation rates.
All the fruits on a plant before the capsules were exploded, were examined in Mt. Yae population at the end of March or the beginning of April from 2009 to 2011. The number of flowers examined was shown on each bar. * Fruit predation rates in the mass-flowering year (2010) were significantly low with the chi-square test ( Text S3 ).
Discussion
From 32-year data, it is confirmed that Strobilanthes flexicaulis is a periodical plant with a six-year mass-flowering cycle ( Fig. 1 ). Detailed observations of six populations during 2008 and 2011 revealed that all plants (branches) had died after flowering ( Table 2 ). These results suggested that S. flexicaulis was a periodical plant that maintained a periodical mass-flowering on Okinawa Island. However, the synchronicity of flowering is variable among local populations ( Table 2 ). Mt. Iyu, Mt. Yae, Mt. Katsuu and Mt. Oppa show high levels of synchrony, while there is much lower synchrony at Nago and Awa. These results indicate that the synchronicity in S. flexicaulis is relatively variable depending on the locality. Thus S. flexicaulis is confirmed a monocarpic periodical plant as in bamboos, but with some variation in the levels of synchronization.
Unlike S. flexicaulis , S. tashiroi was confirmed to be a polycarpic perennial plant; each plant produces flowers at least for several years without withering ( Fig. 2 ). No mass-flowering had been observed in S. tashiroi . Both S. flexicaulis and S. tashiroi belong to Parachampionella group, a small closely related taxonomic group consisting of four species [11] , [12] . These two species are genetically closely related [11] . This means that two distinctively different life histories, a polycarpic perennial plant and a monocarpic periodical plant, have been occurred in a small closely related taxonomic group. From the uniqueness and rarity of monocarpic periodical life history, we suspect that perennial life history is the ancestral trait for monocarpic periodical life history. Because of its shorter life cycle, and the variation of life history in its taxonomic group, S. flexicaulis could be a good model organism for the study of the evolution of monocarpic periodical life history, and periodical plants in general [1] .
For the evolution of periodicity and synchrony we need to evaluate two traits: (1) an internal clock, set for six years in this case, and (2) a selection to keep populations (cohorts) synchronous [1] . The latter synchrony comes from lower fitness of out-of-step plants (from higher selective mortality of seeds produced in off years, or from failure to pollinate if flowering in off years). A rarer, second out-of-step cohort growing at the same site, which would be expected to arise easily enough, would be easily eliminated due to lower fitness in the same manner, if selection favors synchrony [1] .
It is important to consider the evolutionary relationships between periodicity and synchrony [1] . Even though periodicity and synchrony are separate traits, masting or mass-flowering should be achieved by simultaneous evolution of synchrony and periodicity to some extent [7] , [8] . Koenig et al. [13] showed some selective advantages of simultaneous periodicity and synchrony in iteroparous plant species. Such selective advantages should be much higher in semelparous species with strict masting [7] , since all the adult plants die after flowering (reproduction), e.g., Bamboos [1] , Cerberiopsis [14] , [15] , Isoglossa [16] , Stenostephanus [9] , Strobilanthes [10] , Tachigali [17] , [18] . As Janzen [1] pointed out, outliers in strict masting are usually common, i.e., out of periods and asynchronous flowering. The current recorded data of S. flexicaulis shows many off-year flowerings. Our seedling experiments also show three out of 20 plants with one-year delay. Thus, in the current case the six-year periodicity is relatively well established, while synchrony still involves variation depending on the locality. We therefore suspect that the evolution of periodicity might have preceded the selection for synchrony in S. flexicaulis . There are several factors affecting the reproductive success of mass-flowering, e.g., pollination, seed predation [1] , [7] , [8] .
Mass-flowering may increase the efficiency of pollination [1] , [7] . It is often suggested that mass-flowering is highly effective in wind pollination [8] , [19] . In contrast, mass-flowering is suggested to have less or no effect under insect pollination [8] , [20] . Interestingly the current case might be an exception in insect pollination [7] , since this plant flowers in winter. Winter flowers have extremely few insect pollinators, often resulting in no pollinators in isolated flowers. We observed the number of pollinators in mass-flowering monocarpic S. flexicaulis and related polycarpic S. tashiroi before/during/after mass-flowering year at the same locality ( Table 3 ). We find almost no pollinators before and after mass-flowering, but many pollinators during mass-flowering. Pollinator activity may be highly crucial in winter flowering when the number of pollinators is extremely low. Mass-flowering may be able to attract these few pollinators; otherwise impossible.
Seed predation might be lowered significantly with masting [1] , [7] , [17] , [21] , [22] . We compare the rates of fruit predation in mass-flowering monocarpic S. flexicaulis and related polycarpic S. tashiroi before/during/after mass-flowering year at the same locality ( Fig. 3 ). In both S. flexicaulis and S. tashiroi , predation rates are low when mass-flowering, but much higher a year before and after. This suggests that mass-flowering at least reduces the level of predation considerably. Thus the selection favoring synchronicity may be promoted by pollinator activities [7] , [23] and predator avoidance [1] , [7] .
The lack of synchrony in some populations may be basically errors by individual plants, e.g., deleterious mutations and recombination. These out-of-synchronous plants should be weeded out by natural selection later. The current variation in the level of synchrony may be a result of such stochastic factors in small populations and intermittent errors in plant cycle length, that have been appearing temporally before the selection weeds out.
Materials and Methods
Ethics Statement
We were permitted to collect and observe in the Okinawakaigan Quasi-National Park from the Governor of Okinawa, and the Natural Conservation Zone of Mt. Katsuu, Mt. Awa and Mt. Yae from the Okinawa Prefectural Board of Education.
Species and morphology
Strobilanthes flexicaulis is a subshrub distributed in the Ryukyu Islands, Japan, and Taiwan [12] . Strobilanthes tashiroi is a perennial herb endemic to the Ryukyu Islands [12] . These species are closely related and morphologically very similar [12] , [24] . Around the Motobu peninsula on Okinawa Island in the Ryukyu Islands, S. flexicaulis and S. tashiroi grow sympatrically. We identified these species based on the bract shape and the difference in length between longer (anterior) and shorter (posterior) pairs of stamens, which are didynamous ( Text S2 , Figure S1 , Table S4 ).
Study sites and observation of life histories
We recorded mass-flowering years of S. flexicaulis and S. tashiroi around the Motobu Peninsula on Okinawa Island from 1980 to 2011. Quantitative research of mass-flowering species was performed in six populations ( Table S4 ). All flowering individuals of S. flexicaulis were enumerated in the definite zones along paths at Mt. Katsuu, Mt. Yae, and Mt. Nago from 2008 to 2011, and at Awa, Mt. Oppa, and Mt. Iyu from 2009 to 2011. The difference in the synchronicity among populations was checked by the chi-square test ( Text S1 ). Some representative flowering individuals of S. flexicaulis were labeled in winter from 2008 to 2010 ( Table S4 ). After flowering, life or death of labeled individuals was checked in the following summer. We labeled some flowering individuals of S. tashiroi at Mt. Nishime, Mt. Yae and Mt. Nago in 2008 and checked whether these labeled individuals survived and flowered, survived but did not flower, or died in the two successive years, 2009–2010 ( Table S4 ). The seedling experiments starting in 1998, 2004 and 2005 were performed in the greenhouse in Medicinal Botanical Garden, Faculty of Pharmaceutical Sciences, Setsunan University (Hirakata, Japan). A total of 20 individuals were sown and cultivated until flowering and dying.
Pollinator activities and fruit predation
Pollinator observations and surveys of fruit predation rates are performed in high synchronous Mt. Yae population from 2009 to 2011. Because pollinators were not seen in nighttime in preliminary observations, we observed pollinators in daytime in mostly S. flexicaulis in the mass-flowering year and mostly S. tashiroi in the previous and following years, as these are the available flowers in these years, respectively. To evaluate the fruit predation rates, we examined all the fruits on a plant before the capsules were exploded, at the end of March or the beginning of April from 2009 to 2011. The difference in the fruit predation rates in each species between the mass-flowering year and before/after was checked by the chi-square test ( Text S3 ).
Supporting Information
Text S1
The chi-square test was performed to examine the difference in synchronicity among populations.
(DOC)
Text S2
Strobilanthes flexicaulis and S. tashiroi were identified by the combination of the bract shape and the difference in length between longer and shorter pairs of stamens.
(DOC)
Text S3
The chi-square test was performed to examine the difference in fruit predation rates between the mass-flowering year and off years.
(DOC)
Figure S1
The bract shape (length/width) and the difference in length between longer and shorter pairs of stamens are shown. Blue circles are Strobilanthes flexicaulis individuals and red squares are S. tashiroi .
(TIF)
Table S1
The contingency table for the chi-square test to examine the difference in synchronicity between Mt. Yae and Mt. Katsuu.
(XLS)
Table S2
The contingency table for the chi-square test to examine the difference in synchronicity between Mt. Yae, Mt. Katsuu and Mt. Nago.
(XLS)
Table S3
The contingency table for the chi-square test to examine the difference in synchronicity between Mt. Yae, Mt. Katsuu and Awa.
(XLS)
Table S4
Localities and geographical coordinates of observed or/and sampling populations, where the extent of observed areas in each population and the sampling numbers of Strobilanthes flexicaulis and S. tashiroi in pure populations are shown.
(XLS)
Table S5
The contingency table for the chi-square test to examine the difference in fruit predation rate between the mass-flowering year and off years in Strobilanthes flexicaulis .
(XLS)
Table S6
The contingency table for the chi-square test to examine the difference in fruit predation rate between the mass-flowering year and off years in Strobilanthes tashiroi .
(XLS)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
In the Funding section, the grant number from the funder National Natural Science Foundation of China, China is listed incorrectly. The correct grant number is: 81200716. The publisher apologizes for the error.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Bladder cancer (BCa) is the most common neoplasm of the urinary tract and the fifth most prevalent malignancy worldwide. High-grade bladder tumors are more likely to progress to muscle-invasive disease and have a higher tendency to undergo distant metastasis. In contrast, low-grade tumors rarely invade the bladder musculature and metastasize [ 1 – 3 ]. However, both non-muscle-invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC) have a high propensity of recurrence, with a 50–90% probability of recurrence within five years [ 4 ]. In addition, metastatic BCa is considered incurable [ 5 ]. Several established risk factors related to higher risks of disease progression include tumor grade, tumor size, and tumor multiplicity; however, these risk factors are insufficient to address important prognostic indicators such as recurrence rates and progression in individual cases [ 6 , 7 ]. Thus, there is an urgent need to identify novel, more reliable prognostic factors for BCa.
Cancer stem cells (CSCs), also known as tumor-initiating cells, are a subpopulation of undifferentiated yet tumorigenic cells within a neoplasm that are capable of tumor initiation, self-renewal, and proliferation, which are thought to be responsible for tumor progression, relapse, metastasis, and heterogeneity [ 8 , 9 ]. CSC expressions have been identified in multiple human solid tumors, including breast, prostate, ovarian, and lung cancers, and are significantly associated with metastasis-free survival and other clinical outcomes [ 10 ]. Bladder cancer stem cells (BCSCs) were first identified using markers for isolation of normal stem cells in 2009 [ 11 ]. Since then, BCSCs have emerged as a growing field of research, with genome-wide screening methods and platforms for establishing therapeutic targets for tumor-initiating cell populations [ 12 ]. A more profound understanding of BCSCs and their effects on BCa may provide helpful prognostic tools and novel therapeutic targets. However, the clinical impacts of BCSC expressions and functions have not been fully elucidated yet. Hence, this systematic review aims to evaluate all available evidence regarding BCSCs and their roles in predicting the risks of metastasis and recurrence in BCa.
Materials and methods
Objectives
This article aims to provide a systematic review of primary clinical studies to identify the BCSCs markers, which have played a role as prognostic factors in BCa patients.
Study design
This systematic review was created in accordance with the guidelines for Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) [ 13 ]. We determined inclusion criteria, data synthesis methods, and outcomes in advanced in a protocol registered with PROSPERO (CRD42021268964).
Search strategy
A literature search for clinical studies evaluating CSCs as a prognostic indicator in BCa published from January 2000 to February 2022 was conducted from several databases, such as Pubmed, Scopus, EMBASE, Science Direct, Proquest, CINAHL, and The Cochrane Library. The search following key terms used based on the PICO were applied to identify eligible publications: (“Bladder Cancer” OR “Transitional Cell Carcinoma” OR “Urothelial Carcinoma”) AND (“Stem Cell” OR “Stem Gene”) AND (“Metastasis” OR “Recurrence”). Initially, study titles and abstracts were screened. Subsequently, full text analysis of selected articles was done based on pre-set eligibility criteria. In addition, the reference lists of included studies were further evaluated to identify potential studies. Literature screening and analysis were undertaken separately by two independent researchers.
Eligibility criteria
Our inclusion criteria were as follows: (1) studies evaluating the impacts of BCSC expressions on BCa recurrence and/or metastasis; (2) prospective or retrospective cohort studies or case control studies; and (3) publications written in English. Animal and in vitro studies were excluded.
Outcomes
The primary outcomes of this systematic review are the effect of BCSC expressions toward recurrence and metastasis in BCa. In addition, survival analyses from several studies were also presented.
Data extraction
Two authors independently run the systematic search and screened the articles. From all eligible studies, data were also extracted independently, and any disagreement were resolved through discussion among all authors. Data recorded from each study were as follow: author’s name, year of publication, study design, number of study’s participant, intervention given to the participant, method used for gene expression analysis, outcomes (recurrence-free survival and metastasis-free), and mean or median year follow-up. The effect measures used were the hazard ratios and their respective 95% confidence intervals for both univariate and multivariate analyses.
Quality assessment
The qualities of the selected studies were assessed using the Newcastle-Ottawa Scale (NOS). Using this tool, selected studies were assessed based on three aspects: the selection of the study groups; the comparability of the groups; and the ascertainment of either the outcome of interest for case-control or cohort studies, respectively. Good quality studies have 3 to 4 stars in the selection component, 1 to 2 stars in the comparability component, and 3 stars in the outcome component. Fair quality studies have 2 stars in the selection component, 1 to 2 stars in the comparability component, and 2 to 3 stars in the outcome component. Poor quality studies have 0 to 1 star in the selection component, 0 star in the comparability component, and 0 to 1 star in the outcome component. The quality assessment showed in Table 1 .
10.1371/journal.pone.0269214.t001
Table 1 Risk of bias assessment using Newcastle Ottawa Score (NOS).
No
Study
Selection (Max *)
Comparability (Max **)
Outcome (Max *)
Score
Representativeness of exposed cohort
Selection of exposed cohort
Ascertainment of exposure
No outcome of interest at start
Comparability of cohorts based on design or analysis
Assessment of outcome
Was follow up long enough for outcomes to occur
Adequacy of follow up of cohorts
1
Ruan et al, 2012 [ 6 ]
*
*
*
*
*
*
*
*
8
2
Keymoosi et al, 2014 [ 13 ]
*
*
*
*
*
*
*
*
8
3
Wei et al, 2015 [ 14 ]
*
*
*
*
*
*
*
*
8
4
Senol et al, 2015 [ 15 ]
*
*
*
*
*
*
*
*
8
5
Sedaghat et al, 2016 [ 2 ]
*
*
*
*
*
*
N/A
*
7
6
Siddiqui et al, 2019 [ 16 ]
*
*
*
*
*
*
*
*
8
7
Chiu et al, 2020 [ 17 ]
*
N/A
*
*
N/A
*
N/A
*
5
8
Shen et al, 2015 [ 1 ]
*
*
*
*
**
*
*
*
9
9
Xu et al, 2015 [ 18 ]
*
*
*
*
*
*
*
*
8
10
Heo et al, 2020 [ 19 ]
*
*
*
*
*
*
N/A
*
7
11
Shaifei et al, 2019 [ 20 ]
*
*
*
*
*
*
N/A
*
7
12
Kallifatidis et al,2019 [ 21 ]
*
*
*
*
*
*
*
-
7
Result
Study selection
The flow diagram in the form of PRISMA diagram for study selection is shown in ( Fig 1 ). We included 12 clinical studies evaluating effects of BCSCs expression on tumor recurrence and/or metastasis, consisted of cohorts and case controls studies, involving at least 2230 patients (one study did not specify the sample size) with BCa and 68 non-tumor tissue for control in this systematic review. All the eligible studies were published between 2012 and 2020.
10.1371/journal.pone.0269214.g001
Fig 1
Study Selection using PRISMA flow diagram.
Articles were identified and screened for eligibility, 12 clinical studies were included.
Study characteristics
All of the studies selected were reviewed and the result were displayed on Table 1 . We reviewed the intervention given to the patient, gene expression analysis, and outcomes which was consisted of recurrence and metastasis. Eleven out of 12 studies assessed the recurrence-free survival related to the BCSCs and five studies out of 12 studies assessed the metastasis associated with BCSCs. Only four studies which analyzed both tumor recurrence and metastasis. Only three studies provided data about mean or median follow-up time Table 2 .
10.1371/journal.pone.0269214.t002
Table 2 Study characteristics.
Study (year)
Design
Participants
Control
Intervention
Gene Expression Analysis
Outcomes
Mean/median follow-up
Recurrence
Metastasis
Ruan et al., 2012 [ 6 ]
Case-control
32 BCa tissues
32 corresponding normal tissues
-
qRT-PCR IHC
High SOX2 expression significantly played a role in predicting the recurrence-free survival in T1 BCa patients
-
2 years
Keymoosi et al., 2013 [ 13 ]
Prospective Cohort
159 patients with urothelial carcinoma
-
TURBT with no prior chemotherapy or radiation therapy
IHC on TMA slides
High ALDH1 and CD44 expressions were correlated with a significantly increased rate of recurrence (P = 0.013)
-
46 months
Wei et al., 2015 [ 14 ]
Prospective cohort
130 BCa patients
-
Cystectomy or Transurethral resection of the bladder tumor.
qRT-PCR IHC
High Cripto-1 was significantly associated with expression and tumor recurrence or metastasis (P = 0.007)
-
Not available
Senol et al., 2015 [ 15 ]
Prospective cohort
163 cases of urothelial carcinomas of the bladder (UCB)
-
-
IHC
ALDH1 expression was significantly associated with disease recurrence (P<0.001), however, CD44 was not significantly associated (P = 0.688)
-
23.60±16.88 months
Sedaghat et al., 2016 [ 2 ]
Retrospective cohort
140 tissues from transitional cell carcinoma samples
-
-
TMA-based IHC
OCT4 expression had no correlation with tumor recurrence (P = 0.32) or CD133 (p = 0.71)
-
Not available
Siddiqui et al., 2019 [ 16 ]
Prospective cohort
112 histopathologically proven BCa
Bacillus Calmette Gurein (BCG), non-BCG, radical cystectomy with and without adjuvant therapy
IHC
High CD44 and NANOG expression were significantly associated with lower tumor recurrence (P<0.001)
-
Not available
Chiu et al., 2020 [ 17 ]
Retrospective cohort
For patients with transitional cell carcinoma of the urinary bladder, sample size not specified
-
-
IHC staining with SOX2 antibody
High SOX2 and IGF1R expression was correlated with poor recurrence-free survival and was increased in "poorly differentiated" malignant grade tumors (P = 0.0187)
-
Not available
Shen et al., 2015 [ 1 ]
Retrospective cohort
309 patients with transitional cell carcinoma of the urinary bladder
-
-
IHC
High Sox4 expression was significantly associated with higher tumor grade (more likely to recurrent). (P = 3.71E-10)
High Sox4 expression was significantly associated with invasiveness (more likely to spread to other parts of the body). (P = 7.00E-04)
Not available
Xu et al., 2015 [ 18 ]
Prospective Cohort
227 patients with bladder urothelial cell carcinoma (118 non-invasive and 109 invasive)
-
118 patients with non-invasive bladder carcinoma: 11 underwent radical cystectomy and 107 underwent intravesical chemotherapy after transurethral resection 109 patients with invasive disease: 69 underwent radical cystectomy. 20 underwent partial cystectomy; and 20 underwent transurethral resection
IHC using ALDH1A1 antibody and secondary antibody from EnVision System
ALDH1 expression was significantly associated with tumor recurrence (P ≤ 0.05).
ALDH1 expression was significantly associated with lymph node (P = 0.008) and tumor distant metastases (P = 0.018)
52-months
Heo J et al., 2020 [ 19 ]
Retrospective cohort
400 patients with urothelial carcinoma
-
TURBT
IHC
p-TFCP2L1 and CDK1 expression were not associated with recurrence (P = 0.563)
High levels of co-expression of p-TFCP2L1 and CDK1 were associated with distant metastasis (P = 0.442)
Not available
Shaifei et al., 2019 [ 20 ]
Case-control
472 bladder tumors
16 matched adjacent non-cancerous normal tissue
TURBT with no prior neoadjuvant treatment before surgery
IHC on TMA slides
DCLK1 expression was not associated with recurrence (P = 0.314)
DCLK1 expression was significantly associated with distant metastasis (P = 0.042)
Not available
Kallifatidis et al., 2019 [ 21 ]
Retrospective cohort
43 bladder tumors in cohort 1; 43 bladder tumors in cohort 2
20 normal bladder
Cohort 2 receiving gemcitabine + cisplatin
qRT-PCR
ARRB1 transcript levels in bladder tumor specimens from patients who developed metastasis were 7.7-fold elevated compared to the normal bladder and 5.2-fold elevated compared to BCa specimens from patients who did not develop metastasis
Not available
ICH, Immunohistochemistry; NMIBC, non-muscle-invasive bladder cancer;(tissue microarray); qRT-PCR, Real-Time Quantitative Reverse Transcription Polymerase Chain Reaction; TURBT, Transurethral Resection of Bladder Tumor
Association between BCSCs expression with clinicopathological parameters
We identified several BCSCs in this review, which included SOX2 (Sry related HMG-Box 2), SOX4 (Sry related HMG-Box 4), ALDH1(Aldehyde-dehydrogenase 1), CD44, Nanog, Cripto-1 (Cysteine rich dommain), OCT4 (Octamer Binding Transcription Factor 4), CD133 (Prominin 1), β-arrestin-1 (ARRB1) and β-arrestin-2 (ARRB2), IGF1R, p-TFCP2L1 (Transcription Factor CP2-like Protein 1), and CDK1 (Cyclin Dependent Kinase 1). Eight out of 12 studies included performed Kaplan-Meier survival analysis to compare the effect of respective BCSCs expression on clinicopathological parameters. All of the genes were found to be significant prognostic factors based on univariate analysis. Moreover, multivariate analysis using Cox regression also showed that the majority of gene expression were independent prognostic factors; thus, it may play a role as a potentially valuable marker in predicting the recurrence-free, metastasis-free, or disease-free (recurrence/metastasis-free) with P<0.05 Table 3 .
10.1371/journal.pone.0269214.t003
Table 3 Univariate and multivariate analysis of recurrence and metastasis as prognostic factors in patients with bladder carcinoma.
Outcome
Gene expression
Study (year)
Univariate analysis
Multivariate analysis
HR (95% CI)
P value
HR (95% CI)
P value
Recurrence-free
SOX2
Chiu (2020)
2.467 (1.292–4.709)
0.0062
2.966 (1.451–6.064)
0.0029
Ruan (2013)
4.2 (1.827–9.654)
0.001
3.187 (1.130–8.990)
0.029
ALDH1
Xu (2015)
2.84 (1.19–7.14)
0.040
-
-
CD44/Nanog
Siddiqui (2019)
32.52 (9.79–107.99)
<0.001
25.45 (6.71–96.50)
<0.001
ALDH1
Senol (2015)
-
-
4.590 (2.042–10.319)
0.001
CD44
-
-
0.548(0.283–1.059)
0.074
Metastasis
ARRB1
Kallifatidis (2020)
1.35 (1.06–1.71)
0.0137
1.07 (1.01–1.13)
0.015
ARRB2
0.03 (0.35–0.003)
0.005
0.13 (0.86–0.02)
0.006
Disease-free (recurrence/metastasis )
Cripto-1
Wei (2015)
2.678 (1.280–5.605)
0.009
2.306 (1.055–5.039)
0.036
DCLK1
Shafiei (2019)
1.642 (1.063–2.534)
0.025
1.564 (1.004–2.434)
0.048
The expression of SOX 2 was significantly correlated with poorer recurrence free prognosis in the studies by Chiu (P = 0.0062 Univariate and P = 0.0029 Multivariate) and Ruan et al. (P = 0.001 Univariate and P = 0.029 Multivariate) [ 6 , 17 ], Similarly, ALDH1 was also shown to be significantly associated with poorer recurrence free survival with a univariate P value of 0.04 and a multivariate P value of 0.001 from the studies by Xu and Senol et al. respectively [ 15 , 18 ]. In contrast, studies assessing the expression of CD44 with recurrence free survival reported conflicting findings, with Siddiqui et al. reporting univariate and multivariate P values of <0.001 whereas Senol et al. reported a multivariate P value of 0.074 [ 16 ].
With regards to incidence of metastasis, Kallifatidis et al. reported that expressions of both ARRB1 and ARRB2 were significantly associated with increased metastasis with univariate findings of P = 0.0137 and P = 0.005 and multivariate findings of P = 0.015 and P = 0.006 respectively [ 21 ].
Both recurrence and metastasis were significantly marked in patients expressing Cripto-1 in a study by Wei et al. with results from univariate analysis showing P = 0.009 and multivariate analysis P = 0.036 [ 14 ]. Along the same line, DCLK1 was also demonstrated to be significantly associated with increased recurrence and metastasis with univariate and multivariate results showing P = 0.025 and P = 0.048 respectively in a study by Shaifei et al. [ 20 ].
Discussion
Cancer cells utilize normal stem cell self-renewal for long-term proliferation and tissue-repair pathways for invasion; therefore, CSCs expression in cancer may be associated with disease prognosis and treatment outcomes [ 1 ]. Association between tumor biology and CSCs has been addressed in various types of cancer, including breast cancer, colorectal cancer, and bladder cancer. Many studies have revealed that CSCs were considered as an important factor leading to tumor recurrence and metastatic; however, its exact mechanisms are still unclear and may have a different pathway one to another [ 22 ]. By identifying and understanding the molecular mechanism of recurrent and metastatic BCa, numerous stem cell phenotypes, such as beta arrestins, SOX2, SOX4, transcription factor CP2 like 1 (TFCP2L1), and doublecortin-like kinase 1 (DCLK1), have all been identified and described.
Stem Cells govern homeostasis in human tissue. Specific gene expressions are regulated by transcription factors (TFs) and chromatin regulatory proteins. Embryonic stem cells express TFs such as Octamer-binding transcription factor 4 (OCT-4), Nanog Homeobox (Nanog), SRY-box2 (SOX-2), TFCP2L1, and Sry-Related HMG-BOX-4 (SOX4). These TFs are not expressed in already differentiated somatic cells; they suppress stem cells from differentiating. In patients with carcinoma, irregular gene activation causes stem cell proliferation. Cripto-1 is an example of an embryogenic gene overexpressed in bladder cancer.
Since multiple genes related to cancer are already identified in The Cancer Genome Atlas (TCGA), a model proposed to comprehend molecular characterization of muscle-invasive bladder cancer. Robertson et al. proposed a division between 5 subtypes of bladder cancer based on mRNA expression clustering: (1) luminal-papillary, (2) luminal-infiltrated, (3) luminal, (4) basal/squamous, (5) neuronal. Luminal papillary subtypes identified with FGFR3 mutations, TACC3 fusions, or papillary histology amplification. This subtype responds poorly to cisplatin-based therapy. However, Tyrone kinase inhibitor of FGFR3 may be beneficial, as proved by early clinical trials. Luminal-infiltrated subtype identified with the expression of miR-200, EMT, and myofibroblast markers. This subtype may also be resistant to cisplatin-based therapy and thus can be beneficial as a negative predictive biomarker for chemotherapy response. The luminal subtype has a high expression of KRT20 and SNX31. The therapy targeting these specific mutation profiles may be beneficial. The basal-squamous subtype was identified with increased expression of CD274 (PD-L1) and CTLA4 immune markers. Therefore, immune checkpoint therapy and cisplatin-based therapy are appropriate options. The neuronal subtype is identified by the expression of neuronal genes and neuroendocrine. Etoposide-cisplatin therapy is proposed to be beneficial for this variant [ 23 ].
By identifying and understanding the molecular mechanism of recurrent and metastatic BCa, numerous stem cells phenotypes, such as beta arrestins, SOX2, SOX4, transcription factor CP2 like 1 (TFCP2L1), and doublecortin-like kinase 1 (DCLK1), have all been identified and described.
Kinase is an enzyme that catalyzes the transfer of phosphate groups. Mutation of kinase can cause cellular irregularities and lead to abnormal growth. Cyclin-dependent kinase 1 (CDK1) is a protein that functions as a threonine protein kinase for cell cycle regulation. However, CDK1 facilitates phosphorylation of TFCP2L1, activating embryonic stem cells in bladder cancer and driving tumorigenesis [ 19 ].
Two studies reporting a significant relationship between high expression of SOX2 with poor recurrence-free survival also found that SOX2 was highly expressed in tumors with poor pathological differentiation; thus, marking its role in BCa malignancy. SOX2 plays a role in promoting cell proliferation and enhancing cell survival during low-serum stress. BCa cancer cells’ survival and spheroid-forming capability enhancement were induced by AKT phosphorylation due to IGF2/IGF1R induction, which was thought to be involved in molecular mechanism of SOX2 expression leading to poor tumor prognosis. Its mechanism made SOX2 a potential therapeutic target for BCa treatment [ 6 , 17 ].
CD44 was one of the most stem cells which has been widely studied and was commonly expressed in BCa with a poor prognosis. Hu et al . [ 19 ] conducted a meta-analysis about the prognostic value of CD44 expression in BCa and found that CD44 expression may be associated with advanced T stage, tumor grade, and lymph node metastasis, but not with recurrence-free survival and overall survival of the disease. ALDH1 was also commonly reported to be a significant prognostic factor in tumor recurrence and metastasis. Xu et al . [ 18 ] and Senol et al . [ 15 ] further conducted univariate and multivariate survival analyses and also observed a statistically significant association between ALDH1 expression and recurrence-free survival (P<0.05). Moreover, Xu et al . also found that ALDH1 expression was related to distant tumor metastasis [ 18 ].
An inverse expression of ARRB1 and ARRB2 both significantly correlated with tumor metastasis. Kallifatidis et al . [ 21 ] conducted univariate and multivariate analysis and found that up-regulation of ARRB1 and down-regulation of ARRB2 both played a role as functional biomarkers to predict metastasis (P<0.05). Kallifatidis et al . [ 21 ] reported that ARRB2 negatively regulated the activation of STAT3, a transcription factor regulating the self-renewal nature of BCSCs. Conversely, ARRB1 was found to positively regulate BMI-1 and ARRB-1 were linked with poorer prognosis in BCa [ 21 ]. Overexpression of DCLK1 which was previously reported to be remarkable in cell progression and metastasis of colorectal cancer, was also found to be a significant prognostic factor on BCa. Multivariate Cox regression analyses conducted by Shafiei et al . [ 20 ] showed DCLK1 protein expression was an independent prognostic factor to poor disease-specific survival in BCa patients. However, the molecular mechanism of the protein expression was not yet well-established. Cripto-1 or teratocarcinoma-derived growth factor-1 (TDGF-1) was found to have a significant association with tumor recurrence/metastasis in BCa patients (P = 0.007) and also as an independent prognostic factor identified with multivariate Cox regression analysis (P = 0.036), which validated its role to be a valuable marker as a disease-free predictor in BCa patients [ 14 ].
This systematic review showed that most BCSCs expressions were significantly associated with tumor recurrence and metastasis, suggesting its important role in patients’ prognosis. CSC-specific cell-surface markers represent potential therapeutic targets. By knowing stem cell expression in BCa, therapeutic strategies could be set and implemented to improve disease outcomes. However, the mechanism of each CSCs was reported to be different due to its heterogeneity in the level of stem cells. Many cell-surface markers and signaling pathways are distinct in quiescent cells and proliferating cells; thus, microenvironmental interactions can alter stem cells’ marker expression and signaling pathways [ 1 ]. Therefore, a major consideration for this approach remains the specificity of these markers [ 2 ]. In addition, specific CSC phenotypes appear to be correlated with disease outcomes, including risks of recurrence and metastasis [ 2 ]. These findings are comparable to those derived from another independent cohort of samples from the PanCancer Atlas which also shows that the expression of isolated markers is correlated with poor outcomes in bladder cancer [ 23 ]. Based on these findings, further studies regarding CSCs, especially their molecular mechanism, are warranted and may have significant contribution to the overall management of BCa.
Our study has several limitations. The majority of studies included did not show the mean or median follow-up time to determine the outcome. Each study also had different patients’ characteristics, tumors’ profiles, and treatment plans, which may also affect the recurrence and metastasis. We only presented a systematic review without further analysis; thus, we only can show that many studies have shown the beneficial impact of identifying BCSCs, and further studies are required.
Larger multicenter studies are needed to assess each factor that contributed to the recurrence and metastasis of BCa. However, statistical analysis sometimes has poor accuracy and is not applicable individually; thus, artificial intelligence has been further developed and may answer this problem. There have been several research that stated that artificial intelligence was believed to accurately predict cancer behavior, overall survival, and disease recurrence on BCa. Furthermore, Artificial intelligence can provide patient-tailored instruments for diagnosing and managing BCa [ 24 ].
Conclusions
The detection of cancer stem cell expression offers a promising modality in predicting the prognosis of BCa. However, much is lacking in the molecular mechanisms underlying these processes. Hence, future research in this area is warranted and may highly contribute to the overall management of BCa.
Supporting information
S1 Checklist
PRISMA 2020 for abstracts checklist.
(DOCX)
S2 Checklist
PRISMA 2020 checklist.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The ability to predict epileptic seizures provides an opportunity to intervene in order to attenuate their effects, or if possible prevent them. In this study we focus on EEG manifestations of seizures, which are characterized by sudden hypersynchronization of neurons and last from seconds to minutes. [ 1 ] Recently published studies on seizure prediction use a wide variety of approaches, from time series analysis (e.g. phase synchronization [ 2 ] or bivariate phase synchrony [ 3 ]) and spectral features of EEG signals [ 4 , 5 ] to physiological models of neural activity (e.g. neural mass models [ 6 ]) or circadian models [ 7 ]. We focus on spectral measures of EEG signals since they have been successfully used as features for seizure prediction, and are easily interpretable. [ 4 , 8 , 9 ]
In the field of seizure prediction there are certain conceptional, computational and data-related challenges. First, using a large number of features for prediction makes it difficult to interpret their individual contribution. [ 9 ] Secondly, the algorithms for seizure prediction in a clinical setting need to be computationally efficient. Due to hardware constraints, this applies to closed-loop EEG devices for seizure prediction and intervention in particular, which have been a recent focus in the field. [ 8 – 11 ] Finally, data encountered in the field of seizure prediction can be high dimensional and heterogeneous (e.g. recorded using many different channels and types of measurements in addition to EEG, like ECG, EOG etc), yet suffer from class imbalance (patients spend more time in interictal than in preictal states) and limited in the number of labeled samples. This is particularly challenging for the design of a patient-specific model.
In this study we address these issues by developing an easy-to-use, computationally efficient method for patient-specific seizure prediction. In order to achieve that, we extract a small set of interpretable features from power spectra that distinguish a baseline (interictal) EEG activity from a state leading up to a seizure (preictal state). Interictal states are regular brain activity between seizures, which can sometimes be interrupted with interictal spiking. [ 1 , 12 ] Since seizures are characterized by strong synchronization, they are very prominent in power spectra of EEG signals. Although preictal states are not clearly visible in raw EEG signals, multiple studies confirmed the presence of distinct preictal states using spectral [ 4 , 13 , 14 ], as well as information measures. [ 15 – 17 ] For a detailed discussion, see [ 8 ] and [ 9 ].
Although power spectra capture relevant changes in frequency over time, they can be very noisy and contain outliers. We thus use nonnegative matrix factorization (NMF) [ 18 , 19 ] to decompose power spectra into dominant time and frequency components, which are later used for seizure prediction.
To mitigate class imbalance, we employ synthetic minority over-sampling technique (SMOTE) [ 20 ], together with linear SVM with L1 regularization, to assign weights for contributions from each individual channel and eliminate uninformative channels. A software implementation of the presented method is available online at: https://github.com/ostojanovic/seizure_prediction . The method is applied to a part of the Freiburg EPILEPSIAE dataset [ 21 ], and compared to the Epilepsyecosystem dataset [ 22 ]. The developed method is computationally inexpensive and produces good results while providing insights into the structure of preictal states.
Materials and methods
Data preparation
Freiburg EPILEPSIAE dataset
The data consist of heterogeneous EEG recordings of five pre-surgical patients (one female; median age: 29.2) [ Table 1 ] and form a part of the bigger Freiburg EPILEPSIAE database. [ 21 ] Recordings are made at the University Medical Center Freiburg, over the course of several days (three to nine), between 2003 and 2009. The sampling frequency varies between 256Hz and 1024Hz. The electrodes that are used in the recordings include intracranial (depth, strip and grid) and surface electrodes, together with special electrodes (e.g. ECG, EMG and EOG), whose number varies between 31 and 122, depending on the diagnosis. In order to investigate preictal states thoroughly, only intracranial EEG recordings are used.
10.1371/journal.pone.0228025.t001
Table 1
Detailed information about patients the from EPILEPSIAE database.
[ 21 ] The number of preictal intervals is the same as the number of seizures.
Patient’s number
age
sex
number of channels
sampling frequency (Hz)
number of preictal intervals
number of interictal intervals
1
34
male
48
256
16
88
2
37
female
26
512
6
44
3
18
male
94
1024
8
80
4
42
male
38
1024
6
110
5
15
male
91
256
14
9
Since the ability to predict a seizure five minutes before its onset can be useful for patients with uncontrolled epilepsy [ 23 ], we focus on five minute intervals of preictal and interictal states. In the case of a preictal state, an interval of five minutes leading up to a seizure, with a 30 seconds seizure horizon is extracted. Seizure onsets are hand-labeled at the University Medical Center Freiburg. Since preictal states directly precede seizures, seizure prediction can be realized by classification between preictal and interictal states.
In the case of an interictal state five minutes intervals are extracted, which are at least 11 minutes before or after any other seizure. We refer to these intervals of extracted signals as individual measurement periods. The data are filtered with the Parks-McClellan optimal equiripple finite impulse response filter to remove 50Hz line noise.
The dataset is separated into training (70%) and validation set (30%) during a 100-fold cross-validation procedure.
Epilepsyecosystem dataset
The dataset consists of intracranial EEG recordings of three patients (all females; median age: 50). [ Table 2 ] Recordings are made at the St Vincent’s Hospital in Melbourne, Australia as a part of the world-first clinical trial of the implantable NeuroVista Seizure Advisory System. [ 24 ] In total, 16 electrodes are used for each patient and sampling frequency is 400Hz. The dataset consists of the public and the private (benchmark) set. Since labels of preictal and interictal states are known only for the public set, it is used for developing a model, while the benchmark set is used in the final stage for comparison with other algorithms for seizure prediction. [ 22 ]
10.1371/journal.pone.0228025.t002
Table 2
Detailed information about the Epilesyecosystem dataset (after excluding corrupted files).
[ 22 ] The number of preictal intervals is the same as the number of seizures in the public dataset, while for files in the benchmark dataset labels are not publicly known.
Patient’s number
age
sex
number of preictal intervals
number of interictal intervals
number of files (benchmark set)
percentage of excluded files
1
22
female
225
500
162
14.9%
2
51
female
216
1688
941
7%
3
50
female
251
1896
679
1%
Preictal intervals are ten minute segments which are cut out of recordings covering one hour prior to seizure with a five minute seizure horizon. (i.e. from 1:05 to 0:05 before seizure onset). Interictal intervals are also ten minute segments cut out from one hour of recording, which is at least four hours away from any seizure. Some of the files contain data dropouts which happen when the intracranial brain implant temporarily fails to record data. This manifests in zero values of iEEG across all channels at a given time sample. All files that contain more than 50% of data dropouts are excluded from the further analysis. For files that contain less than 50% of data dropouts, the corrupt data are deleted and the rest of the signal is concatenated. The data are filtered with the Butterworth infinite impulse response filter to remove 50Hz line noise.
The public dataset is separated into training (70%) and validation set (30%) during a 100-fold cross-validation procedure.
Deriving time and frequency components
To identify stereotypical behavior between and ahead of seizures, spectrograms of each channel [ Fig 1 ] (for the Freiburg EPILEPSIAE dataset) are obtained using the multitaper method [ 25 ] with time windows of 10 seconds (which is calculated by using 50% overlap of a 20 seconds window). For the Epilepsyecosystem dataset, spectrograms of each channel are calculated using the Fast Fourier Transform. To correct for baseline activity across frequencies, relative power is calculated by dividing spectrograms of each channel by the average interictal spectrogram.
10.1371/journal.pone.0228025.g001
Fig 1
Example spectrograms of preictal and interictal states.
Baseline corrected spectrograms of a preictal ( A ) and an interictal ( B ) individual measurement period of channel HR1 from patient 1. This channel and individual measurement period will be used throughout the paper for illustrative purposes, if not stated otherwise.
Due to the clinical setting and patients’ diagnoses, the sampling frequency varies among different patients from the two datasets. As a result, the highest frequency in the spectrograms varies between 128Hz and 513Hz. However, this difference is unproblematic due to the fact that we develop patient-specific models. After obtaining spectrograms of every individual measurement period for every channel, they are visually inspected, and in the case of anomalies (e.g. electrode detachments, sudden amplitude jumps), excluded from the data.
Time-frequency decomposition
To examine changes in power spectra, spectrograms of each channel and each individual measurement period are decomposed into a time and a frequency component using nonnegative matrix factorization. Originally proposed under the name “positive matrix factorization”, it is a variant of factor analysis [ 18 ], which is first used on environmental data [ 26 ] and later popularized in the application to face recognition under the current name. [ 19 ] For both tasks, NMF is successful in learning interpretable parts-based representation (e.g. concentrations of elements, as in [ 26 ] or parts of faces, as in [ 19 ]) and shown to perform better than independent component analysis, principal component analysis or vector quantization. [ 27 – 29 ] In the field of seizure prediction, NMF has been used to develop a method for automatic localization of epileptic spikes in children with infantile spasms [ 30 ] and for automatic detection and localization of interictal discharges. [ 31 ]
Nonnegative matrix factorization decomposes a nonnegative matrix V into two nonnegative low-rank matrices W and H [ 19 ]:
V ∼ V ˜ n × m = W n × r × H r × m V ˜ i j = ∑ a = 1 r W i a H a j
The outer product V ˜ = W H can be interpreted as a low rank parts-based approximation of the data in V . [ 19 ] We decide on a factorization of rank r = 1 to get the most constrained model with two vectors, one of which represents temporal evolution (time component H ) and one of which represents distribution of frequencies (frequency component W ). [ Fig 2 ]
10.1371/journal.pone.0228025.g002
Fig 2
Time and frequency components and its models.
An example of decomposed time (solid blue lines) and frequency components (solid red lines) and their respective models (dashed lines) of a preictal state ( A , C ), as well as an interictal state ( B , D ). In a preictal state, the time component ( A ) increases as a seizure is approaching, while the frequency component ( C ) has an increase in low frequencies. Both interictal components ( B , D ) are steady and are an order of magnitude lower than their respective preictal components ( A , C ).
To lessen the influence of outliers and to remove noise in the NMF components, they are modeled with smooth basis functions using robust regression. The time component is modeled by a polynomial of second order, while the frequency component is modeled by nonlinearly logarithmically spaced B-splines of sixth order to consider the frequency resolution which decreases in higher frequencies. [ Fig 2 ] By modeling each component with smooth basis functions, the most relevant information is preserved in both domains, while noise is removed.
By calculating the outer product of modeled NMF components as shown in Fig 3 , time-frequency models can be reconstructed. They capture the most important information while leaving out the noise and thus provide simplified intermediate representation of the data, which can be visually compared to the corresponding spectrograms (see S1 Fig in the appendix). The coefficients of the modeled time and frequency components therefore convey relevant information about structure of both states.
10.1371/journal.pone.0228025.g003
Fig 3
Obtaining a time-frequency model from the respective components.
The NMF components are shown with solid red and blue lines for frequency and time, respectively, while their models are shown with dashed lines. The time-frequency model (center) is an outer product of modeled time and frequency components.
Prediction and performance measures
To classify between preictal and interictal states, linear support vector machines [ 32 ] are used. We combine the coefficients of both of the modeled NMF components across all channels into a feature vector. For example, recordings of patient 1 in the EPILEPSIAE dataset contain 48 channels with 12 NMF parameters (9 parameters for the frequency component and 3 parameters for the time component) each, leading to a dimensionality of 48 ⋅ 12 = 576. To account for the risk of overfitting due to the high number of features, L1 regularization is used. L1 regularization shrinks coefficients of less important features to zero by adding the absolute value of magnitude of coefficients as a penalty term to the loss function. [ 32 ]
In both datasets, interictal states are more frequent than the preictal ones, which leads to an imbalance of classes (c.f. Tables 1 and 2 ). To account for this, the SMOTE oversampling technique is used. [ 20 ] It creates synthetic samples of the minority class, based on k neighboring points of minority samples (in our case k = 5). This means that the new synthetic preictal sample is created based on the five closest preictal samples.
To ensure good generalization of the algorithm, 100-fold cross-validation is used on a training set (70%) and a validation set (30%). Average measures (accuracy, sensitivity, specificity, positive and negative predictive values) are reported. Since the classifier should neither miss nor falsely predict a seizure, we report sensitivity sensitivity and specificity, as well as positive and negative predictive values. [ 33 ] In the benchmark dataset the area under the curve (AUC) is used for comparison among other algorithms.
Sensitivity is the probability of a positive test result among those having the target condition (i.e. the proportion of correctly classified preictal states), while specificity is the probability of a negative test result among those without the target condition (i.e. the proportion of correctly classified interictal states). [ 33 ] The positive predictive value (PPV) is the probability of the target condition, given a positive test result (i.e. the measure of how likely it is that, if the classifier predicts a preictal state, a patient is experiencing it), while the negative predictive value (NPV) is the probability of not having the target condition, given a negative test result (i.e. the measure of how likely it is that, if our classifier does not predict a preictal state, a patient is not experiencing it). [ 33 ] Full expressions are given below:
Accuracy = T P + T N all samples Sensitivity = T P T P + F N Specificity = T N T N + F P PPV = T P T P + F P NPV = T N T N + F N
where:
TP is a number of samples classified as true positive
TN is a number of samples classified as true negative
FP is a number of samples classified as false positive
FN is a number of samples classified as false negative.
Results and discussion
Interpretability of the model
Fig 2 shows representative preictal and interictal components (of the EPILEPSIAE dataset), where the modeled NMF components show differences between the states. Model of the frequency component of a preictal state exhibits a peak of high activity in lower frequencies, relative to baseline activity. This is in line with previous findings of a structure below 30Hz (gamma range), which is informative for seizure prediction. [ 13 , 14 ] These structural differences are also visible in recovered time-frequency models (see S2 and S3 Figs in the appendix).
Average preictal and interictal components of all measurements and electrodes differ in both datasets, as shown in S4 and S5 Figs in the appendix. On average, time components of preictal states in the EPILEPSIAE dataset have higher intensity, and frequency components show increase in lower frequencies ( S4 Fig ). Equivalent average components in the public Epilepsyecosystem show slightly different behavior. Time components of interictal states have somewhat higher intensity, and frequency components have an increase in lower as well as in higher frequencies. Since labels for the private Epilepsyecosystem dataset are not available, it is not possible to analyze the benchmark dataset in the same way.
Fig 4 shows normalized histograms of maximum values of frequency components of preictal and interictal states for both datasets. In the EPILEPSIAE dataset most preictal components have maximum in lower frequencies, and interictal states have maximum in both lower and higher frequencies (above 100Hz). On the other hand, most maxima of preictal and interictal components in the public Epilepsyecosystem dataset are below 50Hz as well as between 150Hz and 200Hz.
10.1371/journal.pone.0228025.g004
Fig 4
Distribution of maximum of frequency components.
Results of the EPILEPSIAE dataset are shown in the upper row for preictal ( A ) and interictal states ( B ). The lower row shows results for the Epilepsyecosystem dataset ( C for preictal and D for interictal states).
This difference in components between datasets can exist due to various reasons. The part of the EPILEPSIAE dataset used here might have too few measurements from an each patient. The Epilepsyecosystem dataset has more measurements, but it still contains data for only three patients. For a better assessment more data from different patients should be analyzed. In addition, it should be noted that the part of the EPILEPSIAE dataset used here contains data of pre-surgical patients and seizures recorded in this setting might not always be representative of typical epileptic seizures. As it is shown in [ 34 ], features of intracranial EEG signals show high variability after implantation of electrodes and spatial variability of lower frequency power bands across channels decreases over time. On the other hand, the Epilepsyecosystem dataset contains recordings from the world-first clinical trial of the human-implanted NeuroVista seizure advisory system [ 24 ], which might also be more distinguished than other clinical trials. Lastly, in the EPILEPSIAE dataset the 11-minutes buffer for interictal periods is used, which might be too short. The study in [ 35 ] reveals existence of “pre-cursors” to seizures (energy bursts in iEEG signals), which suggests that epileptic seizures might start hours in advance (also shown in [ 24 ]). Considering all of this, the best assessment of differences in preictal and interictal states would be in a closed-loop seizure prediction setting in real-time, for which the proposed method would, with appropriate adjustments (e.g. calculating spectrograms of consecutive time windows instead of short segments) be suitable.
Predictive performance
On the EPILEPSIAE dataset, similar accuracy is achieved for all patients (above 90%). The lowest performance is for the patient 5 (90.4%) and the highest for the patient 4 (100%), as shown in Fig 5 and Table 3 . Sensitivity is between 0.8 and 1, while specificity ranges from 0.98 to 1, as can be seen in Fig 5 . A combination of high values of sensitivity and specificity is achieved for all patients. Similarly, positive predictive values are between 0.98 and 1, while negative predictive values are between 0.85 and 1 (c.f. Fig 5 and Table 3 ).
10.1371/journal.pone.0228025.g005
Fig 5
Evaluation of prediction performance.
Results on the EPILEPSIAE dataset are shown in the upper row( A-C ). Results on the public Epilepsyecosystem are shown in the middle row ( D-F ) and the results on the private Epilepsyecosystem dataset (benchmark) are shown in the lower row ( G-I ). Performance of each patient is represented by a circle, for accuracy ( A , D , G ), specificity-sensitivity plot ( B , E , H ) and negative and positive predictive value ( C , F , I ). Identical colors are used to represent each patient across all nine subplots. The hatched area represents results attainable by a random classifier.
10.1371/journal.pone.0228025.t003
Table 3
Performance measures for all patients from the EPILEPSIAE dataset (upper section), from the Epilepsyecosystem public dataset (middle section) and Epilepsyecosystem benchmark dataset (lower section).
Patient’s number
accuracy (%)
sensitivity
specificity
positive predictive value
negative predictive value
1
99.7
0.99
1
1
0.99
2
97.5
0.97
0.98
0.98
0.97
3
99.5
1
0.99
0.99
1
4
100
1
1
1
1
5
90.4
0.8
1
1
0.85
1
74.1
0.75
0.73
0.73
0.75
2
73
0.57
0.81
0.63
0.77
3
78.5
0.75
0.82
0.81
0.77
1
71
0.44
0.76
0.25
0.88
2
61
0.37
0.63
0.06
0.94
3
69.2
0.37
0.72
0.11
0.92
Predictions on the public Epilepsyecosystem dataset are lower than on the EPILEPSIAE dataset (around 70% for all patients; c.f. Fig 5 and Table 3 ). The lowest performance is for the patient 1 (74.1%) and the highest for the patient 3 (78.5%). Sensitivity, specificity, positive and negative predictive values for all patients are still higher than attainable results by a random classifier, but still considerably lower than on the EPILEPSIAE dataset, which can be seen in Fig 5 . Sensitivity is between 0.57 and 0.75, while specificity ranges from 0.73 to 0.82. Positive predictive values are between 0.63 and 0.81, and negative predictive values are between 0.75 and 0.77.
On the benchmark dataset, the highest achieved accuracy is for the patient 1 (71%), and the lowest for the patient 2 (61%). However, other performance measures drop significantly (sensitivity and positive predictive value are below 0.5). This drop in performance happens with most of other algorithms that are evaluated on the Epilepsyecosystem dataset [ 22 ], but the difference is not always as big. There might be various reasons for this. In general, it is the harder task to train a model on one dataset, and then evaluated it on the unseen set. Furthermore, the class imbalance between the sets might differ, which would explain the big difference between sensitivity and positive predictive value. It is also possible that SMOTE algorithm learns noise when oversampling the minority class in the public dataset. Finally, patients who have a higher seizure frequency (i.e. seizures per day) seem to have worse seizure prediction performance based on the original clinical trial. [ 24 ]
As mentioned in the Prediction and performance measures , the AUC is used for comparison with other algorithms on the benchmark set. The average reported AUC is 0.57 (0.62 for the patient 1, 0.52 for the patient 2 and 0.58 for the patient 3), which places the proposed algorithm on the 65th place (out of current 102 evaluated algorithms). For comparison, the algorithm with the best performance on the benchmark dataset (which is the combination of extreme gradient boosting, k-nearest neighbours, generalized linear model and linear SVM) has AUC of 0.8. [ 22 ]
The reasons for the overall lower performance on both Epilepsyecosystem datasets can lie in the fact that there are more seizures and more data per patient, making prediction possibly more challenging by potentially adding more variability to the data. It should also be noted that the data of three patients from the Epilepsyecosystem dataset correspond the ones whose seizures are the most difficult to predict [ 24 ].
Conclusion
Since patients with uncontrolled epilepsy prefer to be advised a few minutes before a seizure onset [ 23 ], we decided to use intervals of five minutes, extracted from longer recordings of the EPILEPSIAE dataset. However, this method is easily extensible to longer periods of time, since the length of intervals has no effect on dimensionality of modeled time components, which is shown by comparing the proposed method on the Epilepsyecosystem dataset.
Data from additional patients as well as more data from the same patient could, if available, lead to a better generalization of the model. This however is a challenge for patient-specific models in general, where data from a single patient should suffice, and a large number of labeled training examples is not available.
Overall, this study demonstrates the use of nonnegative matrix factorization of power spectra for a seizure prediction task. The proposed model is conceptually simple, interpretable and has shown good accuracy on two representative datasets and lower performance on the benchmark set where improvements in the direction of coping with class imbalance should be made. A similar approach could be used for similar tasks such as detection of sleep stages in EEG or the detection of irregularities in ECG.
Supporting information
S1 Fig
Time-frequency models and corresponding spectrograms of preictal and interictal states.
An outer product of modeled time and frequency components ( A , C ) and corresponding spectrograms ( B , D ). A preictal state is shown in the upper row ( A-B ) and an interictal state is shown in the bottom row ( C-D ).
(PDF)
S2 Fig
Models of preictal states.
Models shown here are for different channels ( A-I ) from the same individual measurement period for patient 1.
(PDF)
S3 Fig
Models of interictal states.
Models shown here are for different channels ( A-I ) from the same individual measurement period for patient 1.
(PDF)
S4 Fig
Average models of time and frequency components of all channels and all measurements for preictal and interictal states of the EPILEPSIAE dataset.
Models of time components are shown in the upper row ( A-E ), and models of frequency components are shown in the bottom row ( G-K ). Preictal states are indicated with a dashed line and interictal states are indicated with a line marked with + in blue for models of time and red for models of frequency components, respectively.
(PDF)
S5 Fig
Average models of time and frequency components of all channels and all measurements for preictal and interictal states of Epilepsyecosystem dataset.
Models of time components are shown in the upper row ( A-C ), and models of frequency components are shown in the bottom row ( D-F ). Preictal states are indicated with a dashed line and interictal states are indicated with a line marked with + in blue for models of time and red for models of frequency components, respectively.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The role of non-neuronal cells such as glia in neural computation has been the topic of increasing interest over the past decade [ 1 – 3 ]. In the mammalian brain, glia comprise a significant proportion of all cells, comparable to that of neurons [ 4 ]. However, their functional role has traditionally been viewed as one of maintaining the basic physiological needs of neurons [ 5 – 7 ]. This view has now been repeatedly challenged owing to a decades-long stream of evidence that these cells directly modulate neuronal signaling [ 8 , 9 ]. The simple premise here is that the computational power of the brain should be conferred by all cells collectively populating the brain, and not merely by neuronal activity [ 8 – 11 ]. That is, the effects of glia on brain function and neuronal activity exist, and hence must matter. This notion opens up richer and more expansive hypotheses regarding the mechanisms underlying brain computation, including ways by which neuromodulation of networks may be achieved and mapped to function.
In the current work, we zero our attention on astrocytes, a prominent type of glial cell within the nervous system. Collective work in the field of astrocyte biology has repeatedly provided evidence on the instrumental role of astrocytes in controlling neuronal functions such as synaptic wiring, synaptic activity, synaptic memory, and neuronal excitability [ 12 – 20 ], reflecting the potential of astrocytes to control key computational loci in the brain. However, directly probing the role of astrocytes in brain computation has been virtually impossible at the experimental level due to limited knowledge surrounding their rules of engagement and signaling mechanisms, combined with their non-binary rules of ‘excitability’, and our inability to specifically target their neuron-bound modulatory functions without affecting their more general ‘homeostatic’ roles. Additionally, the multiplex nature of astrocytes, whereby a single astrocyte is capable of a multitude of inhibitory, excitatory, or modulatory outputs, distinguishes them in general from neurons. Lastly, the incomplete toolkit available to manipulate them exacerbates the challenge. On the contrary, computational neuroscience provides an ideal playground to probe the role of astrocytes in circuit computation by way of mathematical and algorithmic modeling. So far, the absence of a consensus framework on how to conceptualize astrocyte’s contribution to brain computation in a reductionist way has made it difficult to meaningfully abstract astrocyte functions in computational models. Interestingly, a new hypothesis called “contextual guidance” was recently introduced that potentially alleviates these issues [ 8 ]. It posits that astrocytes act as a contextual switchboard that actively conveys information about the environment and physiological state of the organism to neuronal networks. More generally, accounting for astrocytes, and other glial, in neural computation theory may close gaps in how neural circuits learn and implement functions in a manner sensitive to context. For example, an extant issue in theoretical neuroscience pertains to how different but functionally overlapping tasks may be embedded in a single neuronal circuit [ 21 – 23 ]. Such a scenario would seemingly require mechanisms by which different neuronal dynamical regimes may be learned and then recruited, in a context/task-dependent fashion. The goal of this paper is thus to introduce computational modeling and analysis to probe how astrocytes may enrich the computational capability of neural circuits toward such objectives.
Astrocytes contain distinct physiological features relative to neurons. They have slow time-scales of activation, on the order of seconds or slower. Indeed, while approximately 9% of spontaneous astrocyte intracellular calcium events have kinetics of hundreds of milliseconds, most calcium events are documented in the time-scale of seconds and astrocytic outputs and responses to stimuli commonly extend over several tens of seconds. This fact makes them easy to dismiss from the perspective of fast computation. However, these slow time-scales of astrocytes may in fact be a computationally-relevant feature in light of the specific ways in which astrocytes interact with neurons. In particular, neural network function is often viewed through the lens of synaptic connectivity, wherein specific synaptic ‘weight’ configurations are associated with different tasks [ 24 – 27 ]. However, a single astrocyte can impinge on dozens of neurons and hundreds of thousands of synapses, and, for decades now, physiological experiments have indicated that astrocytes possess the capability to gate and influence synaptic plasticity [ 14 , 28 – 30 ]. This astrocyte-induced synaptic plasticity belongs to a form of meta-plasticity as outlined in [ 31 ]. Along these lines, the involvement of astrocytes in meta-plasticity was further theoretically formalized and modeled in [ 32 ]. These prior works substantiate the notion that astrocytes can impact important physiological learning processes. In the current paper, we set forth to examine astrocytic meta-plasticity at a network scale that can be linked to complex functional settings.
Such a framework would represent a shift from common conceptualizations of neural computation that rely on homogeneous neural units, and thus explain how information processing mechanisms may be enacted over different temporal scales. This, in turn, may better reconcile models of algorithmic learning with the physiological realities of the brain. In fact, recent work has argued that astrocytes may implement a transformer-like model of attention in multi-task adaptation and learning in feedforward architectures [ 33 ]. In [ 34 , 35 ], it is shown that neuron-astrocyte interactions can lead in turn to distinct patterns of neural activity in working memory tasks through mean-field network model analyses. In [ 36 ], neuron-astrocyte interactions are modeled within neuromorphic spiking neural network architectures, also in the context of memory. There, the model is deployed in image recognition tasks via supervised learning, where it is shown that the presence of slow astrocytic calcium signaling can improve memory performance. Other biophysical and phenomenological models of neuro-astrocyte interactions have also been considered [ 37 – 39 ], however, most of these models are focused on one precise astrocyte output or function (such as glutamate release) or on explaining or recapitulating circuit-level phenomena (e.g., neuronal firing rate activity), rather than connecting to higher-level functions. In the current paper, we focus our attention on the network dynamics of neuron-astrocyte interactions in a rate-based recurrent network and reinforcement learning scenario. Specifically, we study neuron-astrocyte interactions with a focus on two dimensions: (i) the dynamics of recurrent interplay of neuronal activity and astrocytic modulation, and (ii) the functional salience of such dynamics in reinforcement learning scenarios. The correlation between network dynamics, e.g., vector fields, attractors, etc., and different functions is itself a crucial area of study in theoretical neuroscience [ 40 ]. Furthermore, there is recognition that leveraging the multiple time-scales and heterogeneous structures of recurrent neural networks to design models for learning multiple, sequential, and temporal tasks [ 41 – 44 ]. As such, adding astrocytes to traditional recurrent neural network architectures could thus further expand the expressiveness of these networks [ 45 – 47 ]. Our goal here is to further explore this emerging question.
Motivated by the above, we seek to develop and study a simplified dynamical systems model of neuron-astrocyte interaction in order to gain fundamental insight into how the time- and spatial-scale separation between astrocytes and neurons may enrich the repertoire of neural dynamics and activity. Our models are built in a bottom-up fashion, using well-established biophysical paradigms combined with validated theories of astrocytic modulation of neuronal dynamics and synaptic plasticity, such as described above. Furthermore, we seek to understand how astrocyte-driven dynamics may enable learning over disparate time-scales and in context-dependent task scenarios, consistent with the contextual guidance hypothesis [ 8 ]. For the latter, we choose to focus on decision-making problems and reinforcement learning (RL) scenarios, given their relevance and ubiquity in algorithmic learning and prior observations that astrocytes can participate in the encoding of reward information [ 48 , 49 ].
We proceed to formulate a novel bio-inspired dynamical systems model of neuron-astrocyte interactions, and then embed this model in algorithmic optimization frameworks to solve context-dependent bandit tasks. Our major contributions include the dynamical systems analysis of this model, and understanding astrocytic modulation as a pseudo-bifurcation parameter that can switch neural and synaptic dynamics between different dynamical regimes via meta-plasticity. Herein, astrocytes will form a ‘second-order’ modulation on the time-evolution of synaptic weights, resulting in different generative dynamics of neural activity. We furthermore show that the structure and time-scale separation of astrocytes relative to neurons is enabling in terms of learning non-stationary bandit problems, exceeding the learning performance of well-established algorithms in this domain. It is worth emphasizing that our goal is not to introduce specific dynamics to the model and ascribe these to astrocytes. Rather, we ground our model in extant biological evidence regarding neural-astrocyte interaction which can be linked to specific hypotheses from the contextual guidance framework regarding the role of said interactions.
Results
Neuron-astrocyte interactions constitute a hypernetwork with multi-scale dynamics
We proceed to develop a reduced model of neuron-astrocyte interaction that captures key aspects of neurobiology while enabling fundamental analysis regarding dynamical expressiveness and links to function.
Neuron-astrocyte structure as a hypernetwork
Classically, biological interactions between neurons, astrocytes, and synapses have been conceptualized in terms of the tripartite synapse structure [ 11 , 50 , 51 ] (as shown in Fig 1A ). Within this framework, astrocytes interact with neurons at synapses, modulating synaptic efficacy [ 52 ] and controlling synaptic plasticity [ 53 ]. Such interactions may occur in a higher-order and ‘closed-loop’ fashion, wherein astrocytes respond to neurotransmitters released during pre- and post-synaptic neuronal activity (see Section A in S1 Appendix for detailed description) and this has been the mainstream assumption in past work attempting to model astrocytes. While this description may capture an important dimension of neuron-astrocyte interaction, it is increasingly clear that astrocytic modulation of neuronal activity is more general and multifaceted. The contextual guidance hypothesis [ 8 ] espouses that astrocytes regulate synaptic activity not only in response to synaptic activity itself, but also as an adaptative response to external drives, such as vigilance state, sensory salience, metabolic load, or underlying pathology (see Fig 1A ). As such, astrocytes may actively ‘control’ neural dynamics in a state-dependent manner [ 34 , 35 ]. While we focus here on effects at the synapse via the release of astrocyte-derived neuroactive transmitters, in principle astrocyte modulation can also occur at cell bodies via the alteration of ionic conditions, notably potassium levels [ 6 , 19 , 54 – 56 ].
10.1371/journal.pcbi.1012186.g001
Fig 1
A. In a tripartite synapse, the presynaptic axon and postsynaptic dendrite are surrounded by an astrocyte [ 11 , 50 , 51 ], enabling multifaceted effects of neurotransmitters and (astrocyte-derived) gasotransmitters. B. A graphical illustration of the neuron-astrocyte hypernetwork: the circles and stars represent neurons and astrocytes respectively; the colored triangles denote the hyperedges and represent the multiplexed intralayer interactions. C. Schematic representation of the feedback interconnections between subsystems in the multi-scale neuron-astrocyte network model.
The above schema of neuron-astrocyte interactions is difficult to capture as a traditional graphical network representation. As a result, we introduce the framework of a hypernetwork to describe the neuron-astrocyte architecture (see Fig 1B for the illustration and detailed description in Section B in S1 Appendix ). We distinguish neurons and astrocytes by representing them on two different layers of the network. The interlayer relationships are all hyperedges, which embody the ability of astrocytes to modulate neuronal synaptic activity and therefore neuronal activity indirectly.
Multi-scale neuronal and astrocytic dynamics
The hypernetwork formulation alone does not capture the full complexity of neuron-astrocyte interaction, as it does not explicitly contain information about the time-scales and dynamics of neuronal and astrocyte activation. For this, we introduce a set of ordinary differential equations (ODEs) overlaying the hypernetwork:
τ n x ˙ i = - a i x i + ∑ j = 1 n w i j ϕ ( x j ) + u i , i = 1 , … , n , (1a) τ w w ˙ i j = - b i j w i j + c i j ϕ ( x i ) ϕ ( x j ) + d i j ψ ( z k ) , i , j = 1 , … , n , (1b) τ a z ˙ k = - e k z k + ∑ l = 1 m f k l ψ ( z l ) + g k , k = 1 , … , m , (1c) g k = h k ϕ ( x i ) ϕ ( x j ) + v k . (1d)
These dynamical equations are based on firing rate descriptions of neural activity (see Methods for modeling details). Here, x i describes the rate of the neuron i = 1, …, n , w ij is the weight of the synapse (i.e., the synaptic efficacy) between neurons i and j , and z k represents the activity (abstracted from calcium activity) of astrocytes k = 1, …, m . Here we emphasize that z k embeds a graded but non-linear transformation between the inputs to astrocytes and their output onto neurons. There exist many models for describing the dynamics of neurons, and the one we use is, in essence, a continuous-time rate-based recurrent neural network (RNN) [ 57 ]. For the edge weights between neurons, we prescribe a Hebbian plasticity rule wherein weight changes are dependent on the correlation ϕ ( x i ) ϕ ( x j ). The signal u i conveys external inputs onto neural dynamics.
To distinguish astrocytes from neurons, we use a different activation function (i.e., ψ (⋅) ≠ ϕ (⋅)) and, most crucially, will assume that the time-scale τ a is slower than that of neurons. Specifically, a larger value of τ n , τ w , and τ a implies a slower rate of time-evolution [ 58 ] of the associated activity variables. Thus, the multiple time-scale feature of neuron-astrocyte processes is readily captured in equations (1) , with a suitable choice of the values of these parameters. Completing the model, f kl denotes interactions between astrocyte l and k , allowing for potential gap junctions-mediated communication between neighboring astrocytes [ 59 ]. An important feature of the model is that astrocytes may be sensitive to contextual information in accordance with [ 8 ], via g k . Here, we postulate two forms of context as specified in (1d) . First, we consider a ‘circuit’ context, such that the astrocyte may have a sensitivity of second-order neuronal activity via the coefficient h k . Second, we formulate an external context, motivated by the contextual guidance hypothesis, conveyed by the exogenous ‘contextual signal’ v k . Such a signal may originate, for example, from the sensory periphery [ 60 ]. The neuronal exogenous input u i may also contain such contextual information.
The model above attempts to balance expressiveness, interpretability, and tractability. In particular, we have not fully captured the spatial scale distinctions of astrocytes relative to neurons here, since we restrict ourselves to only the case of two neurons within the domain of a single astrocyte. We have, however, captured several important features of the astrocytic contextual guidance hypotheses: (i) the presence of multiple, nested loops of feedback between neurons and astrocytes, providing a diversity of mechanisms by which contexts can propagate through astrocytes and affect neuronal activity, and (ii) the potentially orders-of-magnitude separation in time-scales between neuronal activity and astrocytic modulation thereof. Unlike previous abstractions such as [ 34 ], we do not assume spike-like dynamics within astrocytes, since these cells are electrically inactive on a cell-wide scale. In total, astrocytes modeled here: (a) produce slow, graded activity (as a surrogate for calcium) that (b) modulates neuronal excitability and synaptic plasticity and (c) is responsive to the circuit and external context via feedforward and feedback signaling paths. It is of note that the above neuron-astrocyte model is well-behaved from a dynamical systems perspective since solutions exist, are unique, and are restricted to a bounded subspace (see Section C in S1 Appendix ).
From a systems-level perspective, the dynamics of the neuron-astrocyte network can be understood as the interaction between three subsystems, forming two closed-loops as shown in Fig 1C . The first closed-loop consists of the subsystem of neurons (1a) and synapses (1b) . The second closed-loop involves the subsystem of astrocytes (1c) , which transfers information from neurons to synapses. By forming these closed-loops, the astrocytic process not only directly modulates synaptic plasticity based on neural activity but also indirectly modifies synaptic connections, shaping the dynamics of the network as a whole. This mechanism can facilitate the formation and evolution of attractors (e.g., fixed points) in the neural subsystem state space, as elaborated below.
Astrocytic modulation acts as a pseudo-bifurcation parameter that changes meta-plasticity and neural circuit dynamics
To analyze the dynamics of (1) , we reduce it to its simplest motif, i.e., the interaction of two neurons and a single astrocyte. Here, we assume that the neurons form a reciprocal excitatory-inhibitory loop, itself a common canonical motif for cortical interactions between pyramidal and inter-neurons. From (1) , the neuron-astrocyte motif amounts to a set of 5 ODEs:
τ 1 x ˙ 1 = - a 1 x 1 + w 2 ϕ ( x 2 ) + u 1 ( t ) τ 1 x ˙ 2 = - a 2 x 2 + w 1 ϕ ( x 1 ) + u 2 ( t ) τ 2 w ˙ 1 = - b 1 w 1 + c 1 ϕ ( x 1 ) ϕ ( x 2 ) + d 1 ψ ( z ) τ 2 w ˙ 2 = - b 2 w 2 + c 2 ϕ ( x 1 ) ϕ ( x 2 ) + d 2 ψ ( z ) τ 3 z ˙ = - ϵ z + h ϕ ( x 1 ) ϕ ( x 2 ) + v ( t ) . (2)
The dynamics of this system are asymptotically bounded (see Section D in S1 Appendix ). Within this bounded set, the motif may exhibit a unique fixed point, or multiple fixed points, depending on parameterization. Fig 2B and 2C show the case of three fixed points under the assumption that astrocytes evolve at a time-scale two orders of magnitude slower than neurons and synapses (i.e., τ 3 = 100 τ 1 , τ 2 ). Fig 2D illustrates the time evolution of a specific trajectory within this landscape. As expected, z evolves much slower than the other variables. Notably, this slowly-changing astrocytic activity variable seems to drive neural variables to transit between nearly stationary regimes corresponding to phases 1 and 3 of astrocytic states in the lower plot of Fig 2D , suggesting that astrocytes can systematically ‘control’ stationary neural activity.
10.1371/journal.pcbi.1012186.g002
Fig 2
Neuron-astrocyte network motifs and dynamic properties.
A. Graphical representation of the network motif; u 1 , u 2 , v include inputs from other nodes of the hypernetwork as well as those from external sources. B, C. Several examples of trajectories in the state space ( x 2 , w 2 , z ) of the network motif system, where the boxes show the starting points. The parameter conditions are a 1 = 0.7, a 2 = 0.6, b 1 = 1.6, b 2 = 1.7, c 1 = 12, c 2 = −10, d 1 = −4, d 2 = 5, ϵ = 0.6, h = 6, and τ 1 = τ 2 = 0.01, τ 3 = 1. The system has three fixed point points, of which two are stable (red dots) and one is unstable (blue dot). The system dynamics converge to these two stable fixed points. D. Trajectory associated with the thick phase curve from B, C. illustrating two stationary regimes (indicated by phases 1 and 3 in the lower plot). E. depicts the bifurcation diagram of the neural dynamics with respect to the astrocyte output ψ ( z ), where the red curve shows that one branch of fixed point always exists, while the blue curve shows how the other branch of fixed points changes via the saddle-node bifurcation. F, G. Vector fields of the neuron-synaptic dynamics to either side of the saddle-node bifurcation.
In order to understand this phenomenon in more detail, we performed a singular perturbation analysis (see Section E in S1 Appendix ) to better clarify the mechanisms by which astrocyte signals may be modulating neural dynamics. This analysis treats the astrocyte state as a fixed parameter, premised on its relatively slow evolution relative to the neural dynamics. We can then study how this parameter affects the vector field and attractor landscape of the neural subsystem. Fig 2E provides the pseudo-bifurcation diagram of the above motif by showing the position of the fixed points in the x 1 -dimension as a function of the ψ ( z ). When ψ ( z ) is small, there is only one fixed point (the red line, also see Fig 2F ). When ψ ( z ) is large, the neural subsystem manifests three fixed points by means of a saddle-node bifurcation (see Fig 2G ). In other words, at the bifurcation point, there is a fundamental change in the shape of the neuronal-synaptic vector field and hence dynamics. Thus, astrocytic modulation can drastically alter the flow of neuronal and synaptic activity as a function of time. We hypothesize this mechanism may be particularly powerful for the contextual guidance premise as it may enable astrocytes to reshape the dynamics of synaptic adaptation and hence neural computation, based on exogenous contextual signals, e.g., via v ( t ). Thus, astrocytes form, in essence, a pathway for context-guided meta-plasticity and targeted neuromodulation. Below, we probe this hypothesis within the reinforcement learning task paradigm.
Neuron-astrocyte networks are able to learn context-dependent decision-making problems
We apply the proposed multi-scale neuron-astrocyte network model to context-dependent decision-making problems. We focus specifically on multi-armed bandits (MABs), a well-known class of reinforcement learning problems, wherein an agent aims to maximize its cumulative reward over time by selecting actions (arms) from a set of available options [ 61 ]. MABs find applications in various domains, including recommendation systems, clinical trials, and cognitive tasks in neuroscience, as they provide a powerful framework for decision-making under uncertainty [ 62 ]. While well-studied, this class of problems nonetheless poses persistent challenges when environments are non-stationary. Our prevailing hypothesis is that the disparate time-scales of signaling emanating from astrocytes can enable learning in such settings.
A standard MAB assumes a constant environment, in which the probabilities of reward associated with different arms are stationary. Our goal, however, is to study the capacity of our proposed neuron-astrocyte networks, by virtue of their time-scale separation, to learn in more challenging non-stationary and/or context-dependent settings. Thus, we designed both stationary and non-stationary Bernoulli bandit environments (see Fig 3 and Multi-armed bandit tasks in Methods ) within which to evaluate learning efficacy.
10.1371/journal.pcbi.1012186.g003
Fig 3
A. The architecture of the learning algorithm. The three plots on the right represent a stationary Bernoulli bandit scenario where the arm means remain (0.4, 0.8, 0.1) constantly over time, a flip-flop non-stationary Bernoulli bandit where Arm 2’s mean alternates between 0.92 and 0.042, and a smooth-change non-stationary Bernoulli bandit where all arm means change according to a smooth periodic function, respectively. The left figure shows the architecture of the learning algorithm. B, C, and D show the learning performance of the neuron-astrocyte (abbreviated as “Neuro-astro” in plots) method relative to other learning methods for the stationary bandit task, where B is for a single run, C is the average result for 10 runs and D is the mean and standard deviation of the asymptotic regrets (the UCB method is not compared in C and D, as it performs much worse than other methods).
Learning metrics
In MABs, a common figure of merit is the (pseudo) cumulative regret, which is defined specifically in Bernoulli bandits by
R T = ∑ t = 1 T ( max a i ∈ A μ i - E [ r t ] ) , (3)
where T is the total rounds, μ i is the mean of the action a i , which belongs to the action set A , and r t is the reward derived by the agent at trial t with E [ · ] denoting the expected value. A lower value of (3) indicates less accumulated loss and equivalently higher accumulated reward. Additionally, we consider the convergence speed of the algorithm, which measures the time taken by the agent for R T to reach an optimal value. Faster convergence is generally desirable as it signifies more efficient learning by the algorithm.
Learning algorithm architecture
In order to evaluate the proposed neuron-astrocyte model in these tasks, we require a learning/optimization method. For this purpose, we make several implementation assumptions. First, we assume that the network emits an output via a softmax operation, a typical form of network readout in neural network architectures. Second, we assume that networks have access to a signal that contains information about the environmental context (e.g., a change in arm probabilities, without overtly specifying the probabilities themselves). Upon this architecture, we deploy a reinforcement learning method to optimize all parameters of the model (see Methods ). The architecture of our learning algorithm is depicted in Fig 3A . Briefly, during a typical learning episode, the network outputs a policy for action selection, i.e., a probability distribution over the possible actions (at the output of the softmax). The bandit environment provides a reward to the agent in response, which is then fed into an analytical loss function, for which a gradient can be defined and hence network parameters updated. Crucially, this learning paradigm is agnostic to the specific network being learned, i.e., we can train vanilla RNNs and other architectures with the same methodology. This will allow us to make direct comparisons between the proposed neuron-astrocyte network and other standard neural networks.
Performance comparison
We conducted a comprehensive learning performance analysis of the proposed neuron-astrocyte network in comparison to other neural network architectures (vanilla RNN, LSTM, GRU), all trained the same way using the above method. In addition, we also deployed traditional algorithms for solving bandit problems, the Upper Confidence Bound (UCB) and Thompson Sampling (TS) methods. The specific learning procedures for all neural network-based methods are similar, as described in the above section.
Stationary case . Fig 3B–3D illustrates the comparison of the learning performance of different methods (neuron-astrocyte, LSTM, TS, vRNN, GRU, UCB) in a stationary bandit task with arm probability settings of (0.4, 0.8, 0.1). Each method requires exploration of the environment, resulting in high regret during the initial time steps. However, all methods eventually converge with comparable rates and cumulative regret of the same order of magnitude. In particular, the neuron-astrocyte architecture performs similarly to the other network-based implementations in this case. Single-run simulation results show that the neuron-astrocyte method uses less time to converge (see Section F.1 in S1 Appendix ). In addition, this method tends to be robust as the tasks become more challenging due to the small distance between arm probabilities (see Section F.2 in S1 Appendix ).
Non-stationary case . However, in the presence of non-stationarity, the neuron-astrocyte architecture displays significant gains in capability. Indeed, these networks can achieve almost stationary regrets over time as shown in Fig 4A and 4B , with the former depicting results for the flip-flop bandit and the latter for the smooth changing bandit. In contrast, other methods consistently result in escalating regrets. It is important to emphasize again that the setup for learning here is identical across all networks. These results are consistent across different non-stationary scenarios, evident in both individual and multiple runs (see Section F.3 in S1 Appendix ). In addition, similar learning performance is observed in scenarios with a different number of actions (see Section F.5 in S1 Appendix ), which suggests the generality of the neuron-astrocyte method. These observations indicate that the neuron-astrocyte network is able to leverage contextual signals and adapt its actions to the changing environment.
10.1371/journal.pcbi.1012186.g004
Fig 4
Learning performance.
Performance comparison of the neuron-astrocyte method relative to other learning methods for non-stationary bandits: A. flip-flop switching and B. smooth changing. C, D. neuron-astrocyte learning performance for different time-scale separation. E. Single learning traces for τ = 1 and τ = 0.01, highlighting the role of time-scale separation in enabling RL over contexts. F. Astrocyte and synaptic activity projections for both contexts (indicated as a and b ) in early, middle, and late phases of learning, highlighting the formation of distinct synaptic weight trajectories.
Time-scale separation is necessary for context-dependent learning
In order to probe the mechanisms by which the neuron-astrocyte network achieves context-dependent learning, we first focus on the time-scale separation between neurons and astrocytes. In our analysis above, we showed how astrocytic modulation may function, in essence, as a form of meta-plasticity wherein the time-scale separation enabled pseudo-bifurcations that could allow neuronal dynamics to traverse different functional regimes. The question at hand is whether this mechanism confers utility for context-dependent learning. To assess this, we varied the time-scale separation (via τ ) between astrocytes and neurons in our network, to probe the impact of this feature on learning performance. As shown in Fig 4C (see also Section F.4 in S1 Appendix ), different τ have significant impacts on learning performance, to the extent that without time-scale separation learning simply does not occur. This is seen for the case τ = 1, in which astrocytes and neurons have the same time-scale, indicating that the performance of the neuron-astrocyte network does not simply stem from the presence of additional units. Here, the cumulative regret does not converge. When τ = 0.1, the agent can sometimes achieve stationary asymptotic cumulative regret. This learning performance improves for greater time-scale separation. For τ ≤ 0.01 (that is, time scale separation greater than 2 orders of magnitude), the agent can always adapt to the environment. Moreover, for greater time-scale separation with smaller values of τ , there is less variability in the asymptotic regret (see Fig 4D ).
To understand the mechanism underlying this effect, we more closely examined the learning dynamics of individual model instances over the different τ values, especially the τ = 1 and τ = 0.01 cases. As shown in Fig 4E , in the case of τ = 1, the network is able to learn solutions in each context; however, upon switching, regret again accumulates, indicating an overwriting of prior strategies as comparable to the phenomenon of catastrophic forgetting. On the other hand, neuron-astrocyte networks with time-scale separation are able to reliably learn the flip-flop bandit, indicating that they are able to gradually associate the contextual information with the environment and protect previously learned trajectories. As shown in Fig 4F , the astrocyte-mediated meta-plasticity appears to be engaged during the process of learning. Specifically, we projected the trial-wise network activity along population vectors associated with astrocytes ( PC z ) and synaptic weights ( PC w ). We observed that during learning, the network forms distinct synaptic trajectories that asymptotically approach a fixed weight configuration. The time-scale separation between astrocytic and synaptic activation is apparent when tracing the initial stages of the trajectories. The astrocyte output is less sensitive overall to learning, likely an important factor in preventing the context-wise overwriting of prior dynamics (see also Discussion ). Furthermore, another projection along the second principal component of astrocytic activity shows that the sign of astrocytic modulation changes across trials and contexts (see Section F.8 in S1 Appendix ), which indicates that astrocytic modulations of meta-plasticity can be heterogeneous depending on task circumstances.
Discussion
Toward a fuller accounting of brain circuit dynamics
In this paper, we have examined the potential role of neuron-astrocyte interactions in context-dependent learning, with a specific focus on reinforcement-based bandit problems. We began by forming a simplified model of such interactions in the form of a dynamical system, leveraging canonical descriptions of neural firing rate activity and several abstractions of astrocytic activity and modulation that are based on extant neurobiological theory. In particular, we simplified the dynamical description of astrocytes and focused on key aspects: (i) their orders-of-magnitude time-scale separation from neurons, (ii) their modulation of synaptic processes, (iii) their indirect modulation of neuronal firing rates, and (iv) their ability to engage in response to contextual information. Our goal was to understand whether these aspects of neuron-astrocyte interaction, which are known to exist in the brain, matter for network computation.
Contextually-guided meta-plasticity and slow modulatory dynamics
From this perspective, our analysis indicates the potential for astrocytes to reshape neural and synaptic vector fields in significant ways, such as in the formation of multiple stationary regimes of activation and changing the geometry of synaptic weight evolution. Perhaps most notably, astrocytes can modify the dynamics of synaptic plasticity, effectively switching the network between slow and fast weight adaption regimes (see Fig 4F ). This forms a powerful mechanism by which astrocytes can use external and internal contextual information [ 8 ] to shift networks between different modes of learning, which we view as a form of meta-plasticity in the wide sense of that term.
One new and central premise to our work is the use of a contextual signal that is accessible by astrocytes and neurons, with the premise that such a signal may embed task-relevant information and/or other circuit contexts, which is highly consistent with the body or work showing astrocytes ability to detect and respond to functionally salient physiological covariates such as neuromodulators (e.g., dopamine), hormones (e.g., glucocorticoids), or local cytokines. Our abstraction of this signal may be viewed as overly strong, insofar as it presents ‘clean’ context information to the network. From this perspective, we emphasize that all our alternative architectures, and especially the neuron-astrocyte model without time-scale separation, had access to this information. Thus, it is not merely the presence of contextual signaling that augments learning performance in our model, but the specific astrocyte-dependent dynamical mechanisms by which this information alters neurons and synapses.
Following the above, we emphasize that our goal was not simply to add slow modulation to neuronal networks, since this could be done in a myriad of ways. Rather we chose a specific, biologically motivated and hypothesized pathway (involving astrocytes). While we cannot exclude the role of all other slow processes, we do believe that the astrocyte meta-plasticity pathway has unique advantages. To illustrate this, we pose a simple null model where we eliminated astrocytes but added a slow, passive potassium gating to neurons (see Section F.7 in S1 Appendix ). This null model enacts a slow time-scale, but without the nested feedback loop structure that we believe is key to neural-astrocyte interaction. After the same training procedure as in our primary results, we find that this null model can only learn the stationary (but not the non-stationary) version of the task (see Section F.7 in S1 Appendix ). This lends credence to the idea that the unique interactivity properties of astrocytes with neurons are important to the hypothesized functional benefits.
Consistency with biology of spatiotemporal neuron-astrocyte interactions
In this paper, we have focused most of our attention on the temporal separation of neuronal and astrocytic activity and have discussed the consistency of our parameters with known biophysics in this regard. However, equally interesting are spatial features such as the ratio of astrocytes to neurons. In biology, the ratio of astrocytes to neurons is believed to be between 1:1.5 to 1:2 and we used the latter end of this range in our results. However, we also performed a sensitivity analysis to examine whether dynamics change appreciably as a function of this ratio. In the stationary case, an increased ratio positively influences learning performance, leading to a reduction in asymptotic cumulative regret as more astrocytes are introduced. However, the dynamic nature of the flip-flop bandit introduces a subtle impact on the ratio: optimal cumulative regret occurs at intermediate ratios, whereas both excessively low and high ratios detrimentally affect learning by increasing regret (see Section F.6 in S1 Appendix ). It turns out that the optimal ratio is around 7/10, hence our simulation observation is consistent with the biology and indeed predicts a functional optimum within this range.
Astrocytic activity as a stabilizer of catastrophic forgetting
Catastrophic forgetting is a phenomenon in artificial neural networks that arises when networks are tasked with learning multiple tasks sequentially [ 63 ]. In this scenario, it often is the case that previously encountered tasks are ‘overwritten’ when the algorithmic optimization (i.e., learning) strategies are deployed to update the network parameters/weights to meet new task demands. Our results indicate that astrocytic modulation of neuronal and synaptic dynamics mitigates catastrophic forgetting. Here, we believe that the slow time-scale of astrocytes is instrumental in protecting previously learned network outputs upon encountering of a new context. As described above, the slow activation of astrocytes makes them generally less sensitive to parametric adjustment relative to neurons and synapses. Thus, their effects are more stable context-to-context. Furthermore, as we have seen, astrocytes have the effect of controlling neuronal and synaptic dynamics, such that those faster processes can occupy distinct regions of state space depending on astrocytic modulation. The combination of these two phenomena means that astrocytes can effectively insulate the learned trajectories/dynamics of one context preventing overwriting when learning is engaged for a subsequent context. These findings underscore the importance of dynamical heterogeneity in the brain and support the functional advantages that astrocytes may confer.
Clearly, an important next step for these models will be to validate them with appropriate experiments. As mentioned in the Introductions, tools for in vivo study of astrocyte function have lagged relative to those for neurons. For those tools that do exist, e.g., to disrupt astrocyte function [ 64 ], studies such as ours can identify salient behavioral paradigms within which experiments may be conducted. Specifically, our model suggests that astrocytes contribute to learning in context-dependent or, potentially, multi-task settings, more than they might in simpler behavioral paradigms. This can be tested using adequate history-dependent learning tasks and popular astrocyte silencing tools such as CalEx [ 65 ]. Hence, available tools could be deployed to test formally the role of astrocytes in such scenarios. For example, by examining the learning efficacy of rodents engaging multi-arm bandit paradigms [ 66 ].
Insights into algorithmic learning systems
While our goal in this paper has been to explore new theories regarding the potential significance of neuron-astrocyte interactions in the brain, it is nonetheless interesting to consider the implications of these results in the domain of algorithmic systems. We have already commented on the fact that traditional algorithmic methods of learning bandit tasks have difficulty in context-dependent settings, even in the presence of informative signaling. This begs the question of whether neuron-astrocyte type architectures may have utility beyond the bandit/reinforcement learning settings.
In this regard, there certainly exist recurrent neural networks designed to deal with multiple time-scale features, notably LSTMs [ 67 ] and hierarchical RNNs [ 68 ]. The LSTM has an internal memory cell state that enables it to deal with tasks that involve long-term dependencies. In hierarchical RNNs, multiple layers of RNNs are stacked on top of each other, where each layer captures information at a different level of temporal abstraction. The lower layers focus on short-term dependencies, while the higher layers focus on longer-term dependencies. The multi-scale neuron-astrocyte network considered here is in the form of a feedback-connected multi-layered network with different embedded time-scales, and hence may blend the different features of these extant machine learning architectures. It is thus possible that this framework may be extendable to other machine learning domains, especially ones involving disparate time scale requirements such as meta-learning [ 69 ].
Limitations and features not explained
Our model suggests a key role of slow astrocytic modulation of synaptic plasticity in enabling learning over long time-scales. This model effect was premised on prior theory and empirical findings, including [ 14 ]. However, astrocyte interactions with synapses are heterogeneous across and within brain regions (see [ 70 ] for instance) both in extent (number of synapses impinged upon) and nature (pre-, postsynaptic, or both). Hence the effects we show in this paper should be interpreted as a demonstration of sufficiency rather than necessity, and certainly not monolithically across brain areas. Importantly, we do not imply that all slow brain dynamics are enacted by astrocytes. Indeed, recent evidence indicates that slow oscillatory activity in the entorhinal cortex may enable information processing across minutes or longer [ 71 ]. Our findings demonstrate that: (i) astrocytes are particularly apt to convey active, calcium-mediated modulation of synaptic dynamics, and (ii) this modulation is particularly potent in learning scenarios, relative to more diffuse and passive slow dynamics (e.g., potassium, as discussed previously above). As also pointed out above, these results set up clear potential for future empirical work aimed at the disruption of astrocytic calcium signaling in complex function.
Other important facets of astrocytes are their gap junction coupling, which is believed to be a basis for a form of functional inter-connectivity, as well as their structured arrangement in non-overlapping domains throughout the entire brain, something referred to as tiling [ 72 ]. While our model described nested feedback loops of astrocyte-neuron modulation (i.e., connectivity), we did not explicitly explore the role of tiling and gap-junction couplings, leaving this question as future work.
Methods
Multi-scale modeling of neuron-astrocyte network dynamics
In general, neural dynamics can be described by recurrent neural network models. Here, we consider the biology-inspired continuous-time RNN (CTRNN) [ 57 , 73 ]. Consider a group of n neurons where each neuron is connected to some other neurons via synapses. Let x i ∈ R be the state of the unit i , which denotes the mean membrane potential of the neuron. Then, the model of CTRNN is defined by ODEs
τ n x ˙ i = - a i x i + ∑ j = 1 n w i j ϕ ( x j ) + u i , i = 1 , … , n , (4)
where τ n > 0 and a i > 0 are the time constant and decaying parameter respectively, and u i is the external input to unit i . ϕ ( x j ) is the activation function. It is noted that each unit i collects the outputs ϕ ( x j ) (i.e., short-term average firing frequency) from all the connected neural units in the network, weighted with the synaptic connection coefficients w i j ∈ R , where the positive or negative w ij indicates an excitatory or inhibitory synapse respectively.
Synapses are capable of modifying their strength via synaptic plasticity, which is usually formulated as a learning rule where the change of a synaptic strength w ij depends on the correlation between the firing rate of a presynaptic neuron j and the firing rate of the postsynaptic neuron i . We consider the Hebbian learning rule: the weight between two neurons strengthens when they are correlated and weakens otherwise. This rule is defined mathematically by the equation [ 74 ]
τ w w ˙ i j = - b i j w i j + c i j ϕ ( x i ) ϕ ( x j ) , (5)
where b ij > 0 is the decaying parameter; τ w > 0 is the time constant; c i j ∈ R is a parameter which indicates an existing synaptic connection when it is non-zero. When c ij takes a positive value, (5) is called the Hebbian learning , and the case with c ij < 0 is anti-Hebbian learning .
In principle, neurons influence astrocytes by releasing neurotransmitters that induce calcium ion elevations within astrocytes. Biophysically, the increase in calcium levels within individual astrocytes can propagate to neighboring astrocytes over long distances, forming calcium waves [ 75 ]. The mechanisms of this propagation may involve astrocyte to astrocyte gap junctions, which are well validated biologically [ 59 , 76 ] and are believed to form spatially contiguous groups of astrocytes referred to as a network [ 77 ] (but also see astrocyte syncytium [ 78 ]). Current biophysical mathematical models for astrocytes, including gap junction connectivity, are excessively complex and not easily translatable for analytical and computational purposes. In the development that follows, we propose a simplified model to describe astrocyte dynamics based on [ 79 ], which abstracts the mathematical description of astrocyte-to-astrocyte connectivity within a network formulation.
Consider a group of m astrocytes. Let z k ∈ R be the state of astrocyte k which denotes the activity of calcium wave. For the glial node z k , we assume the output of astrocyte calcium wave is similarly defined by an activation function. To distinguish it from the neuron, we use a different function, for instance, the hyperbolic tangent function ψ ( z k ) = tanh( z k ). Then, in the absence of neuron-astrocyte interactions, the dynamics of z k is described by
τ a z ˙ k = - e k z k + ∑ l = 1 m f k l ψ ( z l ) + v k , k = 1 , … , m , (6)
where τ a is a constant time parameter; f kl denotes the weight of the (network) connection from astrocyte l to k ; v k captures other external inputs. The usage of this phenomenological model can be justified with analogous arguments in [ 34 ], where a neuronal leaky integrate-and-fire model is used for astrocytes. Such a model is easy to modify to incorporate the neuro-synapse-astrocyte interactions and greatly facilitates the numerical and analytical investigation as shown in the first subsection of Results.
Stacking all the equations of neurons, synapses, and astrocytes together, we will arrive at the mathematical model for the neuron-astrocyte network as a whole.
τ n x ˙ i = - a i x i + ∑ j = 1 n w i j ϕ ( x j ) + u i , i = 1 , … , n , (7a) τ w w ˙ i j = - b i j w i j + c i j ϕ ( x i ) ϕ ( x j ) + d i j ψ ( z k ) , i , j = 1 , … , n , (7b) τ a z ˙ k = - e k z k + ∑ l = 1 m f k l ψ ( z l ) + h k ϕ ( x i ) ϕ ( x j ) + v k , k = 1 , … , m , (7c)
where the additional terms d ij ψ ( z k ) and h k ϕ ( x i ) ϕ ( x j ) with d i j , h k ∈ R are present to capture the high-order interaction between neurons, astrocytes and synapses according to the description in tripartite synapse structure. In system (7) , there are n and m equations for x and z respectively. The number of synaptic connections is flexible and denoted by o with m ≤ o ≤ n ( n − 1). Therefore, the dimension of system (7) is actually ( m + n + o ).
It is known that the activities of neurons, synapses, and astrocytes evolve on different time-scales. Neural firing occurs in milliseconds, synapse plasticity changes at a slower speed, and astrocyte processes take even longer, ranging from seconds to minutes. These varying time-scales significantly impact information processing in neuron-astrocyte interactions. To investigate the effects of these differences, we need to set the time-scale parameters, denoted as τ n , τ w , and τ a , to different values. To make the speeds of the evolution of these variables distinguishable, we have the assumption: 0 < τ n ≪ τ w ≪ τ a , with ≪ indicating the former entity is much smaller than the latter. As the main goal of this work is to study neuron and astrocyte computation, we set τ n = τ w for simplicity when applying the neuron-astrocyte model to solving the tasks.
Dynamic context-dependent multi-armed bandit tasks
In the setting of a stochastic MAB, there is a set of actions (arms) A to choose from, and the bandit lasts T rounds in total. In each round t , an agent (decision-maker) chooses one action a t ∈ A and obtains a reward r t . The goal of the agent is to optimize the accumulated reward, i.e., max a t ∈ A ∑ t = 1 T r t . We consider the Bernoulli bandits which belong to stochastic MABs. In the context of Bernoulli bandits, the reward of each action is binary, either 1 or 0 depending the outcome is a success or failure. The reward r i of the i -th action is drawn from a Bernoulli distribution, i.e.,
r i ∼ Bernoulli ( μ i ) , i = 1 , … , n ,
where μ i ∈ [0, 1] is a constant denoting the mean of the distribution. Different actions have different μ i where a larger value represents a higher probability of the successful outcome and thus a higher expectation of the reward. The reward sequence up to time T is a random process
{ r t ∼ { Bernoulli ( μ i ) } i = 1 n , t = 1 , … , T . } (8)
In the Bernoulli bandit, the goal of optimizing the accumulated reward is equivalent to minimizing the cumulative regret (3) . The standard Bernoulli bandit is stationary where all μ i are fixed over time. In addition to the stationary case, we further consider non-stationary variants by making the means changeable and time-dependent. Two subcases are considered in this work:
Flip-flop switching: the means μ i of actions remain constant for a certain period of time, and then abruptly transit to different values μ i ′ ∈ [ 0 , 1 ] at certain time instants.
Smooth changing: the means change according to a continuous function of time. Here, we use the periodic function
μ i ( t ) = μ i * S ( Q sin ( 2 π t P + 2 π i n ) ) , (9)
where μ i * is a fixed value in [0, 1]; S (⋅) is the sigmoid function; P is used to control the period of this function and the term 2 π i n makes that the action with the highest expected reward can change between the available actions over time. When Q is large, this type of function is dominated by an approximately constant value, such that it looks like a smooth square wave. We set P and Q to 10000 and 100 respectively.
In dynamic bandits, when the arm means change over time and the action with the highest mean switches, contextual information can be revealed to the agent. This contextual information represents the changes in underlying contexts. Therefore, the tasks we considered become context-dependent. We define the contextual signals as a scalar in all the simulations presented in this work. However, it is important to note that these signals can also be expanded into a multi-dimensional vector to accommodate more general settings.
Discrete-time neuron-astrocyte network
For simplification, we assume that the self-decay parameters are all one and the time-scales of neurons and synapses are the same. Then, the neuron-astrocyte network model without inputs can be rewritten in the compact form
τ x ˙ = - x + W ϕ ( x ) τ W ˙ = - W + C Φ ( x ) + D ψ ( z ) z ˙ = - z + F ψ ( z ) + H Φ ( x ) , (10)
where x = [ x 1 , …, x n ] ⊤ and z = [ z 1 , …, z m ] ⊤ are state vectors for neurons and astrocytes; W = [ w ij ] is the matrix for synapse weights and W ˙ denotes the element-wise derivative of W ; ϕ ( x ) = [ ϕ ( x 1 ), …, ϕ ( x n )] ⊤ and ψ ( z ) = [ ψ ( z 1 ), …, ψ ( z m )] ⊤ are vectors of activation functions while Φ( x ) is the flatten vector of the matrix [ ϕ ( x i ) ϕ ( x j )]; C , D , F , and H are the parameter matrices with corresponding entries.
In (10) , we have set the time constant for astrocytes to the unit, while time constants for neurons and synapses are both τ ≪ 1. In this way, τ is dimensionless and represents the time-scale difference rate between neurons and astrocytes. Note that (10) can be rewritten equivalently by a change of time, so that τ appears on the right hand side of z ˙ .
By using the first-order Euler discretization method [ 80 ], we can transfer the continuous-time neuron-astrocyte model to the discrete-time approximated version
x t = ( 1 - γ ) x t - 1 + γ W t - 1 ϕ ( x t - 1 ) W t = ( 1 - γ ) W t - 1 + γ ( C Φ ( x t - 1 ) + D ψ ( z t - 1 ) ) z t = ( 1 - γ τ ) z t - 1 + γ τ ( F ψ ( z t - 1 ) + H Φ ( x t - 1 ) ) , (11)
where γ is the discretization step size. In the following simulations, γ and τ are set to be 0.1 and 0.01 respectively. We use the sigmoid function ϕ ( x ) = 1/(1 + e − x ) and the hyperbolic tangent function ψ ( z ) = tanh( z ) for neural and astrocyte layer in the simulations.
We incorporate this discrete-time neuron-astrocyte model as the hidden layer within the entire learning network, where a pair of linear input and output layers are placed before and after the hidden layer according to the machine learning convention. The input I ∈ R | u | and the output y ∈ R | y | are feed into and read from neuron-astrocyte network after multiplied by matrices W in 1 , W in 2 and W out . Therefore, the network as a whole is represented by
x t = ( 1 - γ ) x t - 1 + γ ( W t - 1 ϕ ( x t - 1 ) + W in 1 I t ) W t = ( 1 - γ ) W t - 1 + γ ( C Φ ( x t - 1 ) + D ψ ( z t - 1 ) ) z t = ( 1 - γ τ ) z t - 1 + γ τ ( F ψ ( z t - 1 ) + H Φ ( x t - 1 ) + W in 2 I t ) y t = W out x t + b out , (12)
where b out the bias vector with the corresponding dimension.
Reinforcement learning procedure
A key step of our study is the implementation of our model to reinforcement-learning paradigms. Within this functional setting, at each trial, the agent (i.e., the neuron-astrocyte network) is presented with a new reward. This reward is used to algorithmically optimize (i.e., train) the parameters of the model in a trial-wise fashion. In other words, at the conclusion of the trial, the current parameters and outputs of the model, along with the current reward, are used to evaluate a loss function (see below) that determines future parameter adjustments. The research question at hand is whether the neuron-astrocyte architecture and dynamics enable this form of learning to be efficacious. Table 1 summarizes all parameters, both fixed and trainable, and their values in the model and training process.
10.1371/journal.pcbi.1012186.t001
Table 1
Parameters in the neuron-astrocyte model and model training.
Symbols
Description
Values
n
number of neurons
128
m
number of astrocytes
64
τ
time-scale parameter
0.01
γ
discretization step
0.1
I
contextual cues
stationary case: {1} flip-flop: {−1, 1} smooth: {−1, 0, 1}
W
i
n
1
,
W
i
n
2
input weight matrices
Trained
C , D
matrices associated with synapses
Trained
F , H
matrices associated with astrocytes
Trained
W out
output weight matrix
Trained
b out
output bias vector
Trained
The neuron-astrocyte network architecture comprises 128 neurons and 64 astrocytes (except for simulations where we vary the neuron-astrocyte ratio), with randomly initialized connections within each layer and interlayer hyperedges. The complete learning framework is depicted in Fig 3A . We first initialize the matrices C , D , F , H with entries drawn randomly from normal distributions with the zero mean, i.e.,
M i j ∼ 1 N M N ( 0 , 1 ) ,
where N M is the dimension of the focal matrix M . The elements of input and output matrices W in 1 , W in 2 , W out and bias vector b out are initialized from a uniform distribution U ( - 1 N M , 1 N M ) , where N M is again the dimension.
The dimension of the output y t is the same as the number of actions in the bandits, i.e., 3 in most simulations. After multiplied by the readout matrix and plus the bias, the output is fed to a softmax function, and it produces a probability distribution over the available actions p t = [ p t 1 , p t 2 , p t 3 ] . The probability of selecting the action a i ∈ A is
p t i = e y i ∑ 1 3 e y j , i = 1 , 2 , 3 . (13)
An action a t is then sampled from this probability distribution and subsequently executed by the agent. The bandit environment will provide the agent with a reward, represented as r a t . And according to [ 81 ], we use the loss function
L = ( r ¯ t - r a t ) log p t i ,
where r ¯ t is the average of rewards up to t and log p t i is the logarithm of the probability.
We adopted the traditional policy-based RL algorithm REINFORCE for the network training [ 81 ]. The gradient of the loss function L is calculated and used to update the network’s free parameters via the backpropagation (BP). During BP, we use the Adam method to optimize the aforementioned matrices and vectors with the default learning rate of 0.001.
In the case of other RNN-based methods as described in the comparison section below, we simply replace the neuron-astrocyte network with the alternative network models. To ensure a fair comparison of a comparable magnitude of training parameters, all these conventional RNNs are configured with 2 stacked layers, each consisting of 128 units. In the 2 stacked layers structure, the first layer is forward-connected to the second layer: the external input (contextual cues) is fed to the first layer, which has default trainable intra-connection weights; the output of the first layer is fed to the second layer as the input associated with a trainable matrix; and then the output of the second layer is further used to generate actions. The weights are initialized using a default method, and the training procedure remains consistent.
The network models and training procedures are implemented using PyTorch in Python.
Learning performance comparison
Numerous machine learning algorithms have been developed to tackle MABs. Among them, Upper Confidence Bound (UCB) and Thompson Sampling (TS) are widely recognized as the most prominent approaches for standard MABs. Discounted UCB (DUCB) and switching-window UCB (SWUCB) have been devised to handle changing environments in non-stationary scenarios. In addition to these canonical bandit algorithms, some neuro-bandit algorithms that utilize feedforward or recurrent neural networks to model the agent’s policy have been developed in recent years.
To perform a thorough yet not overly exhaustive assessment of learning performance, we analyze the asymptotic cumulative regret of our approach in comparison to selective algorithms across various scenarios. For stationary MABs, we evaluate our method against the UCB and TS algorithms, as well as RNN-based models including LSTM, vRNN, and GRU. In the context of non-stationary MABs, our method is compared to DUCB, SWUCB, and other RNN-based algorithms. It’s worth noting that the training procedures for all RNN-based models remain consistent with the previously described methodology.
Supporting information
S1 Appendix
The supplementary appendix file contains the mathematical analysis of the models and extensive simulation results stated in the main text.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Hearing loss is a common sensorineural disorder affecting one out of 500 live births, with increasing prevalence into adolescence [1] . While there are many environmental causes of hearing loss, such as viral infections, acoustic trauma, and ototoxic drugs, approximately half of all cases are hereditary [2] . Genetic causes of hearing loss can be detected by sequence analysis, which helps clinicians and patients to delineate the characteristics of disease. In addition, hearing loss occurring in early childhood can affect the linguistic development [3] , so it is quite important to improve our techniques to find genetic alterations in patients for further clinical care of this disease.
Nonsyndromic hearing loss (NSHL) contribute 70% of inherited hearing loss, and most of NSHL were autosomal recessive up to 80%, in comparison to 20% autosomal dominant and less than 1% X-linked or mitochondrial disorders [4] . About 46 genes have been identified and causally related to nonsyndromic hearing loss. However, over 100 loci have been mapped for monogenic hearing loss, with specific genes yet to be pinpointed [1] . The sheer complexity of the auditory system accounts for the large number of genes and loci linked to hearing loss.
Studies in hearing loss genes have increased rapidly with the advent of next-generation sequencing (NGS). NGS allows whole genome sequencing to be done quickly at a much lower cost than Sanger sequencing. Whole genome, whole exome, and targeted gene sequencing has become far more feasible, allowing for easier identification of disease genes [5] . Identification of deafness genes has several clinical implications. Because hearing loss is oftentimes monogenic, genetic testing can accurately predict the deafness phenotype. Genetic testing using NGS could provide an accurate, definitive answer with eliminating the need for further expensive testing [5] . Identification of genes and genetic testing will also allow the specific cause of a patient's hearing loss to be uncovered, which cannot be identified by Universal Newborn Hearing Screening [5] . Knowing the cause of hearing loss will allow prediction of the efficacy of certain therapeutic approaches, such as cochlear implantation, and in the future might allow for the development of further therapeutics and protective medications [3] .
Here we report the new diagnostic pipeline combining Sanger sequencing and targeted resequencing to find mutations in familial NSHL cases. Screening mutations in all exons in 80 reported deafness genes could detect candidate mutations in 13 (65%) out of 20 familial NSHL cases. Together with Sanger sequencing against four NSHL genes, the mutation detection rate was increased to 78.1% (25/32).
Materials and Methods
Ethics statement
This study was approved by the Institutional Review Boards (IRBs) at Seoul National University Bundang Hospital (IRB-B-1007-105-402 and IRB-B-1111-139-015) and Seoul National University Hospital (IRBY-H-0905-041-281). We obtained a written informed consent from all the participants in this study. In case of children participants, the written informed consent was obtained from the parents or guardians on behalf of them.
Patient selection
Among 145 hearing impaired probands who visited our tertiary referral center and who were willing to participate the genetic test from May 2010 through April 2012, 30 probands with syndromic features were excluded. Among the remaining 115 probands, 31 families with at least two or more hearing-impaired members without any syndromic feature (multiplex families) were selected and blood samples were taken. Medical histories were collected including age at onset of hearing loss, degree and progression of hearing impairment, and other relevant clinical manifestations.
DNA preparation, Sanger sequencing and targeted resequencing
Genomic DNAs were extracted from peripheral blood as described previously [6] . Sanger sequencing was performed using specific primers for each exon as described ( Table S1 ). Targeted exome sequencing was done by Otogenetics (Norcross, GA). Briefly, genomic DNA was used for NimbleGen capture methods (Roche NimbleGen Inc., Madison, WI) against 80 known deafness genes ( Table S1 ). An additional 50 bp of flanking intronic sequence were added to each exon and genomic intervals were merged using Galaxy software ( http://galaxy.psu.edu ). In total, we targeted 1,258 regions comprising 421,741 bp using NimbleGen methods.
Alignment, coverage calculation and variant detection
Reads were aligned to UCSC hg19 reference genome using BWA-0.6.1 with default settings [6] . To process sam/bam files and mark duplicates, Samtools and Picards were used. Local realignment around indels and base quality score recalibration were done for each samples and variants were called by unified genotyper in GATK-1.3. Perl script and Annovar were used to annotate variants and search the known SNPs and indels from dbsnp135 and 1000 genome draft. Coverages were calculated by GATK.
Model of independent uncaptured exons
To evaluate the correlation of capture performance between each samples, we compared experimental and expected distribution in number of exons that were commonly uncaptured within from 0 to 20 samples. Expectation number was calculated by the model assuming that uncapturing of exons occur independently between samples. Binomial model was not used because the difference of the numbers of uncaptured exons was not ignorable. We defined that uncaptured exon is an exon within which % of bases above 10, 50 or 100 of read depth is less than 1%.
p k : ratio of uncaptured exons in k th sample
P(# = n): probability that number of samples having uncaptured exons in common is n
Here, the number of samples is 20 and the number of exons is 1254. (sum for all combinations where, )
Due to too large number of combinations, each probability(P) was calculated with permutation 1,000 times with Python using a module “decimal” for precision, instead of summing all the combinations (but P(# = 0) was calculated directly.). Then, expected counts were obtained.
Results
We have collected 145 sensorineural hearing loss cases for a molecular genetic diagnosis in SNUH and SNUBH. Especially 32 multiplex familial cases were focused to find genetic aberrations for diagnosis and genetic counseling because we can validate the causative mutation through co-segregation in the family. We established the new diagnostic pipeline combining PCR and targeted resequencing. Eleven cases showed either clearly defined phenotype related to the mutations in SLC26A4 , POU3F4 and mitochondrial DNA genes ( Table 1 and Fig. S1 ). Temporal bone CT was taken to rule out any abnormality of the inner ear. Cases with characteristic radiologic markers such as bilateral enlarged vestibular aqueduct (5 probands) or incomplete partition type III (5 probands) were directly subject to further Sanger sequencing of the corresponding candidate genes, SLC26A4 and POU3F4 , respectively. Mitochondrial DNA was sequenced for one family that showed characteristic maternal inheritance of hearing loss. In these eleven families, we could successfully find mutations by PCR sequencing, which were mostly located in the reported sites ( Table 1 ). GJB2 sequencing was performed for the remaining 21 hearing impaired probands because the mutation in GJB2 was most frequent among familial NSHL cases. We found two cases (SJ19-19 and SH35-75) with known pathogenic mutations in GJB2 gene.
10.1371/journal.pone.0068692.t001 Table 1
Mutations of SLC26A4, POU3F4, GJB2 and MTRNR1 in 12 familial NSHL found by PCR-Sanger sequencing.
Patient
Characteristic phenotype
Gene
Mutation type
GeneBank No.
Chr
Exon
Nucleotide
Protein
MAF *
dbSNP135
SB02-6
Incomplete partition type III
POU3F4
Nonsynonymous
NM_000307
X
exon 1
c.686A>G
p.Gln229Arg
-
-
SB07-18
Incomplete partition type III
POU3F4
Frameshift deletion
NM_000307
X
exon 1
c.1060delA
p.Thr354GlnfsX115
-
-
SB08-19
Incomplete partition type III
POU3F4
Frameshift insertion
NM_000307
X
exon 1
c.950dupT
p.Leu317PhefsX12
-
-
SB09-21
Incomplete partition type III
POU3F4
Nonsynonymous
NM_000307
X
exon 1
c.632C>T
p.Thr211Met
-
-
SB13-29
Incomplete partition type III
POU3F4
stopgain
NM_000307
X
exon 1
c.623T>A
p.Leu208X
-
-
SB16-34
Nonsyndromic EVA
SLC26A4
Nonsynonymous
NM_000441
7
exon19
c.A2168G
p.H723R
0.001
rs121908362
SB23-54
Nonsyndromic EVA
SLC26A4
Nonsynonymous
NM_000441
7
exon19
c.A2168G
p.H723R
0.001
rs121908362
SB28-61
Nonsyndromic EVA
SLC26A4
Nonsynonymous
NM_000441
7
exon19
c.A2168G
p.H723R
0.001
rs121908362
SJ07-7
Nonsyndromic EVA
SLC26A4
Nonsynonymous
NM_000441
7
exon19
c.A2168G
p.H723R
0.001
rs121908362
SJ20-20
Nonsyndromic EVA
SLC26A4
Nonsynonymous
NM_000441
7
exon19
c.A2168G
p.H723R
0.001
rs121908362
SH07-19
Maternal transmission
MTRNR1
Nonsynonymous
Mt
1,555A>G
-
-
SJ19-19
no specific phenotype
GJB2
Frameshift deletion
NM_004004
13
exon2
c.299_300del
p.H100RfsX14
-
-
Frameshift deletion
NM_004004
13
exon2
c.235delC
p.L79CfsX3
-
-
* MAF: minor allele frequency from 1,000 Genome.
Next, we applied targeted resequencing for 20 probands of the remaining familial NSHL cases including one GJB2 positive multiplex family (SH35) to screen all 80 reported NSHL-related genes ( Fig. 1 ). We have captured 1254 exons of 80 genes ( Table S2 ) spanning 480 kb in 20 probands from multiplex families for targeted exome sequencing. Mean read depth in 20 cases was 218.2±56.1 and 88.9±3.7% of bases was read in more than ×10 coverage ( Table S3 ). About 90% exons in all patients were captured with ≥99% of bases at ≥10 of read depth. Missed or low-coverage exons were shared between samples, though different experimental procedures shared different uncaptured exons ( Fig. S2 ). This ensures that most of captured exons were shared through samples, which does not disturb the following analysis of variant detection. The fraction of well-captured exons was much more than expectation by the model of independent uncaptured exons ( Figs. S3 and S4 ).
10.1371/journal.pone.0068692.g001 Figure 1
Analysis flow of NSHL-80 targeted resequencing on familial NSHL.
Targeted resequencing data from 20 familial NSHL cases were filtered through five steps to select candidate SNVs in NSHL genes.
We selected rare single nucleotide variations (SNV) or indels following five steps of filtering to find candidate mutations related to hearing loss in each patient ( Table 2 ). In a basic filtering step, variations with a quality score of less than 20 were discarded, and for heterozygous alleles, only the alleles with a ratio (coverage of variant over the total coverage) of 20% or more were included. The average number of variants was 4.8±0.42 per patient after basic filtering. As a second step, we checked inheritance pattern of multiplex family of each proband ( Fig. 1 ), and excluded the variants which were not matched with the patient's inheritance pattern. According to the information on the inheritance, we could significantly reduce the average numbers of candidate mutations to 1.95±0.29 per patient (t-test p = 2.8×10 −6 ). In three families, all the mutations were not matched with the inheritance pattern. In the third step, we validated 39 variants from the 18 probands by Sanger sequencing and confirmed 36 variants (92.3%) ( Fig. 2 ). When we checked 80 normal hearing control subjects for the variants by Sanger sequencing, seven variants were also found in Korean population. As a final step for the filtering, we investigated segregation and/or phenotype matching to confirm the causality of the variant for deafness. Nineteen variants were examined in nine families by Sanger sequencing in all the family members to exclude 8 variants. We also examined patient's audiogram to match the candidate genes with the patient's phenotype and ruled out eleven variants, too. Especially, in cases where the segregation study could not be performed, we relied upon the audiogram configuration. Molecular genetic diagnosis was made in four subjects (SB61-109, SB55-102, SB50-94 and SB47-91) despite the lack of segregation study results, since their audiograms were well matched the previously reported characteristic audiogram configuration. Finally, we were able to find a most likely causative mutation in 13 out of 20 multiplex hearing loss families ( Table 3 ).
10.1371/journal.pone.0068692.g002 Figure 2
Validation of candidate mutations by PCR-Sanger sequencing.
Candidate mutations in 9 autosomal dominant and 4 autosomal recessive NSHL families were shown in chromatogram of Sanger sequencing.
10.1371/journal.pone.0068692.t002 Table 2
Number of candidate SNVs in 20 familial NSHL through five filtering steps.
Patient
1) basic filtering
2) inheritance pattern
3) Sanger sequencing
4) Control
5) Clinical feature
Final
Segregation
Audiogram profile matching
ADNSHL
SB14-30
8
1
1
1
-
1
1
SB40-77
6
3
3
2
-
0
0
SB41-78
3
1
1
0
-
-
0
SB50-94
3
2
1
1
-
1
1
SB54-101
5
4
4
3
1
1
1
SB55-102
1
0
-
-
-
0
0
SB60-107
6
3
2
2
1
1
1
SB61-109
4
2
2
1
-
1
1
SH14-37
4
2
2
1
1
1
1
SH20-47
5
1
1
1
1
1
1
SH21-50
4
3
3
2
1
1
1
SH40-89
6
5
5
5
2
1
1
SH41-90
2
1
1
1
0
0
0
ARNSHL
SB04-11
8
2
2
2
2
2
2
SB38-75
4
2
2
2
2
2
2
SB47-91
5
3
3
2
-
2
2
SH10-28
5
2
1
1
-
-
0
SH23-52
3
0
0
-
-
-
0
SH27-61
7
0
0
0
-
-
0
SH35-75
7
2
2
2
-
?
2
10.1371/journal.pone.0068692.t003 Table 3
List of final candidate SNVs in 13 familial NSHL.
Patient
Gene
Type
GeneBank No.
Chr
Exon
Nucleotide
Protein
Coverage of Ref
Coverage of Var
Quality score
1000g
dbsnp135
ADNSHL
SB14-30
WFS1
Nonsynonymous
NM_006005
4
exon8
c.T1235C
p.V412A
118
131
99
0.0037
rs144951440
SB50-94
COCH
Nonsynonymous
NM_001135058
14
exon4
c.T341C
p.L114P
18
10
99
-
-
SB54-101
OTOR
stopgain
NM_020157
20
exon2
c.G223T
p.E75X
47
49
99
-
-
SB60-107
MYO6
stopgain
NM_004999
6
exon8
c.C613T
p.R205X
63
55
63,55
-
-
SB61-109
COL11A2
Nonsynonymous
NM_080680
6
exon30
c.C2336T
p.P779L
85
69
99
0.0005
rs150877886
SH14-37
COCH
Nonsynonymous
NM_001135058
14
exon3
c.G113A
p.G38D
79
78
99
-
-
SH20-47
EYA4
Nonsynonymous
NM_172103
6
exon11
c.C909G
p.F303L
69
52
99
-
-
SH21-50
MYO6
stopgain
NM_004999
6
exon8
c.C613T
p.R205X
41
51
99
-
-
SH40-89
GJB3
Nonsynonymous
NM_001005752
1
exon2
c.G250A
p.V84I
125
123
99
0.0018
rs145751680
ARNSHL
SB04-11
OTOF
Frameshift deletion
NM_194322
2
exon24
c.3133delC
p.R1045Gfs*28
10
7
99
-
-
OTOF
stopgain
NM_194322
2
exon8
c.C1122G
p.Y374X
75
52
99
-
-
SB38-75
STRC
stopgain
NM_153700
15
exon20
c.C4057T
p.Q1353X
0
39
81.2
-
rs2614824
SB47-91
MYO3A
Nonsynonymous
NM_017433
10
exon7
c.C580A
p.P194T
117
109
99
-
-
MYO3A
Frameshift insertion
NM_017433
10
exon16
c.1582_1583insT
p.Y530Lfs*9
22
13
99
-
-
SH35-75
GJB2
Frameshift deletion
NM_004004
13
exon2
c.299_300del
p.H100Rfs*14
110
86
99
-
rs111033204
GJB2
Frameshift deletion
NM_004004
13
exon2
c.235delC
p.L79Cfs*3
123
104
99
0.0023
rs80338943
Among 32 familial NSHL cases, we could detect mutations in 25 probands (79.1%) by Sanger and targeted exome sequencing in total. Breaking into the results depending upon the inheritance pattern, we were able to make a molecular genetic diagnosis from 9 (69.2%) of 13 autosomal dominant families on. seven genes such as WFS1 , COCH , EYA4 , MYO6 , GJB3 , COL11A2 and OTOR . Molecular genetic diagnosis was possible in 9 (75.0%) of 12 recessive families. The four probably or possibly damaging mutations that we found were in SLC26A4, GJB2, MYO3A, OTOF, and STRC . We also found one case with MRNR1 mutation with maternal inheritance, and five cases of POU3F4 mutation with X-linked inheritance. However, we could not detect candidate mutations in seven probands, in which the number of variants from basic filtering was not correlated with read depth in 20 probands ( Fig. 3A ). The number of called variants, sequencing depth and mean coverage was not different from those with candidate mutations detected ( Figs. 3B and 3C ).
10.1371/journal.pone.0068692.g003 Figure 3
Interpretation of targeted resequencing in 20 probands.
(A) An average number of candidate SNVs with standard errors were shown at five filtering steps. (B) The relationship between the numbers of candidate SNVs and read depth were plotted in 20 probands. (C) Candidate SNV-found patient group (Found) was compared with patient group without candidate SNV (Not-found) in the number of candidate variants, read depth and called SNVs.
Discussion
Genetic cause of sensorineural disorders such as mental retardation, retinitis pigmentosa, and congenital hearing loss is extraordinarily heterogeneous. It is hard to detect disease-causing mutation in each patient because we have to screen all the candidate genes. However, the genetic diagnosis by high-throughput sequence analysis helps clinicians and patients to delineate the characteristics of disease. Especially early intervention of hearing loss in children might provide better clinical outcomes in the linguistic development. In this study, we could enhance the efficiency to find genetic alterations in familial NSHL patients. A candidate gene approach using conventional PCR sequencing against the candidate genes related to a certain phenotypic marker can cover only 10–20% of familial NSHL cases. Methodologies that enable us to effectively screen these common mutations related to the certain phenotypic markers, such as multiplex SNaPshot minisequencing, have been introduced in our population [7] . Howerver, a substantial portion of NSHL cases without any phenotypic marker still remains unanswered in terms of the molecular genetic etiology.
Therefore, we propose new diagnostic pipeline with high sensitivity to detect candidate mutations ( Fig. 4 ).
10.1371/journal.pone.0068692.g004 Figure 4
Proposed decision procedure for the genetic diagnosis of familial NSHL.
We have recruited 145 sensorineural hearing loss patients, Among 115 NSHL cases, we started with 32 familial NSHL because we could check the inheritance patterns in the family. First, we excluded 12 cases with typical clinical features by PCR-Sanger sequencing. In the remaining 20 familial NSHL probands, we found candidate SNVs in 13 probands. In further study, we can find SNVs by whole exome sequencing (WES).
Through the targeted resequencing of the 20 families, we found most likely responsible genes for nine out of thirteen AD families and four genes from seven AR families. Nevertheless, seven cases still need to find the final candidate mutation. The probands SB41-78 and SB55-102 have a relatively subtle phenotype considering their age (35 years old and 59 years old, respectively). Therefore, it is possible that their hearing loss is just a phenocopy. The proband SB40-77 showed characteristic mid frequency hearing loss', which rendered us to focus upon the candidate autosomal dominant genes such as TECTA or COL11A2 that has been reported to cause ‘mid frequency hearing loss. However, we could not detect a candidate variant in those genes. Rather, we found a potentially pathogenic variant (c.G5054A:p.R1685Q) in the MYH14 gene, a known deafness gene in DFNA4 locus from this proband. This variant has been detected neither in the 160 normal Korean control chromosomes nor in 1000 genomes. It was predicted to be ‘probably damaging’ by the Polyphen. In addition, the R1685 residue was conserved among many species including several mammalians, frog and zebrafish. However the audiogram pattern is not compatible with the previous reports [8] , [9] and we were not able to check the segregation of the variant due to other family members’ reluctance to participate in this study ( Table 2 ). Mutations in the different domains in the MYH14 might lead to different audiogram configurations. Currently, we are thinking that this MYH14 variant might account for the phenotype but did not count this as a causative mutation for this study. As for the family SH41, p.A2T variant of the OTOR gene (Accession No. AF233261) was detected after the basic filtering. This variant was not detected neither of normal 160 Korean chromosomes nor 1000 genomes. Robertson et al. (2000) proposed OTOR's possible role in human deafness based upon its preferential and abundant expression in the cochlea [10] . However, this variant did not co-segregate with hearing loss in one of the member in the family SH41.
Recently, narrow bony cochlear nerve canal (nBCNC) has been recognized and spotlighted as the most frequent inner ear anomaly [11] . Our group has postulated that the bilateral nBCNC may have a genetic etiology while the unilateral nBCNC is least likely to have a genetic contribution [12] . However, we were not able to find any candidate variant among the 80 deafness genes in the family SH27 where there was a sibling pair with bilateral nBCNC. The family SH23-52 segregates hearing loss presumably in an autosomal recessive manner, since the parents of three affected children showed perfect normal hearing. It is likely that the causative mutations for hearing loss in these families reside in genes other than 80 genes in this panel. We will further analyze the mutation in all exome of each family, because it may not be present in 80 candidate genes studied here.
Another reason for the detection failure may be due to the technical incompleteness. Coverage of targeted sequencing is not perfect to miss some exons, but usually considered good enough or not for the further analysis, especially in the experiments with many targets. This study also showed that 10% of exons were not properly captured. Capturing efficiency will be increased by new technologies for next generation sequencing. Recently, new enrichment technologies such as a semi-automated PCR amplification or a microdroplet PCR- based approach replacing the conventional hybridization-based enrichment technique have been successfully utilized in combination with next generation sequencing for genetic diagnosis of familial autosomal recessive deaf patients [13] , [14] . However, the diagnostic yield in these studies was not greatly different from those in ours and previous studies utilizing the hybridization –based enrichment technique [15] , [16] , rendering us to believe that technical incompleteness in capturing cannot solely account for the detection failure.
We found most likely responsible genes for nine out of thirteen AD families. Among seven autosomal recessive NSHL cases, we could detect the mutations in four genes such as GJB2, MYO3A, STRC and OTOF in four cases. One of them was mutations in GJB2, which were used as a positive control because it is well known for causing severe prelingual hearing loss as for the patient SH35-75. The sequence analysis of candidate genes may be easier to use PCR method, but clinical decision for candidate gene sequencing may be more difficult for all the clinician. To this end, we propose simple single step sequence analysis using targeted exome sequencing.
Supporting Information
Figure S1
Audiogram and pedigree for 12 familial NSHL used in PCR-Sanger sequencing.
(DOCX)
Figure S2
Audiogram and pedigree for 20 familial NSHL for targeted resequencing.
(DOCX)
Figure S3
Heatmap for percentage of bases ≥ depth 10, 50 or 100 within all target exons and samples. Most exons were uncaptured in common samples and samples were grouped by the common uncaptured exons.
(DOCX)
Figure S4
Barplot for comparison between experimental and expected distribution in number of exons that were commonly uncaptured within from 0 to 20 samples. Note that the exons in n = 0 were captured in all 20 samples and at n = 0, the experimental counts are greater than the expected ones due to common uncapured exons.
(DOCX)
Table S1
Primer sequences used for PCR-Sanger sequencing.
(DOCX)
Table S2
List of 80 genes related to NSHL for targeted resequencing.
(DOCX)
Table S3
Qualities of targeted resequencing in 20 familial NSHL.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
In recent years, with the development of clinical intensive care medicine, mechanical ventilation (MV) orotracheal intubation has been widely used in intensive care unit (ICU) [ 1 ]. However, the oral microbiota of MV patients begins to change on the first day of admission to the ICU, and pathogenic microorganisms quickly accumulate in the oral cavity, eventually leading to ventilator-associated pneumonia (VAP) [ 2 ]. VAP is defined as a pneumonia that occurs after 48 to 72 hours or later of endotracheal intubation. The diagnosis of this disease requires the fulfillment of ventilator-associated conditions (VAC) or infection-related ventilator-associated complications (IVAC) criteria, along with positive respiratory cultures or histopathological evidence [ 3 ]. It has been reported that 5%-40% of patients receiving invasive mechanical ventilation for more than 2 days develop VAP [ 4 ]. VAP remains an important disease in the ICU with significant economic burden, not only increasing mortality rates but also resulting in substantial costs. A recent cost evaluation from the Spain that the attributable cost of VAP was €20,965 [ 5 ]. Patients receiving mechanical ventilation are prone to rapid colonization of lower respiratory tract pathogens after intubation. Numerous studies over the past 20 years have shown that bacterial colonization of the oral-pharyngeal region is a key factor leading to respiratory infections in mechanically ventilated patients [ 6 – 9 ]. During intubation, colonized bacteria in the oral cavity and throat can migrate directly to the respiratory tract via the endotracheal tube, leading to VAP. Changes in saliva composition and function, altered oral pH, and increased plaque also occur in mechanically ventilated patients after intubation, which increases colonization of pathogenic bacteria in the oral-pharyngeal region and becomes the main source of pathogenic bacteria related to VAP [ 10 ].
Reduction of microbial counts in the oral cavity improves lung migration and colonization. As a common medicinal product, chlorhexidine (CHX) mouthwash is very effective in reducing pathogenic microorganisms including Streptococcus mutans and plaque [ 11 ]. However, with further research on CHX, many meta-analyses have found that CHX is only effective for patients undergoing major cardiac and vascular surgery [ 11 , 12 ]. Moreover, CHX has side effects such as tooth staining, unpleasant taste, dry mouth, allergies, and burning sensation [ 13 ]. Recent reports have suggested that CHX may have potential adverse effects on oral mucosa and decrease bacterial sensitivity. There is also a potential association between CHX oral care and increased mortality [ 14 ]. A recent study conducted a hospital-wide retrospective observational cohort analysis of the impact of CHX oral care on hospital mortality and confirmed this association [ 15 ].
Currently, the "US guidelines for the prevention of VAP" have downgraded CHX oral care from a routine recommendation in all hospitals to only being recommended for hospitals that have implemented more basic prevention measures but still have a high incidence of VAP [ 16 ]. Doctors and critically ill patients in the ICU tend to look for other mouthwashes with beneficial effects similar to CHX while reducing adverse effects. Herbal mouthwash has been considered a suitable alternative to CHX due to its lower side effects of its extracts and ability to reduce bacterial count [ 17 , 18 ].
To assess the effectiveness of herbal extracts and natural oral care products in lowering the incidence of VAP among ventilated patients, we performed a network meta-analysis (NMA) of randomized clinical trials. NMA is a statistical method that allows for the integration of direct and indirect comparisons of multiple treatment modalities within a single analysis, enabling estimation and ranking of their relative effectiveness. Our aim was to compare the efficacy of currently available herbal oral care products and provide guidance for future research and safer, more effective oral care protocols for ICU patients.
Methods
The NMA has been prospectively registered with the International Prospective Register of Systematic Reviews (CRD42023398022). In addition, this study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Network Meta-Analysis (PRISMA-NMA) guidelines [ 19 ], and we adhered to the recommendations of the Cochrane Handbook for Systematic Reviews of Interventions.
Search strategy
Two investigators independently searched relevant studies in this field from online databases, including China National Knowledge Infrastructure (CNKI), Wanfang Database, China Biology Medicine (CBM), PubMed, Web of Science, EMbase, and Cochrane Library, from the inception of the databases up to September 2022. Our data mining was based on keyword searches and combinations, including Oral care (1), Oral health (2), Oral hygiene (3), Oral decontamination (4), Intubation (5), Mechanical ventilation (6), VAP (7), VAP (8), Intensive care unit (9), ICU (10), herb (11), Herbal mouthwash (12), Herbal mouth rinse (13), Natural product (14), Chinese medicine (15), Chinese materia medica (16), Chinese herb (17), RCT (18). The retrieved records were imported into EndNote 21 for screening and deduplication. Any discrepancies between the two investigators were resolved by a third reviewer.
This study employed the PICOS heuristic method to identify eligible studies, which include Population/Participants, Intervention, Comparator, Outcome, and Study design criteria. Eligible participants (P) must be adults receiving mechanical ventilation in the ICU, regardless of gender, region, or any unspecified baseline characteristics. Patients with respiratory infections or other oral diseases were excluded. The eligible intervention (I) in randomized controlled trials involved the use of any natural products for oral care, including raw materials, herbal extracts, finished products, or products containing active ingredients such as traditional Chinese medicine, Miswak, chamomile extract, etc. The eligible comparator (C) in randomized controlled trials included the use of a placebo or standard treatment for oral care, such as CHX, saline, sodium bicarbonate, povidone-iodine, hydrogen peroxide, etc. The eligible outcome (O) in randomized controlled trials was the comparison of VAP incidence (primary outcome) and oral microbial quantity (secondary outcome) among patient groups. In addition to the above criteria, the eligible study design (S) must be a randomized controlled trial. Observational cohort studies, case-control studies, case reports, and reviews were excluded.
Study quality assessment
We utilized the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) tool [ 20 ] to assess the quality of the available evidence. The GRAD-Eprofiler software was employed to assign a rating of high, moderate, low, or very low evidence quality for each included outcome, based on study design, risk of bias, inconsistency, indirectness, imprecision, and publication bias. These factors may increase or decrease the quality of the evidence: (1) risk of bias (downgraded once when less than 75% of analyzed studies are at low risk of bias); (2) inconsistency (downgraded once when I 2 > 50%); (3) indirectness of evidence (e.g., indirect population, intervention, comparison, or outcome); (4) imprecision with wide confidence intervals; and (5) presence of publication bias also reduces the quality of the evidence.
Statistical analysis
The meta-analysis adhered to the PRISMA checklist ( S2 Table ). For binary data, odds ratios (ORs) were used to express treatment effects, while weighted mean differences (WMDs) were used for continuous (mean difference) data. Direct meta-analyses used the 95% confidence interval (CI), while the credible interval (CrI) was used for network meta-analyses. We conducted a network meta-analysis using Bayesian methods [ 21 ] in R (version x64 4.2.2) with the gemtc package (version 0.8–2) and rjags package (version 4–6), selecting the model based on comparisons of the DIC, ratio, and I 2 . Inconsistency of results was assessed by the node-splitting method and its Bayesian p-value, comparing the direct and indirect estimates for each comparison. A p-value < 0.05 indicated significant inconsistency. Potential ranking probabilities of treatments were estimated by calculating the rank probability matrix and the SUCRA values, with higher values indicating better efficacy.
Results
Selection and characteristics of the studies
A total of 538 articles were identified from databases including CNKI, Wanfang, CBM, PubMed, Web of Science, EMbase, and Cochrane Library, with the search period up to September 2023. We used EndNote software to remove duplicates and screened the titles and abstracts to exclude irrelevant studies. Out of 45 articles selected for full-text review, we included 29 studies that met the inclusion and exclusion criteria [ 22 – 50 ]. ( Fig 1 ). Among them, 24 articles reported incidence of VAP [ 22 – 45 ] and six articles reported colony number [ 40 , 46 – 50 ]. The included herbal mouthwashes were Miswak (stem of Salvadora persica), Persica ® (alcoholic extract of S. persica, Achillea millefolium, and Menthaspicata), Matrica ® (Chamomile extract), Listerine ® (main components include Menthol, Thymol, and Eucalyptol), Chinese herbs (extracted from multiple herbs), Orthodentol (extract of Khouzestani Savory, containing 30% carvacrol), and other oral care products including CHX, Normal saline, Sodium bicarbonate, Povidone-iodine, and Hydrogen peroxide. Three three-arm studies [ 22 , 29 , 48 ] and one four-arm study [ 46 ] were included. ( Table 1 ) ( Fig 2 ).
10.1371/journal.pone.0304583.g001
Fig 1
Literature search: Preferred Reporting Items for Systematic Reviews and Meta‑Analyses (PRISMA) consort diagram.
10.1371/journal.pone.0304583.g002
Fig 2
Network geometry graphs for changes on ventilator‑associated pneumonia(a), Colony Number (b).
10.1371/journal.pone.0304583.t001
Table 1 Characteristics of studies included.
author, year
age
Design
Treatment description
Berry 2013 [ 22 ]
arm1: 58.82(16.7)
Single blind RCT
arm1:NS+Rinsing,12/d+toothbrushing,3/d
arm2: 54.93(19.5)
arm2:Sodium bicarbonate+Rinsing,12/d+toothbrushing,3/d
arm3: 59.96(18.0)
arm3:Listerine ® +Rinsing,2/d+toothbrushing,3/d
Maarefvand 2015 [ 23 ]
arm1: 51.63±12.18
RCT
arm1:NS+Children’s toothbrush brushing+ Cleaning with cotton swabs,2/d
arm2: 45.93±14.11
arm2:Matrica+Brushing with children’s toothbrush+ Cleaning with cotton swabs,2/d
Irani 2019 [ 24 ]
arm1: 34.83±13.95
Single blind RCT
arm1:0.2%CHX+cotton swab Wiping,2/d
arm2: 33.65±13.50
arm2:Miswak+Scrubbing,2/d
Kawyannejad 2020 [ 25 ]
arm1: 46.7±14.76
Double blind RCT
arm1:0.2%CHX+Rinsing, 3/d
arm2: 47.78±11.22
arm2:Orthodentol+Rinsing,3/d
Shen ML 2020 [ 26 ]
arm1: 60.1±3.5
RCT
arm1:0.2%CHX+tooth brushing,3/d
arm2: 59.7±4.6
arm2:Chinese herb+ tooth brushing,3/d
Gao SM 2019 [ 27 ]
arm1: 66±5.3
RCT
arm1:0.12%CHX+scrubbing with cotton swabs, 4/d
arm2: 68±4.9
arm2:Chinese herb+ scrubbing with cotton swabs,4/d
Gao SM 2020 [ 28 ]
arm1: 58.92±7.02
RCT
arm1:0.12%CHX+scrubbing with cotton swabs, 4/d
arm2: 56.45±6.77
arm2:Chinese herb+ scrubbing with cotton swabs, 4/d
Gong CQ 2018 [ 29 ]
arm1: 42.52±7.84
RCT
arm1:NS+Wiping, 4/d
arm2: 46.35±10.28
arm2:NS+Brush teeth with a soft bristled toothbrush, 4/d
arm3: 46.35±10.28
arm3:Chinese herb+ Rinsing and Brush teeth with a soft brisrled toothbrush, 4/d
Han J 2021 [ 30 ]
NR
RCT
arm1:NS+Cleaning with cotton swabs, 4/d
arm2:Chinese herb+ Cleaning with cotton swabs, 4/d
He YQ 2012 [ 31 ]
arm:12–93
RCT
arm1:NS+Cleaning with cotton swabs,4/d
arm2:Chinese herb+ Rinsingand Cleaning with cotton swabs, 2/d
Jiang JJ 2021 [ 32 ]
arm1: 61.29±7.02
RCT
arm1:NS+Cleaning with cotton swabs, 2/d
arm2: 62.17±8.14
arm2:Chinese herb+ Cleaning with cotton swabs, 2/d
LI QY 2017 [ 33 ]
arm1: 76.6±9.8
RCT
arm1:Chinese herb+ Cleaning with cotton swabs, 3/d
arm2: 76.5±9.6
arm2:NS+Cleaning with cotton swabs,3/d
Liang YD 2020 [ 34 ]
arm1: 64.85±10.72
RCT
arm1:NS+tooth brushing,2/d
arm2: 66.79±9.42
arm2:Chinese herb+ tooth brushing, 2/d
Lu MY 2015 [ 35 ]
arm:25–85
RCT
arm1:NS+Rinsing,2/d
arm2:Chinese herb+ Rinsing, 2/d
Ruan LJ 2013 [ 36 ]
arm:36–89
RCT
arm1:NS+Rinsing,2/d
arm2:Chinese herb+Rinsing,2/d
She YX 2020 [ 37 ]
arm:68.9
RCT
arm1:Wiping+0.12%CHX,2/d arm2:Chinese herb+ Wiping, 2/d
Shi CH 2017 [ 38 ]
arm1: 57.3±4.2
RCT
arm1:NS+Rinsing,3/d
arm2: 56.1±3.9
arm2:Chinese herb+Rinsing,3/d
Wang L 2020 [ 39 ]
arm1: 59.6±7.2
RCT
arm1:NS+Wiping,2/d
arm2: 58.6±6.3
arm2:Chinese herb+ tooth brushing+ Rinsing, 2/d
Yang ZX 2014 [ 40 ]
arm1: 58.45±5.21
RCT
arm1:2%Hydrogen peroxide+ Cleaning with cotton swabs,4/d
arm2: 58.01±5.67
arm2:Chinese herb+ Cleaning with cotton swabs, 4/d
Yao YL 2020 [ 41 ]
NR
RCT
arm1:NS+Cleaning with cotton swabs,3/d
arm2:Chinese herb+ Cleaning with cotton swabs, 3/d
Zhang B 2020 [ 42 ]
arm1: 75.58±10.57
RCT
arm1:NS+Cleaning and Rinsing,3/d
arm2: 77.68±9.94
arm2:Chinese herb+ Cleaning and Rinsing, 3/d
Zhang CF 2016 [ 43 ]
arm1: 66.7
RCT
arm1:0.12%CHX+Rinsing and Wiping,4/d
arm2: 66.3
arm2:Chinese herb, Rinsing and Wiping, 4/d
Rezvani 2018 [ 44 ]
arm1: 59.18±20.15
Double blind RCT
arm1:0.2%CHX+Rinsing and scrubbing with gauze,2/d
arm2: 60.78±18.41
arm2:Matrika+Rinsing and scrubbing with gauze,2/d
Hafez 2015 [ 45 ]
arm1: 38.10±19.759
RCT
arm1:0.12%CHX+tooth brushing,6/d
arm2: 45.65±19.381
arm2:Miswak+tooth brushing,6/d
Haidari 2013 [ 46 ]
arm1: 49,6±1,31
Double blind RCT
arm1:0.2%CHX+Rinsing, 4/d
arm2: 52,35±1,51
arm2:0.12%CHX+tooth brushing,6/d
arm3: 50,45±1,13
arm3:10%Matrica ® +Rinsing, 4/d
arm4: 52,7±1,24
arm4:NS+Rinsing, 4/d
Baradari 2012 [ 47 ]
arm1: 47.52±7.1
Double blind RCT
arm1:0.2%CHX+Rinsing, 4/d
arm2: 47.56±8.6
arm2:Matrica ® (Chamomile extract) +Rinsing, 4/d
Taraghi 2011 [ 48 ]
arm1: 49.6±1.31
Double blind RCT
arm1:0.2%CHX+Rinsing, 4/d
arm2: 52.35±1.51
arm2:NS+Wiping, 4/d
arm3: 52.7±1.24
arm3:10%persica ® Wiping, 4/d
XU DQ 2021 [ 49 ]
arm1: 58.01±5.67
RCT
arm1:0.1%Povidone-iodine+Rinsing and Brushing with children’s toothbrush, 3/d
arm2: 58.45±5.21
arm2:Chinese herb+ Rinsing and Brushing with children’s toothbrush, 3/d
Zhang HY 2014 [ 50 ]
arm1: 70±8.68
RCT
arm1:NS+Rinsing and Wiping, 3/d
arm2: 70±8.99
arm2:Chinese herb+ Rinsing and Wiping, 3/d
Note: NS means Normal saline; CHX means chlorhexidine; RCT Randomized Controlled Trial.
Incidence of VAP
The direct and indirect comparison results showed consistency. Chinese herb was superior to Orthodentol (OR: 0.11, 95% CI: 0.02–0.51), CHX (OR: 0.32, 95% CI: 0.17–0.59), Listerine ® (OR: 0.16, 95% CI: 0.05–0.59), Chamomile extract (OR: 0.33, 95% CI: 0.13–0.79), Normal saline (OR: 0.16, 95% CI: 0.1–0.24), and Sodium bicarbonate (OR: 0.2, 95% CI: 0.07–0.52). Miswak was superior to Orthodentol (OR: 0.11, 95% CI: 0.02–0.64), Sodium bicarbonate (OR: 0.19, 95% CI: 0.14–0.93), Normal saline (OR: 0.16, 95% CI: 0.14–0.53), Listerine ® (OR: 0.17, 95% CI: 0.03–0.97), and CHX (OR: 0.33, 95% CI: 0.1–0.92). The other comparisons were not significant. ( Table 2 ).
10.1371/journal.pone.0304583.t002
Table 2 Indirect comparison of oral care solutions.
A)
VAP rate
Sodium bicarbonate
Normal saline
Chamomile extract
Listerine ®
Chlorhexidine
Miswak
Orthodentol
Chinese herb
Sodium bicarbonate
-
1.18(0.41,3.2)
0.43(0.09,1.99)
1.17(0.32,4.07)
0.5(0.15,1.75)
0.13(0.02,0.75)
1.51(0.23,10.99)
0.2(0.07,0.54)
Normal saline
0.85(0.31,2.41)
-
0.37(0.11,1.15)
0.99(0.27,3.42)
0.43(0.2,0.92)
0.11(0.02,0.48)
1.27(0.25,7.05)
0.17(0.11,0.25)
Chamomile extract
2.31(0.5,10.66)
2.71(0.87,8.8)
-
2.68(0.5,14.67)
1.15(0.5,2.79)
0.31(0.06,1.36)
3.47(0.65,19.68)
0.45(0.15,1.35)
Listerine ®
0.85(0.25,3.15)
1.01(0.29,3.65)
0.37(0.07,1.99)
-
0.43(0.1,1.92)
0.11(0.01,0.78)
1.29(0.17,10.84)
0.17(0.05,0.63)
Chlorhexidine
2(0.57,6.79)
2.34(1.09,5.01)
0.87(0.36,2.01)
2.34(0.52,9.97)
-
0.27(0.07,0.91)
2.98(0.72,13.95)
0.39(0.2,0.75)
Miswak
7.55(1.33,47.46)
8.92(2.06,42.9)
3.25(0.74,16.56)
8.75(1.29,66.76)
3.75(1.1,14.97)
-
11.41(1.75,87.4)
1.49(0.36,6.87)
Orthodentol
0.66(0.09,4.39)
0.79(0.14,4.04)
0.29(0.05,1.55)
0.78(0.09,5.92)
0.34(0.07,1.38)
0.09(0.01,0.57)
-
0.13(0.02,0.63)
Chinese herb
5.08(1.84,14.38)
5.97(4.07,9.05)
2.2(0.74,6.52)
5.97(1.6,21.2)
2.55(1.34,4.98)
0.67(0.15,2.76)
7.61(1.59,40.96)
-
B)
Colony Number
Chinese herb
Chamomile extract
Povidone-iodine
Hydrogen peroxide
Persica ®
Chlorhexidine
Normal saline
Chinese herb
-
1.64(0.82,3.33)
3.32(2.09,5.26)
1.04(0.56,1.91)
1.63(0.82,3.2)
1.04(0.53,2.03)
1.91(1.04,3.5)
Chamomile extract
0.61(0.3,1.22)
-
2.02(0.87,4.62)
0.63(0.25,1.58)
0.99(0.68,1.41)
0.63(0.46,0.85)
1.17(0.8,1.67)
Povidone-iodine
0.3(0.19,0.48)
0.49(0.22,1.15)
-
0.31(0.15,0.66)
0.49(0.22,1.1)
0.31(0.14,0.7)
0.58(0.27,1.22)
Hydrogen peroxide
0.96(0.52,1.77)
1.58(0.63,4.01)
3.19(1.51,6.82)
-
1.56(0.63,3.87)
1(0.4,2.47)
1.84(0.78,4.33)
Persica ®
0.62(0.31,1.22)
1.01(0.71,1.46)
2.04(0.91,4.64)
0.64(0.26,1.58)
-
0.64(0.47,0.87)
1.18(0.85,1.62)
Chlorhexidine
0.97(0.49,1.9)
1.58(1.18,2.16)
3.21(1.42,7.2)
1(0.41,2.49)
1.57(1.16,2.12)
-
1.85(1.35,2.52)
Normal saline
0.52(0.29,0.96)
0.85(0.6,1.25)
1.74(0.82,3.69)
0.54(0.23,1.28)
0.85(0.62,1.17)
0.54(0.4,0.74)
-
Note: A) Random-effects model. Negative values indicate improvement in clinical response. Changes that are statistically significant are shown in bold. B) The upper right triangle provides the pooled risk ratios from pairwise comparisons (column interventions relative to row interventions), while the lower left triangle summarizes the standardized mean differences and raw mean differences from network meta-analysis (row interventions relative to column interventions). Bold values indicate statistical significance with p<0.05.
Colony number
The direct and indirect comparison results showed consistency. In terms of reducing colony count, Chinese herb (OR: 0.3, 95% CI: 0.19–0.48) and CHX (OR: 0.54, 95% CI: 0.4–0.74) were superior to Normal saline. Hydrogen peroxide was superior to Povidone-iodine (OR: 0.31, 95% CI: 0.15–0.66). Chinese herb was superior to Povidone-iodine (OR: 0.33, 95% CI: 0.19–0.48). CHX was superior to Chamomile extract (OR: 0.63, 95% CI: 0.46–0.85), Povidone-iodine (OR: 0.31, 95% CI: 0.14–0.7), and Persica ® (OR: 0.64, 95% CI: 0.47–0.87). The remaining comparisons were not significant. ( Table 2 ).
Treatment ranking
Based on the research results, rank probabilities were used to generate a ranking graph ( Fig 3 ). The rank. Probability function was used to determine the top two interventions for each outcome measure and compared with the SUCRA (Rank) results. The ranking of interventions in reducing VAP incidence were as follows: Miswak (94.25%), Chinese herb(88.25%), Chamomile extract (62.94%), CHX (57.72%), Sodium bicarbonate (31.14%), Listerine ® (25.2%), Normal saline (21.97%), Orthodentol (18.23%). The ranking of interventions in reducing Colony Number were as follows: Chinese herb (82.48%), CHX (81.93%), Hydrogen peroxide (76.56%), Persica ® (43.48%), Chamomile extract (42.07%), Normal saline (20.94%), Povidone-iodine (2.54%).
10.1371/journal.pone.0304583.g003
Fig 3
Treatment ranking for each assessed outcome: Incidence of ventilator‑associated pneumonia(a), Colony Number (b).
Publication bias
After evaluating the funnel plot and Egger’s test (VAP rate: z = -2.3167, p = 0.0205; Colony Number: z = -0.5110, p = 0.6094), the results showed significant asymmetry in the funnel plot. Although the test result showed no bias in VAP rate, we used the trim-and-fill method and found that 6 studies with non-significant results were missing, indicating some degree of publication bias ( S1 Fig ).
Recommendation of evidence
According to the GRADE evaluation guide [ 20 ], grade-profiler was used to evaluate literature quality. All levels of evidence supporting this result were rated moderate to very low. Due to the limitations of the study, the included herb related studies were all from China and Iran, and there was no double-blind study on Chinese herb related studies. Therefore, the results presented in this NMA should be treated with caution. ( S1 Table ).
Discussion
After screening, we included the following herbal mouthwashes: Miswak (stem of Salvadora persica), Persica ® (alcoholic extract of S. persica, Achillea millefolium, and Mentha spicata), Matrica ® (Chamomile extract), Listerine ® (main components include Menthol, Thymol, and Eucalyptol). Fig 4 shows the two-dimensional plot of efficacy and colony count of all herbal and natural oral care products.
10.1371/journal.pone.0304583.g004
Fig 4
A two-dimensional graph showing the incidence rate (OR) and Colony Number (OR) of Herbal oral care products.
Consistent with the results of clinical double-blind trials [ 44 ], we found that Matrica ® significantly reduced microbial counts but did not affect the incidence of VAP. In contrast, traditional Chinese medicine and Miswak were more effective than CHX in reducing the incidence of VAP, according to our network meta-analysis comparison. Therefore, we consider Miswak and Matrica ® as the most promising herbal oral care products to potentially replace CHX. However, the safety and feasibility of traditional Chinese medicine require further high-quality research for confirmation. It is important to note that our study provides directions for future research rather than definitive conclusions.
Initially, it is important to clarify that although Chinese herb showed superiority in our NMA, the quality of research on Chinese herb is extremely low, with most studies lacking blinding and randomization. Chinese herb is a complex multi-herb preparation, containing several herbs such as honeysuckle [ 51 ] and peppermint [ 52 ], and we speculate that the efficacy of Chinese herb may come from these herbs. However, Chinese herb requires boiling and residue removal from multiple herbs, making it difficult to ensure quality and may lead to differences in efficacy and potential safety risks. Furthermore, it is not feasible for large-scale promotion since the included studies mostly used self-made Chinese herb preparations. Therefore, we do not recommend the use of Chinese herbs. All studies have not mentioned any potential side effects and there is no related research on the matter. We believe that the safety of Chinese herbs is questionable.
According to the node-splitting method, there was no significant inconsistency among the included studies, and the overall network structure was reliable (P>0.05). ( S2 and S3 Figs).
Our results are similar to some of the findings in a recent systematic review [ 53 ]. However, our study showed that Chamomile extract had a significant advantage over CHX in reducing Colony Number (OR: 0.63, 95% CI: 0.46–0.85), but no significant advantage over CHX in reducing the incidence of VAP (OR: 1.15, 95% CI: 0.5–2.79), which is different from the clinical trial results [ 23 ]. We believe that the reason for this contradictory conclusion may be due to limitations in our study’s inclusion criteria, as we did not conduct a detailed meta-analysis of Colony Number for VAP pathogens such as Staphylococcus aureus and Pseudomonas aeruginosa [ 54 ]. We only compared the total number of colonies, which may explain the difference in our results compared to some clinical trials. In addition, the number of articles included in our study was limited, with most coming from China and Iran.
Apart from multi-ingredient mouthwashes, we believe that Miswak is an oral care product that can replace CHX. Although there are currently no long-term studies or large-scale prospective studies on the use of Miswak, Miswak has a long history of use as a traditional oral care product [ 55 , 56 ]. In addition, the World Health Organization recommends and encourages the use of Miswak as an effective oral hygiene practice. Most importantly, Miswak can simultaneously act as a toothbrush and oral care solution, and its effectiveness in improving oral health in the general population has been demonstrated [ 57 ].
Despite being less effective than CHX in both NMA and clinical trials, Matrica ® has shown to be more effective than saline in reducing Colony Number and VAP incidence in ICU patients, consistent with clinical results [ 47 ]. Considering the antimicrobial properties of Matrica ® herbal mouthwash demonstrated in current and previous studies, it has been shown to have a beneficial effect in reducing plaque and gingivitis [ 58 ], with stronger resistance and significantly fewer side effects compared to chemical mouthwashes. Chamomile extract is considered safe for oral care, with the only potential short-term severe side effect being allergic reactions [ 59 ]. Further research is needed to investigate long-term side effects and mortality rates. Chamomile extract may have the potential to prevent VAP in ICU patients.
As for the three herbal oral care products Listerine ® , Orthodentol, and Persica ® , our results are similar to clinical studies [ 22 , 25 , 47 ]. Although they are also superior to Normal saline in terms of antibacterial ability, there are currently significant issues with these products, and we do not recommend them. Firstly, the safety of Listerine ® is very questionable as it contains alcohol, and its component thymol has cytotoxicity. Oral microorganisms can metabolize ethanol into acetaldehyde, causing DNA damage, and may even be associated with oral cancer, raising concerns about the dosage and concentration used [ 60 ]. Secondly, Orthodentol is an herbal mouthwash with a natural formula, and its component Khouzestani Savory is a native plant in Iran that contains 30% carvacrol. Although studies have shown that it has many antibacterial effects and significant effects on toothache [ 61 ], there are no short-term and long-term in vivo toxicology data available for carvacrol. In vitro experiments have provided sufficient evidence of mild to moderate toxicity on the cellular level [ 62 ]. Additionally, Persica ® herbal mouthwash contains three medicinal plants, Persica Salvadora, Milenrama, and mint. These plants can be consumed without short-term safety concerns, which is an advantage. However, some studies have shown that it does not seem to have an effect on gram-negative bacteria, and our NMA results also show that it is not ideal [ 63 ].
Despite numerous studies demonstrating the side effects of CHX and guidelines recommending against its use, it is still widely used in various patient populations, especially ICU patients [ 64 , 65 ]. Currently, although there are many herbal oral care products available, there is still limited research on herbal oral care products for ICU mechanically ventilated patients, particularly with regards to long-term follow-up studies on side effects and mortality rates. Therefore, we believe that further research is needed to explore Miswak, chamomile extract, and Chinese herbs in greater depth. While the safety of Persica ® can be ensured, its efficacy appears to be lower. Further research is needed to confirm the safety of using Listerine ® and Orthodentol in MV patients.
Limitation
This study investigates the efficacy of herbal oral care products on ICU patients. However, several limitations need consideration. Firstly, the data for this study is derived from published literature, and potential publication bias was assessed through funnel plot and Egger’s test. Secondly, the broad inclusion criteria resulted in varying study qualities, with only six studies employing blinding methods. Some studies lacked clarity on randomization methods, potentially impacting the reliability of results. Thirdly, variations in the frequency of herbal oral care product usage and accompanying oral care protocols introduce heterogeneity, potentially affecting their effectiveness in preventing ventilator-associated pneumonia and reducing oral microbial counts. Lastly, the study exclusively evaluates the impact of herbal oral care products on the incidence of ventilator-associated pneumonia and oral microbial counts. Important outcomes related to oral care, such as long-term safety, oral mucosal damage, gingivitis, were not addressed. These outcomes may influence the prognosis and comfort of mechanically ventilated patients, necessitating further clinical research to fill these data gaps.
Therefore, the conclusions drawn from this study should be interpreted cautiously. Future research demands high-quality, large-sample, multicenter, double-blind randomized controlled trials to assess the long-term effects and safety of herbal oral care products. Additionally, investigating other relevant outcomes is crucial for providing robust evidence supporting the application of herbal oral care products in mechanically ventilated patients.
Conclusion
Based on our network meta-analysis, we have observed that Chinese herbal medicine and Miswak are superior to CHX in reducing the incidence of VAP. However, the safety and feasibility of traditional Chinese herbal medicine require further high-quality research for validation. Simultaneously, Matrica ® demonstrates a significant reduction in microbial counts but does not exhibit a significant advantage in lowering the incidence of VAP. This observation aligns with the results of clinical double-blind trials. Therefore, we identify Miswak and Matrica ® as promising herbal oral care products with the potential to replace CHX. It is essential to emphasize that our study provides guidance for future research rather than conclusive determinations.
Supporting information
S1 Fig
Funnel plot of the network meta-analysis indicating publication bias.
(TIF)
S2 Fig
Node-splitting P value.
(TIF)
S3 Fig
Inconsistency test result.
(TIF)
S1 Table
Quality assessment by GRADE.
(DOCX)
S2 Table
PRISMA 2020 checklist.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Hepatic Glycogen Storage Diseases (GSD) are genetic disorders caused by deficient activity of one of the enzymes involved in the glycogenolysis pathway. The global incidence is estimated at 1 case per 20,000–43,000 live births. The most common types of GSD are GSD I, GSD III and GSD Ixα [ 1 ].
In GSD I, glucose-6-phosphate cannot be dephosphorylated to free glucose. There are two major subtypes of GSDI: Ia (~80%), caused by mutations in the G6PC gene, and GSD Ib (~20%), caused by mutations in the SLC37A4 gene. The proteins produced from G6PC (catalytic activity) and SLC37A4 (transporter) work together [ 2 ]. GSD Ia involves glycogenolysis and gluconeogenesis, and the clinical manifestations are increased weight, hepatomegaly, failure to thrive, fasting hypoglycaemia, high lactate, hyperuricemia, nephromegaly and hyperlipidaemia [ 3 ]. In addition to the features presented in GSD Ia, GSD Ib also presents with susceptibility to recurrent infections, impaired neutrophil and monocyte function, and inflammatory bowel disease (Crohn’s-like IBD) [ 1 ].
Mutations in the AGL gene cause GSD type III, in which the defective glycogen debranching enzyme blocks glycogenolysis, stopping the conversion of glycogen to glucose-1-phosphate [ 4 ]. At the same time, gluconeogenesis is enhanced to help maintain endogenous glucose production. Hepatomegaly in type III GSD generally improves with age, but affected individuals may develop chronic liver disease (cirrhosis) and liver failure later in life [ 5 ].
GSD IX is caused by the inability of phosphorylase b kinase (PHKA) to break down the glycogen in liver and/or muscle cells. Type IXα glycogenosis is an X-linked disease caused by mutations in the alpha subunit of PHKA . The signs and symptoms typically begin in early childhood, but GSD IX is usually milder than the other types [ 6 ].
The treatment for the aforementioned types of GSD involves nutritional adjustments primarily, with the periodic and frequent administration of large amounts of uncooked cornstarch (UCCS) and restriction of simple carbohydrates [ 7 ] to maintain normoglycaemia and avoid glycogen storage. Usually, higher and frequent doses of UCCS are prescribed for type Ia patients and lower doses for type IX patients. The dose is adjusted according to weight and metabolic demand [ 8 ]. GSD III and IX patients may require a hyperproteic diet with fewer restrictions for simple sugars. Sometimes additional medications may be necessary.
During the last decades, our understanding of the human being has changed. We know now that the eukaryote cells encoded by our genome are not the only component of our body. Symbiont prokaryotic cells inhabiting many cavities of our body provide metabolic functions far beyond the scope of our own physiological capabilities [ 9 ]. These cells play an important role in health and disease states [ 10 ]. The gut microbes are the most studied human associated microbial communities and consists of trillions of microbes and millions of functional genes [ 11 ]. Healthy humans present a remarkable microbial diversity but with similar functions indicating that different microbial communities are associated with a healthy microbiome [ 12 ]. The gut microbiome can be influenced by diet, lifestyle, drugs and genetics of the host [ 13 ], and has been related to several features present in GSD patients including obesity, IBD and liver disease [ 14 ]. This work aimed to investigate possible associations between GSD and the gut microbiota.
Methods
This study was a cross-sectional, observational convenience sampling study, which included 24 GSD patients (Ia = 15, Ib = 5, III = 1, IXα = 3) and 16 healthy controls. All patients were recruited from the outpatient clinics of the Medical Genetics Service at Hospital de Clínicas de Porto Alegre (MGS-HCPA), Brazil from Jan/2016 to May/2017. As inclusion criteria, the subjects (patients and controls) were ≥ 3 years old and not on antibiotics. The GSD patients also were required to: a) have a genetic diagnosis of GSD and b) be on treatment with UCCS. The healthy controls were recruited by invitation as they came to routine appointments at Santa Cecília Basic Health Unit, Porto Alegre, Brazil. All subjects received a kit and printed instructions for stool collection, storage, and transport. They were also provided with printed instructions to record three days of dietary information. Each participant collected their own frozen fecal sample and three-day dietary record and submitted them to an outpatient clinic during their next routine check-up. Upon returning to the clinic, each participant answered a brief questionnaire about personal features including weight and height, eating habits, intestinal habits, medicines of recent and/or continuous usage and lifestyle. The study protocol was approved by the Ethics Committee of Hospital de Clínicas de Porto Alegre (HCPA). All participants and/or legal guardians signed an informed consent.
As a routine, GSD patients seen at the MGS-HCPA who are on UCCS therapy also receive a multivitamin prescription. Despite optimum dietary treatment other drugs could also be prescribed, mainly for type I patients, such as allopurinol, to prevent gout and urate nephropathy; angiotensin converting enzyme inhibitors, to slow-down or prevent further deterioration of renal function; citrate, to preventing or ameliorating urolithiasis and nephrocalcinosis, in addition to correcting lactacidaemia; statins to treat hypercholesterolaemia [ 15 ]; and mainly for Ib patients, G-CSF to treat neutropenia, neutrophil dysfunction and IBD; and the intestinal anti-inflammatory mesalazine (5-amino-salicylic acid), also to treat IBD [ 16 ].
Nutritional assessment, clinical data and statistical analysis
Macro and micronutrients intake by the subjects were estimated from the three-day food records through the Nutribase software (NB16Cloud, CyberSoft, Inc., Phoenix, AZ, USA). The daily nutrient intake of each participant was the sum of the nutrients of each food item. The average of the three-day intake was used for further analysis. Multivitamin consumption and other medications were not included in the nutritional assessment but were considered as variables that potentially were modifying the gut microbial composition, so they were tested by Permutational Multivariate Analysis of Variance. Clinical data, such as IBD and other relevant conditions, were accessed from medical records. BMI-for-age and Z-scores were calculated within the World Health Organization (WHO) AnthroPlus software suite. A qualitative classification for this data followed the WHO criteria [ 17 ].
Statistical analysis among the groups was performed using PASW Statistics for Windows software (Vs18.0, 2009, SPSS Inc., Chicago, USA). Numerical variables were compared using the Mann-Whitney U test. Categorical variables were compared using X 2 , Fisher’s exact test or Continuity Correction, when necessary (with statistical significant determined by the threshold p ≤ 0.05). Statistical analyses with the microbiome feature are described below.
Bacterial DNA extraction, 16S rRNA gene amplifications and sequencing
The bacterial DNA was isolated from 0.3 mg of frozen faecal sample with QIAamp DNA Stool Mini Kit (Qiagen, Valencia, CA, USA) (Qiagen) according to manufacturer instructions and stored at -20°C until use. The NanoVue system (GE Healthcare, Chicago, IL, USAGE Healthcare) was used to assess the quality of extractions for downstream applications. For the sequencing step, the library was prepared following the procedures described by Barboza et al. [ 18 ]. Briefly, region V4 of 16S rRNA gene was amplified with the barcoded bacterial/archaeal primers 515F and 806R [ 19 ] in a reaction containing 2U of Platinum Taq DNA High Fidelity Polymerase (Invitrogen, Carlsbad, CA, USA), 4 μL 10X High Fidelity PCR Buffer, 2 mM MgSO4, 0.2 mM dNTPs, 0.1 μM of both the 806R barcoded primer and the 515F primer, 25μg of Ultrapure BSA (Invitrogen, Carlsbad, CA, USA) and approximately 50 ng of DNA template in a final volume of 25 μL. After an initial denaturation step of 5 min at 95°C, 30 cycles of 94°C for 45 s, 56°C for 45 s and 72°C for 1 minute were performed, followed by a final extension step of 10 min at 72°C. After visualization on agarose gel 1.5%, the PCR products were purified with the Agencourt AMPure XP Reagent (Beckman Coulter, Brea, CA, USA) and the final concentration of the PCR product was quantified with the Qubit Fluorometer kit (Invitrogen, Carlsbad, CA, USA) following the manufacturer's recommendations. Finally, the reactions were combined in equimolar concentrations to create a mixture composed of 16S gene amplified fragments of each sample. This composite sample was used for library preparation with the Ion OneTouch 2 System using the Ion PGM Template OT2 400 Kit (Thermo Fisher Scientific, Waltham, MA, USA). Sequencing was performed with Ion PGM Sequencing 400 on the Ion PGM System using Ion 318 Chip v2.
16S profiling data analysis
The Fastq files exported from the Ion PGM System were analysed with the BMP Operating System (BMPOS) [ 20 ] according to the recommendations of the Brazilian Microbiome Project [ 21 ]. Briefly, an Operational Taxonomic Unit (OTU) table was built using reads truncated at 200 bp and quality filtered with a maximum expected error of 0.5. After removing singletons, the sequences were clustered into OTUs at cutoff of 97% similarity, and chimeras were checked and removed to obtain representative sequences for each microbial phylotype. Taxonomic classification was carried out in QIIME version 1.9.1 [ 22 ] based on the UCLUST method against the SILVA ribosomal RNA gene database version v132 [ 23 ] with a confidence threshold of 80%. Downstream analyses were carried out with dataset rarefied to the minimum library size [ 24 , 25 ] in the R environment [ 26 ] using the phyloseq package [ 27 ] and vegan package [ 28 ]. The online software Microbiome Analyst [ 29 ] was used to further detect microbial biomarkers associated with GSD patients. After Cumulative Sum Scaling (CSS) normalization [ 30 ], the dataset was analysed by the non-parametric factorial Kruskal-Wallis (KW) sum-rank test followed by Linear Discriminant Analysis [ 31 ]. To make sure the biomarkers observed were not only driven by IBD-like patients, we performed one analysis using the full dataset and another analysis excluding all four IBD-like patients and matched controls.
Faecal calprotectin assay and pH measurement
Frozen faecal samples of patients and controls were thawed and aliquoted at room temperature (20°C) to perform the pH measures and calprotectin assay. To determine the faecal pH, the samples were diluted 1:10 (w/v) in distilled water. After homogenization and incubation for 5 min at room temperature, the faecal pH was measured by an electronic pH-meter (K39-1014B, KASVI, PR, Brazil) three minutes after complete electrode immersion.
The faecal calprotectin was quantified from 100 mg of faecal sample with the RIDASCREEN Calprotectin test (R-Biopharm AG) according to the manufacturer’s instructions. Calprotectin is a calcium-/zinc-binding protein, highly stable and resistant to degradation by intestinal contents (pancreatic secretions, proteases, and bacterial degradation). It is mainly produced by neutrophils in inflammation and has been amply confirmed in intestinal inflammatory diseases [ 32 ]. Calprotectin was evaluated to verify gut inflammation across groups and its influence over the number of OTUs. Due to the small sample size of GSD III and IXα, just the subtypes Ia and Ib (groups containing >15% of total sample) were compared. Results for GSD Ia and GSD Ib patients were presented as median (Q1-Q3) and as min-max to GSD III and IXα. To test the correlation among calprotectin and OUT richness, patients who were on mesalazine were excluded from analysis.
Results
The characteristics of the patients and controls are summarized in Table 1 . The nutrient intake varied significantly between groups ( S1 Table ); the largest variation observed was the higher total carbohydrate and calorie intakes in the GSD group due to UCCS usage. The amount of protein consumed (g) and the number of calories derived from proteins did not differ between patients and controls. However, the percentage of total caloric intake from proteins was lower in patients. Patients ingested less fat (g and Kcal/day) and had a lower percentage of fat in the diet. Regarding micronutrients, patients’ diet was poor in calcium and sodium, and in vitamins B3, H, D and E in comparison to the control group’s diet.
10.1371/journal.pone.0214582.t001
Table 1 Sample characterization, analysis of potential confounding variables and their effect on microbial communities.
Variable 1
Patients (n = 24)
Controls (n = 16)
p-value 1
Microbial community difference between patients and controls
Euclidian Metric
Bray-Curtis Metric
r 2
p-value
r 2
p-value
Sex (M/F)
14/10
07/09
0.561
0.02942
0.287
0.02964
0.267
Age (yr)
12 (10–19.75)
12.5 (10–23.25)
0.579
0.02895
0.302
0.02775
0.340
Faecal pH
6.23 (5.42–7.16)
7.41 (7.10–7.98)
0.001
0.05938
0.005
0.08507
0.001
Inflammatory Bowel Disease (yes/no)
04/20
00/16
0.136
0.06746
0.009
0.05152
0.003
Abdominal pain complaint (yes/no)
09/15
01/15
0.032
0.05590
0.010
0.04845
0.009
Nutritional status* (Obese or Overweight/Normal)
18/06
06/09 †
0.044
0.05199
0.004
0.03423
0.121
UCCS intake (g/day)
309.50 (373.7–245.3)
00
0.001
0.03698
0.114
0.05594
0.001
Drugs (yes/no):
-Allopurinol
4/20
0/16
0.136
0.02477
0.436
0.02426
0.517
-Antibiotic usage (last 6 months)
10/14
3/13
0.241
0.03047
0.252
0.03200
0.179
-ACE inhibitor
11/13
0/16
0.001
0.03351
0.203
0.03919
0.054
-Filgrastim (G-CSF)
5/19
0/16
0.071
0.06654
0.002
0.05377
0.008
-Mesalazine
3/21
0/16
0.262
0.03089
0.290
0.03389
0.109
-Multivitamin
22/2
1/15
0.001
0.04034
0.070
0.05545
0.003
-Potassium Citrate
3/21
0/16
0.262
0.02248
0.516
0.02407
0.551
-Proton Pump Inhibitors
2/22
0/16
0.508
0.03068
0.318
0.03087
0.173
-Statins
1/23
0/16
1.000
0.03312
0.286
0.02542
0.486
UCCS: uncooked cornstarch; ACE: Angiotensin-converting-enzyme inhibitor (enalapril maleate); G-CSF: G-colony stimulating factor. Significant (p<0.05) events are highlighted in bold.
1 Numeric variables were reported as medians (Q1-Q3). Due to the not-normal distribution, numeric variables were subjected to the Mann-Whitney test. Qualitative variables were reported as absolute frequency and tested by X 2 , Fisher’s test or Continuity Correction, as appropriate.
† Data for one control was missing. Weight and height were measured when subjects delivered the sample. In this case, a relative drove the sample to the hospital, thus we were unable to do so.
The intakes of macro and micronutrients were similar among all the GSD types, with some kcal variation from carbohydrate intake due the difference in UCCS consumption among groups ( S2 Table ).
Overall 16S rRNA sequencing results, sequence quality control and control for confounding variables
After quality filtering of the 16S rRNA reads, a total of 1,786,582 high-quality sequences longer than 200 bp were obtained. To analyse whether the number of sequences from each sample was representative of the underlying bacterial community, sequence coverage was calculated ( S3 Table ). An average of 44,664 sequences per sample was obtained with average sequence coverage of 0.99 at the 3% dissimilarity level. This sequencing depth was sufficient to obtain excellent representation of the microbial community in these samples.
Results for suspected confounding variables that potentially were modifying the gut microbial composition are presented at Table 1 and S1 Table . The gut microbial communities were not affected by sex, age, nor the nutritional status of the subjects tested. Faecal pH was lower in patients (6.23) than in controls (7.41), and this variable affected the presence/absence and abundance of the gut microbes, with a reduced OTU count in lower pH. Only 18% of controls (n = 3) and 41% of patients (n = 10) used antibiotics within the 6 months prior to data collection. The use of antibiotics within the 6 months prior to sampling did not affect the presence/absence of microbes ( p = 0.252) nor microbial relative abundance ( p = 0.179) in these samples.
Hepatic GSD is associated with an abnormal gut microbial community
The analysis of overall microbial community structure revealed significant differences between patients and controls ( Fig 1 ). According to the PERMANOVA, the microbial community structure between patients and controls differed by the presence and absence of taxa (r 2 = 0.182; p = 0.003) and by their relative abundances (r 2 = 0.166; p = 0.001). The analysis indicated that the relative abundance of taxa contributed 16% of the variation in the microbial community between patients and controls while the presence/absence of specific taxa contributed 18% to that variation.
10.1371/journal.pone.0214582.g001
Fig 1
Principal coordinates analysis (PCoA) based on Bray Curtis distance matrix (A) and Euclidean distance matrix (B) show the separation of gut microbiomes between GSD patients and controls. Each point represents a microbial community from one subject; colours indicate the treatment.
Microbial diversity as measured by richness of OTUs and by the Shannon diversity index also differed significantly (p < 0.01) between patients and controls ( Fig 2 ). On average, control stool samples possessed 184 OTUs while the patients had only 100 OTUs. The average Shannon diversity index was 3.49 and 2.48 in controls and patients, respectively. Together, these beta and alpha diversity analyses indicated that the GSD gut microbiome is characterized by low diversity and distinct microbial structures.
10.1371/journal.pone.0214582.g002
Fig 2
Alpha diversity measurements of microbial communities in the GSD patients and control groups.
Each panel represents one alpha diversity measure: Richness = total number of OTUs observed, Shannon = microbial index of diversity. Boxes span the first to third quartiles; the horizontal line within the boxes represents the median. Whiskers extending vertically from the boxes indicate variability outside the upper and lower quartiles. *** indicates a statistical difference between treatments at cutoff p ≤ 0.001.
Defining the main taxa associated with the gut microbiota of patients and controls
Specific microbial phylotypes present within the gut community might drive the main differences observed in GSD patients. To find those microbes, biomarker screening analysis was performed at different taxonomic levels. A total of 14 phyla were detected within these samples. However, more than half of the community was dominated by only three phyla: Bateroidetes (58% in controls; 47% in patients), Firmicutes (34% in controls; 39% in patients) and Proteobacteria (5.8% in controls; 10% in patients) ( Fig 3 ). All of the other phyla had very low relative abundances. LEfSe analysis identified three microbial phyla as biomarkers with Actinobacteria and Proteobacteria overrepresented in patients while Euryarchaeota was underrepresented. In particular, Proteobacteria presented a very high LDA score (more than 3.9 orders of magnitude), reflecting a marked increase in relative abundance in patients and consistently low abundance in controls. Firmitutes had a marginally-significant difference between patients and controls ( p = 0.043 and LDA score = 4.53 but FDR = 0.07).
10.1371/journal.pone.0214582.g003
Fig 3
The average relative abundance of phyla found in GSD patients and healthy controls.
Phyla followed by an asterisk (*) are different, both in terms of statistics and biological consistency, between patients and controls at p and FDR ≤ 0.05: Euryarchaeota (LDA score = 1.75), Actinobacteria (LDA score = 3.06) and Proteobacteria (LDA score = 3.94). Firmicutes was marginally significantly different with p = 0.064, LDA score = 4.52 and FDR = 0.112.
At the genus level, nineteen microbial biomarkers were different, both in terms of statistics and biological consistency, between patients and controls ( Table 2 ). Those genera were higher in controls. In patients, those genera were in low abundance and in some cases totally absent. The lack of those microbes might be reflected in the alpha and beta diversity results as mentioned previously (Figs 1 and 2 ). Besides, Lactobacillus and Escherichia/Shigella were found to be dominant in patients with a very high LDA score (4.36 and 3.89, respectively), highlighting the biological importance of those microbes in GSD. To remove any biases caused by patients with IBD-like symptoms (n = 4), all IBD-like patients and their respective controls were removed from the dataset and a new biomarker analysis was performed ( Table 2 ). Similar trends as observed within the full dataset were still present in this reduced dataset. However, the Lactobacillus genus, found previously in higher abundance in patients was not observed within the dataset without IBD-like patients. On the other hand, Escherichia/Shigella was still found to be more abundant in patients than in controls (LDA score = 3.85).
10.1371/journal.pone.0214582.t002
Table 2 Microbial biomarkers differentiating patients with hepatic glycogenosis diseases and healthy controls.
Microbial genus
Patients
Controls
p -values
FDR
LDA score
Relative abundance (%)
(log 10)
Full dataset
n = 24
n = 16
Lactobacillus
11.31
0.04
0.009
0.025
4.36
Escherichia/Shigella
6.70
0.96
0.003
0.013
3.89
Alistipes
2.77
9.12
0.005
0.018
-3.22
Subdoligranulum
1.59
1.00
0.012
0.029
2.42
Lachnospiraceae NK4A136 group
1.44
0.89
0.003
0.013
2.48
Faecalibacterium
1.00
3.52
0.016
0.036
-2.98
Ruminococcaceae UCG 002
0.98
3.09
0.001
0.007
-2.79
Bifidobacterium
0.78
0.19
0.004
0.018
3.1
Ruminococcus gnavus group
0.70
0.14
0.007
0.022
3.03
Phascolarctobacterium
0.53
1.31
0.015
0.035
-2.56
Blautia
0.26
0.53
0.002
0.012
-1.55
Odoribacter
0.25
0.53
0.011
0.028
-1.87
Barnesiella
0.22
0.98
0.009
0.025
-2.46
Roseburia
0.18
1.19
0.002
0.011
-2.78
Christensenellaceae R 7 group
0.14
0.80
0.000
0.002
-2.22
Ruminococcaceae UCG 003
0.10
0.60
0.000
0.003
-2.27
Lachnospiraceae UCG 008
0.04
0.26
0.004
0.018
-1.78
Ruminococcaceae UCG 005
0.03
0.25
0.000
0.002
-1.9
Eubacterium hallii group
0.02
0.08
0.000
0.002
-1.39
Anaerostipes
0.01
0.11
0.001
0.009
-1.55
Coprococcus 1
0.01
0.03
0.000
0.005
-0.95
Family XIII AD3011 group
0.01
0.05
0.000
0.002
-1.21
Family XIII UCG 001
0.00
0.03
0.001
0.007
-1.13
Methanobrevibacter
0.00
0.17
0.001
0.007
-1.78
Ruminococcaceae NK4A214 group
0.00
0.08
0.001
0.007
-1.5
Dataset without IBD-like patients *
n = 20
n = 14
Escherichia/Shigella
6.47
0.92
0.003
0.027
3.85
Alistipes
2.97
9.76
0.008
0.039
-3.28
Ruminococcaceae UCG 002
1.12
3.07
0.004
0.028
-1.38
Bifidobacterium
0.81
0.08
0.003
0.027
3.2
Phascolarctobacterium
0.22
1.38
0.004
0.028
-2.74
Christensenellaceae R 7 group
0.17
0.76
0.001
0.016
-2.16
Blautia
0.14
0.39
0.001
0.017
-2.08
Ruminococcaceae UCG 003
0.11
0.61
0.001
0.016
-2.3
Roseburia
0.10
1.15
0.004
0.028
-2.83
Lachnospiraceae UCG 008
0.04
0.19
0.011
0.047
-1.57
Ruminococcaceae UCG 005
0.03
0.20
0.001
0.016
-1.76
Eubacterium hallii group
0.02
0.07
0.000
0.016
-1.32
Anaerostipes
0.01
0.07
0.010
0.047
-1.28
Coprococcus 1
0.01
0.02
0.008
0.039
-0.77
Family XIII AD3011 group
0.01
0.04
0.001
0.017
-1.14
Family XIII UCG 001
0.00
0.03
0.001
0.016
-1.15
Methanobrevibacter
0.00
0.17
0.003
0.027
-1.81
Ruminococcaceae NK4A214 group
0.00
0.08
0.003
0.027
-1.53
* Four IBD-like (Inflammatory Bowel Disease) patients and matched controls were excluded from the dataset to make sure the biomarkers observed were not only driven by these patients.
Correlations between the gut microbiota, diet, faecal pH and gut inflammation
Spearman correlations were calculated between the microbiome, diet, faecal pH and calprotectin ( Fig 4 ).
10.1371/journal.pone.0214582.g004
Fig 4
Correlations between the microbiota and diet, faecal pH, and gut inflammation.
The faecal pH values varied between patients and controls ( Table 1 ), and this was important for shaping their respective differences in gut microbiomes. Differences were determined with the Euclidian distance matrix (for presence/absence of taxa) and the Bray Curtis distance matrix (for relative microbial abundance). Faecal pH was correlated with the total number of microbial OTUs such that higher faecal pH seemed to support more OTUs.
Microbial richness correlated negatively with total carbohydrate but positively with simple carbohydrates (sugar). Calprotectin seemed to have no influence over the microbiome in terms of the number of OTUs ( Fig 4 ). In addition, there was no correlation between this inflammatory marker and gut microbial richness.
Discussion
This is the first study about the fecal microbiota of GSD patients. In hepatic GSD, high and periodic amounts of UCCS plus dietetic restriction of fast-digestion carbohydrates are the main way to treat the genetic impairment in the glycogenolytic pathway. Our data suggest that the overload of UCCS can lead to low fecal pH by favouring some bacterial genera capable of utilizing complex carbohydrates in detriment of others. The low fecal pH, in turn, also acts as an environmental selection factor to the bacteria in the lumen. Dysbiosis has been associated with IBD and obesity. IBD includes inflammatory bowel diseases of unknown aetiology and has two main forms: ulcerative colitis and Crohn’s disease (CD). CD is a chronic disease that can affect any region in the digestive tract but is more likely to involve the small and large intestines and the perianal region [ 33 ]. Enteropathy is related to type I patients, and despite GSD Ib patients are classically described as prone to IBD-Crohn’s-like due the impaired neutrophil activity, this does not explain why patients with GSD Ia also displayed serologic markers altered for IBD, even if asymptomatic [ 34 ]. Its not clear if UCCS is the cause of obesity in GSD patients [ 35 ], but the microbiome might be associated with it. Here we discuss why the changes in microbiota could be considered as a factor of influence in the phenotype of these patients and why the UCCS usage, even though not exclusively, is an important factor that contribute to that.
Since the introduction of UCCS treatment for GSD, the focus changed from mortality to morbidity and control of long-term complications [ 36 ], such as metabolic syndrome and related symptoms [ 37 , 38 ]. GSD type I patients are prone to obesity, and it is suspected that UCCS contributes to the aforementioned features [ 35 , 39 ]. GSD I patients also are subject of heavier doses of UCCS and more restrict diet in comparison with types III and IX [ 35 ]. Regarding antibiotics, although its usage clearly drives changes in the gut microbial community, subjects who were treated with antibiotics within 6 months prior to data collection, but not during the study itself, were not affected by the previous antibiotic usage.
We found that the phyla Actinobacteria and Proteobacteria were overrepresented in patients while the Euryarchaeota was underrepresented. The microbiome of GSD patients present low diversity and was highly dominated by Escherichia/Shigella .
One possible driver of the differences in gut microbiomes between patients and controls is UCCS overload, which creates an acidic environment [ 34 , 40 ]. In the human body, acids are generated by regular metabolic activities and through the daily intake of food [ 41 ]. Fecal pH was lower in patients than controls and stool acidification might lead to an alteration in the relative abundances of fermenting bacteria, decreasing the conversion of unabsorbable starches to short chain fatty acids (SCFAs) [ 34 ].
SCFAs, including butyrate, are compounds made by bacteria in the gut that affect several physiologic functions and serve anti-inflammatory roles [ 42 ]. Fecal pH was associated with beta diversity and bacterial families belonging to the Clostridia class, an important producer of butyrate in the gut. Several genera of SCFA-producing bacteria— Coprococcus , Blautia , Anaerostipes , Odoribacter and Faecalibacterium —were decreased in patients. Those genera were also identified in paediatric patients with Crohn’s Disease [ 43 ]. Besides, Coprococcus and Faecalibacterium were found to have significantly low abundance in patients with nonalcoholic fatty liver disease, independently of body mass index and insulin resistance [ 43 ].
The bacterial species residing within the mucous layer of the colon may influence whether host cellular homeostasis is maintained or inflammatory mechanisms are triggered. A mutualistic relationship between the colonic microbiota, their metabolic products and the host immune system is likely involved [ 44 ]. The phylum Proteobacteria was more abundant in patients than in controls while the phylum Euryarchaeota was less abundant. Proteobacteria is a gram-negative phylum with an outer membrane mainly composed of lipopolysaccharides (LPS), which are known to sustain systemic levels of low-grade inflammation [ 45 ]. Higher levels of Proteobacteria can be considered a strong marker of dysbiosis [ 46 ]. This phylum is prevalent in patients with liver cirrhosis [ 47 ]. Several serological markers for IBD were altered in GSD-Ia patients [ 34 ], and GSD Ib patients are prone to IBD CD-like. Despite the fact that calprotectin seemed not to influence the number of OTUs gut inflammation (calprotectin >50μg/g) was verified in several patients. GSD type Ib patients have shown a concentration of calprotectin ≤50μg/g, except for one patient, who had an active IBD diagnosed in the same week. This might be due to a remission state and the use of anti-inflammatory mesalazine by these patients.
In general, dysbiosis can be categorized as a) loss of beneficial organisms, b) excessive growth of potentially harmful organisms and c) loss of overall microbial diversity. These three categories often occur at the same time [ 48 ]. Dysbiosis has been implicated in a wide range of diseases, including IBD, liver disease and obesity, that are secondary manifestations in GSD patients [ 49 ]. The reason for dysbiosis remains unclear, but the overload of UCCS contributes to those characteristics. The food intake records showed a difference in the intake of calories, mainly due to the administration of UCCS in patients, as well as a difference in microbial signature that is known to be related to obesity. It is not known whether these microbiome changes are a cause or a consequence of the pathophysiologies. However, correcting the dysbiosis can improve health in some patients [ 50 – 52 ]. Dysbiosis can also provide biomarkers for disease detection and management [ 53 ].
Conclusion
In this study, we reported significant alterations in the intestinal environments of GSD patients versus healthy controls. Microbiota can be affected by abiotic and biotic factors, namely pH and inflammation, and the differences in these factors between patients and controls might be linked to both genetic disease and UCCS consumption. Several bacterial taxa were different in GSD patients than in controls, and those groups are consistent with the secondary phenotypic manifestations of GSD. The microbiome patterns of these patients may reinforce immune-metabolic pathways that already are altered by genetic impairment, and may also be a factor in the differential individual response to treatment. Patients may gain health and quality of life from the restoration of gut microbial diversity that has been diminished by high UCCS intake. Future research therefore should investigate ways to manipulate the gut microbiome and clarify the possible effects of doing so.
Supporting information
S1 Table
Differences in nutrient mean daily intake between healthy controls and GSD patients and their effect on microbial communities.
*Absolute number means that the estimative of ingestion was constant for all the subjects of the group. 1 Mann-Whitney U test. 2 Bray-Curtis. Significant (p<0.05) events are highlighted in bold.
(PDF)
S2 Table
Summary of the finding of the GSD patients (n = 24).
OTU: operational taxonomic unit; UCCS: uncooked cornstarch; ACE: Angiotensin-converting-enzyme inhibitor (enalapril maleate); G-CSF: G-colony stimulating factor. 1 Numeric variables were reported as Median (Q1-Q3) for GSD Ia and Ib and as Min- Max for GSD III and IXα. Qualitative variables were reported as absolute numbers. P-value was accessed to differences between the groups Ia and Ib. † Calprotectin and number of OTUs for patients on and without mesalazine were reported as Min-Max.
(PDF)
S3 Table
Overall description of the 16S rRNA sequencing results among subjects.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Highly pathogenic H5N1 avian influenza viruses (HPAIVs), first isolated in Guangdong, China in 1996 [1] , have spread from southern China across Southeast, East and Central Asia, to the Middle East, Europe and Africa [2] – [5] . Millions of domestic and wild birds have either died or been culled because of the outbreaks caused by these H5N1 viruses. H5N1 HPAIVs have also proven to be highly lethal in humans: 262 fatalities from 440 reported cases, representing a mortality rate of about 60%, as of August 31, 2009 [6] . Recently emerging family clusters in Indonesia and Pakistan have increased concern about human-to-human transmission of H5N1 viruses [7] , [8] , an essential step for epidemics and pandemics in human populations, which has not yet occurred. These human cases of H5N1 infection have raised a great concern for possible emergence of another novel influenza virus from reassortments between H5N1 HPAIVs and the 2009 H1N1 pandemic virus, which probably emerged in Mexico and spread rapidly through human-to-human transmissions and is causing the ongoing pandemic [9] . Such a reassortment between H5N1 and H1N1 viruses could increase the mortality and severity of the current influenza pandemic.
While H5N1 HPAIVs are geographically widespread, some countries have seen greater incidence in poultry and humans than others. H5N1 influenza in Vietnam was first identified in poultry in 2001 and in humans in 2004 [10] , and has since caused 111 cases and 56 deaths in humans, as of August 31, 2009. During the 2003/2004 H5N1 outbreaks, nearly the entire poultry population of Vietnam was culled.
Phylogeographic studies reveal the potential evolutionary and migration history of H5N1 viruses. Reassortment events among gene segments derived from A/turkey/England/50–92/1991 (H5N1)-like HA gene and low pathogenic avian influenza viruses probably occurred in southern China in the early 1990s [11] – [14] . Since that time, H5 HA genes underwent dramatic changes: at least 10 distinct major clades and even more sub-clades have emerged during the past decade [15] . Frequent reassortments are occurring not only between H5N1 and other subtypes of AIVs but also within H5N1 AIVs [10] , [12] , [14] . Analyses of the evolutionary history of H5N1 viruses in Vietnam show close links with viruses in southern China [16] and suggest that introductions took place along the shared border between Vietnam and Yunnan and Guangxi Provinces [14] , [17] .
Since 2001, at least six clades/subclades of H5 HPAIV HA (Clade 0, 1, 2.3.2, 2.3.4, 3, and 5) and nine reassortants of H5N1 HPAIV emerged in Vietnam [10] , [18] , [19] . These H5N1 viruses formed two phylogenetic clusters across both northern and southern Vietnam after they were introduced into northern Vietnam and spread to southern Vietnam [10] , [16] . Prior analysis of the geographic spread of evolving influenza viruses across Vietnam has been more descriptive than statistical. In this study, we sought to determine how location and spread of H5N1 HPAIVs through geographic space interacted with viral evolution in Vietnam between 2003 and 2007, using spatial statistics to map potential correlative changes between geographic and genetic distances.
The geography of Vietnam, as well as the genetic history of H5N1 avian influenza in the country, makes it particularly appropriate for a spatial analysis of H5N1 evolution. Vietnam is long (north-south) and narrow (east-west). Human population densities are highest around Hanoi and the Red River delta in the north and Ho Chi Minh City and the Mekong River delta in the south. Poultry densities are also higher in these areas. There also exist distinct regional differences in economic and agricultural patterns between the north and south. These patterns allow us to easily characterize regional variation between the north and the south. H5N1 viruses isolated in Vietnam that belong to Clade 1 genotype VN3 (Genotype Z) are from a single introduction [16] , [17] , show extreme phylogenetic clustering [14] , [20] , and have all eight genes descendant from one progenitor virus, A/duck/Hong Kong/821/2002 (H5N1) (HK821-like virus) [10] . The possibility of genes isolated at the same geographic point but derived from different precursor viruses, leading to spurious conclusions about interactions between geographic and genetic distance, are thus diminished.
Our results show that genetic evolution of VN3 (HK821-like) H5N1 viruses in Vietnamese domestic poultry is highly correlated with the location of those viruses in geographic space. This correlation varies by scale, time and gene, though a classic isolation by distance pattern is observed. This study is the first to characterize the geographic structure of influenza viral evolution at the sub-national scale in Vietnam, and can shed light on how H5N1 HPAIVs evolve in certain geographic settings. It also lends strength to the supposition that domestic bird populations are primary drivers of influenza viral evolution, although more studies need to be carried out to assess the potential roles of wild birds and domestic animals (e.g. pigs) in facilitating influenza genetic diversity.
Results
Bi-Modal Influenza Genetic-Geographic Distance Distribution
Visual data diagnostics revealed that the frequency distribution of geographic distance between case pairs was bi-modal with a cluster of relatively short and another of relatively long distances ( Figure 1 ). This bi-modal pattern is linked to the distribution of H5N1 incidences in northern or southern Vietnam, with few isolates obtained in the central regions of the country. The distribution of genetic distances is also somewhat bi-modal, with case-pairs at both ends of the geographic scale exhibiting small and large genetic distances.
10.1371/journal.pone.0008631.g001 Figure 1
Genetic versus geographic distance of HK821-like HPAIVs in Vietnam.
The least squares line is plotted in grey.
Individual gene scatterplots stratified by year revealed a marked difference in patterns by year and gene ( Figure 2 ). Specifically, the bi-modal pattern of distances emerges for only specific year and gene combinations. In 2003 across all eight genes, the case pairs are all clustered at the low end of the geographic scale. This reflects a characteristic of our dataset, in which all 2003 cases occurred in northern Vietnam, geographically close to one another. In 2004 there is greater spread of points across the geographic scale, but relatively little across the genetic scale. By 2005, the bi-modal nature of influenza occurrence in the northern and southern regions of Vietnam is established and case pair distances cluster at the high and low ends of the geographic scale. Genetic distance between case pairs is higher in 2005 and 2007 than in the first two years of the dataset, an indication that these influenza viruses have evolved as local epidemic strains in domestic poultry after sweeping over Vietnam during the 2003–2004 outbreaks.
10.1371/journal.pone.0008631.g002 Figure 2
Genetic versus geographic distance of HK821-like HPAIVs, stratified by year and gene segment.
Within- Versus Between-Region Genetic Variation
Three-way ANOVA regressions ( Table 1 ) indicate strongly significant main effects, two-way and three-way interactions within the dataset. The sampling unit for this ANOVA was pairs of observations. The response was genetic distance, and there were three binary predictors: gene, year, and region. These predictors were coded true when the observations pair had the same value, and coded false when the pair had different values. Though the strict ANOVA assumption of independence of observations is not met because distance matrices are based upon paired observations, this analysis provides an initial look at the trends present in the data and from which we expanded into more appropriate Mantel regressions that account explicitly for autocorrelation. The two-way interaction between region and year, controlling for gene, indicates that relative genetic distance between regions changes as you look across years ( Figure 3 ). In 2003, when all cases were detected solely in northern Vietnam, there are no between-region case pairs, and the genetic distance in within-region case pairs is quite low. In 2004 and 2005, genetic distances among between-region cases are higher than within-region cases. By 2007, however, genetic distances are similar for the within- and between-region categories. The significant three-way interaction ( Table 1 ) indicates that genetic variation in Vietnamese HK821-like cases is systematically different between regions, years and gene segments, and that these three factors interact in ways that affect genetic distances between virus case pairs. While the true significance of the relationships tested with the three-way ANOVA is masked by the non-independence of observations, the ANOVA results and the relationships plotted in Figure 3 led us to believe that interesting correlations existed between genetic and geographic distances among H5N1 viruses in Vietnam. Mantel test results, reported below, overcome the non-independence of matrix observations, and allowed us to accurately assess the relationship between geography and molecular evolution.
10.1371/journal.pone.0008631.g003 Figure 3
Boxplots of genetic versus geographic distance for the within- and between-region pairs of H5N1 HPAIVs in Vietnam.
The solid black circle is the median genetic distance in each grouping. Hollow circles represent outliers.
10.1371/journal.pone.0008631.t001 Table 1
Three-way ANOVA of genetic distance between case pairs of H5N1 highly pathogenic avian influenza viruses by region, segment (Gene), and year.
Source
df
SumSquare
MeanSquare
F
p-value
Region
1
0.07
0.07
2099.99
0.0000
Gene
7
0.06
0.01
293.44
0.0000
Year
1
0.01
0.01
471.87
0.0000
Region:Gene
7
0.00
0.00
14.23
0.0000
Region:Year
1
0.00
0.00
115.13
0.0000
Gene:Year
7
0.01
0.00
30.77
0.0000
Region:Gene:Year
7
0.00
0.00
11.44
0.0000
Residuals
16840
0.52
0.00
Significant Correlation between Genetic and Geographic Distances of Gene Segments
The results of the Mantel tests for correlation among matrices indicated significant, positive correlations between geographic and genetic distance for all eight influenza gene segments ( Table 2 ). The PB2 and NS genes had the highest correlations.
10.1371/journal.pone.0008631.t002 Table 2
Mantel tests measuring correlation of geographic and genetic distances for H5N1 highly pathogenic avian influenza viruses from 2003 to 2007.
Gene
Mantel r
p-value
CI 2.5%
CI 97.5%
PB2
0.2310
0.001
0.2011
0.2678
PB1
0.1758
0.001
0.1509
0.2062
PA
0.2101
0.001
0.1800
0.2418
HA
0.1957
0.001
0.1698
0.2284
NP
0.2082
0.001
0.1779
0.2416
NA
0.2167
0.001
0.1865
0.2525
MP1
0.1951
0.001
0.1658
0.2300
NS1
0.2864
0.001
0.2565
0.3279
Correspondingly to the ANOVA results reported above, the multiple regressions using Mantel tests (MRM) indicated significant effects of geographic distance between viruses on genetic distance, while controlling for the effect of temporal distances, for all eight genes ( Table 3 ). Simultaneously, year effects were shown for all genes while controlling for the effect of geographic distance. The R 2 in the MRM analysis are also all statistically significant, though much higher for some gene segments (HA) than others (NA).
10.1371/journal.pone.0008631.t003 Table 3
MRM results of genetic distance on spatial lag and temporal lag for the VN3 subset.
Gene
Coefficient
GenDist
p-value
R2
p-value
PB2
(Intercept)
6.15E-03
1
0.347
0.001
GeogDist
1.77E-06
0.001
Year
2.82E-03
0.001
PB1
(Intercept)
7.38E-03
1
0.351
0.001
GeogDist
1.19E-06
0.001
Year
3.05E-03
0.001
PA
(Intercept)
6.18E-03
1
0.336
0.001
GeogDist
1.29E-06
0.001
Year
2.33E-03
0.001
HA
(Intercept)
9.78E-03
1
0.419
0.001
GeogDist
2.05E-06
0.001
Year
5.08E-03
0.001
NP
(Intercept)
5.95E-03
1
0.216
0.001
GeogDist
1.29E-06
0.001
Year
1.65E-03
0.001
NA
(Intercept)
1.08E-02
1
0.136
0.001
GeogDist
2.53E-06
0.001
Year
2.04E-03
0.001
MP1
(Intercept)
5.98E-03
1
0.191
0.001
GeogDist
1.60E-06
0.001
Year
2.06E-03
0.001
NS1
(Intercept)
6.48E-03
1
0.365
0.001
GeogDist
2.66E-06
0.001
Year
3.09E-03
0.001
GenDist indicates the genetic distance variable, GeogDist indicates the geographic distance variable, Year indicates the temporal distance variable.
Although the Mantel and MRM tests show that there is a significant relationship between genetic and geographic distance, the Mantel correlograms indicate that this relationship is not the same across all geographic distances ( Figure 4 ). While the precise patterns of the correlograms vary by gene, the overall pattern is one of less genetic distance among viruses at scales of zero to approximately 1,100 km. At distances greater than 1,100 km, all genes exhibit strong and significant larger than expected genetic distances. Thus, compared to the null hypothesis of no relationship between geographic and genetic distance, measured genetic distances between case pairs are either somewhat less than expected at small spatial lags or significantly greater than expected at large spatial lags.
10.1371/journal.pone.0008631.g004 Figure 4
Mantel spatial correlograms, stratified by influenza gene segment.
Correlograms show the relationship between geographic distance (x-axis) and the Mantel r correlation score (y-axis) of HK821-like HPAIVs. Under the null hypothesis of no relationship between geographic location and genetic similarity, all points would be on the zero line. Points above the zero line indicate lower genetic distance between case pairs. Points below the zero line indicate greater genetic distance between case pairs. Solid symbols are statistically significant, hollow symbols are not. The sharp rise to the furthest point in the correlograms is an artifact of edge effects caused by the spatial structure of the data, and does not indicate genetic similarity at the highest geographic distances between viruses.
Discussion
We characterized the geographic and temporal structure of molecular evolution of highly pathogenic H5N1 HK821-like avian influenza viruses in Vietnam. Our conclusions illustrate the relationships between space, time and genetics among Vietnamese H5N1 viruses in domestic poultry which extends beyond previous findings about the evolution of these H5N1 viruses in Vietnam [1] . A positive, statistically significant relationship between geographic, temporal and genetic distance exists for all eight influenza genes and across all four years of analysis.
H5N1 avian influenza incidence in Vietnam is bi-modal, as seen in the distribution of case pairs ( Figure 1 ). The clustering of avian influenza in the north, surrounding Hanoi, and the south, surrounding Ho Chi Minh City, was observed in the dataset and has been indicated in other studies [21] . This bi-modality varies greatly by year and gene ( Figure 2 ), however, revealing that some genes evolved and moved across the landscape at higher rates than others.
Previous studies have shown that H5N1 viruses appeared in northern Vietnam and spread to southern Vietnam, and that regional genetic mutation took place [10] . We chose to investigate these regional differences further by explicitly dividing the dataset into cases that took place within-region and those that spanned the distance between northern and southern Vietnam. The statistical ( Table 1 ) and graphical results ( Figure 3 ) indicate that year of incidence has a strong effect on the importance of region of incidence and gene in correlating with genetic change. In support of these findings, the results from the MRM analyses shows that the temporality on genetic evolution is important; they also indicate that, when controlling for the strong influence of time of incidence, the geographic distance between cases is a strong predictor of the genetic distance between cases.
Since very little is known about H5N1 in central Vietnam, and since our dataset consists primarily of isolates from the north and south, we cannot conclude whether the viruses were transmitted by passing over central Vietnam or rather moved gradually southward across the length of the country. The Mantel correlograms ( Figure 4 ) demonstrate, however, that viral genetic mutation did not occur uniformly across geographic space. At distances of less than 1,100 km between viruses (approximately the distance between Hanoi and Ho Chi Minh City), the genetic sequences of viruses tended to be more similar, though the statistical significance of this similarity is affected by low numbers of observations at these middle distances. Case pairs that were greater than 1,100 km apart, however, exhibited high degrees of genetic dissimilarity. This dissimilarity was statistically significant for all eight genes. These results suggest that high levels of genetic dissimilarity are observed at large spatial scales (i.e. between north and south Vietnam), and that H5N1 influenza in Vietnam between 2003 and 2007 did not evolve gradually as it spread south, but rather that significant isolation by distance occurred. Viruses with similar ancestral lineages, e.g. the HK821-like viruses from China, evolve at different rates and in different ways in domestic poultry in northern compared to southern Vietnam, producing viruses that are genetically distant both within- and between-regions by 2007.
Our results clarify and improve understanding of how H5N1 HK821-like avian influenza evolves in space and time in poultry in Vietnam. We were able to observe the sequential establishment of closely related genotypes at two distinct population centers, and follow their differentiation due to relative isolation of the viral hubs over a period of several years. We found several patterns that suggest one general model of evolution in this viral system: 1) within regions, viral mixing in poultry moves toward heterogeneity and the emergence of local types; 2) differentiation was centered around regional viral hubs located at centers of human and bird population density; and 3) evolution occurs because of relative isolation of the hubs, most likely fed by the abundant supply of domesticated poultry (and people) at the hubs. The analysis thus suggests that at the scale of neighboring city hubs and the intervening hinterland, evolution of H5N1 follows the pattern described by classical theory of genetic differentiation due to isolation by distance [22] , wherein gradual differential evolution of previously similar populations is driven by geographic distance and time. While the interactions between time, space and genetic evolution could potentially be circumvented by repeated introductions of H5N1 strains of different lineages into Vietnam, we would still expect to see the northern region around Hanoi and the southern region around Ho Chi Minh City acting as important sites of genetic differentiation rather than the central region of the country. Further investigation of the local-level ecosystems of northern versus southern Vietnam will shed light onto what human and environmental factors are driving the place-specific evolution of H5N1 avian influenza viruses.
Materials and Methods
Viral Genetic and Geographic Dataset
The dataset consists of 125 H5N1 HPAIVs isolated from domestic poultry in Vietnam between 2003 and 2007, each of which has the full-length or nearly full-length genomic sequences for all 8 segments ( Table S1 ) [10] . All eight genes in the viruses belong to the Clade1/VN3 (HK821-like) line. In the context of this work, an influenza gene is referred to a gene segment. These viruses have a wide geographic distribution and were isolated in 28 provinces primarily across northern and southern Vietnam, with a few isolates from Vietnam's central provinces ( Figure 5 ). Each virus was assigned a unique identification number, allowing us to link geographic location, genetic sequence and temporal data in later analyses, and the dataset was sorted in ascending order by this unique ID. All matrices subsequently generated have identical ordering.
10.1371/journal.pone.0008631.g005 Figure 5
Geographic distribution of H5N1 highly pathogenic avian influenza viruses (HPAIVs) used in this study.
Darkened provinces indicate locations of virus isolation.
Genetic, Geographic, and Temporal Distance Matrices Creation
In order to generate the genetic distance matrices, we constructed a maximum likelihood phylogenetic tree using nucleotide sequences as described before [10] . Then the patristic distance, which is the sum of branch length on a path between a pair of taxa in the phylogenetic tree, was extracted using software PATRISTIC [23] . Genetic distances were arranged into matrices for each of the eight influenza viral gene segments.
The database included the province for each virus isolate. The ID number and province listing for each of the 125 viruses was imported into ArcGIS geographic information systems (GIS) software and linked to a map of Vietnam's 64 provinces. Each virus was then assigned the latitude and longitude coordinates of the centroid of the province where the virus was isolated. While a more precise virus isolation location is preferable, only the province of incidence was available. While assigning viruses to the province centroid is somewhat arbitrary, the relatively small area of Vietnamese provinces makes this decision less problematic but does prevent within-province analysis. The list of viruses with their attached geographic location was then used to generate a geographic distance matrix. We calculated the ground distance in kilometers between viruses using the great circle distance measure, which takes the curvature of the earth's surface into account. Five geographic distance matrices were generated: 2003, 2004, 2005, 2007, and one for the entire study period.
Using the year of isolation and the unique ID numbers, a temporal distance matrix was created for all 125 viruses.
Statistical Analysis
All matrices were analyzed with R2.7.1 [24] . Using the eight genetic and single geographic distance matrices, we generated distributional scatterplots and fit a least squares line. Potential differences in genetic versus geographic distance relationships according to gene segment led us to stratify and plot the data, first by gene then further by genes in individual years.
The clustering of cases geographically raised the question of whether differences in genetic distance between viruses exist for between-region case pairs versus within-region case pairs. To explore this possibility, we used a distance of 800 kilometers (approximately half the length of the country) to assign case pairs as either taking place within the same region (e.g., both in northern Vietnam) or between regions (i.e., one in the north, the other in the south). A three-way analysis of variance (ANOVA) measured genetic distance as a function of gene, year and within- versus between-region designation, testing whether underlying genetic distributions differ regionally. ANOVA can, in some cases, be a problematic test for non-independent matrix observations, but the inclusion of the region variable helps to overcome spatial dependence in the dataset. The statistical significance of the relationships we tested using the ANOVA, however, are rendered suspect by the dependence that exists between observations in a distance matrix. Boxplots complement the ANOVA results by showing the nature of the interaction between geographic lag (within- vs. between-region) and year in their explanation of genetic distance among viruses.
As is frequently the case with spatial data, there was dependence in the geographic distance matrix due to clustering of the sample locations. Additionally, matrix observations are non-independent, based upon paired datapoints. The findings from the ANOVA and boxplot results suggested that an interesting relationship between the molecular evolution of viruses and geographic space existed, so we implemented Mantel testing. Mantel tests are used to test for correlation between distance matrices when the underlying probability distribution of the test statistic is unknown and when dependence is present. Mantel tests overcome the lack of observation independence by randomly shuffling the values in one of the matrices multiple times and calculating correlations between the shuffled and original matrices. The probability distribution of the test statistic (the Mantel r) is generated by this random permutation process and used as a basis for assigning a probabilistic interpretation of the true correlation statistic between the observed response and predictor matrices. Mantel tests were conducted for each of the influenza virus' eight genes (aggregated by year), comparing genetic and geographic distance matrices.
We built multiple regression models using Mantel tests (MRM) to simultaneously but separately test the effect of time and space on genetic distance. MRM allows for the analysis of two or more matrices [25] , where the response matrix (genetic distance) is regressed on multiple explanatory matrices (geographic and temporal distances), while also controlling for the effect of those explanatory matrices. In other words, MRM enabled us to determine the statistical significance and relative importance of each explanatory variable (space and time) individually while also acknowledging the effect of both.
Mantel tests indicate whether genetic and geographic distances are related, but they tell us little about the form of this relationship. In order to explore the question of how viral evolution occurs across space we generated Mantel correlograms (also known as spatial Mantel correlograms) stratified by gene. In a Mantel correlogram, the geographic distance between observations is divided into lags, and Mantel statistics (including significance) are calculated for case pairs who fall within each lag. Significance tests at each spatial lag are dependent on the sample size of points that fall within each threshold, so for distances where there were few observations (i.e. because of the bias in our dataset for sampling points in the north and south) the findings were statistically insignificant. Correlograms thus display the degree of likeness or difference among viruses at specific geographic distances. Points above the zero line exhibit positive autocorrelation, those below have negative autocorrelation. The Mantel correlograms allowed us to assess whether the degree of genetic dissimilarity among viruses corresponded to the scale of geographic distance between the viruses.
Supporting Information
Table S1
Summary of the H5N1 AIVs used in this study.
(0.02 MB PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The growth of tumor depends on its surrounding vascular supply, which is commonly stimulated by the overexpression of tumor-secreted pro-angiogenic factors including VEGF [ 1 ]. Given the importance of the pro-angiogenic pathway downstream of VEGF, inhibiting the VEGF signaling axis has proven an effective therapy for patients with solid tumors and neovascular age-related macular degeneration; drugs such as bevacizumab and aflibercept which sequester circulating VEGF have shown efficacy in some of these clinical situations [ 1 , 2 ]. Research has also found that human and animals can produce various anti-angiogenic molecules. An example of a potent endogenous anti-angiogenic protein is thrombospondin-1 (TSP-1) [ 3 ]. TSP-1 was the first protein identified as a naturally occurring inhibitor of angiogenesis. It is a large matricellular protein that interacts with various ligands and receptors, including components of the extracellular matrix, growth factors, cell surface receptors and cytokines [ 4 ]. One of the major pathways that TSP-1 negatively regulates to inhibit angiogenesis is the VEGF-VEGFR2 axis. It is reported that secreted TSP-1 binds to its high affinity receptor CD47 and disrupts the association of VEGFR2 with CD47, thereby downregulating the pro-angiogenic signals downstream of VEGF; another mechanism proposed to explain the inhibitory effect of TSP-1 on VEGF-mediated angiogenesis involves the TSP-1 receptor CD36 and endothelial cell apoptosis pathways [ 5 ]. Although the detailed mechanism of action of TSP-1 as an anti-angiogenic protein is not fully understood, the potential of TSP-1 and its analogs as therapeutics against cancer has already been demonstrated by several preclinical and clinical studies [ 6 – 9 ].
The expression of TSP-1 in tumors is often found dysregulated. In some tumors with negative TSP-1 expression, tumor vascularity is significantly higher and this is associated with worse prognosis than patients with TSP-1 positive tumors [ 10 , 11 ]. Because of its strong anti-angiogenic effect, the role of TSP-1 in ischemic vascular diseases has also been investigated. Interestingly, in the plasma and tissue samples collected from patients with peripheral arterial disease (PAD), TSP-1 is highly upregulated [ 12 , 13 ]. Hypoxia is also reported to increase TSP-1 synthesis in non-tumor conditions in various cell types including endothelial cells (ECs), fibroblasts, renal tubular epithelial cells and vascular smooth muscle cells [ 14 – 17 ]. This effect may be parallel to the induction of VEGF in hypoxic conditions, suggesting a potential negative feedback loop that limits angiogenesis in certain conditions.
Besides the direct intervention of TSP-1/VEGFR/CD47 interactions on the cell surface, another potential therapeutic strategy to harvest the anti-angiogenic potential of TSP-1 that is underexplored is the modulation of its intracellular synthesis [ 5 ]. In addition to the transcriptional regulation by promoters and repressors such as HIF-2α and Myc, TSP-1 expression is also tightly regulated by several microRNAs including miR-18a [ 14 , 18 – 20 ]. The HIF-let7-AGO1 pathway is shown to limit microRNA biogenesis in hypoxic conditions and is likely a contributing factor to the downregulation of miR-18a in hypoxia [ 21 – 23 ]. The abundance of miR-18a is also regulated by Myc while Myc expression is repressed by HIF-1 through multiple mechanisms [ 24 , 25 ]. Therefore, formulating and analyzing the signaling axis that connects HIF, Myc, microRNA and TSP-1 in hypoxia may provide insights into the complex dynamics of TSP-1 induction and help screen therapeutic strategies that can efficiently modulate TSP-1 synthesis to regulate angiogenesis.
TSP-1 can activate the latent TGFβ (transforming growth factor beta) molecule, a multifunctional cytokine that plays a key role in inflammation, wound healing, cell proliferation and immune response [ 26 ]. The ligand TGFβ promotes the synthesis of TSP-1 via a positive feedback, possibly through downstream SMAD signals [ 27 ]. Another possible mechanism of how TGFβ mediates TSP-1 synthesis is through the influx of calcium upon TGFβ ligation and the subsequent calcium-mediated activation of NFATc1 (nuclear factor of activated T-cells 1) which is found to be a TSP-1 promoter [ 28 , 29 ]. Summarizing both the hypoxic and TGFβ stimulation of intracellular TSP-1, our mechanistic model presented in this study is the first computational model that considers pathway interactions between the different modes of TSP-1 regulation discussed above. Previous models of TSP-1 studied its interaction with receptors on the cell membrane or TGF-β in the extracellular matrix and paid minimal attention to the complex story of TSP-1 regulation within the cell, but we consider it very relevant to TSP-1 dysregulation in diseases [ 30 , 31 ]. Thus the focus of this work is restricted to hypoxia- and TGFβ-mediated pathways that regulate TSP-1 expression in ECs. We also explored the potential application of the model in more than one cell type, because of the fact that different groups of cells might be responsible for the synthesis of TSP-1 in different pathological conditions [ 32 ]. Assisted by the model, we have identified several key characteristics of intracellular TSP-1 regulation, focusing on the interactive signaling events during receptor activation and hypoxia, as well as the hierarchical regulation of TSP-1 mRNA orchestrated by different intermediate species and microRNAs. We also simulated the model under selected conditions that mimic certain protein profiles observed in tumors and PAD and tested different therapeutic interventions to restore the dysregulated TSP-1 expression back to baseline. The findings presented in this study should help design future experimental and computational research to further investigate the mechanistic regulatory networks that contribute to the abnormal TSP-1 expressions in cancer and in ischemic vascular disease.
Results
Model formulation and assumptions
The computational model presented in this study describes intracellular synthesis of TSP-1 in ECs under the control of multiple signaling axes ( Fig 1 ). The detailed reaction networks are divided into two subparts, (A) intracellular TSP-1 regulation and (B) TGFβ activation of TSP-1, and the diagrams are shown in Fig 2A and 2B . TGFβ pathways have been reported to play profound roles in cancer and cardiovascular diseases; in both situations, the anti-angiogenic effect of the downstream target TSP-1 can be harnessed therapeutically [ 33 – 35 ]. Established models of TGFβ signaling are available in the literature and they cover a wide range of biological details including TGFβ receptors, SMADs and phosphatases in different cellular compartments [ 36 – 38 ]. Due to model complexity concerns, the TGFβ signaling pathway in our model is an adapted version of the work by Nicklas and Saiz, where they included receptor binding, trafficking, SMAD activation, shuttling and feedback [ 39 ]. In addition, we implemented a different module of SMAD7-induced feedback and added the detail of SMAD7-mediated SMAD4 degradation, while SMAD4 is a co-SMAD that binds receptor-regulated SMADs (R-SMADs) [ 40 , 41 ]. Also, a component of the TGFβ-induced calcium signaling network is included in our model with a few rule-based reactions dictating the rate of calcium influx and outflux upon TGFβ activation (see S1 Fig and S1 Table ). Calcium binds and activates calmodulin and calcineurin sequentially, and activated calcineurin rapidly dephosphorylates the inactive NFATc1 in the cytoplasm [ 42 ]. Dephosphorylated NFATc1 is then shuttled into the nucleus and it promotes TSP-1 transcription; NFATc1 may be phosphorylated again and it aggregates in the cytoplasm in its inactive form [ 28 , 43 ]. The current model does not include the potential contribution of calcium to the TGFβ-dependent SMAD activations or the direct binding between calcium and TSP-1 [ 44 , 45 ]. It is important to note that although the signaling events downstream of TSP-1/receptor ligation are not covered in this model, they are reported to be the major effectors of the anti-angiogenic and pro-inflammatory properties of TSP-1 by regulating various molecules including but not limited to reactive oxygen species, Myc, nitric oxide, cyclic guanosine monophosphate (cGMP) and cyclic adenosine monophosphate (cAMP) [ 26 , 46 – 50 ].
10.1371/journal.pcbi.1005272.g001
Fig 1
Induction of TSP-1 via multiple mechanisms by hypoxia and TGFβ signaling.
An increase in the abundance of translatable TSP-1 mRNA in hypoxia results from the regulation by different pathways. Arrow symbol denotes activation, ⊣ symbol denotes repression. MicroRNAs that target TSP-1 (e.g. miR-18a) are less abundant in hypoxic conditions, together with the activation of different TSP-1 promoters, lead to an increase in intracellular TSP-1 production.
10.1371/journal.pcbi.1005272.g002
Fig 2
Reaction diagram of TSP-1 regulation by hypoxia and TGFβ signaling.
(A) HIF stabilization in hypoxia, induction of let-7 and regulation of TSP-1 mRNA by miR-18a. Transcription of TSP-1 gene is modulated by different factors. (B) TGFβ signaling and calcium-mediated activation of NFATc1. Species whose names end with an N subscript are located inside the nucleus; reactions that point to red signs indicate degradation. The symbols v# in the two subparts (A and B) refer to the chemical reactions listed in S1 Table .
The intracellular regulation of TSP-1 synthesis in the model is primarily driven by hypoxia, an important stress signal in tumors and in PAD, through multiple signaling cascades that connect to the HIFs. The oxygen sensing module is similar to the one described by Zhao and Popel, in which they included hydroxylation of HIF mediated by iron, 2-oxoglutarate, PHD (prolyl hydroxylase domain-containing protein) and FIH (factor inhibiting HIF) as key species and processes during HIF stabilization [ 22 ]. The mechanism of HIF-2α stabilization in hypoxia is similar to that of HIF-1α, but the HIF-2 dimer, compared to HIF-1 dimer, is suggested to be a more dominant activator of TSP-1 transcription [ 14 , 51 , 52 ]. Hypoxia-driven induction of HIF-1α promotes the transcription of let-7, a hypoxia-responsive miR (HRM), while the ability of HIF-2α to induce HRMs is similar to that of HIF-1α and thus is not included considering model complexity reduction [ 23 ]. Myc and tumor protein 53 (p53), whose expressions are shown to be affected by hypoxia, have been identified as upstream regulators of TSP-1 with opposing impacts [ 53 , 54 ]. Accumulated HIF-1α potently regulates the expression of Myc by directly promoting its degradation and inducing the MXI-1 (MAX interactor 1) protein which downregulates the transcriptional activity of Myc [ 25 ]. We assumed that MXI-1 exerts opposite transcriptional activity with respect to Myc on all of its target genes in the model. Myc is considered a weak transcriptional repressor of TSP-1 [ 18 ]. Besides this direct interaction, the downregulation of Myc can significantly contribute to TSP-1 induction by upregulating Prosaposin (PSAP) which leads to increased expression of p53, a positive promoter of TSP-1 transcription, and by downregulating the microRNAs that target TSP-1 mRNA [ 55 , 56 ]. HIF-1α accumulated in hypoxia represses the proteasomal degradation of p53 [ 57 , 58 ].
The microRNAs described in the model include let-7 and miR-18a. MicroRNA let-7 plays a master role in the regulation of AGO1 and Dicer which together strongly limit the global microRNA biogenesis in hypoxia [ 23 , 59 , 60 ]. Myc also negatively regulates the abundance of let-7 by inducing the Lin28B (Lin-28 Homolog B) protein which impairs the processing of let-7 primary transcripts in the nucleus [ 61 , 62 ]. The other microRNA, miR-18a, is included in our model to represent the few confirmed TSP-1-targeting miRs and it is reported to be a direct repressor of TSP-1 mRNA in ECs, colonocytes and cardiomyocytes [ 19 , 63 , 64 ]. It is found that expression of miR-18a strongly depends on the transcriptional activity of Myc, which might be part of an indirect mechanism in the Myc-mediated TSP-1 repression [ 65 ]. All the biochemical reactions involving the mechanistic activities of miRs follow the detailed miR biogenesis/targeting mechanisms modeled previously by Zhao and Popel [ 22 ]. The two model subparts converge on the gene transcription of TSP-1, which depends on the activities of multiple transcription factors including HIF-2α, Myc, nuclear phosphorylated SMAD2-SMAD4 complex, nuclear active NFATc1, and p53 in an multiplicative manner [ 14 , 18 , 27 , 28 , 55 ]. One potential connection between the intracellular TSP-1 regulation and the TGFβ activation of TSP-1 is through the activation of the TGFβ pathway which represses the activity of Myc [ 66 ]. In the model, the synthesis of Myc is regulated by the signal downstream of TGFβ activation, which is simplified as the nuclear phosphorylated SMAD2-SMAD4 complexes [ 67 , 68 ]. Another model assumption is that only the proteins/miRs located in the cytoplasm can undergo degradation, and the phosphorylation of SMADs takes place only in the cytoplasm. In microarray data that profile mRNA expression in C57BL/6 mouse with or without experimental hindlimb ischemia, TSP-1 and NFAT are among the top 5% most upregulated genes and MDM2 (Mouse double minute 2 homolog, E3 ubiquitin-protein ligase), which promotes p53 degradation, is in the top 5% most downregulated genes in the ischemic group compared to the non-ischemic group; in addition, MYCT1 (Myc target protein 1), whose transcription is directly influenced by Myc availability, is also modestly downregulated in the ischemic group [ 69 , 70 ]. This evidence supports our model formulation hypothesis that NFAT, Myc and p53 are potential key players in the intracellular regulation of TSP-1. The model contains over 100 species and nearly 200 parameters (see S1 Table and Methods for details); except the small portion of parameters whose values have been measured and calculated in previous studies, the rest of the parameters are estimated by conducting model optimization and validation against literature experimental data (a total of 41 time-course expression trajectories of pathway signature molecules including over 200 data points).
Model optimization and validation
Model parameters are optimized as described in the Methods Section (see S1 Table ) and model simulations are compared with experimental data obtained by different research groups. Valdimarsdottir et al. quantified the phosphorylated SMAD1 and SMAD2 in bovine aortic endothelial cells (BAECs), pretreated with and without the protein synthesis inhibitor cycloheximide (CHX), in response to 1 ng/ml TGFβ (4e-5 μM) [ 39 , 71 ]. In the simulation, the protein synthesis rates of all species are set to zero to mimic the effect of CHX. Fig 3A–3D compare the model simulation with experimental data, and the results imply that CHX treatment prolongs the plateaus of phosphorylated SMAD1 and SMAD2. Fig 3E compares the model-generated dose response curve of total phosphorylated SMAD2 with data obtained in BAECs [ 72 ]. Fig 3F–3L compare the model simulations of time-course protein expressions of various species including HIF-1α, HIF-2α, AGO1, Dicer, p53 and TSP-1 with corresponding experimental data obtained in human ECs [ 14 , 16 , 23 , 59 , 73 ]. Both the experimental data and our model simulations show that HIF-1α, HIF-2α, p53 and TSP-1 protein expressions are induced in hypoxia while AGO1 and Dicer protein levels are downregulated. The simulated calcium and NFAT dynamics are compared to HUVEC (human umbilical vein endothelial cell) data in S1 Fig in which the simulated overall trend of NFAT activation following a single calcium transient mimics the experimental data [ 74 ]. To show that the basic EC model can be further modified to explain fibroblast data as a proof-of-concept analysis, additional model calibration using a different set of parameters optimized against experimental data obtained from fibroblasts are shown in S2 Fig .
10.1371/journal.pcbi.1005272.g003
Fig 3
Model optimization against experimental data in ECs.
(A-D) Experimental measurement and model simulation of phosphorylated SMAD1 and SMAD2 protein in response to 1 ng/ml TGFβ treatment in BAECs: (A) normalized phosphorylated SMAD1 protein level without CHX treatment, (B) normalized phosphorylated SMAD2 protein level without CHX treatment, (C) normalized phosphorylated SMAD1 protein level with CHX treatment, (D) normalized phosphorylated SMAD2 protein level with CHX treatment. The time-course experimental data from Valdimarsdottir et al. and simulation results are both normalized with respect to the corresponding peak value [ 71 ]. (E) Experimental data from Goumans et al. and model-generated dose response curve of total phosphorylated SMAD2 protein measured at 60 minutes after TGFβ treatment. X-axis is in log scale. Values are normalized against the maximum of each dataset [ 72 ]. (F) HIF-1α protein stabilization and (G) AGO1 protein downregulation are observed in HUVECs in hypoxia (2% O 2 ) by Chen et al [ 23 ]. (H-I) Dicer protein is downregulated by hypoxia (1% O 2 ) in HUVECs; hypoxia results in accumulation of HIF-2α protein in human dermal microvascular ECs from Ho et al [ 59 ]. (J) P53 protein expression is induced in hypoxic conditions (1% O 2 ) in HUVECs; experimental data from Lee et al [ 73 ]. (K) TSP-1 protein expression is increased in hypoxia in HUVECs (2.7% O 2 ); data from Phelan et al [ 16 ]. (L) TSP-1 protein expression is induced in human pulmonary aortic endothelial cells in response to 24 hours of hypoxia; data from Labrousse-Arias et al [ 14 ]. (A-L) Results (data-point values and simulations) are normalized against the maximum (or the normoxic baseline values) in each dataset. Quantification of experimental data from the literature is done in ImageJ following standard protocols.
Given the novelty and the complexity of the model, validation is carried out in a way that the model simulations should reach qualitative agreements with uncalibrated experimental data obtained from a variety of different cell types (ECs, cancer cell lines, etc.), in order to partially resolve the issue of model parameter uncertainties. We have gathered additional experimental data from literature on the expression profiles of pathway signature molecules and the comparisons are shown in Fig 4A–4O . Without further calibration against these data, our model that runs with the parameter set obtained from the optimization process discussed above produces simulations that are able to match the trends and relative expression changes of those key molecules with satisfying accuracy. The “test dataset” shown in Fig 4 and the results in S2 Fig together suggest that the dynamics of key molecules in our model are qualitatively consistent in ECs, fibroblasts and certain cancer cell lines. This proof-of-concept step serves as a concrete theoretical basis for future experimental validations of our experiment-based computational model.
10.1371/journal.pcbi.1005272.g004
Fig 4
Qualitative model validation against experimental data in different cell types.
(A-C) Experimental data (symbols) of total phosphorylated SMAD1 and SMAD2 protein in BAECs and mouse embryonic ECs when stimulated by different amount of TGFβ [ 71 , 72 ]. Model simulations are presented by solid curves. (D-F) HIF1/2 protein expression data (symbols) measured in HUVECs and HCT116 colon carcinoma cells at different oxygen tensions [ 59 , 75 , 76 ]. Model simulations presented by solid curves show rapid induction of HIFs in hypoxia. (G) Dicer protein level (symbols) is decreased in hypoxia (1% O2) and measured in human dermal microvascular ECs [ 59 ]. Model simulation is presented by the solid curve. (H-I) Data on MXI-1 mRNA in H460 lung cancer cells and PSAP protein in HCT116 colon carcinoma cells indicate upregulation of both molecules in hypoxia [ 77 , 78 ]. Model simulations confirm the trend. (J-K) Data on Myc protein in hypoxic conditions (symbols) in HCT116 cells, H460 cells, U2OS osteosarcoma cells and Hela cells [ 75 , 77 ]. The downregulation of Myc is also captured by model simulations (solid curves). (L-M) Levels of miR-18a quantified in HCT116 cells, MGC-803 and HGC-27 gastric carcinoma cells [ 21 , 79 ]. Model simulation of miR-18a downregulation in hypoxia agrees with experimental findings. The expression of miR-19a, another miR that potentially targets TSP-1, is also downregulated in hypoxia [ 19 ]. (N) Let-7a, together with other members of the let-7 miR family, belongs to the HRM group, and its induction is captured by model simulation (solid curve) and supported by experimental data (symbols) in HUVECs [ 23 ]. (O) Experimental data (symbols) in NRK52E rat kidney cells indicate that TSP-1 protein expression is stimulated by TGFβ signals [ 27 ]. Model simulation is shown by the solid curve. (A-O) Experimental data-point values and simulation curves are all normalized with respect to the peaks (or normoxic values in the bar-graphs). Literature data are quantified by ImageJ following standard protocols.
Dose dependency of TSP-1 and SMAD7-mediated feedback
The SMAD proteins are the major effectors downstream of TGFβ. The R-SMADs, which typically refer to SMAD1/5 and SMAD2/3, are represented by SMAD1 and SMAD2 in the model [ 41 , 80 , 81 ]. SMAD7 induction follows the activation of the TGFβ pathway, and it associates with the R-SMAD-receptor complex to prevent phosphorylation of these R-SMADs by internalized receptors ( Fig 5A ). SMAD7 induction leads to a downregulation of SMAD4 in the cell by promoting its degradation ( Fig 5B ) [ 40 ]. In response to the rapid build-up of SMAD7 resulting from TGFβ receptor ligation, total SMAD4 level experiences an initial decay followed by a phase of slow restoration as TGFβ signal diminishes ( Fig 5B ). The tail expression of phosphorylated RSMAD-SMAD4 after 20 hrs in Fig 5C and 5D is an outcome of the reduced inhibitory effect of SMAD7: when SMAD7 expression is reduced, some of the sequestered R-SMAD-receptor complex is freed and is able to re-initiate the activation signaling cascade. SMAD7 primarily exerts its inhibitory effect during the peak of TGFβ activation, so a block of its synthesis should enhance R-SMAD phosphorylation and prolong the R-SMAD activation signal following the peak ( Fig 5D ). The tail expression of nuclear phosphorylated RSMAD-SMAD4 is not present when SMAD7 synthesis is inhibited due to the reduced binding between SMAD7 and R-SMAD-receptor complex. Given the dependency of TSP-1 promoter (SMAD2, NFATc1) activation on TGFβ-mediated signaling events, the TSP-1 synthesis curves produced by the model in response to different doses of TGFβ have similar trends compared to the time-course activation of R-SMADs, and the peak TSP-1 levels evaluated at around 10 hrs are shown to be dose dependent ( Fig 5E and 5F ). Model simulations suggest that TSP-1 protein synthesis is significantly elevated compared to the baseline level at TGFβ doses greater than 1 ng/ml, which supports the experimental findings by Nakagawa et al [ 27 ].
10.1371/journal.pcbi.1005272.g005
Fig 5
Dynamics of SMADs and TSP-1 protein synthesis.
(A) SMAD7, whose expression is induced by TGFβ activation, binds the RSMAD-receptor complex and prevents the phosphorylation of the RSMAD in the complex. (B) SMAD4 abundance is negatively regulated by SMAD7 since accumulated SMAD7 promotes SMAD4 degradation. (C) RSMAD-receptor complex sequestered by SMAD7 upon TGFβ activation is released when SMAD7 expression decreases, giving rise to the tail expression of activated RSMAD-SMAD4. (D) Inhibition of SMAD7 protein synthesis removes the tail expression but significantly prolongs the RSMAD-SMAD4 activation signals. (E) TSP-1 protein synthesis stimulated by different doses of TGFβ and (F) the corresponding dose response curve produced by the model. The peaks of total intracellular TSP-1 levels are normalized with respect to the maximum peak observed in the simulation of 20 ng/ml TGFβ stimulation. X-axis is in log scale.
Hypoxia mediates TSP-1 synthesis by promoting its transcription and inhibiting its mRNA repression
The dynamic cooperation between transcriptional and posttranscriptional regulation of TSP-1 may be critical in its induction in response to hypoxia. Fig 6A shows the different TSP-1 induction profiles under different oxygen tensions. The increase in transcriptional activity gives rise to the increase in TSP-1 mRNA available for translation in hypoxia ( Fig 6B ). Another factor that the model hypothesizes to have contributed to the high expression of TSP-1 in hypoxia is the repression of microRNAs (e.g. miR-18a) that target TSP-1 ( S3A Fig ) [ 19 , 21 ]. According to the simulations, downregulation of miR-18a in hypoxia is associated with a decrease in the production rate of miR-18a primary transcript due to repressed Myc expressions ( S3B Fig ) as well as decreased quantities of miR processing molecules, Dicer and AGO1 ( Fig 6C–6E ). Less TSP-1 mRNA is under repression in hypoxia while more mRNA is ready for translation due to the increased transcription and decreased miR targeting ( Fig 6F ). Since the stabilization of HIF is highly nonlinear with respect to oxygen tension and HIF initiates TSP-1 activation by multiple mechanisms, we wonder if there is also a switch-like behavior in the synthesis of TSP-1 at different oxygen tensions [ 82 ]. Fig 6G shows the normalized TSP-1 protein level as a function of percent oxygen, and the model predicts a nonlinear relationship when TSP-1 level is measured at both 24 hours and 48 hours; the threshold for induction of TSP-1 is centered around 6–8% oxygen. Since the model is constructed based on in vitro data, it should be noted that physiological tissue oxygen tension in vivo is usually much lower than 21% oxygen ( in vitro normoxia), and such a discrepancy may affect our model conclusions when compared with in vivo experimental observations [ 83 ].
10.1371/journal.pcbi.1005272.g006
Fig 6
Low oxygen induces TSP-1 expression by different mechanisms.
(A) TSP-1 protein expression is increased along with decreased oxygen availability. (B) Hypoxia increases the transcriptional activity of TSP-1 gene through the induction of its promoters. (C-D) miR-18a, a TSP-1 targeting miR, is downregulated in hypoxia due to the downregulation of its promoter Myc, and (E) miR processing molecules AGO1 and Dicer. (F) Hypoxia de-suppresses the TSP-1 mRNA that was originally under miR-mediated repression, which contributes to the increase of total translatable TSP-1 mRNA. The ratio of repressed TSP-1 mRNA over free mRNA decreases more than tenfold in the condition of 2% oxygen. (G) Model prediction of a nonlinear behavior in TSP-1 synthesis with respect to oxygen tension. This relationship may be partially contributed by the switch-like nature of cellular oxygen sensing [ 22 , 82 , 84 ]. (B-G) Results are normalized with respect to the normoxic baseline values.
Therapeutic strategies targeting the TGFβ-HIF-miR-TSP1 axis in diseases
Research on TSP-1 has established its promising role as future therapeutics in cancer and vascular disorders [ 85 – 88 ]. Computational studies such as our model may help design experiments to select the best strategy to modulate TSP-1 expression in these pathological conditions by running in silico experiments and assessing the results. In the following two subsections, we investigate how different factors contribute to TSP-1 dysregulation in tumors and in PAD and compare the efficacy of different model-motivated therapeutics.
Tumor
Studies have confirmed that the enforced expression of TSP-1 in certain types of cancer, including lung and breast cancer, were associated with reduced tumor growth and metastasis [ 89 – 91 ]. Since Myc oncogene is often found overexpressed in tumors and given the connection between Myc and TSP-1, one potential mechanism contributing to the Myc-induced tumorigenesis is via the downregulation of TSP-1 [ 18 , 92 – 94 ]. Therefore, enhancing TSP-1 expressions in these scenarios would provide anti-tumor benefits putatively by limiting angiogenesis and downregulating Myc [ 95 ]. The model suggests that in normoxia, Myc overexpression significantly downregulates the abundance of TSP-1 proteins ( Fig 7A ). However, in hypoxia, the contribution of Myc hyperactivity to TSP-1 downregulation is less significant and a simulated knockdown of Myc induces TSP-1 with a smaller fold increase (~1.5 folds at 48 hrs) compared to the substantial upregulation in normoxia (~5 folds at 48 hrs) ( Fig 7B and 7C ). This can be explained by the observation that hypoxia strongly downregulates Myc so the additional effect of reduced Myc synthesis on TSP-1 expression is less significant. In normoxia, hyperactive Myc does not repress TSP-1 transcription significantly, while it dramatically induces miR-18a production which results in increased TSP-1 mRNA repression ( Fig 7D ). Myc-induced miR-18a upregulation is an outcome of elevation in both transcriptional activation and AGO1 abundance, which helps to stabilize miRs and is controlled by the Myc-Lin28B-let7 axis ( S4 Fig ) [ 23 , 61 ]. In order to reverse the downregulation of TSP-1 in simulations mimicking Myc-induced tumors, we test and assess several pathway-related therapeutic interventions including small molecule inhibitors of Myc, miR-18a antagonists and TGFβ treatments (Figs 7E–7G and S5 ) [ 96 ]. In a span of 24 hours, various doses of Myc inhibitors and miR-18a antagonists both elevated TSP-1 production by different amounts, while a rapid inhibition of Myc activity using a very high dose (200 nM) of Myc inhibitors successfully enhanced TSP-1 expression beyond normoxic steady-state level (simulation starting point) ( Fig 7E and 7F ). Forced TGFβ stimulation of TSP-1 may also be an effective strategy in the cases of Myc hyperactivity, since TGFβ signaling delays Myc synthesis (Figs 7G and S5D ). The simulation results suggest that miR antagonists and TGFβ stimulations are more effective in terms of restoring TSP-1 expression in the early phase, while Myc inhibitors could give a higher TSP-1 expression in the long term (evaluated at 24 hrs). Although hypoxia suppresses Myc expression, the model predicts that hyperactivity of Myc due to gain-of-function mutations/deregulations still hinders the hypoxic induction of TSP-1 ( Fig 7B ), suggesting that cells produce insufficient anti-angiogenic factors (e.g. TSP-1) in these scenarios which fails to counteract the strong hypoxia-driven activation of pro-angiogenic factors (e.g. VEGF), and that this imbalance might be one possible reason that turns on the angiogenic switch [ 97 , 98 ].
10.1371/journal.pcbi.1005272.g007
Fig 7
Different therapeutic interventions to restore TSP-1 level in Myc-induced tumors.
(A) Hyperactivity of Myc is simulated by increasing its rate of production (baseline rate multiplied by 5), which results in a significant downregulation of TSP-1 protein level. Experimental evidence indicates that Myc expression could be 2–10 times higher in tumor samples compared with control samples [ 99 – 102 ]. (B) The effect of TSP-1 repression by Myc overexpression is less obvious in hypoxia. (C) In the cases of Myc hyperactivity, knockdown of Myc synthesis (protein synthesis rate reduced to 10% of the hyperactive value) leads to a more notable increase in TSP-1 protein abundance in normoxia compared to hypoxia. (D) Hyperactive Myc engages both the direct transcriptional and posttranscriptional pathways (miRs) to repress TSP-1 protein expression. (E) Simulating the application of a Myc inhibitor at different doses under the condition of Myc hyperactivity. (F) Simulating the application of a miR-18a antagonist at different doses under the condition of Myc hyperactivity. Simulation results suggest that miR antagonists reach a maximum efficacy at around 40 nM. (G) Direct transcriptional stimulation by different doses of TGFβ results in a significant early increase in TSP-1 protein expression under Myc hyperactivity compared to the other two strategies. (A-G) Results are normalized with respect to the normoxic steady-state values calculated with baseline Myc synthesis rate. (E-F) The simulations assume that Myc inhibitors potently bind and sequester cytoplasmic Myc with a Kd of 1 nM, and that miR-18a antagonists bind and sequester miR-18a RISC with a Kd of 1 nM [ 103 ].
Peripheral arterial disease
Research has suggested a detrimental role of TGFβ signaling in the progression of PAD by measuring the levels of TGFβ and its receptors in ischemic tissues and in peripheral blood collected from patients [ 104 – 106 ]. We are interested in the effect of TGFβ-induced TSP-1 synthesis and potential interventions to reduce TSP-1 in both normoxic and hypoxic conditions, since TSP-1 levels are found upregulated in PAD patients [ 13 , 107 ]. In the most severe form of PAD, critical limb ischemia (CLI), arterial blood flow is restricted so severely that tissues are constantly suffering from hypoxia and ischemia [ 108 ]. Therefore, we simulated the condition of CLI as a combination of hypoxia and TGFβ which together contribute to elevated TSP-1 expressions ( Fig 8A ). Fig 8B–8D explore different potential treatment strategies during a 24-hr simulation timespan; under the simulated condition of CLI, inhibition of p53 is shown to be more effective in limiting TSP-1 production than antagonizing let-7 or overexpressing miR-18a (directly targets TSP-1 mRNA). Fig 8E compares time-course TSP-1 protein expressions in response to different combination therapies that this work proposed. A combination of both p53 and NFAT inhibition, designed to limit TSP-1 production by inhibiting both the hypoxia-induced and TGFβ-induced pathways, can most effectively reduce TSP-1 protein expression below the normoxic steady-state level (without TGFβ treatment) at the end of 24 hrs. NFAT inhibitors (e.g. VIVIT) specifically block the calcineurin-NFAT interaction and abolish the gene activation downstream of NFAT, which is assumed in the model to be activated upon TGFβ-induced calcium influx [ 109 , 110 ].
10.1371/journal.pcbi.1005272.g008
Fig 8
Different therapeutic interventions to reduce TSP-1 levels in PAD.
(A) Hypoxia and TGFβ together contribute to a much higher intracellular TSP-1 protein expression in the simulated condition of CLI than the normoxic control condition (with no TGFβ). TSP-1 expression curves in response to different doses of (B) let-7 antagonists, (C) miR-18a mimics, and (D) p53 inhibitors. The effect of let-7 antagonist peaks at around 40 nM. (E) A combination of inhibiting both p53 production and NFAT activities achieve the most significant downregulation of TSP-1 in the simulated condition of CLI. In the simulations, the strength of NFAT inhibitor VIVIT is estimated from literature data to be a 70% decrease in the rate of calcineurin-mediated NFAT dephosphorylation when applied at micromolar doses [ 111 ]. (B-E) MiR-18a overexpression is simulated as an increase in the initial condition of precursor miR-18a; let-7 antagonists bind and sequester let-7 RISC with a Kd of 1 nM; small molecule inhibitor of p53 binds and sequesters cytoplasmic p53 with a Kd of 1 nM.
Model sensitivity analysis
We performed global sensitivity analysis using the techniques of Partial Rank Correlation Coefficient (PRCC, see Methods ) under different simulated conditions to identify parameters that most significantly control the key species in the model [ 112 ]. S6 Fig displays the distribution of model parameters and the corresponding experimental measurements [ 113 , 114 ]. Most of the parameter values after optimization are within one-two orders of magnitude compared to the experimental median values. Certain parameter values that deviate significantly from the experimental median are calculated based on literature data, such as the constitutive degradation rates of TGFβR (0.0278 min -1 ) in S6B Fig and the calmodulin concentration (5.9371 μM) in S6D Fig [ 36 , 115 ]. From the sensitivity analysis, we observed that the HIF-1 dimer level, an indicator of HIF-mediated transcriptional activities in hypoxia, is negatively regulated by an increase in the affinity between HIF-1α and its two hydroxylases, FIH and PHD, which will subsequently promote HIF-1α degradation; as expected, increased binding between oxygen and FIH/PHD-DG-Fe complex (parameters kf4, kf8) speeds up the degradation of HIF ( Fig 9A ). Interestingly, although increased dimerization between HIF-1α and HIF-1β (parameter kf19) increases HIF-1 dimer levels, it downregulates the total HIF-1α protein level within the cell, presumably due to a higher synthesis of the HIF-destabilizing protein TTP (parameter kf18) as described by previous studies (Figs 9A and S7A ) [ 116 , 117 ]. Phosphorylation rate of cytoplasmic SMAD2 (parameter kf75) and SMAD4 shuttling rate (from cytoplasm to nucleus, parameter kf79) are the two most influential factors that positively regulate the levels of active, phosphorylated SMAD2-SMAD4 complex in nucleus, which represents the signaling strength of TGFβ pathways ( Fig 9B ). Besides the rates relating to SMAD7 feedback, increases in other factors such as the binding between TGFβ and its receptor (kf73) and the degradation of SMAD4 (vm34) are both correlated with less total activation of R-SMADs ( Fig 9B ). Sensitivity analysis of factors that control TSP-1 synthesis indicate that parameters relating to the abundance of transcription factors, including Myc, p53, HIFs, and NFAT are more influential ( Fig 9C and 9D ). Besides the strategies of TSP-1 or HIF gene therapies (parameters vm1, vm2, vm20), the model suggests that small molecule inhibitors against Myc and miR-18a can effectively restore and enhance TSP-1 protein expressions in tumorigenic conditions provoked by Myc overexpression ( Fig 9C ). In Fig 9D , manipulating the expression of miRs (let-7 or miR-18a alone) in simulated CLI conditions modulates TSP-1 production to a lesser extent compared to the approaches that directly target the transcription factors, and the results in Fig 8E also support the conclusion derived from the sensitivity analysis that targeting NFAT and p53 together may be a more efficient strategy. Additional results ( S7 Fig ) of model sensitivity in simulated conditions different from the ones presented in Fig 9 are consistent with the results discussed here. Given the potential interactions between different transcription factors (e.g. NFAT, HIF, p53, Myc) at the DNA level and the limited knowledge on how the influence of each individual promoter/inhibitor converge during TSP-1 transcription, the conclusions from the sensitivity analysis are biased by the simplification we made when translating the complex transcriptional activities into mathematical equations. The fact that our model assumed a multiplicative effect of different transcription factors is reflected by the results in Fig 9C and 9D that the TSP-1 protein level is relatively more sensitive to the parameters that control the abundance of its transcriptional promoters/repressors.
10.1371/journal.pcbi.1005272.g009
Fig 9
Global sensitivity analysis of model parameters.
Global sensitivity analysis of parameters that control (A) the area under curve (AUC) of HIF-1 dimer in a span of 48 hours in 2% oxygen, (B) AUC of activated SMAD2-SMAD4 complex in nucleus in a span of 24 hours upon 2.5 ng/ml TGFβ stimulation, (C) AUC of TSP-1 in a span of 24 hours under the condition of normoxia plus Myc hyperactivity and (D) AUC of TSP-1 in a span of 24 hours upon 2.5 ng/ml TGFβ stimulation plus hypoxia (2% oxygen). (A-D) Rate descriptions—kf1: HIF-1α translocation into nucleus; kf2: HIF-1α binds FIH complex; kf4: oxygen binds FIH complex; kf7: HIF-1α binds PHD complex; kf8: oxygen binds PHD complex; kf11: HIF1α-OH-FIH dissociation; kf13: HIF1α-OH-PHD dissociation; kf18: TTP synthesis; kf19: HIF-1α binds HIF-1β; vm1: HIF-1α synthesis; vm2: HIF-2α synthesis; vm3: Myc synthesis; vm5: let-7 synthesis; vm6: MXI-1 synthesis; vm7: miR-18a synthesis; vm9: LIN28B synthesis; vm11: PSAP synthesis; vm13: p53 synthesis; vm18: AGO1 synthesis; vm19: Dicer synthesis; vm20: TSP-1 synthesis; kf66: TGFβR internalization; kf72: TGFβR degradation; kf73: TGFβ binds its receptor; kf74: receptor dimer binds R-SMAD; kf75: R-SMAD phosphorylation; kf76: R-SMAD translocation into nucleus; kf77: phosphorylated R-SMAD (pR-SMAD) binds SMAD4; kf78: pR-SMAD-SMAD4 complex translocation into nucleus; kf79: SMAD4 translocation into nucleus; kf82: pSMAD2-SMAD4 dephosphorylation; kf89: SMAD7 sequesters activated R-SMAD; vm27: TGFβ-mediated calcium influx; vm29: NFAT dephosphorylation; vm32: SMAD7 synthesis; vm34: SMAD4 degradation; vm36: SMAD4 synthesis.
Materials and Methods
Formulation of reactions
We constructed the model based on ordinary differential equations (ODE) with a total of 109 species, 195 kinetic parameters and 138 reactions ( Fig 2 ). Description of reactions, parameter values ( S1 Table ), and initial conditions for all species ( S2 Table ) are available in the appendices. The model allows translocation for certain species, especially the receptor and SMAD complexes, and distinguishes them by cellular locations–in cytoplasm, nucleus or endosome, since their functions are different in different cellular compartments. Transcriptional activation/repression and Dicer cleaving are modeled as Hill-type or Michaelis-Menten kinetics. Most interactions captured by the model are based on literature evidence. All data including reactions, rates, rules and initial conditions used in the model are compiled using MATLAB SimBiology toolbox (MathWorks, Natick, MA). Simulations are performed using the ode15s and sundials method, which are both ODE solvers provided in MATLAB. Since hypoxia is a focus of the study, the initial conditions of all species are their respective steady-state levels in simulations assuming normoxia (21% O 2 ) and no TGFβ treatment. For miR treatments simulated by the model, overexpression of the miR mimic increases the initial condition of the corresponding precursor miR; miR silencing is described as the association of miR antagonist with miRISC to form a complex that cannot function. Our ODE-based computational model inherently considers time delays in biological events and is designed to simulate the average dynamical behavior of different biomolecules considering the stochasticity of cellular activities (binding, transcription, etc.), given the reasons stated in one of our previous works [ 22 ]. Although it is suggested that stochasticity plays a critical role in gene transcription, many signaling network studies that used deterministic approaches to model transcriptional events have been able to generate insightful results that are further validated by experiments [ 118 – 123 ]. Another reason why we did not use the stochastic approach to model transcription is that our study focuses primarily on the dynamic signal transduction and pathway cooperation within the network that together contribute to the induction/inhibition of TSP-1 protein expression in different circumstances, instead of the details in the transcription factor binding process at the DNA level. ImageJ software (NIH) is used to perform densitometry analysis according to the blot analysis protocol in order to obtain the experimental data showed in the model optimization and validation sections.
Estimation of model parameters and initial conditions
Due to the limited literature on miR and TSP-1 modeling and the fact that this model is the first that describes the complex regulation responsible for TSP-1 synthesis under different physiological conditions, we paid considerable attention to parameter estimation and optimization during model construction. Many of the rate parameters and initial conditions used in the TGFβ signaling subpart are taken from the work by Nicklas and Saiz in which they calculated the values based on experimental measurements [ 39 ]. Parameters used in the component describing calcium-mediated NFAT activations are estimated and then optimized to reproduce the qualitative experimental behaviors of calcium and NFAT observed in ECs [ 74 ]. Intracellular concentrations of calcium are estimated based on data from [ 124 ]. For the initial conditions of miRs, we compared literature data and assumed that miR levels are on the order of 10 3 to 10 4 copies per cell in normoxia; the concentrations (in microMolar) used as initial conditions in the model are computed using 1 pL cell compartment volumes based on literature measurements [ 125 – 128 ]. Absolute levels of the different proteins in the model are estimated to be on the order of 10 4 to 10 6 copies per cell based on experimental measurements of several pathway-related proteins including Myc, p53 and calmodulin [ 115 , 129 , 130 ].
We estimated the decay rates of mRNA (1.2e-3 min -1 ), miRNA (1e-4 min -1 ), protein (2.5e-4 min -1 ), translation rate per mRNA (2.33 min -1 ), transcription/mRNA synthesis rate (1.92e-7 μM/min), and the levels of mRNA (2.8e-5 μM) and protein (0.08 μM) in normoxia so that the final values are within ±2 orders of magnitude compared to the median values (normalized by cell compartment volumes, and indicated in the brackets) reported by global quantification studies [ 22 , 113 , 114 , 131 , 132 ]. The rest of the parameters and initial conditions are estimated based on previous computational studies (summarized in S1 and S2 Tables) [ 22 , 39 , 84 ]. The volume concentrations of surface TGFβR are calculated by assuming that the receptors are distributed uniformly within a space of 1 pL given an estimated flat EC surface area of 1000 μm 2 [ 133 , 134 ]. We used the Levenberg-Marquardt algorithm within the lsqnonlin function in MATLAB for model optimization. Since the related time-course data in ECs are limited, the parameters are optimized by minimizing the sum of squared errors between normalized model simulations and experimental measurements (see Fig 3 in Results for details). The same protocol is repeated in the optimization of the model against the fibroblast dataset.
Sensitivity analysis
Global sensitivity analysis is performed using the PRCC algorithm, a sampling-based method developed by Marino et al. to quantify uncertainty in the model. The outputs of interest in the sensitivity analysis are the time integrals of the signals computed in the form of AUC over certain durations, and a sample size of 1000 runs is chosen for each module of sensitivity analysis. The distribution of each parameter tested is within a two orders of magnitude range with a center at the parameter’s original value (e.g. x/10 to 10x). Details and examples of the PRCC algorithm can be found in [ 112 ].
Discussion
In this study, a detailed mass-action based computational model of multiple signaling pathways connecting to TSP-1 regulation is presented. The comprehensiveness and trustworthiness of the model is supported by a careful analysis of literature during model formulation and extensive efforts of model training/validation against experimental data. This work is a continuation of a previous model presented by our group, while in that model VEGF is the major focus [ 22 ]. The scope of the current model is not limited to intracellular signaling since we included the module of TGFβ/receptor signaling as an important path of TSP-1 activation. TSP-1 is long known to be an activator of TGFβ, but the potential role of TGFβ on TSP-1 activation has not received much attention [ 135 ]. The model connects independent literature evidence and hypothesizes both a direct and indirect TSP-1 activation path initiated by TGFβ stimulation via SMADs and calcium regulation. This potential positive feedback loop that amplifies both TGFβ1 and TSP-1 expression might be an explanation to the paired high TGFβ1 and TSP-1 levels observed in certain pathological conditions [ 28 , 136 ]. Although the regulatory roles of TSP-1 in tumor progression is highly cell-type specific, the undesirable anti-angiogenic effect resulting from high TSP-1 expression in PAD is a major interest to cardiologists and vascular biologists [ 107 , 137 , 138 ]. Our model proposed that TGFβ might be an underlying factor driving the high expressions of TSP-1 in PAD patients, given the experimental evidence of TGFβ1 elevation in ischemic tissues [ 106 , 139 ]. It is interesting to note that hypoxia also upregulates TGFβ1 production in smooth muscle cells in addition to the direct transcriptional induction of TSP-1 via HIFs, suggesting another layer of crosstalk between the pathways that control TSP-1 expression [ 14 ]. The biology of hypoxia-induced TSP-1 seems contradictory to the need of angiogenesis when cells are exposed to insufficient oxygen, however, this phenomenon may be more likely an endogenous feedback control developed by the body to contain the angiogenesis driven by pro-angiogenic factors (e.g. VEGF) that are radically produced upon hypoxia [ 140 ].
The potential therapeutic interventions tested in this study to enhance TSP-1 production in simulated conditions of tumors are based on the assumption that these tumors are induced by Myc hyperactivity. Given the profound role of Myc in growth, proliferation, tumorigenesis and stem cells, the focus of our study, TSP-1, is only one of the many potential downstream targets of Myc that have correlations with tumor progression [ 92 ]. It is worth noting that Myc can induce miR-17/92 cluster which targets key proteins in TGFβ signal transduction and represses gene regulation downstream of TGFβ in multiple cancer cell lines, and that the pro-tumorigenic property of Myc overexpression is lost in TGFβ-deficient xenograft models of colorectal cancer; such evidence suggests that Myc may promote tumor growth primarily by repressing the anti-tumorigenic gene expression (including TSP-1) activated by TGFβ signaling, at least in the context of colorectal cancer [ 141 , 142 ]. Moreover, the mutually inhibitory relationship between TSP-1 and Myc may further amplify the signal of one molecule and suppress the other in diseases [ 95 ]. VEGF is also shown to be a target activated by Myc [ 143 , 144 ]. Although research has shown that TSP-1 overexpression can effectively reduce tumor metastasis, the position of TSP-1 in the entire network of cancer-related genes is relatively downstream, which might imply that targeting TSP-1 to attack tumor may be less efficacious than targeting the genes (e.g. Myc) that are more central in the network, since cancer is notorious for developing compensatory pathways to resist targeted therapies [ 145 , 146 ]. The failure of TSP-1 analog (ABT-510) in phase II trials against metastatic cancer should not discourage the continuum of research that aims to explore the therapeutic potential of TSP-1, especially in cardiovascular diseases where its importance has emerged in recent years; on the other hand, multiple phase I studies that explore the targeting of CD47 in cancer, given its inhibitory effect on the immune response, are now under way [ 147 – 150 ]. Still, our simulations proposed that increased TSP-1 synthesis is a possible downstream effector of the tumor-suppressive property of TGFβ signaling, specifically in Myc-dependent tumors; however, the exact role of TGFβ in cancer is quite complex and controversial given its bipolar control of tumorigenesis [ 151 – 155 ].
Sensitivity analysis indicates that TSP-1 production stimulated by hypoxia and TGFβ is strongly influenced by the activity of several transcription factors, namely HIFs, p53 and NFAT. Although the model assumes that HIF-1 does not directly promote TSP-1 transcription, an increase in its abundance, as shown by the sensitivity analysis, has a notable influence on TSP-1 levels comparable to that of HIF-2, which directly activates TSP-1 [ 14 ]. The indirect activation of TSP-1 by HIF-1α might be undesired for gene therapies that use adenoviral HIF-1α to improve angiogenesis and limb perfusion in patients with ischemic vascular diseases [ 156 ]. This might also explain the finding that the angiogenic potency of adenoviral HIF-1α is significantly lower than that of adenoviral VEGF [ 157 ]. To date, the roles of p53 and NFAT, the two potential therapeutic targets identified by our simulations, in PAD are largely unknown. The limited evidence in the literature agrees with the model hypothesis that NFAT, with TSP-1 as one of its effector molecules, potently participates in the cellular response to hypoxia/ischemia: inhibition of NFAT is found to suppress atherosclerosis in diabetic mice, while a significant increase in NFAT expression is observed in ischemic rat brain [ 158 , 159 ]. Tumor protein p53 is long known to be a critical factor in suppressing tumorigenesis and initiating apoptosis; its pro-apoptotic property might render it a promising target in PAD given multiple clinical observations of the increased level of apoptotic events in the serum and tissue of PAD and CAD (coronary artery disease) patients [ 160 – 163 ]. Still, the robustness and reliability of our model-based conclusions can be further enhanced by additional model training, calibration and validation when more experimental measurements (e.g. data of HIFs, miRs, AGOs, Myc, SMADs and TSP-1 expressions in different physiological condition/stimulation) become available in the near future.
The current model describes the dynamics of let-7 and miR-18a in intracellular regulation of TSP-1, but the model is set up in a way that incorporating additional miRs and their targets is feasible. Besides the miRs that target HIFs such as miR-155, many other miRs could be potential candidates to consider for future computational models of TSP-1 regulation [ 164 ]. In the current model simulations, we assume that inhibition of p53 is achieved by the binding of small molecule inhibitors (e.g. Cyclic Pifithrin-α), while p53 is reported to be a target of several miRs including miR-125b and miR-504 [ 165 – 167 ]. The p53 protein also regulates the expression of certain miRs; an example is miR-194, a p53-responsive miR which targets TSP-1 in colon cancer cell lines [ 56 ]. Liao et al. identified let-7g, one of the HRMs, as a factor that improves endothelial functions with targets including TSP-1 [ 168 ]. Future models could also consider alternative pathways relating to TSP-1 regulation that have implications in vascular disorders, such as the axis involving VEGF activation of NFAT in ECs [ 169 ]. The signaling pathway involving PI3K/AKT/PTEN is also shown to mediate TSP-1 expression in both cancer cells and ECs [ 170 ]. Including the VEGF signaling pathway as a part of intracellular TSP-1 regulation seems to be an exciting next step in the future development of our model, since most studies have focused on the regulatory effect of TSP-1 on VEGF but not the other direction. Besides hypoxia, many other factors relating to cancer and PAD including radiation, high glucose and aging have also been shown to affect TSP-1 expression [ 171 – 176 ]. So far, the model is mostly formulated and validated based on knowledge and data of pathways in ECs, which are shown in experimental studies to express TSP-1 at high levels and that the EC-secreted TSP-1 is critical in certain physiological processes [ 28 , 177 ]. Given the fact that the amount of ECs is only a small percentage of all the cells in tissues of tumors or PAD, and other types of cells including smooth muscle cells, stromal fibroblasts and immune cells also secrete TSP-1, there is a need for further model validation in order to sustain and extend our model conclusion, at least qualitatively, to other cell types of interest [ 26 ]. We have already demonstrated the feasibility of this by conducting additional validation against experimental data from other cell types (e.g. cancer cells, fibroblasts etc.), but the related data in smooth and skeletal muscle cells are relatively scarce. A goal of this study is to raise carefully-formulated hypotheses that stimulate future research to produce additional experimental results that either corroborate or refute our predictions. In summary, our model is the first computational study that investigates the complex network of intracellular TSP-1 regulation mediated by hypoxia, microRNA-targeting and receptor signaling. It is an important complementary study to the active research that focuses on the interaction between TSP-1, VEGF and their receptors at the cell surface, and it also provides insights, from the perspective of intracellular control, into the search for therapeutic strategies that adjust TSP-1 activity in order to modulate angiogenesis in cancer and vascular diseases. Combined with additional modules of pharmacokinetic analysis, our computational model can help identify optimal treatment strategies and design appropriate dosing schemes that progressively reduce or enhance TSP-1 expression in patients depending on the specific indication.
Supporting Information
S1 File
Compiled file of all supporting information documents.
(PDF)
S1 Table
Reaction descriptions, reaction rates and kinetic parameters of TSP-1 model.
(PDF)
S2 Table
Differential equations and species initial conditions.
(PDF)
S1 Fig
Model of TGFβ-induced calcium regulation and downstream activation of NFAT in ECs.
(PDF)
S2 Fig
Additional model calibration against published fibroblast data using different parameter values.
(PDF)
S3 Fig
De-suppression of TSP-1 mRNA and downregulation of Myc in hypoxia.
(PDF)
S4 Fig
AGO1 upregulation following Myc overexpression.
(PDF)
S5 Fig
Testing different therapeutic strategies in hypoxia and hyperactive Myc conditions.
(PDF)
S6 Fig
Model parameter distributions.
(PDF)
S7 Fig
Additional sensitivity analysis.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Spinocerebellar ataxia type 3, also known as Machado-Joseph disease (SCA3/MJD), is the most common dominantly inherited ataxia [1] . It is a member of the polyglutamine (polyQ) neurodegenerative disease family which includes Huntington's disease (HD), spinal and bulbar muscular atrophy (SBMA), dentatorubral- pallidoluysian atrophy (DRPLA), and spinocerebellar ataxias 1, 2, 3, 6, 7, and 17 [2] – [4] . It has been demonstrated that polyQ expansion increased the cellular toxicity of the proteins and was responsible for the diseases. In normal individuals, the length of the CAG repeat varies between 12 and 37 trinucleotides whereas in SCA3/MJD patients it varies between 49 to 86 repeat units which located near the carboxy-terminus of SCA3 gene (MJD1) on chromosome 14q32.1 [5] , leading to the toxic translational product of polyQ-expanded ataxin-3. The pathology of SCA3/MJD includes severe neuronal loss in the spinal cord and specific brain regions, such as dentate nuclei (cerebellum), pontine nuclei (brainstem), and substantia nigra (basal ganglia) [6] – [7] . Nuclear inclusions are detected in both affected and unaffected neurons of SCA3/MJD patients [8] – [9] . It is unclear if these aggregates contribute to neuronal dysfunction or possibly represent a protective mechanism, although some recent models suggest an inverse correlation between accumulation of aggregates and neuronal loss [10] – [11] .
Recently, post-translational modifications have been shown to play a major role in the pathogenesis of polyQ diseases. There is increasing evidence demonstrating that different target proteins can be post-translational modified by SUMOylation. And the modified proteins are possible to involve in numerous neurological diseases including polyQ disorders [12] . SUMO is an ubiquitin-like protein with 20% identity to ubiquitin [13] . In vertebrates, the SUMO family has at least four members, SUMO-1, SUMO-2, SUMO-3, and SUMO-4 [14] – [17] . SUMO modification may have altered the function, activity or localization of its substrates [14] , [18] – [20] . The conjugation of SUMO proteins, or SUMOylation, is a post-translational modification process that shares common ancestry and core enzymological features with ubiquitination but has distinct functional roles. SUMOs initially exist in an inactive form, which is processed by the SUMO specific protease to expose the glycine residues at their carboxy-terminal that are required for the formation of SUMO–protein conjugates. SUMOylation is a multistep process, which involves an activating enzyme E1 (SAE1 and SAE2), a conjugating enzyme E2 (Ubc9) and, in some cases, a ligating enzyme E3 [21] – [22] .
SUMOylation is thought to modify the interactions in multiprotein complexes [23] . Beside its role as a covalent modifier, SUMO can bind non-covalently to SUMO-interacting motifs, which have been identified in many proteins [24] , among which several are related to polyQ diseases such as androgen receptor, huntingtin, ataxin-1, and ataxin-7 [25] – [28] . SUMO and ubiquitin share a common three-dimensional structure, except that SUMO has an additional short amino terminal extension [29] . It has been reported that SUMO modification of some proteins on a lysine residue blocks ubiquitination at the same site, resulting in an inhibition of protein degradation and an alteration of protein function [26] , [30] . In HD, SUMOylation of mutant huntingtin increases the stability of the protein and exacerbate neurodegeneration.
In our previous study, SUMO-1 had been identified as a novel ataxin-3-interacting protein by yeast two-hybrid technology. Both co-immunoprecipitation and immunofluorescence staining results proved that ataxin-3 was a target for SUMOylation both in vitro and in vivo [31] , [32] . In order to reveal the exact role of SUMOylation in the pathogenesis of SCA3/MJD, here we report that the major SUMO-1 binding site was identified, which located on lysine 166 (K166) of the mutant-type ataxin-3. SUMOylation did not influence the subcellular localization, ubiquitination or aggregates formation of mutant-type ataxin-3, but partially increased its stability and the apoptosis rate of the cells. Our findings are the first to indicate the effect of SUMOylation on the stability and cellular toxicity of mutant ataxin-3 and implicate the role of SUMOylation in SCA3/MJD pathogenesis.
Results
Ataxin-3 was modified by SUMO-1 on lysine 166
Firstly, the potential SUMOylation motifs on ataxin-3 were predicted by software, “SUMOplot™ prediction” ( www.abgent.com/doc/sumoplot ). The result suggested at least three consensus SUMOylation sequences in ataxin-3, which were K8 in EKQE, K166 in VKGD and K206 in HKTD. Based on these outputs, we constructed three mutants of ataxin-3, ataxin-3 K8R , ataxin-3 K166R , and ataxin-3 K206R , in which the lysine 8, lysine 166 or lysine 206 were all converted to arginine (R). As shown in Figure 1 , slow migrating bands were observed using both ataxin-3 K8R and ataxin-3 K206R as binding substrates of SUMO-1 while no migration was observed when ataxin-3 K166R was used. The results presented in Figure 1 clearly showed that only the conversion of lysine 166 to arginine abrogated the SUMOylation of ataxin-3, meaning lysine 166 was the SUMOylation site in ataxin-3.
10.1371/journal.pone.0054214.g001
Figure 1
Identification of the SUMO-1 modification sites in ataxin-3.
(A) HEK293 cells were used to co-express ataxin-3, ataxin-3 K8R , ataxin-3 K166R , or ataxin-3 K206R with SUMO-1. 10% lysates were precipitated by TCA and subjected to immunoblotting. #ataxin-3 main bands, ## ataxin-3 relative levels of modified and unmodified (whole-cell–TCA precipitates). (B) The tagged proteins were enriched with NTA magnetic nickel columns and detected by immunoblotting with SUMO-1 antibody. **, ataxin-3-20Q modified by SUMO-1, ##, ataxin-3-68Q modified by SUMO-1.
SUMO-1 modification of ataxin-3 did not affect its subcellular localization
SUMOylation has been reported to be able to regulate the subcellular localization of several proteins. To examine whether SUMOylation of ataxin-3 affects its localization, we compared the localization of ataxin-3 and the SUMOylation deficient variant ataxin-3 K166R in transiently transfected HEK293 cells ( Figure 2A ). As previously described, overexpression of ataxin-3-20Q showed a diffusive distribution both in the nucleus and in the cytoplasm. The subcellular localization of ataxin-3-20Q K166R was similar to that of ataxin-20Q. Interestingly, aggregates formation in overexpressed ataxin-3-68Q was also the same as that in ataxin-3-68Q K166R transfected cells. There was significantly difference between wild-type and mutant type ataxin-3, the latter one formed aggregates but the former one did not ( Figure 2A ). The proteins levels in the fractions of cytoplasm and nucleus in the cells transfected with above plasmids suggested again that SUMO-1 modification did not affect subcellular localization of ataxin-3 ( Figure 2B ).
10.1371/journal.pone.0054214.g002
Figure 2
SUMO-1 modification did not affect the subcellular localization of ataxin-3.
HEK293 cells were transfected with plasmids expressing GFP-tagged ataxin-3 or mutant ataxin-3 K166R in the presence of endogenous SUMO-1. Both ataxin-3-20Q and ataxin-3-20Q K166R were localized in the nucleus and cytoplasm uniformly, and the aggregates that formed expressed ataxin-3-68Q and ataxin-3-68Q K166R (A). Immunoblotting analysis of subcellular fractionation of ataxin-3 shows no differences between the various groups (B).
SUMO-1 modification did not affect ataxin-3 ubiquitination or aggregate formation, but partially increased ataxin-3-68Q stability
It has been reported that SUMOylation alters the function or subcellular localization of some proteins, and the competition between SUMO-1 and ubiquitin for identical binding sites protects some proteins from degradation [33] . To determine whether SUMO-1 modification would affect the ubiquitination of ataxin-3, we transiently expressed GFP-ataxin-3 or GFP-ataxin-3 K166R in HEK293 cells and performed immunoprecipitation assays using anti-GFP antibodies. The ubiquitination of ataxin-3 and ataxin-3 K166R was not significantly different, which suggested that SUMO-1 modification did not affect the ubiquitination of ataxin-3, and lysine 166 might not be the ubiquitination site ( Figure 3A, 3B ).
10.1371/journal.pone.0054214.g003
Figure 3
SUMO-1 modification did not affect ataxin-3 ubiquitination.
(A) HEK293 cells were co-transfected with GFP-ataxin-3 and Flag-SUMO-1. The cells were treated with 10 µM MG132 for 12 h and subject to immunoprecipitation analysis using rabbit polyclonal antibodies against GFP. The immunoprecipitants were subject to immunoblotting analysis with the indicated antibodies. (B) HEK293 cells were transfected with GFP-ataxin-3 or GFP-ataxin-3 K166R . The cells were treated with 10 µM MG132 for 12 h and subject to immunoprecipitation analysis using rabbit polyclonal antibodies against GFP. The immunoprecipitants were subject to immunoblotting analysis with the indicated antibodies.
Since SUMO modification may regulate the stability of proteins [33] – [34] , we speculated that SUMO-1 modification might alter the stability of ataxin-3. The levels of sumoylated and un-sumoylated proteins were examined in cells transfected with ataxin-3 or ataxin-3 K166R . Firstly, we detected the soluble and insoluble fractions of cell lysate by western blot separately. The results showed that the bands of insoluble fraction of mutant-type ataxin-3 were stronger than that of the wild-type, which suggested that stabilized mutant ataxin-3 led to aggregate formation and induced the disease of SCA3/MJD. In addition, both bands of soluble and insoluble fraction of ataxin-3-68Q were denser than those of ataxin-3-68Q K166R , indicating SUMOylation might increase the stability of ataxin-3-68Q ( Figure 4A ). Subsequently, we investigated whether the enhanced protein fraction of sumoylated ataxin-3-68Q was related with the increased aggregate formation. To address this possibility, we quantified aggregate formation cells and immunoflurescence density of aggregates by fluorescence imaging and imageJ computational analysis. Unfortunately, there was no significant difference existed between either ataxin-3-20Q and ataxin-3-20Q K166R or ataxin-3-68Q and ataxin-3-68Q K166R (P>0.05) ( Figure 4B, 4C ). Finally, the chase experiment was used to understand the effect of SUMOylation on degradation of ataxin-3. As shown in Figure 4D and Figure S1 , the level of ataxin-3-68Q was significantly higher than that of ataxin-3-68Q K166R especially at 15 h after CHX treatment, while the level of ataxin-3-20Q was similar to that of ataxin-3-20Q K166R . These data suggested that SUMOylation of ataxin-3-68Q might partially enhance the stability of ataxin-3-68Q.
10.1371/journal.pone.0054214.g004
Figure 4
SUMO-1 modification partially increased ataxin-3-68Q stability.
HEK293 cells were transfected with GFP-ataxin-3 or GFP-ataxin-3 K166R . Immunoblotting analysis showed difference between the soluble (S) and insoluble (I) ataxin-3 in 20Q and 68Q with or without K166 (A). At 48 h after transfection, both aggregates formation cells and its immunoflurescence density were quantified. Plasmid groups: 1. GFP-ataxin-3-20Q; 2. GFP-ataxin-3-20Q K166R ; 3. GFP-ataxin-3-68Q; 4. GFP-ataxin-3-68Q K166R . Statistical significance was assessed with a one-way ANOVA. The amount of aggregates formation cells: 1 and 3: P<0.05 (*); 1 and 2: P>0.05 (**); 3 and 4: P>0.05 (***) (B). Immunoflurescence density of aggregates: 1 and 3: P<0.05 (*); 1 and 2: P>0.05 (**); 3 and 4: P>0.05 (***) (C). At 24 h after transfection, cells were treated with CHX (100 µg/ml) to prevent protein synthesis. Cells were harvested at 0, 1, 3, 7, 15 h after CHX treatment, subject to 12% SDS-PAGE, and analyzed by immunoblotting with anti-GFP antibody (D).
SUMO-1 modification increased cytotoxicity of ataxin-3-68Q
We examined the cytotoxicity effects by flow cytometry analysis using PI/Annexin V-FITC staining in HEK293 cells transfected with myc-ataxin-3 or myc-ataxin-3 K166R . Relatively high percentages of early apoptosis rate were found in ataxin-3-68Q transfected cells compared to that of ataxin-3-68Q K166R transfected ones (P<0.05), suggesting SUMO-1 modification might had a cytotoxic effect. However, there was no significant difference between the early apoptosis rates of ataxin-3-20Q and that of ataxin-3-20Q K166R group (P>0.05) ( Figure 5 ).
10.1371/journal.pone.0054214.g005
Figure 5
Early apoptosis rate in HEK293 cells.
Plasmid Groups: 1. pcDNA3.1-myc-His(-)B; 2. pcDNA3.1-myc-His(-)B-ataxin-3-20Q; 3. pcDNA3.1-myc-His(-)B-ataxin-3-20Q K166R ; 4. pcDNA3.1-myc-His(-)B-ataxin-3-68Q; 5. pcDNA3.1-myc-His(-)B-ataxin-3-68Q K166R . Statistical significance was assessed with a one-way ANOVA: 2 and 4: P<0.05 (*); 2 and 3 P>0.05 (**); 4 and 5: P<0.05 (***).
Discussion
Recent studies have revealed that some neurodegenerative disease proteins, such as androgen receptor (AR) [25] , huntingtin [26] , ataxin-1 [27] , ataxin-7 [28] , DJ-1 [35] , tau [36] , and a-synuclein [36] , are also modified by SUMOs, implying that SUMOylation of these disease-related proteins may participate in the regulation of their functions and thereby be associated with their pathogenic role. Tau, an Alzheimer's disease associated protein, has been reported to be SUMOylated by SUMO-1, and to a much lesser extent by SUMO-2 or SUMO-3 simultaneously [36] . Many substrates were reported to be multi-SUMOylated by SUMO-1 on several residues, for example, ataxin-1 modified at least on five and huntingtin on three lysine residues [26] – [27] . SCA3/MJD is the most common spinocerebellar ataxia diseases. In our previous research, we found that ataxin-3 was also a substrate of SUMO-1 [32] . In order to identify the motif residue, mutagenesis analyses were carried out to converse lysine 166 residues to arginine, which lies within a SUMO consensus sequence, VKGD, in ataxin-3. This conversion completely blocked the SUMOylation of ataxin-3. However, the conversion of other lysines, K8 and K206, which also lie within the SUMO consensus sequence in ataxin-3, did not affect SUMOylation of ataxin-3. These data suggest that K166 in ataxin-3 is the major SUMOylation binding site.
Modification by SUMO has been shown to play critical roles in subcellular localization, and protein degradation, which ultimately contribute to regulation of the cell cycle, cell growth, and apoptosis [37] . In order to examine whether SUMOylation of ataxin-3 affects its subcellular localization, we compared the localization of ataxin-3 in transiently transfected HEK293 cells. In agreement with previous studies, we found that the wild-type ataxin-3 protein was diffusively distributed in both nucleus and cytoplasm, while mutant-type ataxin-3 protein formed aggregates in nucleus. However, when we compared ataxin-3 and its SUMOylation deficient variant, we could not detect any difference in the subcellular localization of ataxin-3 in both immunofluorescent staining and immunoblot analysis, which indicates SUMOylation of ataxin-3 does not change its subcellular distribution. The similar result was also observed in SCA7, that SUMOylation on K257 of ataxin-7 does not influence its subcellular localization [28] .
As we know, abnormal accumulation of mutant ataxin-3 in affected neurons reflexes that mutant protein may not be properly degraded. We found the insoluble fraction of ataxin-3-68Q was more than that of ataxin-3-20Q, which supported that mutant-type ataxin-3 protein was stable and easy to form aggregates. As SUMO modification of proteins is involved in protein degradation, it is possible that sumoylation of ataxin-3 may regulate its degradation process. Since SUMO-1 modifications target the same lysine residue as ubiquitin, many researches have revealed a dynamic interplay between the related ubiquitination and SUMOylation pathways [38] . We first performed immunoprecipitation assays to detect the ubiquitination differences between ataxin-3 and ataxin-3 K166R . However, we didn't find any evidence that SUMOylation of ataxin-3 affect ataxin-3 ubiquitination, which also indicate there is no competition between SUMO-1 and ubiquitin for binding site K166.
Subsequently, the soluble/insoluble and total protein level of sumoylated and un-sumoylated proteins were also examined, both bands of soluble and insoluble fraction of ataxin-3-68Q were denser than those of ataxin-3-68Q K166R indicating the SUMOylation modification of mutant-type ataxin-3 might enhance the stability of the protein and participate in the pathogenesis process of SCA3/MJD to a certain degree. In addition, we further confirmed SUMO-1 modification decreased the degradation and enhanced the stability of mutant-type ataxin-3 by chase assay. Therefore, we have no reason to doubt that although SUMO-1 modification on K166 does not influence the UPS pathway but probably affect other processes such as autophagy for mutant-type ataxin-3 degradation. Increased polyQ-expanded ataxin-3 stability might leads to multiple consequences. On the one hand, polyQ-expanded ataxin-3 is more easily gathered to form aggregates. On the other hand, the concentration of the monomer or oligomer of polyQ-expanded ataxin-3 might increases as huntingtin (26), leading to increased cytotoxicity, promotion of apoptosis, and acceleration of the pathological process in SCA3/MJD pathogenicity.
PolyQ disorders are characterized pathologically by the accumulation of protein aggregates within neurons. Whether the microscopically visible inclusions play a causal role in disease pathogenesis or protect neurons from the affects of toxic proteins remains unclear [26] , [39] . Therefore, as a central pathological event in polyQ disorders, aggregation needs to be better understood, particularly from a therapeutic point of view. In agreement with previous studies [40] , we found the amount of aggregate formation cells in mutant-type ataxin-3 as much higher than that in normal control; demonstrating polyQ expansion could induce the formation of aggregates. Although there was no significantly difference in both aggregate cell counting and density quantification between ataxin-3-68Q and ataxin-3-68Q K166R , we could found the tendency that aggregate density of ataxin-3-68Q was slightly higher than that of ataxin-3-68Q K166R , which support the results of insoluble fraction detection and indicate that SUMOylation of mutant-type ataxin-3 might partially increase its stability and probably promote aggregate formation.
It has been reported that protein aggregates could sequester polyQ proteins which affects their normal biological function [39] and finally result in polyQ diseases. SUMOylation of the polyQ proteins might influences their aggregation and toxicity. For example, SUMOylation of the polyQ-expanded AR decreases the amount of the SDS-insoluble aggregates [41] , and study on huntingtin proposed that SUMOylation may explain the intriguing cell death observed in polyQ disorders [42] . As what we show in Figure 5 , SUMO-1 modification of mutant-type ataxin-3 increased the early apoptosis rate of the neurons, indicating that SUMOylation might enhance the stability of mutant-type ataxin-3, thus increase its cytotoxicity, however the concrete mechanism still needs intensive study in future.
In conclusion, our study demonstrated that SUMOylation on K166, the first described residue of SUMO-1 modification of ataxin-3, partially increased the stability of mutant-type ataxin-3, and the rate of apoptosis arisen from the cytotoxicity of the modified protein. Those support the hypothesis that SUMO-1 modification has a toxic effect on mutant-type ataxin-3 and participates in the pathogenesis of SCA3/MJD. Further studies in Drosophila models should be done to confirm these findings.
Materials and Methods
Plasmid construction
Plasmids for myc-ataxin-3 and SUMO-1 in pcDNA3.1-myc-His(-)B (Invitrogen) have been described previously [32] . Ataxin-3 K8R , ataxin-3 K166R , and ataxin-3 K206R were all generated by site-directed mutagenesis using long primers and overlap methods with primers M1/M2, M3/M4, M5/M6, respectively. GFP-ataxin-3 and GFP-ataxin-3 K166R were constructed by subcloning the PCR product amplified using primers M1/M2 with pcDNA3.1-myc-His(-) B-ataxin-3 into pEGFP-N1 (Invitrogen) at Sal I/ Bam HI sites respectively. The p3×FLAG-myc-CMV-24-SUMO-1 plasmid was kindly provided by Professor Wang Guanghui. All constructs were confirmed by sequencing. Primers used in this study are shown in Table 1 .
10.1371/journal.pone.0054214.t001
Table 1
Primers for amplification.
Primers *
Sequence
W1
5′-ACGGGATCCGCCACCATGGAGTCCATCTTCCACG-3′
W2
5′-CCCAAGCTTGGGCATGTCAGATAAAGTGTGAAGG-3′
M1
5′-ACGGGATCCGCCACCATGGAGTCCA-3′
M2
5′- ATCTTCCACGAGAGACAAGGTACG-3′
M3
5′-TTTGTTGTTAGAGGTGATCTGCCAG-3′
M4
5′-CAGATCACCTCTAACAACAAATATAG-3′
M5
5′-AGAGTCCATAGAACAGACCTGGAACG-3′
M6
5′-AGGTCTGTTCTATGGACTCTTTGCTC-3′
*
Primers used are described in Experimental Procedures.
Cell culture and transfection
HEK293 cells were cultured overnight in Dulbecco's modified Eagle's medium (DMEM) (Gibco) supplemented with 10% fetal bovine serum (FBS) (Gibco) and antibiotics penicillin/streptomycin at 37°C under 5% CO 2 , and then transfected with expressing plasmids using Lipofectamine™ 2000 reagent (Invitrogen) according to the manufacturer's protocol in DMEM without FBS. The same volume of DMEM containing 10% FBS was added to the culture medium 6 h after transfection. Forty eight hours after transfection, the transfected cells were observed using an inverted system microscope IX71 (Olympus) or used for immunofluorescent staining, immunoblot analysis, or co-immunoprecipitation.
Preparation of cell extracts and NTA precipitation
Thirty hours after transfection, cells were lysed in 1 ml of lysis buffer (6M guanidine hydrochloride, 100 mM NaH 2 PO 4 , and 10 mM Tris [pH 7.8]). After sonication, 90% lysate was incubated with 25 µl of Ni–nitrilotriacetic acid (NTA) magnetic agarose beads (Qiagen). The beads were washed twice with washing buffer (pH 7.8) containing 8 M urea, followed by washing with a buffer (pH 6.3) containing 8 M urea. After a final wash with phosphate-buffered saline (PBS), the beads were eluted with 2×SDS sample buffer for immunoblot analysis. Then 10% lysate was subjected to trichloroacetic acid (TCA) precipitation and used as a whole cell extract (WCE). The proteins were analyzed by Western blotting using the appropriate antibodies as described recently [33] .
Fluorescence
HEK293 cells were plated onto cover slips in a 12-well plate. The following day they were transfected using Lipofect2000™ (Invitrogen). Forty-eight hours after transfection, they were incubated 10 µg/ml Hoechst 33258 (Sigma) to visualize the nucleus for 5 min at 37°C. Analysis was performed using an inverted system microscope IX71 (Olympus).
Subcellular fractionation
HEK293 cells transfected with expression plasmids were fractionated into cytoplasmic and nuclear fractions 24 h after transfection. After being washed twice with pre-cold PBS, cells were lysed in fractionation buffer containing 10 mM Tris-HCl (pH 7.5), 1 mM EDTA, 0.5% NP-40 and complete mini protease inhibitor cocktail, for 30 min at 4°C. Following centrifugation at 600×g for 10 min at 4°C, the supernatant was collected as the cytoplasmic fraction. The pellets, resuspended with pellet buffer containing 2% SDS, as the nuclear fraction.
Immunoprecipitation
HEK293 cells were collected 48 h after transfection. The cells were sonicated in TSPI buffer (50 mM Tris-HCl [pH 7.5], 150 mM sodium chloride, 1 mM EDTA, 1 µg/ml of aprotinin, 10 µg/ml of leupeptin, 0.5 µM Pefabloc SC, and 10 µg/ml of pepstain) containing 1% NP-40. Cellular debris was removed by centrifugation at 12,000×g for 15 min at 4°C. The supernatants were incubated with the antibodies in 0.01% BSA for 4 h at 4°C. After incubation, protein G Sepharose (Roche) was used for precipitation. The beads were washed with TSPI buffer four times, and then bound immunoprecipitants were eluted with 2×SDS sample buffer for immunoblot analysis.
RIPA-soluble and RIPA-insoluble fraction
For serial extraction in RIPA and formic acid, cells were washed twice in PBS and then lysed in 600 µl RIPA buffer and centrifuged for 20 min at 40,000 g at 4°C. Supernatant was collected as the soluble protein for Western blot, and the pellet was resuspended in 100 µl 70% formic acid with sonication until clear. Formic acid samples were then neutralized by adding 1.9 ml 1 M Tris base and diluted 1∶3 in H 2 O as the insoluble protein for Western blot.
Immunoblot analysis
Proteins were separated by 10%/12% SDS–PAGE and then transferred onto polyvinylidene difluoride membrane (Millipore). The following primary antibodies were used: monoclonal anti-c-myc antibody (Sigma), monoclonal anti-SUMO-1 antibody (Zymed), monoclonal anti-GFP antibody (Santa Cruz), monoclonal anti-ubiquitin antibody (Santa Cruz), monoclonal anti-GAPDH (Chemicon). Sheep anti-mouse IgG-HRP antibody was used as the secondary antibody. The proteins were visualized using an ECL detection kit (Amersham Pharmacia Biotech).
Quantitation of aggregates and immunoflurescence density of aggregates
HEK293 cells were transfected with plasmids expressing GFP-tagged ataxin-3 or mutant ataxin-3 K166R . Forty-eight hours after transfection, the transfected cells were observed using an inverted system microscope IX71 (Olympus). All cells were counted in fields selected at random from five different quadrants of the culture well. Counting was performed by an investigator blind to the experimental conditions. By ImageJ software, the immunoflurescence density of aggregates (mean) was measured. The assay was carried out in triplicate.
Degradation assay
HEK293 cells were transiently transfected with GFP-ataxin-3 and GFP-ataxin-3 K166R . Twenty four hours after transfection, cells were treated with 100 µg/ml cycloheximide (CHX) (Sigma) to inhibit protein synthesis and were harvested at 0, 1, 3, 7, 15 h or 0, 15, 15 h after CHX treatment. The same volume of lysates was analyzed by immunoblotting using anti-GFP antibody (Santa Cruz).
PI/Annexin V-FITC double-stained flow cytometry
HEK293 cells were transfected with expressing vectors. Forty eight hours after transfection, apoptosis were assessed by using FACSCalibur flow cytometer (Becton Dickinson) according to the manufacturer's instructions. Briefly, cells were collected by centrifugation in a refrigerated microcentrifuge and resuspended with 200 µl binding buffer. After the addition of 10 µl annexin V-fluorescein isothiocyanate (FITC) and 5 µl propidium iodide (PI), cells were kept in dark at room temperature for 15 min. Then the cells were added to another 300 µl binding buffer and assessed within 1 h. Data analysis was performed with CellQuest software.
Supporting Information
Figure S1
SUMO-1 modification increases ataxin-3-68Q stability. HEK293 cells were transfected with GFP-ataxin-3 or GFP-ataxin-3 K166R . At 24 h after transfection, cells were treated with CHX (100 µg/ml) to prevent protein synthesis. Cells were harvested at 0, 15 h after CHX treatment, subject to 12% SDS-PAGE, and analyzed by immunoblotting with anti-GFP antibody.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
In Table 3, the column detailing gender was missing. Please view the corrected Table 3 here:
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Tuberculosis (TB) is a severe, often chronic, lung disease causing nearly nine million illnesses and more than one million deaths each year. [1] With appropriate treatment, approximately 90% of patients with active TB disease can be cured, even in patients with HIV infection. [2] , [3] Despite the discovery of the first TB drug over 50 years ago, current treatment regimens for susceptible TB still require the use of a combination of potentially toxic antibiotics for a minimum of six months to ensure eradication. [4] , [5] For multi-drug resistant TB (MDR-TB), available regimens are less potent (but more noxious), requiring longer treatment durations. [6] , [7] , [8] As such, patients' health-related quality of life (HRQL), i.e., valued aspects of life, may be diminished by side effects from medication, prolonged treatment duration, and in some cultures, social stigma attached to the disease. [9] , [10] , [11] TB disease itself may also have a negative impact on TB patients' self perceived health status. [12] , [13]
Health systems in Thailand are increasingly cost effective. In an effort to respond better to patients' needs, healthcare providers are integrating services. [14] , [15] , [16] Economic and decision analyses are frequently conducted to inform resource allocation decisions. HRQL data are widely recognized as an important input in such exercises, particularly for chronic diseases. [17] HRQL can be derived using generic and specific instruments. [9] Generic instruments collect descriptive data and generate health utilities: preference-based, numeric representations of overall health that are the most commonly used measures for evaluating HRQL in economic analyses. [18] While health utilities, as input in cost-utility analyses, allow comparison between populations and across diseases, settings, and countries, information generated from specific instruments focus on problems associated with single disease states, patient groups, or areas of function and do not allow broad comparison. [9] , [18] , [19]
Our systematic review showed that data on formal assessment of HRQL in TB patients are rather sparse, particularly in the Thai setting. To date, there were only two studies conducted in Thai populations. [20] , [21] The first study administered a generic instrument (Medical Outcomes Study 36-Item Short-From Health Survey) to measure HRQL among 84 pulmonary TB patients in Yunnan province of China and Southern Thailand. [20] Findings of this study, however, were published in Chinese, compromising dissemination of the results to non-Chinese speaking researchers. In 2008, Kittikraisak et al. conducted a prospective observational study to evaluate the impact of TB and HIV treatment on HRQL among 849 TB patients in Thailand. This study, however, focused on the patients who were HIV-infected and used a study specific questionnaire to collect HRQL data. [21] The investigators found impairment in physical and mental health when Thai HIV-infected patients studied were first diagnosed with TB. Additionally, completing TB treatment relieved some physical symptoms, but had little impact on mental health. Data, however, cannot be used for economic modeling purposes because neither generic nor specific instruments were employed to collect data. With increasing interest in identifying cost-effective interventions that are responsive to patients' needs, HRQL data collected using standardized instruments are urgently needed.
The main purpose of this study was to collect health utility data, using EuroQol 5D (EQ-5D) and EuroQol visual analogue scale (EQ-VAS) instruments from Thai TB patients and those cured or having completed treatment. The data were collected for use in our economic evaluation analysis of screening and diagnostic algorithms for pulmonary TB among HIV-infected patients in Thailand (to be published). In this study, we explored how socio-demographic characteristics and co-morbidity such as HIV infection affect TB patients' health utility and whether health utilities of patients with different medical conditions were different. Further, we examined concordance of health utilities measured using the two instruments. Lastly, we examined how health utility of patients with two morbidities calculated using multiplicative approach (U CAL ) differed from the measured utilities.
Methods
Ethics statement
This study was approved by ethical review committees of Chiang Rai Regional Hospital and Bamrasnaradura Infectious Diseases Institute. Involvement of Centers for Disease Control and Prevention (CDC) investigators in this study was determined not to meet the definition of engagement in human subjects research per U.S. human subjects research regulations and additional review by the CDC institutional review board was not required. All participants had provided written informed consent.
Study setting and population
From August to October 2009, we conducted a cross-sectional survey and recruited consecutive patients from respective clinics at Chiang Rai Regional Hospital and Bamrasnaradura Infectious Diseases Institute. These two hospitals were part of our multi-site population-based TB surveillance conducted in six provinces in Thailand and were chosen because they serve a high number of TB, HIV-infected TB, and HIV patients. [22] Eligible patients were those diagnosed with TB (including MDR-TB) and/or HIV ≥2 weeks before study enrollment to allow time for them to cope with the diagnosis(es), aged between 18–70 years old, who were able to communicate in Thai and did not require assistance from family members regarding communication, were not pregnant, and were not in the priesthood. Patients with HIV were recruited regardless of their anti-retroviral therapy (ART) status. After giving written informed consent, patients were enrolled and assigned into mutually exclusive groups according to their medical condition: 1) TB patients receiving TB treatment (TB TX ), 2) MDR-TB patients receiving MDR-TB treatment (MDR TX ), 3) patients who had been successfully treated for TB or MDR-TB according to World Health Organization definition and finished treatment for ≥6 months ( any TB C ), 4) HIV-infected patients at any stage who had not been diagnosed with TB ( any HIV), 5) HIV-infected TB patients receiving TB treatment (TB TX /HIV), and 6) HIV-infected patients who had been successfully treated for TB or MDR-TB and finished treatment for ≥6 months ( any TB C /HIV). Sample size that gave 80% power at the 0.05 level of significance (two-sided) was calculated as specified in the study protocol, with parameters estimated based on literature. An effect size of 7% for the difference between utilities of patients in different medical conditions was used in the calculation. Sample size of 32 patients per group was required for patients with TB only (groups 1–3). For HIV-infected patients regardless of TB status (groups 4–6), the required sample size was 49 per group. For groups 1, 2, and 5, we restricted enrollment to patients who were on TB or MDR-TB treatment for ≥2 weeks at enrollment, a long enough duration to experience side effect(s) from medications (if any). Each patient was compensated with 100 Thai Baht (∼3US$) for his/her time, except those in groups 3 and 6 who received additional 400 Baht for travel expenses because they were no longer visiting the hospitals for medical care.
Data collection and instruments
At enrollment, trained study nurses administered: 1) structured questionnaire to collect socio-demographic characteristic data, 2) EQ-5D, and 3) EQ-VAS instruments (with permission to use from the developer [the EuroQol Group]) to collect patients' HRQL data. The two instruments are recommended by the Thai Health Technology Assessment Guidelines to be used to value health utility for economic purposes. [23] EQ-5D consists of five domains relating to: 1) mobility, 2) self-care, 3) usual activities, 4) pain/discomfort, and 5) anxiety/depression and was originally used as a self-administered questionnaire. [24] According to the developer, it can also be administered and used in postal surveys, in clinics, and face-to-face interviews. We conducted face to face interview and for each domain staff asked patients to assess their current health and respond with one of the following options: “no problem”, “moderate problem”, and “severe problem”. A color coded (green, yellow, and red) flip chart was used as a supplemental tool to enhance patients' understanding of the context. An EQ-5D health state of each respondent was recorded. For example, the “32211” health state implied the patient perceived a severe problem with mobility, a moderate problem with self-care and usual activities, but no pain/discomfort or anxiety/depression. Following EQ-5D administration, we asked patients to indicate their health status using EQ-VAS. The EQ-VAS is a standard visual analogue scale including a vertical line, 20 cm in length, anchored at “0” (death) at the bottom and “100” (full health) on top. We asked each patient to mark on the scale a rating of their health and well-being on the interview day. The instrument measures an individual's valuation of their current overall health status.
Statistical analysis
The analysis was divided into four parts. First, we described socio-demographic and health characteristics. Second, we calculated EQ-5D utility value (U EQ-5D ) by assuming that health utilities were additive and that the health utility of a person declined when his/her health deteriorated. [25] We transformed U EQ-5D from the five-digit coded health states using an additive formula, including coefficients and a constant derived from the Thai utility score algorithm established from a recent national household survey in Thai general population. [26] Theoretically, U EQ-5D ranges from 0.0 (death) to 1.0 (full health) value scale; scores less than zero representing states worse than death are possible. [27] We obtained EQ-VAS utilities (U VAS ) by transforming them from a 0 to 100 directly to a 0 to 1 scale; scores lower than zero are not possible. Concordance between U EQ-5D and U VAS was estimated using “concord,” a user-written program for Stata by Steichen and Cox. [28] , [29] Bland-Altman approach was used to examine agreement between the two scales used to measure utilities. [30]
Third, we fitted tobit regression models to examine associations between socio-demographic characteristics and U EQ-5D and U VAS . We assessed whether there was any difference in health utility scores of patients in different medical conditions. Tobit regression models are designed to estimate linear relationships between variables when there is upper- or lower-censoring in the dependent variable. [31] We used an upper-censoring limit of one for analyses of U EQ-5D and U VAS (U EQ-5D and U VAS cannot exceed one), and a lower limit of zero for analysis of U VAS (U VAS cannot be lower than zero). The p-value of likelihood ratio chi-square was used as a guide to the model's goodness of fit. We used a two sided p-value of ≤0.05 to indicate statistical significance. However, to reduce the chances of a type I error, we employed a Bonferroni's adjustment when determining which groups of patients were different. [32] In this case, a p-value of ≤0.003 (0.05/15 comparisons) was considered significant. Lastly, we calculated health utilities of patients with two co-morbidities (e.g., TB and HIV) using multiplicative formula shown below and compared them with the actual data measured using EQ-5D and EQ-VAS. [33] All statistical analyses were conducted using Stata, version 10 (StataCorp LP, College Station, TX, USA). Where health utility ( U ) = 1−disability weight ( DW ) DW, disability weight; U, Health utility; subscript 1 stands for a more severe condition and subscript 2 stands for a milder condition.
Results
Socio-demographic characteristics of the population
During the enrollment period, 223 patients with TB and/or HIV were enrolled in the study. Of these, 222 were analyzed. We excluded MDR TX /HIV from the analysis because there was only one patient in this group. The analytic dataset included 32 TB TX , 11 MDR TX , 32 any TB C , 49 any HIV, 49 TB TX /HIV, and 49 any TB C /HIV. Of the 222 patients, 138 (62%) were male, 128 (58%) were married/cohabitating, and 172 (77%) finished either primary or high school. [ Table S1 ] The median age at enrollment was 40 years (interquartile range [IQR], 35–47), 79 patients (36%) were laborers, and 203 (91%) were covered by health insurance. The median monthly household income was 6,000 Baht (IQR, 4,000–15,000). Among HIV-infected patients, 46 any HIV (94%), 26 TB TX /HIV (55%), and 38 any TB C /HIV (100%) reported currently receiving ART. These patients were diagnosed with and on treatment for HIV for a median of 36 months (IQR, 18–95) and 24 months (IQR, 2–51), respectively. TB patients were diagnosed with TB for a median of 3 months (IQR, 1–5). They were initiated treatment at the time of diagnosis, resulting in a median treatment duration of 3 months (IQR, 1–5). MDR-TB patients were diagnosed with the disease for a median of 9 months (IQR, 4–11); effective treatment was initiated relatively quickly after diagnosis. Table 1 shows response of 222 patients to EQ-5D instrument stratified by medical conditions.
10.1371/journal.pone.0029775.t001 Table 1
Response of 222 Thai patients with various medical conditions to EuroQol 5D instrument, August to October 2009.
All (n = 222)
TB TX (n = 32)
MDR TX (n = 11)
any TB C (n = 32)
any HIV (n = 49)
TB TX /HIV (n = 49)
any TB C /HIV (n = 49)
n (%)
n (%)
n (%)
n (%)
n (%)
n (%)
n (%)
Mobility
No problem
172 (77)
21 (66)
5 (45)
28 (88)
41 (84)
32 (65)
45 (92)
Moderate problem
50 (23)
11 (34)
6 (55)
4 (12)
8 (16)
17 (35)
4 (8)
Severe problem
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
Self care
No problem
210 (95)
26 (81)
9 (82)
32 (100)
48 (98)
47 (96)
48 (98)
Moderate problem
12 (5)
6 (19)
2 (18)
0 (0)
1 (2)
2 (4)
1 (2)
Severe problem
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
Usual activities
No problem
161 (73)
22 (69)
3 (27)
29 (91)
40 (82)
26 (53)
41 (84)
Moderate problem
57 (26)
9 (28)
6 (55)
3 (9)
9 (18)
22 (45)
8 (16)
Severe problem
4 (1.8)
1 (3)
2 (18)
0 (0)
0 (0)
1 (2)
0 (0)
Pain/discomfort
No problem
95 (43)
10 (31)
0 (0)
18 (56)
23 (47)
12 (24)
32 (65)
Moderate problem
122 (55)
21 (66)
9 (82)
14 (44)
25 (51)
36 (73)
17 (35)
Severe problem
5 (2)
1 (3)
2 (18)
0 (0)
1 (2)
1 (2)
0 (0)
Anxiety/depression
No problem
125 (56)
15 (47)
3 (27)
25 (78)
27 (55)
23 (47)
32 (65)
Moderate problem
93 (42)
17 (53)
7 (64)
7 (22)
20 (41)
26 (53)
16 (33)
Severe problem
4 (2)
0 (0)
1 (9)
0 (0)
2 (4)
0 (0)
1 (2)
TB TX , TB patients receiving TB treatment; MDR TX , MDR-TB patients receiving MDR-TB treatment; any TB C , patients who had been successfully treated for TB or MDR-TB for ≥6 months; any HIV, HIV-infected patients at any stage; TB TX /HIV, HIV-infected TB patients receiving TB treatment; any TB C /HIV, HIV-infected patients who had been successfully treated for TB or MDR-TB for ≥6 months.
Health utility measured using EQ-5D instrument
U EQ-5D of the 222 patients ranged from −0.02 to 1.0 (median, 0.7; IQR, 0.6–1.0). One of eight MDR TX (9%) perceived his overall health was worse than death (U EQ-5D , −0.02). This patient was diagnosed with and had been taking medication for MDR-TB for 26 months. By contrast, 7 TB TX (22%), 16 any TB C (50%), 17 any HIV (35%), 6 TB TX /HIV (12%), and 27 any TB C /HIV (55%) perceived they were in full health (U EQ-5D , 1.0). Table 2 shows median U EQ-5D of the 222 patients stratified by medical conditions. Overall, the median U EQ-5D was highest among any TB C (1.0; IQR, 0.7–1.0) and lowest among MDR TX (0.5; IQR, 0.4–0.7).
10.1371/journal.pone.0029775.t002 Table 2
Health utilities of 222 Thai patients with various medical conditions measured using EuroQol 5D and EuroQol visual analogue score instruments, August to October 2009.
Patients
TB treatment
% receiving anti-retroviral therapy
N
Health utility by instrument
Median EQ-5D (IQR)
SD
Median EQ-VAS (IQR)
SD
TB
On TB treatment
Not applicable
32
0.69 (0.57–0.77)
0.22
0.80 (0.70–0.90)
0.15
MDR-TB
On MDR-TB treatment
Not applicable
11
0.51 (0.39–0.73)
0.21
0.60 (0.40–0.80)
0.25
TB or MDR-TB
Cured or completed treatment
Not applicable
32
0.88 (0.67–1.00)
0.17
0.85 (0.80–1.00)
0.15
HIV
Not applicable
94%
49
0.73 (0.63–1.00)
0.19
0.80 (0.70–0.90)
0.15
TB with HIV
On TB treatment
55%
49
0.67 (0.57–0.73)
0.16
0.70 (0.60–0.80)
0.16
TB or MDR-TB with HIV
Cured or completed treatment
100%
49
1.00 (0.69–1.00)
0.18
0.80 (0.70–0.90)
0.14
222
0.73 (0.62–1.00)
0.21
0.80 (0.70–0.90)
0.17
TB, tuberculosis; MDR-TB, multi-drug resistant tuberculosis; EQ-5D, EuroQol 5D instrument; EQ-VAS, EuroQol visual analogue scale instrument; IQR, interquartile range; SD, standard deviation.
Health utility measured using EQ-VAS instrument
The U VAS ranged from 0.0 to 1.0 (median, 0.8; IQR, 0.7–0.9). [ Table 2 ] The same MDR TX with U EQ-5D of −0.02 rated his health condition equivalent to death on EQ-VAS. Three TB TX (9%), 10 any TB C (31%), 10 any HIV (20%), 2 TB TX /HIV (4%), and 7 any TB C /HIV (14%) perceived they were in full health. The median U VAS was relatively high in all groups of patients, except MDR TX (0.6; IQR, 0.4–0.8) and TB TX /HIV (0.7; IQR, 0.6–0.8) whose median scores were slightly lower.
Concordance between U EQ-5D and U VAS
Concordance correlation coefficient between U EQ-5D and U VAS was 0.6 (95% confidence interval [CI], 0.5–0.7). Twenty-two patients (10%) rated equivalent health utilities on the two instruments. Eighty-six patients (39%) rated U EQ-5D higher than U VAS ; 114 patients (51%) rated U EQ-5D lower than U VAS . Bland-Altman plot in Figure 1 shows the differences between health utilities measured using EQ-5D and EQ-VAS instruments in relation to their means. Moderate agreement was observed. The 95% limits of agreement were shown at −0.32 and 0.38. The line of the average of the observed differences (average bias) was shown at 0.03.
10.1371/journal.pone.0029775.g001
Figure 1
Bland-Altman plot showing the differences between health utilities measured using EuroQol 5D and EuroQol visual analogue scale instruments in relation to the mean of the two measurements in the 222 Thai patients, August to October 2009.
Predictors of health utility
In tobit regression analysis, factors independently predictive of U EQ-5D included age and monthly household income. [ Table 3 ] Patients aged ≥40 years old rated U EQ-5D significantly lower than those aged <40 years old, adjusting for medical condition and monthly household income (estimate, −0.08; CI, −0.14 to −0.004). We found a dose response effect when examining impact of household income on U EQ-5D . Patients with monthly household income ≥5,000 Thai Baht rated U EQ-5D significantly higher than patients with lower income, adjusting for medical condition and age. [ Table 3 ] In the analysis of U VAS , we did not find any factor predictive of U VAS .
10.1371/journal.pone.0029775.t003 Table 3
Multivariate tobit regression analysis examining determinants of health utility of 222 Thai patients measured using EuroQol 5D instrument, August to October 2009.
Estimates
95% confidence interval
p-value
Lower
Upper
Patient group
TB TX
−0.24
−0.37
−0.10
<0.01
MDR TX
−0.41
−0.58
−0.24
<0.01
any TB C
ref
ref
ref
any HIV
−0.13
−0.26
0.01
0.04
TB TX /HIV
−0.27
−0.39
−0.15
<0.01
any TB C /HIV
−0.01
−0.14
0.13
0.93
Age group (years)
<40
ref
ref
ref
≥40
−0.08
−0.15
−0.003
0.04
Monthly household income (Thai Baht) *
<5,000
ref
ref
ref
5,000–9,999
0.09
0.004
0.17
0.04
10,000–19,999
0.13
0.03
0.23
0.01
≥20,000
0.17
0.06
0.28
<0.01
TB, tuberculosis; MDR-TB, multi-drug resistant tuberculosis; TB TX , TB patients receiving TB treatment; MDR TX , MDR-TB patients receiving MDR-TB treatment; any TB C , patients who had been successfully treated for TB or MDR-TB for ≥6 months; any HIV, HIV-infected patients at any stage; TB TX /HIV, HIV-infected TB patients receiving TB treatment; any TB C /HIV, HIV-infected patients who had been successfully treated for TB or MDR-TB for ≥6 months; ref, referent group.
*32 Thai Baht = 1 US$.
Difference in health utility
We found that U EQ-5D were highest in any TB C , followed by any TB C /HIV, any HIV, TB TX , TB TX /HIV, and MDR TX , adjusting for age and monthly household income. With Bonferroni's adjustment, patients could be divided, according to the fitted utilities, into three non-mutually exclusive groups: 1) any TB C , any TB C /HIV, any HIV; 2) any HIV, TB TX , TB TX /HIV; and 3) TB TX , TB TX /HIV, MDR TX . Figure 2 shows estimated differences in health utilities of patients with various medical conditions, adjusting for age and monthly household income. In the analysis of U VAS , we found that U VAS of patients in different groups ranked in the same order as that seen in the analysis of U EQ-5D . However, because the model fitted poorly we did not further examine if and how each group of patients differed from one another.
10.1371/journal.pone.0029775.g002
Figure 2
Schematic diagram showing estimated difference in health utilities of 222 Thai patients with various medical conditions adjusting for age and monthly household income and measured using EuroQol 5D instrument, August to October 2009.
U CAL versus the actual data
The median health utility of 49 TB TX /HIV was relatively high (0.7 for both U EQ-5D and U VAS ). U CAL for patients with TB and HIV co-infection (0.8) was statistically different from the measured U EQ-5D (p<0.01), and U VAS (p<0.01). Of the 49 TB TX /HIV, 43 (88%) rated U EQ-5D lower than U CAL . Six (12%) rated U EQ-5D higher than U CAL . None of TB TX /HIV rated U EQ-5D equal to U CAL . Likewise, 31 patients (63%) rated U VAS lower than U CAL . Nine (18%) rated U VAS higher than U CAL . Nine patients (18%) rated U VAS equal to U CAL .
Discussion
In this study, we found that the Thai language EQ-5D and EQ-VAS instruments could be used for measuring and evaluating health utility in a selected group of TB and HIV patients. Further, patients' age and monthly household income were found to be determinants of U EQ-5D . TB and MDR-TB treatment may impact health utilities of patients receiving such treatment. This effect diminished after successful treatment of the disease. Health utilities of patients with HIV and TB calculated using multiplicative model for two co-morbidities overestimated the directly measured utilities.
To our knowledge, this is the first study that elicited health utility in HIV-infected TB patients and compared health utilities between HIV-infected and HIV-uninfected TB patients. Our study is also the first study demonstrating feasibility of the Thai language EQ-5D and EQ-VAS instruments in measuring health utility in a Thai TB population regardless of HIV-infection. The English versions of the instruments were recommended for use in all groups of patients and the Thai versions have recently been ratified by the EuroQol Group's Translation Committee. [34] , [35] Both instruments could identify differences in health utilities among patients with different medical conditions. In this study, more than half were elderly or adults with co-morbidity who had finished basic schooling. All successfully completed the study task using the EQ-5D and EQ-VAS. We believe that this would not be possible with the original English language self-administered instruments. It should be noted that assistance from study personnel was only to read questions on instruments to participants who had difficulty reading. Color-coded supplemental tool was used to help participants who had trouble remembering answer choices. Further, a study by Puhan et al has documented that administration formats do not have a meaningful effect on repeated measurements of patient-reported HRQL outcomes. [36] While our study was conducted among a sample of TB patients, the socio-demographic and health characteristics of our population were similar to those of the population-based TB surveillance network in Thailand. This suggests our findings may be generalisable to the wider Thai TB population. [22] , [37]
Consistent with previously studies, higher HRQL, the U EQ-5D in our study, was correlated with younger age and higher household income, likely because of better prognosis. [12] , [38] , [39] , [40] Yet, we did not find significant associations between sex, education, health insurance coverage, and HRQL as found in other studies or any predictor for U VAS . [39] , [41] This may be due to different characteristics of the populations studied. Our population was out-patients receiving services at hospitals and more than 90% were covered by health insurance. In contrast, those in Duyan's study were hospitalized TB patients with low levels of education and no social insurance coverage. [39] Additionally, those in both Duyan's and Nyamathi's studies reported insufficient housing conditions. [39] , [41]
The U EQ-5D and U VAS obtained from this study were in line with other studies which suggested that impaired health utility occurred during TB and MDR-TB treatments. [21] , [38] , [42] Nearly half of our TB patients were still in the intensive phase of TB treatment, making them more prone to disutility. Moreover, 63% of MDR-TB had been on treatment for more than six months with one patient being on treatment for more than two years. This finding together with those from our previous study suggests that provision of a more holistic approach to medical care not limited only to HIV and TB treatment may be beneficial to the patients. [21] Interventions focusing on symptom management and coordination of care may help relieve symptoms and improve patients' ability to tolerate medical treatment as well as help them gain the strength to carry on with daily life. In this study, we also found that improved health utility after TB treatment was more pronounced in HIV-infected patients than those uninfected in nearly all domains. This is likely due to relief of some TB symptoms and adverse events from HIV and TB drug interactions. [21] , [43] Our study did not measure markers of disease progression (e.g., CD4+ T-lymphocyte) among those HIV-infected. Nonetheless, studies have reported that HIV-infected patients with or without AIDS appeared to have similar levels of HRQL in the era of highly active ART. This could likely be explained by the effectiveness of medication in reversing the progression of disease in individuals with AIDS and accommodation to the stress of living with the disease. [44] , [45] In fact, over 80% of HIV-infected patients in our study were receiving ART and were among those whose U EQ-5D were highest. This finding implies that ART delivery in the public sector of Thai healthcare system may have an impact not only on patients' survival as has been found in other studies, but also HRQL and ability to function in society. [46] , [47] , [48]
It is noteworthy that health utilities of persons with two co-morbidities calculated using multiplicative model were overestimated compared to those measured directly using EQ-5D and EQ-VAS. Because co-morbidities are common, this finding warrants further research of how best to estimate utilities of patients with such conditions.
There are a number of limitations in our study. First, enrollment of patients was not done in a random or systematic manner due to operational constraints. As mentioned, socio-demographic and health characteristics of our patients were similar to those in a multi-site population-based TB surveillance system, suggesting interviewed patients may be broadly representative. Second, there was only one MDR TX /HIV enrolled in our study; this patient was subsequently excluded from the analysis because of small sample size. This implies the rarity of this sub-population in Thailand. HRQL in this particular group remains an open question that needs to be addressed by future research in settings where MDR TX /HIV is more prevalent. Further, the required sample size for MDR TX was not met, prompting caution when interpreting data of this particular group of patients. Third, we did not further stratify patients based on sputum smear microscopy results because of the restriction to enrol only patients who had received TB treatment for ≥2 weeks. Some of these patients were expected to have a conversion by the interview time. In India, Dhingra and Rajpal have documented difference in HRQL between smear positive and negative TB patients using a TB-specific instrument. [49] We were unable to investigate if this difference existed in our study. Fourth, screening for active TB among our HIV-infected patients may not have been optimal. It is possible that some patients may have had undiagnosed TB, resulting in misclassification. However, patients in our study were routinely asked if they had coughed along with other symptoms. This information was passed to attending physicians. Therefore, we believe that number of undiagnosed TB should be small. Lastly, as for other HRQL instruments, the EQ-5D and EQ-VAS reflect patients' opinions. Different individuals assign different values to the same health state, and consequently vary in their preferences. Further, as pointed out by Aghakhani et al, the overall responses to the EQ-5D instrument (which has three possible answers) may be forced to the mid-range category because few patients endorse the ‘severe’ value and some limitation is often present. This possibly results in diverting responses away from the ‘no limitation’ option. [50] The EQ-5D has been critiqued as less sensitive than disease-specific measurements resulting in possible overestimation of patients' HRQL. Nonetheless, because one of the study goals was to identify utility values that could be used for economic modelling in the future, the ability to compare across diseases outweighed the sensitivity concerns.
In resource-limited settings, economic analysis is increasingly carried out to inform practice guidelines, funding decisions, and research initiatives. Utility data collected from our study may be incorporated into cost-effectiveness and cost-utility analyses. These in turn allow TB control strategies to be compared more directly with other public health interventions, with respect to both costs and consequences and whether the interventions are of benefit in relation to HRQL. Our findings also suggest that the EQ-5D and EQ-VAS have discriminative power and are responsive to clinically important changes related to TB treatment.
Supporting Information
Table S1
Socio-demographic and health characteristics of 222 Thai patients with various medical conditions, August to October 2009. *
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The major cell surface glycoconjugate of Leishmania is the lipophosphoglycan (LPG), implicated in a wide range of functions, both in vertebrate and invertebrate hosts [ 7 ]. In the invertebrate host, LPG variations are important for Leishmania specificity to the sand fly [ 8 ], where attachment of the parasite to a midgut receptor is a crucial event [ 9 ]. In the vertebrate host, the main functions of this virulence factor during the earlier steps of infection include: protect the parasite from complement-mediated lysis, attachment and entry into macrophages [ 10 ], able to inhibit phagolysosomal fusion [ 11 ], modulation of nitric oxide (NO) production [ 12 ] and inhibition of protein kinase C (PKC) [ 13 ]. Interestingly, although L . major LPG mutants ( lpg1 - ) were highly susceptible to complement mediated lysis, they were able to invade macrophages reinforcing the role of other molecules and the host defenses during the interaction [ 11 ].
Many functions have been attributed to L . amazonensis LPG including induction of neutrophil extracellular traps (NETs) [ 14 ], induction of protein kinase R (PKR) [ 15 ], triggering and killing of the parasite via Leukotriene B4 (LTB4) [ 16 ]. Although L . amazonensis LPG is important in many steps of host infection, its role during the interaction with macrophages and sand flies remains unknown.
LPG structures have been described for several dermotropic and viscerotropic Leishmania [ 17 – 26 ]. LPGs have a conserved glycan core region of Gal(α1,6)Gal(α1,3)Gal f (β1,3)[Glc(α1)-PO 4 ]Man(α1,3)Man(α1,4)-GlcN(α1) linked to a 1- O -alkyl-2- lyso -phosphatidylinositol anchor. The salient feature of LPG is another conserved domain consisting of the Gal(β1,4)Man(α1)-PO 4 backbone of repeat units ( n = ~15–30). The distinguishing feature of LPGs that is responsible for the polymorphisms among Leishmania spp. is variable sugar composition and sequence of branching sugars attached to the repeat units and cap structure [ 27 ]. For example, the LPG of Leishmania major (Friedlin) has β-1,3 galactosyl side-chains, often terminated with arabinose, whereas the LPGs of Leishmania donovani (Mongi) and L . infantum (PP75 and BH46 strains) possess β-glucoses in their repeat units [ 17 , 20 , 24 ]. However, there is no available information on the degree of variability in the LPG structure for L . amazonensis .
The L . major LPG was identified as potent agonist of Toll-like receptor 2 (TLR2) in human natural killer (NK) cells and murine macrophages, triggering the production of TNF-α and IFN-γ through MyD88 [ 28 , 29 ]. Recently, the LPGs of two New World species ( L . infantum and Leishmania braziliensis ) differentially activated TLR2. In this case, L . braziliensis LPG was more pro-inflammatory being able to induce the translocation of NF-κB to the nucleus [ 30 ].
As a part of a wider project on the glycobiology of New World species of Leishmania , we evaluated the role of L . amazonensis LPGs (PH8 and Josefa strains) during the interaction with host cells and the sand fly L . migonei . The present study might help to improve our understanding on the immune modulation mediated by glycoconjugates of L . amazonensis , the etiological agent of diffuse cutaneous leishmaniasis (DCL).
Materials and Methods
Ethics statement
The animals were kept in the Animal Facility of the Centro de Pesquisas René Rachou/FIOCRUZ. All animals were handled in strict accordance with animal practice as defined by Internal Ethics Committee in Animal Experimentation (CEUA) of Fundação Oswaldo Cruz (FIOCRUZ), Belo Horizonte, Minas Gerais (MG), Brazil (Protocol P-82/11-4). This protocol followed the guidelines of CONCEA/MCT, the maximum ethics committee of Brazil. Knockout mice handling protocol was approved by the National Commission of Biosafety (CTNBio) (protocol #01200.006193/2001-16).
Parasites, growth curves, and molecular typing
World Health Organization Reference strains of L . amazonensis (IFLA/BR/1967/PH8 and MHOM/BR/75/Josefa) were used. The PH8 strain was originally isolated from the sand fly L . flaviscutellata from Pará State, Brazil, and the Josefa strain was isolated from a human case from Bahia State, Brazil. Promastigotes were cultured in M199 medium supplemented with 10% fetal bovine serum (FBS), penicillin 100 units/mL, streptomycin 50 μg/mL, 12.5 mM glutamine, 0.1 M adenine, 0.0005% hemin, and 40 mM Hepes, pH 7.4 at 26°C until late log phase [ 21 ]. Parasites were seeded in triplicate (1 x 10 5 cells/mL), and growth curves of PH8 and Josefa strains were determined daily using a Neubauer improved haemocytometer until cells reached a stationary phase. Both strains exhibited a similar division profile reaching stationary phase after 7 days of culture. For this reason the 6 th day was chosen for harvesting parasites for LPG extraction and molecular typing ( S1A Fig ).
For molecular typing, genomic DNA was extracted from log-phase Leishmania using the phenol/chloroform method (1:1) for amplification of the HSP70 fragment prior to digestion with HaeIII as previously described [ 31 ]. Positive controls included DNA from L . braziliensis (MHOM/BR/75/M2903), L . infantum (MHOM/BR/74/PP75), Leishmania guyanensis (MHOM/BR/75/M4147) and L . amazonensis (IFLA/BR/67/PH8). After PCR-RFLP both L . amazonensis strains were confirmed ( S1B Fig ).
Extraction and purification of LPG
For optimal LPG extraction, late log phase cells were harvested and washed twice with PBS prior to extraction of LPGs ( Fig 1 ). The LPG extraction was performed as described elsewhere with solvent E (H 2 O/ethanol/diethylether/pyridine/NH 4 OH; 15:15:5:1:0.017) after a sequential organic solvent extraction [ 32 ]. For purification, the solvent E extract was dried under N 2 evaporation, resuspended in 2 mL of 0.1 M acetic acid/0.1 M NaCl, and applied onto a column with 2 mL of phenyl-Sepharose, equilibrated in the same buffer. The column was washed with 6 mL of 0.1 M acetic acid/0.1 M NaCl, then 1 mL of 0.1 M acetic acid and finally 1 mL of endotoxin free water. The LPGs were eluted with 4 mL of solvent E then dried under N 2 evaporation. LPG concentrations were determined as described elsewhere [ 33 ]. Prior to use on in vitro cells cultures, LPGs were diluted in RPMI. All solutions were prepared in sterile, LPS-free distilled water (Sanobiol, Campinas, Brazil). All extractions and purifications procedures are depicted in Fig 1 .
10.1371/journal.pntd.0004848.g001
Fig 1
Procedures for extraction, purification, preliminary characterization of L . amazonensis LPG, interaction with vertebrate cells and L . migonei .
Late log phase cells were harvested and washed with PBS. For studies with vector, L . migonei midguts were dissected on days 2 and 4 post feeding containing L . amazonensis from each strain. Parasite cell pellets were subject to extraction with organic solvents as described elsewhere. For purification, the solvent E extract was dried under N 2 evaporation and applied into a phenyl-Sepharose column. The purified LPG was used for biological and immunological assays.
Immunoblotting and preliminary characterization of LPGs
Purified LPGs (5 μg) were subjected to dot-blot, blocked (1 h) in 5% milk in PBS and probed for 1 h with monoclonal antibody (mAb) CA7AE (1:1000), that recognizes the unsubstituted Gal(β1,4)Man repeat units [ 34 ]; mAb LT22 (1:1000) that recognizes β-glucose side chains and WIC 79.3 (1:1000) that recognizes β-galactose side chains [ 21 , 35 ]. After three washes in PBS (5 min), the membrane was incubated for 1 h with anti-mouse IgG conjugated with peroxidase (1:5,000) and the reaction was visualized using luminol.
Purification of murine peritoneal macrophages and cell culture
Thioglycollate-elicited macrophages were extracted from C57BL/6 and C57BL/6 knockouts TLR2 (-/-) and TLR4 (-/-) by peritoneal washing with ice cold RPMI and enriched by plastic adherence (1 h, 37°C, 5% CO 2 ). Cells (3 x 10 5 cells/well) were washed with fresh RPMI then culture in RPMI, 2 mM glutamine, 50 U/mL of penicillin and 50 μg/mL streptomycin supplemented with 10% FBS in 96-well culture plates (37°C, 5% CO 2 ). Cells were primed with interferon-gamma (IFN-γ) (3 IU/mL) for 18 h prior to incubation with LPGs from both strains (10 μg/mL), live stationary Leishmania parasites (MOI 10:1) and lipopolysaccharide (LPS: 100 ng/mL) [ 30 , 36 ].
Cytokine and nitrite measurements
For CBA multiplex cytokine detection, cells were plated, primed as describe above and incubated with LPGs and live stationary promastigotes (MOI 10:1) for 48 h. LPS was added as a positive control and medium as negative control. Supernatants were collected and IL-1β, IL-6, IL-10, IL-12p40 and TNF-α were determined using BD CBA Mouse Cytokine assay kits according to the manufacturer’s specifications (BD Biosciences, CA, USA). Flow cytometry measurements were performed on a FACSCalibur flow cytometry (BD Bioscience, Mountain View, CA, USA). Cell-QuestTM software package provided by the manufacturer was used for data acquisition and the FlowJo software 7.6.4 (Tree Star Inc., Ashland, OR, USA) was used for data analysis. A total 1,500 events were acquired for each preparation. Results are representative of six experiments in duplicate. Nitrite concentrations were determinate by Griess reaction (Griess Reagent System, 2009).
MAPKs and NF-κB translocation assay
For MAPKs, peritoneal murine macrophages were obtained as described above. They were applied on 24 wells tissue culture plates (10 6 cells/well) for 18 h prior to assay. The cells were washed with warm RPMI and incubated with LPG from both species for different times (5, 15, 30, 45 and 60 min) or with medium (negative control) or E . coli extracts (100 ng/mL, only 45 minutes) as positive control. p-p38, p-JNK, p-IκBα and total p38 were assayed as previously described [ 25 ]. p-IκBα antibody was provided by Dr. L. P. de Sousa. NF-κB translocation using CHO reporter lines (a kind gift by M. A. Campos) was determined as described elsewhere [ 30 ]. CHO reporter cells were plated (1 x 10 5 cells/well) in 24-well tissue culture dishes and the LPG (0.02 and 0.2 μg/mL) from both strains was added in a total volume of 0.25 mL medium/well. The cells were examined by flow cytometry (BD Biosciences, CA, USA) and the analyses were performed using CellQuestTM software.
Sand fly in vivo infection
Lutzomyia migonei (Baturite strain) sand flies were kept under laboratory conditions and were fed on 30% sucrose solution for 3–4 days prior to experiments. The insects were artificially fed using a chick skin membrane in a glass-feeder device. The chick skin membrane was provided by the Animal Facility of Centro de Pesquisas René Rachou/FIOCRUZ under the Protocol LW 30/10. Heparinized mouse blood (drawn intracardially from Balb/C), with penicillin (100 U/mL) and streptomycin (100 μg/mL) (37°C) containing 2 x 10 7 /mL logarithmic phase promastigotes (PH8 and Josefa strains) offered for 5 h under dark conditions [ 5 ]. Blood engorged flies were separated and maintained at 26°C with 30% sucrose. Engorged sand flies had their midguts dissected on days 2 and 4 post feeding. The midguts were homogenized in 30 μl of PBS and the number of viable promastigotes determined by counting under a Neubauer improved haemocytometer [ 24 ].
Statistical analyses
For nitrite, cytokine measurements and in vivo sand fly experiments, the Shapiro Wilk test was conducted to test the null hypothesis that data were sampled from a Gaussian distribution [ 37 ]. For the non-parametric distribution, it was performed the Mann-Whitney test. Data were analyzed using GraphPad Prism 5.0 software (Graph Prism Inc., San Diego, Ca). P < 0.05 was considered significant.
Results
The LPGs from L . amazonensis strains display intraspecific polymorphism
The purified LPGs from L . amazonensis PH8 and Josefa strains were differentially recognized by the mAbs CA7AE and LT22 ( S2 Fig ). LPG from PH8 strain was recognized by CA7AE and LT22 as well as the positive control represented by L . infantum (BH46). However, a different recognition profile was observed for the Josefa strain since its LPG was weakly recognized by LT22 but not by CA7AE, indicating the presence of side-chains branching-off the repeat units. Because CA7AE recognizes Gal(β1,4)Man unsubstituted repeat units in LPG [ 34 ], these results indicate that at least some of the repeat units are indeed unsubstituted in the LPG of PH8 strain. On the other hand, the presence of side-chains suggestive of glucoses, due to LT22 reactivity, was detected in the LPGs of PH8 and Josefa strains. However, LT22 also recognized the galactose-branched repeat units of L . major (strains FV1 and LV39) indicating cross-reactivity of the antibodies, thus suggesting the presence of either glucose or galactose as side chains ( S2 Fig ). These data suggested an intraspecific polymorphism in the LPGs of L . amazonensis strains.
LPGs from L . amazonensis strains equally activate NO and cytokine production via TLR4
We investigated whether LPGs purified from different strains could have an impact on the parasite’s interaction with host cells, the ability to elicit NO and cytokine production by murine macrophages. LPGs from both strains were incubated with murine peritoneal macrophages from C57BL/6 and respective knockouts for TLR2 (-/-) and TLR4 (-/-). We did not detect any production of the cytokines IL-1β, IL-10 and IL-12 ( S3A–S3C Fig ). Both LPGs and respective parasites were able to activate through TLR4, resulting in NO, TNF-α and IL-6 production ( Fig 2A–2C ) (P < 0.05). As expected, LPS (positive control) activated TLR4 in the TLR2 (-/-) ( Fig 2A–2C ).
10.1371/journal.pntd.0004848.g002
Fig 2
Nitrite (A) and cytokines TNF-α (B) and IL-6 (C) production by IFN-γ primed macrophages stimulated with LPG and live parasites.
Cells were pre-incubated with IFN-γ (3 IU/mL) for the 18 h then 10 μg/mL of LPG, and supernatants used for cytokine and nitrite measurements were collected 48 h latter. Fresh medium alone was used as negative control cells and LPS (100 ng/mL) as a positive control. Nitrite concentration was measured by Griess reaction and cytokine concentrations were determined by flow cytometry. C = negative control; LPG PH8 = L . amazonensis LPG PH8 strain; LPG Jos = L . amazonensis LPG Josefa strain; La PH8 = L . amazonensis PH8 live promastigotes and La Jos = L . amazonensis Josefa live promastigotes. Results represent the mean ± SD of 6 experiments in duplicate, * = P< 0.05 was considered significant.
LPGs from L . amazonensis equally activate MAPKs and the NF-κB inhibitor p-IκBα via TLR4
No difference in MAPKs phosphorylation (p38 and JNK) and p-IκBα was observed after incubation with LPGs from both strains. In peritoneal murine macrophages this activation was mainly via TLR4 ( Fig 3A and 3B ). We also evaluated if the LPGs from these strains were able to translocate NF-κB in CHO cells. No activation of NF-κB was detected in those cells ( Fig 4 ).
10.1371/journal.pntd.0004848.g003
Fig 3
Activation of p38/JNK (A) and p-IκBα (B) in peritoneal murine macrophages (C57BL/6, TLR2 -/- and TLR4 -/-) by L . amazonensis LPGs (PH8 and Josefa).
Macrophages were stimulated for 5, 15, 30, 45 and 60 min with 10 μg/mL of LPG from L . amazonensis PH8 and Josefa strains. Dually phosphorylated MAPKs (p38 and JNK) and p-IκBα were detected by Western blot analysis. C- = negative control; C+ = E . coli extract, positive control (100 ng/mL, 45 min). Total p38 content was used as the normalizing protein.
10.1371/journal.pntd.0004848.g004
Fig 4
LPGs purified of L . amazonensis do not induce translocation of NF-κB through TLRs.
CHO cells expressing TLR2 (TLR2+), TLR4 (TLR4+), or neither (TLR2-/TLR4-) were either untreated (purple line) or treated (green line) with LPGs from both strains of L . amazonensis . Legend: PH8 and Josefa LPGs (0.2 and 0.02 μg), Controls: LPS (TLR4 control) and S . aureus (S.a.) (TLR2 control). CD25 expression was measured by flow cytometry 18 h after stimulation. Results shown as percentage of CD25 expression on stimulated cells minus percentage of CD25 expression on non-stimulated cells.
Leishmania amazonensis strains equally infected the sand fly L . migonei
In vivo midgut infections of the sand flies were determined on days 2 and 4 post feeding, in order to evaluate the number of parasites after the blood meal digestion, as well as, after its excretion on day 3, where non-attached parasites are lost. Although a higher parasite density was detected for PH8 strain on day 2 (P < 0.05), no statistical differences in the parasite densities from both L . amazonensis strains were observed on day 4, and both strains were able to colonize L . migonei midgut (P > 0.05, Fig 5 ).
10.1371/journal.pntd.0004848.g005
Fig 5
Development of L . amazonensis (PH8 and Josefa strains) in Lutzomyia migonei .
Sand flies were infected with promastigotes (2 x 10 7 parasites/mL) of PH8 and Josefa strains. Day 2 (2 nd ) parasites counted before blood excretion; Day 4 (4 th ) parasites remaining after blood excretion. Results are representative of two experiments and * = P < 0.05 was considered significant.
Discussion
Leishmania amazonensis , etiologic agent of the cutaneous and anergic diffuse leishmaniasis, is characterized by disseminated non-ulcerative skin lesions and constantly proportion of negative delayed hypersensitivity skin-test (DTH), resulting in a high resistance of this disease to any type of chemotherapy [ 1 , 38 , 39 ]. In the Old and New World, parasite glycoconjugates have being implicated in a variety of events during parasite-host interactions [ 40 , 41 ]. More recently, the role of LPG and GIPLs in the L . braziliensis and L . infantum was determined, suggesting that two distinct LPGs were able to differentially modulate macrophage functions [ 30 , 41 ]. Regarding L . mexicana complex, from where L . amazonensis is a member, a recently study has demonstrated the inflammatory role of LPG [ 42 ]. This glycoconjugate naturally exposed to the host immune system could contribute to the maintenance of infection by interfering with the assembly immune response, like modulation of cytokine production and non-activation of effectors cells. In the present work, we investigated whether LPGs from two L . amazonensis strains would account for differences in the interaction with macrophages and L . migonei .
LPG polymorphisms are common in the composition of branching sugars attached to the conserved repeat units of its backbone. While in the Old World species, a wide spectrum of sugar composition and structure is commonly observed, in New World species only glucose residues in the side chains of Leishmania were documented to date [ 17 , 21 , 23 , 24 , 43 ]. Our preliminary characterization of the repeat units using specific antibodies suggested the existence of intraspecies polymorphism in L . amazonensis LPGs with differences in the side-chains and in the level of glycosylation. The LPG of PH8 strain strongly reacted with CA7AE, that recognizes the basic backbone of the repeat units is Gal(β)Man-PO 4 [ 21 , 34 ]. However, Josefa LPG did not reacted with this antibody, thus suggesting the existence of sugars as side-chains in the repeat units. This feature is commonly found in the LPG of L . major reference strain FV1, which does not react with CA7AE [ 17 ]. In order to evaluate the quality of the sugars branching-off the repeat units, LT22 and WIC.79.3 antibodies were used to detect the presence of glucose and galactose, respectively [ 21 , 35 ]. Based on L . major LPGs used as controls, they were either recognized by those antibodies, suggesting cross-reactivity. Moreover, those data reinforced the presence of either glucoses or galactoses as side-chains in L . amazonensis LPGs. A fully detailed biochemical analysis must await the results of further investigations.
Understanding variations and the LPG structures are crucial for the comprehension of the mechanisms of how parasites survive under extremely adverse conditions. Although the role of LPG in the interaction with the vertebrate host immune system has been studied, it is still unclear how its polymorphism affects the parasite survival. L . amazonensis LPG induces release of NETs and LTB4 production by neutrophils, thus contributing to diminish parasite burden in the Leishmania inoculation site [ 14 , 16 ]. Additionally, L . mexicana LPG induce TNF-α and IL-10 in monocytes, modulates IL-12 production and diminishes NF-κB nuclear translocation [ 44 ]. Here we show that LPGs from both L . amazonensis strains stimulates NO and cytokine production (TNF-α and IL-6) by peritoneal murine macrophages via TLR4. A similar cytokine production was also observed for other species such as L . braziliensis LPG, another important dermotropic species. However, this activation was primarily via TLR2 [ 30 ]. The NO production by macrophages play a central role in determining intracellular killing of Leishmania [ 45 ] and the intact structure of LPG appears to be important for this activation [ 12 , 29 ]. In many models, NO synthesis is dependent on a combination of IFN-γ and TNF-α via TLR-dependent mechanisms as an important leishmanicidal effector complex to macrophages [ 46 ]. In conclusion, the preliminary variations in the sugar motifs of LPG, did not result in any difference in macrophage activation/signaling thus suggesting the role of conserved motifs such as the lipid anchor [ 29 ].
Previous studies have demonstrated that different macrophage receptors mediate the uptake and phagocytosis of Leishmania . The early recognition of pathogens by cells capable of synthesizing cytokines is crucial for the adequate control of intracellular pathogens. Gene knockout studies in mice have suggested that TLR signaling is essential for the immune response against Leishmania parasites. Moreover, Leishmania LPGs and GIPLs are agonists of TLR2 and TLR4 [ 28 – 30 , 41 , 42 ]. Glyconjugates can modulate the host immune response and their activity seems to be structure dependent. The L . braziliensis LPG exerts a pro-inflammatory interaction with TLR2, inducing the production of NO and cytokines (IL-1β, TNF-α and IL-6). On the other hand, the L . infantum LPG was shown to be immunosuppressive and did not induce NO, cytokines and NF-κB translocation [ 30 ]. Our results indicate that LPG from both L . amazonensis strains induce the production of NO and cytokines in IFN-γ-primed macrophages via TLR4. However in other members of the L . mexicana complex, L . mexicana LPG activates either TLR2 or TLR4 leading to ERK and p38 MAPK phosphorylation and production of cytokines in human macrophages [ 42 ].
Thus, although it has been shown that LPG of Leishmania activates TLRs and that the engagement of these receptors is important for the infection, the complete intracellular processes that are involved in this activation remain unknown. Here we bring some light into the effects of LPG on MAPK and NF-κB signaling, a kinase and transcription factor known for their crucial role in immune defense against pathogens [ 44 , 47 – 49 ]. According to previous reports, infection by L . amazonensis altered phosphorylation of ERK1/2 in response to LPS in murine macrophages [ 50 ] and also activates a transcriptional repressor of the NF-κB [ 48 , 51 ]. Consistent with those observations, here LPGs from both L . amazonensis strains also activated p-IκBα, a NF-κB translocation inhibitor, via TLR4. Since no further NF-κB translocation was detected in the CHO cells, a possible mechanism that has been suggested favors its inhibition by p50/p50 NF-κB homodimer [ 55 ]. Moreover, L . donovani and L . major infection caused inactivation of ERK1/2 and p38, respectively, which was accompanied by the inhibition of transcription factors also modulation of cytokine production [ 52 , 53 ]. In contrast to GIPLs (with fail to activate MAPKs) [ 41 ], our data show that LPG from both L . amazonensis strains is equally activating MAPKs (p38 and JNK) and p-IκBα in peritoneal murine macrophages via TLR4 ( Fig 3 ). On the other hand, these LPGs do not activate the NF-κB translocation. These and our results strongly suggest that Leishmania species have distinct mechanism of modulating the signaling pathways during immunopathological events.
The role of LPG during the interaction with the invertebrate host is a very controversial subject and it has been extensively investigated using in vitro and in vivo models [ 8 , 21 , 24 , 54 , 55 ]. Although the in vitro system has limitations [ 56 ], this model provided important evidence for parasite attachment in the sand fly midgut using many restricted and specific vector as classified elsewhere [ 57 , 58 ]. For example, successful binding to the midgut was reported using the Old World pairs L . major/Phlebotomus papatasi [ 8 , 54 ], L . major/Phlebotomus duboscqi [ 59 ] and L . tropica/Phlebotomus sergenti [ 60 ]. Perhaps, due its similarity to L . major LPG, who also possesses terminal β-galactosyl residues, L . turanica LPG may also be important for development in P . papatasi [ 61 , 62 ]. Moreover, the role of LPG has been questioned in permissive vectors such as Lutzomyia longipalpis and Phlebotomus perniciosus , where LPG mutants of L . mexicana and L . major were able to sustain infection in those vectors [ 63 ]. Recently, an alternative mechanism was suggested that flagellar protein FLAG1/SMP1 has been also implicated as an attachment binding candidate for specific and restricted vectors. In this work, a competitive binding assays using an antibody against FLAG1/SMP1 inhibited interaction using the pair L . major and P . papatasi . However, no effect was observed for permissive L . longipalpis [ 64 ].
The significance of LPG modifications was investigated during in vivo interaction of L . amazonensis with L . migonei . Although L . amazonensis is naturally transmitted by L . flaviscutellata , the absence of a colony led us to use an alternative sand fly, which had been previously shown to successfully harbor this parasite and L . braziliensis [ 5 ]. Since this species, although suspected, is not yet considered a natural proven vector of L . amazonensis , a high parasite doses was artificially offered to the sand flies. In spite of a loss after the 3 rd day, parasite multiplication inside the alimentary tract of the L . migonei was successful for both L . amazonensis strains. To survive, the parasites need avoid a number of barriers including the lethal effects of digestive enzymes in the early blood-fed midgut and the excretion with the digested blood meal [ 5 , 7 , 65 , 66 ]. The strong correlation between the excretion of blood meal and the sudden loss of promastigotes suggests that the inability of Leishmania strains to persist in an inappropriate sand fly is related to their failure to remain anchored to the gut wall via specific attachment sites [ 22 , 67 ]. Nevertheless, L . migonei was able to sustain infection with both of the L . amazonensis strains tested, regardless of the type of LPG. It seems likely that L . migonei together with L . longipalpis might be considered a permissive vector as previously suggested [ 57 , 58 , 68 ]. However, the fully development of those two L . amazonensis strains should be further investigated.
Some studies have determined that polymorphisms in the phosphoglycan domains of LPG might be crucial for Leishmania promastigotes to attach to the midgut and to maintain vector infection after blood meal excretion [ 9 ]. Additional support is based on the altered behavior of LPG deficient L . donovani and L . major mutant promastigotes (lpg-) who showed diminished capacity to maintain infection within the sand fly midgut [ 54 , 69 ]. Furthermore, it was recently presented the occurrence of intraspecies polymorphism in L . infantum LPG. Also, the biological role of the three LPG types (I, II and III) was studied during the interaction with the vector L . longipalpis [ 24 ]. Consistent with our results, all strains could successfully sustain infection in this vector, indicating that LPG polymorphisms did not affect this process. In spite of having a strong evidence for the existence of a midgut receptor for LPG, there is no current information in L . migonei . Indeed, the only known receptor was described for L . major , a galectin receptor found in the midgur of P . papatasi binding to LPG β-galactose residues [ 9 , 70 ]. The existence of midgut glycoproteins bearing terminal N-acetylgalactosamine in sand fly was also suggested as a putative parasite ligand [ 71 ].
Here we describe for the first time the immunomodulary properties of two LPGs isolated from different hosts. Those LPGs were equally able to trigger NO and cytokine (TNF-α and IL-6) production via TLR4. The preliminary differences in carbohydrate structure did not seem to affect the interaction of these strains with macrophages and the sand fly vector.
Supporting Information
S1 Fig
Growth curves of L . amazonensis .
(A) L . amazonensis (PH8 and Josefa strains) were grown in M199 medium and counts determined daily (initial concentration of 1 × 10 5 /mL). (B) Restriction fragment length polymorphisms of 120 bp kDNA amplicons from Leishmania obtained with restriction enzyme Hae III and analyzed on silver-stained 10% polyacrylamide gel. MM: 50 bp molecular size marker; lanes: Lb– L . braziliensis (MHOM/BR/75/M2903), Li– L . infantum (MHOM/BR/74/PP75); La– L . amazonensis reference (IFLA/BR/67/PH8), PH8 – L . amazonensis PH8 (IFLA/BR/67/PH8) and Jos– L . amazonensis Josefa (MHOM/BR/75/Josefa).
(TIF)
S2 Fig
Dot-blots of Leishmania LPGs using different mAb antibodies.
Purified LPGs from L . amazonensis strains (PH8 and Josefa), L . infantum (BH46 strain) and L . major strains (FV1 and LV39) were probed with the mAbs CA7AE (1:1000), LT22 (1:1000) and WIC 79.3 (1:1000). Peroxidase-conjugated anti-mouse IgG (1:5000) was used as secondary antibody. The reaction was developed with luminol.
(TIF)
S3 Fig
Cytokine production by IFN-γ primed macrophages stimulated with LPG and live parasites.
Cells were pre-incubated with IFN-γ (3 IU/mL) for the 18 h then 10 μg/mL of LPG, and supernatants used for cytokine IL-10 (A), IL-1β (B) and IL-12 (C) measurements were collected 48 h latter. Fresh medium alone was used as negative control cells and LPS (100 ng/mL) as a positive control. Cytokine concentrations were determined by flow cytometry. C = negative control; LPG PH8 = L . amazonensis LPG PH8 strain; LPG Jos = L . amazonensis LPG Josefa strain; La PH8 = L . amazonensis PH8 live promastigotes and La Jos = L . amazonensis Josefa live promastigotes. Results represent the mean ± SD of 3 experiments in duplicate, * = P< 0.05 was considered significant.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The coronavirus disease 2019 (COVID-19), caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), was declared a global pandemic by the World Health Organization (WHO) on March 11 th , 2020 [ 1 , 2 ]. This highly contagious unprecedented virus has impacted government and public institutions, strained the health care systems, restricted people in their homes, and caused country-wide lockdowns resulting in a global economic crisis [ 3 – 5 ]. Moreover, as of November 2 nd , 2020, nearly 46 million COVID-19 cases in 213 countries and territories have been reported, including more than 1.2 million deaths [ 6 ]. The social, economic, and psychological impact of this pandemic on much of the world’s population is profound [ 7 – 13 ].
Soon after its initial rapid spread in China, the first case of novel coronavirus beyond China was reported in Thailand on January 13 th , 2020 [ 14 ]. The first case in the USA was not identified until January 20 th , 2020 followed by the detection of the first cases in the European territory on January 24 th , 2020 [ 15 , 16 ]. The COVID-19 pandemic has since spread to every continent. While some countries like New Zealand and Australia have steadily suppressed the COVID-19 spread, reporting less than 150 cases per day as of November 2 nd , 2020, other countries like Brazil, India, and the USA still struggle to contain the increasing number of cases [ 17 ]. Subsequently, considerable COVID-19 outbreaks have occurred in Latin America since late February 2020.
The WHO declared Latin America the new epicenter of the COVID-19 on May 22 nd , 2020 [ 18 ]. Latin America has paid a high toll during the COVID-19 pandemic, with some of the worlds’ highest death rates [ 19 – 21 ]. While home to less than 10% of the world population, Latin America accounts for about one-third of all reported global deaths (~370 thousand) [ 6 ]. Several socioeconomic, demographic, and political factors make control of the pandemic in Latin America particularly challenging [ 22 – 25 ]. Most countries in the region are now facing the stark social and economic costs imposed by large-scale non-pharmaceutical interventions while largely failing to control the epidemic's spread [ 13 , 24 , 26 ]. Despite these unique conditions, the region has received relatively little attention from researchers globally [ 19 ]. As of November 1 st , 2020, the highest number of cases have been reported in Brazil (5,516,658), followed by Argentina (1,157,179), Colombia (1,063,151), Mexico (918,811), Peru (900,180) and Chile (510,256) [ 17 , 27 ]. Adjusted by population, Chile’s COVID-19 outbreak is among the worst globally, with more than 26,000 cases and 980 deaths per million inhabitants [ 28 ].
The first case of SARS-CoV-2 in Chile was identified on March 3 rd , 2020. While the initial cases were imported from southeast Asia and Europe, the COVID-19 case counts have expanded in this country, placing Chile in phase 4 of the pandemic on March 25 th , 2020 [ 28 , 29 ]. Chile was the fifth country in Latin America after Brazil, Mexico, Ecuador and Argentina to report COVID-19 cases. The first six imported cases were reported in Talca and in the capital of Chile, Santiago [ 28 ]. However, since the early phase of the outbreak, Chile has employed an agile public health response by announcing a ban on public health gatherings of more than 500 people on March 13 th , 2020, when the nationwide cumulative case count reached 44 reported cases [ 30 ].
Moreover, the Chilean government announced the closure of all daycares, schools, and universities on March 16 th , 2020. These closures were followed by the announcement to close country borders on March 18 th , 2020, and the declaration of national emergency on the same date, accompanied by several concrete interventions to further contain the outbreak in the region [ 31 ]. In particular, these included a night-time curfew in Chile starting on March 22 nd , 2020, and localized lockdowns (i.e., intermittent lockdowns at the municipality level depending on total cases and case growth) starting on March 28 th , 2020 in two municipalities in Southern Chile and seven municipalities in Santiago [ 32 ]. These initial containment strategies kept the COVID-19 case counts lower than regional peers; Brazil, Peru, and Ecuador until the end of April 2020. However, the government started to ease the COVID-19 restrictions in late April by reopening the economy under the “Safe Return” plan, including the televised opening of some businesses and stores, as new infections had reduced between 350–500 per day by the end of April, implying an only apparent flattening of the COVID-19 curve [ 33 – 35 ]. Moreover, imposing and lifting lockdowns in small geographical areas (municipalities) proved unsuccessful in areas with high interdependencies such as the Greater Santiago [ 36 ]. This strategy resulted in a new wave of infections; with the virus spreading from more affluent areas of Chile to more impoverished, crowded communities, forcing the government to reimpose lockdown measures in Santiago in mid-May ( Fig 1 ) [ 23 , 37 , 38 ]. By mid-July, the government implemented the “step by step” strategy, considering five stages of gradual opening, at the municipality level, based on the periodic monitoring of epidemiological and health system indicators. The case counts continued to increase, averaging ~4943 cases per day in June 2020, and started to decline thereafter. The mid-June peak of infections resulted in intensive care units (ICU) reaching saturation levels of 89% nationally and 95% in the Metropolitan Region [ 39 ]. Thus far, Chile has accumulated 513,188 reported cases including 14,302 deaths as of November 2 nd , 2020. The majority (~58%) of COVID-19 cases are concentrated in Region Metropolitana (mostly in Chile’s capital, Santiago), with 297,423 reported cases, followed by 30,498 cases in Valparaiso located in coastal central Chile, and 30,934 cases in Biobio located in southern Chile [ 40 , 41 ]. Moreover, the crude case fatality rate in Chile (~2.8%) resonates with the global average case fatality rate (~2.6%) [ 17 , 42 ].
10.1371/journal.pntd.0009070.g001
Fig 1
Timeline of the milestones of the COVID-19 pandemic in Chile as of November 2 nd , 2020.
In this study, we estimate the transmission potential of COVID-19, including the effective reproduction number, R , during the early transmission phase of the COVID-19 epidemic in Chile and around the mid of the epidemic, by July 7 th , 2020. We also estimate the instantaneous reproduction number throughout the epidemic in Chile. The reproduction number can guide the magnitude and intensity of control interventions required to combat the COVID-19 outbreak [ 43 , 44 ]. We examine the effectiveness of control interventions in Chile (see Table 1 ) on the transmission rate. To do this, we conduct short-term forecasts using phenomenological growth models [ 45 ] calibrated using the early trajectory of the epidemic and by the mid of the epidemic (as of July 7 th , 2020) to anticipate additional resources required to contain the epidemic. These phenomenological growth models are useful in capturing the epidemic’s empirical patterns, especially when the epidemiological data are limited, and significant uncertainty exists around infectious disease epidemiology [ 46 ]. These models provide a starting point for forecasting the epidemic size and characterizing the temporal changes in the reproduction number during the epidemic [ 47 ].
10.1371/journal.pntd.0009070.t001
Table 1 Timeline of the implementation of the social distancing interventions in Chile as of November 2 nd , 2020.
Date
Control interventions
March 13 th , 2020
Ban on large social gatherings implemented in Chile [ 30 ]
March 16 th , 2020
Closures of daycares, schools, and universities in Chile [ 32 , 48 ] Mandatory quarantine of high-risk individuals returning from Iran, China, West Europe and South Korea [ 32 ]
March 18 th , 2020
Declaration of national emergency [ 29 ] Closure of country borders [ 29 ] Telework implemented
March 19 th , 2020
Closure of mall and department stores with the exception of supermarkets, pharmacies, banks and grocery stores [ 31 ]
March 21 st , 2020
Closure of non-essential business including theatres, restaurant, bars and gyms [ 31 ]
March 22 nd , 2020
Night time curfew implemented [ 32 ]
March 26 th , 2020
Intermittent lockdown initiated (implemented at municipality level) [ 32 ]
April 8 th , 2020
Orders on mandatory use of facemasks in public transport [ 49 ]
April 17 th , 2020
Orders on mandatory use of facemasks in all public spaces [ 31 ]
April 30 th , 2020
First shopping mall is reopened in Santiago and then closed the next day [ 50 ]
May 5 th, 2020
Total lockdown in Antofagasta [ 31 ]
May 15 th , 2020
Total lockdown imposed in all municipalities of Santiago [ 51 ]
June 12 th , 2020
Total lockdown in Valparaiso [ 31 ]
July 19 th ,2020
Step by step gradual reopening of the country [ 31 ]
Methods
COVID-19 incidence and testing data
We obtained updates on the daily series of new COVID-19 cases as of November 2 nd , 2020, from the publicly available data from the GitHub repository created by the Chile’s government [ 27 ]. Incidence case data by the date of reporting per day, confirmed by PCR (polymerase chain reaction) tests from March 3 rd –November 2 nd , 2020, were analyzed. The daily testing and positivity rates available from April 9 th –November 2 nd , 2020, were also analyzed.
Models
We utilize two phenomenological growth models, the generalized growth model (GGM) and the generalized logistic growth model (GLM) that have been validated by deriving short-term forecasts for multiple infectious diseases in the past, including SARS, pandemic Influenza, Ebola, and Dengue [ 52 , 53 ].
Generalized growth model (GGM)
We generate short term forecasts using the generalized growth model (GGM) that characterizes the early ascending phase of the epidemic by estimating two parameters: (1) the intrinsic growth rate, r ; and (2) a dimensionless “deceleration of growth” parameter, p . This model allows to capture a range of epidemic growth profiles by modulating parameter p . The GGM model is given by the following differential equation:
d C ( t ) d t = C ′ ( t ) = r C ( t ) p
In this equation C ′( t ) describes the incidence curve over time t , C ( t ) describes the cumulative number of cases at time t and p ∈[0,1] is a “deceleration of growth” parameter. This equation becomes constant incidence over time if p = 0 and an exponential growth model for cumulative cases if p = 1. Whereas if p is in the range 0< p <1, then the model indicates sub-exponential growth dynamics [ 54 , 55 ].
Generalized logistic growth model (GLM)
The generalized logistic growth model (GLM) is an extension of the simple logistic growth model that captures a range of epidemic growth profiles, including sub-exponential (polynomial) and exponential growth dynamics. GLM characterizes epidemic growth by estimating (i) the intrinsic growth rate, r (ii) a dimensionless “deceleration of growth” parameter, p and (iii) the final epidemic size, k 0 . The final epidemic size is sensitive to small variations in the deceleration of growth parameters [ 56 ] and would vary as the epidemic progresses. The deceleration parameter modulates the epidemic growth patterns, including the sub-exponential growth (0< p <1), constant incidence ( p = 0) and exponential growth dynamics ( p = 1). The GLM model is given by the following differential equation:
d C ( t ) d t = r C p ( t ) ( 1 − C ( t ) k 0 )
Where d C ( t ) d t describes the incidence over time t and the cumulative number of cases at time t is given by C ( t ) [ 45 ]. This simple logistic growth type model typically supports single peak epidemics in the number of new infections followed by a burnt-out period, unless external driving forces such as the seasonal variations in contact patterns exist. This model can underestimate the peak timing and the duration of outbreaks. This model can also underestimate the case incidence before the inflection point has occurred [ 45 , 47 , 53 , 57 ].
Calibration of the GGM and GLM model
We calibrate the GGM and the GLM model to the daily incidence curve by dates of reporting in Chile using time series data from March 3 rd –March 30 th , 2020, and from May 9th–July 7 th , 2020, respectively ( Fig 2 ). The period from March 3 rd –March 30 th , 2020, includes the initial interventions made by the Chilean government, whereas the period from May 9 th -July 7 th , 2020, comprises the reimposition of lockdowns after a brief reopening of society under the “new normal” ( Fig 1 ).
10.1371/journal.pntd.0009070.g002
Fig 2
Daily incidence curve for all COVID-19 confirmed cases in Chile as of November 2 nd , 2020 [ 27 ].
Model parameters are estimated by a non-linear least-square fitting of the model solution to the incidence data by the date of reporting. This is achieved by searching for the set of model parameters Θ ^ = ( Θ ^ 1 , Θ ^ 2 , … . Θ ^ m ) that minimizes the sum of squared differences between the observed data y ti = y t 1 , y t 2 ,…. y tn and the corresponding mean incidence curve given by f ( t i , Θ ^ ) = C ′ ( t ) : where Θ ^ = ( r , p ) corresponds to the set of parameters of the GGM model and Θ ^ = ( r , p , k 0 ) corresponds to the set of parameters of the GLM model. In both cases, the objective function for the best fit solution of f ( t i , Θ ^ ) is given by:
Θ ^ = a r g m i n ∑ i = 1 n ( f ( t i , Θ ) − y t i ) 2
where t i is the time stamp at which the time series data are observed and n is the total number of data points available for inference. The initial condition is fixed to the first observation in the data set. This way, f ( t i , Θ ^ ) gives the best fit to the time-series data y t i . Next, we utilize a parametric bootstrapping approach assuming a negative binomial error structure for the GGM and GLM model to derive uncertainty in the parameters obtained by non-linear least-square fit of the data as previously described [ 54 , 58 ]. The variance is assumed to be three times the mean for GGM and 96 times the mean for the GLM. The model confidence intervals of parameters and the 95% prediction intervals of model fit are also obtained using the parametric bootstrap approach [ 54 ].
Reproduction number, R , from case incidence using GGM
The reproduction number, R , is defined as the average number of secondary cases generated by a primary case at time t during the outbreak. This is crucial to identify the intensity of interventions required to contain an epidemic [ 59 – 61 ]. Estimates of effective R indicate if the disease transmission continues ( R >1) or if the active disease transmission ceases ( R <1). Therefore, in order to contain an outbreak, we need to maintain R <1. We estimate the reproduction number by calibrating the GGM to the epidemic’s early growth phase (27 days) [ 55 ]. We model the generation interval of SARS-CoV-2, assuming gamma distribution with a mean of 5.2 days and a standard deviation of 1.72 days [ 62 ]. We estimate the growth rate parameter, r , and the deceleration of growth parameter, p , as described above. The progression of local incidence cases I i at calendar time t i is simulated from the calibrated GGM model. Then in order to estimate the reproduction number, we apply the discretized probability distribution of the generation interval denoted by ρ i to the renewal equation as follows [ 43 , 44 , 63 ]:
R t i = I i ∑ j = 0 i ( I i − j ρ j )
The numerator represents the total new cases I i , and the denominator represents the total number of cases that contribute to generating the new cases I i at time t i . Hence, R t , represents the average number of secondary cases generated by a single case at time t . Next, we derive the uncertainty bounds around the curve of R t directly from the uncertainty associated with the parameter estimates ( r , p ) obtained from the GGM. We estimate R t for 300 simulated curves assuming a negative binomial error structure where the variance is assumed to be three times the mean [ 54 ].
Reproduction number, R , from case incidence using GLM
In order to estimate the reproduction number by July 7 th , 2020 (after the reimposition of lockdowns in Santiago and Valparaiso), we calibrate the GLM from May 9th–July 7 th , 2020 [ 55 ]. Next, we model the generation interval [ 62 ], estimate the model parameters ( r , p , k 0 ) from GLM and the reproduction number from the renewal equation as described above [ 43 , 44 , 63 ]. The uncertainty bounds around the curve of R t are derived directly from the uncertainty associated with the parameter estimates ( r , p , k 0 ). We estimate R t for 300 simulated curves assuming a negative binomial error structure [ 54 ] where the variance is assumed to be 96 times of the mean calculated by averaging mean to variance ratio calculated from the data (by binning data points and calculating directly from the data itself).
Instantaneous reproduction number, R , using the Cori method
We estimate R by the ratio of number of new infections generated at time t ( I t ), to the total infectiousness of infected individuals at time t, given by [ 64 , 65 ]:
∑ s = 1 t I t − s w s
In this equation, w s represents the infectivity profile of the infected individual, which depends on the time since infection ( s) , but is independent of the calendar time ( t ) [ 66 , 67 ].
More specifically, w s is defined as a probability distribution describing the average infectiousness profile of an individual after infection. Distribution of w s is affected by individual biological factors such as symptom severity or pathogen shedding. The equation ∑ s = 1 t I t − s w s indicates the sum of infection incidence up to time step t − 1, weighted by the infectivity function w s . The distribution of the generation time can be utilized to approximate the infectivity profile, w s , however, since the time of infection is rarely observed, it becomes difficult to measure the distribution of generation time [ 64 ]. Hence, time of symptom onset is usually used to estimate the distribution of serial interval, which is defined as the time interval between the dates of symptom onset among two successive cases in a transmission chain [ 68 ]. The infectiousness of a case is a function of the time since infection. This quantity is proportional to w s if we set the timing of infection in the primary case as the time zero of w s and assume that the generation interval equals the SI. The SI was assumed to follow a gamma distribution with a mean of 5.2 days and a standard deviation of 1.72 days [ 62 ]. Analytical estimates of R t were obtained within a Bayesian framework using EpiEstim R package in R language [ 68 ]. R t was estimated at 7-day intervals. We reported the median and 95% credible interval (CrI).
Results
Case incidence data
The Ministry of Health Chile reported a total of 481,342 COVID-19 cases as of November 2 nd , 2020 [ 27 ]. The epidemic curve showed an increasing trajectory from April-June 2020 and declined thereafter. On average, ~443 (SD: 133.6) new cases per day were reported in April 2020, ~2697 (SD:1342) new cases per day were reported in May 2020 and ~4943 (SD:972.2) new cases per day were reported in June 2020, the maximum number of cases reported per day during the epidemic. The per-day cases declined starting July, with ~2456 (SD:581) new cases reported per day in July 2020, ~1808 (SD:258) new cases per day reported in August 2020, ~1706 (SD:294) new cases per day reported in September 2020, and ~1521 (SD:275) new cases per day reported in October 2020. Fig 2 shows the daily incidence data of all confirmed cases in Chile as of November 2 nd , 2020.
Initial growth dynamics and estimate of the reproduction number using GGM
We estimate the reproduction number for the first 27 epidemic days incorporating the effects of the social distancing interventions, as explained in Table 1 and Fig 1 . The incidence curve displays sub-exponential growth dynamics with the scaling of growth parameter (deceleration of growth parameter), p , estimated at 0.77 (95% CI: 0.73, 0.81) and the intrinsic growth rate, r , estimated at 0.81 (95% CI: 0.67, 1.00). During the early transmission phase the reproduction number was estimated at 1.80 (95% CI: 1.60, 1.90) ( Fig 3 ).
10.1371/journal.pntd.0009070.g003
Fig 3
Reproduction number with 95% CI estimated using the GGM model.
The estimated reproduction number of the COVID-19 epidemic in Chile as of March 28 th , 2020 is 1.80 (95% CI: 1.60, 1.90).
Growth dynamics and estimate of reproduction number using GLM by July 7, 2020
We also estimate the reproduction number from May 9 th - July 7 th , 2020, incorporating the effects of the reimplementation of localized lockdowns in Santiago, Antofagasta, and Valparaíso. The incidence curve displays a nearly linear growth trend with the deceleration of growth parameter, p , estimated at 0.51 (95% CI: 0.47, 0.56). The deceleration parameter in the GLM model helps modulate the trajectory of the epidemic, depicting a linear growth trend. The intrinsic growth rate, r , was estimated at 22 (95% CI: 13, 31) and the final epidemic size, k 0 , estimated at 3.4 e+05 (95% CI: 3.1 e+05, 3.7 e+05). The reproduction number was estimated at 0.87 (95% CI: 0.84, 0.89) as of July 7 th , 2020 ( Fig 4 ).
10.1371/journal.pntd.0009070.g004
Fig 4
Reproduction number with 95% CI estimated by calibrating the GLM model from May 9 th -July 7 th , 2020.
The estimated reproduction number of the COVID-19 epidemic in Chile as of July 7 th , 2020 is 0.87 (95% CI: 0.84, 0.89).
Estimate of instantaneous reproduction number using Cori method
Utilizing the Cori method based on a sliding weekly window, we observe that the reproduction number peaked on March 16 th , 2020, with an estimate of R~ 6.19 (95% CrI = 5.84, 7.08). The reproduction number declined thereafter and reached ~1.00 (95% CrI: 0.99, 1.04) on April 17 th , 2020. From April 18 th -June 18 th , 2020 the reproduction number fluctuated between 1.01–1.75. This was followed by a decline in the reproduction number to less than 1.0 between June 19 th -August 9 th , 2020. Since then, the reproduction number has fluctuated around 1.0 with the most recent estimate of R ~ 0.96 (95% CrI: 0.95, 0.98) ( Fig 5 ).
10.1371/journal.pntd.0009070.g005
Fig 5
Estimate of instantaneous reproduction number ( R ) for the COVID-19 epidemic in Chile as of November 2 nd , 2020 using the Cori method.
The most recent estimate of R ~ 0.96 (95% CrI: 0.95, 0.98) as of November 2 nd , 2020. Black solid line represents the mean R and the gray shaded region represents the 95% credible interval around mean R .
Assessing the impact of social distancing interventions
To assess the impact of social distancing interventions in Chile given in Table 1 , we generated a 20-day ahead forecast for Chile based on the daily incidence curve until March 30 th , 2020. The 28-day calibration period of the GGM model yields an estimated growth rate, r , at 0.80 (95% CI: 0.60, 1.00) and a deceleration of growth parameter, p , at 0.80 (95% CI: 0.70, 0.80), indicating early sub-exponential growth dynamics. The 20-day ahead forecast suggested that the early social distancing measures significantly slowed down the early spread of the virus in Chile, whose effect is noticeable about two weeks after implementing an intervention, as shown in Fig 6 . A case resurgence was observed in Chile in mid-May 2020. As a consequence of this case resurgence, a total lockdown was imposed in Greater Santiago (representing ~52% of total COVID-19 cases during the epidemic) on May 15 th , 2020. The quarantine in Santiago was gradually eased from August 17, 2020, and was lifted on September 28, 2020, as a part of the move to phase three of a five-step plan of deconfinement that would allow movement on regional transportation and reopening of non-essential businesses and schools [ 31 , 69 , 70 ]. We generated a 20-day ahead forecast based on the daily incidence curve from May 9 th -July 7 th , 2020. The 60-day calibration of the GLM model yields an estimated scaling of the growth parameter, p , at 0.52 (95% CI: 0.47, 0.57), representing an almost linear growth pattern. The 20-day ahead average forecast utilizing the GLM model showed that Chile could accumulate ~45,160 cases (95% CI: 27,934–67,600) between July 8 th -July 27 th , 2020 ( Fig 7 ). Our forecast results approximate closely the ~46,798 cases reported between July 8 th -July 27 th , 2020 by the Ministry of Health, Chile.
10.1371/journal.pntd.0009070.g006
Fig 6
20-days ahead forecast of the COVID-19 epidemic in Chile by calibrating the GGM model until March 30 th , 2020.
Blue circles correspond to the data points; the solid red line indicates the best model fit, and the red dashed lines represent the 95% prediction interval. The vertical black dashed line represents the time of the start of the forecast period.
10.1371/journal.pntd.0009070.g007
Fig 7
20-days ahead forecast of the COVID-19 epidemic in Chile by calibrating the GLM model from May 9 th -July 7 th , 2020.
Blue circles correspond to the data points; the solid red line indicates the best model fit, and the red dashed lines represent the 95% prediction interval. The vertical black dashed line represents the time of the start of the forecast period.
COVID-19 Testing rates and positivity rate
Daily testing and positivity rates for the time period April 9 th –November 2 nd , 2020, by the reporting date are shown in Fig 8 . The total number of tests performed for this time period were 4,325,617, amongst which 10.9% (47,597) had positive results. The average number of daily tests was estimated at ~5,460 for April 2020 and ~12,959 for May 2020, a 137% increase. The testing rate in Chile further increased in June 2020, testing on average ~17,578 individuals per day, followed by a slight decline in July 2020, testing on average 16,587 individuals per day. However, the testing rates continued to increase in August (average ~26,079 tests per day), September (average ~29,663 tests per day), and October (average ~31,821 tests per day), indicating an expanding testing capacity of the country. The positivity rate (percentage of positive tests among the total number of tests) has fluctuated from a monthly average of ~9.07% (SD: 2.3) in April 2020 to a monthly average of ~4.87% (SD: 0.65) in October 2020.
10.1371/journal.pntd.0009070.g008
Fig 8
Laboratory results for the COVID-19 tests conducted in Chile as of November 2 nd , 2020.
The blue color represents the negative test results, and the yellow color represents the positive test results. The solid orange line represents the positivity rate of COVID-19 in Chile.
Discussion
The estimates of the early transmission potential in Chile for the first 27 days of the epidemic indicate sustained local transmission in the country with the estimate of reproduction number R at ~1.80 (95% CI: 1.60, 1.90) which is also in accordance with the estimate of the reproduction number obtained from the Cori method (R~2.2 95% CrI (2.14, 2.28)). The estimates of R from our analysis agree with the estimates of R retrieved from studies conducted in the surrounding Latin American countries including Peru and Brazil [ 71 , 72 ]. Other countries including Korea, China, South Africa and Iran also exhibit similar estimates of R that lie in the range of 1.5–7.1 [ 73 – 80 ]. In contrast, some other countries including Singapore and Australia have reported much lower estimates of R ( R <1) that can be correlated with the implementation of early strict social distancing interventions in these countries [ 81 , 82 ].
The initial deceleration of the growth parameter in Chile indicates a sub-exponential growth pattern ( p ~0.8), consistent with sub-exponential growth patterns of COVID-19 that have been observed in Singapore ( p ~0.7), Korea ( p ~0.76) and other Chinese provinces excluding Hubei ( p ~0.67) [ 78 , 81 , 83 ]. In contrast, studies conducted in Peru, a Latin American country, and Iran have reported a nearly exponential growth pattern of the COVID-19 whereas an exponential growth pattern has been reported in China [ 72 , 75 , 83 ].
Although the initial transmission stage of COVID-19 in Chile has been attributed to multiple case importations, Chile quickly implemented control measures against the COVID-19 epidemic, including border closures on March 18 th , 2020, to prevent further case importations. The 20-day ahead forecast of our GGM model calibrated to 28 days suggest that the social distancing measures, including closure of schools, universities and day cares, have helped slow down the early virus spread in the country by reducing population mobility ( Table 1 and Figs 1 and 6 ) [ 84 ]. The commixture of interventions, including localized lockdowns, night-time curfew, school closures, and the ban on social gatherings in Chile, can probably be attributed to preventing the disease trend from growing exponentially during the early growth phase, as has occurred elsewhere [ 3 , 4 ]. However, the significant increase in case incidence observed in mid-May can probably be attributed to the relaxation of social distancing measures and reopening of society in late April, in the context of the “Safe Return” plan [ 31 ]. As the virus reached the lower-income neighborhoods in Chile, the pandemic quickly exploded [ 23 , 38 , 39 , 85 ]. While the COVID-19 case incidence exhibited a relative stabilization in case trajectory for April 2020 (with an average of ~443 cases per day), highlighting the positive effects of early quarantine and lockdowns, the reopening of society and early economic reactivation in late April 2020 probably resulted in the surge of cases resulting in an acceleration of the epidemic with estimates of R higher than 1.0. The total lockdown comprised of stay-at-home orders imposed in Greater Santiago (which accounted for about 77% of cases in the country) on May 15 th showed an effect in slowing the virus's transmission. Similar lockdowns were imposed in Antofagasta on May 5 th and in Valparaíso on June 12 th , though these regions together represent only ~10% of cases in Chile [ 31 , 28 ]. The deceleration of growth parameter, p , has been estimated at ∼0.51 (95% CI: 0.47, 0.56) after the reimposition of lockdowns and social distancing measures in May, consistent with a linear incidence growth trend, indicating deceleration of the epidemic.
Moreover, we estimated a reproduction number, R , of ~0.87 (95% CI: 0.84, 0.89) in early July, indicating a decline in transmission of the virus consistent with the stay-at-home orders. This reproduction number corresponds to the instantaneous reproduction number estimated during the course of the epidemic utilizing the Cori method, which also indicates a decrease in disease transmission with R ~0.8 as of early July. The instantaneous reproduction number has fluctuated around ~1 since early August with the most recent estimate of reproduction number, R ~0.9 as of November 2 nd , 2020. The 20-day ahead forecast calibrating data to the GLM model (from May 9 th -July 7 th , 2020) has reasonably indicated a declining trend in case incidence. The forecast results also imply that approximately ~45,160 cases (95% CI: 27,934–67,600) could be observed in Chile from July 8 th -July 27 th , 2020. The actual case count by for this time period, as reported by the Chilean government indicated 46,798 cases [ 40 ], closely approximating our mean GLM forecast, falling within the 95% PI. Therefore, based on the most recent estimates of R ( Fig 5 ), it can be implied that maintaining social distancing measures, limiting social gatherings, and reducing population mobility have served as essential ways to containing the spread of the virus [ 86 , 87 ].
Though the number of reported cases in Latin America remains low compared to the United States, official data for many Latin American countries are incomplete. However, Chile has tested a higher percentage of its residents than any other Latin American nation, lending confidence to its reliability [ 88 ]. Chile’s testing capabilities have been greatly expanded during the pandemic, in part from a coordinated effort lead by the Ministry of Science to include testing from public and private laboratories. For instance, the average number of COVID-19 tests performed in Chile per day per thousand people is 1.52 compared to the neighboring South American country, Peru (~0.2 tests per thousand people) as of November 2 nd , 2020 [ 89 ]. The average positivity rate for the whole span of the epidemic in Chile is estimated at ~12.98%. However, the average monthly positivity rate of COVID-19 in Chile is estimated at ~5.90% and ~4.88% for September and October, respectively, compared to ~20.09% in May 2020. The high positivity rate at the beginning of the epidemic indicates that the government failed to cast a wide enough net to test the masses early in the pandemic, and there were probably many more active cases than those detected by epidemiological surveillance, underestimating the epidemic growth curve [ 90 – 92 ]. The most recent lower positivity rates indicate that Chile is testing a comparatively larger segment of the population. This positivity rate for Chile is also consistent with the positivity rate obtained from India, the United States, Canada, and Germany that exhibit moderately high positivity rates (4–8%) for COVID-19, indicating overall limited testing in these countries [ 89 , 93 ]. In comparison, some countries like Mexico and the Czech Republic exhibit very high positivity rates (30–51%) [ 89 ]. Other countries like Denmark, New Zealand, Australia, Singapore, and South Korea have reached very low positivity rates (0–3%) by testing the masses with South Korea’s large testing capacity combined with a strategy that tracks infected people via cell phones [ 88 , 89 , 94 ]. Moreover, studies suggest there is asymptomatic transmission of SARS-CoV-2 [ 66 , 95 , 96 ], which means we could have underestimated our estimates based on the daily incidence’s growth trend from symptomatic cases [ 97 – 99 ]. On the other hand, preliminary results of a study have shown the relative transmission of asymptomatic cases in Santiago to be almost ~3% [ 100 ]. While our study highlights the effectiveness of broad-scale social distancing and control interventions in Chile, it also underscores the need for persistent isolation and social distancing measures to stomp all active disease transmission chains in Chile. In the absence of pharmacological interventions and considering the advent of second waves in Asia and Europe, non-pharmacological interventions such as the ones implemented in Chile are the available options for countries to address the pandemic before large segments of the population are immunized with effective and safe vaccines. In this scenario, real-time metrics that characterize the transmission dynamics and control are crucial to face the future challenges that the pandemic will impose.
This study has some limitations. First, our study analyzes cases by the dates of reporting while it is ideal to analyze the cases by the dates of onset or after adjusting for reporting delays. On the other hand, a substantial fraction of the COVID-19 infections exhibits very mild or no symptoms at all, which may not be reflected by data [ 101 , 102 ]. Moreover, the data are not stratified by local and imported cases; therefore, we assumed that all cases contribute equally to the transmission dynamics of COVID-19. Finally, the extent of selective underreporting, and its impact on these results is difficult to assess.
Conclusions
In this study, we estimate the transmission potential of SARS-CoV-2 in Chile. Our current findings point to sustained transmission of SARS-CoV-2 in the early phase of the outbreak, with our estimate of the reproduction number at R ~1.8. The COVID-19 epidemic in Chile followed an early sub-exponential growth trend ( p ~0.8) which has transformed into an almost linear growth trend ( p ~0.5) as of July 7 th , 2020. The most recent estimate of reproduction number, R , is ~0.9 as of November 2 nd , 2020, indicating a decline in the virus transmission in Chile. The implementation of lockdowns and apt social distancing interventions have indeed slowed the spread of the virus. However, the number of new COVID-19 cases continue to accumulate, underscoring the need for persistent social distancing and active contact tracing efforts to maintain the epidemic under control.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
By combining graph theory and statistical physics, complex network theory provides a powerful tool to investigate the structure and function of complex systems with a large number of interacting elements. The development and characterization of complex networks [1] – [3] makes their application suitable to analyze a wide range of systems from nature to economy, from engineering to society [4] . Beside the well-established applications to Internet and World Wide Web, neural connections and social dynamics [5] , complex networks have been successfully used to study many different phenomena such as, for example, human migration [6] , cancer metastasis [7] and earthquake occurrence [8] .
The extension of complex network theory to climate sciences is a very recent area yielding climate networks , which usually rely on gridded time series of meteorological preprocessed variables. The nodes of the network are identified by geographical regions corresponding to single points of measurement on the spatial grid of the underlying climate database. Each node has a measured state variable which varies in time. A link between two nodes exists if there is a significant statistical interdependence between their time series. The linear cross-correlation function is typically used as the simplest possible measure of the statistical interdependence of the temporal series. However, the influence of the choice of an association measure on the topology of the climate network has been studied, by accounting how the temporal complexity of time series influences the absolute correlations [9] and proposing alternative criteria based on the nonlinear mutual information [10] , [11] .
Until now, attention has been mainly devoted to networks based on the global surface temperature field to understand the influence of El Niño and La Niña events in regions which are far from the El Niño-Southern Oscillation (ENSO) area. Although the temperatures in different zones of the world are not significantly affected by El Niño and La Niña, it was surprisingly found that the climate network during these events is sensitively influenced by showing a different structure and a consistent amount of broken links [12] – [14] . The community structure [15] as well as the dynamics of interacting networks [16] have been investigated using the surface temperature fields and related variables (e.g., sea level pressure and equipotential heights).
Apart from some works where different meteorological variables such as equipotential heights [15] – [17] , sea surface temperature, humidity, and related data [18] are also analyzed, the great part of climate network literature deals with surface air temperature data only. In particular, to the best of our knowledge, a global precipitation analysis has not yet been performed by using the complex network theory. Only Malik et al. [19] recently carried out a complex network study of local extreme monsoonal rainfall in South Asia, while Bayesian networks have been employed to analyze local precipitation in the Iberian peninsula [20] , [21] .
The gap concerning global precipitation in climate networks is evident even though precipitation teleconnections, especially related to the ENSO occurrence [22] , [23] , have been recognized for being a crucial aspect in climate and hydrological changes [24] , and are increasingly examined for Asian monsoons [25] – [28] , European [29] , [30] , African [31] and American [32] , [33] rainfall events.
The present work arises in this scenario and aims at being a first step towards filling this gap in climate network analysis; in fact, precipitation, together with surface temperature and wind, atmospheric pressure and humidity, is one of the most important meteorological variables in defining the climate dynamics. The global annual precipitation over seventy years (1941–2010) is analyzed by means of the complex network theory. We use the Global Precipitation Climatology Centre (GPCC) Database [34] – [36] , which is one of the most reliable precipitation datasets providing land-surface precipitation from rain-gauges over the period 1901–2010. Pre-processing of the data is performed to (i) define nodes corresponding to geographical regions (cells) with the same area, and (ii) consider data mainly based on in situ observations (rather than on interpolated values). The precipitation network is built identifying a node with a geographical region, which has a temporal distribution of measured precipitation, and using the linear correlation function to evaluate possible links between nodes. If the statistical interdependence between two nodes is above a suitably chosen threshold, a link between the two nodes is established. The precipitation network is described through classical tools of the complex network theory - such as the degree centrality, the betweenness centrality and the clustering coefficient - as well as measures introduced here for the first time: the weighted average topological distance, which generalizes the average topological distance definition, and the average physical distance of a node from the rest of the network. To analyze the sensitivity of the network to the physical closeness of the nodes, short-term connections are broken and edges between physically distant nodes only survive. In so doing, the unavoidable spatial correlation between physical neighbors is left aside in favor of highlighting the possible interdependence between not confining regions.
Materials and Methods
In this section the database used to define the precipitation network is described. Details on the pre-processing analysis of data are then offered. Afterwards, we summarize some fundamental concepts in complex network theory [3] , [5] . We only introduce measures which are relevant to the present analysis. In the end, starting from the spatio-temporal global precipitation distribution, details on how to build the precipitation network will be given.
GPCC Full Database Description
The present investigation uses the Global Precipitation Climatology Centre (GPCC) Full Data Reanalysis Version 6, which consists of monthly land-surface precipitation data from rain-gauges built on Global Telecommunication System (GTS) and historic data [34] – [36] . The GPCC Full Database covers the period from January 1901 to December 2010 and is based on both real-time raingauge data as well as non real-time sets of data. The new extended database version was released in December 2011 and the data coverage per month varies from 10,700 to more than 47,000 stations. Non real-time data, coming from dense national observation networks of individual countries and other global and regional collections of climate data, are integrated in the GPCC Full Database using the GPCC Precipitation Climatology Database as analysis background (for more details, see the online documentation [36] ).
The monthly global precipitation is reported on a regular grid with a spatial resolution of × , × , and × latitude by longitude. The global gridsystem goes from ( W, N) to ( E, S). For every gridcell three kinds of data are given: (i) monthly precipitation [mm/month]; (ii) mean monthly precipitation [mm/month] based on the GPCC Precipitation Climatology; (iii) number of gauges per grid. We obtain the annual precipitation by summing the monthly precipitation values.
The GPCC Full Database provides one of the most accurate and complete in situ precipitation data sets. In fact, the wide spatial and temporal coverage makes this database suitable for verification of models [37] , [38] , for analysis of historic global precipitation [39] – [43] , and for research concerning the global water cycle [44] , [45] , e.g., trend and time-series analyses [46] , [47] . Moreover, the number of stations per gridcell is an additional useful piece of information, which can be exploited when evaluating the spatio-temporal precipitation distribution.
Data Pre-Processing
Although the GPCC Full Database is a very accurate and detailed dataset, there are two main concerns to use it, as it is, to define a complex network. In this section we discuss these two issues and propose our solutions to overcome them.
First, the regular grid based on the angular partition of the terrestrial surface (i.e., the geographic coordinate system) leads to the definition of gridcells with different geometric area. This heterogeneity, which becomes more evident by approaching regions far from the equator, may induce substantial bias and spurious correlation when building the precipitation network. One way to avoid this bias is to use a tessellation technique [48] to divide the gridcells into suitable two-dimensional structures. An alternative axiomatic scheme, based on the idea of node splitting invariance to obtain consistent weights for the most commonly used network parameters, was proposed by Heitzig et al. [49] .
Here, we adopt a simpler approach: we build a new grid system with all square cells having a fixed area, × . We focus on the equator, where the original GPCC grid system with × gridcell resolution yields a square cell dimension ( × ) of 278.3×278.3 km 2 . The new graticule, which is here proposed, is built maintaining this cell dimension ( × ) fixed for all latitudes. As a consequence, the number of the new cells is variable over the latitude and, in particular, the new graticule has fewer cells than the original GPCC grid system when moving far from the equator. The number of raingauges of a new cell is the sum of all the raingauges present in the GPCC cells which are completely contained by the new cell. The precipitation value of the new cell is instead the average of the precipitation values measured by those GPCC cells which are entirely included into the new cell. To avoid possible overlaps in the longitudinal direction between cells of the new graticule, we adopt the following convention: when two new cells share an edge falling into an original GPCC cell, the contribute of the original cell (in terms of number of raingauges and precipitation value) is fully allocated to the cell of the new grid system which covers most of the GPCC cell area.
The second concern is about the spatio-temporal distribution of measurement stations. In the GPCC Full Database, in fact, there exists an amount of cells for which no measurement is available over several months (i.e., the number of raingauges for the gridcell is 0). However, in these cases, the global precipitation information are recovered through the interpolation of global and mean data offered by the GPCC Precipitation Climatology Database. In so doing, the spatio-temporal coverage is complete but some precipitation values can be fully based on interpolated data. This aspect becomes important when the oldest data are analyzed, since fewer measurements were available. In order to consider data mainly based on in situ observations, we define a grid cell as active for a fixed year if there is at least one measurement for every month of the year. Otherwise the grid cell is not active. We then restrict the analysis to a temporal window of 70 years starting from 1941 to 2010, and consider only cells which are active (also not consecutively) for at least 50 years over the temporal window of 70 years. In this way, more than 90% of the stations will be included in the spatio-temporal precipitation distribution and, in the worst cases, less than 30% of the distribution (20 years over 70) will rely on interpolated data. The number of active cells is .
Complex Network Tools: Definitions and Structural Properties
A network (or graph) is defined by a set of nodes and a set of edges (or links) . Here, we assume , that is the number of nodes of the network, , can be equal or lower than the number of active cells, . Moreover, we suppose that only one edge can exist between a pair of nodes and no self-loops are allowed. The adjacency matrix , : (1) takes into account whether a link is active or not between nodes and . Since the network is considered as undirected, is symmetric. Since no self-loops are allowed, .
The degree centrality of a node is defined as (2) and gives the number of first neighbors of the node , normalized over the total number of possible neighbors ( ). The degree distribution , , defines the fraction of nodes in the graph having degree . In other words, the degree distribution is the probability that a node in the network is connected to other nodes.
We here propose the weighted average topological distance of a node as (3) where the shortest path length, , is the minimum number of edges that have to be crossed from node to node , and is the set of all neighbors of . is the number of nodes connected to node ( ). The first ratio of the right hand side of Eq. (3) accounts for the mean topological distance of node with respect to all the nodes linked to it, while the second ratio is a weight coefficient considering how strongly node is connected to the rest of the graph. We introduce this notation to generalize the classical average topological distance definition [50] , . In fact, when the graph has disconnected components the average topological distance definition diverges. Relation (3) is identical to the average topological distance when the graph is completely connected ( for every node). In the case of a graph with disconnected nodes, instead, does not diverge and can vary in the interval . The extreme value is reached when two nodes, and , are directly connected one to each other ( ), but disconnected from the other nodes ( ). A large value means that node is topologically far from the rest of the network.
The local clustering coefficient of a node is (4) where is the set of first neighbors of , is the number of edges connecting the vertices within the neighborhood , and is the maximum number of edges in , . The local clustering coefficient gives the probability that two randomly chosen neighbors of are also neighbors. The global clustering coefficient is the mean value of , .
The betweenness centrality of a node is (5) where are the number of shortest paths connecting nodes and , while gives the number of shortest paths from to crossing node . If node is traversed by a large number of all existing shortest paths (that is, if is large), then node can be considered an important mediator for the information transport in the network.
Building the precipitation network
The number of active cells, as previously described, is . Once the time series of the annual precipitation is obtained for each active cell, we can evaluate the cross correlation between all pairs of them. We use the linear Pearson correlation as it is the simplest possible measure to quantify the degree of statistical interdependence between the temporal series. Moreover, Donges et al. [10] found a high level of similarity between Pearson correlation and mutual information networks. The correlation coefficient is given by an element of the correlation matrix, , which is symmetric and estimates the strength of a linear interdependence between two temporal series, and ( , by definition). The correlation coefficient can vary between −1 and 1. A large positive value means the temporal series are strongly correlated, while a large negative value indicates a strong anti-correlation. Since we are interested in both large positive and negative correlation values, the absolute value of will be used to build the precipitation network [10] . Moreover, the physical distance, , is evaluated in kilometers as the shorter great circle path between nodes and and stored in the symmetric matrix .
The average physical distance of an active cell (node) is defined here as (6) where is the number of active cells and is the physical distance between nodes and defined above ( by definition).
The edge density is defined as: (7) where is the number of active links (edges) when the absolute value of (in the following we abbreviate the correlation with ) is above the threshold , while is the cumulative distribution function of the correlation .
To define the adjacency matrix, , and therefore the network, we refer to Eq. (1) and define that an edge, , between nodes and exists when . In so doing, the resulting precipitation network is undirected (A is symmetric) and unweighted (all the values above the threshold correspond to ). The selection of the threshold is a non-trivial aspect of building a climate network [10] , [12] , [13] . Since we are primarily interested in highlighting strong correlated and anti-correlated connections, we set . With this threshold value, the number of active links is and the number of nodes is (57 nodes out of 1731 are not connected with any other node of the network). The chosen threshold, , corresponds to an edge density value, , equal to . We graphically make the nodes coincide with the center of each active cell, see the blue symbols in the left panel of Fig. 1 .
10.1371/journal.pone.0071129.g001 Figure 1
Nodes and links of the network.
(left) Nodes of the network. Each node graphically coincides with the center of an active spatial cell. Red symbols correspond to the nodes of the network ( ) with the additional physical constraint, km, while blue symbols correspond to the nodes of the network ( ) without this constraint. (right) Number of links, , as a function of the physical distance, [ km]. The scale of values is in km, the red line indicates the physical threshold, km.
Results
The properties of the precipitation network are here presented and discussed. Among the network measures introduced in the Materials and Methods Section, particular attention will be paid to the degree centrality and the weighted average topological distance. In fact, these two parameters reveal to be the most meaningful for the present analysis.
As mentioned, the precipitation network is made of nodes. However, not all of them are completely connected one to the other. This means that the graph has disconnected components. This aspect justifies the choice of proposing a different definition of the mean topological distance, see Eq. (3) . In particular, nodes are completely linked and form a big subnetwork, while the remaining nodes contribute to create 14 smaller micro-networks. The size (number of nodes) of each micro-network varies from 9 to 2. Although visible through the betweenness centrality, the best parameter which physically individuates the nodes of these smaller networks is the weighted average topological distance, .
The data analysis is completed by describing the properties of another precipitation network with an additional constraint: an edge, , between nodes and exists when and . A suitably large value of the threshold leads to define a new precipitation network, where edges only exist between nodes which are physically far from each other. In so doing, the unavoidable spatial correlation between physical neighbors is left aside in favor of highlighting the possible interdependence between not confining regions (i.e., ).
To focus on possible links between regions physically far from each other, we set km and obtain a new network with fewer nodes, , and links, , than the network previously discussed. The American and Asiatic continents are the regions which much suffer of the reduction of nodes, while Western Europe and Australia still have an appreciable number of nodes (see the red symbols in the left panel of Fig. 1 ). The number of active links, , is reported in the right panel of Fig. 1 as a function of the physical distance, (the scale of values is in km, the red line represents the threshold km).
Properties of the precipitation network
We start considering the degree centrality for the basic network, see Fig. 2 . One clearly distinguishes two regions with the highest degree centrality values: the Sahel region in Africa, and Eastern Australia. The nodes in these regions are directly connected to a great number of other nodes of the network, therefore they are usually referred to as supernodes . Beside the supernode areas, there exist regions with fairly high degree centrality values: Northern Europe, Central Asia, Southern Africa, Western US, and Northeastern Brazil. It should be noted the great difference in terms of degree centrality values, , when moving from the West to the East Coast of the US, as well as from Northern Europe towards the Mediterranean Sea. We recall that the degree centrality is the ratio between the number of cells directly linked to a fixed cell, normalized over the total number of possible neighboring cells ( ). Speaking in terms of the physical area directly connected to a cell, the maximum degree centrality value here reached, , corresponds to a directly connected physical area of about km 2 , equivalent to the total area of the countries in the European Union.
10.1371/journal.pone.0071129.g002 Figure 2
Degree centrality, . Large values correspond to highly connected nodes.
Real networks are often scale-free, that is power-law degree distributions are displayed [5] , with exponents ranging between −2 and −3. These networks usually result in the simultaneous presence of few nodes highly connected to the others (i.e., supernodes ) and a great amount of barely connected cells. However, because of the finite size of the network, data can have a rather strong intrinsic noise. To smooth the fluctuations generally present in the tails of the distribution, it is often verified if the cumulative distribution function, , presents a power-law behavior. Figure 3 reports the exceedance probability of the node degree, . A power law decay with exponent equal to −2 is clearly observable in the intermediate range, (see the red line in Fig. 3 ).
10.1371/journal.pone.0071129.g003 Figure 3
Exceedance probability of the node degree, .
The weighted average topological distance, , is represented in the top panel of Fig. 4 . The scale of values is restricted to the interval [6] , [12] , while higher values are reported without distinction with the grey color. The grey-colored regions individuate the nodes of the micro-networks. In fact, as mentioned in the Materials and Methods Section, disconnected components of the graph have extremely high values. In the worst case, when a micro-network has two nodes only, . This situation occurs for 18 nodes out of 1674.
10.1371/journal.pone.0071129.g004 Figure 4
Weighted average topological distance and average physical distance.
(top) Weighted average topological distance, . The scale of values spans in the interval [6] , [12] . Higher values are reported without distinction with grey color. (bottom) Average physical distance, . The scale of values is in km.
As a first comment, there is quite a notable correspondence between high degree centrality and low weighted average topological distance, and viceversa. This is especially evident, on one hand, for large values as in the supernode areas, whose nodes have the lowest values. On the other hand, regions with the highest values (grey and red colored zones in the top panel of Fig. 4 ) have . Few remarkable exceptions to this inverse correspondence are the Atlantic Coast of South-America, the Mongolian area and the Indonesian archipelago. For these regions medium-low values do not correspond to appreciably high degree centrality values. It should be noted that the North-American and European regions have quite the same low values.
At this stage, it is useful to compare the weighted average topological distance map, , with the average physical distance map, , offered in the bottom panel of Fig. 4 (the scale of values is in km). Some regions (Northwestern Europe, Central Asia and Mongolian area, African Sahel region) have both measures with quite low values and one can infer that the strong topological connection is partially due to the high physical closeness of the nodes involved. However, leaving aside these regions, for the rest of the network there exists an inverse correspondence between the weighted average topological distance and the average physical distance. This aspect is more marked in the Southern Hemisphere, where the average physical distance between active cells is in general higher and, at the same time, nodes are often topologically close one to each other. To this end, one can refer in particular to the Atlantic Coast of South America and Eastern Australia, but also South-Africa, Mid-Western US, Northern South-America and the Indonesian Archipelago. Nevertheless, the inverse proportionality between and is visible in topologically low connected areas such as Eastern Asia, which is instead a region whose cells on average are not physically far from the rest of the network.
To carry out a sensitivity analysis of the network with respect to the physical neighborhood of the nodes, we here define that an edge between nodes and exists if and, at the same time, the additional physical constraint, km, is satisfied. In so doing, a different network is specified where edges between nodes distant less than cannot exist.
The degree centrality and the weighted average topological distance are presented in the left and right panels of Fig. 5 , respectively. The current scenario deeply emphasizes the role of the supernode areas of the original network (the African Sahel region and Eastern Australia), which are still the regions with the highest degree centrality and the lowest weighted average topological distance. All the other previously existing links are instead weakened or, in several cases, even broken, meaning that their correlation was due to the physical closeness of the nodes involved. In particular, it is worth noting that the US and Western Europe show now very different patterns. Indeed, Western Europe still preserves a large number of highly connected nodes, while the US have a small amount of nodes which are scarcely connected to the rest of the network.
10.1371/journal.pone.0071129.g005 Figure 5
Long-range network, km. (left) Degree centrality. (right) Weighted average topological distance. The scale of values spans in the interval [8] , [16] , while higher values are reported without distinction with grey color.
The evident absence of nodes in Asia and North-America (and the poor connection of the few remaining cells) can be thought in terms of the presence of strong precipitation variation on a relatively short spatial scale, thereby leading to the emergence of high precipitation gradients. A high precipitation gradient can be, for instance, enhanced by the occurrence of regional extreme events (e.g., tropical cyclones, monsoons, tornadoes, blizzards, heat waves) which are usually localized in time and space. The Asian and North-American continents, due to their huge land mass extension, experience the most imponent extreme phenomena [51] . Therefore, in these places precipitation strongly varies on regions which are relatively close one to the other, allowing only short-range links to survive, which are eventually broken by the additional physical constraint, km. Examples of these high precipitation gradient areas are South-East and North-West China, Central Asia (including Mongolia, Kazakhstan and Central Russia), India, Nepal and Pakistan, as well as Western (Washington, Oregon and California), Central (Texas, Louisiana, Oklahoma, Arkansas, Kansas, Nebraska, Missouri, Iowa and Minnesota), and Southeastern (Florida, Mississippi, Alabama) United States. In all these cases, the high precipitation gradient makes extremely dry and extremely wet regions coexist in a few hundred kilometers range.
In the supernode regions (Sahel, Eastern Australia and Western Europe) extreme events in general occur less frequently. Medium (Western Europe) or very low (Sahel, Eastern Australia) rainfall spread out more uniformly on a continental scale, the precipitation gradient is weaker and nodes remain connected in the long-range.
Coming back to the original network with nodes, the local clustering coefficient and the betweenness centrality are presented in the top and bottom panels of Fig. 6 , respectively (note that a logarithmic representation is adopted for the betweenness centrality). These two measures are a little less significant and weakly related to the degree centrality and the weighted average topological distance maps. In fact, both distributions in Fig. 6 are spotted worldwide without evidencing regions of particular interest.
10.1371/journal.pone.0071129.g006 Figure 6
Local clustering and betweenness centrality measures.
(top): local clustering coefficient, . (middle): Probability density function, , of the local clustering coefficient, C. The red line represents the global clustering coefficient, . (bottom): betweenness centrality ( plot is shown).
From a qualitative point of view, the local clustering coefficient presents a strong heterogeneity in the central part of Africa, in Eastern Asia and South America. Some patterned zones with high values are found on coastal regions: Brazil, Eastern Australia and Eastern Africa. More in general, seems to mainly vary around the values represented by the green and yellow tones ( ). As plotted in the middle panel of Fig. 6 , the probability density function of the local clustering coefficient, , quantitatively confirms this behavior by showing a moderately trimodal distribution. The central mode, by far the most probable one, lies very close to the global clustering coefficient value, (see the red line in the middle panel of Fig. 6 ), which is the arithmetic mean of the values. Going back to the meaning of this parameter, the present results can be interpreted as follows: on average, there is about 52% of chances that two randomly chosen neighbors of node are also neighbors.
The betweenness centrality unveils the importance of a node in the network. Bottom panel of Fig. 6 summarizes that the nodes of the micro-networks and, more in general, nodes which are poorly connected to the rest of the network are the least important ones. These regions are depicted in dark blue. This result is not trivial, since the contrary (highly connected nodes are important) is not true. In fact, the importance of supernode areas and regions with low values is not detectable at all from the betweenness centrality map.
The shortest path distribution for 4 significant nodes can be observed in Fig. 7 . We recall that the shortest path length, , is the minimum number of edges that have to be crossed from node to node . The four nodes are chosen as examples of relevant behaviors. Two nodes in the Northern and Southern Europe regions are shown in panels A and B, respectively. The Northern Europe node is closely linked to the whole European region and Western Russia. Meanwhile, a topological connection of the same strength is found with Northern and Central America. This pattern can be qualitatively associated to the Gulf Stream impact and to the atmospheric circulation induced by the North Atlantic Oscillation (NAO) [52] . Important connections are also visible with the Australian and African Sahel regions, while the farthest nodes are located in Eastern Asia. Although not physically distant from the Northern Europe node, the Southern Europe node presents a quite different scenario (see panel B). A strong connection is evident with the European and African Sahel regions only, while the links with the American and Australian continents are weaker. The node in the Southern part of America (panel C) is not deeply related to the confining Brazil region, but rather with the Caribbean America as well as with the Indonesian archipelago and Australia. Regions which are fairly linked with this node are the African Sahel, Western Russia and the Mongolian area. As a last example of shortest path length, we consider a node in Eastern Asia (panel D), which is a region with the highest weighted average topological distance (see the top panel of Fig. 4 ). The Eastern Asia node is topologically well connected only to those nodes which are also physically close to it. The whole European and American continents, which are physically distant, are furthermore vaguely linked to this node. In this case only, the topological distance is somehow related to the physical distance.
10.1371/journal.pone.0071129.g007 Figure 7
Shortest path, , of 4 significant nodes of the network. The measured nodes are represented in each panel by a pink square. (A) Northern Europe, node coordinates: ( E, N). (B) Southern Europe, node coordinates: ( E, N). (C) South America, node coordinates: ( W, S). (D) Asia, node coordinates: ( E, N).
Some of the patterns linking mid-latitude to tropical nodes (e.g., North-American and European nodes with Sahel, Indonesian and Central America nodes) seem to suggest the impact of the propagation of planetary waves on precipitation at a global scale level [53] . Beside the influence of stationary planetary waves on the precipitation at local scale [54] , [55] , it was recently found that extreme events simultaneously occur worldwide in concomitance with the amplification of trapped planetary waves [56] .
Nevertheless, some links can be qualitatively related to the oceanic and atmospheric circulation [51] . In addition to the Gulf Stream and the NAO effects revealed by the shortest path of Northern Europe node ( Fig. 7A ), the South-Equatorial Current (linking the Pacific Coast of South America to Eastern Australia and Indonesian Archipelago) and the Brazil Current (linking the Atlantic Coast of South America to the Atlantic Coast of Africa) can be individuated for the South America node ( Fig. 7C ). The Australia node (see Fig. S1 ) is affected as well by the South-Equatorial Current branch going from Australia to South Africa. The Africa node (see Fig. S2 ) is related to the Atlantic Coast of South America through the Brazil Current.
The different patterns expressed by the four nodes in Fig. 7 can be also observed through the physical area connected to each node as a function of the topological distance, see Fig. 8 . Although the European trends are similar, the Southern node has significantly lower values in the range . Moreover, there is a striking difference between the South America and Asia nodes. In fact, for , the area connected to the South-American node is up to six/seven times larger than the area linked to the Asia node.
10.1371/journal.pone.0071129.g008 Figure 8
Physical area connected to a node as a function of the topological distance, . The four nodes of Fig. 7 are displayed.
In ( Text S1 ), an animated representation of the shortest path - linking a node to the rest of the network - is displayed for the nodes presented in Fig. 7 (see Movies S4 , S5 , S6 , S7 ) and for other meaningful nodes of the network (see Figures S1 , S2 , S3 and Movies S1 , S2 , S3 ).
We conclude this section offering a possible climatological interpretation of two measures, the weighted average topological distance, , and the betweenness centrality, , which both rely on the concept of shortest path. As in complex network theory these two parameters are used to measure the information flow, here they should be intended as indexes of short- and long-range connections. Nodes with low values are in general connected on a larger topological scale where precipitation varies more uniformly (e.g., Sahel region, Eastern Australia, Western Europe), while high values describe regions with short-range connections, due to their higher precipitation variability (e.g., Southeastern Asia). The interpretation of the betweenness centrality is not so straightforward since it represents a mediator of both long- and short-range connections, which equally contribute to the final value of a node they pass through. As a consequence, the information on short- and long-range connections is partially lost. This is the reason why the betweenness centrality map is spotted, without remarkably patterned regions, and it is not very meaningful for the precipitation network.
Conclusions
The recent development of complex network theory is offering new quantitative tools to disentangle the global climate dynamics. In spite of teleconnections having long been studied in climatology, the idea to read climatic correlations among different Earth regions as forming a complex network is relatively new. Starting from this point of view, we have focused on the global precipitation network. To this aim, we have used reliable datasets of measured land-surface precipitation that have only recently become available. Paying attention not to introduce spurious correlations due to uneven partitions of the Earth surface and to have a sufficient number of measured data in each cell, correlation analysis (with cut-off at 0.5) performed on a 70 year-window has yielded an undirected and symmetric network with 1674 nodes and 9481 links. We have investigated the structure of this precipitation network by some topological properties of nodes (degree centrality, local clustering coefficient, etc.) and, in particular, we have introduced a weighted form of the average topological distance in order to prevent some misleading behaviors when the graph has disconnected components.
Some key aspects of the precipitation network clearly emerge. Firstly, supernodes (i.e., highly connected nodes) occur in the Sahel region in Africa and in Eastern Australia, and a scaling-law behavior is revealed in the node degree distribution. Sahel and Eastern Australia regions are some of the most arid areas in the world. Very low rainfall is uniformly distributed on continental scales and huge extreme events are rare. As a consequence, the precipitation gradient tends to weaken, making these regions well connected on a large spatial scale. This long-range connection is confirmed by the fact that Sahel and Eastern Australia remain supernode regions also in the long-range network (see Fig. 5 ).
Strongly connected zones are evident also in Northern Europe, Central Asia, Western US, and Northeastern Brazil, even if they are not always characterized by a high betweenness centrality. Remarkably, strong connection differences occur between Northern and Mediterranean Europe as well as between Western and Eastern US. The European differences can be explained by the fact that Northern regions are mostly influenced by the oceanic circulation of the Gulf Stream and the atmospheric effects of the NAO (as observed in Fig. 7A ), while Southern Europe is more affected by the Mediterranean, Saharan and Caucasian circulation (as shown in Fig. 7B ). Differences for the US connections reflect the emergence of high precipitation gradients, which are partially due to frequent extreme events, and the ENSO influence on Western [32] , [57] and Eastern [58] US precipitation. Nevertheless, Eastern and Western Coasts are reached by very different ocean patterns: the Gulf Stream on the Atlantic Coast and the North Pacific Current on the Western Coast.
A second key point is the high topological distance of the Pacific region of the Asian continent with respect to the rest of the network. Notice that this behavior is not spuriously due to a geographic reason - namely, the decrease of lands above sea level close to the considered region - but it seems a peculiar feature of the precipitation network. A possible explanation of this network characteristic is the dramatic emergence of monsoons, tropical cyclones and heat waves which reduce the correlation only the to short-range scale. A similar disconnected behavior, even if less marked, is also observable for some coastal regions on the Southern Mediterranean and in Ethiopia. It should be noted as well that islands such as Iceland, Madagascar, New Zealand, and Japan, are completely disconnected from the rest of the network, forming different micro-networks. In fact, islands, which are in general more subject to extreme events [59] , are not reached by the influence of the continental land mass and are furthermore exposed to the specific oceanic currents.
Finally, the shortest path distribution proved to be a powerful tool to unveil how information about precipitation in a node is linked to the network. Some geographic regions are embedded in very connected portions of the network and a few jumps between neighboring nodes are sufficient to cover large geographic regions; vice versa , other nodes turn out to be quite isolated from the rest of the network. The same tool is also useful to highlight the geography of this information flux about precipitation. For example, we have shown the case of two nodes in Europe that, in spite of their closeness, exhibit two rather different connection areas.
The reported results highlight that the complex network approach can be an useful framework to explore the huge amount of climatic data that have been collected in the last years and, then, shed light on climate dynamics.
Supporting Information
Text S1
Shortest path of significant nodes of the network.
(PDF)
Figure S1
Shortest path of the Australia node (coordinates: E, S) represented by a pink square.
(EPS)
Figure S2
Shortest path of the Africa node (coordinates: W, N) represented by a pink square.
(EPS)
Figure S3
Shortest path of the US node (coordinates: W, N) represented by a pink square.
(EPS)
Movie S1
Animated representation of the shortest path of the Australia node (coordinates: E, S) displayed in Fig. S1 . The node is indicated by a pink square.
(AVI)
Movie S2
Animated representation of the shortest path of the Africa node (coordinates: W, N) displayed in Fig. S2 . The node is indicated by a pink square.
(AVI)
Movie S3
Animated representation of the shortest path of the US node (coordinates: W, N) displayed in Fig. S3 . The node is indicated by a pink square.
(AVI)
Movie S4
Animated representation of the shortest path of the Northern Europe node (coordinates: E, N) displayed in Fig. 7A of the main text. The node is indicated by a pink square.
(AVI)
Movie S5
Animated representation of the shortest path of the Southern Europe node (coordinates: E, N) displayed in Fig. 7B of the main text. The node is indicated by a pink square.
(AVI)
Movie S6
Animated representation of the shortest path of the South America node (coordinates: W, S) displayed in Fig. 7C of the main text. The node is indicated by a pink square.
(AVI)
Movie S7
Animated representation of the shortest path of the Asia node (coordinates: E, N) displayed in Fig. 7D of the main text. The node is indicated by a pink square.
(AVI)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Replication-deficient viral vectors such as recombinant adeno-associated virus (rAAV) are increasingly being used to introduce ‘therapeutic’ genes into neural cells, a method that allows targeted supply of neuroprotective and/or growth-promoting molecules to the injured or degenerating CNS [1] – [5] . In the eye, vitreal injection of rAAV serotype 2 (rAAV2) or other viral vectors encoding growth factors increases retinal ganglion cell (RGC) survival and axonal regeneration after optic nerve (ON) injury [6] – [9] .
The gene therapy approach holds much promise for the treatment of neurodegenerative diseases and retinal dystrophies [10] – [13] , as well as potentially enhancing repair after neurotrauma [14] . However, what is not yet clear is the extent to which long-term constitutive expression of transgenes changes the structure and function of transduced neurons. This is especially relevant when using genes that encode, for example, secretable neurotrophic factors because these peptides are known to alter dendritic architecture, synaptic density and plasticity, cause down-regulation of cognate receptors and modulate activity of signaling molecules [15] – [19] . Thus persistent over-expression of some transgenes may alter local circuitry and neuronal responsiveness to endogenous neuroactive factors [1] , [20] , [21] .
We therefore set out to determine whether soma size and dendritic architecture is altered after prolonged rAAV2 vector transduction, and whether any such changes depend on the type of gene that is introduced into CNS neurons. Because rAAV based gene therapy will potentially be used in post-injury as well as neurodegenerative conditions, we chose to use our established visual system regeneration model to quantitatively analyze changes in adult rat RGCs that had been axotomized and then induced to regenerate. RGC viability and long-distance axonal regeneration was promoted by grafting an autologous peripheral nerve (PN) segment onto the cut ON [22] , [23] . Four vectors were tested: rAAV2-GFP alone, and bi- cis tronic rAAV2 vectors encoding either brain-derived neurotrophic factor (rAAV2-BDNF-GFP), a secretable form of ciliary neurotrophic factor (rAAV2-CNTF-GFP), or a non-secreted protein growth-associated protein 43 (rAAV2-GAP43-GFP). A saline-injected control group was also included. BDNF, CNTF, and GAP-43 have all been shown to influence adult RGC viability and axonal regeneration [23] , and the impact of each of these genes when encoded in rAAV vectors has been documented previously in rat PN-ON graft studies [8] , [9] .
Five to eight months after the initial surgery, regenerated RGCs were identified by retrogradely labeling them with fluorogold (FG) injected into the distal end of each PN graft. Living retinas were subsequently removed, wholemounted, and regenerated FG positive ( + ) RGCs were intracellularly injected with Lucifer Yellow (LY). Expression of GFP in all rAAV vectors permitted identification of transduced, regenerated RGCs. GFP + and non-transduced GFP negative ( − ) RGCs that had regrown an axon into the PN graft were filled. After immunoprocessing for LY, soma size and dendritic morphology were analyzed and quantified using Neurolucida software. We observed gene-specific changes in the morphology of identified, regenerating adult RGCs after long-term rAAV2 therapy, not only in transduced RGCs but also in non-transduced RGC populations. Furthermore, some changes appeared to be subtype specific, seen only in large, type RI-like RGCs.
Results
Impact of transgenes on RGC morphology
FG + RGCs that had regenerated an axon to the distal end of PN grafts were identified and photographed under UV light ( Fig. 1A, G ) and also at 488 nm to determine whether the RGCs were transduced (GFP + ) or non-transduced (GFP − ; Fig. 1B, K, L ). In all groups, transduced RGCs were most frequently seen in temporal retina, in the vicinity of the initial rAAV2 vitreal injection ( Fig. 1D ) [14] , [24] . RGCs were injected iontophoretically with LY and a photograph taken under 488 nm ( Fig. 1C, I, M ) for subsequent identification of individual RGCs from immunolabelled and Neurolucida traces ( Fig. 1E, F, J, N ). This procedure was repeated on 20–50 RGCs per retina ( Fig. 1D ). The number of regenerated FG + RGCs that were analyzed in each control or vector group is shown in Table 1 . For clarity, throughout the text for each of the 4 vectors used in this study we will denote transduced and non-transduced (nt) FG + RGCs as GFP/ntGFP, BDNF/ntBDNF, CNTF/ntCNTF or GAP43/ntGAP43 respectively.
10.1371/journal.pone.0031061.g001
Figure 1
Injection, tracing and identification of individual RGCs from rAAV2-injected retinae.
Photomicrographs and Neurolucida-traced images showing the procedure for labeling and identifying regenerating retinal ganglion cells (RGCs). Regenerating RGCs were first identified based on retrograde labeling (Fluorogold; A), and transduced cells were based on GFP expression (GFP + RGC is shown in B). After filling the dendrites with Lucifer yellow (C), the RGC was again photographed. This procedure was repeated on 20–50 RGCs per retina (D). The visualization of dendritic architecture was further enhanced with Lucifer yellow immunohistochemistry, and individual cells with complete fills (E) were traced using Neurolucida software. The Neurolucida trace (F) was compared to images of the cells taken immediately after the Lucifer yellow injection (C) to allow each cell to be classified as transduced (GFP + ) or as a non-transduced “bystander” neuron (GFP − ). G–N: Representative images of RGCs that were retrogradely labeled with Fluorogold (G, K), identified as GFP − (H) or GFP + (L), injected with Lucifer yellow (I,M) and traced using Neurolucida software (J,N). Scale bars: 100 µm.
10.1371/journal.pone.0031061.t001 Table 1
Number of RGCs analyzed in each control and experimental group for all RGCs, and for those identified as Type RI cells.
All cells
Type RI
Saline
68
17
BDNF
73
40
ntBDNF
14
7
BDNF
59
33
CNTF
117
42
ntCNTF
62
22
CNTF
55
20
GAP43
63
19
ntGAP43
16
4
GAP43
47
15
GFP
54
20
ntGFP
14
5
GFP
39
15
Total analysed
375
138
No axon/FG-
66
Incomplete fill
59
Grand Total
500
In some vector groups, a significant proportion of the 375 fully analyzed RGCs possessed one or more highly abnormal dendritic morphologies, including either very sparse dendrites or unusually tangled processes ( Fig. 2A ). The proportion of RGCs with such abnormal dendritic morphologies was, compared to saline, not significantly different in rAAV2-GFP and rAAV-GAP43-GFP groups, but was significantly increased in the rAAV2-BDNF-GFP and rAAV2-CNTF-GFP injected groups (X 2 = 130; p<0.0001; Fig. 2B ).
10.1371/journal.pone.0031061.g002
Figure 2
Increased prevalence of RGCs with abnormal morphology in retinae injected with rAAV2-BDNF-GFP and rAAV2-CNTF-GFP.
A: Representative Neurolucida traces showing retinal ganglion cells (RGCs) with abnormal morphology. The term ‘aberrant morphology’ was used to describe RGCs with one extremely long dendrite that was exceptionally tortuous or abnormally sparse or asymmetric. Arrows indicate axons. The rAAV2 treatment group is indicated under each cell; nt = non-transduced RGC. B: Pie charts showing frequency distribution of RGCs with aberrant morphology in control and experimental groups.
Discriminant analysis (summarized in Table 2 )
10.1371/journal.pone.0031061.t002 Table 2
Summary of main dendritic morphological changes of long-term gene therapy affected retinal ganglion cells (RGCs) compared to appropriate GFP transduced or non-transduced control groups.
ALL RGCS
ALL RGCS
TYPE I-LIKE RGCS
BDNF
Increased soma area (P<0.0001), longer dendrites (p = 0.02), larger dendritic field (p = 0.003), increased total and average nodal distance (p = 0.008; p = 0.03). Sholl analysis: denser dendrites. Deeper stratification (p = 0.004). Increased prevalence of aberrant morphology.
BDNF transduced
Increased soma area (p<0.0001), increased field size (p = 0.003), deeper stratification (p = 0.049).
Increased soma area (p = 0.04)
BDNF non-transduced
No change
Increased soma area (p = 0.04), increased field area (trend: p = 0.063), deeper stratification (p = 0.01).
CNTF
Increased soma area (p<0.0001). Sholl analysis: denser dendrites. Deeper stratification (p = 0.003). Increased prevalence of aberrant morphology.
CNTF transduced
Increased soma area (p<0.0001), deeper stratification (p = 0.02).
Increased soma area (p = 0.0002), deeper stratification (p = 0.02).
CNTF non-transduced
Increased soma area (p<0.0014), deeper stratification p = 0.02
Increased soma area (p = 0.001), reduced complexity (nodes (p = 0.006), branch order (p = 0.03), increased terminal/nodal distance (total and mean: p = 0.05), deeper stratification (p = 0.03).
GAP43
Increased complexity (fractal count: p = 0.02). Increased dendritic density (p = 0.02). Deeper stratification p = 0.01
GAP43 transduced
Increased complexity (fractal count: p = 0.04). Deeper stratification (p = 0.002).
Increased tortuosity (p = 0.01). Deeper stratification (p = 0.0009)
GAP43 non-transduced
Increased dendritic density (p = 0.03)
No change
To quantify the morphological changes in regenerate RGCs induced by transgene expression, we applied multivariate statistical analysis to take into account all of the 15 parameters that were measured using Neurolucida (see Materials and Methods for further detail). We used discriminant analysis, a statistical technique used for differentiating groups using multiple quantitative variables [25] . The analysis extracts canonical scores, which represent a transformation of the original measurements into an expression of maximal differences between groups. Where significant differences were detected between the canonical scores of experimental and control groups, we then performed post-hoc analysis of the morphological measurements to determine the nature of these differences.
Discriminant analysis of the 5 treatment groups confirmed significant differences between all treatment groups. Most of the differences relative to control groups (saline and rAAV2-GFP injected retinae) were contained in canonical scores 1 and 2 ( Fig. 3A ). In addition, canonical score 3 demonstrated that RGCs from rAAV2-GFP transduced retinae differed from saline injected controls (p<0.0001; Fig. 3B ). Because the canonical scores represent a transformation of the data from the morphological parameters they cannot be attributed to specific morphological differences. We therefore performed ANOVA and posthoc analysis on all of the measurements to identify how GFP expression affected morphology. However, no individual parameter was found to be responsible for this difference either in transduced or non-transduced cells, suggesting the effects of GFP on RGC morphology were relatively minor. Nonetheless in all of the following analyses we compared treatment groups to rAAV2-GFP controls and not to saline injected controls.
10.1371/journal.pone.0031061.g003
Figure 3
Evidence for morphological differences in RGCs from retinae injected with rAAV2 encoding different transgenes.
A: Plot showing canonical scores 1 (Y axis) and 2 (X axis) from a multivariate discriminant analysis of dendritic morphology of all retinal ganglion cells (RGCs). Plots show the first two canonical scores that together represent more than 80% of the variance. Axes represent units of standard deviation. Circles represent the 95% confidence region to contain the true mean of the treatment groups. Black lines show the coordinate direction (i.e. morphological parameters measured in Neurolucida) in canonical space. Note that the length of the lines is not representative of effect size due to the multidimensional nature of the analysis. B–D: Box plots showing median and quartiles for selected morphological parameters that were significantly different between treatment groups. B: values for the third canonical score which accounts for the significant difference between Saline and rAAV2-GFP injected groups. Means for soma area (C) and dendritic field area (D) are shown for all treatment groups. * p<0.05; **p<0.001; *** P<0.0001. E: Sholl analysis of all RGCs by treatment group. Error bars = standard error of the mean. Asterisk (*) indicates significant (p<0.05) difference between rAAV2-BDNF-GFP and GFP; # indicates significant (p<0.05) difference between rAAV2-CNTF-GFP and GFP.
In rAAV2-BDNF-GFP injected retinae, the soma size of all LY-injected, regenerating RGCs was significantly increased compared to LY labeled, FG + RGCs in rAAV2-GFP controls (P<0.0001; Fig. 3C ). In addition, dendrites were longer (p = 0.02), field size was larger (p = 0.003; Fig. 3D ), and total and mean nodal distances were longer (p = 0.008 and p = 0.003 respectively). When transduced and non-transduced RGCs were analyzed separately, most of the changes could be attributed to transduced BDNF neurons, which had larger somata (P<0.0001) and larger field size (p = 0.03) compared to their GFP transduced counterparts. No significant differences were observed in regenerate ntBDNF compared to ntGFP RGCs.
In rAAV2-CNTF-GFP injected retinae, the only difference compared to rAAV-GFP injected retinae was a significant increase in the soma size of RGCs (p<0.0001). The increase was observed in both CNTF and ntCNTF RGCs compared to GFP (p<0.0001) and ntGFP (p = 0.0014) RGCs respectively. In addition, CNTF RGCs had significantly larger soma size compared to ntCNTF RGCs ( Table 2 ; p = 0.0085).
For RGCs in rAAV2-GAP43-GFP injected retinae, the fractal count [26] was significantly increased compared to rAAV2-GFP controls (p = 0.02), and dendrite density was also increased (p = 0.02). Analysis of transduced versus non-transduced neurons revealed that the difference in fractal count was found only in transduced GAP43 RGCs (p = 0.04), whereas an increase in dendritic density was found only in ntGAP43 RGCs (p = 0.03).
Sholl analysis
To further characterize the morphological changes induced in RGCs by the different transgenes, we also performed a Sholl analysis, the most commonly used method to measure dendritic field density and structure [27] . The method assesses the distribution of dendrites as a function of eccentricity using the cell body as the centre. Analysis of the total cell population revealed that RGCs in rAAV2-BDNF-GFP and rAAV2-CNTF-GFP injected retinae had significantly denser dendrites from 200 µm to 400 µm (BDNF) and from 320 µm to 400 µm (CNTF) distal from the cell body compared to rAAV2-GFP controls ( Fig. 3E ). No significant differences were detected when transduced and non-transduced RGCs were analyzed separately.
Assessment of morphological changes in one putative RGC subtype
The analysis above demonstrates that each transgene had a distinct and significant impact on dendritic morphology of RGCs. However, it is well established that there are multiple different subtypes of RGCs within the rat retina with characteristic morphologies [28] , [29] . It is therefore desirable to analyze these subtypes separately to confirm that morphological differences between subtypes have not masked differences due to transgene expression. However, unlike in mouse, in which there is a correspondence of specific molecular markers with RGC cell subtype [30] , [31] , in rat there is no definitive method of classifying all RGC subtypes independent of morphology. The only phenotypic marker consistently used in rat RGCs is melanopsin, which is expressed in about 3% of RGCs. However, in the specific context of RGC axonal regeneration into a PN-ON graft, murine melanopsin expressing RGCs do not regenerate more frequently than other RGCs [32] , thus in our rat study we would expect only about 10–15 regenerate RGCs to be melanopsin positive, clearly too few for any justifiable analysis.
Nonetheless, there is consensus across many studies that rat RGCs can be effectively classified into three or four main subtypes based on cell body size, number of dendrites, dendritic field size and stratification [29] , [33] – [36] . Most importantly, this morphological classification has also been applied to RGCs in peripheral nerve grafted animals [37] , [38] , suggesting that regenerating RGCs maintain core structural features of their subtype despite injury-induced alterations in dendritic architecture [39] . Consistent with this, the outcome of our whole population analysis of the 375 LY-labeled RGCs did not reveal any overall change in the number of primary dendrites emanating from RGCs in any vector group. Using only morphological criteria, the best consensus across all studies is the identification of the RGC 1 subtype, known as RI in regenerating RGCs [37] , [38] . These cells have a large soma, 4–6 primary dendrites and large dendritic field area with a very typical dendritic branching pattern [29] , [36] . These same criteria identify type RI RGCs following a PN graft [37] , [38] . Types 2 and 3 have smaller cell bodies than type 1 cells, fewer than 3 primary dendrites, and are differentiated primarily by the number of branch points but this measure becomes unreliable in regenerating RGCs in which branching density is greatly reduced [37] .
Taking this previous literature into account, we conservatively assessed whether it was possible to identify RI-like RGCs within our general population using the number of primary dendrites, cell body size and typical dendritic branching pattern as the essential criteria. We identified all RGCs that had large somas, more than 4 primary dendrites and a “typical” type RI dendritic branching pattern (examples are shown in Fig. 4A ). All RGCs from each treatment group were then plotted on separate scatterplots with the number of dendrites on the X axis and cell body size on the Y axis. RGCs that were identified as “RI-like cells” are shown as grey circles to identify them relative to the other cells (white diamonds; Fig. 4B ). RI-like cells clustered together to the upper right of the scatterplot for all treatments, suggesting that they formed a coherent group. Most importantly, even though BDNF and CNTF RGCs had consistently larger cell bodies than the other treatment groups, the clustering was still apparent. To further confirm that our identification was robust, we performed a discriminant analysis on all RGCs, where cells were either classified as “RI” or “other”, and these cells were found to cluster into two distinct groups, regardless of treatment ( Fig. 4C ).
10.1371/journal.pone.0031061.g004
Figure 4
Classification of RI-like RGCs.
A: Neurolucida traces of representative RI-like retinal ganglion cells (RGCs) from control and experimental rAAV2 groups. Transduced and non-transduced (nt) FG + RGCs in the 4 rAAV2 groups are labeled as GFP/ntGFP, CNTF/ntCNTF, BDNF/ntBDNF or GAP43/ntGAP43 respectively. B: Clustering of transduced and ntRI-like RGCs based on the number of primary dendrites (X-axis) and soma area (Y axis) in control and experimental groups. RI-like RGCs are denoted by grey circles and remaining cells by white diamonds. C: Plot showing canonical scores 1 (X axis) and 2 (Y axis) from a multivariate discriminant analysis of dendritic morphology of RI-like RGCs and “other” RGCs. Plots show the first two canonical scores that together represent more than 80% of the variance. Axes represent units of standard deviation. Solid black circles represent the 95% confidence region to contain the true mean of the group. Black lines show the coordinate direction (i.e. morphological parameters measured in Neurolucida) in canonical space. Note that the length of the lines is not representative of effect size due to the multidimensional nature of the analysis.
Analysis of RI-like RGCs
Plots showing canonical scores 1 (X axis) and 2 (Y axis) from multivariate discriminant analysis of dendritic morphology for transduced and non-transduced RI-like RGCs in each vector group are shown in Figure 5A ; the data are summarized in Table 2 . As seen in Figure 5B , BDNF and ntBDNF RI-like RGCs possessed soma areas that were significantly increased compared to GFP and ntGFP cells respectively (p = 0.04 for both). In ntBDNF RGCs, there was also a strong trend for a larger dendritic field area compared to ntGFP cells (p = 0.06; Fig. 5B ). In addition, transduced and non-transduced RGCs in rAAV2-BDNF-GFP injected eyes were significantly different from each other, ntBDNF RGCs having a larger dendritic field area compared to BDNF RGCs.
10.1371/journal.pone.0031061.g005
Figure 5
Evidence for morphological differences in type RI-like RGCs from retinae injected with rAAV2 encoding different transgenes.
A: Plots showing canonical scores 1 (X axis) and 2 (Y axis) from multivariate discriminant analysis of dendritic morphology for RI-like retinal ganglion cells (RGCs). Plots show the first two canonical scores that together represent more than 80% of the variance. Axes represent units of standard deviation. Circles represent the 95% confidence region to contain the true mean of the group. Black lines show the coordinate direction (i.e. morphological parameters measured in Neurolucida) in canonical space. Note that the length of the lines is not representative of effect size due to the multidimensional nature of the analysis. B: Box plots showing median and quartiles for selected morphological parameters that were significantly different between treatment groups. Transduced and non-transduced (nt) FG + RGCs in the 4 rAAV2 groups are labelled as GFP/ntGFP, CNTF/ntCNTF, BDNF/ntBDNF or GAP43/ntGAP43 respectively. Asterisk (*) indicates groups that are significantly different from ntGFP RGCs (p<0.05) and # indicates groups that are significantly different from GFP RGCs (p<0.05).
In CNTF and ntCNTF RI-like RGCs, soma area was also significantly increased compared to GFP and ntGFP cells respectively ( Fig. 5B ). In addition, dendrites of ntCNTF RGCs were less complex than those of ntGFP cells, with fewer nodes and reduced branch order, as well as increased total and mean terminal/nodal distances, suggesting growth of terminal segments. As observed in rAAV2-BDNF-GFP injected eyes, transduced and non-transduced RGCs in rAAV2-CNTF-GFP injected eyes were significantly different; compared to CNTF RGCs, ntCNTF RGCs had significantly fewer nodes and lower order branching, confirming the loss of complexity in these non-transduced, or bystander, neurons.
Morphological differences in RI-like RGCs from rAAV2-GAP43-GFP injected eyes were attributed to an increase in tortuosity in GAP43 transduced cells compared with GFP transduced controls ( Fig. 5B ). Note here that Sholl analysis of RI-like RGCs failed to reveal any significant differences (data not shown).
Stratification of dendritic tree
For the majority of FG + RGCs, the dendritic tree extended within the inner half of the inner plexiform layer (on average within 30 µm of the cell body) suggesting that they were ON centre cells [40] . A small number of cells appeared to be either OFF or bistratified (ON/OFF) cells but these were not analyzed statistically due to low numbers (0–3 cells per treatment group).
Within the population of presumed ON cells, sampled across a similar range of retinal eccentricities, the depth profile of dendritic trees was increased in all treatment groups, and this was mainly due to branches extending to an abnormal depth within the IPL (BDNF: p = 0.004; CNTF: p = 0.003; GAP43: p = 0.01; all compared to GFP controls Fig. 6A–C ). BDNF affected only transduced RGCs (p = 0.049), CNTF affected both transduced and non-transduced neurons (p = 0.02 for both), and GAP43 affected only transduced RGCs (p = 0.002) compared to appropriate transduced or non-transduced GFP controls. In the population of RI-like cells, RGCs in rAAV2-BDNF-GFP injected eyes were not affected, whereas dendrites of rAAV2-CNTF-GFP transduced and non-transduced, and rAAV2-GAP43-GFP transduced RGCs ramified more deeply (p = 0.02; p = 0.03; p = 0.0009 compared to appropriate transduced or non-transduced GFP controls). In RGCs with grossly abnormal dendritic morphology, while dendritic branches were occasionally seen at the border of the inner plexiform and inner nuclear layers (INL), these processes were not seen to penetrate the INL itself ( Fig. 6D, E ).
10.1371/journal.pone.0031061.g006
Figure 6
Abnormal stratification in RGCs from retinae injected with rAAV2 encoding different transgenes.
A–C: Side views of Neurolucida traces showing examples of abnormal stratification in each rAAV2 treatment group. Arrows indicate axons and “s” is immediately below the soma. The scale bar in A applies to the three traces. D: Neurolucida trace of an rAAV2-GAP43-GFP transduced retinal ganglion cell (RGC). Ai and Aii are side views. E: confocal image of the same cell as in D; pink is Cy3 immunofluorescence for lucifer yellow, wholemount counterstained with Hoechst 33342 (blue). Areas a and b are shown in more detail. Area a is shown in panels ai-aiv. Panel ai shows a higher power magnification of a and panels aii and aiii show the XZ and YZ projections respectively at the crosshairs shown on panel ai. Panel aiv shows the full YZ projection of panel a and shows that dendritic processes descend through the inner plexiform layer and run along the border of the inner nuclear layer (INL), but they do not enter the INL. Area b is shown in panel bi as an XZ projection showing the relationship of this dendritic branch to the INL. GCL, ganglion cell layer. Scale bar in ai: 25 µm.
Discussion
The potential clinical benefits of using gene therapy to deliver growth factors to treat neurological or retinal dysfunction are currently under investigation [10] , [13] , [41] – [45] . Post-injury delivery of appropriate viral vectors may also be an effective treatment after neurotrauma [14] . Depending on circumstances, such therapies may need to be long term and it therefore seems prudent to examine whether there are any unintended consequences of sustained delivery of such factors on neuronal structure and function. We have shown previously that intraocular injection of rAAV-BDNF-GFP or rAAV-CNTF-GFP influences adult RGC survival and regeneration after injury [8] , [9] . Here we show that, in FG + RGCs that had regrown an axon into PN grafts and therefore had comparable access to factors expressed by the grafted PN tissue, intravitreal delivery of vectors encoding BDNF, CNTF or GAP43 nonetheless resulted in significant and complex changes in the dendritic morphology of the total population of regenerate neurons.
Each factor induced specific structural changes but overall, BDNF and CNTF increased RGC soma size, while BDNF increased dendritic field size and CNTF and GAP43 altered dendritic complexity. These changes did not obviously restore regenerating RGCs towards a more “normal” morphology, but rather added to the effects that were induced post-injury; thus BDNF further increased dendritic field size and CNTF reduced complexity in RGCs whose dendritic arbors had become larger [46] and less complex [37] following optic nerve lesion and PN transplantation. Furthermore, all three transgenes induced abnormal dendrite growth that was not restricted to normal sub-laminae within the IPL. A significant finding was that the morphologies of non-transduced (GFP − ), FG + RGCs in rAAV injected eyes were also differentially affected, thus the total impact of a given transgene is multifarious and extends to bystander neuronal populations [47] .
As described in the Results, others have argued that it is possible to recognize and classify at least some rat adult RGC subtypes that are regenerating an axon into a PN graft [37] , [38] . Regenerating type RI RGCs in particular appear to display the most characteristic and convincing similarities to this same class in normal retina. We identified all RGCs that had large somas, more than 4 primary dendrites and a “typical” type RI dendritic branching pattern and discriminant analysis showed that these cells were always clustered in a separate group, irrespective of vector treatment. The impact of rAAV-CNTF-GFP injections was most obvious in these RGCs, with a significant reduction in the complexity of the dendritic arbors. While additional phenotypic markers for RGC subtypes, other than morphology, would have been helpful in these experimental animals, few such markers are available for rat RGCs. Importantly, characterization of regenerating, type 1-like RGCs has been reported in other species [48] – [50] and in the cat a number of physiological studies on RGCs regenerating axons into PN grafts unequivocally confirms the phenotype of morphologically characterized alpha cells as being Y-type in character although with some altered receptive field properties [51] , [52] .
Growth factors and dendritic morphology
BDNF
Studies of nervous system development and regeneration reveal pleiotropic effects of growth factors on dendritic morphology [17] , [19] , [39] , [53] , [54] . It is generally accepted that BDNF increases dendritic field size and complexity of branching [55] , [56] , although BDNF may also cause a loss of dendrite complexity [57] , [58] . Consistent with these complex effects, we found that RGCs were differentially affected depending on whether RGCs were transduced or non-transduced. Dendritic field area was consistently increased in RGC within rAAV2-BDNF-GFP injected eyes, and was accompanied by increased dendritic length, specifically due to longer nodal segments, suggesting interstitial growth [46] . Perhaps surprisingly, the increased growth was not accompanied by changes in branch density, perhaps due to altered dynamics of loss and formation of dendritic branches; in Xenopus tadpoles, target-derived (tectal) BDNF increases, whereas local (retinal) BDNF decreases, RGC dendrite complexity [58] . The different effect of BDNF depending on whether it is local or target-derived is relevant because in the present study the adult RGCs regenerated into blind-ended PN grafts and thus were exposed to factors expressed in the grafts but they could not re-connect with central targets. The resulting loss of balance between retinal and target BDNF may have contributed to the non-characteristic growth patterns we observed.
A key finding was the effect of rAAV2-BDNF-GFP injections on both transduced and non-transduced, regenerating RI-like RGCs. Interestingly, the effects were most pronounced in the non-transduced, bystander population. BDNF secreted from single cells within brain slices of immature cortex has been shown to act as an “intercellular morphogen”, increasing dendritic growth in neighbouring neurons [16] , [59] . There are many possible mechanisms whereby secreted, transgene-derived BDNF might exert its effects on RGCs: BDNF effects can be mediated by full-length (TrkB-FL) and truncated (TrkB-T1) receptors, and by the p75 receptor. TrkB-FL promotes dendritic growth and complexity via recruitment of PI3-kinase and perhaps MAP kinase signaling pathways [17] , [56] although in our model TrkB-FL signalling may be less relevant because the receptor may have been down-regulated in response to sustained high levels of BDNF [60] . TrkB-T1 has slightly different effects than TrkB-Fl in that it increases dendrite growth in regions distal to the soma and inhibits proximal branching, at least in cortical pyramidal neurons [61] . In addition, BDNF signalling via p75 may also be involved because in hippocampal neurons, p75 overexpression reduces dendrite complexity [57] and NGF activation of p75 increases dendrite length [62] .
CNTF
Sustained expression of CNTF in regenerating RGCs was associated with increased cell body size in all RGCs, but increased aberrant dendritic growth and a loss of dendritic complexity were detectable only in RI-like cells. Interestingly, changes in dendritic architecture were most pronounced in non-transduced RGC populations. Intravitreal delivery of rAAV2-CNTF-GFP results in extensive elongation of RGC axons [8] , [9] but comparatively little is known about the impact of CNTF on dendritic architecture, although CNTF and leukemia inhibitory factor (LIF) have been reported to induce dendritic retraction in cultured sympathetic neurons [15] . The actions of CNTF, LIF and other cytokines are regulated by suppressor of cytokine signaling (SOCS) molecules and SOCS3 deletion enhances RGC axonal regeneration [63] . We previously reported that intravitreal CNTF injection results in a long lasting increase in SOCS3 expression in RGCs [64] . The more pronounced effects of rAAV2 mediated expression of the secretable form of CNTF on the dendritic morphology of non-transduced RGCs may reflect lower levels of SOCS expression in these bystander cells compared to transduced RGCs, the latter therefore having a reduced capacity to respond to the cytokine [24] .
GAP43
AAV-GAP43-GFP expression primarily affected dendritic complexity and branching, and is consistent with the influence of GAP43 on cytoskeletal structure and neurite/axonal growth [65] – [67] . There is also direct evidence that motifs found in the GAP43 protein regulate dendritic growth and branching in cultured hippocampal cells [68] . We observed significant changes in all RGCs, characterized by the development of denser and more complex dendritic trees with more tortuous dendrites. It is unclear how vector induced GAP43 protein expression affected field density of non-transduced RGCs, given that the protein is not normally secreted. However, GAP43 may promote secretion of other factors that alter the growth of neighboring neurons. One candidate is the protease nexin 1 (PN-1), a serine-protease and thrombin inhibitor expressed in glia and neurons in vivo [69] , [70] . Secreted PN-1 alters extracellular protease activity, influencing neuronal development [71] including neurite outgrowth [72] – [75] . GAP43 may influence PN1 secretion [76] , potentially altering the retinal environment and contributing to the dendritic changes described here. Note that PN1 has also been implicated in pathological situations [70] and in Alzheimer's disease [77] where abnormal dendritic morphologies are common.
Abnormal morphologies and stratification
Previous studies of RGC dendritic morphology following ON lesions with or without a PN graft have described a high proportion of “unclassifiable” cells [37] . Abnormal morphologies were most frequent in rAAV2-CNTF-GFP and rAAV2-BDNF-GFP injected retinae, suggesting that these two transgenes promote aberrant arborization and growth beyond what is normally observed following ON crush or PN transplantation [46] , [51] , [78] .
Subsets of lamina-specified RGCs are tuned to distinct visual features [40] , [79] and disruption to stratification leads to compromised visual function. As in the cat [52] , [80] , [81] , the majority of regenerate RGCs possessed dendrites characteristic of ON-responsive cells, however we commonly observed inappropriate extension of dendritic branches into deeper regions of the IPL. Abnormal stratification was seen more frequently in rAAV2-CNTF-GFP and rAAV2-GAP43-GFP injected eyes and resembled the dendritic trees in retinae after exposure to increased glutamate levels [82] , [83] .
Conclusions
While vector-mediated expression of secreted growth factors in neurons and other cells undoubtedly has beneficial effects on cell viability and regenerative growth, we now show that long-term overexpression of such transgenes alters the dendritic morphology of both transduced and non-transduced regenerating neurons, potentially altering the pattern and efficacy of the afferent synaptic input to these cells [84] . To determine whether similar transgene-induced changes are seen in normal RGCs we are currently completing a separate quantitative study examining changes in dendritic architecture of uninjured RGCs. Altered dendritic morphologies may affect the function of any conserved or reconstructed neural circuits, raising important questions for future study. In the visual system such changes may be a benefit or a hindrance to behavior; changes in dendritic field size and/or complexity in different RGC classes may increase detection capabilities in retinae with reduced RGC numbers but may also negatively impact upon visual acuity [52] . Altered dendritic architecture in other CNS regions would obviously be associated with different functional processing issues. While the time-course of the observed dendritic changes and how they relate to axonal regeneration remains to be determined, the present data suggest it is prudent to develop reliable systems that allow the effective regulation of transgenes, especially if gene therapy is to be used to provide neurotrophic support during the treatment and clinical management of neurodegeneration and neurotrauma.
Materials and Methods
Data were obtained from female Wistar rats aged 8–10 weeks at the time of rAAV2 injection and PN-ON surgery. Rats were purchased from the Animal Resources Centre (WA) Experimental work was approved by The University of Western Australia Animal Ethics Committee and conformed to national NHMRC guidelines.
AAV vectors
Each vector was commercially produced by GTC Virus Vector Core (NC, USA) and was generated from either the pTRUF12 plasmid (GFP, BDNF and GAP43) or the pTRUF12.1 plasmid (CNTF; gift of Prof. Joost Verhaagen). For bi- cis tronic rAAV2 vectors the relevant gene sequence also contained a post IRES site that also permitted subsequent GFP expression, hence resulting in the production of two individual proteins (the growth factor protein and the GFP protein as a marker for transduction). Expression was driven by the cytomegalovirus early enhancer chick-β-actin (CMV-CAG) promoter. The CNTF gene sequence included a nerve growth factor signal to allow secretion of the vector produced CNTF protein (gift of Prof. Sendtner). Due to the rAAV2 packaging size limitation this vector was based on the pTRUF 12.1 plasmid with a CMV-CAG promoter that lacked a promoter intron and thus provided sufficient space for the transgene [8] . Previous in vitro and in vivo studies using the CNTF, BDNF and GAP-43 rAAV-2 constructs have shown by Western blot, ELISA and immunohistochemistry that transduced cells express biologically active proteins and promote RGC survival and axonal regeneration [2] , [8] , [9] , [85] .
Surgery
Rats were anaesthetized with an intraperitoneal (ip) injection of a 1∶1 mixture of ketamine (100 mg/ml) and xylazine (20 mg/ml; 1 ml/kg). For vitreal injections, 4 µl of either saline, rAAV2-GFP, rAAV2-BDNF-GFP, rAAV2-CNTF-GFP or rAAV2-GAP43-GFP (n = 5 per group) was injected into the temporal part of the left eye via a glass micropipette inserted just behind the ora serrata, the pipette tip angled in order to avoid damage to the lens. All vector concentrations were 1×10 12 genome copies/ml. Seven to nine days later, rats were again anaesthetized (see above) and the ON was cut about 1.5 mm behind the eye and a segment of PN was grafted onto the stump to enhance regeneration [22] , [23] . The graft consisted of a 1.5 cm segment of autologous tibial nerve sutured onto the proximal stump of the cut ON with 10/0 suture (Ethilon; Johnson & Johnson, Australia). The distal portion of the PN was positioned over the skull and the end sutured to connective tissue using 6/0 suture. Care was taken to avoid damaging orbital blood vessels and the ophthalmic artery lying beneath the ON; vascular integrity of the retina immediately after this procedure was confirmed in each rat. Animals received a subcutaneous injection of buprenorphine (0.02 mg/kg, Temgesic; Reckitt & Colman, UK) and an intramuscular injection of Benacillin (0.1 ml, Troy Laboratories Pty. Ltd. Australia).
Retinal wholemount preparation
Grafted animals survived for 5–8 months before further analysis. The range of post-graft survival resulted from the fact that it was only possible to process a small number of animals at any given time. Intracellular RGC injections were done on 16 different days during that period. To control for any effect of post-operative survival time on RGC morphology, rats from different AAV groups were sampled across the 5–8 month range, 80% of animals sampled between 6 and 8 months after surgery. For example, PN grafted rats injected with rAAV2-CNTF-GFP (n = 5) were injected on weeks 26, 27, 30 (2 animals) or 33 after surgery while rAAV2-BDNF-GFP rats (n = 5) were injected on weeks 25, 26, 31, 32 or 33. Two to three days prior to sacrifice, rats were anaesthetized as above and RGCs were retrogradely labeled with 4% FG (0.5 µl) injected into the distal end of the PN graft, more than 1 cm beyond the suture point with the transected ON. Rats were sacrificed with pentobarbitone (150 mg/kg, ip) and the whole retina was rapidly removed from the eyecup in oxygenated AMES buffer and flat mounted RGC-side facing down onto a glass slide. A circle of black Millipore filter paper was lowered onto the retina. The tissue adhered to the filter paper and was turned over and placed RGC-side up in oxygenated AMES buffer.
Single cell injections
The retina was placed in a slide chamber, immobilized with a small weight and superfused with oxygenated AMES for the duration of the experiment. Glass micropipettes (resistance: 50–300 MΩ) were filled by capillary action with LY (Molecular Probes/Invitrogen; 2% in 0.1 M Tris buffer). RGCs were injected with LY by inserting the micropipette into the cell soma under the control of a micromanipulator. Injections lasted for 2–5 min until dendrites appeared completely filled. The micropipette was slowly removed and the filled RGCs visualized and photographed under UV light and also at 488 nm to determine whether regenerate FG + RGCs were transduced (GFP + ) or non-transduced (GFP − ) by the respective rAAV vectors ( Fig. 1 ). We note that post-IRES GFP expression in AAV vectors can be lower compared to GFP driven directly by a promoter [86] , thus it is possible that some apparently GFP − RGCs were transduced but the GFP was not discernable. However, based on GFP expression we previously determined the transduction efficiency of our bi- cis tronic vectors in normal rat RGCs and found little difference between these vectors and AAV-GFP [8] . Furthermore, the present quantitative analysis consistently revealed significant differences between transduced and non-transduced RGC populations, thus we argue that if any RGCs were incorrectly characterized as non-transduced, their number was very small. Between 20–50 RGCs were intracellularly injected per retina ( Fig. 1D ). In all groups, RGCs were most easily visualized and injected at mid-retinal eccentricities ( Fig. 1D ), thus there was only minimal sampling from central or peripheral retina.
Immunohistochemistry
At the end of each experiment, the retina was removed from the slide chamber, gently peeled off the filter paper and fixed in 4% paraformaldehyde for 2 hrs at room temperature in the dark. The fixed retinae were washed in PBS for 3×10 min and incubated overnight at 4°C with an anti-LY antibody (Santa Cruz Biotechnology), 1∶500 dilution in 1% triton-X-100, 1% BSA in PBS. Following a one hour wash in PBS at room temperature, retinae were incubated with a Cy3 goat-anti-rabbit antibody (Jackson Laboratories), 1∶300 in 1% triton, 1% BSA in PBS, for 4.5 hrs at room temperature in the dark. Retinas were then washed 3 times for 10 min in PBS and mounted in Citifluor, coverslipped and sealed with nail varnish. Slides were stored at 4°C until further analysis.
Quantitative and statistical analysis of retinal ganglion cell morphology
RGCs were manually traced directly from immunolabelled retinal wholemounts using Neurolucida software and analyzed in Neurolucida explorer. The experimenter was blinded to treatment group. In all groups, RGCs were sampled from a similar range of eccentricities. Use of a digitized stage allowed 3-dimensional measurements to be taken directly from the tissue without confocal imaging. The resulting Neurolucida traces contained quantitative data in the Z-axis allowing an estimation of dendritic stratification depth. In selected RGCs with abnormal dendritic trees, stratification relative to the inner nuclear layer was visualized by counterstaining wholemounts with Hoechst 33342 (Sigma) dye. Dendrites were traced at 40× magnification under oil immersion (Uplan Apo 40×/1.00 oil Iris).
Due to various technical problems it was not possible to process retinae from 4 of the 25 PN-ON grafted eyes. In the remaining 21 retinae (4 saline, 4 rAAV2-GFP, 4 rAAV2-BDNF-GFP, 5 rAAV2-CNTF-GFP and 4 rAAV2-GAP43-GFP) a total of 500 cells in the ganglion cell layer were injected and their dendrites manually traced using Neurolucida. Of these, 125 (25%) were discarded, either because (i) after fixation and processing some were found not to be FG + or did not have a clearly labeled axon projecting to the optic disk and therefore may not have been RGCs, (66 cells), or (ii) because the dendrites were extremely abnormal in appearance and it was not possible to be certain whether this was due to incomplete LY fills or genuine morphological changes (59 cells).
Analysis of morphology
As previously described [37] , we observed RGCs with morphologies comparable to those of normal intact cells [29] , as well as RGCs with one or more abnormal structures including either very sparse dendrites or unusually tangled processes. To determine whether different transgenes were more likely to induce these abnormal morphologies, we compared the proportion of cells in each treatment group with abnormal dendrites using a Chi squared test.
To further characterize differences between RGCs we performed multivariate analysis (Discriminant analysis; JMP) of key morphological RGC parameters [87] . Parameters measured for each cell were: soma area, number of nodes, number of primary dendrites, total dendrite length, dendritic field area, average segment tortuosity, maximum order branching, fractal count, dendritic density (dendritic field area/total dendritic length), total terminal distance, average terminal distance, total nodal distance, average nodal distance, total terminal/nodal distance and average terminal/nodal distance.
A multivariate analysis was carried out to determine whether RGC morphology was significantly different between treatment groups. This analysis provided 4 canonical scores for each cell which were analyzed by ANOVA and confirmed to be significantly different between treatment groups (Tukey's post hoc test; p<0.0001 for all treatment groups for at least one of the canonical scores). Measurements of morphological parameters (eg. cell soma size, dendrite length etc) were then compared by ANOVA (Tukey's post hoc test; significant when p<0.05) to identify which parameters were responsible for the differences.
To determine whether transgenes had differential impact on transduced and non-transduced cells in each treatment group, transduced cells in experimental retinae were compared to transduced cells in rAAV2-GFP injected retinae and non-transduced cells in experimental retinae were compared to non-transduced cells in rAAV2-GFP injected retinae. Transduced and non-transduced cells were also compared within treatment groups. This analysis provided 3 canonical scores for each treatment, that were analyzed by ANOVA and post hoc tests (Tukey). If post hoc analysis of canonical scores revealed significant differences between groups, measurements of morphological parameters were then analyzed by ANOVA as described above.
Dendritic tree morphology was also assessed using Sholl analysis [27] , which involves counting the number of dendrites intersecting concentric circles of increasing diameter (20 µm intervals) centered on the cell body. The radial distribution of dendrites was analyzed by ANOVA (repeated measures, Bonferroni post-hoc test). All 375 Neurolucida traces are available on the public data base NeuroMorpho.Org.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Freshwater availability has emerged as a global problem, given that more than four billion people currently experience some periods of severe water scarcity [ 1 ]. Water security challenges are further exacerbated when biodiversity risks and declines in freshwater ecosystem services are considered, with more than 80% of global citizens affected [ 2 ]. Balancing competing anthropogenic and ecological demands on finite water resources is a crucial issue for water and land managers [ 3 – 4 ], and frameworks are emerging to address the complex trade-offs associated with water management decision-making. Recent studies have emphasized the critical role of establishing a hydrologic foundation to inform water management decisions–including “Ecological Limits of Hydrologic Alteration, ELOHA” [ 5 ] and “Eco-Engineering Design Scaling, EEDS” [ 6 ].
A hydrologic foundation provides a key building block for any study of freshwater management, water security planning, or assessment of ecological water demands [ 5 ]. It requires an understanding of ongoing hydrologic processes, water availability, evapotranspiration, effects of land management and water infrastructure on streamflow, and the role of spatial and temporal variability in these processes. Empirical data analysis and model simulation represent two approaches for developing hydrologic analyses that inform water management decisions. When available, streamflow data provide a means to assess hydrologic processes and classify river flow regimes [ 7 – 10 ]. Predictive hydrologic models are often developed to simulate river flows [ 11 – 14 ], since streamflow observations are often unavailable or collected at insufficient spatial or temporal resolution. These models can assist forecasting large-scale changes in basin conditions associated with changes in land use and climate [ 15 ].
Tropical watershed hydrology has proven difficult to model due to high rainfall variability [ 16 – 17 ], high evapotranspiration rates [ 18 ], variation in forest canopy interception and storage [ 19 – 20 ], and uncertain hydrologic inputs from fog condensation in cloud forests [ 21 ]. In addition to these geographic and hydrologic uncertainties, social factors also complicate the feasibility of finding a balance between human and ecological needs in tropical regions. Approximately 40% of the global population currently resides in the tropics, with projected increases of up to 50% by 2050 due to continued population growth [ 22 – 23 ]. Given multiple compounding and/or competing water management challenges, assessing the amount of water available in these tropical environments is the first step to ensuring that there is enough water to meet increasing demands [ 24 ].
The objective of this paper is to build an empirical hydrologic foundation to inform water management decisions in the El Yunque National Forest (EYNF) of Puerto Rico as the amount of water in streams draining the forest is unknown. To facilitate this effort, mass-balance and observation-based water budgets for nine local watersheds were developed using a novel assemblage of remotely sensed rainfall data, gaged streamflow observations, and municipal water withdrawal rates which were not available for previous water budgets developed for this region.
Materials and methods
Study site
Tropical islands can provide unique examples of trade-offs in freshwater management due to the combined challenges of high amounts and extreme variability of rainfall, limited options for meeting anthropogenic freshwater demand, and highly productive ecosystems [ 25 ]. Puerto Rico provides a particularly compelling case study of water management trade-offs and challenges since it is among the wettest islands in the Caribbean, with one of the highest densities of both human population [ 26 ] and dams in the world [ 27 ]. The Luquillo Mountains in eastern Puerto Rico have been extensively studied and provide an ideal setting for research into the hydrologic cycle of tropical islands [ 26 , 28 ]. This study focuses on nine watersheds draining the EYNF (latitude 18°18’N, longitude 65°47’W) which is the only tropical rainforest owned and managed by the U.S.D.A. Forest Service. We selected these watersheds to build on the results of previous water budgets developed for this region [ 29 – 30 ]. The nine watersheds were delineated based on streamflow gage locations and used as the primary focal point of all subsequent analyses ( Fig 1 and Table 1 ).
10.1371/journal.pone.0213306.g001
Fig 1
Study area in eastern Puerto Rico with the focal watersheds in the EYNF outlined.
The spatial distribution of average annual rainfall (cm) for 2005–2013 from gridded NWS data is shown in the background.
10.1371/journal.pone.0213306.t001
Table 1 U.S. Geological Survey streamflow gages used in this study. Data quality for gaging stations varied between gages and between years over the period of analysis. According to the USGS, a gage is excellent if 95% of discharge readings are within 5% of true values, good if 95% of discharge readings are within 10% of true values, fair if 95% of discharge readings are within 15% of true values, and poor if accuracy is less than fair .
USGS Gage Number
Watershed Name
Area Served (km 2 )
USGS Data Quality Rating
Longitude
Latitude
50064200
Grande
19.01
Fair to good
-65.8413
18.3433
50075000
Icacos
3.25
Poor to fair
-65.7855
18.2752
50076000
Blanco
31.84
Poor to good
-65.7847
18.2272
50063800
Espiritu Santo
22.41
Poor to good
-65.8133
18.3584
50071000
Fajardo
38.43
Fair
-65.6946
18.2969
50061800
Canovanas
26.51
Poor to fair
-65.8888
18.3169
50055750
Gurabo
57.21
Fair to good
-65.8847
18.2320
50065500
Mameyes
17.61
Poor to fair
-65.7508
18.3274
50067000
Sabana
10.12
Poor to fair
-65.7306
18.3291
Within the study area, the Luquillo Mountains rise over 1,000 m above sea level in less than 10 km from the ocean. Consequently, there is a strong precipitation gradient ranging from 100 cm per year (in the lowlands on the leeward side of the mountains) to 450 cm per year (at high elevations on the windward side, Fig 1 ). The EYNF is composed of numerous forest types, each of which experience many natural disturbances such as hurricanes, landslides, and droughts [ 31 ]. Watersheds are steep, particularly in the upper portions, and are characterized by large boulders and low amounts of fine sediment [ 29 ]. Previous studies have shown groundwater loses in these watersheds are minimal [ 30 ], likely as a result of geomorphology. Recreation by residents and tourism by non-residents are important uses of the forest that can lead to land use modifications including clearing of vegetation for road and trail development to support picnicking, swimming, and hiking activities [ 32 ].
More than 30 water withdrawal structures extract freshwater for municipal and agricultural use from streams draining the forest and adjacent downstream regions [ 29 – 30 ]. These streams also host a variety of native migratory taxa that use both freshwater and estuarine ecosystems to complete their life cycle, including shrimps [ 33 – 34 ], fishes [ 35 ], and snails [ 36 ]. Many of these migratory organisms play important roles in regulating ecosystem processes such as primary production [ 37 – 38 ], and decomposition [ 39 ]. Thus, migratory biota and associated stream ecosystem processes are vulnerable to losses in riverine connectivity from dams and water withdrawals [ 27 , 40 ]. The ecological effects of water abstraction from streams draining the EYNF have been a focus of prior studies [ 41 – 43 ], and ecologically sensitive water management remains a critical topic for research [ 44 ].
Water budget
Water budgets provide useful tools for identifying key sources and sinks of water within a river basin, as well as examining the connections between the hydrologic cycle [ 17 ]. These simple mass-balance assessments quantify natural hydrologic processes taking place in an environment by partitioning total precipitation ( P ) between physical processes including evapotranspiration ( ET ), surface runoff ( RO ), and water withdrawals ( W ) averaged over sufficiently longtime scales. Here, we do not include groundwater, as previous water budgets indicate that it is not significant in this region [ 30 ]. The highly simplified linear mass balance ( Eq 1 ) expresses all parameters as volume of water (L 3 ) applied to a basin drainage area (L 2 ), resulting in units of average water depth (L).
P − E T − R O − W = 0
(1)
Applying the principle of conservation of mass, any combination of three parameters may be specified to compute the fourth. Because of the varying time scales of the hydrologic processes represented, the mass balance approach is not appropriate for determining instantaneous distribution of water in the environment. Instead, its application requires averaging parameters over a sufficient time so that the volume of water flowing through each compartment is captured at the magnitude of the longest temporal scale which is monthly for this study.
Importantly, the mass balance model simplifies the complex hydrologic cycle to include those components most relevant in the EYNF system. A full accounting of hydrologic processes is not specifically included because some processes occur at temporal scales much smaller than the monthly time scale applied (e.g., for canopy interception of rainfall) and some are assumed to be at steady-state because of their much longer temporal scale (e.g., for soil and groundwater storage) [ 45 ]. Another potentially important hydrologic process–cloud condensation and resulting fog drip–has previously been shown to minimally contribute to the overall water budget [ 46 ].
Our water budget constructs a mass balance model for nine watersheds in the EYNF draining approximately 70% of the total forest area as shown in Fig 1 . It extends prior EYNF water budgets [ 29 – 30 , 47 ] by incorporating new data sets that were unavailable during prior analyses. Specifically, we determine precipitation using gridded radar data provided by the National Weather Service, rather than estimating rainfall based on elevation. Additionally, our water withdrawal estimates are based on data provided by the water utility company on the island and we limit our study to intakes within the EYNF boundaries. This budget also provides an alternative approach to simulation-based water budgets for the region [ 48 ].
Precipitation data
Monthly precipitation estimates for the EYNF from January 2005 to December 2013 were calculated from rainfall data provided by the National Weather Service (NWS) Advanced Hydrologic Prediction Service [ 49 ]. These data sets contain quality controlled, spatially distributed precipitation estimates based upon multi-sensor observed data–including radar and ground based rain gage network sites. Fig 1 shows the spatial extent of annual rainfall across the EYNF, and clearly shows the annual rainfall differential for the windward (northeast) side of the mountains compared to the leeward (southwest) side [ 50 ].
Runoff data
Runoff was quantified using observed streamflow data from nine long-term U.S. Geological Survey (USGS) gaging sites and associated watersheds ( Table 1 and Fig 1 ) [ 51 ]. USGS gage locations were used to define watershed boundaries and served as the primary nodes of analysis for the water budgets. Daily averaged river discharge was used to compute monthly averaged flow for January 2005 to December 2013 (which corresponds to the time period of rainfall observations) and runoff volumes were converted to depths using the gage drainage area. Because observed runoff volumes at the EYNF boundary include the impact of municipal withdrawals inside the forest, observed runoff volumes at gages were corrected to account for upstream municipal withdrawals to represent total runoff.
Municipal withdrawals
The Puerto Rican Aqueduct and Sewer Authority (PRASA) has the responsibility and authority to provide potable water for public and private customers throughout Puerto Rico. The streams of EYNF provide a source of raw water serving local communities ( Fig 2 ). PRASA operational records for three years (including the calendar years 2013–2015, summarized in Table 2 “Withdrawals”) were obtained to determine abstraction rates in the study basins. Complete records from years prior to 2013 were not available, so it was necessary to assume that withdrawal rates during these three years are representative of other years in the study period.
10.1371/journal.pone.0213306.g002
Fig 2
Puerto Rican Aqueduct and sewage authority intake locations within and adjacent to EYNF.
10.1371/journal.pone.0213306.t002
Table 2 Summary of hydrologic budget for the EYNF. All values are expressed as depth in centimeters and presented as monthly and annual totals. Forest-wide averages are expressed as area-weighted quantities for the nine focal watersheds. Several instances occur where runoff and withdrawals exceeded precipitation, leading to negative ET values. All negative ET values are highlighted using gray shading.
Basin
J
F
M
A
M
J
J
A
S
O
N
D
ANN
Rainfall (cm)
Grande
21.7
11.7
8.9
21.1
19.5
26.5
19.0
25.1
29.5
26.8
21.4
20.4
252
Icacos
31.2
18.0
13.1
29.8
29.8
39.3
30.2
33.6
35.1
39.3
31.7
27.4
359
Blanco
25.8
14.3
9.9
23.8
24.9
34.8
27.2
29.8
32.4
36.6
28.4
23.7
312
Espiritu Santo
27.7
15.5
11.9
26.7
24.6
30.1
22.6
29.8
31.6
29.9
24.9
24.2
299
Fajardo
28.0
15.3
11.4
27.0
29.1
35.7
28.9
29.8
30.3
35.4
32.0
26.3
329
Canovanas
15.6
8.2
6.4
16.0
15.4
22.4
17.0
21.4
26.5
24.8
18.9
16.3
209
Gurabo
14.7
7.0
4.9
13.3
16.3
24.9
22.1
23.5
25.5
28.4
21.3
14.1
216
Mameyes
33.9
19.7
14.8
33.2
33.2
40.6
31.2
35.3
35.2
38.6
32.9
29.3
378
Sabana
33.2
18.5
14.4
32.8
34.9
40.0
32.1
34.9
33.1
37.1
34.2
29.8
375
EYNF Average
23.0
12.4
9.2
22.0
23.0
30.5
24.4
27.5
29.5
31.6
25.8
21.3
280
Runoff (cm)
Grande
20.3
11.3
17.0
15.0
28.2
15.4
20.7
21.0
23.4
20.3
20.9
22.4
236
Icacos
37.6
26.5
32.3
33.9
56.6
38.2
39.0
36.9
40.0
43.4
43.4
37.7
465
Blanco
18.4
10.2
16.6
19.1
30.3
23.2
23.2
27.1
31.1
33.3
22.6
22.7
278
Espiritu Santo
27.3
14.7
18.4
20.8
31.9
16.8
22.2
23.1
24.0
20.5
26.5
29.4
275
Fajardo
16.4
7.7
12.8
13.7
23.6
19.2
16.4
17.2
21.3
25.1
20.2
16.5
210
Canovanas
9.9
5.6
6.7
6.9
11.1
7.4
9.1
13.6
15.0
14.2
12.5
10.1
122
Gurabo
3.6
2.1
2.4
3.3
6.7
5.6
6.8
5.9
10.2
13.3
6.6
4.6
71
Mameyes
25.7
16.5
22.0
21.9
36.5
21.7
23.3
23.8
27.6
27.2
27.7
28.9
303
Sabana
12.7
7.9
10.8
13.2
24.8
16.1
12.6
16.4
21.4
22.2
19.5
18.0
196
EYNF Average
14.6
8.1
11.5
12.5
21.0
14.4
15.4
16.7
20.0
21.1
17.4
16.4
189
Withdrawals (cm)
Grande
0.9
0.7
0.8
0.8
0.8
0.8
0.7
0.7
0.7
0.8
0.8
0.8
9
Icacos
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0
Blanco
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
1
Espiritu Santo
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
1
Fajardo
3.3
2.9
3.0
2.7
2.9
3.0
2.9
2.8
2.7
2.9
3.1
3.2
35
Canovanas
0.4
0.3
0.4
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.4
4
Gurabo
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.4
4
Mameyes
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0
Sabana
2.5
2.3
2.4
2.4
2.4
2.2
2.4
2.5
2.5
2.6
2.6
2.6
29
EYNF Average
0.9
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.9
0.9
10
Estimated Evapotranspiration (cm)
Grande
0.6
-0.3
-8.9
5.3
-9.4
10.3
-2.4
3.4
5.3
5.8
-0.2
-2.9
6
Icacos
-6.4
-8.5
-19.2
-4.1
-26.8
1.1
-8.8
-3.3
-4.9
-4.1
-11.6
-10.3
-107
Blanco
7.3
4.0
-6.8
4.6
-5.5
11.5
3.9
2.6
1.1
3.2
5.7
0.9
32
Espiritu Santo
0.4
0.8
-6.5
5.8
-7.3
13.3
0.3
6.7
7.6
9.4
-1.7
-5.3
23
Fajardo
8.3
4.7
-4.4
10.6
2.5
13.6
9.5
9.7
6.4
7.4
8.8
6.5
84
Canovanas
5.4
2.3
-0.7
8.7
4.0
14.7
7.7
7.5
11.3
10.3
6.0
5.9
83
Gurabo
10.8
4.7
2.2
9.8
9.3
19.0
15.0
17.3
15.0
14.7
14.4
9.2
141
Mameyes
8.2
3.2
-7.1
11.3
-3.3
19.0
7.8
11.6
7.6
11.5
5.1
0.4
75
Sabana
18.0
8.3
1.2
17.2
7.7
21.6
17.1
16.0
9.2
12.2
12.1
9.2
150
EYNF Average
7.5
3.6
-3.2
8.7
1.1
15.3
8.2
10.0
8.7
9.6
7.5
4.0
81
Evapotranspiration estimation
Evapotranspiration estimates were computed by solving Eq 1 using the monthly estimates of precipitation, runoff, and withdrawals described earlier. The term “evapotranspiration” typically describes combined water sinks related to physical (evaporation) and biological (transpiration) processes. However, with the mass balance approach applied here, ET is a derived quantity to describe these and all other water sinks/sources that are not stream discharge or municipal withdrawals, including changes to long-term surface or soil storage, infiltration to groundwater, and hydrologic contributions from other unidentified sources of water, possibly including artesian springs or cloud condensation.
Results
A computed monthly water budget for each of the nine focal watersheds is generated using observed data covering the period 2005–2013 ( Table 2 ). Annual precipitation was characterized with a forest area-weighted average of 280 cm per year. However, precipitation was temporally variable within the annual cycle with all nine watersheds exhibiting ranges over 20 cm between the wettest and driest months. Rainfall was also spatially variable given that average annual rainfall varied from 209–378 cm across the Canovanas and Mameyes watersheds, respectively.
As expected, runoff variation between basins generally follows precipitation amounts. However, differences in basin-specific hydrologic processes relating to topographic and geologic variation resulted in non-linear runoff responses. Observed runoff varied from 31% (Gurabo) to more than 130% (Icacos) of average annual precipitation.
Municipal withdrawals were calculated to be 10 cm per year on average at the forest-wide scale, or 3.6% of total precipitation and 5.3% of total runoff. However, results are highly dependent on the watershed of interest with average withdrawal rates ranging from 0% (Icacos and Mameyes) to 17% (Fajardo) of total runoff, indicating that municipal withdrawals have the potential to extract enough water from the EYNF streams to create negative (worst case) or unintended (best case) hydrological or ecological consequences.
As shown in Table 2 , several instances occur where runoff and withdrawals exceed precipitation, leading to negative ET values (28 of 108 month-watershed combinations). This suggests that more water is being discharged from these basins than is falling as precipitation–an outcome not possible without a secondary water source possibly from artesian groundwater inputs and/or fog condensation. All calculated ET values that were negative or zero are denoted by gray shading in Table 2 .
Discussion
Approximate decadal water budgets have been compiled for the EYNF and adjacent downstream regions outside of the EYNF [ 29 – 30 , 48 ]. While all budgets have relied on the mass balance equation (i.e., Eq 1 ), different methods were used to estimate parameter values (i.e., P , RO , W , and ET ) with each successive analysis. Improved temporal and spatial resolution is now possible as additional observational and remotely sensed data have become available Table 3 provides a comparison of the current water budget presented here with two prior budgets. Although methods vary substantially between budgets, estimates of precipitation, evapotranspiration, and runoff are generally similar in that they do not differ more than 20% between individual budget estimates.
10.1371/journal.pone.0213306.t003
Table 3 Comparison of our current hydrologic budget for EYNF with two prior water budgets. It is important to note the difference in geographic scope of the three budgets with respect to accounting for water withdrawals. Our current study accounts for all water withdrawals within the EYNF. In contrast, the previous two studies considered large water withdrawals outside of the EYNF.
Hydrologic Parameter
Overview of Method
Water Budget Estimates (cm)
Naumann (1994)
Crook et al. (2007)
This study
Naumann (1994)
Crook et al. (2007)
This study
Precipitation (P)
Developed using unit area rainfall based upon elevation [ 46 ]
Developed using regression equation [ 52 ] with temporal pattern derived from rain gage at the El Verde [ 64 ]
Gridded radar rainfall data for period 2005 to 2013 [ 49 ]
338
358
280
Runoff (RO)
Calculated: RO = P–ET–W
Developed using data from 7 USGS gage stations using available data through 2002
Developed using data from 9 USGS gage stations for 2003 to 2013 to coincide with rainfall data
194
228
189
Withdrawal (W)
Estimated using gravity flow capacity of known intake pipes within the EYNF & including withdrawals from large additional intakes outside the EYNF
Estimated using permitted withdrawal capacity or gravity flow capacity of intake pipes within the EYNF & including withdrawals from large additional intakes outside the EYNF
Quantified using PRASA operational data for 2013 to 2015. This analysis only considers withdrawals from intakes within the EYNF
12
25
10
Evapo-transpiration (ET)
Developed using unit area rainfall based upon vegetation type and life zone [ 46 ]
Calculated: ET = P–RO–W
Calculated: ET = P–RO–W
132
130
81
Precipitation was observed using a high-resolution gridded dataset [ 49 ], that was unavailable for prior water budgets. Increased data resolution led to an overall reduction in rainfall estimates and illustrates the high spatial variability in rainfall within EYNF. The variability in precipitation is attributable to elevation [ 52 ], basin aspect, and prevailing winds (i.e., windward and leeward sides of the mountain range, Fig 1 ) [ 50 ]. Previous water budgets for the region used precipitation estimates that did not take these variables into account, indicating the water budget presented here is based on the highest resolution of precipitation data available.
Estimates of municipal withdrawal rates are markedly different between the three water budgets, particularly given that Nauman (1994) and Crook et al. (2007) consider water withdrawal amounts from intakes outside of the EYNF boundary (Tables 3 and 4 ). Prior budgets indicate that 37 to 41 MGD day (1.62 to 1.80 m 3 /s) are withdrawn outside of the EYNF boundary (i.e. > 5-fold increase relative to water withdrawn within the EYNF; Table 4 ). However, prior budgets also estimated the volume of withdrawal based on either the capacity of known water intake structure [ 29 ] or the regulatory permit limit assigned to the water utility [ 30 ]. In contrast, the summary of operational data reflected in our budget shows significantly reduced withdrawal estimates, which we believe to be more realistic estimates of municipal impacts. Additionally, hydropower intakes within the EYNF boundary were excluded from our water budget in contrast to the previous two budgets ( Table 4 ), because the water used for hydropower generation is returned to the river channel. Nonetheless, all three of these studies indicate that water withdrawals represent a substantial alteration of the hydrologic budget and should be included in analyses [ 48 ].
10.1371/journal.pone.0213306.t004
Table 4 Comparison of water withdrawn inside and outside of the EYNF Boundary across studies.
Budget
Water Amount Withdrawn Inside EYNF Boundary (m 3 /s)
Water Amount Withdrawn Outside EYNF Boundary (m 3 /s)
Water Amount Withdrawn for Hydropower Included in Budget (m 3 /s)
Naumann (1994)
0.32
1.79
0.13
Crook et al. (2007)
1.25
1.65
0.16
This study
0.32
Not included
Not included
Based on analysis of observed data, runoff and withdrawals exceeded precipitation in some basins, resulting in negative estimates of evapotranspiration. This physical impossibility indicates an additional but unidentified source of water within these basins. Crook et al. (2007) also observed this artifact in the Icacos basin but offered no explanation for the observed condition. Our analysis found that this discrepancy occurs in multiple watersheds, particularly those with substantial portions of their upper basins within the cloud forest (generally defined as forest located above 600 m elevation as shown in Fig 1 ). A plausible and likely explanation of the larger volume of runoff compared to precipitation is the contribution of moisture through condensation within the cloud forest [ 50 ]. Previous estimates of the contribution of condensation to the forest hydrology suggest that an additional 1–10% of total precipitation can be gained through fog condensation in these high elevation areas [ 50 , 53 – 55 ]. Data presented here suggest that this percentage could reach as high as 23% of total annual precipitation, depending on the proportion of cloud forest cover in a given basin. Methodology used in this report is insufficient to adequately explain this hydrologic uncertainty which represents an important opportunity for future research in tropical hydrology. However, despite the aforementioned uncertainty, water budget estimates of annual forest-wide ET (81 cm/yr) fall within prior estimates of annual ET values derived for the ENYF, namely: 75 cm/yr [ 56 ], 115 cm/yr [ 57 ], 172 cm/yr [ 52 ], and 175 cm/yr [ 58 ].
A monthly time step is useful when determining large-scale water budgets as it allows for the inclusion of hydrologic processes which function at variable time scales. However, intra-month variability can be highly relevant to ecological processes and significant in interpreting results. For instance, Fig 3A presents daily streamflow data for the Espiritu Santo watershed from 2005–2013 along with annual and monthly averages for this period. The importance of time scale is clearly shown when withdrawal capacity is included. Currently, withdrawals upstream of the streamflow gage in the EYNF are reasonably small (dashed red line), but significant withdrawals exist between the gage and the ocean (dashed black line). Furthermore, the intake capacity (dashed blue line) at these withdrawals is higher than the current rate. Withdrawal capacity exceeds daily streamflow rates in rare cases (~1% of the time). At a monthly averaging interval, the withdrawal capacity is always less than the observed runoff (solid blue line) which erroneously suggests a lack of water deficits throughout the average year. However, at a daily interval, the annual withdrawal rate is shown to exceed available runoff, indicating the potential for over-abstraction from the river, as has been observed in several previous studies during periods of low precipitation [ 34 , 41 , 43 ]. Thus, conclusions regarding the effects of withdrawals should be considered not only on an average basis, but also on the basis of exceedance probability ( Fig 3B ). This is a major challenge given that monthly intervals are commonly used for long-term water budgets to rectify the time scale of different processes (i.e., precipitation, runoff, and evapotranspiration) and they do not reflect the loss of hydrologic and ecological connectivity at finer temporal scales.
10.1371/journal.pone.0213306.g003
Fig 3
Sample hydrograph from Rio Espiritu Santo (USGS Gage No. 50063800) for 2005–2013.
This study contributes to the growing body of literature demonstrating that ecological effects of increased water extraction are likely, even in areas of plentiful rainfall [ 25 , 38 , 41 , 43 , 59 – 61 ]. The water budget methodology employed here provides a hydrologic foundation for informing these trade-offs in water management and raises interesting questions about other important but currently unquantified sources of water that may not be persistent in a changing climate.
Ecological connectivity within a given watershed has been shown to be intimately connected to hydrology [ 43 ]. Studies indicate that some upstream migrations can persist even under heavy withdrawal or extended droughts provided that some free flow of water (no matter how small) continues to overtop dam faces or spillways during at least some time of the year [ 27 , 41 ]. Therefore, estimates of municipal withdrawals like those contained here can inform the operation of these structures along with decisions regarding future permit applications to better insure hydrological and ecological connectivity. In 2015, Puerto Rico experienced a historic drought which resulted in increasing pressure for additional water withdrawal from the EYNF [ 62 ]. The water budget we present here provides a basis for informing the trade-offs associated with challenging decisions about when and where to withdraw additional water supply, while satisfying migratory species needs for freshwater.
Finally, Puerto Rico’s tropical ecosystems have experienced a variety of changes over time due to deforestation and reforestation [ 25 – 26 ], increased water abstraction [ 30 ], and natural disturbances [ 32 ]. In coming years, shifts in population demographics and associated water needs along with climate change could induce significant changes in local hydrology and further exacerbate water management trade-offs [ 63 ]. A hydrologic foundation of ongoing processes within the EYNF will not only be helpful confront these challenges but will be crucial to the continued management of the forest, its stream ecosystems, and nearby water supplies.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Of the 24 neglected tropical diseases (NTDs) and conditions listed by WHO, snakebite is among the top killers [ 1 ]. Tens of thousands of people die each year as a result of snakebite envenoming, with close to 50,000 deaths in India alone [ 2 ] and up to 32,000 in sub-Saharan Africa [ 3 ]. Yet there are few sources of effective, safe, and affordable antivenoms. The regions that bear the highest snakebite burden are especially underserved [ 4 ].
The Fav-Afrique antivenom, produced by Sanofi Pasteur (France), is considered safe and effective and is one of the few antivenoms to be approved by a Stringent Regulatory Authority (French National Regulatory Authority), although limited formal evidence has been published [ 5 , 6 ]. It is polyvalent, targeting most of the medically important snake species in sub-Saharan Africa. In particular, it is highly effective in treating envenoming by Echis ocellatus , the West African saw-scaled viper [ 5 – 7 ] that causes great morbidity and mortality throughout the West and Central African savannah. The venom of E . ocellatus may induce systemic haemorrhage, coagulopathy, and shock, as well as extensive local tissue damage. In the absence of treatment, the case fatality rate is 10%–20% [ 8 ]. Médecins Sans Frontières (MSF) uses Fav-Afrique in its projects in sub-Saharan Africa, notably in Paoua in Central African Republic (CAR), where E . ocellatus envenoming is frequent [ 9 ]. Worryingly, MSF has been informed that the production of Fav-Afrique by Sanofi Aventis will be permanently discontinued. The last batch was released in January 2014, with an expiry date of June 2016. All the vials produced have already been sold by Sanofi Pasteur.
Although several alternative antivenom products target a similar list of species as Fav-Afrique, there is currently no evidence of their safety and effectiveness. We aimed to review the evidence for the efficacy and safety of existing and in-development snake antivenoms, and to list the alternatives to Fav-Afrique in sub-Saharan Africa.
Search Strategy
We searched clinical trial registries (National Institutes of Health clinicaltrials.gov and WHO International Clinical Trial Registry Platform) and a publication database (EMBASE) to identify ongoing and completed clinical trials. The registries were searched by condition using the keywords “snakebite” OR “snake bite” OR “snake envenom*” OR “envenom*” OR “bite.” Publication database search strategy was based on the Medical Subject Heading (MeSH) terms “clinical trial” AND “snake bites” AND “polyclonal antiserum OR snake venom antiserum OR venom antiserum.” All terms were explored, and results were limited to studies conducted in humans. No time limits were imposed. Searches were conducted in September 2014 and included all records from the launch of the databases. Only those studies with a design compatible with that of a clinical trial (prospective, comparative, and interventional) and with the definition given by the CONSORT glossary were included. Prospective, single-arm cohorts were not considered as clinical trials.
Search Results
The registry searches yielded 29 records, four of which were observational studies. Among the interventional studies, 12 investigated antivenom as an intervention (eight were retrieved out of 176,201 records in clinicaltrials.gov and 12 out of 254,285 in ICTRP). Table 1 summarises the characteristics of the 12 trials. Four trials were sponsored by pharmaceutical companies and the remainder, by an individual researcher or academic institution. Four trials were open for recruitment and five were completed or terminated. A total of 11 different antivenoms were being investigated, most in only one trial.
10.1371/journal.pntd.0003896.t001
Table 1 List of clinical trials investigating snake antivenom published in clinical trials registries.
Trial ID number
Title
Sponsor
Type of funding
Location
Year of trial registration
Recruitment status
Results published
NCT00303303
The Efficacy of Crotaline Fab Antivenom for Copperhead Snake Envenomations
Carolinas Healthcare System
Government
United States
2006
Terminated
No
NCT00636116
Phase 3 Multicenter Comparative Study to Confirm Safety and Effectiveness of the F(ab)2 Antivenom Anavip
Instituto Bioclon S.A. de C.V.
Industry
US
2008
Completed
No
NCT00639951
Study to Evaluate the Efficacy of Two Treatment Schemes With Antivipmyn for the Treatment of Snake Bite Envenomation
Instituto Bioclon S.A. de C.V.
Industry
Mexico
2008
Recruiting
NA
NCT00811239
A Controlled Clinical Trial on The Use of a Specific Antivenom Against Envenoming by Bungarus Multicinctus
Hanoi Medical University
Government
Vietnam
2008
Completed
Yes [ 21 ]
NCT00868309
A Comparison of Crotalinae (Pit Viper) Equine Immune F(ab)2 Antivenom (Anavip) and Crotalidae Polyvalent Immune Fab, Ovine Antivenom (CroFab) in the Treatment of Pit Viper Envenomation
Instituto Bioclon S.A. de C.V.
Industry
US
2008
Completed
Yes [ 22 ]
ISRCTN01257358
Clinical trial of two new anti-snake venoms for the treatment of patients bitten by poisonous snakes in Nigeria
Nigeria MoH
Unknown
Nigeria
2009
Completed
Yes [ 23 ]
SLCTR/2010/006
Low dose versus high dose of Indian polyvalent snake antivenom in reversing neurotoxic paralysis in common krait ( Bungarus caeruleus ) bites: an open labelled randomised controlled clinical trial in Sri Lanka
Individual researcher
None
Sri Lanka
2010
Not recruiting
No
ACTRN12611000588998
A randomised controlled trial of antivenom and corticosteroids for red-bellied black snake envenoming
Individual researcher
Government
Australia
2011
Not recruiting
No
NCT01284855
Comparison of Two Dose Regimens of Snake Antivenom for the Treatment of Snake Bites Envenoming in Nepal
University of Geneva
Government
Nepal
2011
Not recruiting
No
NCT01337245
Emergency Treatment of Coral Snake Envenomation With Antivenom
University of Arizona
Government
US
2011
Recruiting
NA
ACTRN12612001062819
A randomized controlled trial (RCT) of a new monovalent antivenom (ICP Papuan taipan antivenom) for the treatment of Papuan taipan ( Oxyuranus scutellatus ) envenoming in Papua New Guinea
University of Melbourne
Government
Papua New Guinea
2012
Recruiting
NA
NCT01864200
A Randomized, Double-Blind, Placebo-Controlled Study Comparing CroFab Versus Placebo With Rescue Treatment for Copperhead Snake Envenomation (Copperhead RCT)
BTG International Inc.
Industry
US
2013
Recruiting
NA
The publication database search yielded 97 results ( Fig 1 ). After cleaning, 82 records were retained, of which 30 had a design consistent with clinical trials. The remainder included 26 reviews or commentaries, 18 cohorts or cases series, four retrospective analyses of medical records, two case studies, one diagnostic study, and one cross-sectional survey. A search of references yielded an additional 11 reports of clinical trials. Of the 41 clinical trials thus identified, 32 investigated antivenom as an intervention. The locations of the 32 studies were Latin America (Brazil n = 3, Columbia n = 5, Ecuador n = 1); Asia (India n = 4, Thailand n = 5, Sri Lanka n = 3, Myanmar n = 1, Malaysia n = 1); Africa (Nigeria n = 5), and US ( n = 4). 27 were sponsored by a public organization (e.g., university or public hospital). Most trials ( n = 20) were conducted before 2000, the oldest dated from 1960 [ 10 ]. A total of 30 antivenoms were investigated; half were investigated in only one trial.
10.1371/journal.pntd.0003896.g001
Fig 1
Flow diagram of the selection process used in this study.
The search was conducted on 15 September 2014. Merging the search results gave a total of 41 clinical trials investigating the efficacy or safety of snake antivenoms, of which four were active. A total of 36 different antivenoms were investigated (see Table 2 ). Based on the trial design (Phase I to IV), ten products were considered still “under development,” although development appears to have stalled for most of them. Our search strategy appears robust; a report conducted in 2010 identified a total of 43 randomized controlled trials on snakebite envenoming, 28 of which investigated antivenom properties [ 11 ]. We retrieved all except two of these trials [ 12 , 51 ]; the discrepancy could be due to differences in the criteria used to define clinical trials.
10.1371/journal.pntd.0003896.t002
Table 2 List of antivenoms investigated in clinical trials published in peer-reviewed journals or on public registries.
Product name
Other name/product specifications
Manufacturer
Development stage 1
Target region
Publications
Clinical trials registry number
CroFab
Polyvalent ovine antivenom (Fab) against Crotalid
Protherics
Phase III–IV
North America
[ 22 , 24 , 25 ]
NCT00303303 NCT00636116 NCT00868309 NCT01864200
Anavip
Polyvalent equine antivenom (Fab2) against Crotalinae (pit viper)
Instituto Bioclon S.A.
Terminated after Phase III
North America
[ 22 ]
NCT00868309 NCT00636116
Antivypmin
Polyvalent equine antivenom (Fab2) against Crotalinae (pit viper)
Instituto Bioclon S.A.
Phase III
North America
None
NCT00639951
NA
Polyvalent equine antivenom (Fab2) against North American Coral snakes ( Micrurus )
University of Arizona
Phase III
North America
None
NCT01337245
Tiger snake antivenom
Monovalent equine (Fab) against Notechis scutatus
CSL
Phase III–IV
Australia
None
ACTRN12611000588998
Taipan antivenom
Monovalent equine (Fab) against Oxyuranus scutellatus
CSL
Phase I–II
Australia
None
ACTRN12612001062819
Antibotropico IVB
Instituto Vital Brazil
Phase II
Latin America
[ 26 ]
None
Antibotropico Butantan
Polyvalent equine antivenom against Bothrops species
Instituo Butantan
Phase II–III
Latin America
[ 26 – 29 ]
None
Antibotropico FUNED
Fundação Ezequiel Dias
Terminated
Latin America
[ 26 ]
None
Antibotropico-laquetico Butantan
Bothrops-Lachesis polyvalent equine antivenom
Instituo Butantan
Phase II
Latin America
[ 30 ]
None
Antiofiodico botropico polivalente
Polyvalent equine antivenom (IgG) against Bothrops asper , Bothrops atrox , and Bothrops xanthogrammus
Instituto Nacional de Higiene y Medicina Tropical "Leopoldo Izquieta Pérez"
Phase II–III
Latin America
[ 28 ]
None
Monovalent B . atrox equine antivenom
Instituto Clodomiro Picado
Terminated
Latin America
[ 31 , 32 ]
None
Monovalent B . atrox equine antivenom
Instituto Nacional de Salud
Terminated
Latin America
[ 29 ]
None
B . atrox–Lachesis antivenom
Polyvalent equine antivenom (IgG) against B . atrox and Lachesis muta muta
Fundação Ezequiel Dias
Terminated
Latin America
[ 30 ]
None
Polyvalent Antivenom
Polyvalent equine antivenom (IgG) against B . asper , Crotalus durissus , and L . muta
Instituto Nacional de Salud
?
Latin America
[ 28 ]
None
Polyvalent antivenom ICP
Polyvalent equine antivenom (IgG or Fab2) against B . asper , Crotalus simus , and Lachesis stenophrys
Instituto Clodomiro Picado (University of Costa Rica)
Phase II
Latin America
[ 31 – 34 ]
None
EchiTab
Monovalent ovine antivenom (Fab) against Echis oscellatus
Therapeutic Antibodies/Micropharm
?
Sub-Saharan Africa
[ 35 ]
None
EchiTab Plus
Polyvalent equine antivenom against Bitis arietans , E . oscellatus , and Naja nigricollis
Instituto Clodomiro Picado (University of Costa Rica)
Phase I–II
Sub-Saharan Africa
[ 23 , 36 ]
ISRCTN01257358
EchiTab G
Monovalent antivenom (IgG) against E . oscellatus
Micropharm
Phase I–II
Sub-Saharan Africa
[ 23 , 36 ]
ISRCTN01257358
EgyVac antivenom
Equine polivalent antivenom against B . arietans , E . oscellatus , and N . nigricollis
Vacsera Ltd
Terminated after Phase I
Sub-Saharan Africa
[ 36 ]
None
Ipser Africa Antivenom
Polyvalent equine (Fab2) antivenom against B . arietans , Bitis gabonica , Echis leucogaster , N . nigricollis , Naja haje , Naja melanoleuca , Dendroaspis viridis , Dendroaspis jamesoni , and Dendroaspis augisticeps
Institut Pasteur
?
Sub-Saharan Africa
[ 35 ]
None
Monospecific antivenom against E . oscellatus
Institut Pasteur
?
Sub-Saharan Africa
[ 37 , 38 ]
None
SAIMR Echis antivenom
Monovalent equine antivenom (IgG or Fab2) against Echis carinatus / ocellatus
South African Vaccines Producer
?
Sub-Saharan Africa
[ 38 ]
None
North and West African polyvalent antivenom ( Echis , Bitis , Naja )
Behningwerke
?
Sub-Saharan Africa
[ 37 , 38 ]
None
Malayan pit viper antivenom
Monovalent equine antivenom against Calloselasma rhodostoma
Queen Saovabha Memorial Institute
Phase I–II
South East Asia
[ 11 , 39 – 41 ]
None
Malayan pit viper antivenom
Monovalent caprine antivenom against C . rhodostoma
Twyford Pharmaceutical
Phase I–II
South East Asia
[ 39 – 41 ]
None
Malayan pit viper antivenom
Monovalent equine antivenom against C . rhodostoma
Thai Government Pharmaceutical Organisation
Phase I–II
South East Asia
[ 39 – 41 ]
None
Monocellate cobra antivenom
Monovalent equine antivenom against aja. kaouthia
Queen Saovabha Memorial Institute
?
South East Asia
[ 42 ]
None
Green pit viper antivenin (QSMI)
Polyvalent equine antivenom (Fab2) against green pit vipers
Queen Saovabha Memorial Institute
Phase I–II
South East Asia
[ 41 , 43 ]
None
B . multicinctus and B . candidus antivenom
Polyvalent equine antivenom (Fab2) against Bungarus multicinctus and Bungarus candidus
Vietnam Poison Control Center, Hanoi Medical University
Phase I–II
South East Asia
[ 21 ]
NCT00811239
Monospecific antivenom against D . russelii
Myanmar Pharmaceutical Factory
?
South East Asia
[ 44 ]
None
ProlongaTab
Monovalent ovine antivenom (Fab) against Daboia russelii
Therapeutic Antibodies Inc
Terminated
South Asia
[ 45 , 46 ]
None
SII Polyvalent ASV IP
Polyvalent equine antivenom (Fab2) against Naja naja , E . carinatus , D . russelii and Bungarus caeruleus
India Serum Institute
?
South Asia
[ 47 – 49 ]
None
Snake antivenin IP
Polyvalent equine antivenom (Fab2) against N . naja , E . carinatus , D . russelii and B . caeruleus
Haffkine Biopharmaceutical Corporation Ltd
Phase II
South Asia
[ 45 , 46 , 50 , 51 ]
None
Snake venom anti-serum
Polyvalent equine F(ab)2 against B . caeruleus , N . naja , D . russelii and E . carinatus
VINS bioproducts
Phase II
South Asia
None
SLCTR/2010/006 NCT01284855
Snake venom antiserum
Polyvalent equine F(ab)2 against B . caeruleus , N . naja , D . russelii and E . carinatus
Bharat Serum and Vaccines Ltd
Phase II
South Asia
None
SLCTR/2010/006
1 Not all publications mentioned the trial phase, and development status was established based on trial design, primary objectives, and number of subjects. This classification, though, bears some limitations, especially with regards to snake antivenoms development, in which Phase I with healthy volunteers are generally not conducted.
Urgent Need for More Research
Our results highlight the paucity of adequately conducted clinical trials and corroborate previous findings on the scarcity of safe, effective, and quality-assured snake antivenoms [ 4 ]. Comparison with dengue fever, which has a similar burden (11.97 Disability-Adjusted Life Years (DALYs) per 100,000 [4.99–20.46] versus venomous animal contacts 39.62 DALYs per 100,000 [22.46–69.74]) [ 13 ], is particularly revealing. In 2011, of 79 identified trials on dengue fever, 27 were recruiting patients, with six new products in development [ 14 ]. By contrast, the research pipeline for snakebite remains desperately dry, despite numerous calls for action [ 15 – 17 ].
Antivenoms in Sub-Saharan Africa
To determine how many antivenom products are currently available in sub-Saharan Africa, we searched WHO “Venomous snakes and antivenoms database” and held bilateral discussions with snakebite experts and pharmaceutical companies. We found that 12 antivenom products were commercially available in sub-Saharan countries as of September 2014 ( Table 3 ), only three of which had been tested in at least one clinical trial, and many of which may lack efficacy [ 18 ].
10.1371/journal.pntd.0003896.t003
Table 3 Available snake antivenom products in sub-Saharan Africa, as of September 2014.
Product
Company
Country of production
Antivipmyn-Africa
Instituto Bioclon/Silanes
Mexico
ASNA-C
Bharat Serums and Vaccines
India
ASNA-D
Bharat Serums and Vaccines
India
EchiTabG
MicroPharm
United Kingdom
EchiTabPlus
Instituto Clodomiro Picado
Costa Rica
Fav-Afrique
Sanofi Pasteur
France
Inoserp PanAfrica
Inosan
Spain
SAIMR Boomslang antivenom
South African Vaccine Producers
South Africa
SAIMR Echis antivenom
South African Vaccine Producers
South Africa
SAIMR Polyvalent Snake antivenom
South African Vaccine Producers
South Africa
Snake Venom Antiserum (Pan-African)
VINS Bioproducts
India
Snake venom antiserum Echis ocellatus
VINS Bioproducts
India
Case study: The MSF experience in Central African Republic
The experience of MSF in CAR suggests that there are indeed significant variations in the efficacy of antivenoms against African snake venoms. MSF has been using Fav-Afrique to manage patients presenting with features of snakebite envenoming in Paoua, CAR, since 2008. In the first half of 2013, Fav-Afrique was temporarily unavailable, and an alternative product was identified, directed against the venoms of 11 species of African snakes, including E . ocellatus . This antivenom was used for six months, with the same criteria for therapy as for Fav-Afrique. Although a methodologically sound study could not be conducted, a retrospective analysis of MSF medical records showed that the case fatality rate increased from 0.47% (three of 644 treated patients) with Fav-Afrique [ 9 ] to 10% (five of 50 treated patients) with the alternative antivenom. While more than 80% of patients were successfully treated with only one dose of Fav-Afrique, more than 60% treated with the alternative antivenom (31 of 50) required more than one dose to control envenoming. Worryingly, the first dose of the alternative antivenom was not able to alleviate spontaneous bleeding at admission in ten of 13 patients, and the administration of additional doses was required. These field data need cautious interpretation. However, they echo findings on the availability of ineffective and potentially harmful antivenoms in sub-Saharan Africa and support the conclusion that post-marketing surveillance is crucial [ 18 ]. They also call for a more robust and systematic evaluation of marketed products by regulatory authorities in the affected countries.
The Way Forward
Sanofi Pasteur urgently needs to disclose its plan to mitigate the negative impact of the decision to stop producing Fav-Afrique. Over the longer term, the multi-component strategy described by the Global Snakebite Initiative must be fully financed [ 19 ]; both innovations for better products and interventions and access to quality care need to be enhanced. The vast majority of the trials that we identified were sponsored by public organizations. The snakebite antivenom market so far appears poorly lucrative, unpredictable, and fragmented, hindering investment from pharmaceutical companies [ 4 ]. A major donor needs to step in, provide support, and, importantly, encourage existing global health initiatives, such as Drugs for Neglected Diseases initiative (DNDi), the Global Alliance for Vaccine and Immunization (GAVI)-Alliance, or the European and Developing Countries Clinical Trials Partnership (EDCTP), to extend their remits to life-saving treatments for snakebites. Finally, WHO should fully include snakebite envenoming in its list and programme of NTDs, support national regulatory authorities in performing adequate evaluations of existing antivenom products, and establish partnerships for access to existing and future antivenoms. Snakebite envenoming has been a most neglected disease for far too long.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Spanning 73 countries and territories and placing an estimated 1.39 billion individuals at risk of infection, lymphatic filariasis (LF) presents a considerable risk to global health [ 1 ]. Similarly, with an estimated 198 million malaria infections and 584,000 malaria-related deaths in 2013, the global burden of human malaria is staggering [ 2 ]. Yet despite the wide ranging impacts of these diseases, global elimination efforts have made significant strides, spearheaded by mass drug administration (MDA) programs supported by large pharmaceutical donors [ 3 – 5 ] and the widespread use of insecticidal bed nets [ 6 – 9 ]. As a result, disease prevalence in many locations has decreased dramatically, enabling a growing number of countries to discontinue their treatment efforts for LF [ 5 , 10 ] and spurring the creation of an increasing number of malaria elimination programs [ 11 – 13 ]. However, lessons learned as a result of LF elimination efforts have shown that the cessation of MDA, recommended after the successful passing of a transmission assessment survey [ 14 ], results in an additional set of programmatic challenges. Foremost in such post-intervention settings is the issue of post-MDA surveillance, as vigilant monitoring is required to ensure that recrudescence of disease has not occurred [ 15 ]. This monitoring is costly and current efforts for LF are centered upon the periodic sampling of the human population in order to examine circulating levels of filarial antigen [ 16 – 17 ]. While effective, these efforts require blood sampling of the human population. The invasive nature of this practice, coupled with the requirement of informed consent, results in participation challenges [ 14 ] that logically increase as populations become further removed from the time of widespread disease transmission. While still largely of future concern, similar challenges likely await the malaria community as control efforts continue to reduce the burden of disease, making this programmatic obstacle one of utmost global importance.
Molecular xenomonitoring (MX), the testing of vectors for the presence of parasite genetic material, has been proposed as a non-invasive means of conducting post-MDA surveillance for LF [ 14 , 17 – 18 ]. Although precise correlations between levels of parasite within the vector population and levels within the human population have not been conclusively established, parasite presence within the vector population is indicative of the potential for disease transmission. Furthermore, when monitoring for LF in locations endemic for the Wuchereria bancrofti parasite, a pathogen without a known zoonotic host [ 19 ], presence is directly indicative of active human infection. Yet despite its many advantages, MX is costly and when used for monitoring in a post-MDA setting, typically requires the collection and sampling of many thousands of mosquitoes [ 18 , 20 – 21 ]. Therefore, as a growing number of countries continue to enter the surveillance phases of their LF eradication programs, alternative methodologies for streamlining, simplifying, and reducing the costs associated with post-MDA monitoring will be required.
As an alternative to traditional approaches to MX, excreta and feces produced by mosquitoes potentially harboring parasites can be tested for the presence of pathogen DNA. Previous work has demonstrated that vector feces-monitoring for the PCR-based detection of Trypanosoma cruzi can be used as a means of surveying insect host infection status [ 22 ]. Similarly, it has been shown that genetic material from the Brugia malayi parasite can be successfully detected in the excreta and feces collected from individual mosquitoes [ 23 ]. Building upon these findings, we describe methodological proof-of-principle for the real-time PCR-based monitoring for B . malayi parasite DNA in pools of mosquito excreta/feces as a platform for the surveillance of large numbers of insects. While unconventional, excreta/feces monitoring has the potential to provide significant time, cost, and labor savings over traditional MX methodologies due to its exceptionally high-throughput nature. Furthermore, as excreta/feces collection would likely prove readily adaptable to a variety of both passive and active trapping practices and platforms, its potential feasibility as an exceedingly low cost, long-term surveillance tool is great. Equally promising, initial experiments have demonstrated that this approach to MX can be applied to the detection of Plasmodium vivax DNA, indicating its possible usefulness for the monitoring of both unicellular and multicellular eukaryotic pathogens. Given these encouraging findings, the further exploration of mosquito excreta/feces testing as a new method for disease surveillance purposes is warranted and efforts to adapt this alternative MX approach to other mosquito-borne illnesses should be pursued.
Materials and Methods
Mosquito Rearing for the Accumulation of Excreta/Feces
Accumulation of excreta/feces from mosquitoes potentially infected with B . malayi
Mosquito cartons containing the excreta/feces from female Aedes aegypti mosquitoes potentially infected with the B . malayi parasite were received from the Filarial Research Reagent Resource Center (FR3) located at the University of Georgia, College of Veterinary Medicine, Athens, GA. Rearing and infection of mosquitoes occurred in accordance with “SOP Number 8.3” available on the FR3 website ( http://www.filariasiscenter.org/parasites-resources/Protocols/materials-1 ).
Accumulation of excreta/feces from mosquitoes potentially infected with P . vivax
Mosquito cartons containing the excreta/feces from female Anopheles stephensi mosquitoes potentially infected with the P . vivax parasite were received from the Centers for Disease Control and Prevention, Atlanta, GA, USA. Rearing of mosquitoes occurred in accordance with the protocols described in chapter 2.4 of the “Methods in Anopheles Research” manual [ 24 ] available on the BEI Resources website ( https://www.beiresources.org/Catalog/VectorResources.aspx ). Infection of mosquitoes occurred in accordance with previously described methodologies [ 25 ].
Accumulation of feces from uninfected mosquitoes
Moist filter paper rafts containing Culex quinquefasciatus eggs were received from BEI Resources ( www.beiresources.org ). Eggs were rinsed into open-topped plastic vessels containing approximately 1 L of tap water at a depth of approximately 5 cm and a small volume of standard flake-based fish food was added to the water in each container. Upon maturation into pupae, 50 mosquitoes were transferred into plastic containers, approximately 5 cm in diameter, containing 1 cm of tap water. These containers were then placed into waxed cardboard cartons (approximately 18 cm in diameter by 14.5 cm in height). Cartons were covered with standard mesh tulle and mosquitoes were allowed to emerge as adults. Upon emergence, a cotton ball soaked in 10% sucrose was placed on top of each carton and this solution was refreshed daily. Mosquitoes remained within the cartons producing feces until they expired naturally (10–20 days). At this time, the expired mosquitoes were removed and the cartons were collected, flattened, and stored at 4°C.
Extraction of DNA from Mosquito Excreta/Feces
Preliminary experiments were designed to determine the effectiveness/efficiency of extracting DNA from the excreta/feces of mosquitoes potentially infected with the B . malayi parasite. To make this determination, various extraction protocols and techniques were tested in order to evaluate their efficiency ( Table 1 ). Because the FR3-derived mosquito cartons containing excreta/feces from potentially infected insects were non-waxed, initial samples were either scraped off of the cartons using a metal spatula, or strips of the carton material (hereafter referred to as carton strips) were directly used as the starting material for the extraction procedure. The amplification of B . malayi parasite DNA from all extracts was evaluated using the previously described real-time PCR primer-probe pairing [ 26 ]. Results demonstrated that DNA extractions performed using the QIAamp DNA Micro Kit (Qiagen, Valencia, CA) provided the most consistent and effective detection of parasite DNA. For this reason, this kit was used in all subsequent experiments.
10.1371/journal.pntd.0004641.t001
Table 1 Evaluation of extraction methods for the isolation of DNA from mosquito excreta/feces.
DNA Extraction Method
Quantity of Excreta/Feces (Mosquito Excreta/Feces/Days) *
# of Samples (# of Positives)
Qiagen DNeasy Blood and Tissue (Qiagen, Valencia, CA)
62.5
2 (0)
Overnight Soak in 1 x PBS
62.5
8 (0)
Overnight Soak in 1 x TE
62.5
8 (0)
Phire Plant Direct PCR Kit (Thermo Fisher Scientific, Vantaa, Finland)
62.5
8 (0)
Published Insect Feces Extraction [ 22 ]
1–2
5 (0)
Published Insect Feces Extraction [ 22 ] + Phenol/Chisam Purification
1–2
5 (0)
QIAamp DNA Micro Kit Extraction (Qiagen, Valencia, CA)
1–2
5 (5)
Nucleospin Blood DNA Kit (Macherey-Nagel, Bethlehem, PA)
1–2
5 (0)
Nucleospin Blood DNA Kit (Macherey-Nagel, Bethlehem, PA)
1–2
5 (0)
* Mosquito Excreta/Feces/Days are defined as the estimated quantity of excreta/feces produced by a single mosquito over a 24 hour period.
To adapt the Qiagen protocol for use with the bulky, brittle mosquito carton material, minor modifications were made to the manufacturer’s suggested instructions for DNA extraction from bloodspots. Briefly, carton strips were soaked in 360 μl of Buffer ATL for 1 hour prior to incubation with Proteinase K at 56°C. Additionally, following incubation at 70°C, samples were centrifuged at maximum speed for 5 min and supernatants were transferred to new 1.7 ml microcentrifuge tubes. Tubes were centrifuged for an additional 5 min at maximum speed to pellet residual debris and the supernatants were transferred to QIAamp MinElute columns. Lastly, all samples were incubated in Buffer AE at room temperature for 5 min prior to the elution of samples from the columns.
Evaluation of Positivity of Excreta/Feces from Mosquitoes Potentially Infected with B . malayi
Although preliminary experiments demonstrated that excreta/feces derived from vector mosquitoes fed on B . malayi microfilaria (mf)-positive blood resulted in the amplification of parasite DNA, the availability of mf-containing blood does not guarantee that all mosquitoes will feed or ingest parasites while feeding. Additionally, as the FR3’s standard operating procedure (SOP 8.3) requires that mosquitoes spend three to five days as adults prior to the time an infective blood meal is introduced, a substantial volume of parasite-negative feces was produced and deposited into mosquito cartons prior to blood feeding. Furthermore, as mosquitoes are known to excrete while taking a blood meal [ 27 ], it is likely that excreta would be deposited before parasite DNA had reached/been incorporated into the voided material. Therefore, a portion of the voided material collected from mosquitoes provided with mf-positive blood would likely not contain parasites and would therefore not result in a positive PCR. For this reason, a large panel of potentially positive excreta/feces samples was tested in order to estimate the rates of sample positivity. In total, 59 independent samples were tested, with each sample consisting of a 0.48 cm 2 carton strip. Based upon observations of the volume of excreta/feces produced by single mosquitoes housed in 50 ml conical tubes, it was estimated that the volume of excreta/feces on each carton strip was equivalent to the average volume produced by one to two mosquitoes over a 24 hour period. Negative control extractions were performed on similar volumes of mosquito feces collected from uninfected C . quinquefasciatus . All samples underwent DNA extraction using the modified Qiagen procedure described above and were analyzed by 45 cycles of real-time PCR using the published reagent concentrations and cycling protocol [ 26 ]. 2 μl aliquots of each DNA extract were tested in triplicate and samples returning two or more positive results were considered positive for B . malayi parasite DNA.
Assay Sensitivity Testing
In order to determine detection limits for the presence of B . malayi -infected excreta/feces in large pools of uninfected mosquito feces, a titration of samples was created, with each sample containing a 0.48 cm 2 strip from a carton used to house mosquitoes provided with a B . malayi- positive blood meal mixed with various volumes of uninfected mosquito feces. Feces from uninfected C . quinquefasciatus mosquitoes were removed from cartons using a cotton swab, and the feces-covered cotton was added to each sample. As 50 uninfected mosquitoes were raised in each carton, and adult mosquitoes were observed to survive for a minimum of 10 days (with most surviving considerably longer), it was conservatively estimated that each carton contained a minimum of 500 mosquito feces/days (i.e. the amount of feces produced by 500 mosquitoes in one 24 hour period, or the amount of feces produced by a single mosquito over a 500 day period). While the distribution of feces within cartons was not precisely uniform, by sectioning cartons based upon total internal surface area (approximately 1,050 cm 2 ), it was possible to roughly estimate the number of mosquito feces/days being added to each sample. Samples estimated to contain approximately 62.5, 125, 250, and 500 feces/days were prepared. Negative control extractions were also prepared using mosquito feces collected from uninfected C . quinquefasciatus . All samples were extracted and tested in duplicate reactions using the same extraction and detection methods as described above for the evaluation of PCR positivity testing.
Adaptation of Excreta/Feces Testing to the Detection of P . vivax DNA
To test whether the detection of mosquito-borne pathogen DNA from mosquito excreta/feces was possible for species other than the B . malayi parasite, a set of samples was created using mosquito excreta/feces produced by carton-raised A . stephensi that had been fed on P . vivax -positive blood. As was done for B . malayi detection, samples were prepared by excising 0.48 cm 2 carton strips containing potentially positive excreta/feces. To establish proof-of-principle, 20 samples were prepared and DNA was extracted using the modified Qiagen protocol described above. DNA extracts from each sample were tested using a previously described primer-probe set for the universal detection of Plasmodium species [ 28 ] with reaction recipes and cycling conditions remaining consistent with the authors’ published protocol.
Results
Evaluation of PCR Positivity of Excreta/Feces from Mosquitoes Potentially Infected with B . malayi
Carton strips were excised from containers used to house A . aegypti mosquitoes provided with B . malayi mf-containing blood and testing was conducted to determine the percentage of excreta/feces samples containing B . malayi DNA. Such testing was necessary since the production of feces can occur prior to the provision of an infective blood meal or before the ingestion of a blood meal. Furthermore, the availability of infective blood does not guarantee that each individual mosquito will feed and, dependent upon the mosquito species, localization of parasite material to voided excreta/feces may take time following blood meal ingestion. Accordingly, DNA was extracted from 59 independent samples, each consisting of a carton strip measuring 0.48 cm 2 and containing excreta/feces from one to two mosquitoes over a 24 hour period (i.e. one to two mosquito feces/days). Real-time PCR testing, using 2 μl of template DNA resulted in positive detection for 21 out of 59 samples tested (35.6%). For positive samples, mean Ct values ranged from 26.62 (+/- 0.24) to 41.98 (+/- 0.03) ( Table 2 ). Because only a fraction of the deposited mosquito excreta/feces would contain parasite DNA, 35.6% may be a true indication of the frequency of positive samples.
10.1371/journal.pntd.0004641.t002
Table 2 PCR positivity of excreta/feces from mosquitoes potentially infected with B . malayi .
Sample #
Ct Value (Std. Dev.)
Sample #
Ct Value (Std. Dev.)
1
Negative
32
Negative
2
27.75 (+/- 0.15)
33
Negative
3
28.67 (+/- 0.17)
34
Negative
4
Negative
35
Negative
5
38.25 (+/- 2.79)
36
Negative
6
Negative
37
Negative
7
Negative
38
Negative
8
36.53 (+/- 2.33)
39
39.07 (+/- 1.07)
9
40.49 (+/- 0.12)
40
Negative
10
Negative
41
Negative
11
34.80 (+/- 0.70)
42
Negative
12
37.55 (+/- 0.91)
43
Negative
13
41.96 (+/- 2.67)
44
40.74 (+/- 2.17)
14
40.54 (+/- 1.26)
45
Negative
15
Negative
46
31.69 (+/- 0.50)
16
Negative
47
Negative
17
Negative
48
38.88 (+/- 0.49)
18
Negative
49
37.69 (+/- 0.69)
19
Negative
50
Negative
20
Negative
51
40.32 (+/- 0.43)
21
Negative
52
Negative
22
Negative
53
37.49 (+/- 1.53)
23
Negative
54
Negative
24
Negative
55
39.48 (+/- 2.37)
25
41.44 (+/- 0.95)
56
26.62 (+/- 0.24)
26
27.91 (+/- 0.54)
57
Negative
27
Negative
58
Negative
28
41.98 (+/- 0.03)
59
Negative
29
Negative
Negative Extract #1
Negative
30
Negative
Negative Extract #2
Negative
31
Negative
Assay Sensitivity Testing
A titration of samples containing potentially positive 0.48 cm 2 carton strips mixed with varying amounts of uninfected mosquito feces was prepared in order to estimate the limits of detection for B . malayi-based excreta/feces testing. In total, five samples containing an estimated 62.5 mosquito feces/days, six samples containing an estimated 125 mosquito feces/days, six samples containing an estimated 250 mosquito feces/days, and two samples containing an estimated 500 mosquito feces/days were assayed. As expected, due to the uncertainty of which samples actually contained B . malayi DNA, a fraction of the samples failed to give positive PCR detection of B . malayi DNA. However, detection of parasite DNA proved possible at all tested levels of sensitivity ( Table 3 ).
10.1371/journal.pntd.0004641.t003
Table 3 Limits for the detection of B . malayi DNA in mosquito excreta/feces samples.
Sample ID
Quantity of Potentially Positive Excreta/Feces (Mosquito Excreta/Feces/Days) *
Quantity of Negative Excreta/Feces (Mosquito Excreta/Feces/Days) *
Ct Value (Std. Dev.)
A
1–2
62.5
Negative
B
1–2
62.5
Negative
C
1–2
62.5
Negative
D
1–2
62.5
37.89 (+/- 2.31)
E
1–2
62.5
38.77 (+/- 0.20)
F
1–2
125
30.98 (+/- 0.20)
G
1–2
125
Negative
H
1–2
125
Negative
I
1–2
125
Negative
J
1–2
125
Negative
K
1–2
125
Negative
L
1–2
250
29.56 (+/- 0.01)
M
1–2
250
35.88 (+/- 0.04)
N
1–2
250
Negative
O
1–2
250
38.73 (+/- 0.91)
P
1–2
250
Negative
Q
1–2
250
Negative
R
1–2
500
Negative
S
1–2
500
38.40 (+/- 1.03)
Negative #1
N/A
62.5
Negative
Negative #2
N/A
62.5
Negative
Negative #3
N/A
62.5
Negative
* Mosquito Excreta/Feces/Days are defined as the estimated quantity of excreta/feces produced by a single mosquito over a 24 hour period.
Adaptation of Excreta/Feces Testing to the Detection of Plasmodium vivax DNA
To explore whether excreta/feces testing would efficiently detect pathogen DNA from species other than B . malayi , testing for the presence of the human malaria-causing parasite P . vivax was performed. To demonstrate proof-of-concept, 20 samples were prepared and tested by PCR. Each sample was comprised of a 0.48 cm 2 carton strip excised from a mosquito container having housed A . stephensi female mosquitoes provided with Plasmodium -positive blood. Real-time PCR testing of DNA extracted from each sample clearly demonstrated the adaptability of excreta/feces testing to the detection of P . vivax since all samples were positive with Ct values ranging from 26.82 (+/- 0.26) to 29.21 (+/- 0.80) ( Table 4 ).
10.1371/journal.pntd.0004641.t004
Table 4 PCR positivity of excreta/feces from mosquitoes infected with P . vivax .
Sample #
Ct Value (std. dev.)
Sample #
Ct Value (std. dev.)
1
28.23 (+/- 0.58)
12
28.77 (+/- 0.09)
2
27.84 (+/- 0.05)
13
28.83 (+/- 0.49)
3
27.85 (+/- 0.15)
14
27.88 (+/- 0.19)
4
27.85 (+/- 0.29)
15
27.36 (+/- 0.14)
5
26.82 (+/- 0.26)
16
26.82 (+/- 0.10)
6
27.98 (+/- 0.28)
17
28.00 (+/- 0.27)
7
26.94 (+/- 0.20)
18
28.48 (+/- 0.29)
8
27.26 (+/- 0.32)
19
28.40 (+/- 0.26)
9
27.19 (+/- 0.32)
20
28.07 (+/- 0.12)
10
28.37 (+/- 0.49)
Negative Extract #1
Negative
11
29.21 (+/- 0.80)
Negative Extract #2
Negative
Discussion
While sensitive and less intrusive to the local population than human sampling, the number of studies implementing current MX practices for the surveillance of LF or malaria has been limited. Although such efforts provide valuable data [ 10 , 17 – 18 , 21 ] the routine use of MX for post-MDA LF surveillance or long-term recrudescence monitoring is not yet standard procedure. Despite the existence of effective molecular tools [ 28 – 29 ], vector monitoring for malaria is even more uncommon and World Health Organization recommendations for infection monitoring and prevalence estimation rely solely on human sampling [ 2 ]. Limited implementation has occurred for multiple reasons, including the need to process and test large numbers of mosquitoes from areas suspected of having low parasite density within the vector population [ 10 , 18 , 21 ]. Difficulties in establishing a concrete correlation between vector-parasite levels and human prevalence have further restricted MX implementation [ 21 ]. Yet despite these shortcomings, MX continues to receive attention as the need for post-intervention disease surveillance continues to grow and mosquito trap designs continue to improve [ 30 – 34 ]. Accordingly, methodologies capable of harnessing the advantageous aspects of MX while making its practice more practical and inexpensive would be of great benefit to global LF and malaria elimination efforts, as well as to monitoring efforts for other vector-borne diseases.
The work presented here provides methodological proof-of-concept for a novel approach to MX with the potential to greatly reduce the cost, time, and labor associated with large-scale surveillance efforts. The successful amplification of parasite DNA from pooled mosquito excreta/feces containing B . malayi genetic material has demonstrated that high-throughput MX for LF is feasible. In the past, real-time PCR-based MX for the presence of the filariasis-causing parasites has been restricted to the testing of pools of 25 or fewer mosquitoes. This is because the biological mass of mosquitoes and high yields of mosquito DNA associated with pools of large size results in the inability to detect the presence of small quantities of parasite DNA [ 35 ]. However, excreta/feces testing enables the sampling of material obtained from vast numbers of mosquitoes, while simultaneously limiting the biological mass associated with each sample. As we have demonstrated, it is possible to detect trace amounts of parasite DNA in pools containing the voided material from as many as 500 uninfected mosquitoes. Future studies implementing this approach will benefit from the drastic reduction in cost of DNA extractions and PCR (approximately 20-fold). Furthermore, as it has been shown that non-vector mosquitoes rid themselves of parasite material more rapidly than vector species (as indicated by a shortened period of time during which parasite detection is possible within non-vectors [ 23 ]), one would expect to find greater quantities of parasite DNA within the excreta/feces of non-vector mosquitoes. Therefore, the testing of mixed pools of vector and non-vector excreta/feces should be possible. While such testing will result in reduced ability to directly correlate the presence of parasite with individual vector species, it will likely increase the sensitivity of detection when surveying for the presence of parasite in post-transmission-interruption settings as both vector and non-vector mosquitoes potentially harboring parasite material will be screened. In addition, it is likely that excreta/feces testing will eliminate the need for the labor intensive and time consuming species-sorting efforts which are commonplace in current MX work [ 10 , 17 – 18 , 21 , 36 ]. By drastically reducing the numbers of pools that must be screened and by eliminating the need for sorting mosquitoes by species, labor-related time and costs are dramatically reduced.
While operationalizing this alternative approach to MX presents some implementation hurdles, adaptation of current passive and active trapping methods to the collection of mosquito excreta/feces is possible. Such adaptation could occur by transferring live mosquitoes from a trap to a holding carton, in which they would be sugar fed using a cotton ball, thereby encouraging the voiding of waste material. Expired mosquitoes would then be removed and additional mosquitoes could be added following further collection from the trap. Periodic testing of the accumulated excreta/feces would enable the high-throughput screening of the voided material from a series of such traps. Any trap with the capacity to maintain live mosquitoes could be used for this purpose including the CDC Gravid Trap, the Ifakara tent trap and others [ 30 , 37 ]. Alternatively, collection of excreta/feces could occur directly within traps of various designs. One such design proving readily adaptable to excreta/feces collection in preliminary experiments is the “Large Passive Box Trap” developed by Ritchie, et al [ 38 ]. While work aimed at evaluating the adaptability of this trap to the collection of various species of mosquitoes is currently ongoing, and further efforts to optimize this trap for the purpose of excreta/feces collection will be required, simply lining the internal surfaces of this passive trap with waxed paper provides an uncomplicated method for collecting the accumulated material voided by trapped mosquitoes ( S1 Fig ). Swabbing the excreta/feces from the waxed paper then enables the PCR analysis of pooled material.
Additional testing will be required to determine the stability of parasite DNA in mosquito excreta/feces over time and under field conditions. However, in the proof-of-concept experiments described in this paper, mosquito excreta/feces containing parasite DNA was allowed to accumulate for 14–16 days prior to transfer to cold storage. In this setting, parasite DNA remained stable and detectable ( Table 3 ). While further validation under conditions mimicking tropical temperatures and humidity will be required, these results are encouraging, as DNA stability within tropical and sub-tropical climates could present another hurdle when operationalizing this method in the field.
Since production of feces can occur prior to the provision of a parasite-positive blood meal and since this provision does not ensure that all mosquitoes will ingest and/or metabolize a parasite, a percentage of the excreta/feces samples collected will likely test negative for parasite DNA. It is therefore difficult using blood-fed mosquitoes to definitively assess the consistency of detection of parasite DNA in excreta/feces. During initial testing, we demonstrated that 21 out of 59 samples comprised of 0.48 cm 2 carton strips derived from containers used to rear mosquitoes with a B . malayi -positive blood source were positive ( Table 2 ). However, although sufficient to fulfill our primary aim of providing methodological proof-of-concept, it cannot be conclusively determined whether the remaining 38 samples were all truly negative for parasite DNA. While spiking uninfected excreta/feces samples with extracted B . malayi genomic DNA would provide clear positive and negative samples, this approach is extremely artificial and has decreased biological relevance since it eliminates any possible effects of mosquito metabolism on the integrity of parasite DNA. Since the major uses of excreta/feces testing will likely center on mapping and long-term, low-cost, post-transmission-interruption recrudescence monitoring, marginally reduced consistency of detection has diminished significance as continuous, sustainable, high-throughput surveillance would enable detection of even low-levels of parasite prevalence. The high-throughput nature of this testing was clearly demonstrated by the positive detection of parasite DNA derived from pools containing various volumes of negative feces up to 500 mosquito feces/days ( Table 3 ). Detection proved possible at all tested sensitivity levels and with overall sample positivity rates similar to those obtained when testing potentially positive excreta/feces samples without the addition of negative feces (36.8% vs. 35.6% respectively). Thus, the inclusion of large amounts of negative feces does not appear to alter detection efficiency. Given these findings, sustainable, high-throughput surveillance efforts using excreta/feces screening could serve as a “first-alert” platform, with positive detection serving as a “red flag” for recrudescence in settings of known transmission interruption. In such a scenario, detection would spur the implementation of more traditional surveillance and monitoring studies.
By successfully detecting P . vivax DNA in pools of excreta/feces produced by Plasmodium -positive-blood fed A . stephensi , we have provided proof-of-principle for the application of this platform to the detection of malaria parasites. Furthermore, the increased rates of sample positivity and decreased Ct values seen when assaying for P . vivax are not entirely surprising and indicate this system may work even better for malaria than LF. Estimates have suggested that the ratio of Plasmodium merozoites to gametocytes within the peripheral blood is as great as 156:1 [ 39 – 40 ]. Given this ratio, the vast number Plasmodium merozoites ingested during a blood meal (up to 32 per infected erythrocyte [ 41 ]), and knowledge that merozoites obtained during blood feeding are unable to undergo further development within the mosquito host (only gametocytes undergo further development [ 42 ]), the great majority of ingested parasites are simply metabolized and/or eliminated by the mosquito. In contrast, while mosquito hosts possess measures that provide partial protection against filarial infection [ 43 – 44 ], and environmental conditions are thought to impact rates of parasite survival [ 45 ], all filarial parasites taken up as part of a blood meal are of the correct lifecycle stage (mf) to potentially undergo further development within the vector host. Therefore, due to the varying natures of their lifecycles, it follows that a greater percentage of filarids ingested during a blood meal are able to successfully develop within the mosquito host as compared to Plasmodium . Since successful parasite development would likely mean the absence of parasite DNA in mosquito excreta/feces, the lower levels of sample positivity and the more modest Ct values observed during B . malayi testing compared to P . vivax testing seem logical.
With its adaptability to both B . malayi and P . vivax , MX of mosquito excreta/feces for various other mosquito-borne pathogens should be explored. Given the successes realized with the detection of these parasites, it is extremely likely that similar detection will prove possible for W . bancrofti and other malaria species. However, the applicability of this new platform to other types of pathogens should also be examined, since improved high-throughput screening for RNA viruses such as Dengue, Chikungunya, and Zika would be welcomed programmatic tools. Furthermore, since all species of biting insects draw from the same reservoir of blood within a target host, the possibility of cross-vector monitoring should also be considered. For example, excreta/feces samples collected from mosquitoes could be monitored for the presence of disease-causing agents having unrelated insect hosts (such as Leishmania ssp. or Loa loa ). Adaptability to various pathogens and the possibility of cross-vector monitoring could also make excreta/feces sampling an attractive strategy for tropical disease integration efforts. In light of these factors, and the potential time, cost, and labor savings associated with such applications, we believe that this proof-of-concept study suggests that further evaluation of this new method is warranted.
Supporting Information
S1 Fig
Mosquito excreta/feces collection in the modified “Large Passive Box Trap”.
Voided material from mosquitoes entering the passive trap collects on wax paper lining the trap surfaces below. Outlined in red is one such mosquito and a series of excreta/feces “spots” which has accumulated.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The majority of HIV infections by heterosexual transmission occur in adult and adolescent women across the female genital tract (FGT) mucosa [ 1 , 2 ]. Understanding factors contributing to HIV acquisition at the main site of infection in women is critical for developing effective biomedical HIV prevention interventions such as pre-exposure prophylaxis (PrEP) strategies, microbicides, and vaccines, as well as for evaluating factors that may alter HIV acquisition risk, such as hormonal contraception.
Within the FGT mucosa, the number and type of cellular targets, primarily CD4 + T cells expressing the cell surface receptor C-C chemokine receptor type 5 (CCR5, the primary HIV co-receptor), predicts susceptibility to HIV infection [ 3 , 4 ]. In the FGT mucosa, these markers are increased compared to the blood and penile mucosa [ 5 – 7 ], potentially explaining women’s increased risk of HIV acquisition during unprotected vaginal sex compared to men [ 8 ]. In addition to HIV co-receptor expression, CD4 T cells are heterogeneous with regards to their HIV susceptibility, and certain T cell phenotypes such as activated T cells [ 9 ] and cells expressing mucosal trafficking markers (such as α 4 β 7, a heterodimeric integrin receptor involved in T cell migration into the gut associated lymphoid tissue) [ 7 ] have been associated with HIV acquisition in experimental models [ 10 , 11 ] and may be over-expressed in the FGT compared to blood.
Despite knowledge of these cellular markers of HIV susceptibility, inter and intra-individual variability in these markers is not well described but is critical in the design and interpretation HIV biomedical prevention studies. Of note, longitudinal studies of these markers have been hindered by challenges associated with mucosal sampling, in which trauma induced by invasive sampling (such as biopsy collection) may affect subsequent mucosal immune characterization. Furthermore, while HIV target CD4 T cells in the systemic, lymphoid, rectal, and female genital mucosal compartments have been frequently studied, detailed knowledge about specific CD4 T cell subsets, phenotypes, and functional characteristics likely to facilitate HIV acquisition in the mucosa remains limited, partially due to limited cellular yield to conduct more in-depth phenotypic and functional analyses. Recently, McKinnon et al compared absolute yields of mononuclear leukocyte subpopulations from genital leukocytes obtained by cervicovaginal lavage (CVL), endocervical cytobrushes, and cervical biopsy and illustrated that cell yields for all leukocyte populations tested from two serially collected cytobrushes were significantly higher than CVL and comparable to those obtained from a single cervical biopsy [ 12 ]. However, T cell yield was significantly lower in cytobrush samples compared to biopsy, raising the question of whether cytobrush samples are suitable for in-depth characterizations of CD4 T cells within the FGT which may require high cellular yields for assays.
Because characterization of FGT-derived CD4 T cells using minimally-invasive sampling methods is important for longitudinal evaluation of these cells during biomedical HIV prevention studies, we compared CD4 T cell yields and in-depth T cell subset phenotypes from cells collected longitudinally using three different non-biopsy, minimally-invasive genital tract sampling methods: CVL, two endocervical flocked swabs (FS), and each of two sequential endocervical cytobrush (CB) collections. Additionally, we determined whether two consecutive cytobrushes provide sufficient T cell yields for in-depth phenotypic characterization of FGT CD4 T cell subsets.
Methods
Study design and participants
We recruited HIV seronegative women between 18–44 years old with an intact uterus and cervix who were either enrolled in the Atlanta site of the Women’s Interagency HIV Study (WIHS) or who met risk factors for HIV acquisition as per enrollment criteria for the cohort (a. injection drug use or use of crack, cocaine, heroin, or methamphetamine; b. diagnosed with a sexually transmitted infection; c. unprotected sex with 3 or more men; d. having sex for drugs, money, or shelter; e. sex with a known HIV-positive man; f. having a partner meeting any of the preceding criteria). Most of the recruited participants were African American, who are disproportionately affected by HIV in the Southern US, and are representative of the demographic of the Atlanta WIHS cohort. Negative HIV serostatus was confirmed at the time of screening using an FDA-approved HIV antibody test. Women were excluded if they were pregnant, currently using systemic hormonal contraceptives, had a symptomatic vaginal infection or genital ulcer disease at screening or treatment for vaginal infection in the preceding two weeks, an active malignancy, were using immunosuppressive medications, had surgery within the preceding 2 months, or had a cervical procedure within the previous 30 days. This protocol was approved by the Institutional Review Board at Emory University and the Grady Research Oversight Committee. Written informed consent was obtained from all participants.
Specimen collection
We collected paired blood and cervicovaginal samples during the follicular (between 7–10 days from the start of the previous menstrual cycle) and luteal (between 21–25 days from the start of the previous menstrual cycle) phases for up to five consecutive visits. Sexual, reproductive, medication, and genitourinary symptom histories were collected at each study visit. Cervicovaginal specimens were collected by the same clinician for all participants during speculum pelvic examination in the following distal to proximal order to reduce any potential effect of one collection method on the subsequent[ 13 ]: 1) CVL with 10ml phosphate buffered saline (PBS) directed toward the cervical os i.e., the opening in the center of the ectocervix, and vaginal walls x 1 minute lavage and placed in 5mL human AB serum (Atlanta Biologicals), 2) two sequential endocervical FS (COPAN Innovation, Italy) placed in 20mL complete media (RPMI -1640, penicillin-streptomycin, L-glutamine (Gibco) with 10% heat inactivated fetal bovine serum [FBS]), 3) two sequential endocervical CB collected in 20mL complete media each. Specimens were collected and immediately transported on ice for real-time processing and immune phenotyping. Presence of visible blood contamination was assessed in all FGT specimens at the time of sample collection, recorded on a scale of 1–10 at the time of processing for FS and CB; CVL was tested for blood and leukocytes using Mutistix 8SG urinalysis strips (Siemens Healthcare, Los Angeles, CA) and for semen using the ABACard p30 antigen detection test (Abacus Diagnostics, West Hill, CA). Vaginal swabs were tested for sexually transmitted infections (STIs) including Neisseria gonorrhea , Chlamydia trachomatis , Trichomonas vaginalis , and HSV-1/2 DNA via multiplex PCR in a CLIA-certified laboratory. Blood was collected in 8 mL-sodium citrate-containing CPT vacutainer tubes (BD, Franklin Lakes, NJ).
Sample processing
All samples were processed within three hours of collection. CVLs and CBs were processed at 4°C as described by McKinnon et al [ 12 ]. Briefly, CVL samples were centrifuged at 1500 rpm for 5min, the cell pellet was resuspended in 10 ml complete media and the cell suspension was strained through 100 μm cell strainer up to 3 times, washed once with 10ml complete medium and resuspended in complete medium at 20 million cells/ml. Two million cells were used for flow cytometry staining. FS were washed in complete media to extract cells into a 50ml conical tube and were strained through a 100 μm strainer once. CBs were inserted into 25ml serological pipette containing 20ml complete media. The CB was moved in and out of the pipette tip to dislodge the cells, while the CB was being washed in complete media to extract cells into a 50ml conical tube. Cells were then strained through a 100 μm strainer once. Both FS and CB cell suspensions were washed once with 10ml complete medium and resuspended in 100μl of complete medium and used for flow cytometric staining. Cell suspensions were placed on ice at all times to preserve cell viability. After determination of cell counts, cells were used immediately for flow cytometry. PBMCs were isolated from whole blood collected in sodium citrate tubes and isolated by density gradient centrifugation according to standard procedures as described previously [ 14 ].
Flow cytometry
Staining on whole blood was done at room temperature while PBMCs and cells collected from CVL, FS, and CB were stained in PBS containing 2% FBS for 30 min at 4°C. In addition, leukocytes obtained from the first and second CB were separately assayed to compare CD4 T cell yield and phenotype between individual CBs. Cells were stained with fluorochrome-conjugated antibodies specific for CD45 (2D1), CD4 (OKT4), CD8 (SK1), CCR5 (2D7), CXCR4 (12G5), CD27 (M-T271), CD69 (L78) from BD Pharmingen (San Jose, CA); CD45RO (UCHL1), FOXP3 (236A/E7) from eBioscience, (San Diego, CA); CD38 (HIT2) from Invitrogen (Grand Island, NY); and α 4 β 7 from the NIH Nonhuman Primate Reagent Resource. Dead cells were excluded from the analysis based on staining for Live/Dead Near-IR dead cell stain from Molecular Probes, Invitrogen. Staining for FOXP3 was performed after cells were stained for surface antigens followed by permeabilization/fixation using the FOXP3 kit and protocol, followed by intracellular staining. Samples were acquired on a LSR Fortessa (BD Biosciences) and all cellular events were collected for mucosal samples while 500,000 events were collected for samples from blood.
Statistical analysis
Data were analyzed using FlowJo software v X.0.7 (Tree Star, Inc., Ashland, OR) after gating out dead cells. CD45 + , CD4 + , and CD8 + T cell yields from each mucosal sampling method were summarized and compared after log transformation. T cell subset frequencies and frequency of cell-surface markers (HIV co-receptors, activation markers, and mucosal trafficking markers) were compared between each mucosal sampling method and blood. Expression of activation and mucosal trafficking markers were compared from CB and blood-derived CD45RO + CCR5 + cells versus CD45RO + CCR5 - cells. Finally, the frequency of FOXP3 + CD4 T regulatory cells and expression of CCR5 on these cells was compared between CB- versus blood-derived CD4 T cells. For the longitudinal data, separate linear mixed effects models with random effects for participant and visit were developed for each T cell subset/ cell surface marker of interest. One thousand bootstrap repetitions were used to calculate the 95% confidence interval for correlations from repeated measures data, and statistical significance was determined by p ≤ 0.05 using correlation analysis of the first observation per participant. Wilcoxon signed rank test was used to compare paired specimens within a participant when data from only one visit was used. Statistical analyses were performed in SAS version 9.4.
Results
We enrolled HIV negative women for up to five consecutive bimonthly visits. We excluded 5 specimens with high visible CVL blood contamination (n = 4) and limited CVL cell yields (less than 100 CD45 + cells, n = 1) from all subsequent analyses. Demographic, clinical, and visit characteristics of the 12 included participants contributing 33 visits are shown in Table 1 . The median participant age was 35 years, all were African American, and the majority reported condomless sex in the last 6 months. Vaginal sex was reported during the week before 48% of study visits. Few genital tract infections were noted during study visits, but the vaginal pH was over 4.5 in over half of study visits. Any visible blood was noted at the time of collection in 6%, 41%, and 88% of CVL, FS, and CB samples, respectively.
10.1371/journal.pone.0178193.t001
Table 1 Demographic, clinical, and visit characteristics among participant visits included in the analysis.
Participant Characteristic
Number of participants (%) or median (IQR) N = 12
Age
35 (33–38)
African American race
12 (100%)
Condomless vaginal sex in the last 6 months
10 (83%)
Self-reported sexually transmitted infection a in the last 6 months
3 (25%)
Visit Characteristic
Number of visits (%) or median (IQR) N = 33
Visit during follicular phase b
15 (46%)
Symptomatic genital infection c
0
Asymptomatic sexually transmitted infection
Gonorrhea
3 (9%)
Chlamydia
3 (9%)
HSV 1/2 DNA
1 (3%)
Trichomonas
1 (3%)
pH
Median pH (IQR)
4.7 (4.0–5.0)
pH > 4.5
17 (52%)
Self-reported vaginal sex within 1 week of study visit
14 (48%)
Semen contamination
3 (9%)
Any visible blood noted
CVL
2 (6%)
Flocked swab
13 (41%)
Endocervical cytobrush
29 (88%)
a Gonorrhea, syphilis, chlamydia, pelvic inflammatory disease, genital herpes, genital warts, or trichomonas.
b Based on self-reported last menstrual period (follicular phase visits scheduled 7–10 days after onset of menses +/- 3 days; luteal phase visits scheduled 21–25 days after onset of menses +/- 3 days).
c Presence of purulent cervicovaginal discharge or ulcerative vaginal lesions during participant visit.
CD4 T cell recovery is highest from endocervical cytobrushes relative to endocervical flocked swabs and cervicovaginal lavages
Fig 1A shows a representative flow plot of the gating strategy used to identify CD4 T cells based on live CD45 + CD8 - leukocytes. For T cell yield analyses, we additionally excluded 8 specimens with inadequate staining for analyses and quantified the absolute number of CD45 + , CD4 + , and CD8 + cells isolated from each collection method from 9 participants over two (2 participants) to three visits (7 participants) for a total of 25 participant visits ( Table 2 ). Parameters in samples excluded versus included in the analysis are shown in S1 Table . The median visual CB blood score (assigned 0–10 at the time of sample processing) was at least 7 for 15% and 8% for included versus excluded samples, respectively.
10.1371/journal.pone.0178193.g001
Fig 1
Highest CD4 T cell recovery from endocervical cytobrushes relative to endocervical flocked swabs and cervicovaginal lavages.
(A) Example analysis of human genital mucosal specimens from 9 women sequentially sampled for cervicovaginal lavage (CVL), flocked swab (FS) followed by two consecutive cytobrushes (CB1, CB2). The cellular components were isolated by centrifugation and incubated with antibodies to identify T lymphocyte subsets at 4°C for 30 minutes. Whole blood from the same individual was stained as a positive control. Scatter plots show data for nine participants over two-three independent visits at either the follicular or the luteal phase of the menstrual cycle. Participants are color coded for correlation plots. (B) CD4 T cell yields significantly higher from CB sampling relative to either CVL or FS, CD4 T cell yields from CB are not associated with yields from CVL. (C) CD4 T cell yields from FS correlate with CB yields, and CB1 and CB2 yields strongly correlate with each other. (D) CD4 T cell distribution profile across sampling methods, frequency of CD4 + cells similar and correlated between CB1 and CB2. ***, p < 0.001; **, p < 0.01, *, p < 0.05.
10.1371/journal.pone.0178193.t002
Table 2 Median (IQR) cell numbers of leukocyte subsets and ratio of CD4 to CD8 T cells in sequential cervicovaginal lavages (CVL), two consecutive flocked swabs (FS), and first and second endocervical cytobrushes (CB1 and CB2), respectively. Data are from 25 participant visits.
CVL
FS
CB1
CB2
CD45 +
8,526 (1,980–16,750)
5,542 (2008–38,910)
67,021 (20,116–407,211)
277,944 (46,142–796,235)
CD4
1,364 (402–4534)
1,697 (644–144,97)
29,501 (7,659–135,699)
85,653 (23,454–294,090)
CD8
782 (351–1794)
864 (291–5560)
8950 (2,505–43,116)
24031 (7,708–167,322)
CD4:CD8
1.7 (0.9–2.8)
1.4 (1.1–2.2)
1.7 (1.1–2.8)
1.6 (1.1–2.9)
Among all three methods tested, CB yielded the highest number of lymphocytes, CD4 T cells and CD8 T cells. The cell yield from CB2 was as good as (or in many cases was higher than) the yield from CB1. Compared to either the first or the second CB, CVL yielded markedly lower total CD45 + cells (median CVL 8,526; CB1 67,021; CB2 277,944; p < 0.001 for CVL vs. either CB1 or CB2), CD4 + cells (median CVL 1,364; CB1 29,501; CB2 85,653; p < 0.0001 for CVL vs. either CB1 or CB2), and CD8 + cells (median CVL 782; CB1 8,950; CB2 24,031; p < 0.01 for CVL vs. either CB1 or CB2). Cell yields from CVL were comparable to those obtained from two combined FS (FS median CD45 + 5,542; CD4 + 1,697; CD8 + 864. ( Fig 1B ).
To ascertain whether prior CVL or FS affected CB cell recovery, we determined correlations between cell yields from CVL, FS, and CB. Consistent with the distinct FGT compartments sampled by CVL and CB (cervicovaginal versus endocervical), we observed no association in CD4 T cell yields between the two methods ( Fig 1B ). In addition, the cellular yields from the CB samples in our study were comparable or higher than those described in other studies where CVLs were not collected prior to CB [ 12 , 13 ], indicating that CVL followed by CB represents a feasible sampling strategy for assessing T cell immunobiology of the genital mucosa. In a comparison with CB samples collected without preceding FS from 6 women over 12 visits, we noted that CB yields obtained were not significantly reduced by a prior FS ( S1 Fig ). CD4 T cell yields from the FS positively correlated with yields from CB indicating similarity in compartments accessed by both procedures ( Fig 1C ). Consistently, CD4 T cell yields from the first and second CB were strongly correlated with each other ( Fig 1C ). In terms of T cell distribution, CD4 T cells (as % of CD45 + ) in CB were comparable to peripheral blood, both of which were enriched for CD4 T cells relative to CVL and FS ( Fig 1D ). We noted a strong correlation in distribution of CD4 T cells between CB1 and CB2, indicating similarity of subpopulation profiles between the two methods ( Fig 1D ). In summary, CB sampling yields higher numbers of CD4 T cells compared to FS and CVL, and two consecutive CBs provide highest recovery of CD4 T cells.
The female genital tract mucosa is enriched with CD4 T cells expressing CCR5 and markers of activation relative to blood across all FGT sampling methods
We next examined CD4 T cells for expression of markers significant for HIV acquisition, HIV co-receptors, binding proteins, and markers of activation, from cells derived from each sampling method. Characteristic of enrichment of antigen-experienced T cells within effector compartments, a majority of CD4 T cells (> 85%) within the genital compartment regardless of sampling method expressed CD45RO relative to an average of 50% within peripheral blood CD4 T cells ( Fig 2A ). Therefore, to obtain an accurate representation of CD4 T cell phenotype between peripheral and mucosal compartments, phenotypic analysis was performed on memory (CD45RO + ) CD4 T cells. We used a cut-off of 100 events within the CD45RO + CD4 + gate as criterion for further phenotypic analysis. Examination of CCR5 (histogram, Fig 2A and 2B ) demonstrated significant enrichment of CCR5 + CD4 T cells in genital CD4 T cells derived from all FGT sampling methods (median CVL 35.1%; FS 37.9%; CB1 39.2%; CB2 32.7%) relative to peripheral blood (median 13.9%; p < 0.001). The frequency of CCR5 + memory CD4 T cells did not differ across the 3 genital sampling methods. In contrast, CXCR4 expression was significantly higher within memory CD4 T cells in the blood (median 36.8%) compared to CVL (median 29.7%; p < 0.05), FS (median 19.6%; p < 0.0001), CB1 (median 21.3%; p< 0.001), and CB2 (median 20.3%; p < 0.001). Within genital sampling methods, CXCR4 expression was significantly higher in CVL relative to FS, CB1, or CB2 (all p< 0.01). Expression of α 4 β 7 was comparable in mucosal samples and blood with a median of 25–35% ( Fig 2B ), but we noted that per-cell expression of α 4 β 7 was higher in blood relative to mucosal samples ( Fig 2A , histogram ) indicating possible down-regulation of α 4 β 7 upon migration to genital mucosa. Consistent with both methods sampling endocervical areas, the expression profiles of CCR5, CXCR4, α 4 β 7 and on CD4 T cells correlated between FS and CB samples ( Fig 2C ) but not between CVL and CB ( S2 Fig ). CD4 T cells derived from CB samples without visible blood had higher CCR5 and α 4 β 7 expression than those with visible blood ( S3 Fig ). Together these data indicate that HIV target cells (CCR5 + CD4 T cells) are highly enriched in the FGT, and CD4 T cells derived from all 3 genital sampling methods express HIV binding proteins.
10.1371/journal.pone.0178193.g002
Fig 2
Enrichment of CD4 T cells expressing CCR5 and markers of activation in genital mucosa relative to blood.
(A) Representative flow plots show that CD4 T cells from CVL, FS and CB are highly enriched for antigen-experienced CD4 T cells as evidenced by expression of CD45RO. Histograms show comparison of expression of CCR5, CD38, CXCR4, CD69, α 4 β 7 and CD27 on genital and whole blood (WB) CD4 T cells. Naive CD4 T cells are overlaid in grey. (B) Distribution of HIV co-receptors CCR5 (n = 16 visits for CVL, 24 visits for FS, CB1, CB2, and WB), CXCR4 (n = 19 visits for CVL, 21 visits for FS, 22 visits for CB1, CB2, and WB), and expression of the integrin α 4 β 7 (n = 22 visits for CVL, 27 visits for FS, CB1, CB2, and WB) on mucosal and blood CD4 T cells. (C) Correlation of CCR5, CXCR4 and α 4 β 7 expression on FS and CB. (D) Distribution of activation markers CD38 (n = 21 visits for CVL, 25 visits for FS, CB1, CB2, and WB), CD69 (n = 14 visits for CVL, 15 visits for FS, CB1, CB2, and WB) and CD27 (n = 19 visits for CVL, 23 visits for FS, CB1, CB2, and WB) on mucosal and blood CD4 T cells (****, p < 0.0001; **, p < 0.01; *, p < 0.05).
Next, we examined expression of markers of T cell activation across compartments. Expression of the cyclic ADP ribose hydrolase, CD38, was significantly higher in CVL-derived CD4 T cells ( Fig 2D , median 21.5%) relative to blood (median 8.2%; p < 0.05), FS (median 9.4%; p < 0.05), and CBs (CB1 median 7.89%, CB2 8.48%; both p < 0.05). Expression of CD69, the acute marker of T cell activation and tissue retention, revealed that a greater proportion of FGT associated CD4 T cells derived from CVL (35.7%), FS (33.6%), CB1 (19.9%), and CB2 (20.3%) expressed CD69 relative to blood (0.97% p < 0.01). This attribute is consistent with the phenotype of T cells in vaginal mucosa, which are typically identified based on expression of CD69 [ 15 ]. Finally, consistent with an effector-memory phenotype, CD4 T cells within the FGT expressed significantly lower frequencies of CD27 relative to cells in the blood (median CVL, 26.5%; FS, 19.25%; CB1, 41.2%; CB2, 42.9%; blood, 53.1%). Within genital sampling methods, CVL-derived CD4 T cells expressed higher CD38, higher CD69, and lower CD27 frequencies relative to FS or CB-derived CD4 T cells. There were no differences in activation marker expression on CD4 T cells derived from CB samples without versus with visible blood ( S3 Fig ). Together, these data show enrichment of activated (CD69 + CD27 lo ) CD4 T cells in the FGT with higher enrichment of this effector subset in CVL samples.
When comparing expression of HIV co-receptors, binding proteins, and activation markers among CB1 samples with versus without visible blood, CD4 T cells derived from CB without visible blood had higher CCR5 and α 4 β 7 expression than those without visible blood.
Phenotype of HIV target CCR5 + CD4 T cells in genital mucosa
To better understand HIV susceptibility markers on CD4 T cells within the genital mucosa, we next performed in-depth phenotypic analyses of CCR5 - and CCR5 + memory CD4 T cells. For these analyses, we combined two sequentially collected CBs based on the preceding findings of higher cellular yield in CB samples. Because cell yields in the CCR5 - and CCR5 + subsets in the CVL and FS were insufficient for in-depth phenotypic comparisons, this analysis was performed in CB samples only. In addition, we contrasted CD4 T cell subsets in CB and blood to capture distinctive characteristics of HIV target CD4 T cells within the genital mucosa.
First, we evaluated whether CCR5 + CD4 T cells were enriched for additional markers of HIV susceptibility. Examination of CXCR4 expression revealed similar distribution profiles within CCR5 - and CCR5 + subsets in blood and FGT compartments ( S4 Fig ). On the other hand, CCR5 + CD4 T cells expressed higher frequencies of α 4 β 7 in both blood and genital compartments ( Fig 3A ), indicating enhanced capacity of these cells to both bind HIV and traffic to the gut associated lymphoid tissue where HIV infection may be established.
10.1371/journal.pone.0178193.g003
Fig 3
HIV target CCR5 + CD4 T cells in female genital mucosa display phenotypic attributes of activation and trafficking to rectal mucosa.
Histograms show comparison of expression of markers between CD45RO + CCR5 + and CCR5 - cells genital and whole blood CD4 T cells for (A) α 4 β 7 , (B) CD27, (C) CD38 and (D ) CD69 (***, p < 0.001, n = 20 participant visits for all markers except CD69 for which n = 12 visits)
Next, we asked whether CCR5 + cells demonstrated an effector phenotype and were enriched for markers of activation. CCR5 + CD4 T cells in the FGT, but not blood, displayed a CD27 lo effector phenotype [ 16 ] compared with CCR5 - cells, indicative of local immune activation ( Fig 3B ). While expression of CD38 was not higher in CCR5 + versus CCR5- CD4 T cells in the FGT ( Fig 3C ) , we found that expression of CD69, which is induced early after activation, was significantly higher in CCR5 + versus CCR5 - CD4 T cells in the FGT but had negligible expression in blood CD4 T cells regardless of CCR5 expression ( Fig 3D ). Together, these data show that CCR5 expression identifies activated effector cells within the FGT and that, among those studied, CD69 is strongly associated with CCR5 expression in the FGT.
CD4 T regulatory cells in the genital mucosa express high levels of CCR5
Based on our data showing heighted T cell activation in the FGT, we next sought to determine the frequency and phenotype of immune regulatory CD4 T cells in the FGT and blood. For this purpose, we examined CB samples from 10 women at a single visit either at the follicular (n = 5) or the luteal phase (n = 5) of the menstrual cycle. CD4 + T regulatory cells (T reg ) are a specialized CD4 T cell subset expressing the forkhead box P3 ( FOXP3 ) transcription factor and maintain immune homeostasis by suppressing proliferation of effector T cells [ 17 ]. Tregs constitute 5–10% of blood CD4 T cells in otherwise healthy individuals, but little is known about whether Tregs can be detected in the lower FGT in humans and how their frequencies compare to that in blood.
Of the 10 women sampled, we were able to detect FOXP3 expressing CD4 T cells in discernable numbers (i.e., greater than 50 events) in CB samples from 8 women. CVL and FS samples were not tested. Fig 4A shows a representative flow plot of FOXP3 expressing CD4 T cells in FGT and blood. We observed significantly higher frequency of Tregs (expressed as % of total CD45RO + CD4 T cells) in CB (median 10%, range 7 to 17%) relative to peripheral blood (median 8%, range 2.5 to 12%). Consistent with phenotype of Tregs, FOXP3 + CD4 T cells were more likely to be CD25 + , relative to non-Tregs ( FOXP3 - CD45RO + ) in both the genital mucosa and in the blood ( Fig 4B ). Because Treg numbers are associated with increased viral acquisition as observed in lymphoid tissue and mucosa in nonhuman primate models [ 18 , 19 ], we also determined if Tregs preferentially expressed CCR5 in the FGT and blood. The data showed that a higher proportion of Tregs relative to non-Tregs expressed CCR5 in both the blood and FGT ( Fig 4C ). Together, these data demonstrate the presence of Treg lineage cells expressing CCR5 in the FGT in discernable numbers in the majority of participants sampled.
10.1371/journal.pone.0178193.g004
Fig 4
Frequency and phenotype of CD4 T Regulatory cells in the genital mucosa.
(A) Representative flow plots showing FOXP3 + CD4 T cells in cytobrush and PBMCs (plots are gated on total CD4 + T cells). Scatter plot shows frequency of Tregs in 8 participants. ( B ) shows expression of CD25 and ( C ) shows CCR5 expression in FOXP3 + and FOXP3 - CD45RO + cells (**, p < 0.01; *, p < 0.05 using a two-tailed, paired non-parametric t test
Discussion
A clear understanding of immune factors that impact HIV acquisition within the FGT is important for systematic evaluation of HIV biomedical prevention strategies. This includes understanding the mucosal impact of candidate HIV PrEP drugs and microbicides, and delineating mucosal immune correlates of vaccine efficacy in HIV vaccine clinical trials for informing the rational design of HIV vaccines. Achieving this goal will require optimizing mucosal sampling strategies that can be used in longitudinal studies for accurate and effective immune profiling within the FGT. Approaches for assessment of humoral responses within the vaginal mucosa are generally well established with several studies demonstrating detection of total and HIV-specific IgG and IgA responses using vaginal swabs [ 20 , 21 ]. CVL sampling is widely used for measurement of mucosal cytokines, chemokines, and antimicrobial factors [ 22 – 24 ] and recently, measurement of cytokines and antibodies by self-sampling using menstrual cups has been demonstrated as another tool for assessment of soluble immune factors in the genital mucosa [ 25 ]. However, there is less clear guidance regarding optimal specimen types for accurate assessment of cellular immune profiles, specifically with regards to in-depth characterization of HIV susceptibility markers on CD4 T cells.
Previous studies have used minimally-invasive methods, such as CVL [ 26 ], endocervical flocked swabs [ 13 , 27 ], cytobrushes [ 7 ], and menstrual cup [ 28 ], for cellular characterization of the FGT, while others have used more invasive cervical or endometrial biopsies [ 7 , 29 ] in order to attain sufficient cellular yields, but few studies exist that compare multiple minimally-invasive methods. In this study, we sought to longitudinally characterize and compare T cell yields and phenotypes using 3 minimally-invasive FGT sampling methods. Our data confirm that CD45 + , CD4 + , and CD8 + T cell recovery was highest from CB relative to FS and CVL and was consistent with one previous multicenter study which demonstrated lower yields with CVL compared to CB or cervical biopsy [ 12 ]. Our study additionally showed that FS yields correlated with CB, and FS-derived cells have similar phenotypes to those from CB with lower probability of blood contamination and may therefore provide an alternative to CB in instances where CB collection is not feasible, blood contamination is of concern, or where cells are required for other assays. The frequency of expression of HIV binding proteins CCR5, CXCR4, and α 4 β 7 were similar among memory CD4 T cells derived from each sampling method and also correlated between FS and CB. Relative to CB, CVL-derived CD4 T cells expressed higher CD38, CD69, and lower CD27, indicative of higher activation, and potentially reflecting the different anatomic compartment (ectocervix and vagina) sampled using this method compared to FS or CB (endocervix). Our study provides the first comparison of these three minimally-invasive methods during longitudinal assessments in the same women, thereby accounting for intra-individual variability when comparing methods, and suggesting that sampling using one method did not preclude sufficient cell yields from the subsequent method.
By demonstrating the utility of minimally-invasive sampling methods for in-depth FGT investigations, our findings add to the previous literature characterizing FGT-derived CD4 T cell phenotypes and functionality in women at risk for HIV acquisition. Utilizing two consecutive CBs, we demonstrate that CCR5 + CD4 T cells are highly enriched within the FGT, CCR5 + CD4 T cells are highly activated relative to CCR5 - CD4 T cells in the genital mucosa. This phenotype resembles cells within the intestinal lamina propria and intraepithelial lymphocytes in humans [ 30 ]. Furthermore, expression of CD69 demonstrates recent activation and is highly consistent with the phenotype of tissue effector memory (TEM) cells, which are highly enriched in mucosa and non-lymphoid immune sites [ 31 ]. These cells rapidly produce effector cytokines upon antigen exposure [ 32 , 33 ] with mouse studies showing their indispensible role in protection against HSV2 and other mucosal infections [ 34 ].
However, due to their activated state and the expression of CCR5, vaginal TEMs represent highly vulnerable HIV target cells, and studies in macaques strongly implicate their role as founder cells contributing to local viral expansion and viral dissemination [ 9 , 35 , 36 ]. With respect to viral dissemination, the expression of α 4 β 7 on activated CCR5 cells in the FGT could also be highly significant. Apart from facilitating HIV binding and entry [ 37 , 38 ], α 4 β 7 directs migration of cells to gut associated lymphoid tissue [ 39 ] and could traffic HIV infected cells from the FGT to sites of active replication in the gut, thereby facilitating establishment of HIV infection [ 40 ]. Increase in frequency of α 4 β 7 + CD4 T cells in the FGT is shown to enhance infection in vaginal explants [ 11 ], and blocking α 4 β 7 protects macaques from vaginal SHIV acquisition [ 10 ]. Therefore, a better understanding of the phenotype and viral permissivity of CCR5, α 4 β 7 co-expression on target cells in the FGT should be evaluated in future human studies and may be done so using these sampling methods. In addition, based on the relative differences in HIV transmission risk across rectal, cervico-vaginal, and penile routes, these methods could also be used to compare CD4 T cell phenotype and relative activation status across these compartments [ 8 , 41 ] or understanding the relative role of macrophages versus T cells in HIV susceptibility in each of these distinct mucosal compartments [ 42 , 43 ].
We demonstrated the presence of Treg lineage cells for the first time in the FGT and in the majority of participants sampled. These Treg lineage cells were more likely to express CCR5, suggesting heightened HIV susceptibility. The frequency of CD4 T regs in blood has previously been shown to correlate with estradiol levels and peak during the late follicular phase of the menstrual cycle when estrogen levels are the highest [ 44 ], and estrogen induces FOXP3 expression in vitro [ 45 ], thereby supporting their potential contribution to hormonally-mediated changes in HIV susceptibility. Although the present study was not designed to capture dynamic changes in Treg frequencies within the FGT in the presence of varying endogenous or exogenous reproductive hormones, the ability to use longitudinal FGT sampling methods is critical to evaluating changes in these cells in the presence of reproductive hormones in future studies.
This study has some limitations. First, our study included African-American women who met certain HIV risk criteria and may not be generalizable to other populations. African-American women are disproportionately affected by HIV in the United States, and inclusion of African-American women in biological studies are critical to understanding HIV risk in this group. We did not, however, include women under 18 years old, and thus results cannot be extrapolated to adolescent women. Second, the small sample size was insufficient to examine the effects of multiple covariates that can affect T cell phenotype in the female genital tract or even cellular yield, including vaginal infections, the presence of semen, mucosal trauma, endogenous or exogenous hormones, and the vaginal microenvironment including the vaginal microbiome which can impact HIV risk, in part, by altering the phenotype of FGT CD4 T cells [ 46 ]. However, since all three sampling strategies were performed on each participant, the presence of these conditions would not have affected between-strategy comparisons. Third, the sequence of sampling was not randomly assigned but rather selected in the order of potential to induce trauma. While it is possible that the sequence affected cellular yields or phenotypes, correlations observed between yields and T cell markers between the second (FS) and third (CB) methods suggest that the order of sampling did not affect results. Additionally, we cannot exclude that the preceding method contributed to additional blood contamination in the subsequent method. However, correlations observed between FS (where less blood contamination was noted) and CB (where more blood contamination was noted) suggest that visible blood contamination does not likely represent peripheral blood, a finding noted in other studies[ 12 , 47 ]. Furthermore, for most T cell markers, differences were not observed between CB with versus without visible blood). Finally, while we cannot completely exclude the presence of peripheral blood as a source of genital tract T cells, studies have shown that microtrauma and bleeding occurs not uncommonly during sex, so these cells are likely to enter the FGT in the setting of potential sexual exposure to HIV and as such are likely still relevant for HIV acquisition.
In conclusion, we show that all three methods may be used for cellular immune profiling in the FGT, but FS and CB sampling methods have the highest cellular yields necessary for in-depth analyses. Notably, CD4 T cells within the FGT express CCR5 and α 4 β 7 and are highly activated, attributes which could act in concert to facilitate HIV acquisition and dissemination within the host [ 48 ]. Furthermore, we identified CD4 Treg lineage cells for the first time in the FGT and noted their high expression of CCR5, suggesting that their role in HIV acquisition should be explored in future studies. These sampling methods will allow for further investigation of strategies to reduce immune activation and/or HIV target cells at the genital mucosa as a means to enhance the efficacy of PrEP, microbicides, vaccines, and other biomedical prevention interventions, as well as to understand the potential effect of the mucosal immunity on these biomedical prevention interventions.
Supporting information
S1 Fig
Comparison of total cell yields from CB1 and CB2 either in setting of a prior flocked swab not performed (-, n = 12) versus performed (+, n = 12).
(TIFF)
S2 Fig
Correlation of HIV binding proteins (CCR5, CXCR4 and α4β7) expression on CB and CVL (ns, p ≥ 0.05 using correlation analysis from the first visit per participant).
(TIFF)
S3 Fig
Comparison of HIV binding proteins (CCR5, CXCR4, and α 4 β 7 ) and activation markers (CD38, CD27) on first cytobrush (CB1)-derived CD4 T cells from specimens without (qualitative blood score < 1) versus with visible blood (qualitative score ≥ 1).
CD69 analysis not performed due to small sample size (**, p < 0.01; *, p < 0.05).
(TIF)
S4 Fig
Comparison of expression of HIV co-receptor CXCR4 on CD45RO + CCR5 + and CCR5 - genital (CB) and whole blood CD4 T cells (ns, p ≥ 0.05 using linear mixed effects model with random effects for participant and visit).
(TIFF)
S1 Table
Comparison of sample parameters and cell yields between included (n = 25), excluded from all analyses (n = 5) due to high visual blood or low cellular yield, and excluded from Table 2 cell yield analyses due to inadequate staining (n = 8).
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Angiogenesis is the growth of new vessels from the existing vasculature in various organs. Angiogenic processes play a critical role in the physiologic conditions of cancer, heart diseases, atherosclerosis, and various eye diseases [1] . The formation of abnormal neovascularization and the growth and spread of vessels in the eye are the most common causes of blindness. Such alterations occur in age-related macular degeneration, central retinal vein occlusion, and diabetic retinopathy. Angiogenic factors such as vascular endothelial growth factor (VEGF), platelet-derived growth factor (PDGF), fibroblast growth factor (FGF), insulin-like growth factor (IGF), and angiopoietin have been implicated in dysregulation of growth, migration, adhesion, and differentiation of retinal vascular cells in retinal vascular diseases [2] – [4] . Nonetheless, the mechanism underlying pathophysiological neovascularization in the retina remains poorly understood.
Growth arrest-specific 6 (Gas6) is a vitamin K dependent protein discovered through the screening of genes upregulated in growth-arrested embryonic mouse fibroblasts [5] . Gas6 is a ligand for the tyrosine protein kinase receptors Axl, Mer, and Tyro3, which have been implicated in vascular homeostasis, cell growth, survival, and platelet thrombus formation [6] . Thus far, Axl and Gas6 signaling is implicated in cell proliferation, migration, and invasion during tumor angiogenesis, as well as in diabetic nephropathy [7] , [8] . Gas6 is a novel growth factor for kidney mesangial cells, and is posttranslationally activated by C-carboxylation in the presence of vitamin K. In diabetic nephropathy, streptozotocin-treated Gas6 knockout mice exhibit less pronounced glomerular hypertrophy and glycoxidized low-density lipoprotein increase in mouse mesangial cells [7] , [9] . However, nothing is known regarding the relationship between Gas6 and neovascularization in the retina and in developing zebrafish embryos.
In the present study, we investigated the effect of Gas6 on cell proliferation and migration under normal growth conditions of retinal microvascular endothelial cells (HRMECs) and the effect of Gas6 in developing zebrafish embryos. We provide evidence that Gas6 plays a crucial role in angiogenesis via regulating extracellular signal-regulated kinase (ERK1/2) phosphorylation in HRMECs and zebrafish. Our data reveal an unexpected function of Gas6 as an angiogenic factor that promotes cell survival and proliferation.
Materials and Methods
Cell culture
HRMECs (Cat. No. ACBRI 181) were purchased from Cell Systems (Kirkland, WA) and used at passages 3–7. Cells were grown in CSC complete medium (CS-4ZO-500; Cell Systems) containing Bac-Off® (antibiotic). Cultures were maintained at 37°C in a humidified 95% air/5% CO 2 atmosphere. Quiescence was induced by incubating the cells in CSC complete serum-free medium for 24 h. Cells were then used for the experiments, unless otherwise indicated.
Zebrafish and angiogenesis assays
Adult zebrafish were maintained under standard conditions at 28.5°C with a 14 h light/10 h dark cycle. Embryos were obtained from crosses between flk:GFP transgenic fish and raised in embryonic water. All experimental protocols for animal care and use were approved by the local ethical board (Korea Institute of Oriental Medicine Animal Care and Use Committee), and animal husbandry and procedures were performed according to institutional guidelines. For angiogenesis assays, dechorionated anesthetized zebrafish embryos in egg water, with Tricaine (0.016% MS222, Sigma-Aldrich, St. Louis, MO) added, were placed on agarose-modified dishes and microinjected (∼2 nl) with recombinant human Gas6 (300 ng/µl, rhGas6; Cat. No. 885-GS-050, R&D Systems, Minneapolis, MN) or recombinant human VEGF (5 ng/µl, rhVEGF, 293-VE-001MG/CF, R&D Systems) into the perivitelline space at 50 h post-fertilization (hpf) and maintained under standard conditions at 28.5°C, as previously described [10] , [11] . Fluorescence images were collected using a confocal microscope (Olympus, FV10i, Center Valley, PA) at 80 hpf. Embryonic stages were determined by hpf.
Proliferation assay
Cell proliferation was measured using the 3-(4,5-dimethyl thiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) assay kit (Cell Proliferation Kit I, Cat. No. 11 465 007 001, Roche, Nutley, NJ). Cells were plated (1×10 4 cells/well) in quadruplicate into 96-well plates with various doses of rhGas6 and rhVEGF. Cell viability was measured at 24 h and 48 h after incubation, and 10 µl MTT solution was added to the wells and incubated for 4 h. After incubation at 37°C, 100 µl solubilization solution was added into each well and incubated for 24 h. Absorbance was measured in a microplate autoreader (BIO-TEK, Synergy HT, Winooski, VT) at 550 nm. All experiments were repeated at least three times.
Wound-healing cell migration assay
HRMECs were plated at 1×10 5 cells/well on a 12-well plate in normal culture medium and allowed to reach 80–90% confluency. An injury line with a width of 0.6∼1 mm was made with a sterile pipette tip and cells were rinsed with phosphate-buffered saline (PBS). Fresh culture medium containing rhGas6 or rhVEGF was placed into the wells, and the cells were incubated for 6 h. To specifically determine the role of Gas6 in migration, cells were pretreated with depletion media containing 0.5% bovine serum albumin (BSA) with or without warfarin (1 µM, Sigma-Aldrich). Cell migration was monitored by visual examination using an inverted microscope (BX51 Olympus), as previously described [12] .
Rat aortic ring-sprouting assay
Aortas were harvested from 6-week-old Sprague-Dawley rats and immediately placed on ice in PBS supplemented with 1% fetal bovine serum. Plates (24-well) were coated with 200 µl Matrigel (Cultrex®, Trevigen Inc., Gaithersburg, MD). After polymerization, the rings were carefully placed in the wells with fine micro-dissecting forceps and sealed in place with an overlay of 100 µl Matrigel. rhGas6 (200 and 400 ng/ml) or rhVEGF (20 ng/ml) was added to the wells in a final volume of 1 ml serum-free medium and incubated at 37°C in 5% CO 2 . rhVEGF was used as a positive control. After 5 days, images of the aortic rings were obtained using an inverted microscope (BX51 Olympus). Sprouting length was quantified as mean maximal sprout length from the perimeter of the aortic ring to the most distal tip of the angiogenic sprout in four quadrants of each aortic ring. Experiments were performed in triplicate.
Tube formation assay
Briefly, 24-well tissue culture plates were coated with 200 µl basement membrane-like extract (BME/Matrigel) and incubated for 30 min at 37°C. HRMECs were seeded at 1×10 6 cells/ml and treated with serum-free medium containing the vehicle (PBS) or 400 ng/ml rhGas6 or 20 ng/ml rhVEGF. The cells were cultured for 9 h at 37°C in 5% CO 2 . The tube networks were stained with calcein AM (Invitrogen Life Sciences, Grand Island, NY) and observed with an inverted fluorescent microscope (BX51 Olympus). Images were captured using a digital camera (DP 70 Olympus). Four fields per well were captured for quantitative analysis. Sprout length was measured under an inverted microscope at 100× magnification and quantified. The digitized images were imported into Image J software. The experiments were independently repeated three times.
Western blot analysis
Aliquots of protein were treated with Laemmli sample buffer (Bio-Rad, Hercules, CA), heated to 100°C for 5 min, and electrophoresed with 20 µg protein/lane on a denaturing sodium dodecyl sulfate polyacrylamide gel. Proteins were then transferred to a nitrocellulose membrane (Whatman, GmbH, Hahnestr, Germany) using a tank blotting apparatus (Bio-Rad). Membranes were probed with 1∶1000 dilutions of polyclonal antibodies against phosphorylated ERK1/2, p38, and c-Jun N-terminal kinase (JNK) (Cell Signaling Technology, Danvers, MA). The membrane was washed and incubated with a horseradish peroxidase-linked goat anti-rabbit IgG (Santa Cruz Biotechnology, Santa Cruz, CA). After washing the membranes three times, the signals were detected with a WEST-one™ enhanced chemiluminescence solution (GenScript, Piscataway, NJ) using Fujifilm LAS-3000 (LAS-3000, Fuji Photo, Tokyo, Japan).
Immunofluorescence staining
To determine the expression levels of pERK1/2 in HRMECs, Gas6 treated HRMECs were fixed in 2% paraformaldehyde. Cells were permeabilized and blocked with 0.1% Triton X-100/PBS containing 2% BSA and 0.5% normal serum to reduce the nonspecific adherence of antibodies. Cells were incubated in primary anti-pERK1/2 at a dilution of 1∶1000 for 1 h at 37°C in a humidified chamber. After incubation with primary antibody, the cells were rinsed and incubated with anti-rabbit IgG -FITC (Santa Cruz Biotechnology). Cell nuclei were stained with Hoechst stain (blue). After washing with PBS and treating with mounting solution (Wako, Kumamoto, Japan), cells were observed under an Olympus IX-81 fluorescence microscope using a 20× objective.
Knockdown by gas6 morpholino oligonucleotide in zebrafish embryos
Transcript and genome sequence data of gas6 reported are available in the NCBI and ensemble database under accession numbers NM199978, ENSDARG00000007804 (zebrafish) and NM000820, ENSG00000183087 (human). Zebrafish show 66% homology with human gas6. To knockdown in zebrafish embryos, morpholino antisense oligonucleotides (MOs) for gas6 EX5-MO (exon 5 splice donor site; 5′ - GCT GAC GGT GTG TTT TTA CCG TTC T -3′), gas6 EX7-MO (exon7 splice donor site; 5′ - TGG TGT TGC TGA AGA CTG ACC TAC A -3′), and standard Ctrl-MO (control morpholino, 5′- CCT CTT ACC TCA GTT ACA ATT TAT A-3′) were designed and synthesized by Gene-Tools, LLC (Corvallis, OR). MOs were resuspended in 1× Danieau's buffer [58 mM NaCl, 0.7 mM KCl, 0.4 mM MgSO 4 , 0.6 mM Ca(NO 3 ) 2 , 5 mM HEPES, pH 7.6] with 0.1% phenol red, and microinjected into embryos at the 1–4 cell stage (1–5 ng/embryo). Injected embryos were incubated until the indicated stage and photographed using a fluorescent dissecting microscope (Olympus SZX16). Efficacy of morpholinos on the pre-mRNA splicing was evaluated using RT-PCR with forward primer designed to exon 1 (5′- GCC ATG AGG GAG CTG GTG TGG AGC-3′) in conjunction with reverse primer designed to exon 8 (5′- CAG ACG ACC GTC ACA GAA GCA GCG-3′). Total RNA from embryos was extracted using a standard protocol with TRIzol reagent (Invitrogen) and reverse transcribed into cDNA using M-MLV reverse transcriptase. The RT product was used as a template for PCR amplification of gas6 and β-actin using the following cycles with 5 pmol primers and Taq DNA polymerase (Elpis, Pusan, Korea): 95°C for 30 s, 56°C for 30 s, and 68°C for 50 s for 30 cycles.
Chemical treatment of zebrafish embryos
Transgenic (flk:GFP) zebrafish embryos were treated with SB203580 (p38 MAPK inhibitor, Calbiochem, Billerica, MA), PD98059 (p42/44 MAPK inhibitor, Calbiochem) and U0126 (MEK/ERK inhibitor, Cell Signaling Technology) that were diluted in dimethyl sulfoxide (DMSO). The embryos were incubated with 40 µM SB203580, 20 µM PD98059, and 40 µM U0126 at 12 hpf and observed using a confocal microscope (Olympus, FV10i) at 30 hpf.
Whole-mount immunostaining
Whole-mount immunostaining was carried out as previously described [13] . Briefly, transgenic (flk:GFP) zebrafish embryos were treated with U0126 (MEK/ERK inhibitor, Cell Signaling Technology). The embryos were incubated with 40 µM U0126 for 14 h, fixed at 26 hpf, and stained with anti-phospho-ERK1/2 antibody (1∶500, Cell Signaling Technology). For fluorescent detection of the antibody, Alexa Fluor 568 anti-rabbit conjugate was used (1∶500, Molecular Probes, Grand Island, NY). All stained embryos were mounted with glycerol and photographed on an Olympus FV10i confocal microscope and their expression was analyzed by Image J software.
shRNA knockdown and quantitative real-time PCR (QRT-PCR)
HRMECs were cultured to 80% confluence and transfected with Axl, Mer, and Tyro3 shRNA. Transfection of plasmids into HRMEC with the Neon Transfection System (Invitrogen) was performed according to the manufacturer's protocol with two pulses of 1400 V and 20 ms. After transfection for 24 h, the cells were collected for real-time PCR. Briefly, total RNA from cells was extracted using a standard protocol with TRIzol reagent (Invitrogen) and first-strand complementary DNA was synthesized in a 20 µl reaction volume using 0.5 µg of total RNA. Reverse transcript products were obtained using a TaKaRa PrimeScriptTM 1st Strand cDNA Synthesis kit (TaKaRa, Mountain View, CA) followed by QRT-PCR on a iQ5 Continuous Fluorescence Detector System (Bio-Rad). The PCR reactions contained 250 nM of primers, 1 µl cDNA (5 ng), and 10 µl 2X SYBR-green Realtime PCR Master Mix (SYBR Premix Ex Taq TM, TaKaRa) in a total volume of 20 µl with at least two independent biological sample repeats and four technical assay repeats. Error bars were generated from the standard deviation (calculated by the Bio-Rad iQ5 Continuous Fluorescence Detector System software). Reactions without template and/or enzyme were used as negative controls.
Statistical analysis
All experiments were repeated at least three times, and representative data are shown. Data are expressed as the mean ± standard error of the mean (SEM). Differences between groups were analyzed using Student's t -test and one-way analysis of variance followed by the Tukey multiple comparison test (PRISM5 software, Graph Pad, La Jolla, CA). Values of p <0.05 were considered statistically significant.
Results
Gas6 promotes proliferation, migration, and tube formation in HRMECs
To examine the effect of Gas6 on the proliferation of HRMECs, we performed MTT assays. As shown in Figure 1A , Gas6 plays a role in promoting HRMEC proliferation in a time- and dose-dependent manner. During the angiogenesis process, endothelial cells were stimulated to migrate and proliferate [14] . To evaluate the impact of Gas6 on HRMEC migration, scratch-wound healing assays were performed. For wound healing, HRMECs were serum starved for 24 h, and conditioned media were incubated with 200 and 400 ng/ml rhGas6 or 20 ng/ml rhVEGF as a positive control. Figure 1B shows the effect of Gas6 on the migration of HRMECs compared with a control group. To examine the effect of Gas6 on the migration, we pretreated with warfarin, which inhibited the activation by Gas6 [15] . As shown as Figure 1B , rhGas6 induced the migration in a dose-dependent manner and the induced migration was inhibited by pretreatment with 1 µM warfarin ( Figure 1C ).
10.1371/journal.pone.0083901.g001 Figure 1
rhGas6 induces proliferation and migration of HRMECs.
HRMECs were incubated in the presence of the indicated concentrations of rhGas6 for 24( A ). Proliferation was determined using the MTT assay, and absorption was analyzed at 550 nm using a microtiter plate reader. The results are presented as the mean ± SEM. (n = 4). *** p <0.001, ** p <0.01, and * p <0.05 vs. control, # p <0.01. ( B ) HRMEC responses to rhGas6 or rhVEGF were determined using a scratch-wound healing assay. Lines indicate the same width of the gap. Representative images are shown at 6 h after generating the scratch. ( C ) Warfarin (1 µM) was preincubated for 30 min prior to rhGas6 addition. Representative images are shown after generating the scratch.
Next, we examined whether Gas6 could induce tube formation, an endothelial function crucial to angiogenesis. The addition of rhGas6 to the incubation medium led to dose-dependent tube formation ( Figure 2 ). Cumulative sprouting length also increased in a dose-dependent manner in rhGas6 treated cells, with rhVEGF used as a positive control ( Figure 2C ).
10.1371/journal.pone.0083901.g002 Figure 2
rhGas6 induces tube formation in HRMECs.
( A ) HRMECs were incubated with rhGas6 or rhVEGF. Representative images at 9 h after treatment with rhGas6 or rhVEGF are shown. ( B ) Tube formation by HRMECs on Matrigels was observed by fluorescence microscopy. Relative density was measured using ImageJ software. Data are expressed as mean ± SEM. (n = 4). ** p <0.01 and * p <0.05 vs. control. ( C ) Cumulative sprout length was quantified as described in the Materials and Methods. Data are expressed as mean ± SEM. (n = 4). ** p <0.01 vs. control.
Gas6 induces vessel outgrowth in the aorta of rats and in zebrafish
To analyze the angiogenic function of Gas6 ex vivo and in vivo , we used rat aortas [16] and transgenic flk:GFP zebrafish embryos in order to provide angiogenesis model systems [17] . As shown in Figure 3A , rhGas6-treated rat aorta rings showed microvessel outgrowth of endothelial tubules and increased vessel sprouting length. The vessel sprouting length of the aorta was 33% (rhGas6, 200 ng/ml) to 63% (rhGas6, 400 ng/ml) longer than that of control sprouting vessels ( Figure 3B ).
10.1371/journal.pone.0083901.g003 Figure 3
Angiogenic responses are induced by rhGas6 in rat aortic rings and in zebrafish embryos.
( A ) Representative images after 5 day incubation with rhGas6 and rhVEGF. Sprouts from rat aortic rings are shown in the Matrigel control, rhGas6, and rhVEGF after treatment for 5 days, as described in the Materials and Methods. ( B ) Quantification of sprout length from aortic rings revealed increased sprout formation after a 5-day treatment with rhGas6 and rhVEGF in Matrigel. Data are expressed as mean ± SEM. (n = 4). *** p <0.001 vs. control. ( C ) Transgenic embryos (flk:GFP) at 50 hpf were injected into the perivitelline space with Texas Red dye (Ctrl), Texas Red dye and rhGas6 (300 ng/µl), or Texas Red dye and rhVEGF (5 ng/µl). After 30 h, embryos were photographed under a confocal microscope. In the controls, the formation of ectopic sprouts (arrow) is never observed, and the injected embryos of the rhGas6- and rhVEGF-treated embryos show ectopic sprouting of subintestinal vessels (SIVs). The experiment was repeated three times (Control, n = 28; rhGas6, n = 24; rhVEGF, n = 24).
In zebrafish embryos, rhGas6 or rhVEGF was injected with Texas Red dye as a tracer into the zebrafish embryo yolks at 50 hpf. After 30 h, the extent of subintestinal vessel (SIV) formation was analyzed. Embryos injected with control medium and Texas Red dye served as controls and showed normal SIV development ( Figure 3C, left panel ). Compared to the control embryos, 75% of rhGas6-injected embryos had significantly increased vessel outgrowth in the SIV, resulting in abnormal SIV patterning ( Figure 3C, middle panel, white arrow; right panel ). Fifty-nine percent of rhVEGF-injected embryos, used as a positive control, were sufficient to cause the outgrowth of additional vessels, and similar results were observed with overexpression of rhGas6 ( Figure 3C ). These results suggest that rhGas6 can promote angiogenesis both ex vivo and in vivo .
Knockdown of gas6 inhibits angiogenesis in zebrafish
To investigate the functional role of gas6 in zebrafish vessel development, we performed gene knockdown studies using antisense (EX5-MO and EX7-MO). MOs were targeted against the exon 5 and exon 7 splice donor site to prevent mRNA splicing, resulting in intron insertion and thus premature translation termination ( Figure 4A ). We examined MOs efficacy by RT-PCR and reduced a product of gas6 transcripts in EX5-MO injected embryos, which was not observed in control MO injected embryos ( Figure 4B ). We also inserted an intron 7 region by blocking of gas6 mRNA splicing in EX7-MO injected embryos ( Figure 4C ). We identified the existence of intron 7 in the exon 7 morphants by sequencing of control- and gas6 -morphant mRNA. Translation of the mRNA sequences in morphants generated a premature stop codon for the exon 5-MO and the exon 7-MO, resulting in a truncated form of the Gas6 protein. After microinjection of gas6 MOs into zebrafish flk:GFP transgenic embryos, blood vessel formation was examined. The Ctrl-MO injected embryos exhibited normal development in the axial vasculature, dorsal aorta, posterior cardinal vein, and the dorsal longitudinal anastomotic vessel ( Figure 4D, left panel ). However, there were defects in intersegmental vessels and dorsal longitudinal anastomotic vessel formation in morphants at 30 hpf ( Figure 4D , EX5-MO and EX7-MO). Next, we examined whether rhGas6 could restore the effect of gas6 morphants. Injection of rhGas6 rescued the phenotypic change using microinjection of gas6 MOs into zebrafish flk:GFP transgenic embryos (Figure S1 in File S1 ). Thus, these data suggest that inhibition of Gas6 signaling resulted in defects in intersegmental vessel formation in zebrafish angiogenesis.
10.1371/journal.pone.0083901.g004 Figure 4
Knockdown of gas6 induces the inhibition of angiogenesis in intersegmental vessels.
( A ) The red arrows (Ex5:In5 and Ex7:In7) indicate morpholino target sites for splicing blocks. Primers (blue arrows, exon 1, forward primer; exon 8, reverse primer) were designed for the testing of morpholino efficacy, as described in the Materials and Methods. ( B , C ) Testing and quantification of morpholino nucleotide efficacy by RT-PCR in standard control MO (Ctrl-MO) and gas6 MO (EX5-MO and EX7-MO) treated embryos at 26 hpf. In control morphants, gas6 mRNA (768 bp, black arrow) is detectable by RT-PCR ( B , first line). In gas6 EX5 morphants, the wild-type gas6 mRNA is undetectable by RT-PCR at doses of 2 and 5 ng/embryo. The morphant mRNA encodes a truncated form of the Gas6 protein. ( C ) In gas6 EX7 morphants, RT-PCR products reveal Ctrl-MO embryos expressing wild-type gas6 transcript, while EX7-MO embryos express two transcript variants at doses of 1 or 2 ng/embryo. The black arrow shows reduced expression of wild-type gas6 mRNA. The red arrow indicates results from aberrant splicing, resulting in a gain of ∼250 base pairs of intron 7, which encodes a premature stop codon that occurs in the Gas6 . ( D ) Angiogenesis defects in gas6 morphants. The flk:GFP transgenic zebrafish embryos were microinjected with Ctrl-MO (n = 30, 4 ng/embryo) and gas6 MO (EX5-MO, n = 35, 4 ng/embryo; EX7-MO, n = 34, 2 ng/embryo), and their blood vessel formation was examined at a cellular level in living embryos at 30 hpf. Normal formation of intersegmental vessels, as shown by GFP-positive endothelial cells, is observed in Ctrl-MO embryos, but severe vascular defects is observed in gas6 MO-injected embryos. The experiment was repeated two times.
ERK signaling pathway in rhGas6-induced HRMECs and in gas6 -morphants
To investigate the mechanism of rhGas6-induced migration in HRMECs, we examined the signal transduction pathway in rhGas6-induced HRMECs. To determine whether mitogen-activated protein kinases (MAPKs) such as ERK1/2, p38, and JNK are involved in rhGas6-induced migration of HRMECs, the phosphorylation levels of MAPKs were measured. As shown in Figure 5A , rhGas6 (400 ng/ml) treatment resulted in a transient phosphorylation of all MAPKs. The levels of phosphorylated ERK1/2 and JNK maximally increased at 15 min after Gas6 treatment, and the levels of phosphorylated p38 increased at 5 min. Gas6 levels were also increased in addition to phosphorylated ERK1/2 levels, as determined by immunofluorescence staining ( Figure 5B ).
10.1371/journal.pone.0083901.g005 Figure 5
Activation of ERK is regulated by Gas6.
( A ) rhGas6 induced signaling pathway in HRMECs. Cells were incubated with rhGas6 (400 ng/ml) at various times, and cell lysates were subjected to western blotting with specific antibodies, as described in Materials and Methods. ( B ) Immunofluorescence staining for pERK1/2 and Hoechst staining was performed, as described in Materials and Methods. Representative images of p-ERK1/2 staining in rhGas6-treated HRMECs. ( C, D ) PD98059, SB203580, or U0126 was preincubated for 30 min, and HRMEC responses to rhGas6 or rhVEGF were determined by western blotting using specific antibodies. The experiment was repeated three times. ( E ) Phospho-ERK1/2 staining in the trunk of 26 hpf embryos [Ctrl-MO embryo (n = 7); gas6 -MO injected embryo (n = 8); U0126-treated embryo (n = 8)] and quantitative analysis of zebrafish p-ERK1/2 expression by whole-mount immunostaining. The gas6 -morphant and U0126 treated embryos showed reduction of p-ERK1/2 protein expression compared to controls in the zebrafish trunk. The experiment was repeated two times. ** p <0.01 vs. Ctrl-MO.
Among the MAPKs, phosphorylated ERK1/2 is mainly observed during the stimulation of receptor tyrosine kinase activity by a growth factor, and is a major participant in the regulation of cell growth and differentiation. This appears to be a common and central component within various signal transduction pathways [18] . In order to determine the key signal transduction pathways involved with Gas6 induced angiogenesis in HRMEC, we evaluated the effects of MAPK inhibitors on MAPK, Akt, and eNOS signaling. Gas6 stimulated the phosphorylation of ERK1/2, p38, pAkt, and eNOS. Inhibitors SB203580 (p38), PD98059 (ERK), or U0126 (MEK/ERK) inhibited the Gas6-induced signaling ( Figure 5C ). The phosphorylation of ERK was associated with Gas6 induced signaling. Furthermore, to investigate the signal transduction pathways of Gas6 in vivo , we examined the effects of Gas6 on phosphorylation of ERK1/2 in developing zebrafish embryos. After microinjection of gas6 MO into zebrafish embryos, we analyzed the phosphorylation of ERK1/2 as a downstream effector of Gas6 signaling. The results of whole-mount immunostaining of embryos with anti-phospho-ERK1/2 antibody revealed that repression of Gas6 reduced phosphorylation of ERK1/2 ( Figure 5E , middle panel, 25% inhibition). Treatment with U0126 also reduced phosphorylation of ERK1/2 ( Figure 5E, right panel , 24% inhibition).
Migration and phenotypic changes mediated by the ERK inhibitor in Gas6-induced angiogenic processes
To investigate the effect of the ERK1/2 inhibitor on migration of Gas6 induced HRMECs, the cells were pretreated with a MEK/ERK kinase inhibitor (U0126) and Gas6 for 6 h. For quantitating relative migration, the number of cells migrating in a field was counted. U0126 treatment significantly inhibited rhGas6 or rhVEGF induced migration ( Figure 6A ). Next, to test the effect of receptor tyrosine kinase inhibitors on migration of Gas6 or VEGF induced HRMECs, cells were pretreated with foretinib (0.1 µM) or SU11248 (0.1 µM) and rhGas6 or rhVEGF for 6 h. rhGas6 or rhVEGF induced the migration and the induced migration was inhibited by pretreatment with foretinib or SU11248 (Figure S2 in File S1 ).
10.1371/journal.pone.0083901.g006 Figure 6
Disruption of Gas6 signaling via the ERK pathway resulted in defective angiogenesis in HRMECs and in zebrafish.
( A ) U0126 was preincubated for 30 min, and HRMEC responses to rhGas6 or rhVEGF were determined using a scratch-wound healing assay. Lines indicate the same width of the gap, and migrating cells are marked with a red asterisk. Representative images at 6 h after generating the scratch are shown. The experiment was repeated three times. *** p <0.001, ** p <0.01 vs. control, ### p <0.001, # p <.05 vs. U0126-treated cells. ( B ) At 30 h, fluorescent images show gross morphology of Ctrl-MO injected control (n = 34), gas6 -morpholino [EX5-MO (n = 35), EX7-MO (n = 30)] injected, and SB203580 (n = 12), or PD98059 (n = 12), or U0126 (n = 12) treated embryos. The experiment was repeated two times. Fluorescent micrographs of live flk:GFP zebrafish Ctrl-MO embryos at 30 hpf. Note the proper formation of the major axial vasculature, dorsal aorta, and posterior cardinal vein, as well as the intersegmental vessel. Representative embryos treated with gas6 -MO oligonucleotide, SB203580, PD98059, or U0126. Note the abnormal and stunted formation of the intersegmental vessel (longitudinal white bar in B ). Representative mild defect of intersegmental vessels in the embryos treated with SB203580. *** p <0.001, * p <0.05 vs. Ctrl-MO.
To further investigate the effect of signaling molecule expression on phenotypic changes in angiogenic processes, we treated with SB203580 (p38 MAPK inhibitor), PD98059 (ERK inhibitor), or U0126 (MEK/ERK inhibitor), and injected EX5-MO and EX7-MO into flk:GFP transgenic zebrafish embryos using previously described methods. In this transgenic system, endothelial cells can be directly observed under a fluorescence stereomicroscope. The specificity of the MEK/ERK inhibitor has already been reported in vivo [19] , including analysis in zebrafish [8] , [9] . As shown in Figure 6B , treatment with inhibitors (SB203580, PD98059, or U0126) at 12 hpf inhibited intersegmental vessel formation compared to control levels (treated with 0.4% DMSO alone) at 30 hpf. Similar results were observed with knockdown of gas6 (EX5-MO and EX7-MO) in embryos ( Figure 6B ). These results suggest that Gas6 can regulate angiogenesis by activation of the MEK/ERK kinase pathway.
Discussion
Gas6 has been shown to play a pivotal role in pathophysiological processes such as atherosclerosis, cancer, and thrombosis through activation of cells ranging from platelets to endothelial cells, as well as vascular smooth muscle cells [6] . Gas6 is secreted or expressed by various cancer cells, smooth muscle cells, retinal pigment epithelial cells, mesangial cells, and endothelial cells [7] , [9] , [20] . Expression of Gas6 and Axl is increased in various types of cancers, and Axl has also been suggested as a rational target for cancer therapy [21] , [22] . In the retina, Gas6 is expressed endogenously by human retinal pigment epithelial cells, and deficiency or inhibition of Gas6 induces platelet dysfunction and protects from thrombosis [6] , [23] . However, there has been no study on the effect of Gas6 on the functions and mechanisms mediated by HRMECs. The results from this study provide evidence that Gas6 stimulates angiogenesis involving proliferation, migration, and sprouting of endothelial cells and of zebrafish embryos via ERK1/2 signaling.
Angiogenesis, a process of new blood vessel growth, is induced by various growth factors such as VEGF, FGF, PDGF, and transforming growth factor-beta (TGF-β), and it is a target for combating diseases characterized by either poor vascularization or abnormal vasculature. Pathological angiogenesis in the retina is a major feature of diseases that lead to blindness, particularly diabetic retinopathy [24] . Inhibiting angiogenesis by targeting specific pro-angiogenic factors such as VEGF has become a major focus of drug development for diabetic retinopathy, and anti-VEGF drugs are used to treat diabetic retinopathy [25] , [26] . The mechanism of action by which Gas6 induces neovascularization in retinal microendothelial cells is still subject to speculation. In this study, Gas6 promoted the in vitro proliferation of HRMECs in a time- and dose-dependent manner ( Figure 1 ). Gas6 also induced migration and tube formation in HRMECs. Furthermore, vessel sprouting on the rat aorta was stimulated by rhGas6, and the outgrowth of endothelial tubules and vessel sprouting length was increased. Knockdown of gas6 using antisense MOs inhibited angiogenesis in zebrafish development ( Figure 4 ). To confirm that the phenotypes observed in gas6 MO-injected embryos were caused by loss of function in gas6, we injected MO and rhGas6 protein. rhGas6 could partially rescue of gas6 MO-induced angiogenesis defect at 30 hpf, suggesting that the defect caused by injecting the gas6 MO was induced caused by a knockdown of gas6 activity (Figure S1). Gas6 protein has been shown to interact with receptor tyrosine kinases Axl, Mer, and Tyro3 [27] . To effectively link Axl, Mer, and Tyro3 expression levels to specific cellular behaviors in HRMECs, we used a collection of Axl, Mer, and Tyro3 targeting shRNAs that reduced Axl, Mer, and Tyro3 expressions in a graded manner. Axl, Mer, or Tyro3 shRNA transfected cells showed reduction of Axl, Mer, or Tyro3 mRNA levels, and Gas6 induced Axl, Mer, or Tyro3 mRNAs levels (Figure S3 in File S1 ). Moreover, inhibitors of receptor tyrosine kinases inhibited rhGas6 induced migration (Figure S2 in File S1 ). Gas6 induced signaling pathway effects on migration may therefore mediate the receptor tyrosine kinase family members, Axl, Mer, and Tyro3, in HRMECs.
Prevention of neovessel growth is a promising strategy of intervention to improve long-term prognosis of visual outcome and quality of life for many patients. Cell migration is essential in pathophysiological processes, such as wound healing and metastasis. VEGF induces neovessel growth and migration in the development of both proliferative diabetic retinopathy and diabetic macular edema, and anti-VEGF agents have emerged as new approaches in the treatment of these devastating diabetic complications [28] . The Ras-dependent ERK1/2 MAP kinase pathway plays a central role in cell proliferation control [29] . Recently, it was shown that inhibition of ERK activation, which occurs immediately after wounding, significantly inhibited the directional migration of fibroblasts [30] . ERK1/2 inhibitors and gas6 MOs reduced phosphorylation of ERK1/2 in zebrafish, and phenotypic changes were also significantly inhibited by gas6 MOs microinjection into transgenic (flk:GFP) zebrafish embryos. This suggests that Gas6 signaling regulates angiogenesis through ERK1/2 kinase.
To our knowledge, this study provides the first evidence that Gas6 can induce proliferation, migration, and angiogenesis in HRMECs and in zebrafish during vessel formation. Moreover, these processes may occur via phosphorylation of ERK1/2.
Supporting Information
File S1
Contains the files: Figure S1. Recombinant human Gas6 rescues the phenotypic changes in gas6 morphant-injected embryos. ( A ) Embryos injected with gas6 -exon5 morpholino. ( B ) Embryos injected with gas6 -exon7 morpholino. Shown above are 30 hpf zebrafish Ctrl-MO-injected embryos (Ctrl-MO), embryos injected with a gas6 morpholino (EX-5 MO, EX-7 MO), injected with a gas6 morpholino and rhGas6 protein [EX-5 MO + rhGas6 (130 ng/µl), EX-7 MO + rhGas6 (130 ng/µl)]. Injection of rhGas6 resulted in a rescue of ISVs defects. ISVs formation is indicated by the white dotted line. Figure S2. Effect of receptor tyrosine kinase inhibitors on rhGas6-induced migration in HRMECs. Wound-healing cell migration assay was performed as described in Material and Methods. Foretinib and SU11248 were preincubated for 30 min, and Gas6 was then incubated for 6 h in the HRMECs. Foretinib and SU11248 inhibit Gas6 induced migration in HRMECs. Figure S3. rhGas6 stimulated the expression of Axl, Mer, and Tyro3 in HRMECs. The HRMECs were transfected with Axl, Mer, or Tyro3 shRNA and treated with rhGas6 (400 ng/ml). shRNA knockdown and quantitative real-time PCR were performed as described in Material and Methods. Gas6 induced Axl, Mer, and Tyro3 mRNA expression. Each bar represents the mean ± SD from three independent experiments (*** p <0.001 vs. control shRNA; ### p <0.001 vs. rhGas6-untreated and control shRNA-treated cells).
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The adult vertebrate skeleton appears to be a rather uniform bony tissue. This apparent uniformity belies its embryonic origins, as a mosaic structure originating from diverse progenitors that arise in different germ layers, and its development via two distinct ontological processes: intramembranous and endochondral ossification [1] , [2] , [3] . In contrast to the direct differentiation of osteoblasts from mesenchymal progenitors that produces the intramembranous bones of the skull and clavicle, the caudal bones of the head, the vertebral column, ribs, and appendicular skeleton develop first as cartilaginous anlagen [4] . Chondroblasts in these skeletal precursors have one of two fates: to undergo a histologically well-defined program of hypertrophy and terminal differentiation in the growing bone or to persist as specialized chondrocytes at the articular surface [5] , [6] . The majority of chondroblasts in prospective skeletal elements proliferate in regular columns as radially flattened immature chondrocytes before maturing through three distinct and spatially organized phases: prehypertrophic, hypertrophic, and mineralizing chondrocytes. The hypertrophic chondrocyte differentiation pathway is tightly regulated, with numerous signalling pathways providing both positive and negative signals at each step in the differentiation process. Some of these signals, like Ihh and Delta, are made by subpopulations of differentiating chondrocytes [7] , [8] , [9] . Others, like members of the TGF-β/BMP and Wnt families, are also made by cells in the perichondrium [10] , [11] , [12] , a tissue that surrounds the growing cartilaginous core and gives rise to the bone collar. Those competing extracellular signals induce or repress transcription factors to regulate chondrocyte proliferation and differentiation; the result is coordinated longitudinal bone growth and joint articulation [6] , [13] , [14] , [15] , [16] , [17] , [18] , [19] , [20] , [21] .
Dlx homeobox genes encode nuclear transcription factors [22] , [23] , [24] . In particular, Dlx5 and Dlx6 are expressed in all anlagen of the endochondral skeleton. Their expression has been noted in precartilaginous limb bud mesenchyme [25] , [26] where each overlaps with expression of Sox9 [27] , [28] . At later stages of skeletogenesis, when proliferating and differentiating chondrocytes are found in spatially distinct regions of the growth plate, Dlx5 and Dlx6 are expressed in the post-mitotic prehypertrophic and hypertrophic zones but not in immature chondroblasts in the resting or proliferating zones [27] , [28] , [29] . Dlx5 and Dlx6 are also expressed in the perichondrium/periosteum in the long bones as well as ribs and vertebrae [27] , [30] , [31] , [32] . Dlx5 −/− and doubly deficient Dlx5/6 −/− mice have revealed requirements for Dlx5 and Dlx6 during chondrogenesis [28] and chondrocyte hypertrophy [27] , [33] . Reciprocally, forced expression of either Dlx5 or Dlx6 alone in chicken limb bud micromass cultures stimulated chondrogenesis [28] and misexpression of Dlx5 in vivo reduced proliferation of epiphysial chondrocytes and resulted in precocious chondrocyte hypertrophy [27] , [29] , [34] . Together, these experiments implicate Dlx5 and Dlx6 as partially redundant positive regulators of chondrocyte differentiation. However, to date, perturbations of Dlx expression have either not been cell type specific or have been done in the context of endogenous Dlx5 expression. This general concern is particularly germane when seeking to elucidate a cell-autonomous function for Dlx5 in chondrocyte hypertrophy given endogenous Dlx5 expression in both the differentiating chondrocytes and in the perichondrium, the site of synthesis of secreted factors that regulate this process in the chondrogenic core. Here, we first describe a transgenic line of mice in which exogenous Dlx5 expression is targeted to immature chondrocytes using regulatory elements from the Col2a1 gene. Tissue-specific misexpression of Dlx5 accelerated chondrocyte hypertrophy and promoted precocious ossification in the endochondral skeleton of these transgenic mice. Visualization of transgene expression in the absence of endogenous Dlx5 expression (achieved by crossing the allele onto a Dlx5/6 null background) indicated that transgene expression was limited to the chondrogenic core and not the perichondrium. The subsequent rescue of endochondral ossification defects in Dlx5/6 −/− ; Col2a1-Dlx5 mice therefore establishes a cell-autonomous function for Dlx5 during chondrocyte hypertrophy and, furthermore, demonstrates functional equivalence of Dlx5 and Dlx6 in the endochondral skeleton.
Results
Generation of Transgenic Mice and Characterization of Transgene Expression
Dlx5 and its cis -linked paralogue Dlx6 function as positive regulators of both chondrogenesis and chondrocyte hypertrophy in the endochondral skeleton [27] , [28] , [29] , [33] . To further investigate the function of Dlx5 during chondrocyte hypertrophy in vivo , we generated transgenic mice in which a Dlx5 cDNA was expressed under control of the promoter and intron 1 enhancer of the Col2a1 gene ( Fig. 1A ) so as to target Dlx5 expression to immature chondrocytes following their differentiation from pluripotent mesenchymal precursors. Following pro-nuclear injection, we obtained four Col2a1-Dlx5 transgenic founders; these four founders (or their hemizygous offspring) had a variable number of copies of the transgene, as measured by semi-quantitative PCR from genomic DNA ( Fig. 1B ). The three founders with the highest number of copies (more than ten) were recovered dead as neonates and their transgenic alleles are referred to as Col2a1-Dlx5 t1 , Col2a1-Dlx5 t2 , and Col2a1-Dlx5 t3 . The viable founder with allele Col2a1-Dlx5 t19 had fewer than four copies as a hemizygote and was used to establish a stable transgenic line expressing epitope-tagged Dlx5 ( Fig. 1C ).
10.1371/journal.pone.0008097.g001 Figure 1
Generation of Col2a1-Dlx5 transgenic mice.
(A) Transgene design. Numbering of promoter and intron elements is with respect to the Col2a1 transcription start site. The start codon of Col2a1 in exon 1 has been mutated to prevent translation in this exon [56] . The asterisk indicates the Flag sequence 5′ to the murine Dlx5 open reading frame, bpA, poly-adenylation sequence from the bovine Growth Hormone gene; SA, splice acceptor. Half arrows indicate the approximate location of primers for genotyping: Col2a1 forward (Cf) plus Dlx5 reverse (Dr), and for RT-PCR: Flag forward (Ff) plus Dr. (B) Transgene copy number. Semi-quantitative PCR was used to compare the approximate transgene copy number in transgenic founder mice: t2/+ and t3/+ represent dead hemizygous founders and t19/+ is a hemizygote neonate from the stable Col2a1-Dlx5 t19/+ line. Control lanes from the left are: wild type DNA (0), wild type DNA mixed with p3000i3020Col2a1-Dlx5 at 1 copy per genome equivalent (1) or 10 copies per genome equivalent (10). Amplification of a genomic fragment of the single copy gene Ihh was used to judge relative amplification of the transgene. (C) Specific amplification of expressed Flag-Dlx5 sequence from transgenic embryos in a reverse transcriptase-dependent manner. RT-PCR analysis of five F4 generation embryonic day (E) 17.5 embryos from a wt x Col2a1-Dlx5 t19/+ mating demonstrates stable heritable expression of the Col2a1-Dlx5 transcription unit. The first lane in the bottom panel (+) shows a positive PCR control for the minus RT experiment. Non-adjacent lanes from the same gel have been spliced together to generate the figure. Lane numbers refer to individual embryos.
To examine the tissue distribution of transgene expression, we examined embryos following whole mount in situ hybridization with a Dlx5 riboprobe. Dlx5 expression was visualized in a number of locations that normally express the definitive chondroblast marker Col2a1 ( Fig. 2 ). In particular, and in contrast to non-transgenic littermates, Dlx5 expression was apparent in somites along the entire rostro-caudal axis and in rib cartilage and throughout the skeletal anlagen of the limb, where it closely matched expression of the endogenous Col2a1 gene in the stylopod and zeugopod ( Fig. 2A–E ). In contrast, at E12.5 ectopic expression of Dlx5 was most obvious in the phalanges, where it coincided with endogenous Col2a1 expression. Endogenous expression of Dlx5 in the otic vesicle, mandibular arch, branchial arches 2 and 3, or in a proximal anterior mesodermal domain in the limb was not altered in transgenic embryos. To confirm that Dlx5 was being expressed ectopically in immature chondroblasts, we examined tissue sections following in situ hybridization. Indeed, Dlx5 was expressed throughout the Col2a1 -expressing zones of the long bones and vertebrae ( Fig. 2F–I and data not shown), including the resting and proliferating zones of the long bone epiphyses, where its expression is not usually detectable (compare Fig 2H,I ). Moreover, in the hypertrophic zone, where Col2a1 transcription is normally down regulated, we saw a parallel decrease in Flag-Dlx5 transcript abundance (compare Fig. 2G,I ). To view transcription of the transgene in the absence of endogenous Dlx5 expression, we introduced the Col2a1-Dlx5 t19 allele into a Dlx5/6 −/− background. Section in situ hybridization to the long bones of the limbs of Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ embryos revealed that transgene expression was restricted to the cartilaginous core of the skeletal anlagen and was not expressed in the surrounding perichondrium ( Fig. 2J,K ). In summary, expression of the Col2a1-Dlx5 t19 allele faithfully replicated endogenous Col2a1 gene expression in chondrocytes and resulted in ectopic Dlx5 expression in immature and proliferating chondroblasts.
10.1371/journal.pone.0008097.g002 Figure 2
Transgene expression in the endochondral skeleton.
(A–E) Whole mount in situ hybridization of Dlx5 (A–C) or Col2a1 (D,E) probes to wild type ( wt ) and Col2a1-Dlx5 t19/+ transgenic ( t/+ ) littermates at E11.5 and E12.5. Arrows point to somites in panel A. Arrowhead points to digit 2 in panel C. (F–I) Section in situ hybridization of Col2a1 (F,H) or Dlx5 (G,I) riboprobes to adjacent sections of the femur of wild type (F,G) or Col2a1-Dlx5 t19/+ transgenic embryos (H,I) at E14.5. Arrowheads indicate the Col2a1 -positive proliferating zones, arrows point to the resting zones, and the bracket demarcates the hypertrophic zone in panels G and I. Proximal is to the right. (J,K) Section in situ hybridization of Col2a1 (J) or Dlx5 (K) riboprobes to adjacent sections of the tibia of Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ transgenic embryos at E14.5. Proximal is to the right. md, mandibular arch; ot, otic vesicle; pam, proximal anterior mesoderm. Scale bar = 1 mm in A, 0.5 mm in B–E,J,K, 0.2 mm in F–I.
Acceleration of Chondrocyte Hypertrophy Following Forced Expression of Dlx5 in Chondroblasts
To examine the functional consequences of misexpressing Dlx5 in chondrocytes, we first examined skeletal preparations of neonates at postnatal day zero (P0) or P1 after staining cartilage and mineralizing bone with alcian blue and alizarin red respectively. All three dead founders, or viable neonates bearing the Col2a1-Dlx5 t19 allele, examined this way had a common phenotype of hypermineralization of the endochondral skeleton ( Fig. 3 ). In its most severe form, seen in three non-viable founders, hypermineralization resulted in shorter stature (compare Fig. 3A, B ). The extent of hypermineralization, as estimated from the extent of alizarin red staining and generalized skeletal dysmorphology, followed the allelic series Col2a1-Dlx5 t19/+ < Col2a1-Dlx5 t19/t19 < Col2a1-Dlx5 t2/+ , Col2a1-Dlx5 t3/+ < Col2a1-Dlx5 t1/+ , and correlated with our estimates of transgene copy number ( Fig. 1B ). We analyzed each component of the skeleton in turn and uncovered a differential sensitivity to transgene expression in the axial and appendicular skeleton.
10.1371/journal.pone.0008097.g003 Figure 3
Dose-dependent hypermineralization in the endochondral skeleton of Col2a1-Dlx5 transgenic neonates.
(A,B) Lateral views of a wild type CD-1 P0 pup (A) and two hemizygous Col2a1-Dlx5 transgenic pups (B), both found dead shortly after birth. (C,D) Dorsal view of the caudal skull of wild type (C) and t19/+ (D) neonates at postnatal day zero (P0) following staining with alcian blue and alizarin red. The arrow points to an area of fusion between the basioccipital and exoccipital bones in panel D. (E,F) Lateral views of a wild type (E) and t1/+ transgenic founder (F) at P0 after staining with alcian blue and alizarin red. (G) Bubble graph of the ossification index of multiple litters of t19/+ hemizygotes and their wild type littermates. Following staining with alcian blue and alizarin red, a score was assigned to each embryo or neonate that reflected the extent of occipital ossification relative to a typical wild type at that stage: 1 = no brain case (exencephalic), 2 = brain case intact but no ossification of the supraoccipital (SO) apparent, 3 = smaller SO than is seen in a majority of wild type specimens (which were assigned a score of 4); 5 = obviously smaller distance between the SO and exoccipital (EO), or between the EO and the basioccipital (BO), compared to a majority of wild type; 6 = unilateral touching of SO and EO or of EO and BO; 7 = bilateral contact between SO and EO, or between the EO and the BO, or fusion of any of these bones. Bubble size is proportional to the number of neonates with a given score; the smallest circles represent a single individual, the largest circle represents 17 individuals. (H–J) Ventral views of wild type (H), hemizygous (I) and homozygous (J) embryos with the Col2a1-Dlx5 t19 allele at P1 following alizarin red staining. Arrowheads point to precociously mineralized vertebrae in panel I. (K,L) Ventral views of wild type (K), and Col2a1-Dlx5 t1/+ transgenic founder (L) at P0 following alcian blue and alizarin red staining. The most caudal thoracic (T13) and sacral (S4) vertebrae are marked with white asterisks. (M–P) Limb skeleton preparations from neonates with the genotypes as shown following alcian blue and alizarin red staining. at, atlas; at/ax*, fused atlas and axis; ax, axis; BO, basioccipital; D, dentary, EO, exoccipital; fe, femur; fi, fibula; h, humerus; IP, interparietal; IP*, interparietal bone with expanded mineralization; r, radius; sc, scapula; SO, supraoccipital; S/EO*, fused supraoccipital and exoccipital bones; t, tibia; u, ulna; wt, wild type. Scale bar = 5 mm in A,B, 1 mm in C,D, 2 mm in E,F,K–P, 0.5 mm in H–J.
The basioccipital, exoccipital and supraoccipital bones of the rostral skull arise from the most cranial somites [35] . Eventually, the occipital bones fuse to surround the foramen magnum but the mineralizing occipital bones of neonates are normally well separated at birth, with alcian blue-staining cartilaginous matrix between the supraoccipital and exoccipital bones and between the exoccipital and basioccipital bones ( Fig. 3C ). Col2a1-Dlx5 t19/+ hemizygotes showed variable degrees of expanded mineralization of the supra- and basioccipital bones at birth, such that these bones sometimes touched or were fused at discrete points with the adjacent edge of an exoccipital bone ( Fig. 3D ). The degree of occipital ossification was scored for wild type (n = 29) and t19/+ hemizygotes (n = 35) at two developmental stages and is shown, along with an explanation of the scoring system, in Fig. 3G . By P0, all transgenic pups examined (n = 24) showed some degree of advanced ossification in the occipital bones, compared to 1 of 18 wild type neonates. In its most severe form, as exemplified by the Col2a1-Dlx5 t1/+ founder, this hypermineralization was so excessive that the supraoccipital and exoccipital bones were completely fused, as were the first two cervical vertebrae and there was hypermineralization in the chondrocranium surrounding the interparietal bone ( Fig. 3F ). Indeed, the entire head was dysmorphic in the recovered t1/+ , t2/+ , and t3/+ founders: deeper dorso-ventrally but foreshortened rostro-caudally with pronounced shortening of the maxillary process and dentary such that the tongue protruded well past the jaws. Non-endochondral components of the skull like the dentary and frontonasal bones were likely to have been secondarily affected by accelerated mineralization in the chondrocranium and were similarly affected in all three dead founders ( Fig. 3A,B and E,F ).
The axial skeleton originates in the somitic sclerotome and vertebrae with a diversity of shapes and functions subsequently form along the rostral-caudal axis. Adjacent somites contribute to single vertebral anlagen, which assemble from three sclerotomal compartments: dorso-medial (dorsal neural arch and spinous process), lateral (laminae and pedicles of the neural arch and ribs at thoracic levels), and ventral (centrum and vertebral discs). Embryonic vertebrae have discrete ossification centres in the centrum and the neural arches, which normally fuse post-natally. The vertebrae of P1 Col2a1-Dlx5 t19/+ hemizygous neonates (n = 8) showed precocious ossification between these centers in the form of narrow bridges of mineralized tissue that joined the ossified centrum and neural arch. This precocious ossification was further advanced in Col2a1-Dlx5 t19/t19 homozygotes (n = 9) such that the majority of centra were completely fused to the adjacent neural arches ( Fig. 3H–J ). The overall morphology of vertebrae and length of the vertebral column were not strongly affected in this line. While all vertebrae of the Col2a1-Dlx5 t1/+ founder were completely fused, vertebrae were additionally dysmorphic and the vertebral axis was significantly shorter in the strongest phenotypes ( Fig. 3K,L ), resulting in the short stature of these founders ( Fig. 3A,B ).
In contrast to the axial skeleton, the limbs of Col2a1-Dlx5 neonates were mildly affected. Col2a1-Dlx5 t19/+ limbs were indistinguishable from wild type in both overall size and in the extent of mineralization. The limbs of homozygous Col2a1-Dlx5 t19/t19 neonates averaged slightly smaller than wild type littermates (e.g. femurs from homozygotes averaged 94% of wild type femur length, n = 4, Fig. 3M,N ) but this was not statistically significant. A limb phenotype became more obvious in the most severely affected founder ( Col2a1-Dlx5 t1/+ ) in which the long bones were shorter and thicker than non-transgenic controls, particularly in the hindlimb, and the relative extent of mineralization was more advanced in the femur compared to wild type controls ( Fig. 3O,P ). Finally, we did not observe truly ectopic mineralization in cartilaginous structures that do not normally ossify, like the trachea or chondrocostal cartilage. In conclusion, our observations of the skeletons from four independent Col2a1-Dlx5 alleles demonstrate that misexpression of Dlx5 in immature chondrocytes promoted endochondral ossification in a dose dependent manner and that, above a certain threshold, this resulted in neonatal lethality.
We next asked whether the timing of onset of mineralization was affected in Col2a1-Dlx5 embryos. At E14.5 there was no mineralization in the centra of either wild type or transgenic littermates, as visualized by alizarin red staining (not shown). Thereafter, the initiation of mineralization in the centrum at a specific axial level in wild type embryos was somewhat variable such that, by E15.5, the number of vertebral centra in which mineralization had commenced was quite disparate among littermates of the same genotype, varying from zero to twelve ( Fig. 4A ). While this variability continued through the next two days of development, we reasoned that, if Dlx5 were affecting the onset of mineralization, transgenic embryos would have consistently more ossification centres in their vertebral centra at a given developmental stage. This was not the case however ( Fig. 4A ). We also measured the proportion of the centrum that had mineralised at different times to ask whether, once initiated, endochondral ossification was accelerated. By E16.5, mineralization was well under way in all thoracic and lumbar vertebrae in all embryos, regardless of genotype. We found no difference in the extent of mineralization at a given axial level at this stage when comparing transgenic or wild type littermates ( Fig. 4D ). By E17.5 however, there was a clear difference in the extent of ossification of the vertebral bodies in Col2a1-Dlx5 t19/+ hemizygotes and additional areas of mineralization were first apparent that bridged the mineralising region of the centrum with mineralised tissue in the neural arches ( Fig. 4B–F ). In contrast to hemizygotes at E17.5, when bridges were faintly detectable in a variable number of vertebrae, all vertebrae from the atlas to the fourth sacral vertebrae were affected in t19/t19 homozygotes and, moreover, the two mineralization centres were completely fused ( Fig. 4G ). We also measured ossification of the ribs and found a 10% difference in the extent of ossification of the ribs at P0 ( P <0.05, n = 9 wild type, n = 12 transgenic). However, no significant difference in the extent of ossification in the long bones of the limbs was apparent up to P0 (data not shown, n = 20 wild type, n = 12 transgenic). Thus, expression of Dlx5 in somitic cartilage did not appear to affect the initial timing of mineralization, but rather contributed to a dose-dependent acceleration of ossification once initiated.
10.1371/journal.pone.0008097.g004 Figure 4
The timing of the onset of mineralization in the vertebral bodies was not affected by Dlx5 but the subsequent rate of ossification was.
(A) Bubble graph depicting the number of vertebral centra in which mineralization was detectable after alcian blue and alizarin red staining of wild type and t19/+ hemizygous embryos at three developmental stages. The smallest bubble represents a sample of one, the largest represents eight individuals. (B,C) Dorsal view of the third and fourth lumbar vertebrae of a wild type and t19/+ hemizygote at E17.5 following alcian blue and alizarin red staining. (D) Quantitation of the area of the L3 centrum that had mineralized, plotted as average area ± sem. Wild type, white bars (n = 17 and n = 18 at E16.5 and E17.5 respectively); Col2a1-Dlx5 t19/+ , grey bars (n = 4 and n = 13 at E16.5 and E17.5 respectively); * P <0.005. (E–G) Ventral views of wild type (E), hemizygous (F) and homozygous (G) embryos with the Col2a1-Dlx5 t19 allele at E17.5 following alizarin red staining. Arrowheads point to precociously mineralizing vertebrae in panel F. Scale bar = 0.5 mm for all photomicrographs.
We next sought to determine whether precocious mineralization of the endochondral skeleton was due to an underlying acceleration of chondrocyte differentiation. We examined hematoxylin and eosin stained sections through the head, trunk and limbs at various stages ( Fig. 5 ). At E16.5, before a mineralization phenotype was apparent in the caudal skull, advanced hypertrophy of the basioccipital bone was apparent. The lateral edges of the basioccipital bone contain small, rounded chondroblasts in wild type embryos, whereas hypertrophic chondrocytes occupied the entire element in transgenic littermates ( Fig. 5A,B ). Transverse sections through the vertebrae of Col2a1-Dlx5 t19/+ transgenic neonates confirmed that mineralization had occurred throughout the vertebrae, whereas the ossification centres of the centrum and neural arches of wild type pups were separated by blocks of cartilaginous tissue that contained both radially flattened and hypertrophic chondrocytes ( Fig. 5C,D ). Finally, we examined the growth plates of three-week-old mice after H&E staining. There were no gross distortions in the overall architecture of the growth plates of transgenic weanlings (n = 3) compared to wild type littermates (n = 4), with both proliferating and hypertrophic zones being of similar size, although dividing chondrocytes in the proliferative zone were stacked somewhat less regularly in transgenic mice ( Fig. 5E,F ). Thus, forced expression of Dlx5 in immature chondrocytes resulted in accelerated chondrocyte hypertrophy in the axial skeleton and a more subtle effect in the growth plate of limbs.
10.1371/journal.pone.0008097.g005 Figure 5
Misexpression of Dlx5 in chondrocytes accelerated chondrocyte hypertrophy.
(A,B) Haematoxylin and eosin (H&E) staining of coronal sections of the head of wild type (A) and Col2a1-Dlx5 transgenic (B) E16.5 embryos. Arrows point to proliferating zones in panel A. Dorsal is up. (C,D) Alizarin red-stained and cleared vertebral columns of wild type (C) and transgenic (D) P1 embryos were embedded in paraffin and sectioned. Arrows point to cartilaginous tissue in panel C. Boxed region is shown at higher magnification in panel C insert. Dorsal is up. (E,F) H&E stained longitudinal sections through the proximal humerus of wild type (E) and transgenic littermate (F). Proximal is up. Scale bar = 0.2 mm in A,B,E,F, 2 mm in C,D.
To further confirm the basis of the phenotype, we examined the expression of markers of prehypertrophic and hypertrophic chondrocytes in transverse sections through the limbs and vertebrae at times before an overt effect on mineralization was apparent. Consistent with the idea that Dlx5 promoted precocious ossification via accelerated chondrocyte hypertrophy, expression of both prehypertrophic ( Ihh ) and hypertrophic ( Col10a1 ) markers was expanded in the limbs and vertebrae of Col2a1-Dlx5 t19/+ embryos compared with wild type littermates ( Fig. 6 ). In summary, expression of Dlx5 in immature chondrocytes promoted chondrocyte hypertrophy and endochondral ossification.
10.1371/journal.pone.0008097.g006 Figure 6
Expansion of prehypertrophic and hypertrophic marker gene expression in Col2a1-Dlx5 transgenic mice.
(A–D) In situ hybridization of Ihh (A,C) or Col10a1 (B,D) riboprobes to adjacent cryosections of wild type and Col2a1-Dlx5 transgenic femurs at E14.5. Proximal is to the right. (E–H) In situ hybridization of Ihh (E,G) or Col10a1 (F,H) riboprobes to adjacent cryosections of wild type and Col2a1-Dlx5 transgenic vertebral centra at E16.5. Dorsal is up. Scale bar = 0.2 mm for all photomicrographs.
The Chondrocyte Hypertrophy Function of Dlx5 Is Cell Autonomous and Is Sufficient to Rescue Endochondral Ossification in Dlx5 /6 Null Embryos
Dlx5 and Dlx6 have shared essential functions in mandibular patterning and AER perdurance and Dlx5/6 null neonates are exencephalic as a result of a failure to ossify the skull vault; in particular, Dlx5/6 neonates lack supraoccipital and interparietal bones at birth [33] , [36] , [37] , [38] . Targeted deletion of both Dlx5 / 6 also revealed shared functions in the endochondral skeleton. Defects that arise as a result of delayed or absent chondrocyte differentiation vary from the complete absence of mineralised endochondral elements, notably the paired supraoccipital bones, to delayed mineralization throughout the vertebral and appendicular skeleton [33] , [36] . While the limb and axial skeleton of Dlx6 −/− mice has not yet been described, Dlx6 is largely redundant with Dlx5 in patterning the first branchial arch [39] . Further understanding of the functional equivalence of Dlx5 and Dlx6 in endochondral ossification requires the substitution of one for the other in an in vivo context. We asked whether cartilage-specific expression of Dlx5 could rescue cartilage maturation defects in Dlx5/6 null embryos by crossing Col2a1-Dlx5 t19 hemizygous mice (CD-1 background) with heterozygous Dlx5/6 +/− mice (C57Bl/6 x DBA background) to specifically reconstitute Dlx5 function in the cartilage of otherwise Dlx5/6 null embryos. Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ neonates showed a rudiment of a mineralized element of varying size, located dorsal to the exoccipital bones, which was interpreted to be supraoccipital in identity ( Fig. 7B , n = 4), and partial rescue of the supraoccipital was sometimes unilateral. In contrast, Dlx5/6 −/− ; Col2a1-Dlx5 t19/t19 homozygotes had mineralised supraoccipital bones that were of a similar size to those in Dlx5/6 +/− or wild type littermates ( Fig. 7C,D , n = 3). Notably, the interparietal, which forms via intramembranous ossification, was not rescued in these animals, nor was transformation of mandibular structures to a maxillary identity ( Fig. 7 ). Finally, we asked whether rescue of endochondral ossification was a general feature of these embryos by examining a more quantitative trait, namely ossification of the vertebral centra. Ossification of the centra of Dlx5/6 null embryos lags that in wild type littermates, being, on average, 76% of that in heterozygous or wild type littermates at E17.5 ( Fig. 7E–H ). Vertebral mineralization in Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ embryos, however, was indistinguishable from Dlx5/6 +/− littermates. Taken together, our data show that Dlx5 can fully compensate for Dlx6 in the endochondral skeleton. While we cannot formally exclude the possibility that Dlx5 was expressed at very low levels in the perichondrium, that were below our detection limits, our results are consistent with a cell autonomous function for Dlx5 during chondrocyte hypertrophy.
10.1371/journal.pone.0008097.g007 Figure 7
Chondrocyte-specific expression of Dlx5 rescued endochondral defects in Dlx5/6 null embryos.
(A–D) Lateral views of the heads of P0 neonates with the genotypes shown after staining with alcian blue and alizarin red. Asterisk indicates a missing supraoccipital bone in panel A. Arrows point to supraoccipital bones in panels B–D. (E–G) Dorsal view of the second to fifth lumbar vertebrae from a Dlx5/6 −/− (E), Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ (F), and Dlx5/6 +/− (G) embryo at E17.5 following alcian blue and alizarin red staining. Since Dlx5/6 −/− embryos are smaller, digital images of L2 to L5 at E17.5 were scaled to the same vertebral size to allow measurements of the relative mineralization of the vertebral bodies. Rostral is at the top. (H) Quantitation of mineralization in the L3 centrum of Dlx5/6 −/− (n = 9), Dlx5/6 −/− ; Col2a1-Dlx5 t19/+ (n = 3), and Dlx5/6 +/− (n = 9) vertebrae, plotted as average area ± sem. * P <0.05; ns, P >0.05. Scale bar = 2 mm in A–D, 0.5 mm in E,F,G.
Discussion
Dlx5 and Dlxt6 as Cell Autonomous Regulators of Chondrocyte Hypertrophy
We have generated a line of transgenic mice in which the transcription factor Dlx5 is expressed in proliferating chondroblasts. Interestingly, the level of expression of exogenous Dlx5 was similar to that of endogenous Dlx5 , as measured by in situ hybridization, with both being expressed at much lower levels than Col2a1 . Our finding is consistent with the observations of others, using essentially the same Col2a1 sequences [40] and likely reflects the absence of cis -acting sequences that contribute to the higher level expression of Col2a1 . We consider this to be an advantage of our mouse model since ectopic expression of Dlx5 is at near physiological levels in Col2a1-Dlx5 t19 hemizygotes. Through cell type-specific manipulation of its expression, we provide evidence that Dlx5 is required cell autonomously for chondrocyte hypertrophy and that Dlx6 has a redundant function in this tissue. This is best demonstrated by the fact that restoration of Dlx5 expression in chondrogenic condensations in the chondrocranium was sufficient to rescue the supraoccipital bones in Dlx5/6 -deficient mice. Rescue of the supraoccipital bones in Dlx5/6 −/− ; Col2a1-Dlx5 t19/t19 neonates further indicates that chondrogenic differentiation events prior to activation of Col2a1 are not defective in Dlx5/6 null embryos but rather that the absence of the supraoccipital bones in Dlx5/6 −/− neonates is due to a subsequent block in chondrocyte differentiation. This chondrogenic function is apparently independent of the role of Dlx5 or Dlx6 as osteoblast differentiation genes or as regulators of craniofacial patterning since Dlx5/6 −/− ; Col2a1-Dlx5 t19/t19 neonates retained intramembranous bone defects and transformation of mandibular to maxillary structures. Similarly, Dlx5 likely has an independent function in the perichondrium where it may promote differentiation to the periosteum. Another study, in which Dlx5 was specifically misexpressed in chondrocytes, documented enhanced chondrocyte hypertrophy in the limb [34] . These tissue-specific manipulations of Dlx5 expression validate the results of more generalized over-expression studies with this gene [29] .
Dlx6 −/− mice display a range of first branchial arch defects that closely resemble those seen in Dlx5 −/− neonates, but that are milder [39] . The non-additive nature of the patterning defects that occur following the combined deletion of Dlx5 and Dlx6 suggest functional redundancy of these genes in the first branchial arch. Similarly, gain-of-function studies point to a quantitatively equivalent function in stimulating multipotent precursors to differentiate into chondroblasts [28] . The functional equivalence of these genes in a specific tissue had not previously been addressed in vivo ; a definitive test of functional equivalency requiring substitution of Dlx5 coding sequences for those of Dlx6 . This requirement is satisfied in embryos with combined Col2a1-Dlx5 t19 and Dlx5/6 −/− alleles and constitutes a test of functional equivalency in cartilage. That both quantitative and qualitative defects in endochondral ossification were rescued by chondrocyte-specific expression of Dlx5 argues that Dlx5 and Dlx6 are functionally interchangeable in chondrocytes and that specific endochondral elements depend on different levels of Dlx5 or Dlx6 activity; vertebral ossification was not completely dependent on either Dlx5 or Dlx6 (since it was not completely blocked in Dlx5/6 null embryos) and was rescued to wild type levels in hemizygous Col2a1-Dlx5 embryos while complete rescue of the supraoccipital bones required homozygosity of the transgene in a Dlx5/6 null background. Nonetheless, it is quite likely that some functions of the Dlx5-Dlx6 locus will depend on diverged protein functions; some genes behave differently following loss of Dlx6 versus Dlx5 in the mandibular arch, for example [39] . This is not altogether surprising given the striking differences in the amino acid sequences of the Dlx5 and Dlx6 proteins and the differential distribution of their transcriptional activation activities [28] .
A Hierarchy of Transcription Factors in Chondrocyte Differentiation
Like Dlx5 , Runx2 (encoding two isoforms that differ at their amino-termini) is a multifunctional regulator of both chondrocyte and osteoblast differentiation [41] and the two factors have some remarkable parallels in both expression and function. Like Dlx5 , Runx2 is expressed in cartilaginous condensations and later, during long bone growth, both genes are expressed in prehypertrophic and hypertrophic chondrocytes. In addition, Dlx5 and Runx2 are expressed in the perichondrium flanking the prehypertrophic and hypertrophic zones and, in both cases, perichondrial expression extends over the proliferating zone ( [27] , [40] , [42] , [43] , [44] , [45] , [46] , [47] and see Fig. 2H ). As this study has demonstrated for Dlx5 , either Runx2 isoform accelerated chondrocyte hypertrophy in the axial and appendicular skeleton when misexpressed in immature cartilage [40] , [48] . Isoform selective deletion further revealed that Runx2-II has a non-redundant function in chondrocyte maturation [49] and Runx2-II , rather than Runx2-I , is specifically expressed in prehypertrophic and hypertrophic chondrocytes [47] . Runx2 isoforms also induced ectopic chondrocyte hypertrophy in transgenic mice, in which persistent cartilage in the trachea and chondrocostal cartilage was diverted to a hypertrophic fate [40] , [48] , and retroviral-mediated misexpression of Runx2-II converted persistent cartilage to a hypertrophic fate in the chicken hyoid skeleton [50] . In contrast, Dlx5 appears unable to induce ectopic mineralization. None of our four Col2a1-Dlx5 alleles caused ectopic mineralization in persistent cartilage. Even in tissues that normally mineralise, Dlx5 misexpression did not appear to affect the timing of the initial appearance of a calcified matrix, but rather contributed to a subsequent acceleration once begun. Indeed, in the vertebrae, the effects of an accelerated deposition of calcified matrix in the centra were not apparent until two days after mineralization was first visible. Similarly, we never observed isolated patches of mineralized vertebral tissue between the centrum and the neural arches; rather, premature mineralization was always contiguous with the major ossifying zones of the centrum and neural arches and their appearance coincided with the first detectable Dlx5 -mediated expansion of the mineralized zones of centra at E17.5. Moreover, although prehypertrophic and hypertrophic zones were expanded in transgenic mice, these zones remained discrete and well defined. We did not observe, for example, isolated cells or patches of cells that were either Ihh - or Col10a1 -positive within the proliferative zone of limb bones. This is in agreement with the observations of Chin et al . (7). In common with Runx2 then, Dlx5 accelerates chondrocyte differentiation but, in contrast, while Runx2 is capable of initiating chondrocyte hypertrophy de novo , Dlx5 is unable to induce ectopic chondrocyte hypertrophy in tissues where it would not normally occur, nor advance the initial timing of mineralization. Thus, it is likely that Dlx5 requires the synthesis of other rate-limiting critical cofactors for its hypertrophic function. This is an interesting situation given that Dlx5 appears to be a direct upstream regulator of Runx2 during osteoblast differentiation [51] , [52] . Nevertheless, our data strongly suggest that Dlx5 is not sufficient for transactivation of Runx2 in chondroblasts. Indeed, the signals and upstream transcriptional regulators of Runx2 in chondrocytes are not yet known.
The Mosaic Nature of the Skeleton
The differential sensitivity of the appendicular and axial skeleton to a given level of ectopic Dlx5 expression is not without precedent; non-uniform effects of Col2a1 -driven transgenes in the endochondral skeleton have been a hallmark of other studies too. For example, the supraoccipital bone and vertebral skeleton was more sensitive to the effects of a dominant negative PTHrP receptor than the limbs, which only exhibited an effect at the highest number of transgene copies and then only in discrete elements [53] . Furthermore, expression of a constitutively active form of Akt in immature chondrocytes led to accelerated chondrocyte hypertrophy and mineralization in the cranial and vertebral skeleton but delayed hypertrophy in limbs [54] . The differential responses of the axial and appendicular skeleton to Dlx5 and other regulators of cartilage development no doubt reflects differences in the molecular milieu of anatomically distinct chondrocytes and further underscores the mosaic nature of the skeleton.
Materials and Methods
Ethics Statement
Prior institutional approval was received for the animal work described in this study from the University of Guelph Animal Care Committee, in accordance with Canadian Council on Animal Care guidelines.
Generation of Transgenic Mice, Genotyping, and RT-PCR
The murine Dlx5 open reading frame was amplified as a 5′ Flag -tagged Hin dIII- Nco I fragment, and shuttled into a modified pSlax13 [55] . The bovine polyadenylation ( bpA ) sequence from p3000i3020Col2α1 [56] was amplified and cloned downstream of Dlx5 as an Nco I- Xba I fragment, and the FlagDlx5-bpA cassette was cloned back into p3000i3020Col2α1 as a Hin dIII- Sal I fragment, replacing the βgeo - bpA sequences. The construct was digested with Not I and Sal I and the linear transgene cassette was microinjected into the pronuclei of fertilized CD-1 oocytes. Genotyping of transgene bearing mice was done with forward primer 5′-AACAGTTCCCCGAAAGAGGT-3′ and reverse primer 5′-GAGCGCTTTGCCATAAGAAG-3′ , which amplified a 1040 bp fragment. The Col2a1-Dlx5 t19 allele was maintained in a hemizygous state on a CD-1 background; hemizygous animals were occasionally bred to generate homozygous embryos or neonates. Offspring from Col2a1-Dlx5 t19/+ x Dlx5/6 +/− crosses were backcrossed to Dlx5/6 heterozygotes to generate transgene-positive, Dlx5/6 null mice. A 438 bp genomic Ihh fragment was amplified with primers: 5-CACTTGTGGTGGAGGATGTG-3′ and 5′-TACCACACGCTTGTCAGCTC-3′ . Dlx5/6 heterozygotes were genotyped with the lacZ primer pair: 5′-GCGTTACCCAACTTAATCG-3′ and 5′-TGTGAGCGAGTAACAACC-3′ [32] . The presence of FlagDlx5 and β - actin mRNA was determined on total RNA prepared from E16.5 embryos using the Trizol Reagent (Invitrogen). 1 µg of RNA was reverse transcribed (Superscript II, Invitrogen) and PCR amplified (Taq, UBI Life Sciences) with standard protocols. Primer sequences used for RT-PCR were: FlagDlx5 (For 5′-GACTACAAGGACGACGATGAC-3′ and Rev 5′-GAGCGCTTTGCCATAAGAAG-3′ ) and β - actin (For 5′-GAGAAAATCTGGCACCACACC-3′ ; Rev 5′-CAGGAAGGAAGGCTGGAAGAG-3′ ).
Skeletal Staining
Alcian Blue and Alizarin Red staining was used to visualize cartilaginous and mineralised skeletal tissues respectively of embryos and neonates. Briefly, eviscerated and skinned (>E16.5) bodies were fixed in 95% ethanol over several days. Embryos were stained in 80% ethanol, 20% acetic acid, 0.01% Alcian blue 8GX for 2–4 days, cleared overnight in 1% KOH, stained 1–2 days in 0.001% Alizarin Red S in 1% KOH, then rinsed in 1% KOH, and a graded series of H 2 O/glycerol to 100% glycerol.
Histology
Tissues were fixed overnight in 4% paraformaldehyde then dehydrated and embedded in paraffin using standard techniques. 7 µm sections were stained with hematoxylin and eosin using a standard protocol.
In Situ Hybridisation
Whole mount and cryosection in situ hybridisation was done as described previously [55] with the following antisense riboprobes: Dlx5 , a 0.87 kb Bam HI – Hin dIII fragment corresponding to the full length open reading frame [28] ; Col2a1 , a 0.4 kb cDNA corresponding to nt 1–402 of Genbank X57982; Ihh , a 1.8 kb partial cDNA Eco RI fragment [57] ; and Col10a1 , a 0.86 kb Apa I - Sal I fragment containing nt 1351–2215 of Genbank X65121.
Imaging
Images were taken using a MicroPublisher colour digital camera on a Leica MZ12.5 Stereomicroscope with Qcapture software (QImaging) or on a Leica DMRA2 upright microscope with Openlab software (Improvision) and processed using Adobe Photoshop.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats) is a powerful gene editing tool capable of performing DNA cleavage with the help of guide RNAs (gRNAs) and the constitutive expression of Cas9 [ 1 ] [ 2 ]. However, substantial numbers of gene knockout (KO) experiments using pLentiCrispr-V1 or pLentiCrispr -V2, or pLenti-Cas9 plus pLenti-Guide-Puro (Version 3)[ 3 , 4 ] have failed. Factors contributing to the failure include low titers of the lentiviral Cas9, less-than-perfect gRNAs, inefficient selective markers for constitutive expression of Cas9, difficult-to-transfect target cells (e.g., primary cells), and wildtype (WT) or WT-like clones overgrowing and taking over edited cell pools.
The gene knockout (KO) efficiency of CRISPR/Cas9 mainly depends on cellular Non-homologous End Joint (NHEJ) repair to maintain genomic integrity if a complementary donor is not present when Cas9 cleaves both double strands at gRNA binding sites. NHEJ is an error-prone cellular mechanism that repairs ends with mismatch or frameshift (i.e., 1-nt and 2-nt indels) at the break sites of genomic DNA, thus disrupts target gene expression. However, NHEJ repair resulting in in-frame change(s) of the coding DNA sequences (cDNAs), may still be detectable by Western blotting (WB) if the antibodies are designed for the C-terminals of proteins, though detectable by T7 endonuclease assay, a semi-quantitative editing assay method. Our experiments (unpublished) indicated that, in cell pools with lower numbers of gene-edited cells, wildtype (WT) or WT-like cells would overgrow and dominate the cell pools, leading to negative Western blot results and gene KO failure. Therefore, selecting stable KO clones through single cloning or single-cell cloning from transduced pools is critical to ensure KO efficiency.
With insertion of a fluorescent marker mOrange into the lentiviral vector (i.e., CRISPR/Cas9 plasmid), lentiviral production and titer could be visually monitored and estimated because HEK293T cells become brilliant orange colored. According to our observations, generation of high-titer viruses is critical for target cells to be successfully transduced. mOrange fluorescence also indicates Cas9 expression and puromycin resistance, because all three are under the control of the same EF1α promoter. When target cells are transduced with lentiviruses, cells expressing Cas9 and puromycin resistance become visible as well, and moreover, cells can be sorted by flow cytometry, and viral titer can be assayed by fluorescence-activated cell sorting (FACS) instead of the lengthy colony formation method. A few studies have used other fluorescent markers, such as green fluorescent protein (GFP) or red fluorescent protein (RFP), which were inserted to replace the puromycin resistance gene [ 5 – 7 ]. However, puromycin selection for positive clones is an easy, economical, and fast means of selecting target cells. Currently there is no pLentiCrispr vector with selection markers for both puromycin and fluorescence.
Many driver genes in human cancer cells are mutated either actively or silently. Currently, gain- and loss-of-function assays are still critical experiments for clarifying if a new mutant gene is a driver mutation. Two challenges exist: first, when a gene of interest is endogenously expressed in target cells with a high basal level and in all cell lines, it is problematic to investigate gain-of-function of the mutant by directly re-expressing the mutant gene in the target cells. To this end, an empirical method would be to knock out the endogenous expression using CRISPR technology and then re-introduce a mutant(s) into the knockout population. Second, when a gene is knocked out in cell lines, a cascade of downstream molecular and cellular markers is often altered. Oftentimes, researchers or reviewers want to know if rescue of the original gene would reverse the downstream changes in order to authenticate the gene function. In both scenarios, reintroducing a mutant(s) and/or rescuing of a wildtype gene would fail because Cas9-gRNA by nature disrupts introduced genes. Here in this paper, we report how to overcome both hurdles by using CRISPR technologies with modified cDNAs for gene rescue.
Materials and methods
Gastric cancer cell lines
Gastric cancer cell lines N87 (CRL-5822), AGS (CRL-1739), KatoIII (HBT-103), Snu1 (CRL-5971) and Snu16 (CRL5974) were ordered from American Type Culture Collection (ATCC) (Manassas, VA). Cell lines Ycc1 (Yonsei Cancer Center, Seoul, South Korea), Ycc2, GT5 (SK-GT5, Memorial Sloan Kettering Cancer Center, New York, NY), MKN45 (CVCL_0434, Cancer Cell Line Encyclopedia Project [CCLE]) were procured earlier and preserved in this research lab.
Insertion of mOrange into the lentiviral plasmid pLentiCrispr-Cas9-V2 to create V2mO
The plasmid pLentiCrispr-Cas9-V2 (Addgene, Cambridge, MA, #52961) is commonly used for gene knockout. In our improved plasmid, a fluorescence gene, mOrange, was amplified with Q5 hifi polymerase (New England Biolabs, Ipswich, MA) with two primers, Crisprv2.mOrange.F and Crisprv2.mOrange.R ( Table 1 ). The forward primer carried a BamHI site and a sequence that overlapped with P2A; the reverse primer carried a Tth111I site, T2A sequence, and a sequence that overlapped the puromycin resistance gene. pLentiCrispr-Cas9-V2 was double-digested with BamHI and Tth111I, and the fragment was gel-purified and ligated with the Q5-PCR product of mOrange using a T4 DNA ligase (New England Biolabs). The resultant plasmids were screened and verified by the double digestion of BamHI and Tth111I, and potential clones were verified by sequencing using hSpCas9.out.F and PuroVar.out.R primers. The end plasmid was V2mO ( Fig 1A ), which is available from Addgene.
10.1371/journal.pone.0228910.g001
Fig 1
Construction of pLentiCrispr-V2-mOrange (V2mO), its lentiviral production in HEK293T and transduced mOrange expression in GC cell lines AGS and GT5.
A. V2mO plasmid construct, with mOrange cistronically inserted between Cas9 and puromycin cDNAs. Between the original pLentiCrispr-Cas9-V2 plasmid (top) and the V2mO plasmid (bottom) are shown the overlapping sequences and restriction enzymes for the construction. B . HEK293T cells expressed mOrange in lentiviral production, which enables estimation of transformation efficiency and viral titer. C & D . Target cell lines AGS (C) and GT5 (D) were transduced with lentiviral V2mO-RhoA.g5 after puromycin selection and mOrange sorting. E. Cell line GT5 was transduced with lentiviral V2mO-Gli1.g4 after puromycin selection and mOrange sorting. F & G . Cell lines AGS (F) and GT5 (G) were transduced with lentiviral V2mO-Gal3.g1 after puromycin selection and mOrange sorting.
10.1371/journal.pone.0228910.t001
Table 1 Primers used in this study.
PCMVfor
5’ CGCAAATGGGCGGTAGGCGTG 3’
T7
5’ TAATACGACTCACTATAGGG 3’
Crisprv2.mOrange.F
5’GATTACAAAGACGATGACGATAAGGGATCCGGCGCAACAAACTTCTCTCTGCTGAAACAAGCCGGAGATGTCGAAGAGAATCCTGGACCGATGGTGAGCAAGGGCGAGGAGAATAACATGGCCATC 3’
Crisprv2.mOrange.R
5’TGGGGACGTCGTCGCGGGTGGCGAGGCGCACCGTGGGCTTGTACTCGGTAGGGCCGGGATTCTCCTCCACGTCACCGCATGTTAGAAGACTTCCTCTGCCCTCCTTGTACAGCTCGTCCATGCCGCCG 3’
hSpCas9.out.F
5’ TGTACGAGACACGGATCGAC 3’
PuroVar.out.R
5’ ACACCTTGCCGATGTCGAG 3’
hRhoA.E3.gRNA5.F
5’ caccgGAACTATGTGGCAGATATCG 3’
hRhoA.E3.gRNA5.R
5’ aaacCGATATCTGCCACATAGTTCc 3’
hRhoA.Dn5Fk.F3.Sac2
5’ aaaCCGCGGaggtggatcggcgtactaga 3’
hRhoA.Dn5Fk.R2.EcoR1
5’ aaaGAATTCagatggcaggatgagaatgg 3’
hRhoA.F.H3.Not1
5’ TGACAAGCTTGCGGCCGCtATGGCTGCCATCCGGAAGAAACTGG 3’
hRhoA.R.Stop.Xba1.BamH1
5’ aaGGATCCTCTAGActaCAAGACAAGGCACCCAGATTTTTTCTTC 3’
hRhoA.Y42C.F2
5’ GTGCCCACAGTGTTTGAGAACTGTGTGGCAGATATCGAGGTGGATG 3’
hRhoA.Y42C.R2
5’ CATCCACCTCGATATCTGCCACAcAGTTCTCAAACACTGTGGGCAC 3’
RhoA.Wt.F.Kpn1
5’ aaGGTACCATGGCTGCCATCCGGAAGAAACTGGTGATTGTTGGTGATG 3’
RhoA.cFlag.R.Apa1
5’ttGGGCCCTTACTTGTCATCGTCATCCTTGTAATCGATGTCATGATCTTTATAATCACCGTCATGGTCTTTGTAGTCCATTCTAGAcaagacaaggcacccagattttttcttcccacg3’
hRhoA.seq.F
5’ cggtctggtcttcagctacc 3’
hRhoA.seq.R
5’ TTAACCGCATAAGGGCTGTG 3’
hGli1.E2.gRNA2.F
5’ caccgGGCTCGCCATAGCTACTGAT 3’
hGli1.E2.gRNA2.R
5’ aaacATCAGTAGCTATGGCGAGCCc 3’
hGli1.E5.gRNA4.F
5’ caccgAGGAAGGCGAGGGCCCTTTT 3’
hGli1.E5.gRNA4.R
5’ aaacAAAAGGGCCCTCGCCTTCCTc 3’
hGli1.seq.g34.F
5’ cagtacttccctgggactgc 3’
hGli1.seq.g34.R
5’ ccagcacccacacctcttta 3’
hGal3.E3.gRNA1.F
5’ caccgCAGACCCAGATAACGCATCA 3’
hGal3.E3.gRNA1.R
5’ aaacTGATGCGTTATCTGGGTCTGc 3’
hGal3.H3Not1.F
5’ aaaAAGCTTGCGGCCGCaatggcagacaatttttcgctc 3’
hGal3.XbaBamH.R
5’ aaaGGATCCTCTAGAttaTATCATGGTATATGAAGCACTG 3’
hGal3.seq.F
5’ tgcctttgccatattcctct 3’
hGal3.seq.R
5’ gataagctccaggtgctcca 3’
hGal3.g12.dg1.F1
5’aaaGCGGCCGCaatggcagacaatttttcgctccatgatgccttatctgggtctggaaacccaaaccctc3’
hGal3.g12.dg2.F2
5’aaaGCGGCCGCaatggcagacaatttttcgctccatgatgccttgtctgggtctggaaacccaaaccctc 3’
hGal3.g12.dg3.F3
5’aaaGCGGCCGCaatggcagacaatttttcgctccatgatgccttgtcagggtctggaaacccaaaccctc 3’
Guide RNA design, lentivirus production, target cell transduction, puromycin selection and mOrange cell sorting
Our gRNA design followed the methods on Dr. Feng Zhang’s website ( http://crispr.mit.edu ; Massachusetts Institute of Technology; site no longer active). We also referred to German Cancer Research Center’s E-Crisp website ( http://www.e-crisp.org/E-CRISP/designcrispr.html ). Every gene was designed with 3–5 targets of gRNA sequences. Based on previous data, guide sequences from three genes, i.e. RhoA gRNA5, Gli1 gRNA2 and gRNA4, and Gal3 gRNA1, were carried out in this study.
To insert gRNA sequences, duplexes were ligated into V2mO. Briefly, guide RNA forward and reverse primers were allowed to form duplexes in a mix of equal molar concentrations and volumes in a heating block heated from 100°C and cooled gradually on its own. The duplex was then used as an insert ligated using T4 DNA ligase (New England Biolabs) into V2mO precut by BsmbI (New England Biolabs). The ligates were transformed into Stbl3 competent cells and the resultant clones were screened by cracking gel for insert sizes and verified by sequencing. The end lentiviral plasmids were called V2mO-RhoA.g5, V2mO-Gli1.g2, V2mO-Gli1.g4, and V2mO-Gal3.g1.
These four plasmids were then co-transfected with either psPax2 or pCMV.Dr8.2 and with either pMD2.G or pCMV.VSV.G in a ratio of 10:10:1 into HEK293T (ATCC, Manassas, VA) cells with ~70% confluency using Lipofectamine 3000 (ThermoFisher Scientific, Carlsbad, CA) or JetPrime (Polyplus, Illkirch, France). mOrange fluorescence was monitored for transfection efficiency and viral titer estimation. Lentiviral supernatants were harvested in 24 hrs and re-harvested 24 hrs consecutively for 2 nd time and 3 rd time, lentiviral supernatants were filtered and concentrated by the PEG8000 method into 10 times concentrated. Target AGS and GT5 cells were seeded in 6-well plates with ~70% confluency. Lentiviral supernatants or concentrates were added to the cells with 8 μg/mL polybrene. Transduced cells were then selected by puromycin for 3–6 days using a concentration based on killing curves. Surviving cells were propagated and sorted for mOrange+ cells with a MoFlo flow cytometer in our institutional Flow Cytometry and Cellular Imaging Facility.
Single cloning, KO verification with Western blot and single clone sequencing
Puromycin-selected and mOrange-sorted pools were subjected to single cloning. The cells were diluted to ~50–100 cells in a 10-cm plate for single cloning. Single clones were propagated and picked in about 1–2 weeks and Western blot was used to verify knockout of RhoA (mAB antibody #2117; Cell Signaling Technology, Danvers, MA, USA), Gli1 (mAB antibody #3538; Cell Signaling Technology) and Gal3 [using an anti-Gal3 antibody made as previously reported [ 8 ] [ 9 ]). A set of 5 to10 single clones with RhoA or Gal3 knockout were mixed to form KOmixes. To verify if those single clones were truly single or mixed, each clone was subjected to genomic DNA extraction using 50mN NaOH and 1M Tris (pH 7.0), followed by phenol and chloroform purification. PCR products were amplified using Q5 hifi polymerase (New England Biolabs). For the RhoA gene, the primers were hRhoA.Dn5Fk.F3.Sac2 and hRhoA.Dn5Fk.R2.EcoR1. For the Gal3 gene, the primers were hGal3.seq.F and hGal3.seq.R. The PCR products were visualized in 2.0% gel and purified for sequencing. Short electropherograms of the areas near the gRNA binding regions were used for further analyses.
Short PCR electropherograms for direct estimation of editing efficiency
The initial puromycin-selected and mOrange-sorted pools were harvested and genomic DNAs were extracted, Q5 hifi PCR performed using the method as described above. For the RhoA gene and Gal3 gene, the primers were the same as above. For the Gli1 gene, the primers were hGli1.seq.g34.F and hGli1.seq.g34.R. PCR electropherograms were used directly from Sanger sequencing. For direct estimation of gene editing efficiency, short PCR electropherograms of the areas surrounding the guide RNA regions of cell pools were selected at regions with aberrations. A horizontal line was drawn to average aberrant nucleotide peaks, and then a second line was drawn to average WT nucleotide peaks. Direct estimation of editing efficiency was calculated as Ratio = height of aberration line /(height of aberration line + height of WT line). If the two lines merged at equal height, then the ratio was estimated to be 50% or more.
Tracking of indels by decomposition (TIDE) and interference of CRISPR edits (ICE) analyses
Tracking of indels by decomposition (TIDE) analyses followed authors’ instruction [ 10 ] to identify the major induced mutations in the projected editing sites and determine their frequencies in the cell populations. Briefly, genomic DNAs were extracted from both parental cell lines and CRISPR/Cas9-selected pools of AGS and GT5 pools, then, short PCR products were amplified as described above. PCR primers for RhoA, Gli1, and Gal3 were the same as those used for single clone sequencing and direct estimation of editing efficiency. These analyses were presented as two graphs: the first graph shows the indel spectrum of CRISPR/Cas9-gRNA transduced cell populations compared to parental cell populations, and the second graph shows aberrant sequence signals of CRISPR/Cas9-gRNA transduced cell populations compared to parental cell populations.
Interference of CRISPR edits (ICE) analysis [ 11 ] used the same two Sanger sequences as TIDE to analyze indel plots and discordance graphs: one from the control or parental sequence, and the other from the Cas9-gRNA transduced cell population. The PCR primers were the same as those used in TIDE analysis.
Construction of expressing vectors p3xFlag-RhoA.Wt, -RhoA.Y42C mutant, and -Gal3.Wt wildtype genes
For the RhoA gene, RhoA.Wt cDNA was Q5-PCR-amplified using hRhoA.F.H3.Not1 and hRhoA.R.Stop.Xba1.BamH1, digested with Not1 and BamH1, gel-purified, and ligated into the expression plasmid p3xFlag-CMV-10. Mutant RhoA.Y42C was generated by overlaying two PCR products into a single PCR product: a) product by hRhoA.F.H3.Not1 and hRhoA.Y42C.R2, and b) product by hRhoA.Y42C.F2 and hRhoA.R.Stop.Xba1.BamH1. The expression plasmids p3xFlag-RhoA.Wt and p3xFlag-RhoA.Y42C mutant were digestion-verified and confirmed by sequencing using hRhoA.seq.F and hRhoA.seq.R. For the Gal3 gene, Gal3.Wt cDNA was amplified using hGal3.H3Not1.F and hGal3.XbaBamH.R and subcloned into p3xFlag-CMV-10, and the final plasmid p3xFlag-Gal3.Wt was digestion-verified and sequenced by PCMVfor. Those three plasmids were aligned against GenBank data to make sure no nucleotides were altered, except the intended RhoA.Y42C mutation.
Construction of rescue plasmids N-terminal Flag p3xFlag-RhoA.dgWt, mutant -RhoA.dgY42C, wildtype -Gal3.dg1, -Gal3.dg2, -Gal3.dg3 and C-terminal Flag p3.1-cFlag-RhoA.dgWt and -RhoA.dgY42C
Because Cas9 cleaves where gRNA binds, one or a few nucleotides in the gRNA5 binding site of the RhoA.Wt, RhoA.Y42C and in the gRNA1 binding site of the Gal3 gene were modified, but without any amino acid being altered. In the RhoA gene, for construction of N-terminal Flag RhoA rescue plasmids, short Q5 PCR products were generated using hRhoA.F.H3.Not1 in combination with hRhoA.dgWt.R for the wildtype and hRhoA.dgY42C.R for the mutant. Clones were screened using digestion by AgeI, BamHI, and EcoRV because modified DNA sequences were absent in the EcoRV site. Positive clones were verified by sequencing. The modified plasmids were called p3xFlag-RhoA.dgWt and p3xFlag-Rhoa.dgY42C; in both of these, three nonconsecutive nucleotides were altered in the gRNA5 binding site, but the amino acid sequences remained unaltered. For construction of C-terminal Flag RhoA rescue plasmids, RhoA.Wt and mutant RhoA.Y42C, three nonconsecutive nucleotides altered RhoA.dgWt and RhoA.dgY42C, were created in the same fashion with primers RhoA.Wt.F.Kpn1 and RhoA.cFlag.R.Apa1. RhoA-Flag PCR products were cloned into pcDNA3.1+. Potential clones were digestion-verified and sequencing-confirmed by T7. The end plasmids were p3.1-cFlag-RhoA.Wt, -RhoA.Y42C, -RhoA.dgWt and -RhoA.dgY42C.
For the Gal3 gene, Q5 PCR paired hGal3.XbaBamH.R with hGal3.g12.dg1.F1, hGal3.g12.dg2.F2, and hGal3.g12.dg3.F3 to generate p3xFlag-Gal3.dg1, -Gal3.dg2, and -Gal3.dg3 plasmids, respectively. Clones were subjected to digestion and sequencing to verify that a single nucleotide, two nonconsecutive nucleotides, and three nonconsecutive nucleotides, respectively, were altered in the cDNAs of gRNA1 binding site of the three plasmids.
Results and discussion
The V2mO plasmid with mOrange fluorescent marker expressed in-frame with Cas9 and puromycin gene
To create our improved lentiviral plasmid, mOrange cDNA was integrated cistronically, as EF1α-Cas9-P2A-mOrange-T2A-puromycin, into pLentiCrispr-V2 (Addgene, Watertown, MA; #52961) [ 3 ]. The resultant V2mO vector contained three cDNAs of Cas9, mOrange, and puromycin resistance gene, all under the control of a single EF1α promoter. This single polycistronic mRNA could be translated and cleaved, at P2A and T2A self-cleaving peptides, into three functional individual Cas9, mOrange, and puromycin proteins. The integrity of the other components of the original pLentiCrispr-V2 was maintained without any alteration.
Sequencing of potential plasmids using hSpCas9.out.F and PuroVar.out.R verified that two clones were correct ( Fig 1A ). The original vector, pLentiCrispr-V2 is 14.873 kilobases (kb) in length and, with insertion of the 708-basepair (bp) mOrange, resulted in V2mO of 15.635 kb. With gRNAs inserted, i.e., with the insertion of 20 bp to replace the 1880 bp filler fragment (a nonfunctional fragment for cloning purpose), the V2mO is actually shorter than the popular pLentiCrispr-V2 empty vector (without gRNA insertion). Theoretically, the bigger a plasmid is, the lower its transfection efficiency will be.
To test our improved vector, V2mO was transformed into HEK293T cells with packaging and envelope plasmids. With one gRNA inserted, V2mO transformed competent HEK293T cells with close to 100% fluorescence observed under a fluorescence microscope ( Fig 1B ). Using the fluorescent marker, we were able to estimate the lentivirus titer in HEK293T cells and the transduction efficiency in target cells. HEK293T cells brightly expressing mOrange not only indicate high-titer viral production, but also add an extra step for target cells’ mOrange sorting by FACS on top of puromycin selection. Our results showed that lentiviral V2mO with gRNAs (e.g., RhoA-gRNA5, Gli1-gRNA4, or Gal3-gRNA1) was also expressed in target cell lines AGS and GT5 (see V2mO-RhoA.g5, Fig 1C and 1D ; V2mO-Gli1.g4 in GT5 cells, Fig 1E ; and V2mO-Gal3.g1, Fig 1F and 1G ), although mOrange expression in target cells was much weaker than in HEK293T cells.
There are a few reports in which puromycin cDNA in the LentiCrispr/Cas9 vector was substituted with a GFP marker, including pSpCas9(BB)-2A-GFP (Addgene #48138) [ 5 ] and pL-CRISPR.EFS.GFP (Addgene #57818) [ 12 ], and with an RFP marker, including LentiCrispr-RFP (Addgene #75162) [ 6 ]. In these studies, GFP and RFP provided good indicators of viral titer and clone selection for target cell screening by flow cytometry. However, puromycin selection is one of the most effective and easiest methods of positive clone selection. Replacement of the puromycin gene by fluorescence may be inconvenient for some laboratories, because they may not have access to a flow cytometer for cell sorting. In contrast, our inclusion of both a puromycin selector and the fluorescent marker mOrange enables inexpensive fast selection and visualization.
RhoA gene knockout by use of V2mO-RhoA-gRNA5
For RhoA gene, Western blot showed that all of our gastric cancer (GC) lines endogenously expressed RhoA protein ( Fig 2A ). Sequencing of those cell lines showed that they are all Y42 wildtype (unpublished data); thus, to reintroduce our intended RhoA.Y42C mutant, endogenous expression knockout had to be performed first. After lentiviral transduction of V2mO-RhoA-gRNA5 into the GC cell lines AGS and GT5, puromycin selection, and mOrange sorting, pools of surviving cells in both cell lines showed that the majority of cells were negative for RhoA expression, although AGS cell pool 2 still showed strong RhoA expression ( Fig 2B ). Of the 11 single clones picked out of AGS pool 1, four remained RhoA positive ( Fig 2C ). All three GT5 pools showed significantly reduced RhoA expression ( Fig 2B ). Out of the eight single clones picked out of GT5 pool 1, only two remained RhoA positive, six clones were negative ( Fig 2D ). Notably, the last GT5 sample which was from the same pool 1 (as in Fig 2B ) passed a few passages used to make single cloning, had turned to RhoA positive at the time of Western blot screening. Therefore, it is imperative that single cloning or single-cell cloning from the transduced pools to single out expression-positive clones and obtain stable knockout clones.
10.1371/journal.pone.0228910.g002
Fig 2
Western blots showing expressions of RhoA, Gli1 and Gal3 in GC cells, their transduced pools and respective single clones isolated from potential pools.
A. Western blot showed elevated expression of endogenous RhoA in all tested gastric cancer cell lines. B. Western blots showed levels of RhoA knockout in cell lines AGS and GT5 transduced with lentiviral V2mO-RhoA.g5 after puromycin selection and mOrange sorting. C & D. In the AGS and GT5 cell lines, respectively, Western blots of RhoA showed single clones from pool 1 that was transduced with V2mO-RhoA.g5. E. Western blots showed Gli1 pools transduced with lentiviral V2mO-Gli1.g2 and g4. F & G. Western blots of Gal3 in AGS and GT5 cell lines, respectively, and their single clones after transduction with lentiviral V2mO-Gal3.g1.
RhoA antibody was generated by immunization with a partial C-terminal peptide of RhoA’s 183 amino acids (unpublished communications, Cell Signaling Technology); therefore, Cas9’s editing of genomic RhoA DNA strands with guide gRNA5 was followed by NHEJ repair. Any repair resulting in an indel of triplicate nucleotides would not change the amino acid sequence in the C-terminal and would therefore still be detectable by Western blot, although some of them may have one or two amino acids changed. However, negative clones detected by Western blot indicated that the RhoA gene had been truly disrupted.
To view genome alteration in the vicinity of RhoA gRNA5 binding region, genomic PCR electropherograms were compared. For RhoA KO single clones, PCR electropherograms of genomic DNAs from AGS cell line showed that one clone had an extra A inserted in the 16 th /17 th nucleotide of gRNA5 binding region, the other clone was a two-clone mixture ( S1 Fig ). In GT5 cell line, one clone KO1 observed an missing A in the 17 th nucleotide of gRNA5 binding region, a second clone KO3 saw a mixture of two clones, and the other two clones KO6 and OK8 saw there were 12-nucleotide and 26-nucleotide deletions, respectively, in gRNA5 binding region ( S2 Fig ).
Gli1 gene knockout by use of V2mO-Gli1-gRNA4
For Gli1 knockout, four gRNAs were initially designed. Preliminary tests showed that two, Gli1-gRNA2 and gRNA4, worked. However, in the GT5 cell line, gRNA2 reduced Gli1 expression and Gli1-gRNA4 reduced it substantially indicating that significant numbers of cells had Gli1 knockout ( Fig 2E ). Puromycin-selected and mOrange-sorted cells were fluorescent in abundant numbers (see Fig 1E ), consistent with sequence aberration on the Gli1 PCR electropherogram ( Fig 3C ). However, it is noted that gRNA4-treated cells that, despite initial Gli1 reduction (early pool), with more cell passages (late pool) Gli1 expression bounced back ( Fig 2E ), suggesting there were wildtype or wildtype-like cells that increasingly became dominant, rendering Gli1 expression detectable through Western blot.
10.1371/journal.pone.0228910.g003
Fig 3
Sequencing electropherograms of PCR products and direct estimation of editing efficiencies around indicated binding sites.
A. RhoA gRNA5 binding site in AGS cells. Editing efficiency ratio = 12.5%. B. RhoA gRNA5 binding site in GT5 cells. Editing efficiency ratio = 28.5%. C. Gli1 gRNA4 binding site in GT5 cells. Editing efficiency ratio = 36.1%. D. Gal3 gRNA1 binding site in AGS cells. Editing efficiency ratio = 26.7%. E. Around the Gal3 gRNA1 binding site in GT5 cells. Editing efficiency ratio = 50%.
Gal3 gene knockout by use of V2mO-Gal3-gRNA1
Our previous experiment showed that Gal3-gRNA1 successfully knocked out Gal3. However, after lentiviral transduction, puromycin selection, and mOrange sorting ( Fig 1F ), Western blot still picked up Gal3 expression in the original pool of AGS cells, albeit at significantly lowered levels ( Fig 2F ). Of seven single clones from that pool, five were negative and two were positive for Gal3 expression. In the GT5 cell line, Western blot showed that all five selected clones were negative for Gal3 expression ( Fig 2G ), which was consistent with our PCR electropherogram data ( Fig 3E ). It is important to note that, in the GT5 cell line, Western blot of Gal3 revealed that the GT5-gRNA1 late pool with more passages expressed Gal3 protein, whereas the GT5-gRNA1 early pool did not. This strongly suggests that, in a mixed pool of cells transduced with lentivirus and subjected to puromycin selection and mOrange sorting, individual cells were still expressing Gal3 or Gal3 wildtype-like proteins detectable by Gal3 antibody. Based on our findings for the RhoA and Gal3 genes, it is clear that, for a gene to be edited by Cas9, a proper gRNA needs to bind to the binding site; second, in a virus-transduced pool, wildtype or wildtype-like cells may overgrow and dominate over passages, thus, it is necessary to perform single cloning in a mixed cell population.
To view how the genome was altered in the vicinity of Gal3 gRNA1 binding region, genomic PCR electropherograms were compared. For the Gal3 KO single clones, PCR electropherograms of genomic DNAs from AGS cells showed that two clones KO3 and KO5 had an extra T inserted in the 17 th nucleotide of gRNA1 binding region. In addition, clones KO1 and KO7 were mixtures of two or more clones, alteration happened even before gRNA1 binding region ( S3 Fig ). In GT5 cell line, one clone KO23 observed the same extra T inserted in the 17 th nucleotide of gRNA1 binding region, the other two clones, KO3 and KO9, were mixtures of two or more clones. Alteration occurred at the 18 th and 19 th nucleotide, respectively, of gRNA1 binding region ( S4 Fig ).
Direct estimation of gene editing efficiency with Sanger electropherograms
We noticed that sequencing Sanger electropherograms of the PCR products flanking the gRNAs gave informative visualization of the gene editing in the cell pools transduced by Cas9-gRNAs. Guide RNAs (gRNAs) were aligned with the Sanger electropherograms, with the gRNA direction pointing to the protospacer adjacent motif (PAM) sequences (NGG). Sequence aberrations, other than noise in the midst of gRNA binding sites before the PAM sequence NGG, were identified, and systemic aberration of those nucleotides compared to wildtype nucleotide peaks was given as direct estimations of the editing efficiencies. Direct estimation of editing efficiency was calculated as Ratio = height of aberration line /(height of aberration line + height of WT line). In the AGS cell line, RhoA gene editing clearly began at the 17th nucleotide of the gRNA5 site ( Fig 3A ), with an editing efficiency of 12.5% of the cell pool. In the GT5 cell line, RhoA gene editing was detected at the 5th single nucleotide and systemic editing started at the 16th nucleotide ( Fig 3B ) and afterwards, suggesting random repair by NHEJ in absence of a donor fragment. The estimated editing efficiency was 28.5% in the GT5 cell line. In the same cell line, Gli1 gene editing was detected at the18th nucleotide of the gRNA4 site ( Fig 3C ), with an editing efficiency of 36.1% of the cell pool. In both the AGS and GT5 cell lines, gene editing of the Gal3 gene was detected at the 16th nucleotide, with editing efficiencies of 26.7% for AGS cells ( Fig 3D ) and 50% for GT5 cells ( Fig 3E ).
On the basis of our results for RhoA, Gli1, and Gal3 gene editing, the following observations were made. First, it appears that most gene editing begins between the 16th bp and 18th bp of the gRNA sites, which is consistent with findings in the literature that Cas9 cuts at gRNA’s 17th nucleotide [ 1 ] [ 13 ]. However, in the absence of donor DNA, cellular repair by NHEJ may randomly produce other forms of editing, such as that we observed at the 5th nucleotide of RhoA gRNA5 in the GT5 cell line. Second, Sanger electropherograms are informative and could be used for direct estimation of gene editing efficiencies of cell pools. Third, because Cas9 is already constitutively expressed as it is integrated into the genome of a cell line, it becomes advantageous that additional gene or genes knockout could be added on top of the gene of interest; only one or more gRNAs in a suitable vector would be required. For example, Lenti-Guide-Puro (Addgene #52963) [ 3 ] with an additional gRNA, which also adds higher viral titer, or Lenti-multi-Guide (Addgene #85401) could be used with multiple gRNAs [ 14 ].
Comparison of the editing efficiency by direct electropherogram estimation, TIDE and ICE analyses
TIDE analysis identifies the major induced mutations in a projected editing site and determines their frequency in a cell population. This analysis uses a sophisticated algorithm adopting only two electropherograms: one from a Cas9-gRNA-treated cell population, and the other from an untreated control or parental cell population. In our study, TIDE analysis provided very informative data showing that RhoA-gRNA5 and Gal3-gRNA1 treatment resulted in statistically significant (p<0.001) insertion and/or deletion frequencies ( Fig 4A1, 4B1, 4D1 and 4E1 , red short columns), with the expected Cas9 cut at the beginning of the aberrant sequences ( Fig 4A2, 4B2, 4C2, 4D2 and 4E2 ). This analysis was consistent with the 16th-18th nucleotide cut detected by direct electropherogram estimations ( Fig 3 ). V2mO-RhoA-gRNA5 gave overall editing efficiencies of 15.6% and 35.7% in the AGS and GT5 cell lines respectively, and V2mO-Gal3- gRNA1 gave overall editing efficiencies of 27.0% and 41.5% in the AGS and GT5 cell lines respectively. The only inconsistency in results between TIDE analysis and direct electropherogram estimation was seen for the GT5 Gli1-gRNA4 cell population, in which direct electropherogram estimation based on gRNA region gave an efficiency of 36.1%, while TIDE analysis gave only 21.1% and a statistically insignificant indel spectrum (p>0.001) ( Table 2 , Fig 4C1 ), although direct electropherogram showed abundant sequence aberration in Gli1-gRNA4 region ( Fig 3C ).
10.1371/journal.pone.0228910.g004
Fig 4
TIDE (tracking of indels by decomposition) analyses showing the indel spectra and aberrant sequences of CRISPR/Cas9-gRNA-transduced cell populations versus untreated parental cell populations.
The graphs on the left analyzed indel frequencies within ±10 bp from theoretical gRNA breakpoints. The graphs on the right depicted PCR sequence aberrations; theoretical gRNA cuts were indicated by blue lines. A1 & A2. The RhoA gRNA5 region in AGS cells. Total editing efficiency = 15.6%. B1 & B2. The RhoA gRNA5 region in GT5 cells. Total editing efficiency = 35.7%. C1 & C2. The Gli1 gRNA4 region in GT5 cells. Total editing efficiency = 21.1%. D1 & D2. The Gal3 gRNA1 region in AGS cells. Total editing efficiency = 27.0%. E1 & E2. The Gal3 gRNA1 region in GT5 cells. Total editing efficiency = 41.5%. Eff, efficiency.
10.1371/journal.pone.0228910.t002
Table 2 Comparison of the editing efficiencies by direct electropherogram estimation, and TIDE and ICE analyses.
Gene targets and cell lines
RhoA
Gli1
Gal3
Method of assessment
AGS
GT5
GT5
AGS
GT5
Direct estimation (%)
12.5
28.5
36.1
26.7
50
TIDE analysis (%)
15.6
35.7
21.1
27.0
41.5
ICE analysis KO score
16
40
17
25
42
ICE analysis ( ice . synthego . com ) [ 11 ] provides an alternative means of analyzing CRISPR gene editing efficiency. ICE uses the same two electropherograms that TIDE uses: one from a Cas9-gRNA-treated cell population, and the other from an untreated control or parental cell population. ICE gives an ICE score (an indel percentage), a KO score (proportion of indels that indicate frameshifts), and an r 2 regression showing the degree of alignment between the treated and control (parental) cell populations ( Fig 5 ). Our r 2 regression results showed that RhoA-gRNA5 in the AGS and GT5 cell lines, Gli1-gRNA4 in the GT5 cell line, and Gal3-gRNA1 in the AGS and GT5 cell lines all achieved r 2 scores of 0.95 to 0.99. The KO scores ranged from 16 to 42 and showed similar trends to those seen in the TIDE analyses of the same sequences on a scale of 1 to 100, with 1 being the least efficient and 100 being the most efficient ( Fig 5A–5E ). ICE analysis gave the GT5 cell line’s Gli1-gRNA4 pool a KO score of 17; these results were similar to those from the TIDE analysis. In the ICE analysis, bad readings on two ends of the Sanger sequences were more pronounced in its discordance graphs than in TIDE analysis, because the latter could adjust parameters to avoid this phenomenon. For the indel percentage, both TIDE and ICE gave very similar patterns and percentage readings for the three genes and two cell lines and thus corroborated each other.
10.1371/journal.pone.0228910.g005
Fig 5
ICE (Inference of CRISPR edits) analyses showing indel and discordance plots.
A. The RhoA gRNA5 region in AGS cells showing a knockout (KO) score of 16. B. The RhoA gRNA5 region in GT5 cells showing a KO score of 40. C. The Gli1 gRNA4 region in GT5 cells showing a KO score of 17. D. The Gal3 gRNA1 region in AGS cells showing a KO score of 25. E. The Gal3 gRNA1 region in GT5 cells showing a KO score of 42.
Direct electropherogram estimation assesses ratios of aberrant nucleotides to wildtype nucleotides near gRNAs binding sites from Cas9 cutting sites; this method encompasses short reads that are generally 65–70 nucleotides in lengths. In contrast, TIDE and ICE analyses use longer Sanger electropherograms, normally 300–450 bp. This explains why, in comparing direct electropherogram estimation results to TIDE and ICE analyses results for Gli1-gRNA4 in the GT5 cell line, direct electropherogram estimation showed a higher editing efficiency, while TIDE and ICE analyses showed lower efficiency scores. Because Gli1-gRNA4 in the GT5 cell line altered a region equivalent to 48 bp (16 amino acids) in the parental wildtype GT5 cells ( Fig 3C , [between the red and black arrows] and Fig 4C2 ), which could well harbor indels that could cause a frameshift and thereby disrupt Gli1 expression.
Edit deconvolution by inference of traces in R (EditR) software ( baseEditR.com ) [ 15 ] which assesses gene editing using Cas9-Cytidine deaminase fusion enzymes, mainly focuses on gene editing with single-base resolution and without the need for double-stranded break induction. Because the latter is the cellular mechanism on which gene editing or gene knockout is relied upon and intended, therefore, EditR was not included in our editing efficiency comparison.
TIDE analysis uses mostly aberrant sequences of the Cas9-gRNA population to give total efficiency calculations, whereas ICE uses discordance of the Cas9-gRNA population in a similar fashion to give KO scores. In comparison, we propose that direct estimations of Sanger electropherograms surrounding gRNA regions, although simple, are more straightforward and effective for estimating how much of a pool contains Cas9-gRNA-edited or disrupted cells. Moreover, direct electropherogram estimation uses only Cas9-gRNA transduced pools, without the need to sequence control or parental cell lines, because wildtype or wildtype-like cells always exist in substantial proportions in those pools.
Gene rescue into KO pools requires cDNA modification
Our study of the RhoA gene was originally intended for reintroducing a clinically prevalent mutant Y42C into the gastric cancer (GC) cell lines AGS and GT5. The Catalogue of Somatic Mutations in Cancer (COSMIC) [ 16 ] shows that, of 1854 primary stomach cancer samples, 26 samples had RhoA genes with Y42C and Y42S mutations, yielding a prevalent mutation rate of 1.40%, the highest point mutations in the RhoA gene.
One difficulty of studying the Y42C mutant is that RhoA is highly expressed in all of the tested GC cell lines ( Fig 2A ), but sequencing (unpublished data) showed that these GC cell lines do not have a Y42C point mutation. When the plasmids N-terminal Flag-tagged p3xFlag-RhoA.Wt and p3xFlag-RhoA.Y42C were transfected into RhoA KOmixes of AGS and GT5 cells, we did not see RhoA expression detectable by Western blots ( Fig 6A ), although Flag tag was strongly expressed. Because Flag was fused with RhoA in its N-terminal, this finding strongly suggested that RhoA cDNA was edited by integrated Cas9, rendering RhoA dysfunctional, as undetected by Western blot, although p3xFlag-RhoA.Y42C already presented a nucleotide change in the gRNA5 binding site ( Fig 6D ).
10.1371/journal.pone.0228910.g006
Fig 6
RhoA wildtype, RhoA.Y42C and Gal3 gene rescue and cDNA modifications.
A & B. Western blots of RhoA, Flag, and β-actin showed before (A) and after (B) three nonconsecutive nucleotide modifications in the gRNA5 binding site. C. Western blots of RhoA, Flag, and β-actin showing RhoA rescue after DNA modifications in AGS and GT5 cells with Flag being at C-terminal. D. Alignment of rescue plasmids p3xFlag-RhoA.dgWt and p3xFlag-RhoA.dgY42C at the gRNA5 binding site was shown aligned alongside the original RhoA genomic DNA sequence. RhoA gRNA5 was shown at the top, and the amino acid translation was shown at the bottom. Note that wildtype is highlighted in pink for codon Y(TAT), while mutant is Y42C is highlighted in green for codon C(TGT). Three nonconsecutive nucleotides modified are in blue. E. Western blot of Gal3 in AGS cells showed the effect of overexpressing p3xFlag-Gal3.Wt and nucleotide modified p3xFlag-Gal3.dg1, dg2 and dg3. F. Alignment of the rescue plasmids p3xFlag-Gal3.Wt and p3xFlag-Gal3.dg1, dg2, and dg3 with, respectively, 1, 2, and 3 nonconsecutive nucleotides modified from the gRNA5 binding site. The amino acid translations are shown at the bottom. Note that the changed nucleotides are in green. KO, knockout.
Re-overexpressing N-terminal Flag-tagged p3xFlag-RhoA.dgWt and p3xFlag-RhoA.dgY42C with an extra three nonconsecutive nucleotides modified in the gRNA5 binding site ( Fig 6D ) into the same RhoA KOmixes of AGS and GT5 cells brought back expression of both Flag and RhoA ( Fig 6B ), which strongly suggested that Cas9 was still cleaving any unmodified, reintroduced RhoA wildtype and Y42C mutant; therefore modifying cDNA sequence within gRNA binding sites becomes a requirement ( Fig 6B, 6C and 6E ). Native RhoA protein is estimated to be 22 kDa in size; the addition of a 3xFlag of 81 nt (27 amino acids) added 27 x 110 Da, or 2970 Da, to the RhoA protein, making it an estimated 25 kDa in size.
To address why Flag tag was expressed but RhoA was not in AGS and GT5 RhoA KOmixes when reintroduced RhoA cDNAs were not modified, RhoA cDNAs with C-terminal Flag, p3.1-cFlag-RhoA.Wt, -RhoA.Y42C, -RhoA.dgWt and -RhoA.dgY42C, were constructed without and with three nonconsecutive nucleotides modification in the gRNA5 binding region. Western blots showed that both RhoA and Flag expressed strongly in transiently transfected RhoA KOmixes of AGS and GT5 only with three nonconsecutive nucleotides modification in gRNA5 binding region ( Fig 6C ). This finding strongly corroborated our assumption that Cas9 was editing reintroduced RhoA cDNA if gRNA binding region was not modified. The remnant bands of Flag and RhoA in AGS by unmodified RhoA probably indicated that introduced RhoA cDNA copies were more than integrated Cas9’s editing capability. Therefore to highly overexpress RhoA protein, three-nonconsecutive modification of RhoA in gRNA5 binding region is required.
Rescue of wildtype Gal3 expression by transfecting p3xFlag-Gal3.Wt into the AGS Gal3-KOmix cell pool also yielded dysfunctional expression undetectable by Western blot ( Fig 6E ). Thus, p3xFlag-Gal3.dg1, Gal3.dg2, and Gal3.dg3 with 1, 2, and 3 nonconsecutive nucleotides modified at the gRNA5 binding site ( Fig 6F ) were transfected into the Gal3-KOmix pool. Western blot showed that all produced the expected 35 kDa Gal3 protein and Flag tag ( Fig 6E ). However, the Western bands of Gal3 were markedly weaker than those for the parental cell population, perhaps because the inserted 3xFlag tag in the N-terminal of the Gal3 altering peptide epitope structure caused reduced sensitivity to our Gal3 antibody. Thus, in the case of the Gal3 gene, CRISPR/Cas9 is very stringent in its on-target editing and avoids off-target editing. Combining this finding with that for the RhoA gene, 3-nucleotide modification in the middle of gRNA binding sites seemingly would ensure the successful rescue of wildtype and mutant gene expression by exogenous introduction.
Off-target editing by CRISPR/Cas9 observed in RhoA gene but not in Gal3 gene
Off-target genome editing refers to nonspecific and unintended genetic modifications, which consist of unintended point mutations, indels, inversions, and translocations [ 17 – 21 ]. In this study, CRISPR/Cas9 editing on DNAs other than gRNA-matched sequences is considered off-target editing. In the Gal3 gene, CRISPR/Cas9 was very stringent in on-target editing around the gRNA1 binding site, as observed by short DNA sequencing in which only one nucleotide modification of gRNA1 binding site prevented Gal3 from being edited ( Fig 6E and 6F ), while straight transfection of wildtype p3xFlag-Gal3.Wt without any cDNA modification led to detection of neither Gal3 nor Flag tag expression by Western blots ( Fig 6E ). In comparison with the rescue plasmid RhoA.Wt and the mutant plasmid RhoA.Y42C, Y42C already presents a nucleotide change in the binding site of RhoA gRNA5, yet, as shown in Fig 6A , reintroduced p3xFlag-RhoA.Y42C did not result in detection of RhoA expression in transiently expressed pool, suggesting an evidence of CRISPR/Cas9 off-target editing, an obvious disadvantage of the CRISPR/Cas9 system.
It has been reported that high frequency off-target activity by presented Cas9-gRNA could rise to >50%, which will be a big concern [ 22 ]. It was identified that an active Cas9 can tolerate mismatches of gRNAs with targets harboring up to five mismatches, which has important implications in research and therapeutic applications [ 23 ]. It has also been shown that 15 off-target sites from 27 different single guide RNAs (sgRNAs), each harboring a single-base bulge and one to three mismatches to the guide strand, showed a significant variety of off-target sites cleaved by Cas9 [ 20 ]. But it was argued that by ChIP-seq that inactivated Cas9 binds to many sites of the genome, but activated Cas9 rarely cleaves off-target sites without matching gRNAs [ 1 ]. To reduce off-target editing, an optimized U6:3 RNA promoter produces the most potent effects of germline and somatic genome engineering in Drosophila in combination with homologous direct repair (HDR) and offsets nicking-based mutagenesis [ 24 ]. Another finding is that truncated gRNAs, shorter than 20 nucleotides in length at the far end of the PAM sequences, can decrease undesired mutagenesis at some off-target sites by ≥5,000-fold without sacrificing on-target genome-editing efficiencies [ 25 ]. However, in comparison with the major zinc finger nuclease (ZFN) and transcription activator-like effector nuclease (TALEN) gene editing methods, and CRISPR/Cas9 produces the least off-target editing [ 26 ]. Elevated Cas9 expression is among those factors affecting off-target effects in the CRISPR/Cas9 method. Therefore, strategies to control Cas9 activation, such as Tet-inducible Cas9, 4-hydroxytamoxifen (4-HT)-inducible Cas9 constructed by fusing a hormone-binding domain of the estrogen receptor (ERT2) to Cas9, or light-activated Cas9 constructed by fusing a light-responsive element to Cas9, have been applied, but these have limitations and drawbacks, such as adding significant complexity for research laboratories. Other simple but efficient approaches, such as using a Cas9 self-targeting system or substituting four amino acids in Cas9, rendering it a “high-fidelity” Cas9 nuclease, are among the best strategies for overcoming off-target effects. SpCas9-HF1, a high-fidelity variant, rendered all or nearly all off-target events undetectable by genome-wide break capture and targeted sequencing methods, thus providing an alternative to wildtype SpCas9 for research and therapeutic applications [ 27 ]. By using structure-guided engineering to improve the specificity of Streptococcus pyogenes Cas9 (SpCas9), a specificity-enhanced variant, eSpCas9, was developed. This variant displayed reduced off-target cleavage while maintaining robust on-target activity; thus, it could be broadly useful for genome-editing applications requiring a high level of specificity [ 28 ]. By using single-molecule Förster resonance energy transfer experiments, a non-catalytic domain REC3 within Cas9 was found and a new hyper-accurate Cas9 variant, HypaCas9, was developed which demonstrated high genome-wide specificity without compromising on-target activity in human cells [ 29 ]. In Cas9, a single point mutation (p.R691A) was identified that HiFi-Cas9-R691A reduces global off-target editing while maintaining high on-target activity [ 30 ]. Compared to the rationally designed eSpCas9 [ 28 ], SpCas9-HF1 [ 27 ], and HypaCas9 [ 29 ], HiFi-Cas9-R691A demonstrated the clinical utility of HiFi-Cas9 for therapeutic genome-editing applications. An engineered SaCas9-HF Cas9 variant from Staphylococcus aureus showed high genome-wide targeting accuracy without compromising on-target efficiency [ 31 ].
Our finding of a discrepancy of off-target editing in the RhoA.Y42C-gRNA5 region but not in the Gal3-gRNA1 region might be attributed to the positioning of the one nucleotide mismatch on the gRNAs; the one-nucleotide change in RhoA.Y42C was on the far end of the gRNA5 PAM sequence, whereas the one-nucleotide change in Gal3.Wt.dg1 was closer to the PAM sequence. According to literature, single and double mismatches are tolerated to varying degrees, depending on their positions along the gRNA-DNA interface [ 23 ] [ 25 ]. Even a target sequence truncated by one or two nucleotides at the very far end of the gRNA PAM sequence was still edited. The closer to the PAM sequence a mismatch is, the less likely it is to be edited. Therefore, raising stringency and specificity, eliminating off-target editing by CRISPR/Cas9 technologies through molecular engineering, and/or discovering novel Cas9-like genes from eubacteria or archaea, such as a high specificity FnCas9 from the bacterium Francisella novicida [ 32 ], will be the focus for future gene editing research before CRISPR technology could be adopted for human genetic and medical applications.
Conclusions
Our V2mO lentiviral vector successfully expressed mOrange in-frame with Cas9 and puromycin cDNAs, which added visualization of viral production and estimation of titers, as well as the capability of cell sorting Cas9+ cells in target cells by FACS, while the other components of the original pLentiCrispr-V2 were maintained without any alteration. Our results indicate that generating high titers of viruses is the first step of successful gene editing or gene knockout. When V2mO was made into lentiviruses, with proper gRNAs, i.e. RhoA-gRNA5, Gli1-gRNA4, and Gal3-gRNA1, they sufficiently knocked out RhoA, Gli1, and Gal3 genes in GC cell lines AGS and GT5 as visualized by short PCR electropherograms around gRNA binding regions and detected by Western blots. Gene editing efficiencies or knockout ratios of target cell pools could be estimated by direct Sanger electropherograms from Cas9-gRNAs transduced cell populations without the sequences of control or parental populations. Analysis software, such as TIDE and ICE, was shown to provide very informative data, including gene editing efficiencies, indel frequencies, and aberrant or discordant sequences against sequences of control or parental populations. Western blots also verified RhoA and Gal3 knockout in both AGS and GT5 cell lines. Single cloning of knocked-out cell pools must be performed to establish stable knockout clones; otherwise, pools of transduced cell lines will be gradually overgrown and dominated by wildtype or wildtype-like cells. This study also proved that rescue and re-overexpression of wildtype and mutant genes into knockout pools require cDNA modification by an extra three nonconsecutive nucleotide changes in the gRNA binding sites without alteration of amino acids, with the changes preferably in the midst of the gRNA binding sites or closer to PAM sequences. We have shown that cDNAs reintroduced into knockout populations without sequence modification in the gRNA binding regions will be edited by Cas9. Cas9’s stringent on-target effect was observed in the Gal3 gene, but off-target editing was observed in the RhoA gene because the RhoA.Y42C mutant already presented a nucleotide change in the gRNA5 binding site. Although we successfully rescued and re-overexpressed RhoA.Wt, RhoA.Y42C, and Gal3.Wt back into knockout cell mixes, our findings and other researchers’ publications suggest that, while CRISPR/Cas9 is a powerful method of gene editing, but that off-target or mismatched editing of target sequences is a big concern, particularly for critical medical or clinical applications.
Supporting information
S1 Fig
Electropherograms of single clones of RhoA knockout (KO) from the AGS cell line aligned with wildtype sequence.
For each electropherogram, the wildtype (WT) sequence is aligned at the bottom along with gRNA sequence. In AGS cell line, RhoA clones KO2 and KO9 showed genomic sequences in the vicinity of gRNA5 region. Clone KO9 had an extra A inserted at the 16 th /17 th nt, causing frameshift. Clone KO2 was not a single clone, but a mixed one, though Western blot showed it was truly RhoA KO (see Fig 2C ).
(DOCX)
S2 Fig
Electropherograms of single clones of RhoA knockout (KO) from the GT5 cell line aligned with wildtype sequence.
For each electropherogram, the wildtype (WT) sequence is aligned at the bottom along with gRNA sequence. In cell line GT5, four RhoA clones KO1, 3, 6 and 8 showed genomic sequences in the vicinity of gRNA5 region. Various alterations were observed at gRNA5 binding region. Clone KO1 had an A missing at the 17 th nt of gRNA5, causing frameshift. Clone 3 was not a single clone, but a mixed one, possibly comprised of 2 clones, though Western blot showed it was truly RhoA KO (see Fig 2D ). Clones KO6 and KO8 had 12 nt and 26 nt deleted at the 16 th or -2 nd nt of gRNA5 binding regions, respectively. Western blot showed both were truly RhoA KO (see Fig 2D ).
(DOCX)
S3 Fig
Electropherograms of single clones of Gal3 knockout (KO) from the AGS cell line aligned with wildtype sequence.
For each electropherogram, the wildtype (WT) sequence is aligned at the bottom along with gRNA sequence. In AGS cell line, Gal3 clones KO1, 3, 5, and 7 showed genomic sequences in the vicinity of gRNA1 region. Clones KO3 and KO5 had an extra T inserted at the 17 th nt, causing frameshift. Clones KO1 and KO7 were not single clones, but mixed ones, though Western blot showed that they were truly RhoA KOs (see Fig 2F ).
(DOCX)
S4 Fig
Electropherograms of single clones of Gal3 knockout (KO) from the GT5 cell line aligned with wildtype sequence.
For each electropherogram, the wildtype (WT) sequence is aligned at the bottom along with gRNA sequence. In cell line GT5, Gal3 clones KO3, 9, and 23 showed genomic sequences in the vicinity of gRNA1 region. Clone KO23 had an extra T/C inserted at the 17 th nt, causing frameshift. Clones KO3 and KO9 were not single clone, but a mixed one, aberration began at 18 th or 19 nt of gRNA binding sequence, though Western blot showed it was truly RhoA KO (see Fig 2G ).
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Vocal learning – the encoding and production of acoustic traits acquired from conspecific or heterospecific tutors – is restricted to very few groups of birds and mammals, and understanding its origins remains a core aim of evolutionary biology [1] , [2] . One of the primary study systems for exploring the adaptive significance of vocal learning is birdsong, yet the distribution of vocal learning in birds is contentious. The prevailing view is that song learning is restricted to three major clades, the oscine passerines (Passeriformes), parrots (Psittaciformes), and hummingbirds (Trochiliformes). In contrast, it is often assumed that songs develop without learning from tutor birds in the suboscines, a diverse clade of passerines containing ∼1150 species. However, the finding that parrots are the sister group to both clades of passerines, the oscines and suboscines [3] , [4] , has cast some doubt on whether suboscines develop songs by learning or not [2] , [5] , [6] . In addition, recent evidence that vocal learning occurs in some species of suboscine suggests that they offer a better system for understanding the origins of this trait than do oscines and parrots, wherein learning is virtually universal.
Previous experimental tests of vocal learning in suboscines have shown that Willow ( Empidonax traillii ) and Alder ( Empidonax alnorum ) flycatchers raised with heterospecific tape tutors [7] , and deafened Eastern Phoebes ( Sayornis phoebe ) [8] , all produce normal adult song. These species also lack the forebrain cell clusters that control song acquisition in oscines [8] , although recent studies have shown that at least one of these species, S. phoebe , has an incipient or vestigial homologue of these cell clusters [2] . It was often inferred on the basis of these studies that suboscines were unable to learn their songs, but this paradigm has recently been challenged. For example, several lines of evidence suggest that Procnias bellbirds develop songs with learning [5] , [9] : (1) mismatch between genetic variation and geographic variation in male songs of Procnias tricarunculata ; (2) intricate changes in song over time within individuals; (3) the production of heterospecific song of a cage mate by a captive Procnias nudicollis ; and (4) the production of bilingual, rather than hybrid dialects. Vocal learning has also been proposed for the long-tailed manakin Chiroxiphia linearis [10] , [11] and screaming piha Lipaugus vociferans [12] , although in both cases the effects of learning appear to be weak, and further supporting evidence is required [13] , [14] .
These studies have led to growing claims that vocal learning may be widespread in suboscine passerines, although all research so far has focused on tyrant-flycatchers (Tyrannidae), manakins (Pipridae) and cotingas (Cotingidae) belonging to the infraorder Tyrannides, sometimes referred to as the non-tracheophone (or “bronchophone”) suboscines [15] . In contrast, almost no attention has focused on the infraorder Furnariides – i.e. tracheophone families such as the antbirds (Thamnophilidae) and ovenbirds (Furnariidae) – that make up about half of suboscine species (and approximately 5% of total global bird diversity). Consequently, the widespread assumption that vocal learning is absent in the tracheophone clade [16] – [18] is based entirely on inference or anecdotal observations.
Direct experiments in tracheophones are clearly a priority, particularly as they may shed light on the role of social mechanisms, including sexual selection, in the evolution of vocal learning. Sexual selection has been proposed to explain variation in learning across oscine passerines, although evidence is contradictory and widespread support is lacking [19] – [23] . Some oscines subject to strong sexual selection have large song repertoires [24] and are often capable of sophisticated mimicry [25] , while others place a higher premium on vocal traits unrelated to learning, such as performance consistency [26] . The link between sexual selection and vocal learning may be easier to detect in suboscines, where all species proposed to learn songs have lek-based or polygamous reproductive strategies, and thus intense sexual selection. This includes cases of learning in bellbirds [5] , [9] , manakins [10] , [11] and pihas [12] . In contrast, the sexual selection hypothesis predicts that learning will be rare or absent in tracheophones as all members of this clade have long-term social bonds and thus apparent low levels of sexual selection [27] . However, experiments testing the mode of song development are still lacking in this major branch of the suboscine radiation.
To address this issue, we raised spotted antbirds ( Hylophylax naevioides ) by hand from the egg in soundproofed aviaries, exposing developing chicks to either silence (no tutor), or the song of a congeneric species from Amazonia, Hylophylax naevius (heterospecific tutor). This species was an ideal tutor as it is closely related to H. naevioides , with a song that is essentially similar in overall tone and pattern, but with a distinctly different cadence that can easily be detected in acoustic analyses. Our aims were two-fold: (1) to quantify the progression of song development in captive-reared birds from hatching to the production of adult song, and (2) to compare the structure of the adult songs of captive-reared birds in our two treatment groups with those of wild conspecifics and heterospecific tutors.
Materials and Methods
Ethical statement
This study was approved by and performed in accordance to The Smithsonian Tropical Research Institute Institutional Animal Care and Use Committee (IACUC # 2009-01-03-17-09). Additionally, field protocols were approved by and performed in accordance to the Autoridad Nacional del Ambiente (ANAM) of Panama (# SE/A-61-09).
Study species
H. naevioides is a tracheophone suboscine passerine found in Central America, north Colombia, and Ecuador. Monogamous pairs form long-term pairbonds where both sexes contribute equally to parental care and defence of stable year-round territories [28] , [29] . Both sexes sing a structurally similar stereotyped song ( Figure 1F, 1G ) [30] . Consistent with other antbirds [31] , [32] , these songs are individually recognizable [30] and given year-round with a slight peak in vocal activity in the breeding season, suggesting that they mediate competition over both mates and territories [33] .
10.1371/journal.pone.0095746.g001 Figure 1
Representative song spectrograms from individuals in each treatment group
. Broadband spectrograms show ( A, B ) captive H. naevioides reared in silence with no tutor, ( C–E ) captive H. naevioides reared with H. naevius tutor, ( F ) wild male H. naevioides , ( G ) wild female H. naevioides , ( H ) wild male H. naevius .
Collection and bird handling
We conducted collection efforts from April to June, 2009–2010, near Soberanía National Park, in the vicinity of Gamboa, Panamá (9°7′N, 79°40′W). As nestling passerine birds can potentially imprint on the songs of adults immediately after hatching [34] , we collected study individuals of H. naevioides at the egg stage. A total of 32 eggs were removed from 16 nests with complete clutches (2 eggs) in the early breeding season. This ensured that siblings could be divided among treatments (see below), and that adult pairs would readily re-nest, minimizing impacts on the population.
Within 1 h of collection, we placed eggs in a Grumbach Compact S84 incubator with automatic temperature and humidity set at 36.5°C and 70%, respectively, based on field data collected from active H. naevioides nests (Gustavo Londoño, unpublished data ). Eggs were automatically turned every hour throughout the duration of incubation. One day prior to hatching, we placed eggs in individual tissue-lined nest cups in a Brinsea TLC-4 brooder (temperature: 36°C; humidity: 70–80%). Over the course of 11 days – i.e. the mean time for chicks to fledge their nest in the wild [28] , [35] – we gradually lowered temperatures in brooders to 28°C (experimental room temperature).
Day 1 after hatching we hand-provisioned chicks with 60–75% of their body mass in lab-reared waxworms and crickets, augmenting this diet with small wild-caught katydids. At day 2, we syringe-provisioned chicks with FoNS (formula for nestling songbirds) gruel [36] , [37] and insects as above every 30–60 minutes depending on intake quantity. Feed rate was adjusted to ensure a daily intake of 60–75% of the chick's body mass up to day 8, rising to 50–60% of body mass at fledgling (approximately day 11). Chicks were weighed every morning (06:30 h) and evening (18:00 h), and following each feeding. Average weight gain and developmental rates were similar to those of wild nestlings [28] .
Upon fledging, we moved chicks to individually isolated 84 cm×46 cm×76 cm flight cages where they were provided ad libitum with a balanced adult diet (mashed soaked EVO cat food, boiled egg, bird vitamins, yoghurt, and live mealworms and crickets). We gradually weaned chicks off FoNS gruel until they demonstrated independent feeding of this mash from self-feed pans. We measured the quantity of food consumed by individuals to ensure a sufficient daily intake of 40–50% of the body mass to maintain a mass of 16–20 g. We changed water twice daily, dropping mats daily, and cleaned perches and cages weekly.
Health was assessed throughout the study by weekly faecal analysis and daily visual inspection for signs of illness (feather expansion, extended eyelid closure, and lethargy). Of 14 eggs collected in 2009, all successfully hatched, 5 fledged, 5 survived to maturity, and 5 completed the experiment in a healthy condition. In 2010, 17 of the 18 eggs successfully hatched, but none completed the experiment due to recurring problems with coccidial infection at the pre-fledgling developmental stage. Given the good health of the individuals exposed to experimental treatments, and their vigorous singing behaviour, we believe that our results are not affected by any challenging physical condition.
Experimental Procedure
At hatching, we randomly assigned individuals to one of two treatment groups: (i) no tutor (i.e., silence) or (ii) heterospecific tutor. Because treatments began seven days after hatching, at which point it is not possible to sex individuals accurately, we assigned treatment groups independently of sex. Of the five individuals reaching maturity, two were males in the no tutor group, and three were females in the heterospecific tutor group. Prior to fledging, we placed individuals in sound attenuation chambers only during their daily tutor sessions, whereas after fledging they were permanently housed individually in flight cages within sound attenuation chambers (see below) for the duration of the experiment.
Tutor sessions began daily at 07:00 h and lasted 1 h. The no-tutor group received silence; the heterospecific tutor group received playback of H. naevius song. To track song development through to the production of adult song, we recorded all individual birds daily from 10 minutes prior to – 1 h after tutor sessions. All digital sound files were 16 bit wav mono files recorded at a sampling frequency of 44.1 kHz. The experimental treatment concluded when individuals had produced a minimum of 10 high quality adult songs (between 83 and 165 days; see Fig. 2B ).
10.1371/journal.pone.0095746.g002 Figure 2
Development of song in Hylophylax naevioides .
( A ) Broadband spectrograms of vocalizations produced by two captive individuals demonstrating different stages towards the production of crystallized adult song. Upper panel, a male bird reared in isolation with no tutor (NT1): subsong at 51 and 59 days post hatching (dph); incipient song at 74 dph; crystallized song at 86 dph. Lower panel, a female bird reared in isolation with a heterospecific tutor (HT1): subsong at 46 and 68 dph; incipient song at 81 dph; crystallized song at 83 dph. ( B ) Ages (dph) at which the five stages of song production (babbling, subsong, incipient song, crystallized song, and soft song) were produced by two male birds reared in isolation with no tutor (NT1, NT2; left of the red line) and three female birds reared in isolation with a heterospecific tutor (HT1, HT2, HT3; right of the red line), until the end of the experiment for each individual (200 dph). See methods for description of song production developmental stages.
Experimental set-up
Sound attenuation chambers
We constructed sound attenuation chambers using a double wall design and open cell polyether polyurethane acoustic foam coated with a moisture and chemically resistant film (Soundcoat soundfoam M with uniseal). This design provides sound attenuation levels of approximately 30 dB. Chambers were ventilated with silicone tubing attached to an air pump and equipped with full-spectrum fluorescent light strips programmed to mimic diurnal cycles in the wild. To record vocalizations, chambers were equipped with cardioid condenser hanging microphones (Audio-Technica U853A) connected to external high-resolution Solid State wav recorders (Edirol R-09HR). Chambers in the heterospecific tutor treatment group were additionally equipped with mini-amp speakers (Radioshack). Chambers were located in two separate climate controlled experimental rooms, one for each treatment group. Treatment groups were split between experimental rooms to avoid the potential risk of individuals hearing vocalizations during routine husbandry.
Playback treatment preparation
We used RavenPro version 1.4 to prepare playback audio (wav) files for the heterospecific tutor group. Playback loops consisted of 10 H. naevius songs separated by 3 sec of silence, followed by 5 min of silence, similar to natural rates of singing in the wild. To avoid pseudoreplication, we used unique playback loops from recordings of songs from different individuals of H. naevius for each individual in the heterospecific tutor group.
Sampling songs of wild individuals
To compare the songs of captive-reared individuals with those of wild conspecifics ( H. naevioides ), we recorded 6 high-quality songs from both males (N = 15) and females (N = 17) in Soberanía National Park. Recordings were made between September 2009 and February 2012 between 06:30 h and 11:00 h. Songs were recorded on compact flash cards as 16-bit wav mono files at a sampling frequency of 44.1 kHz using a Sennheiser ME67-K3U directional microphone (Sennheiser, Hanover, Germany) and Marantz PMD-661 Solid State recorder (Marantz, Kanagawa, Japan). We solicited singing on territories using playback (prepared as above) of male or female songs.
We also compiled recordings of wild heterospecific ( H. naevius ) individuals, with songs of 11 unsexed birds downloaded from a digital archive ( http://www.xeno-canto.org/ ; file accession numbers: XC63522, XC63523, XC90280, XC44136, XC74917, XC39754, XC33214, XC2993, XC90289, XC3724, XC98060) and a further individual extracted from a commercial CD (N = 1; [38] ).
Acoustic analysis
Using RavenPro version 1.4, we produced broadband spectrograms (bandwidth = 61.9 Hz, Hann window size = 1024) of songs of all study individuals, to compare songs produced by captive-reared birds with those of wild H. naevioides and H. naevius . Assessment was conducted qualitatively by visual inspection, and quantitatively by extracting acoustic parameters by hand from spectrograms using on-screen cursors (see Fig. 3 ). This technique is sensitive to errors generated by small changes in the exact placement of on-screen selections, and environmental background noise. We minimized such errors using RavenPro to generate a set of robust statistical estimators for each selection. This approach reduces the effect of outliers by using measures of central tendency and dispersion such as order statistics, the median, interquartile range, and quartile skewness (see [39] ). We used 4 spectral and 4 temporal robust estimators to quantify a total of 19 acoustic parameters for each song (see Table 1 ). By nature of having multiple elements within songs, long note and short note song parameters per song were generated by averaging robust estimators taken for each long note and short note within an individual song, respectively, prior to further analysis. This step was unnecessary for the other acoustic parameters that only had one element per song (i.e., entire song, first half of song, second half of song).
10.1371/journal.pone.0095746.g003 Figure 3
Acoustic analysis of Hylophylax naevioides song.
Broadband spectrogram illustrates an example of one song produced by a wild adult male. Boxes denote the acoustic selections used in this study to calculate acoustic parameters using robust statistical estimators in Raven 1.4. Parameters were calculated as averages across five different subsets of notes separately: the full song, the first half of the song by the nearest note to the middle time, the second half of the song by the nearest note to the middle time, the long notes and the short notes (see Table 1 ).
10.1371/journal.pone.0095746.t001 Table 1
Description of acoustic parameters extracted from songs (see Figure 3 ).
Robust acoustic parameter
Definition
Specific measurement *
Centre Frequency (Hz)
Frequency that divides the selection into two frequency intervals of equal energy †
1. Entire song
2. 1st half of song
3. 2nd half of song
4. Long note
5. Short note
6. Difference between 1st and 2nd half of song
1 st Quartile Frequency (Hz)
Frequency dividing the selection into two frequency intervals containing 25% and 75% of the energy †
7. Long note
8. Short note
3 rd Quartile Frequency (Hz)
Frequency dividing the selection into two frequency intervals containing 75% and 25% of the energy †
9. Long note
10. Short note
Inter-quartile Range Bandwidth (Hz)
The difference between the 1 st and 3 rd Quartile Frequencies
11. Long note
12. Short note
Centre Time (sec)
The duration the selection is divided into two time intervals of equal energy †
13. Long note
14. Short note
1 st Quartile Time (sec)
The duration that divides the selection into two time intervals containing 25% and 75% of the energy †
15. Long note
16. Short note
3 rd Quartile Time (sec)
The duration that divides the selection into two time intervals containing 75% and 25% of energy †
17. Short note
Inter-quartile Range Duration (sec)
Difference between the 1 st and 3 rd Quartile time
18. Long note
*For all parameters except 1–3, mean was calculated for notes within a song. † Power values in short-time spectra and frequency bands that compose the spectrogram are summed to generate aggregate power envelopes and spectra, resulting in a power versus time envelope and power versus frequency spectrum, respectively. The aggregates are normalized and treated as probability density functions with time or frequency being the variate, and density the fraction of the total signal energy. From the distribution function, various measures of central tendency and dispersion are then used to characterize the signal energy distribution in time and frequency. (See [39] ).
Statistical analysis
To compare songs produced by captive-reared individuals with those of wild birds, we first calculated mean individual values for each song acoustic parameter. We then performed a series of rotated principal components analyses (PCAs) with Kaiser normalization on the correlation matrix of mean values of song parameters to reduce the dimensionality of our dataset and to avoid multicollinearity. Four PCAs were conducted on four different subsets of the dataset: (1) all H. naevioides individuals, (2) only male H. naevioides , (3) only female H. naevioides , and (4) both H. naevioides and H. naevius (sexes pooled). In each case, PCA extracted three PC scores with eigenvalues >1, which accounted for >85% of variance in the acoustic datasets ( Table S1 ). We then used PC scores (1) to test for differences in the structure of male and female songs, (2) to test for differences between songs produced by captive-reared and wild birds, (3) for a Discriminant Function Analysis (DFA), (4) for a bootstrap test, and (5) to plot the structure of all songs in relation to one another.
We compared the structure of captive and wild antbird songs in three complementary ways. First, we used an ANOVA to compare PC scores between wild and captive songs. Because we found no significant difference in the structure of male and female H. naevioides songs as described by all three PCs ( Table 2 ), we pooled song data from the sexes for this analysis. Second, we used DFA with cross-validation using the lda function in the MASS package [40] . We ran the DFA on PC scores calculated as above with the addition of 2 PC scores to account for 95% percent of the total variance in the acoustic datasets. We then calculated the proportion of captive-reared treatment individuals that were grouped with the wild birds, heterospecific tutor birds, or as captive-reared birds based on the PC scores generated from the acoustic song datasets. Third, we determined whether the overall structure of the songs produced by captive-reared birds (as defined by PC1, Table S1 ) fell within the natural range of acoustic variation we sampled in wild birds. To do this, we randomly selected data for a single individual in the wild-song dataset, extracted the PC1 score, and repeated (allowing the same individual to be selected more than once) until we had generated the same number of data points as the total number of wild individuals sampled (N = 32). We repeated these steps 10,000 times, then plotted means for each bootstrap replicate, and finally assessed whether the mean PC1 scores for captive-reared birds fell within the null distribution of the PC1 scores for the wild birds. Bootstrap tests were conducted on each treatment (and hence sex) separately.
10.1371/journal.pone.0095746.t002 Table 2
Effect of captive rearing on song structure in Hylophylax naevioides , where song is defined by PC1, PC2 and PC3 (mean ± SD).
Variable
Captive-reared
Wild song
X 2 *
P
PC1
−0.19±2.57
0.03±3.25
0.01
0.93
PC2
−1.03±1.60
0.16±2.44
1.14
0.29
PC3
−0.59±0.85
0.09±1.04
2.28
0.13
*Statistics derive from a Kruskal Wallis test; N 1 = 5 captive-reared birds, N 2 = 32 wild birds (sexes pooled).
Results
Song development
The five captive-reared individuals in this study all followed the same progression in vocalization types towards the production of crystallized adult song, although the timing of each stage varied ( Fig. 2A, 2B ). Three individuals varied only marginally in the timing of the progression to crystallized adult song, between 79 and 83 days post hatching (dph), but this progression was longer for one individual in the no tutor group (106 dph) and one individual in the heterospecific tutor group (138 dph) ( Fig. 2 ). All individuals began ‘babbling’ as their first step in song production between 13 and 47 dph ( Fig. 2B ). Babbling appeared to be a modified version of the contact call produced by all individuals immediately upon fledging. It took from 1 to 30 days to progress from babbling to the second step, which we term ‘subsong’ ( Fig. 2B ). Subsong comprised actual components and groups of notes of a complete adult song. In all but one individual (HT3, in the heterospecific tutor group), initial subsong was similar to the middle section of adult song, but without any inflection in pitch. HT3 took longest to progress from the babbling phase, and was unique in including the introductory section of adult song into her subsong. During the subsong phase, individuals either added more notes to sections or produced different parts of adult song – either the introductory, middle, or terminal section – until they were able to produce an ‘incipient adult song’. The delay between the start and end of the subsong phase varied from 1 to 57 days. Incipient adult songs were often missing a rise and fall in pitch and notes lacked the clarity of those produced during ‘crystallized adult song’. All individuals required some time in the incipient adult song phase (15–43 days) before they were able to produce crystallized adult song. Following the production of crystallized song, individuals continued to produce sub and incipient song for a further 3–21 days until they could produce consistent crystallized adult song ( Fig. 2B ).
Comparison of adult songs in captive and wild birds
Visual inspection of spectrograms of crystallized adult songs produced by captive-reared H. naevioides showed them to be extremely similar in spectral structure and temporal patterning to those produced by wild conspecific adults ( Fig. 1 ). Accordingly, when we plotted the structure of the songs of all individuals in our sample in the acoustic space defined by PC1, PC2 and PC3, we found that captive-reared birds produced songs that fell within the area of acoustic space occupied by songs produced by wild H. naevioides ( Fig. 4 ). Specifically, the songs of the two individuals raised in silence with no tutor fell within the area occupied by wild conspecific adults, as did the songs produced by the three individuals raised in the heterospecific tutor group. These qualitative appraisals of the similarity of songs of captive individuals to the songs of wild conspecifics were corroborated quantitatively in three ways.
10.1371/journal.pone.0095746.g004 Figure 4
Comparison of the structure of the songs of experimental versus wild individuals.
Songs by male (closed symbols) and female (open symbols) wild (circles) and captive-reared (red triangles) H. naevioides grouped together while those produced by H. naevius (closed diamonds) grouped separately. Plot produced according to three principal components generated from acoustic data extracted from spectrograms; PC3 is represented by depth and is not labelled (see Table S1 for factor loadings).
First, we found that there was no significant difference in the structure of crystallized adult songs produced by captive-reared and wild H. naevioides as defined by the first three principal components (PC) extracted from song ( Table 2 ). Second, none of the adult songs produced by captive-reared females and males grouped separately from those produced by wild birds in the DFA. In contrast, all adult songs produced by both captive-reared and wild H. naevioides grouped separately from those of H. naevius in the DFA ( Table 3 ). Finally, the result of the bootstrap test revealed that the mean structure of captive-reared songs (as defined by PC1) did not differ significantly from that of wild songs as determined by the sampling distribution extracted from observed data on wild song structure for both males ( Fig. 5A ) and females ( Fig. 5B ; P >0.1 in all cases). Although our experimental sample (N = 5) is small, we note that our wild population is far better sampled (N = 32), and thus our bootstrapping procedure is able to assess the statistical significance of a relatively small number of observations, in line with previous studies (e.g. [41] ).
10.1371/journal.pone.0095746.g005 Figure 5
Null distributions showing range of acoustic variation in wild birds.
Arrows indicate where the songs of captive-reared individuals fall within the sampling distribution generated from the songs of wild males ( A ) and females ( B ), where songs are described by PC1 ( Table S1 ). Null distributions were generated with 10,000 bootstrap replicates.
10.1371/journal.pone.0095746.t003 Table 3
Discriminant function analysis grouping (proportion) of song profiles by captive-reared no tutor and heterospecific treatment individuals with respect to song profiles by wild conspecific H. navioides and heterospecific wild H. naevius .
No Tutor
Heterospecific Tutor
Wild H. naevioides
Wild H. naevius
Total correct
0
0
1
1
0.9
Wilks lambda
F
Df num
Df den
P
0.11
9.56
15
114
<0.0001
Our analyses reveal a high degree of underlying structural similarity between adult captive-reared and wild H. naevioides , but we note that one individual in the heterospecific tutor group developed songs with apparent anomalous peaks in the second note of each syllable (see Fig. 1E ). Given that this structural feature is reminiscent of the songs of wild H. naevius ( Fig. 1H ), we cannot rule out the possibility that subtle or marginal vocal learning occurs in at least some Hylophylax individuals.
Discussion
We have shown that H. naevioides individuals raised in captivity under experimental conditions produced crystallized adult songs that were statistically indistinguishable from those produced by wild adults. The same result was obtained for individuals raised with no tutor as those with a heterospecifc tutor. These results indicate that normal adult songs of H. naevioides develop in the absence of a conspecific tutor, thus providing the first experimental evidence that tracheophone suboscines can develop songs without learning. By this we mean that songs appear to develop without the need to encode tutor song, and we do not rule out other forms of learning, such as honing the use of vocal motor-control systems (i.e. sensorimotor learning) [42] . Given that study individuals could hear themselves singing, sensorimotor learning may have influenced the rate of progression to adult song in our experiments, but it cannot explain accurate song development in captive-reared individuals lacking conspecific tutors.
Our findings are consistent with a number of anecdotal observations from tracheophones over recent decades. For example, a single captive barred antshrike ( Thamnophilus doliatus ) produced normal song after being reared in silence [7] , although it is not clear at what age it was taken into care. Similarly, one of the only known hybrid tracheophone suboscines was an antpitta ( Grallaria ) that produced a song structurally intermediate between the songs of its putative parent species [43] . This contrasts with the situation in hybrid oscines in which the structure of songs is typically unchanged from one or other of the parental song types, either because hybrid offspring copy songs from the parent male, or produce repertoires containing songs from both parental types (mixed singing) [44] – [46] . It has also been noted that there is little evidence of individual song variation, mimicry, repertoires or dialects in tracheophones, all suggesting an absence of vocal learning [30] , [47] , [48] .
It is plausible that vocal learning may be the ancestral state in passerines and their sister-group, the parrots. Song learning may theoretically have a single ancient origin at the root of this parrot–passerine clade, in which case non-learning suboscines have potentially lost the ability to learn songs during their evolutionary history. This may explain why even non-learning suboscines possess rudimentary substrates for learning [2] , which in turn may explain why we detect a subtle adjustment in the songs of one individual H. naevioides in the heterospecific tutor group ( Fig. 1E ). Regardless of whether suboscines are losing or evolving the ability to learn songs, the distribution of vocal learning in non-tracheophones appears to be restricted to lineages with polygamous or lek-breeding reproductive strategies, and absent from those with social monogamy, suggesting that sexual selection may promote vocal learning [5] . Our results support this idea by indicating a lack of learning in tracheophone suboscines, a group with long-term social bonds, apparently low levels of extra-pair copulations, and thus relatively weak sexual selection [49] , [50] . The positive association between sexual selection and vocal learning across suboscines suggests either that elevated sexual selection promotes the evolution of vocal learning because it allows the development of repertoires known to mediate female choice in birds, or else that reduced levels of sexual selection can lead to the loss of vocal learning over time.
It is worth noting that the ability to learn song is retained in many oscine systems with social monogamy, and that in these cases its function is not exclusively sexual. For example, vocal learning has been shown to facilitate territory defence within species through the matching of song repertoires by neighbours [51] , [52] and may also drive convergence between species when territorial signals mediate interspecific competition [53] . Although it is clear that learning can function in these agonistic contexts, such contexts do not evidently predict song learning in suboscines. So far, evidence for learning in suboscines is restricted to systems without territoriality (e.g. lek-breeding systems), whereas learning is absent in systems with intense territoriality within and between species [17] , [31] . Thus, the evidence from suboscines suggests that song learning is more likely to arise from mechanisms of sexual selection (e.g. mate choice and intrasexual competition for access to matings) than from non-sexual forms of social selection (e.g. inter- and intrasexual competition for resources such as food or territories).
Our demonstration of reduced or negligible vocal learning in a tracheophone suboscine supports the increasing focus on this clade as a study system for testing evolutionary theory [18] , [27] , [54] . Until recently, most studies of birdsong were focused on oscine songbirds, which are potentially unsuitable subjects for some key questions because of vocal learning. For example, habitat-related differences in oscine song are often proposed to result from acoustic adaptation to habitat features [55] – [57] whereas an alternative explanation is that song differences are entirely caused by young individuals learning the songs, or parts of songs, that they perceive most clearly in their natal habitat [55] . Relationships between oscine song structure and environmental variables may therefore lack a genetic basis, and be driven instead by phenotypic plasticity [58] . Likewise, the convergence of oscine song in contact zones has been attributed to convergent character displacement [59] , yet this could be explained by heterospecific copying, where one species accidentally learns the song, or song types, of heterospecifics within the contact zone [45] . Our results therefore add weight to the argument that studies focusing on tracheophone suboscine songs can overcome these problems because non-learning increases our confidence that song variation has a genetic basis [17] , [27] , [60] .
Captive rearing experiments also provide insights into song development after hatching in tracheophone suboscines. All five study individuals followed a similar ontogeny of song development, with an initial babbling stage followed by a subsong stage where various notes and segments of adult song were produced, leading eventually to the production of incipient song and, finally, crystallized adult song. This sequence is similar to that described in several species of oscine [61] . Progress to the first crystallized adult song was relatively rapid and consistent, taking 3–4 months in all but one individual ( Fig. 2 ). This period was longer than that reported in Empidonax alnorum , which can produce a rudimentary form of adult song immediately after fledging the nest [7] , and shorter than that reported for Sayornis phoebe , which has a prolonged period (7–8 months) of incipient song until crystallization is reached at the beginning of the breeding season [2] . Both these species are non-tracheophone suboscines with simple, innate songs. Production of the first crystallized song in H. naevioides do not appear to be timed to the breeding season, with all birds producing adult song at least 2 months prior to the typical nesting period of wild birds. This matches observations in Hypocnemis antbirds, where 3-month-old juveniles begin to sing with their parents to defend family territories outside the breeding season (J. Tobias and N. Seddon, unpublished data ).
We have shown that H. naevioides , a socially monogamous suboscine species with long-term pair bonds, is capable of developing normal adult songs when raised in silence or with heterospecific tutors. In all cases the songs of captive-reared individuals were not significantly different from the songs of wild adults, although we do show evidence that a small amount of imprinting on heterospecific tutors may occur. Taken together, these findings support the view that vocal learning is absent or negligible in the tracheophone suboscine clade, confirming its importance as a model system to address questions in ecology and evolution, including the nature of selection driving signal evolution and speciation. Our results also strengthen support for the hypothesis that the evolution or retention of vocal learning is associated with sexual selection in suboscines, perhaps providing clues for the origin of vocal learning mechanisms more generally.
Supporting Information
Table S1
Factor loadings of acoustic measures for four sets of Principal Components Analyses
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The PPAR-gamma coactivator (PGC) family of transcriptional co-activators have been called ‘master regulators’ of mitochondrial biogenesis because they co-activate the transcription factors nuclear respiratory factor-1 and -2 (NRF-1 ( Nrf1 ), NRF-2 ( Gabpa )), thereby initiating nuclear gene expression for mitochondrial proteins [1] , [2] . The known PGC family gene products (PGC-1α ( Ppargc1a ), PGC-1β ( Ppargc1b ), and PRC ( Pprc )) are variably affected by inflammation; for instance, Ppargc1a mRNA levels may decrease after exposure of cells to LPS, while PGC-1α protein content and stability are increased in mice exposed to LPS [3] , [4] , [5] .
Ppargc1a levels in mice infected with Staphylococcus aureus ( S. aureus ) have been shown to increase as mitochondrial damage develops under the stresses of excessive reactive oxygen and nitrogen species generation [6] , [7] . Mitochondrial damage in sepsis disrupts oxygen homeostasis, leading to a state of high tissue oxygenation but low tissue oxygen utilization sometimes called cytopathic hypoxia [8] , [9] . Increased mitochondrial damage and cytopathic hypoxia are correlated with a high mortality rate [10] . Sepsis is a growing health problem and a leading cause of death in the U.S., and despite aggressive intensive care, still has an in-hospital mortality rate of approximately 30% [11] .
The program of mitochondrial biogenesis is responsible for maintaining adequate mitochondrial mass and quality in the cell, and it is indispensable for energy homeostasis and cell viability during periods of cell stress that increase mitochondrial turnover. Mitochondrial biogenesis is a bi-genomic process that requires coordination between nuclear- and mitochondrial-encoded genes [12] . The cell activates mitochondrial biogenesis in response to sepsis, which helps to counteract the effects of mitochondrial damage and to maintain appropriate oxygen metabolism and ATP availability.
The signal transduction pathways that link sepsis-induced inflammation to the up-regulation of mitochondrial biogenesis are not well understood, but there is some evidence that the Toll-like receptor (TLR) family of innate immune receptors may provide such a link. The TLRs are transmembrane proteins that sense conserved pathogen-associated molecular patterns. When activated, the TLRs signal through various downstream kinases to mobilize pro-inflammatory transcription factors such as NF-κB, AP-1, and IRF3/7 [13] . TLR2 is activated by components of Gram-positive bacteria, while TLR4 is activated by components of Gram-negative bacteria [14] . It has been shown that TLR4 −/− mice, compared with WT mice, show less mitochondrial damage in a model of heat-killed E. coli sepsis, but also show less activation of mitochondrial biogenesis and a slower recovery of mtDNA copy number [15] . Thus, TLR signaling may directly link the innate immune response in sepsis to the regulation of mitochondrial biogenesis.
To better understand the regulatory role of the PGC family in the complex process of mitochondrial biogenesis during the acute inflammatory response, we used an established fibrin-clot model of S. aureus sepsis to test the hypothesis that TLR signaling leads to downstream regulation of the PGC family of genes. We examined the expression of genes of mitochondrial biogenesis program in this sepsis model in the livers of WT, TLR2 −/− , and TLR4 −/− mice in order to compare the wild-type response to the effects of specific innate immune receptor deficiencies on mitochondrial biogenesis. This study focused on the liver because it is both a key metabolic and immune organ (through the presence of Kupffer cells), and because the liver receives the portal circulation and so its immune cells are stimulated by peritoneal infection.
Most of the research on the regulation of mitochondrial biogenesis has focused on the roles of cytosolic kinases and transcription factors that activate and regulate the genes involved, but post-transcriptional mechanisms are probably also important. In fact, several miRNAs have been demonstrated to target genes involved in metabolism, including PGC-1α [16] , [17] . Such miRNAs are the products of non-protein-coding genes that are processed into mature 19–21 bp sequences. For mRNA binding, the proximal 7–8 bp of the miRNA (the seed region) binds to a complimentary sequence in the 3′ UTR of a target mRNA, resulting in either sequestration or degradation [18] . Each miRNA is complementary to hundreds of mRNAs in silico , though comparatively few matches have been shown to result in gene silencing. Thus, along with gene promoter maps, we compared the 3′ UTRs of PGC family genes to locate conserved miRNA binding sites.
Our findings demonstrate a dampening and a delay in mitochondrial biogenesis during Gram-positive inflammation in both TLR-deficient mouse strains, but also demonstrate that Ppargc1a and Ppargc1b are co-regulated independently of PRC. Moreover, we report that Ppargc1a and Ppargc1b gene expression is negatively correlated with mir-202-3p, which is differently expressed in WT, TLR2 −/− , and TLR4 −/− mice.
Materials and Methods
Mouse studies
The studies were conducted in C57Bl/6J mice purchased from Jackson Laboratories (Bar Harbor, ME) and in TLR2 −/− and TLR4 −/− mice on a C57Bl/6J background obtained from Shizuo Akira, Japan [14] , [19] , and backcrossed >10 generations onto the C57Bl/6J background. Mice of either gender weighing 20–30 grams were used under study protocol A262-07-09, which was approved by the Institutional Animal Care and Use Committee.
Mice were anesthetized with an intraperitoneal injection of xylazine and ketamine, and the abdomen was shaved and cleaned with povidone-iodine. Midline laparotomy was performed and an infected fibrin clot was inserted in the peritoneum. The peritoneum and abdomen were then closed with proline sutures. All mice were resuscitated with 1 ml of 0.9% NaCl administered subcutaneously. Mice were sacrificed at 6, 24, 48, or 72 hours PI by overexposure to isoflurane. Livers of healthy control (HC) mice of each strain were also obtained. The livers were harvested immediately and either the mitochondria were isolated at once or the tissue was snap-frozen and stored at −80°C.
To prepare the fibrin clots, Staphylococcus aureus (ssp aureus) was reconstituted and suspended in fibrin according to published methods [6] . The bacteria were sterilely inoculated on agar slants for 18 hours and then resuspended to a concentration of 10 10 cfu/ml based on optical density at 550 nm. Doses of 10 5 , 10 6 , or 10 7 cfu were then suspended in 500 ul fibrin clots.
Cell Studies
Mouse AML12 hepatocytes were purchased from the American Type Culture Collection (Manassas, VA). AML12 cells were cultured in 5% CO2–95% air at 37°C in DMEM/F12 medium (GibcoBRL, Grand Island, NY) containing l-glutamine and 2.438 g/L sodium bicarbonate. The medium was supplemented with 10% FBS, a mixture of insulin, transferrin, selenium (ITS; Sigma, St. Louis, MO) and 40 ng/ml dexamethasone.
AML12 cells were exposed to 10 7 cfu heat-killed S. aureus per ml (prepared from the same strain as that implanted in mice) and were harvested at different time-points. Gene expression was tested by real time RT-PCR.
AML12 cells were transfected at approximately 60% confluency with scrambled siRNA (AllStars Negative Control siRNA, Qiagen) or miRNA mimic for mmu-mir-202-3p (Qiagen) using Lipofectamine RNAiMax transfection reagent (Invitrogen). Transfection to 80% was confirmed with BlockIT fluorescent oligo (Invitrogen). Serum starvation was achieved 24 hours after transfection by replacing cell culture media with media without FBS for 4 hours. mRNA was extracted with Trizol and gene expression was tested by real time RT-PCR.
Mitochondria and mtDNA Isolation
Liver mitochondria were isolated from ∼2 g of fresh tissue using a modified Clark protocol [20] . Briefly, the livers were hand-dounced in a 0.25 M sucrose buffer and then centrifuged at 2,000x g for 3 min at 4°C. The supernatant was centrifuged at 12,500x g for 8 min at 4°C, and the pellet recovered for mtDNA extraction using the mtDNA Extraction Kit (Wako Chemical, Japan) according to the manufacturer's instructions.
Real-Time RT-PCR
RNA was extracted from frozen liver with TRIzol reagent (Invitrogen, Oslo, Norway) and subjected to reverse transcription with the ImProm-II Reverse Transcription System (Promega, Madison, WI) according to the manufacturer's instructions. Mouse-specific primers were designed ( Table S1 ) and real-time PCR was carried out in triplicate as described, using 18 s primers for internal controls [15] . Real-time PCR output for HC mice of each strain was set to one, and the relative quotients at later time points are shown. The mtDNA copy number count was determined in reference to a standard cytochrome b (Cyt b ) plasmid with a Cyt b probe as described [15] .
MicroRNA PCR
MicroRNAs were prepared with an All-in-One™ miRNA qRT-PCR Detection Kit (GeneCopoeia, Rockville, MD) according to the manufacturer's instructions. Briefly, the extracted RNA was reverse-transcribed in the presence of a Poly-A polymerase with an oligo-dT adaptor. Quantitative PCR was then carried out with SYBR green detection with a forward primer for the mature miRNA sequence and a universal adaptor reverse primer.
In silico analyses
To prepare promoter maps, mouse ( Mus musculus NCBI assembly m37) and human ( Homo sapiens NCBI assembly GRCh37) genomes were accessed on Ensembl ( www.ensembl.org ), and aligned using zPicture (zpicture.dcode.org). The alignments were then fed into rVista 2.0 (rvista.dcode.org) and analyzed for transcription factor consensus sequences according to the Transfac Professional library (v10.2) with similarities optimized for function [21] .
MicroRNA binding predictions were made with TargetScan Mouse release 5.1 ( www.targetscan.org/mmu_50 ) [22] . Specific microRNA binding patterns were then predicted on microRNA.org ( www.microrna.org ). mRNA folding and single-strand frequency predictions were made with mfold version 3.2 [23] , [24] , using mRNA sequences from the Ensembl database.
Statistics
All grouped data are presented as means ± SD. The n values indicated in the figure legends are for the total number of mice from each strain. Each time point in the real-time PCR experiments was compared to the healthy control (HC) of its own strain using a one-sided Student's t -test. In addition, the 6 h time points between strains were compared using two-sided Student's t -tests. The level of significance for all tests was set at P <0.05.
Results
Characteristics of the Sepsis Model
The dose-response behavior to S. aureus sepsis was evaluated in all three strains of mice. At all bacterial doses, WT mice showed very low mortality at 72 hours post-implantation (PI). In contrast, TLR2 −/− mice were dose-dependently susceptible to S. aureus , with 10 6 cfu causing approximately 65% mortality by 72 hours PI. Somewhat unexpectedly, TLR4 −/− mice also showed a high mortality in response to 10 6 cfu S. aureus , approaching 100% by 72 hours PI ( Fig. 1a ). We thus chose a dose of 10 6 cfu S. aureus for all remaining experiments.
10.1371/journal.pone.0011606.g001 Figure 1
Survival curves and mtDNA content in S. aureus sepsis.
(A) Survival curves for WT, TLR2 −/− , and TLR4 −/− mice are shown. Doses of 10∧5, 10∧6, and 10∧7 are shown for TLR2 −/− mice, showing dose-response effects. (B) Absolute mtDNA content in WT and TLR2 −/− was measured by Q-PCR of Cyt b in comparison to known standard. TLR2 −/− mtDNA content is lower than WT mtDNA content in S. aureus sepsis at each time point (n = 4 at each time point; #, p = 0.05; *, p<0.01).
Mitochondrial DNA
Mitochondrial DNA copy number was assessed by quantitative real-time PCR (Q-PCR) in WT and TLR2 −/− mice in healthy controls and at 24, 48, and 72 hours PI ( Fig. 1b ). The TLR2 −/− mice had a lower mtDNA copy number than the WT mice at all times tested, indicating either greater mitochondrial damage or a lag in mitochondrial biogenesis, or both.
Mitochondrial Biogenesis Markers
The mRNA levels for Nrf1 , Gabpa , and mitochondrial transcription factor A ( Tfam ) were measured by Q-PCR in WT, TLR2 −/− , and TLR4 −/− mice ( Fig. 2a, 2b, 2c ). There were no between-strain statistical differences in total mRNA levels for these transcription factors. The times-to-peak, however, were different: WT mice achieved peak induction of all three transcription factors by 24 hours PI, whereas the TLR2 −/− mice peaked by 48 hours PI. Notably, the time course for activation for each of the three transcription factors ( Nrf1,Gabpa , and Tfam ) was similar, suggesting the possibility of mutual transcriptional control during acute inflammation.
10.1371/journal.pone.0011606.g002 Figure 2
Nrf1,Gabpa , and Tfam mRNA levels in S. aureus sepsis.
The mRNA levels of Nrf1 (A), Gabpa (B), and Tfam (C) were measured in WT, TLR2 −/− , and TLR4 −/− mice in healthy controls (HC) and at 6 h, 24 h, 48 h, and 72 h PI. n≥3 at each time point for each strain; *, p<0.05, compared to HC of the same strain.
The mRNA levels of cytochrome b ( Cytb ), superoxide dismutase-2 ( Sod2 ), and thioredoxin reductase-2 ( Txnrd2 ) were measured as indices of damage and recovery of mitochondrial respiratory and antioxidant capacity, respectively ( Fig. 3a, 3b, 3c ). The induction pattern of each of these genes was similar in each different genetic strain. Cytb , Sod2 , and Txnrd2 were all maximally induced at 24 hours in WT mice, but the levels of Txnrd2 were significantly lower at 24 hours in both TLR2 −/− and TLR4 −/− mice ( Txnrd2 : WT vs. TLR2 −/− , p<0.001; WT v. TLR4 −/− , p<0.01). Cytb and Sod2 showed trends towards a decrease in the TLR2 −/− and TLR4 −/− mice compared to the WT mice ( Cytb : WT vs. TLR2 −/− , p = 0.1; WT v. TLR4 −/− , p = 0.05; Sod2 : WT vs. TLR2 −/− , p = 0.08; WT v. TLR4 −/− , p = 0.1). Again, this group of genes may be under mutual transcriptional control during acute inflammation, since innate immune system dysregulation impairs their transcription. Since Cytb is a downstream target of Tfam [25] , and Sod2 and Txnrd2 are up-regulated in mitochondrial biogenesis [26] , [27] , impaired induction of the two NRF transcription factors could inhibit up-regulation of the genes for mitochondrial proteins.
10.1371/journal.pone.0011606.g003 Figure 3
Cytb , Sod2 , and Txnrd2 mRNA levels in S. aureus sepsis.
The mRNA levels of Cytb (A), Sod2 (B), and Txnrd 2 (C) were measured in WT, TLR2 −/− , and TLR4 −/− mice in healthy controls (HC) and at 6 h, 24 h, 48 h, and 72 h PI. n≥3 at each time point for each strain; *, p<0.05, compared to HC of the same strain; #, p<0.01, ##, p<0.001 compared to 6 h time point of the indicated other strain (see further descriptive statistics in Results ).
PGC family members
Ppargc1a , Ppargc1b , and Pprc showed differential regulation in the WT, TLR2 −/− , and TLR4 −/− mice ( Fig 4a, 4b, 4c ). Pprc was similarly up-regulated in all three mouse strains, with an approximate 5-fold induction at 24 hours PI (WT: 4.4-fold vs. HC, p<0.05; TLR2 −/− : 5.2-fold vs. HC, p<0.05; TLR4 −/− : 5.4-fold vs. HC, p<0.05). Ppargc1a and Ppargc1b , however, were both maximally induced in WT mice at 6 hours PI ( Ppargc1a : 6.4-fold vs. HC, p<0.01; Ppargc1b : 4.8-fold vs. HC, p<0.05), but failed to be induced in TLR2 −/− mice at 6 hours PI ( Ppargc1a : 1.3-fold vs. HC, p>0.05; Ppargc1b : 1.2-fold vs. HC, p>0.05), and were over-expressed in TLR4 −/− mice at 6 hours PI ( Ppargc1a : 16.7-fold vs. HC, p = 0.001; Ppargc1b : 10.0-fold vs. HC, p<0.001). Comparisons of Ppargc1a and Ppargc1b mRNA between the three strains at 6 h show that each strain behaved significantly differently from the others ( Ppargc1a : WT vs. TLR2 −/− , p = 0.01; WT v. TLR4 −/− , p = 0.01; Ppargc1b : WT vs. TLR2 −/− , p<0.05; WT vs. TLR4 −/− , p<0.01). This unexpected finding demonstrates the differential regulation of the PGC family members, with Ppargc1a and Ppargc1b showing a TLR-dependent response to S. aureus -induced inflammation. Pprc mRNA levels were not affected by deficiencies of TLR2 or TLR4, and were up-regulated at 24 h PI in all three strains. To test whether S. aureus directly activates Ppargc1a in hepatocytes, studies were conducted in AML12 cells, which were shown to express TLR2 by endpoint RT-PCR (data not shown). AML12 cells were exposed to 10 7 heat-killed S. aureus (HKSA) per mL, and Ppargc1a mRNA levels were measured after different periods of exposure. The means of triplicate time-course experiments showed significant increases in Ppargc1a mRNA levels at 1, 2, and 3 hours post-exposure to HKSA ( Figure 4d ).
10.1371/journal.pone.0011606.g004 Figure 4
Ppargc1a , Ppargc1b , and Pprc mRNA levels in S. aureus sepsis and in cells exposed to HKSA.
The mRNA levels of Ppargc1a (A), Ppargc1b (B), and Pprc (C) were measured in WT, TLR2 −/− , and TLR4 −/− mice in healthy controls (HC) and at 6 h, 24h , 48 h, and 72 h PI. n≥3 at each time point for each strain; *, p<0.05, compared to HC of the same strain; #, p<0.05, compared to 6 h time point of the other two strains (see further descriptive statistics in Results ). (D) Ppargc1a mRNA levels were measured in AML12 cells exposed to 10 7 HKSA at several timepoints after exposure. *, p<0.05, compared to control cells. n = 3 independent time-course experiments.
In silico promoter analyses
In silico analyses of the proximal promoters (500 bp upstream of the transcription start site) of the three PGC family members were performed to uncover any obvious transcription factor binding site similarities between Ppargc1a and Ppargc1b that were not also present in the promoter of Pprc that might explain their co-regulation. Using the web-based programs zPicture and rVista, we were able to demonstrate good conservation of the Ppargc1a and Ppargc1b promoters in the mouse and the human, whereas the proximal Pprc promoter region is poorly conserved between mouse and human. There is modest overlap in the predicted transcription factors for the three genes, but no binding sequences that are conserved between Ppargc1a and Ppargc1b but not Pprc ( Figure S1a ). Interestingly, Pprc shows a high level of conservation between mouse and human in the first 1000 bp in intron 1, where there are 5 conserved binding sites for c-myc, USF, and Max. However, Ppargc1a has no such sites, and Ppargc1b shows only two ( Figure S1b ).
Analysis of the 3′UTR of the PGC family members
Since the proximal promoters of Ppargc1a and Ppargc1b did not share any transcription factor binding sites not also shared by Pprc , we next examined the 3′UTRs of each gene to determine if they share binding sites for miRNAs. We found that Ppargc1a and Ppargc1b share multiple miRNA binding sites, but have none in common with Pprc ( Figure S2 ). The miRNA binding sites were analyzed, and the miRNAs mmu-let-7a and mmu-mir-202-3p were predicted to have both good seed-region binding and good downstream binding to their targets in Ppargc1a and Ppargc1b ( Fig. 5a ). In addition, analyses of the 3′UTR seed regions performed using the online software mfold [23] , [24] showed that the two genes will be in a single-strand formation in roughly one-half the predicted folding patterns, implying an availability to bind the miRNA ( Figure S3 ).
10.1371/journal.pone.0011606.g005 Figure 5
mir-202-3p is associated with Ppargc1a and Ppargc1b degradation.
(A) Predicted binding of mir-202-3p to the Ppargc1a 3′UTR at 24 bp and to the Ppargc1b UTR at 9 bp and at 16 bp. (B) mir-202-3p levels were measured by Q-PCR in WT, TLR2 −/− , and TLR4 −/− mice in healthy controls (HC) and at 6 h and 24 h PI. n = 3 at each time point for each strain; *, p<0.05, compared to HC of the same strain. (C) mir-202-3p expression is negatively correlated with the mRNA levels of Ppargc1a (R 2 = 0.56) and Ppargc1b (R 2 = 0.83). (D) AML12 cells were transfected for 24 h with either mir-202-3p mimic or scramble siRNA, and then serum-starved to induce Ppargc1a . mir-202-3p causes a significant decrease in Ppargc1a mRNA. *, p<0.05, **, p<0.01. Transfections and starvations were performed in triplicate.
Mmu-mir-202-3p and PGC family genes
Specific miRNA levels were measured by Q-PCR in WT, TLR2 −/− , and TLR4 −/− mice. Let-7a was tested first and showed no differential regulation among the three genetic strains (data not shown). We then tested mir-202-3p and found that it was significantly increased at 6 h and 24 h in TLR2 −/− mice (6 h: 3.7-fold vs. HC, p<0.001; 24 h: 3.8-fold vs. HC, p<0.01), was unchanged at 6 h, but was increased at 24 h PI with borderline significance in WT mice (6 h: 1.4-fold vs. HC, p = 0.16; 24 h: 3.3-fold vs. HC, p = 0.10), and was unchanged at 6 h but decreased at 24 h in TLR4 −/− mice (6 h: 1.1-fold vs. HC, p = 0.3; 24 h: 0.4-fold v. HC, p<0.05) ( Fig. 5b ). The miR-202-3p levels thus correlated inversely with Ppargc1a and Ppargc1b mRNA levels at 6 h and 24 h PI. To illustrate this, the average fold-inductions of Ppargc1a and Ppargc1b were plotted separately against the fold-induction of mir-202-3p at the same times in the same strains. The best-fit function showed a negative exponential relationship, with R 2 values of 0.56 for Ppargc1a and 0.83 for Ppargc1b ( Fig. 5c ).
In order to confirm the association between Ppargc1a expression and mir202-3p under more general conditions, we obtained a microRNA mimic of mir-202-3p (a dsRNA sequence that matches the mir-202-3p sequence) for transfection into AML12 cells. AML12 cells were transfected with either mir-202-3p mimic or scrambled RNA for 24 hours, and were then serum-starved for 4 hours to induce Ppargc1a mRNA (See Figure 5d ). In this system, Ppargc1a mRNA increased approximately 7-fold after starvation, but this effect was blunted by the presence of the mir-202-3p mimic (starvation only Ppargc1a : 7.1-fold vs. control, P <0.01; mir-202-3p mimic Ppargc1a : 2.1-fold vs. control, P value vs. starvation only: P< 0.01). There was also a significant reduction from the scramble-treated cells (scramble Ppargc1a : 3.3-fold vs. control, P value vs. starvation only: P< 0.05), however this reduction was significantly less than in the mir-202-3p treatments (scramble vs. mir-202-3p mimic P <0.01). The effect of scrambled RNA on the system may have been due to activation of a non-specific dsRNA response. In any event, the microRNA mir-202-3p experiment functionally confirmed a decrease Ppargc1a mRNA levels in cultured cells.
Discussion
This study examined the effects of acute S. aureus sepsis on the expression of the PGC family and other genes that regulate mitochondrial biogenesis in WT, TLR2 −/− , and TLR4 −/− mice and had two major new findings. First, there is a differential regulation of the PGC family members in S. aureus sepsis that occurs downstream of TLR signaling. Specifically, Ppargc1a and Ppargc1b show a peak expression at 6 h PI in WT mice (confirmed in vitro by exposing mouse hepatocytes (AML12 cells) to HKSA) and TLR4 −/− mice, whereas Pprc expression peaks at 24 h PI in all three strains. The fact that Pprc is equally expressed in all three mouse strains indicates that the gene is not being regulated by signals downstream of either TLR2 or TLR4. However, Ppargc1a and Ppargc1b gene expression are modified by changes in TLR signaling, as TLR2 −/− mice fail to up-regulate either gene, but TLR4 −/− mice show greater mRNA up-regulation than WT mice. Second, we discovered that Ppargc1a and Ppargc1b gene expression may be controlled post-translationally in vivo . The microRNA mir-202-3p is specific for both Ppargc1a and Ppargc1b , and its expression correlates negatively with both genes. In addition, mir-202-3p functionally decreases Ppargc1a mRNA levels in AML12 cells. Thus, we report that Ppargc1a and Ppargc1b are co-regulated in the acute phase of S. aureus sepsis by factors that are downstream of innate immune activation, and that this co-regulation is correlated with the expression of mir-202-3p.
TLR2 is activated by components of the Gram-positive cell wall, whereas TLR4 is activated in response to components of the Gram-negative cell membrane [14] . Previous studies have shown that TLR2 −/− mice have an increased susceptibility to S. aureus , and have higher bacterial loads during S. aureus sepsis [28] . Very few studies have examined the role of TLR4 in vivo in the response to live S. aureus . The available data show that TLR4 −/− mice have an impaired ability to respond to S. aureus [29] . Thus, our finding that TLR4 −/− mice respond differently than WT mice to S. aureus sepsis is in line with the literature. In this study, the increased mortality of TLR4 −/− mice (as compared to WT) in the setting of increased Ppargc1a and Ppargc1b gene expression does not rule out the possibility that overexpression of these genes could, in fact, increase survival. The loss of TLR4 has major repercussions for host defense, including impaired resolution of infection. TLR4 −/− mice have dysregulated immune systems such that multiple factors unrelated to mitochondrial biogenesis probably contribute to the increased mortality rate. In general, TLR-deficient mice tend to fail to upregulate inflammation early in response to their TLR-specific pathogens [30] . However, bacterial infections can activate multiple accessory pathways; for example, the receptors NOD1/2, TLR9, and DAI can all be activated in response to bacterial challenge [31] , [32] . Alternative innate inflammation pathways can thus still be up-regulated in TLR-deficient mice. Overall, one limitation of this study is that is does not establish a direct link between the TLRs and Ppargc1a and Ppargc1b gene expression; instead, we showed a link between their deficiency in an otherwise intact physiologic system and differential regulation of Ppargc1a and Ppargc1b . Further studies such as in vitro work on TLR-deficient cells or a in a system of single TLR expression (such as TLR-expressing HEK293 cells) will be important in confirming the roles of TLRs in the regulation of Ppargc1a and Ppargc1b .
The mitochondrial biogenesis transcription factors Nrf1,Gabpa , and Tfam were all up-regulated in response to S. aureus sepsis in all three mouse genotypes, but there was a lag in time-to-peak in TLR2 −/− mice compared with WT mice (48 h vs. 24 h). In addition, TLR2 −/− peak transcript levels trended towards being greater than WT peak transcript levels, implying that the system attempts to rescue the TLR2 −/− phenotype. Furthermore, the genes of the downstream mitochondrial proteins showed less activation ( Txnrd2 ) or a trend towards less activation ( Cytb , Sod2 ) at 24 h PI in TLR2 −/− and TLR4 −/− mice compared with WT mice. Taken together, the delayed or dampened activation of the assayed genes in TLR2 −/− and TLR4 −/− mice implies that appropriate activation of mitochondrial biogenesis requires TLR signaling.
Ppargc1a and Ppargc1b were co-regulated at 6 h PI in the mouse model, so analyses of both the proximal promoter regions and the 3′UTRs of all three PGC family members were performed in silico in order to identify similarities that could account for the observed co-regulation. Several transcription factors are known to regulate either of the genes alone, and several microRNAs have been shown to be involved in the stability of Ppargc1a and other mitochondrial genes in response to exercise [16] , [17] , [33] . In the published literature, however, no factors have yet been reported that coordinately affect the expression of both genes. The proximal promoters of Ppargc1a and Ppargc1b showed no similarities not also shared by that of Pprc , but the 3′UTRs of the two genes showed binding sites for at least three of the same miRNAs, none of which were found in the 3′UTR of Pprc . This narrow subset of miRNAs was further restricted to those that showed good binding characteristics and whose targets in the two mRNAs were predicted to be in single-strand formation.
The microRNA mmu-mir-202 was previously shown to be one of the small subset of microRNAs that are found in the mitochondria, indicating that it may have associations with mitochondrial function [34] . Mir-202-3p was predicted to have good binding to the Ppargc1a and Ppargc1b mRNAs at sites likely to be in a single-strand formation. Measurements of mature mir-202-3p by Q-PCR revealed that mir-202-3p is increased at 6 h and 24 h in TLR2 −/− mice and also at 24 h in WT mice, but is decreased at 24 h in TLR4 −/− mice. This fits with both the lack of up-regulation of Ppargc1a and Ppargc1b mRNA in TLR2 −/− mice at these time points, and their return to baseline levels at 24 h in WT mice. Thus mmu-mir-202-3p expression correlates with decreased Ppargc1a and Ppargc1b in this in vivo model of sepsis. We suspect that this degradation is caused by binding of mir-202-3p to the identified sites in the target mRNAs, leading to their degradation by the RISC. This in vivo finding was confirmed by in vitro reduction of Ppargc1a in response to mir-202-3p, but the differing magnitude of the changes in Ppargc1a may indicate that secondary effects are also at work. MicroRNAs can individually regulate large networks of genes, so the finding that mir-203-3p correlates with the down-regulation of Ppargc1a and Ppargc1b could in theory be due to a secondary effect of mir-202-3p on a different gene. Further in vivo studies of the microRNA may yield additional information on the complete range of functions of mir-202-3p.
It is well-recognized that dysregulation of oxygen metabolism leads to cellular dysfunction and that mitochondrial damage correlates with mortality in sepsis. Moreover, since mitochondrial biogenesis is activated by cell survival pathways, it is a pro-survival event in sepsis. The PGC family members are known co-activators of the mitochondrial biogenesis transcription factors Nrf1 and Gabpa [1] , [2] , [12] . Mice with embryonic deficiencies in either Ppargc1a or Ppargc1b develop minor metabolic abnormalities, but mice deficient in both factors have a severely disrupted perinatal cardiac phenotype [35] . In addition, both factors are independently capable of promoting mitochondrial biogenesis if they are over-expressed [36] . Thus, our findings that Ppargc1a and Ppargc1b are co-regulated in response to sepsis, and that mRNA stability for each correlates inversely with the expression mir-202-3p, indicate that both may serve pro-survival functions in sepsis. This also identifies mir-202-3p as a possible target for therapy, although currently there are no interventions that can target the blockade of specific miRNA in vivo . Future studies should identify the key transcriptional control mechanisms for mir-202-3p. This could open potential new avenues of mitochondrial therapeutics, as a decrease in mir-202-3p levels is associated with increases in Ppargc1a and Ppargc1b gene expression. This would translate to therapeutic potential if Ppargc1a or Ppargc1b gene activation could be shown to improve recovery or lessen mortality in clinical sepsis.
Supporting Information
Figure S1
In silico promoter and intronal analyses of Ppargc1a , Ppargc1b , and Pprc . (A) The proximal 500 bp from the TSS in Ppargc1a , Ppargc1b , and Pprc from both mouse and human were aligned with zPicture and then submitted to rVista, with all transcription factors searched under an optimized matrix. (B) The proximal 1000 bp in intron 1 in Ppargc1a , Ppargc1b , and Pprc from both mouse and human were aligned with zPicture and then submitted to rVista, with all transcription factors searched under an optimized matrix.
(6.63 MB EPS)
Figure S2
In silico 3'UTR miRNA binding analyses of Ppargc1a , Ppargc1b , and Pprc . The 3'UTRs of Ppargc1a , Ppargc1b , and Pprc were analyzed in the mouse using TargetScan Mouse release 5.1. Shown are miRNAs with good seed-binding characteristics. The let-7 family and mir-202-3p were conserved between Ppargc1a and Ppargc1b .
(3.31 MB EPS)
Figure S3
In silico 3'UTR mRNA folding analyses of Ppargc1a and Ppargc1b . The entire mRNA sequences of Ppargc1a and Ppargc1b were separately fed into the program mfold, and predictions were made on average single-strand characteristics of the mRNA. Shown are the regions surrounding the predicted mir-202-3p binding sites. The green bars show the seed-binding regions, while the red bars show the length of predicted mir-202-3p binding.
(1.09 MB EPS)
Table S1
Real-time PCR primer sequences and gene accession numbers.
(0.02 MB XLS)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The recent demonstration that bacterial genome reconstructions are possible from archaeological material [1] presents a significant advancement in the field of infectious disease research by permitting an evaluation of the evolutionary history of human pathogens. Accurate quantification of genetic changes over time and rates of evolutionary change, however, are highly contingent upon our having an accurate representation of the diversity present in modern related organisms; hence, the evolutionary context of ancient pathogen genomes can only be properly interpreted if adequate sequence data are available and are reliably representative of extant pathogenic strains. Rapid and cost-effective DNA sequencing techniques have helped to increase the acquisition of such data, and while we eagerly await the addition of new pathogen genomes to publicly available databases, evolutionary histories can at best be extrapolated based on the limited data that are currently accessible.
Genomic comparisons of ancient and modern strains offer the best resolution for estimating rates of genetic change [2] . Such an approach was used in the analysis of our reconstructed ancient Yersinia pestis genome [1] , where comparisons to all 18 complete modern genomes that are publicly accessible produced tighter rates of change than previous estimates, thus yielding more accurate divergence times on the phylogenetic tree [3] , [1] . This method introduced an inherent limitation, however, since our exclusive use of available modern genomes likely encapsulated only a fraction of the true genetic diversity present amongst extant Y. pestis strains. To identify possible evolutionary events missed by our genomic analysis, we re-evaluated our reconstructed ancient Y. pestis sequence by comparing it against single nucleotide polymorphism (SNP) data publicly available for 289 Y. pestis strains at over 600 positions [3] . Although this method does not provide the resolution necessary for reassignment of our molecular clock values, it could provide a qualitative indication of phylogenetic signals that were lost via our original, more conservative analytical approach based strictly on complete genomes.
Methods
We downloaded the raw SNP data of Morelli et al. [3] from the European Bioinformatics Institute’s ArrayExpress database ( http://www.ebi.ac.uk/ArrayExpress ) with accession number ID E-MTAB-213, consisting of 1217 SNPs in 289 strains. These were joined with the SNP data of Bos et al. [1] that characterise 1761 SNPs in 27 strains. The consolidated dataset resulted in a total of 946 SNPs in 311 strains.
All positions were assigned relative to the Y. pestis CO92 strain (NC_003143.1). This dataset carries several missing calls in individual strains or ambiguous calls that are likely attributed to sequencing issues. Based on these 946 SNP positions we then computed a Maximum Parsimony (MP) tree using MEGA 5 with 200 bootstrap replicates [4] . We chose the partial deletion method with a 95% cut off in MEGA 5 that removed 310 SNPs that had insufficient data before the MP-tree calculation. The resultant tree is, therefore, generated from the 636 remaining SNPs. Table S1 shows the sequence data for all 946 SNPs and identifies those that were either retained ( Table S1a ) or removed ( Table S1b ) by the 95% partial deletion filter. As in [1] , we used the Y. pseudotuberculosis genome as our outgroup.
To infer the MP-tree we used the close-neighbour-interchange method. Inferred branch lengths were used to estimate a relative divergence time of the intermediate clade between the divergence of the rodent pathogen Y.pestis microtus ( Figure 1 , branch A) and the ancient Black Death cluster ( Figure 1 , branches B, C, and D). The ratio of branch length A to the total branch length A+B+C+D gives a divergence for the intermediate cluster of F = A/A+B+C+D = 0.5571 times the total distance. The approximate minimum and maximum divergence times of the intermediate Y.pestis cluster were calculated by multiplying our value for F by the distance between the two minimum and the two maximum ages of Y.pestis microtus (41 AD –480 AD) and the Black Death (1283 AD –1342 AD).
10.1371/journal.pone.0049803.g001
Figure 1
Maximum parsimony (MP) tree showing the relative phylogenetic placement of 311 Y. pestis strains.
Branch 1 and Branch 2 designations as well as Y. pestis group names follow those defined in [9] . Dates in black denote those generated via full genomic analysis [1] , whereas dates in gold are inferred here from relative branch lengths.
Results and Discussion
Our MP tree ( Fig. 1 , Fig. S1 ) confirms the result shown previously that our ancient genome essentially falls at the root of all Y. pestis commonly-associated with human infection, namely the branch 1 and branch 2 strains [1] . Our previous full genome analysis revealed two SNPs common to the branch 1 family that distinguished our ancient sequence from the branch 1 and branch 2 root. Neither of these SNPs, however, are included in the current tree construction since one was absent in the Morelli et al. dataset [3] and was thus removed when the intersection of the two datasets was generated, and the other was removed by the 95% partial deletion filter in MEGA 5. Removal of these two positions makes the ancestral sequence sit directly at the root of the branch 1 and branch 2 strains.
Regardless, our analysis shows three novel observations that become apparent by our inclusion of SNP data from the additional Y. pestis strains. First, there is a cluster of 11 strains that diverge at a position predating the Black Death. These strains correspond to sequences in the 0.ANT1 Y. pestis group defined in Morelli et al. [3] ( Fig. 1 , Fig. S1 ). Our tree shows this group to consist of a single branch, whereas the network of Morelli et al. [3] shows it to be represented by three branches. Again, this discrepancy arises from our conservative approach of using the intersection of both the Morelli et al. and Bos et al. SNP datasets in combination with our partial deletion filter. High coverage full genomic data will be helpful to resolve the degree of genetic diversity present in these strains. It is not currently known if these sequences represent strains that are pathogenic to humans, though the placement of their divergence provisionally suggests a radiation event possibly resulting from a distinct epidemic occurring significantly in advance of the Black Death. Based on similarity in mortality levels, geographic distribution, and recorded symptoms, historians have long suspected that the Plague of Justinian (542–740 AD) might have been caused by the same infectious agent as that responsible for the 14 th -century Black Death [5] . Since several publications have implicated Y. pestis as the principal cause of the Black Death by phylogenetic assignment and evaluation of DNA quality [1] , [6] , [7] , the possibility that the Plague of Justinian may have been responsible for the deep cluster we observe here carries some legitimacy. This is further supported by the placement of the cluster approximately half the distance between the Black Death (1283–1342 AD) and the ancestral rodent strain Y. pestis microtus , which is suspected to have diverged from the soil-dwelling Y. pseudotuberculosis root approximately 2000 years ago (41–480 AD). Our cursory dating analysis based on relative branch lengths reveals a divergence time of 733–960 AD for this cluster, thus placing it in a phylogenetic position expected for a Y. pestis radiation event roughly coincident with the Plague of Justinian. We regard this as an important observation since a Y. pestis involvement in the plague of Justinian seemed unlikely from our previous whole genome analysis [1] . Confirmation of their potential for human infection via isolation of one of these strains, or an as yet uncharacterised close relative, from an infected patient would be helpful to understand their possible relationship to the plague of Justinian.
Second, there is a small cluster of Y. pestis sequences that diverge immediately predating the Black Death, all of which define the 0.ANT3 Y. pestis group [3] ( Fig. 1 , Fig. S1 ). Again, it is not known whether these in fact represent Y. pestis sequences that could be pathogenic to humans; however, the most parsimonious explanation holds that this cluster represents a genetic diversification event immediately preceding the disease’s arrival in Europe. Though this could have occurred either in human populations or in rodent populations, it most likely took place in East Asia. Genetic evidence is accumulating in support of the view long held by historians that the disease responsible for the Black Death likely emerged in China [3] , [8] , even though adequate historical evidence is currently lacking to disclose the precise location. This notion is supported by the fact that the strains in this cluster were all isolated from China [3] . Regardless, this diversification event likely gave rise to the strain(s) responsible for the high mortality in the European pandemic of 1347–1351 AD.
Third, the ancient strain obtained from medieval London clusters with two modern strains of the 3.ANT Y. pestis group [3] ( Fig. 1 , Fig. S1 ) based on SNP profiles, thus revealing that the variant responsible for the Black Death is identical to certain modern strains of Y. pestis for the 636 positions considered in this analysis. Full genomic data are, however, not currently available for these strains; thus, we cannot determine the full extent of their similarity with the ancient genome, and it is likely that the two modern strains discussed here possess additional derived positions to distinguish them from the ancient sequence. In addition, the genetic architecture of these modern strains is currently unknown, and since little is known is about the structure of the ancient genome, we cannot currently assess similarities in gene order between the strains.
We acknowledge that the above conclusions have yet to be confirmed via a more robust full genomic comparison of Y. pestis strains, both contemporary and historical. Specifically, Y. pestis data from human mass burials dating to the Justinian era may hold pertinent information to permit a more thorough evaluation of the evolutionary history of this notorious human pathogen.
Supporting Information
Figure S1
Maximum parsimony tree showing identification names for all Y. pestis sequences considered in this analysis. Branch and group designations match those defined in [9] . Sample identification names match those in the ArrayExpress (E-MTAB-213) dataset. “E. Smith.” refers to the East Smithfield Black Death sequence described in [1] .
(TIF)
Table S1
Sequence data for all SNPs generated from the intersection of the Bos et al. [ 1 ] and Morelli et al. [ 3 ] datasets. Table S1a shows those retained by the 95% partial deletion filter in MEGA 5, and Table S1b shows those removed by the filter.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The key components of a sensitized solar cell are the photoanode, sensitizer (dye molecules or semiconductor quantum dots) and electrolyte. The photoanode both loads the sensitizers and offers channels for the transportation of photogenerated carriers. Metal oxides play an important role in optoelectronics [ 1 , 2 ]. A variety of metal oxides, e.g., titanium dioxide (titania, TiO 2 ), zinc oxide (ZnO) and tungsten oxide (WO 3 ), can be used as the photoanodes [ 3 – 5 ]. So far, TiO 2 is still by far the most suitable semiconductor material for the photoanodes. Due to a variety of appealing physical and chemical properties, ZnO has demonstrated application perspectives in many areas[ 6 – 13 ]. Importantly, it is also considered to be a possible alternative to TiO 2 as the photoanode material in sensitized solar cells[ 14 – 22 ]. The bulk electron mobility of ZnO is more than one order of magnitude larger than that of TiO 2 and this better electron transport property may alleviate the mass transport limitations in the solar cells [ 3 , 23 , 24 ]. Moreover, among all the metal oxides, ZnO also has the richest family of nanostructures, so that the morphology of the photoanode might be more controllable. Especially, in some quantum-dot-sensitized solar cells (QDSSCs), photoanodes based on one-dimensional (1D) ZnO nanostructures, such as ZnO nanorods and ZnO nanowires, are employed to load the QDs and provide the transportation channels for the photoexcited electrons. In comparison with semiconductor nanoparticles, there are relatively fewer grain boundaries in these 1D ZnO nanostructures, thus charge recombination can be suppressed to a certain degree [ 25 – 27 ]. Nonetheless, the capability of attaching QDs and harvesting incident photons of these 1D ZnO nanostructures is limited by their low specific areas. Similar problem also exists in the 1D nanostructure-based photoanodes in the dye-sensitized solar cells (DSSCs). For overcoming this difficulty, hierarchical structured photoanodes have been developed in both DSSCs and QDSSCs [ 28 – 47 ].
Among these works, relatively less effort has been devoted to the QDSSCs than to the DSSCs. In this context, a ZnO hierarchical photoanode with a different configuration from the previous ones has been developed in this lab. The CdS QD-sensitized ZnO photoanodes are usually synthesized by the initial growth of the ZnO nanostructures and the subsequent deposition of the CdS QDs [ 48 – 53 ]. In this work, cone-shaped primary nanostructures were used for a large space between them to accommodate sufficient electrolyte. Spike-shaped secondary nanostructures were used for a specific area as large as possible. Both the primary and the secondary nanostructures were ZnO. These ZnO hierarchical nanostructures were grown using a facile and inexpensive hydrothermal method.
Experimental Section
Each ZnO hierarchical nanostructure array was fabricated in two steps, namely the growth of the primary ZnO nanostructure array and the subsequent growth of the secondary ZnO branches on these primary nanostructures. The primary ZnO nanostructure array was grown on a fluorine-doped tin oxide (FTO) glass substrate using a field-assisted method, whose details can be found in Ref.[ 54 ]. The precursor was an aqueous solution that contained 0.02 M zinc nitrate (Zn(NO 3 ) 2 ) and 0.02 M hexamethylenetetramine (HMTA, C 6 H 12 N 4 ). A beaker that contained this solution was immersed in a water bath and two electrodes were inserted into the solution. The FTO substrate was used as the cathode and a Pt wire was used as the anode. The reaction occurred under 90°C with a 2.5 V voltage applied to the two electrodes. The reaction time was controlled to be 3, 6 or 9 h. Then the FTO substrate with the primary ZnO nanostructures was immersed in a limpid aqueous solution of 0.057 M zinc acetate (Zn(CH 3 COO) 2 ·2H 2 O, ZnAc) and 0.5 M sodium hydroxide (NaOH) under constant stirring at room temperature for the growth of the secondary ZnO branches, so that a hierarchical ZnO nanostructure array was finally obtained [ 34 ]. (It is worth stressing that the limpid solution would become turbid about 5 min after NaOH and ZnAc were dissolved and secondary ZnO nanostructures would not be available if the FTO substrate was inserted in a turbid solution. That is, for obtaining the secondary ZnO nanostructures, the FTO substrate should be inserted to the solution right after the NaOH and ZnAc were dissolved in water.) After this, the sample was repeatedly rinsed using deionized water and ethanol.
For the deposition of CdS QDs[ 48 , 55 ], the FTO substrate was further immersed in a mixed solution of 5 mM cadmium nitrate (Cd(NO 3 ) 2 ) and 5 mM thioacetamide (C 2 H 5 NS) in 100 mL deionized water for 30 min, and then rinsed with deionized water and ethanol. Finally the sample was annealed in air at 400°C for 30 min.
The ZnO hierarchical nanostructure arrays, either with or without the CdS QDs on them, were characterized using such means as scanning electron microscopy (SEM, FEI Quanta 600 microscopy), transmission electron microscopy and energy dispersive X-ray spectroscopy (TEM and EDS, JEM-2100F), X-ray diffraction (XRD, DMAX-2400) and UV-vis spectrophotometry (UV 5000 spectrometers, Gary). For the TEM observation, part of the ZnO nanostructures were scratched down from the photoanodes and dispersed in ethanol. Then some ethanol drops with the ZnO nanostructures were applied to the C thin films on the Cu meshes.
For assembling a QDSSC, the CdS sensitized ZnO electrode and a platinized FTO counter electrode were sealed together with a 60-μm-thick hot-melt surlyn spacer. The ZnO nanostructures loaded with CdS QDs were thus sandwiched between two transparent FTO electrodes. An I − /I 3 − based electrolyte (DHS-E23, Dalian HeptaChromaSolarTech, China) was then injected into the space between the two electrodes through holes in the counter electrode.
Photovoltaic properties of the QDSSCs were measured under AM 1.5 simulated sunlight at 100 mW·cm −2 (Oriel Solar Simulator, Model 91160). The exposed area was 0.25 cm 2 . First, the dependence of the photocurrent density on the photovoltage ( J - V curves) was recorded. Then, the incident-photon-to-current efficiency (IPCE) spectra were analyzed in the wavelength range from 350 to 800 nm. Finally, more properties of the cells were disclosed using the electrochemical impedance spectroscopy (EIS). The EIS measurements were performed at the V OC s under dark conditions.
Results and Discussion
The cone-shaped primary ZnO nanostructures shown in Fig 1A and 1B , which were fabricated in the aqueous solution of Zn(NO 3 ) 2 and HMTA, are all roughly upwards aligned. Hereafter these ZnO nanostructures are referred to as “ZnO nanocones (ZNC(x)s)”, where “x” denotes the growth time in hour. As introduced in the “Experimental” section, the value of “x” can be 3, 6 or 9. As shown in Fig 1A and 1B , the ZnO nanocones grown in 3 h (ZNC(3)s) were around 4.2 μm in height. When the growth time was prolonged to 6 and 9 h, the height increased to 5.7 and 6.6 μm, respectively. ( S1 Fig ) After the reaction in the aqueous solution of ZnAc and NaOH, as shown in Fig 1C and 1D , the previously smooth ZNC surfaces were covered with secondary small protrusions, hereafter referred to as “ZnO nanospikes (ZNSs)”, and became quite rough. The height of the ZNC(3)/ZNS, ZNC(6)/ZNS and ZNC(9)/ZNS arrays increased only slightly to 4.4 ( Fig 1C ), 5.8 and 6.9 μm ( S1 Fig ), respectively, and the thickening was also detectable but not considerable. The process of the two-step growth of such a ZNC/ZNS hierarchical nanostructure array is shown in Fig 1E .
10.1371/journal.pone.0138298.g001
Fig 1
SEM images of the ZNC(3) arrays and the hierarchical ZNC(3)/ZNS arrays.
(A) top view and (B) side view of a ZNC(3) array, (C) top view and (D) side view of a ZNC(3)/ZNS array, and (E) the schematic of the two-step growth process of a ZNC/ZNS array.
The results of the XRD analysis on a ZNC array, a ZNC/ZNS array and a CdS QD-sensitized ZNC/ZNS array are shown in Fig 2A together and little difference is observable between them, demonstrating that the crystal structure of the primary ZNC and the secondary ZNS were similar to each other. The main peaks at 2θ values of 32.2, 34.9, 36.7, 48.0, and 63.3° can be respectively indexed to the (100), (002), (101), (102) and (103) crystal planes of the hexagonal phase ZnO (JCPDS No. 36–1451). The CdS QDs were too small to be detected in the XRD analysis. The TEM images given in Fig 2B and 2C further disclosed the smooth surface of a ZNC and the surface decorated with small protrusions of a ZNC/ZNS nanostructure. These results obtained from individual nanostructures are in agreement with those obtained from arrays shown in Fig 1 . In the HRTEM image shown in Fig 2D , the crystal planes are separated by 0.28 nm, which is in accordance with the interplanar spacing of the (100) crystal planes of the wurtzite ZnO (JCPDS No. 36–1451). In the EDS shown in Fig 2E , except for the Cu peak arising from the supporting mesh, only the Zn and O elements are detectable. This result shows that the residuals on the samples during the reaction, e.g., elemental sodium, were all successfully removed by the rinsing and the purity of the samples could be guaranteed.
10.1371/journal.pone.0138298.g002
Fig 2
The XRD patterns and TEM images of the ZNCs and ZNC/ZNS nanostructures.
(A) XRD patterns of a ZNC array, a ZNC/ZNS array and a CdS QD-sensitized ZNC/ZNS array, (B) TEM image of a ZNC, (C) TEM image and (D) HRTEM image of a ZNC/ZNS nanostructure, (E) EDS of a ZNC/ZNS nanostructure.
As previously reported by this lab [ 54 , 56 ], on the one hand, the ZnO growth could occur even without the application of an external voltage; on the other hand, the application of an external voltage indeed greatly improved the orderliness of the ZnO nanostructures. Therefore, both a hydrothermal growth mechanism and an electrochemical growth mechanism were likely to have contributed to the growth of the ZNCs in this work.
As a common knowledge, Zn 2+ cations occur largely in four-coordination as tetrahedral complexes [ 57 ]. Thus it is believed that Zn(OH) 4 2- s were the precursors of the ZnO growth in this work. That is, the formation of Zn(OH) 4 2– precursors and their subsequent incorporation into the ZnO crystals resulted in the growth of the ZnO crystals [ 58 ]. Obviously, the Zn 2+ s in these precursors came from the Zn(NO 3 ) 2 . The existence of the hydroxide anions (OH - s) is attributed to the use of the HMTA. The HMTA slowly decomposed to produce ammonia (NH 3 ) in a gradual and controlled manner, which could form ammonium hydroxide (NH 4 OH) and provide the OH - s [ 59 , 60 ]. The reactions in the solution can be simplified as described by the following formulae[ 61 – 63 ]:
( CH 2 ) 6 N 4 +6H 2 O → 6HCHO+4NH 3
(1)
NH 3 + H 2 O → NH 4 + + OH −
(2)
Zn 2+ +4OH − → Zn ( OH ) 4 2 −
(3)
Zn ( OH ) 4 2 − → ZnO+H 2 O+2OH −
(4)
In a ZnO crystal, the Zn-terminated (0001) plane had a high surface energy, thus Zn(OH) 4 2– precursors were preferentially adsorbed to the (0001) plane [ 64 ]. Moreover, besides providing the OH - s, the HMTA also played an important role in hindering the growth of some planes [ 65 ]. For example, as proposed by Sugunan et al., HMTA was preferentially attached to some nonpolar planes of the ZnO crystal and thus cut off the access of Zn 2+ s to these planes[ 66 ]. Consequently, the c-axis became the major growth direction.
The application of an external voltage presumably triggered the following electrochemical reactions [ 67 – 69 ]:
NO 3 − +H 2 O+2e − → NO 2 − +2OH −
(5)
NO 3 − +6H 2 O+8e − → NH 3 +9OH −
(6)
Hence, besides reactions (1) and (2), the electrochemical reactions described in (5) and (6) also generated OH - s in the solution, providing more reactants for reactions (3) and (4). With excessive OH - s, the movement of the Zn 2+ s to the ZnO crystal was likely to become the rate-determining step in the ZnO growth. Tena-Zaera et al argued that ZnO nanowires mainly grew along the longitudinal axis if the diffusion of Zn 2+ s was slower than the generation of OH - s [ 70 ]. Under the negative electric field at the cathode surface, on the one hand, more Zn 2+ s moved to the vicinity of the ZnO crystals; on the other hand, the negatively charged Zn(OH) 4 2– precursors generated in reaction (3) received a repulsion from the cathode and became more difficult to be adsorbed onto the ZnO crystal planes. As a result, only the growth along the c-axis continued due to the high reactivity of the (0001) plane [ 64 , 71 ]. The eventual result was that the negative electric field enhanced the growth along the c-axis by attracting more Zn 2+ s and hindered the growth along other directions by repulsing the Zn(OH) 4 2– precursors.
Electric field around a sharp end of a ZnO crystal was stronger than that around a flat end, thus a sharp end could attract more Zn 2+ s and grew faster. Therefore, the primary ZnO nanostructures in this work became cone-shaped under the external electric fields. Moreover, it is possible that the erosion by the OH - s around the boundaries of the (0001) planes also contributed to the tapering of the ZnO nanostructures [ 72 ].
The crucial factor for the growth of the secondary ZNSs was the formation of the “etch pits” on the surfaces of the primary ZNCs. From these etch pits, the nanospikes, which were much smaller than the nanocones, were further developed in the supersaturate aqueous solutions of ZnAc and NaOH [ 34 ]. When an FTO substrate with ZNCs on it was immersed in the limpid solution with high concentration NaOH, the OH - s could erode the already existent ZnO nanostructures [ 72 ]:
ZnO + 2 OH − → ZnO 2 2 − + H 2 O
(7)
A large number of etch pits resulted from this reaction on the ZNC surfaces and constituted the starting places for the growth of the ZNSs.
According to the analysis in Refs. [ 73 – 75 ], the Zn(OH) 4 2– precursors still played a vitally important role in the growth of the secondary ZNSs, whose process can as well be described by formulae (3) and (4). This time the Zn 2+ cations and the OH - anions came from the ZnAc and NaOH, respectively. In comparison with some methods of growing ZnO nanowires and nanoforests, the growth of the nanocones and the nanospikes in this work, which did not involve any organic structure-directing agents, seeding process and heating process, appeared to be simpler and less costly.
Fig 3A and 3B show the TEM and HRTEM images of the ZNC/ZNS nanostructures with some QDs, ~6 nm in size, adsorbed on them. Fig 3B also shows lattice fringes of 0.316 and 0.336 nm, which can be indexed as the (101) and (002) planes of CdS (JCPDS 10–0454). The selected area EDS of the sample, shown in Fig 3C , further confirms the elementary composition of the CdS-coated nanostructures, which, as expected, are Zn, O, Cd and S elements. (The Cu peak arose from the supporting Cu mesh in the TEM observation.)
10.1371/journal.pone.0138298.g003
Fig 3
ZNC/ZNS nanostructures loaded with CdS QDs.
(A) a TEM image, (B) an HRTEM image, and (C) EDS.
The UV-vis absorption spectra of the as prepared ZnO nanostructures and the CdS sensitized ZnO nanostructures, which were calculated from the reflectance spectra and the transmission spectra ( S2 Fig ), are given in Fig 4A .
10.1371/journal.pone.0138298.g004
Fig 4
UV–vis absorption spectra and band gap estimation of ZnO and ZnO/CdS nanostructures.
(A) UV–vis absorption spectra; (B) band gap estimation.
The energy band gaps of the ZnO nanostructures and CdS QDs were estimated using the formula [ 76 ]:
α ⋅ h ν = A ( h ν − E g ) n ,
(8)
where A is a constant, α the absorption coefficient, hν the photon energy and E g the energy band gap. The value of n depends on the type of the semiconductor. Since both CdS and ZnO are direct semiconductors, n is 0.5 [ 76 – 78 ]. As shown in Fig 4B , the linear parts of the dramatic increase in ( α · hν ) 2 were extrapolated to the low photon energy end and the intersections with the abscissa are considered to be the band gap values. Using this method, the band gaps of the ZnO nanostructures and CdS QDs are estimated to be 3.2 and 2.3 eV, respectively. These results are in good agreement with the values given in Ref. [ 79 ], which reported the band gaps of CdS and ZnO to be 2.25 and 3.2 eV, respectively.
As shown in Fig 4 , no matter the nanostructures were covered with the CdS QDs or not, the visible light absorption of the hierarchical ZNC/ZNS nanostructures was stronger than that of the ZNCs, indicating the use of the hierarchical nanostructures could improve the light harvesting efficiency of the photoanode.
The ZNC and ZNC/ZNS nanostructure arrays were loaded with CdS QDs and then respectively used as the photoanodes of the solar cells. Their J - V curves, EIS and IPCE curves are shown in Fig 5 . The short-circuit current density ( J SC ), open-circuit voltage ( V OC ), fill factor ( FF ) and conversion efficiency ( η ) of the cells are listed in Table 1 . The values of the parameters that describe the properties of the ZnO/CdS/electrolyte interfaces are given in Table 2 . The η of the cell based on the ZNC(3)/ZNS photoanode is more than twice that of the cell based on the ZNC(3) photoanode (1.1% vs. 0.53%). This improvement mainly arises from a 76% increase in the J SC of the ZNC(3)/ZNS-based cell over that of the ZNC(3)-based cell (4.67 vs. 2.66 mA·cm −2 ). In total, 20 ZNC(3)/ZNS-based QDSSCs were fabricated and their photovoltaic properties were quite reproducible ( S3 Fig ). The hierarchical nanostructures of the ZNC/ZNS photoanode accommodated more CdS QDs and also trapped more incident light inside the photoanode, thus a larger J SC resulted.
10.1371/journal.pone.0138298.t001
Table 1 Photovoltaic properties of the CdS-sensitized cells with the ZNC photoanodes and the ZNC/ZNS photoanodes.
Photoanode
J SC (mA cm −2 )
V OC (V)
FF
η (%)
ZNC(3)
2.66
0.605
0.33
0.53
ZNC(3)/ZNS
4.67
0.618
0.37
1.1
ZNC(6)/ZNS
5.00
0.624
0.42
1.3
ZNC(9)/ZNS
6.06
0.625
0.37
1.4
10.1371/journal.pone.0138298.t002
Table 2 Values of the resistance and capacitance per unit area across the photoanodes at the ZnO/CdS/electrolyte interfaces obtained from the fitting of the data shown in Fig 5B .
Anode
R 3 (kΩ cm 2 )
C 3 (μF cm −2 )
R 3 · C 3 (ms)
ZNC(3)
0.22
4.3
0.9
ZNC(3)/ZNS
0.24
5.9
1.4
10.1371/journal.pone.0138298.g005
Fig 5
Photovoltaic properties of the CdS-sensitized cells with the ZNC and ZNC/ZNS nanostructures as the photoanodes.
(A) J - V curves, (B) Nyquist plots (The equivalent circuit is given as the inset. The dots are experimental data and the two semicircle lines are the results of fitting.) (C) IPCE curves.
The results given in Fig 5A and Table 1 show that prolonging the ZNC length was effective in raising the conversion efficiency. The conversion efficiencies of the cells based on ZNC(6)/ZNS and ZNC(9)/ZNS were 1.3% and 1.4%, which were 18% and 27% higher than that of the ZNC(3)/ZNS-based cell, respectively. Table 1 further shows that this improvement in the conversion efficiency mainly arose from the increase in J SC (4.67 vs. 5.00 & 6.06 mA cm −2 ). The actual surface areas of the photoanodes were enlarged by prolonging the ZNCs, thus more QDs were adsorbed on them and larger J SC s resulted.
In comparison with other works on the QDSSCs based on ZnO photoanodes, the results obtained in this work is quite competitive ( S1 Table ), confirming the effectiveness of utilizing hierarchical nanostructures in improving the QDSSC performance.
At a ZnO/CdS/electrolyte interface, some of the photoexcited electrons would recombine with the holes around the interface instead of moving forward into the photoanode. This recombination process would lower the V OC of the cell [ 80 ]. Interestingly, although the employment of the hierarchical nanostructures increased the specific area of the photoanode, both the V OC and FF of the ZNC/ZNS-based cell were still higher than that of the ZNC-based cell, though only slightly.
As shown in Fig 1C , the ZnO nanospikes provided more channels for the fast transportation of the photoexcited electrons into the photoanodes, thus these photoexcited electrons had a smaller probability of recombining with holes, giving rise to higher V OC and FF . This slight increase in the V OC and FF is echoed in the EIS results shown in Fig 5B and Table 2 .
Theoretically, the Nyquist plot of a solar cell should contain three semicircles at low, medium and high frequencies [ 81 , 82 ]. The experimental data can be fitted with an equivalent circuit shown in the inset of Fig 5B . R 1 in the equivalent circuit, which can be calculated by analyzing the low frequency part of the Nyquist plot, is a reflection of the Nerst diffusion of the I - and I 3 - anions in the electrolyte [ 81 ]. R 2 and C 2 are the incarnations of the electrochemical reaction impedance at the Pt counter electrode and can be worked out from the left semicircles at high frequencies. R 3 and C 3 , whose value can be estimated using the semicircle at the medium frequencies, are used to represent the charge transfer impedance at the QD/electrolyte/ZnO triple junction. All the recombination events between the electrons and holes in the ZnO photoanode, the electrolyte, and the QDs, which hindered the electron transfer from the QDs to the FTO substrate through the ZnO photoanode, contributed to R 3 [ 30 , 83 , 84 ].
In the Nyquist plots shown in Fig 5B , only the semicircles at the medium frequencies are large enough for analysis, thus only the values of R 3 and C 3 are calculated by fitting. Fig 5B and Table 2 demonstrate that the resistance per unit area across the ZNC(3)/ZNS photoanode was slightly larger than that of the ZNC(3) photoanode (0.24 vs. 0.22 kΩ·cm −2 ), suggesting that the employment of the hierarchical nanostructures also reduced the recombination events, despite of an increase in the actual interface area. As a result, the lifetime of the photoexcited electrons at the interface, i.e., the multiplication of the resistance and capacitance, of the ZNC(3)/ZNS photoanode is longer than that of the ZNC(3) photoanode (1.4 vs. 0.9 ms). As previously reported, photoexcited electrons might diffuse to the FTO via the pathway of the secondary nanospikes and circumvented the primary nanorods [ 34 , 44 ]. On a ZNC/ZNS photoanode, a photoexcited electron had two possible pathways to the external circuit. Firstly, it could entered a ZNS first and then entered the underlying ZNC through the ZNC/ZNS interface. Secondly, it could move along the ZNS surfaces all the way to the FTO substrate. The probability of a photoexcited electron recombining with a hole is high at the ZNC/ZNS interface. The fact that the lifetime of a photoexcited electron in the ZNC/ZNS photoanode was longer than that in the ZNC photoanode (1.4 vs. 0.9 ms, Table 2 ) suggests that most photoexcited electrons followed the conduction channels along the ZNS surfaces. The enlargement of surface area of the ZNC/ZNS photoanode provided more opportunities for the photoexcited electrons to diffuse to the external circuit.
The IPCE spectra in Fig 5C show that the employment of the hierarchical nanostructures did not change the shape of the response spectrum of the cell. Instead, it only greatly raised the response at the wavelengths between 0.4 to 0.5 μm.
Conclusion
Applying hierarchical nanostructures, which can be obtained using a facile and inexpensive hydrothermal method, is an effective approach to improving the performance of a QDSSC based on a ZnO photoanode. The results of comparative experiments disclosed that the J SC , V OC and FF were all raised when a photoanode of ZnO nanocones was replaced by a photoanode of ZnO nanocone/nanospike hierarchical nanostructures. As a result, a more than doubled conversion efficiency was attained. This improvement in the photovoltaic performances indicates that the employment of properly tailored hierarchical nanostructures can not only enhance the capability of a photoanode to load QDs and trap incident light, but also, to a lesser extent, to contain the recombination of the photoexcited electrons.
Supporting Information
S1 Fig
Longer growth time.
(DOC)
S2 Fig
Reflectance spectra and transmittance spectra.
(DOC)
S3 Fig
20 samples.
(DOC)
S1 Table
Comparison.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Ageing [1] and the health transition in low- and middle-income countries (LAMICs) are responsible for an unprecedented increase in the prevalence and societal impact of noncommunicable diseases, including dementia [2] . Large numbers of people with dementia currently live in LAMICs [3] , [4] with prevalence estimates comparable to those of the Western world [5] . At present, disease-modifying drugs are not available [6] and symptomatic medications have been found to have only modest benefit [7] . Primary prevention of dementia is therefore of great importance [8] .
Mild cognitive impairment (MCI) is an intermediate state between normal cognitive ageing and dementia [9] . Identification of MCI is thought to be crucial to early intervention. Indeed, in some studies MCI is associated with an increased risk of dementia [10] , as well as with future disability [11] and mortality [12] . Such associations, however, do vary according to the nature of the sample (clinical versus population-based), the case definition of MCI applied, the assessment procedures used for operationalizing component criteria [13] – [15] , and, potentially, the cultural background of participants [16] , [17] . A recent review also suggested that MCI is associated with neuropsychiatric symptoms, cited as being of potential importance for defining subgroups at higher risk of developing dementia in the future [18] .
In community-dwelling older adults the prevalence of amnestic MCI (aMCI), defined according to Petersen's revised criteria [10] , ranges between 2.1% [19] and 11.5% [20] and is most commonly found to be around 3%–5% [21] – [33] with few exceptions in older samples [20] , [34] – [36] . Reports of the community prevalence of aMCI have been predominantly derived from European and North American populations. To our knowledge, very few population-based studies have been published from LAMICs and those from Asia are controversial. Specifically, estimates of aMCI prevalence were similar to those found in Western countries in Kolkata, India (6%) [37] and in Chongqing, China (4.5%) [29] , but higher prevalences were reported by Lee and colleagues in Malaysia (15.4%) [38] and by Kim et al. in South Korea (9.7%) [39] .
Estimating the population prevalence of MCI in LAMICs is a public health priority as rapid demographic ageing is predicted to result in a large majority of people residing in these regions being at risk of dementia and cognitive decline. If so, this will have significant implications with regard to social support and future health care costs, especially as systems are not in place to cope with increased neurodegenerative disease and health resources at present are already extremely limited.
In this study, using data from the cross-sectional phase of the 10/66 Dementia Research Group (DRG) programme on dementia, noncommunicable diseases and ageing in LAMICs [40] , we operationalized the Mayo Clinic–defined aMCI [10] construct and then estimated the prevalence of this condition in eight LAMICs, in addition to its sociodemographic correlates and associations with disability and neuropsychiatric symptoms.
Methods
Ethics Statement
Written informed consent, or witnessed oral consent in case of illiteracy, or next of kin written agreement in case of incapacity, was obtained from all participants. The appropriate Research Ethics Committees at King's College London and at all local countries approved the study protocol and the consent procedures.
Sample
The 10/66 study has been described previously [40] . In brief, the study consisted of a series of cross-sectional one-phase geographic catchment area surveys, carried out in eight urban and rural sites in Peru, Mexico, China, and India, and in three urban sites in Cuba, the Dominican Republic, and Venezuela, between January 2003 and November 2007. The target sample size was 2,000 participants per country, in order to allow estimation of a typical dementia prevalence of 4.5% (SE 0.9%) with 80% power. All community-resident individuals aged 65+ y were eligible for inclusion. Using a process of full household enumeration, all residents aged 65+ y within catchment areas were approached by means of door-knocking and a reliable informant was required for inclusion. Being younger than 65 y was the only exclusion criteria, and weighted sampling procedures were not applied.
Measurements
All participants completed the 10/66 standardized assessment at their place of residence. This consisted of participant and informant interviews and a physical examination, described in full elsewhere in an open-access publication [40] . Participant interviews included questionnaire measures of sociodemographic status, education and childhood environment, social networks and support, self-report measures of common physical disorders, health service use, and lifestyles (smoking, alcohol intake, diet, exercise), in addition to a fully structured diagnostic interview for mental disorder (Geriatric Mental State [GMS], described below). Physical examinations included measures of resting blood pressure, anthropometric measures, and a structured neurological examination. A battery of cognitive assessments was administered (described below) and an informant interview included structured questionnaires on cognitive decline and neuropsychiatric symptoms (both described below), as well as questions on care arrangements, caregiver strain and distress, financial implications of caregiving, and support received. The 10/66 study protocol was translated into Spanish, Tamil, and Mandarin, and minor adaptations were made by local clinicians fluent in English. Validation statistics for the assessments and procedures have been published [41] . The protocol included the GMS Examination [42] , [43] , an informant interview on all participants, a neurological examination, and a neuropsychological battery that comprised the following:
(1) The participant interview section of the Community Screening Instrument for Dementia (CSI “D”) [44] . This was developed as a screening instrument for dementia for use in cross-cultural settings in combination with the informant interview. The cognitive assessment covers multiple domains, including orientation to time and place, language, memory, praxis, and abstract thinking. It deliberately excludes literacy-dependent items. A memory subscale was derived from the CSI “D” using the items addressing immediate and delayed recall of a three word list, recall of the name of the interviewer, and recall of five elements of a short story (logical memory). (2) The Modified Consortium to Establish a Registry for Alzheimer's Disease (CERAD) ten-word-list learning task [45] . Six words: butter, arm, letter, queen, ticket, and grass were taken from the original CERAD battery English language list. Pole, shore, cabin, and engine were replaced with corner, stone, book, and stick, which were deemed more culturally appropriate for all sites in the 10/66 pilot phase (a wider sample that included the survey sites). In the learning phase, the list is read to the participant. Next, the participant is asked to immediately recall the words that they remember. This process is repeated three times, giving an immediate word list memory score, with a maximum total of 30. After a 5-min delay, the participant is again asked to recall the ten words with encouragement but no cues, giving a word list delayed recall score with a maximum total score of 10.
Demographic correlates analyzed against aMCI were age, gender, education, and number of assets. Participants' gender and stated age were recorded. Age was confirmed by the interviewer from official documentation and informant report, and any discrepancies resolved through further questions and clarification and, ultimately, by consensus within the research team. Illiteracy (inability to read and/or write), level of education (none/did not complete primary/completed primary/secondary/tertiary), and number of household assets (car, television, refrigerator, telephone, plumbed toilet, water, and electricity mains) were also recorded.
The impact of aMCI was quantified through investigating associations with disability and neuropsychiatric symptoms. Participant interviews included the 12-item WHO disability assessment schedule (WHODAS-12) [46] , which assesses five activity-limitation domains (communication, physical mobility, self-care, interpersonal interaction, life activities and social participation). Two questions with scores ranging from 0 (no difficulty) to 4 (extreme difficulty) cover each domain, and the global standardized score ranges from 0 (not disabled) to 100 (maximum disability). Details on the WHODAS 2.0 validity and psychometric properties can be found elsewhere [47] , [48] . The informant interview, as well as administering structured CSI “D” questions (regarding decline in memory or intelligence, activities of daily living, social and occupational functioning used for dementia diagnoses—summarized below and applied as an exclusion criteria), also included the neuropsychiatric inventory (NPI-Q) [49] , and the following binary symptom categories were selected for analyses of associations with aMCI: depression, anxiety, apathy, irritability.
For analyses of associations of aMCI with disability and neuropsychiatric symptoms, the following covariates available in the dataset were used for adjusted models in addition to the four sociodemographic variables described above: depression (GMS), self-reported limiting physical impairments (arthritis, visual difficulties, hearing difficulties, respiratory disorders, heart problems, gastrointestinal problems, fainting episodes, limb paralysis, skin disorders), self-reported hypertension, self-reported stroke, psychotic disorder (GMS), self-reported regular pain.
Case Definition of aMCI
Mayo Clinic–defined aMCI was diagnosed on the basis of the following criteria: (1) objective memory impairment beyond that expected for age; (2) subjective memory complaint; (3) no, or only mild impairment in core activities of daily living, and (4) no dementia. Each criterion was operationalized as follows.
Objective memory impairment
A composite memory score was created using results from the memory subscale of the CSI “D” [44] , immediate and delayed word recall scores from the modified CERAD ten-word list [50] . For all tasks impaired performance was defined as a score 1.5 standard deviation (SD) or more below the mean adjusted for age and education. The 1.5-SD definition stems from that applied to define “abnormal memory performance” by Peterson et al. in 1999 [9] , and has been recently recommended also by a National Institute on Aging-Alzheimer's Association workgroup [51] . Operationalization of MCI in other population-based studies has consistently followed this definition [25] , [33] , [52] , [53] , which has also been used to define other constructs such as “Cognitive Impairment No Dementia” [54] . The CERAD word list has been used in previous research [25] . Although, there have been controversies surrounding the MCI entity itself [55] – [58] , they have not to our knowledge focused on the 1.5-SD threshold. Norms were derived from controls without dementia from the 24-centre 10/66 pilot study, which had found minimal geographic variation [41] . Participants were excluded if hearing impairment had prevented cognitive assessment.
Subjective memory impairment
An ordinal scale ranging from 0 to 6 was created by summing item scores from relevant questions in the GMS including: (1) Have you had any difficulty with your memory (0, no; 1, yes)? (2) Have you tended to forget names of your family or close friends/where you have put things (for each question: 0, no/transient; 1, noticed most days per week; 2, noticed daily)? (3) Do you have to make more efforts to remember things than you used to (0, no; 1, yes)? Using this scale, subjective memory impairment was defined as present when an individual scored three or more: the definition that has been used in all previous research to use this scale [59] , [60] .
Normal activities of daily living/instrumental activities of daily living
On the basis of responses from the CSI “D” informant interview, normal activities of daily living (ADL)/instrumental activities of daily living (IADLs) were defined as very mild or no impairment in either carrying out household chores, pursuing hobbies, using money, feeding, dressing, or toileting. The definition of impairment did not include problems arising only from physical impairments.
No dementia
Diagnoses of dementia were applied using the 10/66 dementia algorithm and Diagnostic and Statistical Manual of Mental Disorders IV (DSM-IV) criteria [61] . Participants meeting either criterion were excluded from the analyzed sample (both aMCI cases and controls).
Statistical Analysis
Analyses were carried out on the 10/66 data archive release 2.1. All analyses used STATA version 10.1 [62] . As mentioned above, participants with dementia were excluded from all analyses as has been standard practice in MCI epidemiological research. Sample characteristics across countries were described including age, gender, education, number of household assets, global disability scores (WHODAS-12) [46] , and NPI-Q symptoms [49] .
In order to determine the potential impact of aMCI we assumed that, while both activities of daily living (ADLs) and instrumental activities of daily living (IADLs) would be expected to be intact in people with aMCI, subtle functional impairment may already be present as well as possibly nonspecific and mild behavioral and psychological symptoms of dementia (BPSD) [18] . Zero-inflated negative binomial regression (ZINB) count models were used to assess the association between aMCI and WHODAS-12 disability and NPI-Q scores using identical models to those previously reported for these samples [63] . We used zero-inflated models to deal with skewness in the distribution of the scores characterized by excessive zeros (inflation). The model distinguishes a group whose members have always zero counts (referred to as “certain zero”), from one in which members have either zero or positive counts. ZINB includes a logistic part to model the probability that a zero comes from the first group versus the second group and a negative binomial part to model the counts within the second group. Log-scale coefficients were exponentiated and 95% confidence intervals back-transformed. We determined the appropriateness of the ZINB model against a standard negative binomial model using the Vuong test postestimation and adjusted for the relevant covariates listed above, followed by Poisson regression models to generate prevalence ratios for NPI-Q symptoms as binary-dependent variables. ZINB models were further compared to zero-inflated Poisson models and in every country the test of the dispersion parameter (labelled alpha in Stata and theta by some other sources) was significant at the 0.001 level, indicating ZINB as more appropriate in all cases. Behavioural/psychological outcomes, depression, anxiety, apathy, irritability were modelled separately against aMCI as an independent variable for illustrative purposes, with no attempt to adjust given symptoms for the other three, accepting that these are related constructs.
Prevalence of aMCI was reported for each country by age and gender and adjusted for household clustering. Direct standardization, using the whole sample as the reference population, was used to compare prevalence estimates across countries after adjustment for age, gender, and education. For each country associations with age (continuous variable), gender, education (ordinal variable), and number of household assets (ordinal variable) on aMCI prevalence were calculated using mutually adjusted (as appropriate) prevalence ratios (PRs), with robust 95% confidence intervals (using the “robust” syntax in Stata to take into account household clustering: model robust standard errors [64] , [65] ), using Poisson working models.
To determine the pooled effects for all analyses, the statistical outputs for each country were combined into fixed-effect meta-analyses. Random effect models were not used as we wished to summarise the countries within this study rather than generalise to a hypothetical population of centres. We then calculated Cochrane Q heterogeneity and Higgins' I 2 (95% CIs). The latter statistics set the degree of heterogeneity between studies that is not explained by chance and is expressed as a percentage with values up to 25%, 50%, and over 75% representing mild, moderate, and high heterogeneity, respectively [66] .
Results
The results were derived from a total of 15,376 participants aged 65+ and without dementia across the different countries. Response rates (i.e., participation rates for all potentially eligible residents within the defined geographic catchments) were higher than 80% in all countries. Missing data on the variables of interest were present in less than 1% of the sample. Descriptive data by country are displayed in Table 1 . Age was not evenly distributed across groups (65–69, 70–74, 75–79, and 80+ y) across countries, the samples from Venezuela, China, and India being slightly younger. In all countries more women participated than men. Educational level was highest in Cuba, and the number of household assets was lowest in Mexico and India.
10.1371/journal.pmed.1001170.t001 Table 1
Sociodemographic characteristics of participants by country.
Characteristics
Cuba
Dominican Republic
Peru
Venezuela
Mexico
China
India
Puerto Rico
Sample size (
n
)
2,620
1,767
1,767
1,820
1,821
2,014
1,802
1,765
Response rate (%)
94
95
82
80
85
83
83
93
Age, n (%) – MV
7
0
1
4
1
0
4
0
65–69 y
738 (28.2)
511 (28.9)
538 (30.5)
813 (44.7)
537 (29.5)
683 (33.9)
703 (39.0)
398 (22.6)
70–74 y
739 (28.2)
483 (27.3)
475 (26.9)
450 (24.7)
552 (30.3)
634 (31.5)
604 (33.5)
439 (24.9)
75–79 y
582 (22.2)
345 (19.5)
368 (20.8)
320 (17.6)
384 (21.1)
417 (20.7)
290 (16.1)
436 (24.7)
80+y
555 (21.2)
428 (24.2)
386 (21.8)
236 (13.0)
348 (19.1)
280 (13.9)
201 (11.2)
492 (27.9)
Gender – MV
0
2
0
33
0
0
15
7
Females, n (%)
1,686 (64.4)
1,154 (65.3)
1,073 (60.7)
1,146 (63.0)
1,143 (62.8)
1,128 (56.0)
974 (54.0)
1,183 (67.0)
Educational level, n (%) – MV
8
19
16
40
2
0
2
0
No education
54 (2.1)
314 (17.8)
103 (5.8)
133 (7.3)
459 (25.2)
743 (36.9)
935 (51.9)
47 (2.7)
Some education
548 (20.9)
916 (51.8)
212 (12.0)
408 (22.4)
802 (44.0)
246 (12.2)
411 (22.8)
313 (17.7)
Complete primary
864 (33.0)
338 (19.1)
654 (37.0)
913 (50.2)
337 (18.5)
532 (26.4)
301 (16.7)
356 (20.2)
Complete secondary
681 (26.0)
126 (7.1)
486 (27.5)
262 (14.4)
117 (6.4)
358 (17.8)
110 (6.1)
661 (37.5)
Complete tertiary
468 (17.9)
66 (3.7)
301 (17.0)
92 (5.1)
104 (5.7)
135 (6.7)
43 (2.4)
383 (21.7)
Three assets or fewer – MV
8
5
0
0
0
1
4
0
n (%)
67 (2.6)
256 (14.5)
83 (4.7)
33 (1.8)
373 (20.5)
104 (5.2)
918 (51.0)
4 (0.2)
Neuropsychiatric symptoms,
n
(%)
41
20
11
103
16
3
29
112
Depression
117 (4.5)
220 (12.5)
86 (4.9)
84 (4.6)
73 (4.0)
3 (0.2)
139 (7.7)
36 (2.0)
Anxiety
158 (6.0)
233 (13.2)
199 (11.3)
263 (14.5)
121 (6.6)
7 (0.4)
77 (4.3)
101 (5.7)
Apathy
117 (4.5)
226 (12.8)
93 (5.3)
138 (7.7)
165 (9.1)
15 (0.7)
18 (1.0)
58 (3.5)
Irritability
583 (22.5)
412 (23.3)
381 (21.6)
383 (21.3)
434 (23.9)
26 (1.3)
227 (12.6)
254 (15.2)
WHODAS-12 – MV
11
15
12
96
3
12
4
9
Mean (SD)
9.69 (14.2)
13.91 (17.3)
9.36 (14.3)
9.18 (13.8)
8.59 (15.3)
5.30 (12.0)
17.44 (17.2)
12.13 (16.6)
Mean (SD) omitting zeros
16.55 (15.2)
21.11 (17.3)
15.91 (15.7)
16.18 (14.8)
18.03 (17.9)
18.39 (16.1)
22.19 (16.4)
21.33 (17.0)
MV, missing values; NPI-Q severity: total severity in neuro-psychiatric inventory.
In each country there was a statistically significant zero-inflation in the distributions of WHODAS-12 scores (Vuong test for the whole sample, z = 45.29, p< 0.001) that confirmed the better fit of ZINB over negative binomial alone. Associations between aMCI, disability, and neuropsychiatric symptoms are summarized in Table 2 along with meta-analytical fixed-effect method-pooled estimates, and between-country heterogeneity. After adjustment, disability was significantly higher in aMCI cases compared to the remainder in Peru, India, and Dominican Republic, although was lower in China. The pooled fixed-effect model meta-analytical estimate indicated a positive association with disability although there was moderate to high heterogeneity in these associations between countries. After adjustment aMCI cases were more likely to have informant-rated anxiety, irritability, and apathy symptoms, with no significant between-country heterogeneity. However, there was no overall association with informant-rated depression in pooled estimates although the individual prevalence ratio was significant in Peru.
10.1371/journal.pmed.1001170.t002 Table 2
Association between aMCI and disability (WHODAS-12), and the association between aMCI and neuropsychiatric symptoms (NPI–Q; depression, anxiety, apathy, and irritability).
Analysis
ZINB (95% CI)
Adjusted a PRs (95% CI)
WHODAS-12 a
Depression b
Anxiety
Apathy
Irritability b
Individual study site estimates
Cuba
0.93 (0.74–1.19)
0.96 (0.23–3.93)
1.74 (0.77–3.94)
1.66 (0.59–4.67)
0.84 (0.44–1.57)
Dominican Republic
1.49 (1.08–2.06)
1.04 (0.47–2.30)
1.75 (1.00–3.05)
1.54 (0.76–3.12)
0.98 (0.52–1.82)
Peru
1.51 (1.17–1.94)
2.14 (1.01–4.54)
1.54 (0.89–2.65)
1.38 (0.57–3.33)
1.28 (0.83–1.96)
Venezuela
0.92 (0.53–1.60)
2.14 (0.47–9.74)
2.49 (1.40–4.42)
3.59 (1.94–6.65)
1.74 (1.06–2.86)
Mexico
1.12 (0.78–1.62)
1.07 (0.35–3.29)
1.59 (0.76–3.31)
0.79 (0.35–1.82)
1.11 (0.73–1.69)
China c
0.67 (0.45–0.99)
NC
NC
10.2 (1.40–74.5)
9.90 (2.57–38.0)
India
1.20 (1.03–1.40)
0.69 (0.31–1.53)
0.81 (0.25–2.57)
1.18 (0.13–10.8)
1.27 (0.82–1.98)
Puerto Rico
1.05 (0.87–1.27)
2.60 (0.90–7.54)
1.85 (0.98–3.49)
1.68 (0.65–4.34)
1.04 (0.61–1.76)
Pooled meta-analysis (fixed-effect method)
d
Combined estimate
1.13 (1.04–1.23)
1.31 (0.91–1.89)
1.75 (1.37–2.25)
1.83 (1.33–2.51)
1.24 (1.03–1.49)
Test for heterogeneity p -value
0.008
0.344
0.753
0.091
0.058
I 2 Higgins (95% CI)
63% (20–83)
11% (0–74)
0% (0–71)
43% (0–75)
49% (0–77)
Association between aMCI and disability is measured by exponentiated coefficients from a zero inflated binomial model and representing the increase in disability of aMCI participants compared to normal. Zero inflation fitted using age, gender, educational level, number of household assets, depression, arthritis, visual problems, hearing problems, cough and breathing problems, heart problems, gastrointestinal problems, fainting, limb and skin problems, hypertension and stroke. The association between aMCI and neuropsychiatric symptomsis measured by the risk ratio from a regression using a Poisson working model and model robust standard errors, and representing the risk for having the symptom in aMCI participants compared to normal.
a
Adjusted for age, gender, and educational level, number of household assets and of physical limiting impairments, psychosis, and stroke.
b
Depression and irritability were additionally adjusted for pain. The four NPI–Q symptoms are all associated but in the four models presented in the table we have not adjusted each of them for the other three.
c
China was not adjusted for psychosis
d
The pooled fixed-effect model meta-analytical estimate for depression and anxiety were done without China.
NC, not calculable due to zero cell sizes.
The prevalence of aMCI ranged from 0.8% in China to 4.3% in India, and changed very little after direct standardization for age, gender, and education level, as displayed in Table 3 . Adjusted PRs (95% CI) from Poisson regression models for independent associations with age, gender, education, and assets are shown in Table 4 . No pooled associations were found with age or education but there was a modest association with male gender and fewer assets. Overall little heterogeneity was found between nations in these associations.
10.1371/journal.pmed.1001170.t003 Table 3
Prevalence of aMCI by country, gender, and age group.
Country and Gender
aMCI Prevalence, % (95% CI)
Crude Prevalence (95% CI)
Standardized Prevalence (95%CI) a
65–69 y
70–74 y
75–80 y
80+y
All Age Groups
All Age Groups
Cuba (
n
)
738
739
582
555
1.8 (1.3–2.3)
1.5 (1.0–1.9)
Males
1.5 (0.0–3.0)
1.8 (0.2–3.4)
0.0 (0.0–0.0)
1.7 (−0.2 to 3.6)
—
—
Females
2.7 (1.3–4.2)
2.6 (1.1–4.0)
1.6 (0.3–2.9)
0.8 (−0.1 to 1.7)
—
—
Dominican Rep. (
n
)
511
483
345
428
1.4 (0.9–2.0)
1.3 (0.7–1.8)
Males
1.7 (−0.2 to 3.6)
2.2 (0.0–4.4)
2.7 (−0.4 to 5.7)
2.9 (0.1–5.7)
—
—
Females
0.9 (−0.1 to 1.9)
1.7 (0.2–3.1)
0.4 (−0.4 to 1.3)
0.7 (−0.3 to 1.7)
—
—
Peru (
n
)
538
475
368
386
3.1 (2.3–3.9)
2.6 (1.9–3.3)
Males
5.4 (2.1–8.6)
2.7 (0.3–5.1)
2.1 (−0.3 to 4.5)
4.4 (1.4–7.4)
—
—
Females
2.3 (0.7–3.8)
1.7 (0.2–3.2)
3.6 (1.1–6.0)
3.4 (0.9–5.9)
—
—
Venezuela (
n
)
813
450
320
236
1.2 (0.7–1.7)
1.0 (0.7–1.4)
Males
1.3 (0.0–2.6)
0.0 (0.0–0.0)
2.6 (−0.3 to 5.5)
0.0 (0.0–0.0)
—
—
Females
1.6 (0.5–2.7)
1.4 (0.0–2.9)
1.5 (−0.2 to 3.1)
0.0 (0.0–0.0)
—
—
Mexico (
n
)
537
552
384
348
3.2 (2.4–4.1)
2.8 (2.0–3.6)
Males
3.7 (0.8–6.7)
4.3 (1.5–7.0)
5.1 (1.6–8.6)
4.0 (0.8–7.2)
—
—
Females
1.3 (0.2–2.5)
4.1 (2.0–6.2)
3.9 (1.4–6.5)
1.0 (−0.4 to 2.4)
—
—
China (
n
)
683
634
417
280
0.8 (0.4–1.2)
0.6 (0.3–0.9)
Males
1.0 (−0.1 to 2.1)
0.4 (−0.3 to 1.1)
1.7 (−0.2 to 3.6)
0.0 (0.0–0.0)
—
—
Females
1.3 (0.2–2.4)
0.6 (−0.2 to 1.4)
0.8 (−0.3 to 2.0)
0.7 (−0.6 to 2.0)
—
—
India (
n
)
703
604
290
201
4.3 (3.3–5.2)
4.6 (3.7–5.4)
Males
7.0 (4.1–9.9)
3.8 (1.5–6.1)
4.8 (1.3–8.3)
1.0 (−1.0 to 2.9)
—
—
Females
3.3 (1.5–5.0)
4.4 (2.2–6.6)
5.6 (1.8–9.5)
1.1 (−1.1 to 3.2)
—
—
Puerto Rico (
n
)
398
439
436
492
3.9 (3.0–4.8)
3.0 (2.2–3.8)
Males
3.9 (0.1–7.8)
5.5 (1.7–9.2)
4.1 (0.8–7.3)
5.5 (2.2–8.9)
—
—
Females
4.4 (2.1–6.8)
3.4 (1.3–5.5)
3.5 (1.3–5.6)
2.3 (0.6–3.9)
—
—
a
Direct standardization for age gender and educational level using the whole sample as the standard population.
10.1371/journal.pmed.1001170.t004 Table 4
Mutually adjusted (95% CI) for the independent effects of age, gender, education, and assets on aMCI prevalence.
Analysis
Adjusted PRs (95% CI) a
Age
Gender
Education
Assets
(Per Year Increment)
(Males Versus Females)
(More Versus Less Years)
(More Versus Less)
Individual study site estimates
Cuba
0.97 (0.92–1.02)
0.63 (0.33–1.21)
0.95 (0.72–1.24)
1.52 (1.00–2.30)
Dominican Republic
1.03 (0.97–1.09)
2.25 (1.04–4.86)
1.27 (0.83–1.96)
0.82 (0.63–1.06)
Peru
1.03 (0.99–1.07)
1.29 (0.75–2.22)
1.08 (0.82–1.42)
0.81 (0.64–1.03)
Venezuela
0.95 (0.88–1.02)
0.79 (0.33–1.90)
0.91 (0.55–1.52)
0.97 (0.83–1.14)
Mexico
1.01 (0.97–1.04)
1.57 (0.94–2.60)
1.24 (0.95–1.61)
0.81 (0.69–0.95)
China
0.97 (0.88–1.06)
1.00 (0.40–2.51)
0.86 (0.64–1.15)
0.80 (0.50–1.27)
India
0.97 (0.94–1.01)
1.19 (0.74–1.93)
1.14 (0.89–1.47)
0.85 (0.72–0.99)
Puerto Rico
0.99 (0.95–1.02)
1.46 (0.91–2.33)
1.04 (0.86–1.26)
0.94 (0.70–1.27)
Pooled meta-analysis (fixed-effect method)
Combined estimate
0.99 (0.98–1.01)
1.25 (1.01–1.54)
1.06 (0.96–1.16)
0.88 (0.82–0.95)
Test for heterogeneity
0.209
0.25
0.619
0.168
Higgins (95% CI)
27% (0–67)
23% (0–64)
0% (0–68)
33% (0–70)
a
Mutually adjusted for age, educational level, gender, and number of assets as appropriate.
Discussion
Using data from a large series of cross-sectional surveys applying standard sampling and measurements, we estimated the community prevalence of Mayo Clinic–defined aMCI in six countries in Latin America, China, and India. To our knowledge this is the first study to attempt to make direct comparisons of prevalence estimates of aMCI across diverse cultures and world regions. Differences in prevalence between countries were marked and ranged from 0.8% (China) to 4.3% (India), i.e., greater than five-fold variation. After direct standardization for age, gender, and education, using the whole population as the reference, these differences were not markedly attenuated.
Inconsistencies in aMCI prevalence observed between the 10/66 study centres are likely to be due to components of the aMCI diagnosis itself. In a cross-cultural context, these support questions previously raised concerning its conceptual basis [67] and/or operationalization outside clinical settings [68] . However, aMCI has been reported to be associated with increased mortality in a prospective study [12] , and differences in aMCI-associated survival between country sites cannot be excluded as a factor influencing variation in prevalence. It should be noted that the 10/66 dementia diagnosis showed much higher sensitivity than the Diagnostic and Statistical Manual of Mental Disorders IV (DSM-IV) criteria in both pilot and clinical validation 10/66 studies [41] , [61] . Compared to numerous aMCI prevalence reports from community-based sites in Finland (5.3%) [26] , Italy (4.9%) [69] , Japan (4.9%) [32] , the US (6%) [30] , South Korea (9.7%) [39] , Malaysia (15.4%) [38] , and India (6%) [37] , both the crude and adjusted aMCI prevalence reported here are relatively low. However, the estimates are similar to those reported by the British MRC CFAS study (2.5%) [15] and to estimates for aMCI prevalence in community samples from Southern France (3.2%) [33] , the US (3.8%) [25] , and Germany (3.1%) [70] . Low aMCI prevalence in our Latin American sites contrast with the aMCI prevalence (ranging between 3.8% and 6.3% depending on age) reported amongst American Caribbean Hispanics [31] . Differential mortality may explain these differences, but a potential role of the environment and lifestyle in the increased risk of MCI amongst Hispanic immigrants in North America cannot be excluded. Crude aMCI prevalence in India (4.3%) is similar to the figure described by Das and colleagues in Kolkata [37] . Prevalence in China was the lowest (0.6%), similar only to that described in the VITA study in Vienna [27] and markedly lower than that reported in a recent study from Chongqing (4.5%) [29] . Overall, the results suggest that there is very little consistency in prevalence of aMCI across world regions. When considered between studies, this may well reflect diagnostic issues arising from a lack of specific criteria for the operationalization of MCI (i.e., cognitive batteries and specific cut-off scores for impairment) as well as unmeasured differences and cultural variations potentially relevant for some components of the aMCI construct (such as subjective memory impairment, as described below). The objective for the analyses here was to standardize the assessments as much as possible in order to gain a clearer idea of international variation. The fact that substantial heterogeneity remains suggests important variation in constructs underlying the definition. These will be considered further below.
Female gender, increased age, lower education, and lower socioeconomic status are associated with dementia [71] and have been described in association with MCI [31] . In our study, however, the effects of age and education on aMCI prevalence were negligible across study sites, with no between-country heterogeneity in this respect. It is important to bear in mind that age- and education-standardised normative data were used to define aMCI and the lack of association supports the robustness of the norms, although for education, it might also reflect lower variance in the exposure or weaker underlying associations between education and other risk factor profiles in these samples. Lower socioeconomic status remained associated with aMCI and this may be an additional marker, beyond education, of relevant social disadvantage. The observed association with male gender contrasts with the higher reported age-adjusted prevalence of dementia in women compared to men [71] , but could reflect the effect of dementia case exclusion consistent with Mayo Clinic Study of Aging reports that women experience a transition from normal cognition directly to dementia at a later age but more abruptly [20] .
As described earlier, a key consideration with aMCI applied as a construct in international research is its cross-cultural validity. An advantage of the 10/66 study was that identical measures were taken and identical algorithms applied for diagnosis across the study sites and the protocols for cognitive assessments in the 10/66 study were the result of a long and painstaking process of development and validation [41] . However, a construct such as subjective memory impairment is potentially subject to cultural influences and may underlie between-site variation. For example, between sites, people with objectively lower performance on cognitive assessments may be more or less likely to admit to memory difficulties. Since this is a component of the most commonly used definitions of aMCI/MCI, these cultural variations may be reflected in differing prevalences. However, despite the differences in prevalences of aMCI between sites, associations with disability were relatively consistent, providing support for the cross-cultural applicability of the aMCI construct. They did not suggest, for example, that only more severe forms of aMCI were being identified in China where prevalence was lowest, compared to India where it was highest (particularly since disability was lower rather than higher in China in those with aMCI compared to the remainder of the sample). Associations between aMCI and disability should be viewed with caution since activities of daily living impairment is an exclusion criterion for the former. Lower likelihood of reporting difficulties in China would be unlikely to account for the negative association observed between aMCI and disability in that site because under-reporting would have to be differential between those with/without aMCI. There is very limited evidence from population-based studies on the occurrence and characteristics of neuropsychiatric symptoms that may accompany MCI [18] . While we did not find any association between aMCI and depressive symptoms, our findings of a significant association between aMCI and anxiety, apathy, and irritability are largely consistent with those from the Cardiovascular Health Study and the Mayo Clinic longitudinal study on aging in the US [72] , [73] , the Kungsholmen study in Sweden [74] , and a small study from Thailand [75] . However, it should be borne in mind that individual behavioural/psychological symptoms were not mutually adjusted as outcomes and the independence of observed associations in Table 2 cannot be assumed.
Strengths of the study include the very large sample size and the wide range of populations sampled in terms of culture, economy, and population characteristics. Moreover, internal validity was maintained through rigorously prevalidated and standardised measurements applied consistently between countries in addition to common algorithms used to define aMCI. There are some limitations. The samples were drawn from specific geographic catchment areas and cannot be assumed to be representative of the source nation/site. No attempt was made to differentiate urban and rural status in this analysis because not all sites recruited from both settings. The study was cross-sectional in design and the impact of survival cannot be evaluated. Furthermore, within the aMCI category, participants who had developed this late in life could not be distinguished from those for whom it was a stable lifetime trait. Finally, aMCI diagnosis was determined without clinical judgement, which is difficult to obtain in large population-based studies and unfeasible in most of our study sites. Although aMCI was originally derived as a diagnosis for secondary or tertiary care clinical settings, it is being increasingly applied in epidemiological research and data from community samples is an important supplement, particularly if future community-level interventions are planned to prevent progression to dementia. Our analysis here is intended to extend this particular evidence base. Follow-up is currently underway in most 10/66 sites, which will provide further data on predictive validity.
This is one of the first studies, to our knowledge, to investigate the prevalence of aMCI in LAMICs, where the large majority of older people and people with dementia currently live [3] , [4] . Longitudinal data are needed to clarify further the predictive validity of the aMCI case-definition applied here and to evaluate the extent to which it can be applied as a risk marker for further cognitive decline or dementia. In addition, further evaluation is needed of the associations with disability and neuropsychiatric symptoms since our findings do suggest higher than expected comorbidity and there are large absolute numbers of older people with aMCI in these rapidly ageing and populous world regions.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Acetabular reconstruction for patients with massive acetabular defects remains one of the most difficult problems in revision total hip arthroplasty (THA) [ 1 , 2 ]. Several factors may cause the loss of acetabular bone, these include the earlier bone removal to accommodate the socket in primary THA, micromotion between the original socket and the pelvis and, lysis caused by wear particles. Acetabular reconstruction for patients with massive acetabular defect makes it difficult to place the new prosthesis in an optimal location biomechanically to achieve sufficient strength for long-term secure fixation. An acetabular reinforcement device (ARD, antiprotrusio cage) is frequently used as a load-sharing device to provide initial stability for the reconstruction and to allow the incorporation of the allograft into the host pelvis without stress [ 3 ] ( Fig. 1 ). The application of ARD not only restores the physiological hip function but also prepares the bone stock for future re-revision [ 1 , 4 , 5 ]. The key to a successful reconstruction is robust fixation of the device to the host acetabulum, which is often very deficient in the revision setting. Inadequate interfacial stability often leads to loosening of the device, absorption of the allograft, and failure of the surgery [ 6 , 7 ].
10.1371/journal.pone.0121588.g001
Fig 1
A schematic drawing showing the acetabular reconstruction for a pelvis with a massive acetabular defect.
An acetabular reinforcement device (ARD) is used to provide the initial stability of the reconstruction and to allow the incorporation of the allograft into the host pelvis without stress.
The stability of the initial fixation of ARDs in revision THA is critical. In particular, it is difficult to achieve stability when the host bone quality is poor or large bony defects are present. Unfortunately, such clinical situations are often encountered in revision procedures [ 8 , 9 ]. The currently commercially available ARDs are all designed with standard (non-locked) screw fixation. The fixation stability of a standard screw mainly relies on the initial screw/bone anchoring strength. As long as the screws are secured, the fixation stability solely depends on the contact friction between the screw head and the implant. The inability of standard screws to resist any external forces acting on the implant in poor-quality bone may create a risk of implant loosening when the contact pressure is low.
Locking screw and plate systems with different designs have long been developed for orthopedic trauma surgery. The earlier age of the locking screw/plate devices consisted of monoaxial locking design that have been extensively studied both clinically and biomechanically [ 10 – 13 ]. The monoaxial devices allow for the fixed-angle locking of screws via the threads in the screw head and the plate in an orthogonal manner. The design of these systems results in a rigid screw/plate interface, thereby increasing the force required to displace the construct. Consequently, the monoaxial locking mechanism sacrifices, to some extent, angularity and compression on bone/plate interface to achieve a higher stiffness on screw/plate interface ( Fig. 2 ). The monoaxial locking devices were later developed to allow polyaxial locking, with different fracture patterns and bone structures taken into account [ 10 ]. The polyaxial locking system enables the fixation screw to be inserted at a variable angle from the perpendicular aspect. Consequently, the polyaxial locking system allows the surgeon to sense the bone quality and to change the screw direction if required. To date, this new generation of polyaxial locking systems has been widely applied for fracture osteosynthesis [ 14 , 15 ].
10.1371/journal.pone.0121588.g002
Fig 2
A schematic drawing showing standard compressive screw/plate (left) and monoaxial locking screw/plate (right) systems.
For the locking screw/plate system, following the insertion of fixation screws, the screw/plate can be considered an integral device that leads to an improvement in the initial stability at the plate/bone interface.
Although locking screw systems have been extensively used in trauma surgery to help enhance bone fixation in osteopenic patients, the usefulness of locking screws for fixation of ARDs has not been investigated so far. In the present study, we hypothesized that a locking mechanism for ARDs might provide better initial stability to allow the incorporation of the allograft into the host pelvis without stress. The compression and angular stability of pelvis reconstructed with novel ARDs equipped with either monoaxial or polyaxial locking fixation mechanisms were tested, and their biomechanical properties were compared with those of standard fixation mechanisms. After evaluation of the biomechanical characteristics, this study might provide a preliminary evaluation of biomechanical performance for ARDs with or without interlocking mechanisms. Future investigations such as finite element analysis and actual clinical trial deserve to be conducted in future studies.
Materials and Methods
Synthetic bone model
Synthetic composite pelvis (model #3409, Large Right Fourth Generation Composite Hemi Pelvis, Pacific Research Laboratory Inc., Vashon Island, WA, USA) was chosen as testing objects. Synthetic bones are commonly used when imitation of the actual strength properties of real bone is required, and they are suitable for testing, comparing, or designing implants and other devices. Consequently, synthetic bones are suitable for a variety of mechanical experiments when cadaver bones cannot be obtained. To simulate pelvic discontinuity, the standard synthetic pelvis was mounted on a custom-made jig; an oscillating saw was then used to create the standardized transverse gap defect through each acetabulum. By using this method, the level of the transverse osteotomy was kept identical in all specimens.
Acetabular reinforcement devices (ARDs)
Three types of stainless ARDs with a uniform thickness of 3 mm were custom-manufactured using a computer numerically controlled (CNC) machine based on ZCA Acetabular Reconstruction Cages (EDI Code: E1141004, Zimmer Inc., Warsaw, Indiana, USA). The ARDs made of 316L stainless-steel were manufactured with different fixation mechanisms: standard (non-locked), monoaxial, and polyaxial locking. The screw holes of standard ARDs were smooth without threads ( Fig. 3A ) and fixed with standard compression screws ( Fig. 3B ). To create the monoaxial locking ARD, the standard ARD was modified with a monoaxial locking mechanism: the head of the locking screw was equipped with a thread, and a matching thread was made in the ARD in a unidirectional manner ( Fig. 4 ). To create the polyaxial locking ARD, the standard ARD was modified with a polyaxial locking mechanism based on the non-contact bridging plate (NCB plate, Zimmer Inc., Warsaw, Indiana, USA). The head of the polyaxial locking screw was contoured to fit congruently into the reciprocal hole, and screw locking was achieved through the use of a locking cap that was threaded into the screw holes ( Fig. 5 ).
10.1371/journal.pone.0121588.g003
Fig 3
A schematic drawing (left) and a photograph (right) showing (A) a standard ARD and (B) a 5.0-mm compression screw.
The screw hole in the ARD is smooth (without threads) and it can be fixed with standard compression screws.
10.1371/journal.pone.0121588.g004
Fig 4
A schematic drawing (left) and a photograph (right) showing (A) a monoaxial locking ARD and (B) a 5.0-mm locking screw.
The locking screw has a thread on the screw head, and the ARD has a matching thread. The system allows fixed-angle locking of screws through the fine threads in the head and ARD in a unidirectional manner.
10.1371/journal.pone.0121588.g005
Fig 5
A schematic drawing (left) and a photograph (right) showing (A) a polyaxial locking ARD and (B) a 5.0-mm screw and a locking cap. (C) The system allows the screw to be inserted at a variable angle from the perpendicular aspect and finally locked by a cap.
The head of the polyaxial locking screw is contoured to fit congruently into the reciprocal hole, and screw locking is achieved with a locking cap threaded into the screw holes.
The fixation screws used for the three types of ARDs were all 5.0 mm in diameter and 31.0 mm in length, with a thread depth of 0.3 mm and a thread pitch of 0.75 mm. The following three combinations of ARDs and screws were tested: (1) standard ARD attached with standard compression screws tightened to 5.0 N·m (no locking); (2) monoaxial interlocking ARD attached with monoaxial screws tightened to 5.0 N·m (monoaxial locking); and (3) polyaxial interlocking ARD attached with polyaxial screws tightened to 3.0 N·m and locked with a locking cap tightened to 5.0 N·m (polyaxial locking).
Each test was conducted on a new synthetic pelvis. The ARD was implanted so that the peripheral flanges were flush with the rim of the acetabulum. All screws were implanted into the corresponding ARDs with orthogonal insertion using a drilling guide ( Fig. 6A ). The drilling guide had a thread matching the ARD thread. It allows fixed-angle placement of screws through the fine threads and ARD in a unidirectional manner ( Fig. 6B ). Prior to the insertion of fixation screws, a pilot hole with a diameter of 4.5 mm was drilled orthogonally to the screw hole as determined by the drilling guide. All screws were tightened with a calibrated torque limiting screwdriver set to 5.0 N·m. Following the implantation of the ARD, each specimen was secured in a rectangular metal frame with acrylic (AcryliMet, South Bay Technology Inc., San Clemente, CA, USA) to constrain the movement of ARD during compression or torsion test. The prepared constructs were then mounted on a biaxial servohydraulic material testing machine (Bionix 858, MTS Corp., MN, USA) to compare the relative construct stabilities among the three different types of ARDs. All tests were performed at room temperature (22°C)
10.1371/journal.pone.0121588.g006
Fig 6
(A) Drilling guide. (B) Drilling guide fixed on ARD for orthogonal drilling of the pilot hole.
The drilling guide has a thread matching the ARD thread. It allows fixed-angle placement of screws through the fine threads and ARD in a unidirectional manner.
Compression test
A total of 60 pelvises were used (45 for compression and 15 for torsion). In compression, 45 pelvises were divided into three groups (standard, monoaxial and polyaxial). For each ARD group, 15 pelvises were divided into three subgroups with different fixation screw numbers (5 pelvises in each subgroup). Compression stiffness of ARDs secured with different numbers of screws (4, 5, and 6 screws) on the flange of the pelvis was examined ( Fig. 7A ). Each prepared construct was secured in the rectangular metal frame as previously described. The rectangular frame was then clamped on a custom-made vice capable of angle adjustment, which is connected to the load cell on the lower side of the MTS frame. The angle of the custom-made vice was then adjusted so that the acetabulum was positioned with a 45° inclination ( Fig. 7B ). An upside-down positioned stem with a 32 mm metal ball was used as the plunger, clamped on the upper side of the MTS wedge grip and connected to the actuator. After positioning the construct, an axial compressive force was applied at a constant crosshead rate of 1 mm/min. The relationship between force and displacement was continuously recorded in 0.1-mm increments (sampling rate: 0.17 Hz) using the MTS Teststar II software. The experimental setup and testing configuration are shown in Fig. 7B .
10.1371/journal.pone.0121588.g007
Fig 7
(A) ARDs secured with 4, 5, and 6 fixation screws on the flange of a pelvis. (B) The experimental setup of the compression test.
The prepared construct was secured so that the acetabulum was mounted with a 45° inclination on the lower side of the MTS frame and connected to the load cell. An upside-down positioned stem with a 32-mm metal ball was used as the plunger, clamped on the upper side of the MTS wedge grip and connected to the actuator.
Torsion test
In torsion, 15 pelvises were divided into three groups (standard, monoaxial and polyaxial). All 15 pelvises were secured with 4 fixation screws. Torsion stiffness of ARDs that were secured with 4 fixation screws on the flange of the pelvis was examined ( Fig. 8A ). Each prepared construct was secured with the acetabulum placed horizontally. A square bar was used as the transmission shaft; it was clamped on the upper side of the MTS wedge grip and connected to the MTS actuator. For all ARD specimens, a square hole was specially made to prevent the square bar from sliding during the torsion test. After the construct was clamped, the torsion test was performed to measure the magnitude of torque by rotating the MTS actuator at a constant rate of 0.1 degree/sec. The relationship between torque and the rotational angle was simultaneously recorded in 0.1-degree increments (sampling rate: 1 Hz). The experimental setup and testing configuration are shown in Fig. 8B .
10.1371/journal.pone.0121588.g008
Fig 8
(A) ARDs secured with 4 fixation screws on the flange of a pelvis. (B) The experimental setup of the torsion test.
The prepared construct was secured with the acetabulum in a horizontal position. A square bar was used as the transmission shaft; it was clamped on the upper side of the MTS wedge grip and connected to the MTS actuator. For all ARD specimens, a square hole was specially made to prevent the square bar from sliding during the torsion test.
Statistical analysis
Means and standard deviations were calculated for descriptive purposes. Multiple comparisons among three different designs of ARD (standard, monoaxial and polyaxial) and different fixation screw number (4, 5 and 6) were performed using a two-way ANOVA test (Minitab 15, Minitab Inc., US), with the significance level set at P = 0.05.
Results
Compression test
Fig. 9A shows a typical force vs. displacement curve for the compression test. Compression stiffness was defined as the slope of the curve at the initial linear phase (straight line passing through the value of force at 0.2 mm in displacement). The mean compression stiffness of various ARDs secured with 4, 5, and 6 screws is shown in Table 1 and Fig. 9B . The results indicated that, regardless of the ARD type (monoaxial, polyaxial, or standard), the compression stiffness increased with increasing screw number, the exception was ARDs secured with 4 and 5 screws in the standard group ( P > 0.05). Additionally, the monoaxial ARDs with the same fixation screw numbers (4, 5, or 6) had the highest stiffness, whereas the standard ARDs had the lowest. The exceptions were that no significant difference was found between 4-screw ARDs of the standard and polyaxial groups ( P > 0.05), and 5-screw ARDs of the polyaxial and monoaxial groups ( P > 0.05).
10.1371/journal.pone.0121588.g009
Fig 9
(A) A typical force vs. displacement curve at the initial linear phase of the compression test (0–1 mm). (B) The average compression stiffness of various ARDs secured with 4, 5, and 6 screws determined in the compression test.
The monoaxial ARD demonstrated the highest compressive stiffness, whereas the standard ARD had the lowest. The groups without significant differences are indicated with ‘‘NS”.
10.1371/journal.pone.0121588.t001
Table 1 The average compression stiffness and torsion stiffness of various ARDs secured with different screw number. Torsion stiffness of ARDs without screw number of 5 and 6 are indicated with “NA”.
ARD Types
Screw Number
Compression Stiffness (N/mm)
Torsion Stiffness (N·m/degree)
Standard
4
381.4 ± 117.2
6.9 ± 2.1
5
487.4 ± 92.4
NA
6
684.2 ± 65.5
NA
Polyaxial
4
552.8 ± 62.8
11.4 ± 1.3
5
765.5 ± 132.4
NA
6
832.7 ± 24.2
NA
Monoaxial
4
672.6 ± 84.1
13.3 ± 1.0
5
838.6 ± 99.8
NA
6
1,012.1 ± 111.2
NA
Torsion test
Fig. 10A depicts a typical torque vs. rotation angle curve for the torsion test. Torsion stiffness was defined as the slope of the curve at the initial linear phase (straight line passing through the value of torque at 0.2 mm in displacement). The mean torsion stiffness of various ARDs is shown in Table 1 and Fig. 10B . Significant differences in torsion stiffness were found among three groups ( P < 0.05). The stiffness of the monoaxial locking construct (mean value: 13.3 ± 1.0 N·m/degree) was 1.18 times higher than that of the polyaxial locking construct (mean value: 11.4 ± 1.3 N·m/degree) ( P < 0.05). The stiffness of the polyaxial locking constructs was 1.65 times higher than that of the standard non-locked constructs (mean value: 6.9 ± 2.1 N·m/degree) ( P < 0.01).
10.1371/journal.pone.0121588.g010
Fig 10
(A) A typical torque vs. angle curve at the initial linear phase of the torsion test. (B) The average torsion stiffness of various ARDs secured with 4 screws determined in the torsion test.
The monoaxial ARD demonstrated the highest torsion stiffness, whereas the standard ARD had the lowest. Significant differences in torsion stiffness were found among the groups ( P < 0.05).
Discussion
Revision THA with severe acetabular bone loss is a reconstructive challenge. Clinically, the most common treatment for acetabular bone loss in revision hip arthroplasty is bone grafting with a bulk allograft that is then stabilized with an ARD. Previous studies recommend using ARDs to protect bulk bone grafts in load-bearing defects [ 16 – 18 ]. The use of an ARD is proposed to improve osseointegration because it protects the underlying graft from excessive mechanical stress while supporting the cup, restoring the limb length, and maintaining better bone stock for future revision [ 4 , 5 ]. Unfortunately, the use of allografts has shown poor clinical outcome [ 19 – 21 ]. The failure of the allograft treatment may be attributed to inadequate fixation and excessive stress on the allograft [ 22 , 23 ]. Initial stability of the ARD is crucial for improving the survival rate because excessive motion at the bone/implant interface may result in the eventual failure of osseointegration, which reduces the success rate of acetabular reconstruction [ 22 , 23 ].
Extensive studies have described the mechanical roles of ARDs in revision THA [ 24 , 25 ]. However, studies addressing the effectiveness of load sharing mechanisms using the concept of locking mechanisms are lacking. According to our results, the standard compression screws exhibited the lowest construct stability in terms of resistance to compression and angular forces. Standard compression screws achieve their fixation based solely on the initial screw/bone anchoring strength. When these screws are secured, they rely on contact pressure between the screw head and the implant to resist external translation and angulation. The compression load from the screw head is transferred only to the bone that is engaged by the screw threads [ 26 ]. Consequently, standard compression screws are incapable of producing significant resistance to any external forces acting on the implant in poor-quality bone. The application of standard compression screws in ARD can cause a severe problem in cases with poor bony purchase secondary to either poor-quality bone or a large defect. In contrast, the results of both the compression and torsion tests revealed that the monoaxial locking mechanism exhibited the highest acetabular construct stability; therefore, it provides the most robust support for the acetabular socket and allows underlying bone grafting in an environment that is protected from excessive stress. However, the monoaxial locking screws were inserted in a unidirectional configuration, which may restrict the ability of the surgeon to determine screw placement. This can result in a problem when there is either poor-quality bone or a large defect present in the screw placement path. Our results indicated that the polyaxial locking screws exhibited construct stability intermediate between the monoaxial and standard screws. The polyaxial locking screws were considered beneficial because they allowed screw insertion at a variable angle while providing compression in the same manner as a standard screw prior to insertion of the locking cap. The use of polyaxial locking screws might resolve the problem of determining the screw insertion angle in relation to good-quality host bone.
Locking screws have widely been used in orthopedic trauma surgery but have had limited application in arthroplasty surgery. In the authors’ opinion, in contrast to ARDs in acetabular reconstruction, the interlocking mechanism is inappropriate for fixation of the acetabular cup in primary cementless THA. For cementless THA, compression is necessary to maximize the contact between the bone and the cup, which increases the friction between the two surfaces and thereby potentially facilitates bone ingrowth. In such cases, the application of an interlocking mechanism between the screw and the cup would increase the screw/cup interfacial stability. However, this mechanism causes an enormous reduction of bone/cup interfacial pressure. In extreme conditions, the screw locking mechanism may even create a gap between the cup and the bone, leading to the failure of bone ingrowth into the acetabular cup.
Numerous reports utilizing synthetic pelvis to investigate the mechanical stability of postoperative pelvis for acetabular reconstruction have demonstrated that the synthetic pelvis is a good alternative for in vitro experiments when human pelvis cannot be obtained [ 25 , 27 ]. Gililland et al. [ 25 ] used synthetic pelvis to evaluate the mechanical stability of three types of acetabular reconstruction constructs: a cup-cage construct, a posterior column plate construct, and a bicolumnar construct. Their results demonstrated that the bicolumnar construct provided improved component stability. Recently, Milne et al. [ 27 ] compared polyaxial compression locking screws with non-locked and cancellous screw constructs for acetabular cup fixation using synthetic pelvis. Their results revealed that polyaxial locking compression screws significantly improved the construct stiffness compared to non-locked or cancellous screws. They concluded that the increase in construct stiffness will likely reduce interfacial micromotion. In the present study, the synthetic pelvis was chosen as a substitute for human pelvis because of the difficulty of obtaining human pelvis and reliability of results acquired using specimens with uniform properties.
Our study has limitations. First, the ARDs were manufactured in house using only one type of metal, shape, and surface finish. Possible effects of variations in the above-mentioned properties were not considered. Second, a synthetic pelvis was used as a substitute for human bone. Although synthetic pelvis provides a platform for comparison of fixation stability, there must be some differences in mechanical characteristics between the synthetic pelvis and actual bone; therefore, extrapolation of our results for clinical application should be done with caution. Third, the evaluation of compressive and angular stiffness did not account for the surrounding muscle forces, which may impact the clinical relevance of the results. However, we believe that the results of compressive and angular stiffness provide a preliminary platform for comparison of the postoperative stability of ARDs with various locking designs in acetabular reconstruction. Fourth, a square bar was used as the transmission shaft for torsion test. During torsion test, the angular distortion of the square bar might affect the measured torsion stiffness of the ARD construct. However, torsion stiffness of all ARD constructs were tested in a reproducible manner using an identical stiffed square bar, and we believe that this study provides a comparison of the mechanical performance of various ARDs in an artificial pelvis. Fifth, this was an in vitro analysis of specimens prepared in a laboratory environment, which did not take into account the effects of temperature and body fluids on compression and torsion tests. Finally, only static loading (compression and torsion in the synthetic pelvis) was conducted without consideration of other types of physiological loading. In real-life situations, the screw/bone, screw/implant and implant/bone interfaces are subjected to dynamic multi-directional loading. Although our loading mode does not necessarily represent the actual physiological loading conditions, all specimens were prepared and tested in a uniform and reproducible manner, and we believe that this study provides information that can be useful to orthopedic surgeons performing acetabular reconstruction. Further investigation of the effects of other loading methods, such as dynamic fatigue testing, might be necessary in the future.
Conclusions
Equipping ARDs with interlocking mechanisms effectively improves the initial stability at the device/bone interface compared to standard non-locked ARDs. Our study demonstrates the potential benefit of adding locking mechanisms to ARDs. Monoaxial ARDs provide the most robust support for the acetabular socket. However, the monoaxial locking screws are inserted in a unidirectional configuration, which may restrict the ability of the surgeon to determine screw placement. Polyaxial ARDs provide the surgeon with more flexibility in placing the screws at the cost of reduced mechanical performance. This in vitro study provides a preliminary evaluation of biomechanical performance for ARDs with or without interlocking mechanisms, future investigations such as finite element analysis and actual clinical trial deserves to be conducted in future studies.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Mammography and ultrasound are established tests in the diagnosis of breast cancer [ 1 – 3 ]. Still, these imaging modalities regularly yield inconclusive findings where the presence of breast cancer cannot be confirmed or excluded. In these cases, image-guided biopsies are performed to establish a diagnosis. Biopsy, however, is challenging if findings are difficult to localize, as is the case in architectural distortions or diffuse and multiple lesions. In this uncommon but diagnostic challenging setting, breast MR imaging can be used as a problem-solving tool, which can either help guide or avoid biopsies due to its high sensitivity and soft tissue contrast. This indication for MR imaging is, however, controversial in the imaging community [ 4 – 7 ]. While breast MR imaging provides a very high sensitivity and negative predictive value, particularly in non-calcified breast lesions [ 5 , 8 ], a recent meta-analysis highlighted important research gaps: problem-solving definitions are not well-defined and the empirical evidence about specific indications, such as architectural distortions, is sparse [ 5 , 9 , 10 ]. In addition, it remains unclear whether MR imaging actually facilitates the work-up of inconclusive cases or not. Additional MR imaging findings that are unrelated to the reason for the original referral require further work-up, which is justified if additional cancer lesions are detected, but is ultimately unnecessary in case of benign lesions.
Consequently, the purpose of this study was to evaluate the diagnostic performance and incidental lesion yield of breast MR imaging if used as a problem-solving tool and align these results with original referral reasons.
Materials and methods
Case selection and reference standard
Eligible for this retrospective, single-center, IRB-approved study were women who were consecutively referred to our institution (Diagnostikum Graz), an independent cross-sectional imaging centre receiving outpatients from multiple referring physicians. Patients were referred for breast MR examinations due to findings on digital mammography and/or ultrasound between March 2013 and December 2014. Specifically, we included those patients in that according with our national health care system regulations received an interim BI-RADS 0 category (further imaging assessment required) assignment. This problem-solving indication includes a variety of findings such as discrepancies between MG and US such as asymmetric densities without US correlate, lesions with discrepant size in both modalities, lesions with equivocal morphology in either both or one of these modalities and multiple lesions. The indication for MR imaging in this setting is assessed by the representative physicians of the medical authorities, in compliance with national health regulations. This study was conducted according to STARD ( S1 File ).
Included in this study were those women who had a reference standard of either histopathology or imaging follow-up at 24 months. Histopathological diagnosis was established either by image-guided biopsy (ultrasound-guided core biopsy or vacuum-assisted biopsy under mammography/MR imaging guidance) or open surgery. Experienced breast pathologists performed the breast tissue specimen work-up. Lesions classified as benign by imaging or histopathology were followed up by imaging (mammography, ultrasound, or MR imaging, as appropriate) for at least 24 months. Excluded were patients with contraindications against MR imaging or incomplete MR imaging scans.
MR imaging
MR imaging was performed on a 3T magnet (Magnetom Skyra, Siemens Medical Solutions®, Erlangen, Germany) using the vendor-supplied, 16-channel dedicated breast coil. Breast MR imaging examinations were generally scheduled in the second week of the menstrual cycle in premenopausal women. Menopausal patients who were receiving hormonal replacement therapy were requested to cease treatment one month before the examination [ 11 ].
The imaging protocol encompassed an axial T2w-TSE sequence (Turbo Spin Echo DIXON fast, TR 6500 ms, TE 81 ms, flip angle 120°, spatial resolution 4 mm 3 , 35 slices, time of acquisition 2:10 minutes), and a readout-segmented, multi-shot echo planar imaging-based, diffusion-weighted imaging sequence (RESOLVE, TR 5500 ms, TE1 56 ms, TE2 88 ms, b-values 0 and 800 s/mm 2 , spatial resolution 1.9 x 1.9 x 5 mm 18 mm 3 , 28 slices, no interslice gap, three orthogonal directions, one average, acquisition time 3:36 min). The scanner software (Syngo MR E11, release number: N_4VE11C) automatically calculated the Apparent Diffusion Coefficient (ADC) maps. T1-weighted images were acquired as follows: FLASH 3D; SPAIR fat saturation; TR 4.89 ms; TE 1.81 ms; flip angle 10°; spatial resolution 0.9 x 0.9 x 1.8 mm 1.5 mm 3 ; and time of acquisition 1:10 minutes per measurement. These were obtained once before and four times after the intravenous injection of 0.1 mmol/kg gadoteridol (Prohance®, BRACCO, Milano, Italy). Between postcontrast measurements 3 and 4, an interleaved, isotropic, high-resolution T1w sequence was acquired (FLASH 3D, SPAIR fat saturation, TR 7.33 ms, TE 3.73 ms, flip angle 15°, spatial resolution 0.9 mm isotropic, time of acquisition 2:26). Image subtractions were calculated in-line by the scanner software (Syngo MR E11, release number: N_4VE11C). Overall acquisition magnet time for this protocol was less than 15 minutes.
MR image interpretation
All imaging data was read during routine clinical practice by one of four board-certified radiologists with >10 years of experience. Results were saved in a prospectively populated database within our institution´s electronic information system. Image interpretation was performed according to the American College of Radiology BI-RADS® lexicon using morphologic and dynamic enhancement criteria [ 12 ] before any histopathological sampling. Signal intensity time curves were measured by Regions-of-Interest placed in the most enhancing part of the lesion [ 13 ]. Image interpretation during routine clinical practice considered the results of prior images as this facilitates image interpretation [ 14 ]. In addition, lesions showing high ADC values were considered benign as suggested in the literature [ 15 – 18 ]. Lesions without contrast-enhancement on MRI were generally considered benign.
Data analysis
Data were extracted from our institutional prospectively populated database into a computerized spreadsheet (Excel: Microsoft, Redmond, WA).
Receiver-Operating-Characteristics (ROC) analysis was performed using BI-RADS as the classification variable and final diagnosis, based on the reference standard (benign versus malignant), as the reference variable. Lesions were considered malignant if image-guided biopsy or surgery, or both, confirmed invasive carcinoma or DCIS. The reference standard for benign lesions was histopathology (biopsy and/or surgery) and imaging follow-up of at least two years or only imaging follow-up of at least two years. Histopathological diagnoses were established by board-certified breast pathologists.
MR imaging reading results were compared to reference standard results to calculate true positive (TP), true negative (TN), false positive (FP), and false negative (FN) findings. For this purpose, BI-RADS category assignments 1–3 were considered negative and 4–5 positive. Additional evaluation of diagnostic parameters was performed considering BI-RADS category assignments 1 and 2 negative and 3–5 positive.
The diagnostic parameters sensitivity TP/(TP+FN), specificity TN/(TN+FP), positive predictive value (PPV) TP/(TP+FP), and negative predictive value (NPV) TN/(TN+FN) were stratified by clinical presentation, conventional imaging findings, and breast density, and were compared using McNemar tests. A P-value ≤0.05 was considered to indicate a significant result.
Results
Patients and lesions
Of 322 patients, 302 women (mean age, 50±12 years; range, 20–79 years) fulfilled the inclusion criteria, and comprised the study cohort. Twenty patients (6%) with MR imaging BI-RADS 2 (n = 18) and MR imaging BI-RADS 3 (n = 2) ratings were lost to follow-up. Thus, 302 patients were included in the analysis. Indications for the examination are listed in Table 1 .
10.1371/journal.pone.0190287.t001
Table 1 Summary of the results of this study: MRI findings are stratified by conventional imaging findings, clinical presentation, and ACR breast composition. Resulting cancer prevalence, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) considering BI-RADS 4 and 5 as positive and BI-RADS 1–3 as negative MRI results.
Total
TP
TN
FP
FN
Cancer prevalence (%)
Sensitivity (%)
Specificity (%)
PPV (%)
NPV (%)
All
318
53
243
20
2
17.3%
96.4% (CI 87.5%-99.6%)
92.4% (CI 88.5%-95.3%)
72.6% (CI 60.9%-82.4%)
99.2% (CI 97.1%-99.8%)
Index lesion
302
48
236
17
1
16.2%
98.0% (CI 89.2%-100%)
93.3% (CI 89.5%-96.1%)
73.9% (CI 61.5%-84.0%)
99.6% (CI 97.7%-100%)
Mammography
Mass
163
24
128
10
1
15.3%
96% (CI 79.7%-99.9%)
92.8% (CI 87.1%-96.5%)
70.6% (CI 52.5%-84.9%
99.2% (CI 85.8%-100%
Mass with architectural distortion
19
8
10
1
0
42.1%
100% (CI 63.1%-100%)
90.9% (CI 58.7%-99.8%)
88.9% (CI 51.8%-99.7%)
100% (CI 69.2%-99.8%)
Mass with microcalcifications
18
4
14
0
0
22.2%
100% (CI 39.8%-100%)
100% (CI 78.8%-100%)
100% (CI 39.8%-100%)
100% (CI 76.8%-100%)
Architectural distortion
69
8
55
6
0
11.6%
100% (CI 63.1%-100%)
90.2% (CI 79.8%-96.3%)
97.1% (CI 28.9%-82.3%)
100% (CI 93.5%-100%)
Architectural distortion with micro
7
1
6
0
0
14.3%
100% (CI 2.5%-100%)
100% (CI 54.1%-100%)
100% (CI 2.5%-100%)
100% (CI 54.1%-100%)
Microcalcifications
26
3
23
0
0
11.5%
100% (CI 29.2%-100%)
100% (CI 85.2%-100%)
100% (CI 29.2%-100%)
100% (CI 85.2%-100%)
Clinical presentation
Palpable
84
22
54
8
0
26.2%
100% (CI 84.6%-100%)
87.1% (CI 76.2%-94.26%)
73.3% (CI 59.2%-84%)
100%
Not palpable
218
26
182
9
1
12.4%
96.3% (CI 81.1%-99.9%)
95.3% (CI 91.2%-97.8%)
74.3% (CI 60.2%-84.6%)
99.5% (CI 96.4%-100%)
Breast composition
ACR a
17
3
14
0
0
17.3%
100% (CI 29.2%-100%)
100% (CI 76.8%-100%)
100%
100%
ACR b
89
20
62
6
1
23.6%
95.24% (CI 76.2%-99.9%)
91.2% (CI 81.8%-96.7%)
76.9% (CI 60.7%-87.8%)
98.4% (CI 90.1%-99.8%)
ACR c
153
17
126
10
0
11.1%
100% (CI 80.5%-100%)
92.7% (CI 86.9%-96.4%)
62.9% (CI 48.4%-76%)
100%
ACR d
43
8
34
1
0
18.6%
100% (CI 63.1%-100%)
97.1% (CI 85.1%-99.3%)
88.9% (CI 53.7%-98.2%)
100%
MRI-only lesion BI-RADS 3–5
16
5
7
3
1
37.5
83.3% (CI 35.9%-99.6%)
70% (CI 34.8%-93.3%)
62.5% (CI 24.5%-91.5%)
87.5% (CI 47.4%-99.7%)
Note: Index lesions refers to the findings that were the reasons for referral to MRI; MRI-only lesions were those lesions additionally detected by MRI and not detected by the initial assessment before MRI; BI-RADS ratings were dichotomized into positive (4/5) and negative (1/2/3) to count true-positive (TP), true-negative (TN), false-positive (FP) and false-negative (FN) lesions, PPV positive predictive value, and NPV negative predictive value. ACR, American College of Radiology.
Breast MR imaging revealed 144 (45.3%) mass lesions that demonstrated a mean size of 15.3 mm ± 14.3 mm (SD; range, 4–95 mm). Further, there were 44 (13.8%) non-mass enhancements with a mean size of 29.2 mm ± 23.1 mm (SD; range, 4–75 mm).
Final lesion diagnoses were malignant in 55 of 318 lesions (44 mass and 11 non-mass), and benign in 263 (100 mass, 33 non-mass lesions, and 130 without contrast-enhancing correlates in MR imaging). Malignant histopathological diagnoses were: invasive ductal carcinoma (IDC) in 43 (35 mass and eight non-mass); and ductal carcinoma in situ (DCIS) in 12 (nine mass and three non-mass) cases. BI-RADS ratings were assigned as follows: BI-RADS 1/2: 184 (0 malignant); BI-RADS 3: 61 (two malignant); BI-RADS 4: 39 (20 malignant); and BI-RADS 5: 34 (33 malignant). Twenty lesions classified as MR imaging BI-RADS 4/5 showed benign histopathology after image-guided biopsy.
The prevalence of malignancy was 17.3% (55 of 318). The malignancy rate tended to be higher in mass lesions (30.6%, 44 of 144) compared to non-mass lesions (25%, 11 of 44) though this was not statistically significant (P = 0.599). None of the 130 conventional findings without an enhancing MR imaging correlate was malignant.
Malignancy rates differed according to clinical and conventional findings ( Table 1 ). Lesions that presented as a mammographic mass with architectural distortion had the highest probability for malignancy (42.1%), while pure architectural distortions and microcalcifications had the lowest malignancy rates (11.6% and 11.5%, respectively). Details on MR imaging results stratified by clinical presentation and imaging findings are given in Table 1 .
Diagnostic performance
ROC analysis ( Fig 1 ) revealed an area under the ROC curve of 0.977 (95% CI: 0.963–0.992). Considering BI-RADS 4 and 5 malignant and 1–3 benign the reference standard revealed 53 true-positive, 243 true-negative, 20 false-positive, and two false-negative breast MR imaging findings ( Fig 2 ). All underlying data are given in the S2 File . The sensitivity, specificity, and positive and negative predictive values were calculated as: 96.4% (95% CI: 87.5–99.6%), 92.4% (95% CI: 88.5–95.3%), 72.6% (95% CI 60.9–82.4%), and 99.2% (95% CI: 97.1–99.0%), respectively. Upon subgroup analysis, no significant differences in diagnostic parameters were found when stratified by conventional imaging findings, clinical presentation, and breast density (P>0.05). Both false-negative findings were assigned BI-RADS 3 and diagnosed due to changes after short-term follow-up.
10.1371/journal.pone.0190287.g001
Fig 1
ROC plot of BI-RADS ratings against the reference standard.
At a cut-off of >BI-RADS 3, the sensitivity and specificity were 96.4% and 92.4%, respectively. In addition, at a cut-off of >BI-RADS 2 the sensitivity and specificity were 100% and 70.3%, respectively.
10.1371/journal.pone.0190287.g002
Fig 2
A 52-year-old patient referred for problem-solving due to newly diagnosed architectural distortion in the left breast (A; white circles on mammography images). 3T contrast-enhanced MR imaging (B; top: T2w image, middle: early contrast-enhanced image, bottom: late contrast-enhanced image) shows the architectural distortion (white circle) demonstrating only mild background enhancement. The lesion was classified as BI-RADS 2, definitely benign. Follow-up of two years did not reveal malignancy.
Considering BI-RADS 3–5 malignant and 1 and 2 benign the reference standard revealed 55 true-positive, 185 true-negative, 78 false-positive, and no false-negative breast MR imaging findings. The sensitivity, specificity, and positive and negative predictive values were calculated as: 100% (95% CI: 93.5–100%), 70.3% (95% CI: 64.4–75.8%), 41.4% (95% CI 36.9–45.9%), and 100%, respectively.
Incidental MR imaging lesions
Incidental lesions identified by MR imaging and assigned BI-RADS 3–5 categories were identified in 16 of 302 patients (5.3%). Of these, six lesions (37.5%, 6/16) proved to be malignant. One lesion was assigned a false-negative BI-RADS 3 and the lesion was upgraded upon six-month follow-up, which was initiated due to the MR imaging findings. Consequently, 10.9% (6/55) of all malignant lesions were detected only by MR imaging and presented multifocal or multicentric disease that did not show any conventional imaging correlates. All these lesions measured ≤10 mm. All other lesions detected outside the area of the conventional imaging findings by MR imaging were given BI-RADS category 2 assignments and not considered as incidental lesions for this analysis.
Discussion
Breast MR imaging yields excellent diagnostic results if used as a problem-solving tool independent of referral reasons. The number of suspicious incidental lesions detected by MR imaging is low, but has a substantial malignancy rate. Our study population included the largest number to date, to our knowledge, of patients undergoing 3 Tesla breast MR imaging for problem-solving. Breast MR imaging had a high negative predictive value of 99.2% and a high PPV of 72.6%. In other words, only three of 10 biopsies recommended based on positive breast MR imaging (BI-RADS 4 and 5) findings were false-positive, and thus, unnecessary, while in eight of 10 patients with a negative MRI result (BI-RADS 1–3), biopsies or further follow-up examinations could be avoided. This came at the cost of two false-negative findings—both of which, however, were detected by short-term follow-up examinations that were initiated due to breast MR imaging BI-RADS 3 lesions, and one of which was detected exclusively by MR imaging. Therefore, all cancers were visualized as enhancing lesions by breast MR imaging. These findings validate that MR imaging is a safe diagnostic instrument if applied as a problem-solving tool in inconclusive cases, as malignancy can be reliably excluded. Furthermore, the number of incidental MR imaging findings that required follow-up or invasive diagnostic procedures was as low as 5.3%, but yielded a substantial malignancy rate of 37.5%. Our multiparametric breast MR imaging protocol is a fast diagnostic test, allowing the examination of up to four patients per hour. Image interpretation and reporting takes only a few minutes and results in high sensitivity and negative predictive values. Thus, breast MR imaging as a problem-solving tool improves patient care by avoiding the anxiety related to follow-up examinations and possibly missed cancer diagnoses.
The indication “problem-solving” for breast MR imaging has been a controversial topic. While a recent meta-analysis corroborated the ability of breast MR imaging to exclude cancer in non-calcified lesions [ 5 ], MR imaging might not be as accurate in lesions that present as microcalcifications [ 8 ]. Although, based on a rather small study population, our results demonstrated excellent sensitivity and NPV in calcified lesions, a fact that might be explained by the more modern equipment, namely, 3 Tesla in combination with a multichannel coil and high-resolution 3D gradient-echo, T2w, and DWI imaging. None of previous studies investigated this clinical setting exclusively with 3 Tesla breast MR imaging [ 5 , 10 ]. These results are particularly interesting as stereotactically-guided biopsies are more invasive and expensive than US-guided biopsies. Here, MRI would be a valuable test for risk stratification to avoid or guide biopsy in case of multiple or equivocal microcalcifications.
The aforementioned meta-analysis concluded that problem-solving definitions are not well-formulated and the empirical evidence about the performance of breast MR imaging in specific imaging findings that lead to MR imaging, e.g., such as architectural distortions, is sparse [ 4 , 5 , 10 , 19 ]. Here, our study directly adds data, as it provides a detailed analysis of the findings that led to the breast MR imaging examination and associates these findings with diagnostic outcomes. Of note, breast MR imaging performance was similar across different indications. Variations were seen regarding malignancy rates and subsequent PPVs, but malignancy could be excluded with high certainty independent of specific indications. Therefore, this study confirms that problem-solving MR imaging is a reliable tool that performs well under varying conditions. Based on our results, we consider all the indications for breast MR imaging investigated in this work as appropriate. Still, we need to stress that breast MR imaging should not generally be used for further evaluation of conventional lesions that can definitely be clarified by US-guided biopsy [ 9 , 10 , 13 ]. Problem-solving in such lesions would be necessary in case of multiple or difficult to localize lesions that are likely to be missed by immediate biopsy. Although a general application of MR imaging in such lesions has demonstrated excellent diagnostic performance, percutaneous biopsy is readily available, easily tolerated, and leads to a faster definite diagnosis [ 19 , 20 , 21 ].
Breast MR imaging is known to detect more cancers than conventional tests, such as mammography and ultrasound. A common concern among breast imagers is that the inherent higher sensitivity of breast MR imaging will lead to the detection of multiple additional lesions that require further workup. Our data show that this number is, first, as low as 5.3% and, second, entails a substantial malignancy rate of 38%. A recent study showed a similar malignancy rate of 33% (7/22) in suspicious MRI-only lesions [ 10 ]. Consequently, our results do not show a relevant number of additional recalls due to the application of MR imaging in the investigated setting. However, suspicious MRI-only lesions warrant further evaluation. Prior studies showed higher rates of incidental recalls between 8.3%-15.6%, with a malignancy rate of these lesions ranging from 0–17% [ 19 , 22 , 23 ]. The lower recall rate and higher prevalence of malignancy in our cohort is likely due to the exclusive use of 3T multiparametric MR imaging.
We are obliged to mention the limitations of the current study. Twenty patients (6.2%) were lost to follow-up. While this rate is well within the acceptable range, false-negative findings could have been missed, resulting in the low possibility of overestimating sensitivity. The retrospective character of this study did not allow an assessment of inter-reader agreement. However, all examinations were read under routine clinical conditions by experienced breast radiologists. The performance of experienced radiologists has been shown to be superior to that of non-experienced radiologists in the interpretation of breast MR imaging [ 24 ]. Further, breast MR images were interpreted considering conventional imaging and clinical findings, an approach that has been recommended, as it improves diagnostic accuracy in non-mass lesions [ 14 ]. Accordingly, our audit reflects the actual clinical setting, and thus, provides more representative results as a retrospective reader study under research conditions. Our results were obtained using a fast (<15 min) multiparametric breast MR imaging protocol (T2w, DWI and DCE-MR imaging) at 3 Tesla, using multichannel coil technology. Again, the retrospective character of our study does not allow conclusions on the respective contribution of individual parameters to the final diagnosis. Our experience is, however, in line with prior publications that showed high specificity when T2w or DWI sequences were applied [ 15 , 16 , 18 , 25 ].
In conclusion, breast MR imaging yields excellent diagnostic results if used as a problem-solving tool independent of referral reasons. The number of suspicious incidental lesions detected by MR imaging is low, but has a substantial malignancy rate.
Supporting information
S1 File
STARD checklist.
(DOC)
S2 File
SPSS table of the raw study data.
(SAV)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
As the polio program nears completion in India, with only a single case in 2011, another disease may be eradicated. However, the road to this goal had been made difficult due to social “resistance” when families refused to vaccinate their children. Though “resistance” to the program due to rumours about the vaccine, frustration with the slow pace of development, and fatigue with repeated doses have been documented [1] , [2] , [3] , [4] , this qualitative study contextualizes some sources of fear and “resistance” to the program during the summer months of 2009 to provide insight for future eradication endeavours.
During the time of the study, monthly administration of the vaccine without an explanation contributed to families growing tired of the program [1] , [2] , [3] , [4] , [5] . Since about 2000, the polio eradication program shifted from vaccinating only a few times a year on National and Sub-National Immunization Days (NIDs) to vaccinating door-to-door every month until children reach the age of six in high-risk areas [5] . Though the door-to-door vaccinations were technically for families who failed to vaccinate their children at established booths, since fewer and fewer families bothered coming to the polio booth, the monthly door-to-door policy had effectively become the standard. This policy shift was also subtly reflected in the UNICEF ad campaign which changed from advocating “do boond zindagi ke liye” or “give two drops for life” to “do boond har baar” or “two drops every time.”
From 2007 through the course of this study in 2009, the Global Polio Eradication Initiative (GPEI) vaccinated almost exclusively with the monovalent oral polio vaccine (mOPV1) in a final push to eradicate P1 [6] . There are three strains of the wild polio virus: P1, P2, and P3, of which P2 has been eradicated using the trivalent oral polio vaccine (OPV) [7] . However, because several studies revealed poor trivalent OPV efficacy for the remaining strains in endemic states, monovalent vaccines were developed [8] , [9] , [10] . Though a bivalent vaccine targeting both P1 and P3 is now in use [11] , eradication strategy at the time of the study focused on P1, resulting in the spread of P3. At the end of 2008, there were 484 cases of P3 and 75 cases of P1 in India, and by the end of 2009 there were 662 P3 cases and 80 P1 [12] . Whereas the strategy to target P1 was probably sound at a biological level, the lack of clear explanation about the use of monovalent vaccines, combined with the confusion due to repetitive rounds, may have contributed to “resistance.”
Possible adverse events associated with OPV may also have contributed to “resistance.” Though the OPV is relatively safe vaccine, as a live-attenuated vaccine, it does have official contraindications including pregnancy or immunodeficiency [13] , and carries a small risk of Vaccine-Associated Paralytic Polio (VAPP). Though VAPP is rare, occurring approximately in one out of every one million children vaccinated, it is clinically identical to paralytic poliomyelitis [7] . Additionally, in India is has been found that the risk may be elevated to 1 per 143 000 infants born [14] , possibly due to the increased coverage, though this has not been confirmed [15] .
Thus, this study was conducted to identify social factors associated with “resistance” to the polio eradication program. The study focused initially on “resistance” regarding rumours of sterilization amongst the Muslim community which will be explained in a separate paper. When it became apparent that “resistance” had spread into the non-Muslim community, the study was broadened to identify knowledge and opinions about the eradication strategy amongst the general public.
Methods
Ethics Statement
Proper informed consent was taken for all interviews and financial compensation provided, in accordance with a protocol approved by the Brown University IRB on March 17, 2009. Both verbal and written consent was obtained from literate respondents who were provided a copy of the consent documents for their records. Only verbal consent was obtained if subjects were not literate, though they were provided a copy of a separate consent document for their records. All consent forms and procedures were approved by the Brown University IRB.
General Protocol
This research took place in the city of Aligarh in Uttar Pradesh, India. Aligarh District was classified as a “high-risk” district for the spread of polio by the GPEI [16] and was the source of the majority of India's polio cases in 2003 [17] .
The methods were conducted in accordance with the principles of rapid ethnography/rapid assessment procedures (RAP) and included in-depth key informant interviews, behavioural observation, and semi-focus groups [18] .The qualitative data was collected over four months, May–August 2009. Methods included participant observation of two GPEI-sponsored clinics and three week-long door-to-door polio rounds, interviews with 27 stakeholders in the polio program, and 80 semi-structured interviews with families who interacted with the polio program. Informal interviews were embedded within the participant observation while the structured interviews were conducted afterward.
Participant Observation and Unstructured Interviews
The researcher used the method of participant observation to collect data both at clinics run by the GPEI and the door-to-door vaccination program. During the door-to-door program, the researcher was embedded with polio vaccination teams as they attempted vaccinating families who had actively refused in the past to accept vaccination, described as “resistant” families. Three polio rounds were conducted during the course of the study, preceded by booth days which started on May 26, July 5, and August 9 of 2009. During the rounds, the researcher was perceived to be a part of the vaccination team, and held onto charts and paperwork while observing interactions between the vaccinators and the families. Field notes were taken of families' reaction to the polio program, health conditions in the neighbourhoods, and the behaviour of the information, education and communication (IEC) teams of the SMNet which consisted of UNICEF Community and Block Mobilizing Coordinators (CMCs and BMCs) as well as medical interns from the Jawaharlal Nehru Medical College and Ajmal Khan Tibbiya College. Participant-observation of the GPEI-run clinics included noting interactions between the patients and staff, and conducting 15 informal interviews with clinicians. The participant observation of the polio rounds was used to gain a stronger understanding of the realities of vaccination on the ground, brainstorm challenges to the program, and compare the local situation to that expected from the initial literature review to fit within the theoretical framework of ‘structural violence’ and health justice.
Participant observation at the GPEI paediatric clinic was conducted to provide insight into local health care and needs, while observation of the door to door Pulse Polio rounds provided insight into both workings of the program and families perception of it. A total of 22 informal interviews were also conducted with the vaccination teams during the course of data collection, and transcribed alongside the field notes. The participant observation was also used to generate further research questions, and determine proper sites for the bulk of interview recruitment which occurred afterward. The participant observation stage of the study was not used for active participant recruitment itself.
Active Recruitment and Structured Interviews
Twenty-seven formal interviews were conducted after active participant recruitment with both grassroots and administrative stakeholders in the polio eradication program. The in-depth structured interviews were conducted with health promoters called Community Mobilizing Coordinators (CMCs) (5), polio booth-workers (5), clinicians who worked in underserved areas (4), medical interns (5), community physicians (5), and administrators with the GPEI (3). These diverse ‘stake holder’ views were conducted to get an insight into how resistance was viewed by those involved with the eradication effort at various levels, and how they saw or sought to address the shifting nature of resistance.
Interview questions were based on data collected from the initial participant observation, and included questions about individuals' knowledge and opinions about the polio eradication program, the oral polio vaccine, causes of “resistance” to vaccination, their confidence in the program's ability to succeed, and their opinions about current or alternative eradication strategies.
Eighty formal, semi-structured interviews were also conducted with families with children who interacted with the polio eradication program in major parts of Aligarh after active recruitment. Individuals from each family were interviewed based on their willingness to participate. All respondents were either the head or co-head of the family as mothers and fathers participated about equally. Though interview questions were geared towards one interviewee, if other family members contributed to the discussion, the interview was allowed to take its course as a semi-focus group. Other than one day of interviews where five “resistant” families were specifically sought out for interview in Jeevangarh, families with children were selected randomly. Participants were recruited by knocking on doors in the major streets/alleys of each ward as determined by neighbourhood informants, asking for families with children who would be willing to interview until a total of around five families were interviewed in each ward as demarcated by local GPEI partner organizations. This occurred in all wards except Jeevangarh which had a total of nine interviewees, four random, and five exclusively “resistant” as described. This partially random selection yielded a diverse number of participants, including several who were “resistant” to vaccination. The number of interviews in each of the wards was Maulana Azad Nagar (5), Jamalpur (5), Civil Lines (3), Jeevangarh (9), Begambagh (5), Devatray (5), K.R. Jain(5), Gandhi Nagar (5), Upper Kot/Upper Fort (5), Bhojpura (6), Shahjamal (6), Indira Nagar (5), Bannadevi (5), PPC (5), and Mehfooz Nagar (6).
These wards represented major blocks of population divided along socioeconomic and religious lines. This sampling was thus sought to get a broad overview of opinions in the community about the polio program, with “resistant” views being well represented. This was why “resistant” individuals were initially sought, though “resistance” was found to be common enough that neighbourhood opportunity sampling yielded resistant families who had interacted with the program. Of all who gave their informed consent to participate in the study, 77 families continued the semi-structured interviews to completion. A total of 3 families from Bhojpura, Jeevangarh, and Mehfooz Nagar decided to stop the interview midway for an unspecified reason. Participants were asked questions about their knowledge and opinions about the polio eradication program, the oral polio vaccine, causes of “resistance” to vaccination, their confidence in the program's ability to succeed, and the provision of health services. Two translators who were familiar with the local environment, fluent in local dialects of Hindustani, and trained to conduct health promotion field activities joined the researcher in conducting interviews. All interviews were conducted in either Hindustani or English by the researcher with the assistance of the translators.
In total, 107 structured or semi-structured interviews were conducted. Participants included 80 families with children in the aforementioned parts of Aligarh, and 27 stakeholders in the eradication program. Each structured interview took 30–40 minutes, and was conducted at a site of the interviewee's choosing. Transcripts were either recorded by hand or with an electronic recorder with the permission of the interviewee. All interviews were made confidential unless the right was specifically waived. Interviews were conducted to the point of saturation, as data repetition occurred at all levels, indicating the views found reflected that of a substantial portion of the studied respondents [19] .
Data Analysis
Analysis of the data was conducted by the researcher independently. All transcription and translation was done by the researcher who is fluent in Hindustani. Due to difficulties with sound quality and local dialects, full transcription was conducted for 20 of the recorded interviews. Partial transcription was done for the remaining 82 recordings in addition to the partial transcription already done in the form of field notes during all structured and semi-structured interviews. Of the structured and semi-structured interviews, 5 were transcribed by hand exclusively. All informal interviews were written alongside field notes. No software was used in the analysis of the data, which was manually coded for causes for “resistance” in the Muslim community, causes for rising “resistance” in the non-Muslim community, gaps in knowledge about polio eradication strategy, behaviour of the polio vaccination teams, and trust of the medical establishment and government. This coding scheme was developed based on previous literature about causes of “resistance” and inductions from the participant observation. The informal interviews from the field notes and transcripts from the interviews were coded and grouped by the described major themes to give a better understanding of “resistance.” Participant observation, semi-structured interviews, and formal interviews based on active recruitment were given equal weight and not differentiated during analysis.
Results
Fatigue and Confusion from Program Intensification
During the course of the study, families in Aligarh showed fatigue from vaccinating their children monthly because they did not understand the need to do so. Though most respondents supported the eradication program and vaccinated their children, many did not seem informed why the program had intensified the frequency of vaccination. Families described that when they asked the door-to-door vaccination teams why they visited them so often, they were usually not given an adequate response. Though the presence of medical interns helped, members of the vaccination team were observed sometimes providing dubious etiological explanations to the families: telling them that polio was “special” and needed a constant boost which other vaccines did not. One clinician who worked with routine immunization services explained that many patients did not understand why, whereas they vaccinated with BCG at fifteen days, DPT at one and a half months, and received a measles injection twice, they had to vaccinate for polio almost twelve times a year until they were about six years old. A respondent from a Christian family in Banna Devi explained that she did not understand the shift, or its purpose. Explaining her frustration and confusion about the frequency of the visits, she said:
It (the polio vaccine) is working. But nowadays they are just overdoing it. Coming all the time and bothering people. Like they are coming in the afternoon, which is just a nuisance. They are also overdosing everyone…before it used to be every month. Now it is every week right?!…This has become too much then right?
Though her reference to the vaccination happening every week was likely due to confusion with the “B-team” that vaccinated children missed during the first rounds, it seemed no one informed her why there was a shift in policy.
Doubts due to Translucent Monovalent Vaccine Policy
Confusion about intensified vaccination was confounded by a lack of information about the strategy to eradicate the P1 strain of the virus through the monovalent mOPV1 vaccine. As families saw polio cases occur despite the intensified rounds, they started to doubt the program. Families were not usually informed of differences between the strains: P1 and P3; and when they saw or heard of polio cases occurring, usually P3 at the time, many came to doubt the efficacy of the monthly administered vaccine, which only targeted P1. Expressing her doubts about the vaccine and frustration with the repeated doses, a Hindu lady in Begambagh explained:
People might think there is no point to the program and refuse to vaccinate on that ground. We think there is no point which is why we don't vaccinate. For the rest, we cannot say. Son, they are giving so much of the vaccine, so much of the vaccine, but still polio is affecting children.
Somewhat informed about the strains, but still confused about the strategy, an educated man in Banna Devi asked:
It is a P3 virus for polio right? It comes up in the newspaper that, despite all the vaccinations, P3 cases are occurring.
He was not informed that the P1 strain was being targeted for vaccination, and was growing tired of the program.
The lack of awareness about the vaccine was often reinforced by the vaccination teams. When providing families with mOPV1 on the polio rounds, vaccinators usually told families that the vaccine protected them from “polio,” and not just the P1 strain. Based on the advice of the vaccinators, families gave their children the vaccine thinking it protected them from “polio” when in fact it protected them only from P1.
If families thought the mOPV1 protected them generically from “polio” it would have proven problematic if P3 cases occurred. Though this likely did not occur amongst selected participants, a Muslim family interviewed in Shahjamal shared a qualitatively similar experience. When family members were asked what they thought of the polio eradication program, the parents responded that they thought the government was trying “to make a fool of the public.” They said that despite vaccinating their son regularly, he became crippled in a manner characteristic of polio. The father said he did not believe in the rumours that the government wanted to sterilize their children, but rather, it was their personal hardship which caused them to lose faith in the program. The full story of the family is outlined below:
Translator: He is saying that the program is making a fool of the public.
Respondent: A fool out of them.
Researcher: Meaning?
Respondent: Making a fool out of them meaning they don't like it. To make crazy…The thing you are trying to say is in front of me…look.
Translator: He is saying to say that this child has polio.
Researcher: He has polio?
Translator: And that he has regularly been drinking the vaccine.
Respondent: And he's been drinking regularly. He's been drinking the vaccine till he was seven years old.
Researcher: Okay okay. And still he got sick?
Respondent: Still he got sick…‥
Researcher: In your view, does this vaccine do any work or what does it do?
Respondent: In our view…they tell us to come and take the vaccine, take the vaccine. Are we not taking the vaccine? No we are not. This is the benefit from the vaccine (pats his crippled child on the back); this is the benefit from it. (Emotionally) This is the benefit from it! Nothing! We shouldn't vaccinate.
Though this man said he once believed in the polio eradication program and vaccinated his children regularly, he had lost faith that the vaccine worked because despite vaccinating, his son became crippled. If the public is not informed of which type of vaccine they receive while cases continue to occur, they may stop supporting health programs.
Fear from Adverse Events Proximate to Vaccination
Though many people had come to doubt the efficacy of the vaccine, as rumours spread about adverse reactions with the polio vaccine, some individuals also became sceptical of its safety. In fact, one family had become so afraid of the vaccine that on the polio rounds, they threatened to kill their own children, call the police, and frame the polio workers if they did not leave their homes. Throughout Aligarh, there were rumours that when some families vaccinated their children against polio, the next day the children contracted a fever, became afflicted with polio, or even died. In Upper Kot, one grandmother screamed that the vaccine gave one of her grandchildren polio and refused to vaccinate the other grandchildren “even if the Prime Minister of India” came to her door. This “resistance” frustrated the interns who, referring to Vaccine-Associated Paralytic Polio (VAPP), remarked “there is one case in a million and that one case causes so many problems.” Often, rumours indicated that adverse effects occurred when the children were vaccinated during a fever, which is not an official contraindication to OPV vaccination. When vaccinating on the polio rounds, CMCs and medical interns tried to counsel patients about this fact, but many remained fearful. One family in Shahjamal explained to us how they refused to vaccinate because they had heard this rumour:
Respondent's Wife: Okay, so if you drink the vaccine when you are sick it can be problematic right? I have heard it has caused problems. Like when a child is getting a fever and they force you to drink the vaccine…
Respondent's Wife: When we used to take the vaccine, many children had problems because of it. Some children even died.
Respondent: Some children even died!
Researcher: Some even died because of the polio vaccine? Okay okay.
Respondent: When they had a fever.
Respondent's Wife: With Typhoid…sometimes they get the child to drink the vaccine during typhoid fever too…and then it is problematic.
Researcher: So people force you to take the vaccine?
Respondent: The vaccine is given by force.
Respondent's Wife: When we refuse, (the polio workers) say that the vaccine won't cause any harm and give the vaccine. Then the health of the child gets compromised. Quite a few cases like this have happened. That is why a lot of people are afraid.
Thus, the family refused to vaccinate because they heard the polio vaccine caused severe adverse effects during a fever. The fact that vaccinators told them the vaccine wouldn't cause any harm before administering what was a lethal vaccine in the rumour caused families to further distrust the program. Even an individual who was himself tragically afflicted with polio, and thus continued to vaccinate his children, expressed some fears because of this risk:
Researcher: Do you feel the vaccine is safe?
Respondent: I feel it is safe but I have heard of two cases which happened in Jeevangarh. I heard that the children had fevers, but that the polio workers forced their way and vaccinated the children, causing harm to the children. There people had forced their way and were very rude.
Researcher: So this happens?
Respondent: Yes, yes it happens. If you are going to vaccinate a child, you should know everything that is going on (with the child) at first.
Many people were afraid to vaccinate their children because they feared that the polio vaccine actually caused adverse affects such as fevers, diarrhoea, and even paralysis.
However, this fear of adverse reactions was not limited to the polio vaccine alone. For example, when a routine immunization camp funded by the GPEI to increase vaccine acceptance was held in Jeevangarh on June 11, 2009, one of the staff workers informed the visiting physicians and vaccinators that they should not push people to vaccinate. Apparently a child had died in the past few days from an adverse reaction to the DPT vaccination, causing widespread fear and refusal to vaccinate.
Distrust of the Vaccination Teams
In addition to fears about the vaccine itself, several families expressed distrust of the vaccination teams. Whether educated or not, many doubted the training of the CMCs and BMCs, and felt that both the instances of polio cases occurring despite vaccination and the stories about adverse effects occurred when vaccinators failed to maintain the proper temperature of the vaccine (cold-chain). One educated Hindu resident of Banna Devi explained that he saw polio cases arising despite repeated rounds. He felt the workers might be to blame, breaking the cold-chain:
Respondent: I don't understand how, despite taking the vaccine from our physicians, and taking these door to door vaccines, symptoms (cases) keep arising….This means something is wrong. Either there is something wrong with the medicine, or the proper temperature which the medicine has to be kept at is not maintained. Either somebody is not looking at the expiration date or the workers are being careless in how they maintain the temperature and handle the icebox.
Translator: The cold-chain is broken.
Respondent: Or they hold it in their hands.
Many people, especially from the upper classes, felt that they would not only put their children at risk of adverse events from the broken cold-chain, but possibly other diseases as the teams vaccinated multiple children with the same dropper. Many of these respondents were not afraid of the vaccine per se and said that they would gladly go to a physician or clinic for the same vaccination. One “resistant” family in Upper Kot explained their position as such:
Respondent: No. We don't trust the workers or whoever comes. If we have to take it at the Medical College we will from the doctors but it does not feel right to take it from the polio workers. They are given fifty rupees to wander around and give drops. They might finish a drop and just throw it, and will use the same mouth piece for everyone.
Researcher: So you will vaccinate there but not here?
Respondent: Yes, we will go, show our children to the doctor at the Medical College get routine care and vaccinate and come back.
Researcher: So you don't vaccinate from the workers, you don't like the workers?
Respondent: The workers aren't educated and are just paid fifty rupees to wander; we don't know what they are vaccinating with, what they are giving.
When a businessman from Jamalpur waited at one of the health clinics for medication, looking at the polio vaccinator, he commented on her training and the danger he felt it posed to children:
She has been holding the vial the whole time, warming it. The people do not handle the drops properly; it should be kept in the ice box. For example, when giving the drops, she might put the dropper into a child's mouth and then use the same dropper with the next child, causing contamination.
Attitudes regarding Policy Transparency
Vaccination teams did not share information with families regarding either the intensification of the vaccine program, use of mOPV1 versus the trivalent OPV, or risks associated with vaccination. Though the rationale for the intensification of the program was simply poorly publicized, officials with the GPEI in Aligarh indicated that it was formal policy to not actively inform the public of other issues discussed in this paper.
Local GPEI officials described that the monovalent strategy to vaccinate against P1 was based on the national strategy, but there no obligation was felt to inform the public of shifts in the type of vaccine provided. As seemed apparent from respondents and observations on the polio rounds, GPEI officials confirmed that the public was not informed that only the monovalent was provided unless they asked though it was, “not as if (they were) hiding the fact either.” However, given that the vast majority of the population was uneducated, unless informed the vaccine protects them from one strain, they would not have been able to ask or find out. As one GPEI official explained:
Look, it is like this, the government has emphasized we should eradicate P1 first, then P3 which is easy. P3 is less virulent, spreads slower, and its residual paralysis effects are weaker. But today we are talking about the community. To the community, these things aren't 100% shared, through the newspapers and other communications we don't always say which vaccine is being used. But if someone asks, it is not as if we are hiding the fact either. Whether paediatricians, private practitioner, or a common man. If someone asks us, we answer and tell them what the strategy is, why P1 is used, why P3 is not, these things aren't hidden on any level, but you can't share it in every community because there are very few people who will understand you.
Responses from key stakeholders regarding sharing risks associated with vaccination, like VAPP, were similar. One of the local community physicians who worked regularly with the polio eradication initiative explained that because people were uneducated, only minor risks associated with vaccination were shared with patients. Major ones like VAPP were ignored because of the risk of rumours. He explained:
They (patients) know about some reactions, but they don't know about serious reactions like paralysis. Nobody talks about paralysis with them, because if we tell them there might be one case, the person will run away, so we generally avoid it. But for minor reactions like development of swelling, that people share.
Researcher: So in general there is the impression is that it is best not to inform then right now?
Respondent: No, we can only inform them if they are literate. Without education, you tell one person, they will tell 100 persons. That will happen.
As seemed apparent from respondents and the polio rounds, GPEI officials confirmed that the public was not informed of the risks associated with vaccination. Instead, vaccinators informed them there was no risk or side effect. As explained by an individual with one GPEI partner organization:
Look, you can explain everything to educated people and they understand everything. But if you talk about VAPP with uneducated or less educated people, about Vaccine Associated Paralytic Polio, it would be taken negatively. For this reason, these technical issues, we do not discuss with them, we just tell them that there is no harm from the vaccine itself, there is no side-effect, and your child will be fully safe.
Paediatricians and physicians who generally worked with the population directly rather than in a more administrative/public health position, were more discouraged by the lack of transparency regarding VAPP. One paediatrician openly complained, “VAPP is kept a secret, people are not told about it, and it remains a rarely discussed issue, even in the medical community.”
Limitations
Due to the ethnographic approach to the study, in addition to the manual coding of the transcribed data and field notes, this study is subject to researcher's bias. Selection bias may have occurred due to the relatively large yet unspecified number of participants who declined to interview. Additionally, as not all interviews were completely transcribed, the full scope of views shared by participants may not have been acknowledged. Difficulties translating local dialects and poor sound quality may also hinder analysis of the data.. Nevertheless, the study provides insight into views and attitudes toward vaccination which were prevalent at the time of the study.
Discussion
Given that polio eradication necessitates almost complete vaccination coverage [5] , unclear communication about vaccination policy seems to have been problematic amongst study participants. Increased transparency and an adverse-effects compensation program may need to be considered to build more trust with the public in future programs.
Intensification of the polio program and lack of transparency about the use of monovalent vaccine seemed to contribute to “resistance” to the program. Families in cities like Aligarh had not been given adequate explanations as to why the polio eradication program was vaccinating every child every month. From the data, it is apparent that this may have contributed to fatigue, if not suspicion of the program. Because the public was uninformed of the strategy to eradicate P1 first as well as differences between P1 and P3, when P3 cases occurred, many saw a “polio case” generically and came to doubt the efficacy of OPV. Simultaneously, the dearth of this information deprived families of the choice to vaccinate against P3, potentially breaching trust between the patient and provider, as qualitatively was the case with the family in Shahjamal whose child developed polio-like conditions despite vaccinating regularly. These families deserved to know what medications they were, or were not being able to provide for their children.
During the course of this study, it was also found that there were rampant rumours that the OPV caused children to develop fevers, sickness, AFP, or even die. There are three possible causes for these beliefs: that they were coincidental, that they were cases of VAPP, or that they were cases of P3. Most officials insisted that the rumours of fevers, paralysis, and death were coincidental: that the conditions existed at the same time as the administration of the OPV or lay dormant in the children before the administration of the vaccine. Because the vaccine was administered monthly, coincidences are likely in a population where death and disease remain, all too common.
However, the fact that the vaccination teams usually told them the vaccine was completely safe made these individuals further doubt the program. This explanation to the public is problematic, if not dishonest. With all vaccines, there is some inherent amount of risk. The OPV, as a live attenuated vaccine, carries the risk of causing either a fever, mild body aches, or even full fledged paralysis if the virus reverts [20] . As these were the very conditions patients described their children had after taking the vaccine, it would beg to ask the question if cases of Vaccine Associated Polio (both with and without paralysis) were occurring.
Though GPEI officials in Aligarh insisted that not a single case of VAPP occurred in the district since the start of the program, VAPP cases have occurred in India [21] , [22] . It has been estimated that there may have been from 83 to as many as 300 cases of VAPP per year in India during the course of the program [4] , [23] . There were also 21 cases of Vaccine Derived Polio Viruses (VDPV) in India during the course of 2009 [12] . A third possible cause for these rumours might have been occurrences of P3 which happened proximate mOPV1 vaccination. Regardless of whether the cases described in this paper were coincidental, cases of VAPP, VDPV, or P3, their effects represent how adverse affects like VAPP cause public apprehension and distrust when full information is not disclosed.
Das and Das have described trust as fundamental for effective immunization, for even with poor information, people rely on the trust with their provider to accept the vaccine [24] , [25] . Though rumours about vaccine failure and the arguments presented for “resistance” are often deemed “unsound from a biomedical perspective,” they are often based on rational arguments and have a strong emotional aspect due to their personal nature [26] . Studies in risk assessment have demonstrated that biased media coverage, and anxiety-provoking incidents, as was the case here, cause uncertainties to be denied and risk perception to be exaggerated [27] . Thus, though the statistics for such cases are small, the emotional impact of each incident is large for the family of an affected child, and has similar reverberations when the story is spread, forming “shared notions of resistance” [28] . As stories spread of adverse events proximate to vaccination it was this trust which was shaken, causing an increase in “resistance.”
Though the policies of not disclosing the risks associated with OPV vaccination or explaining the monovalent strategy were initially done to avoid confusion and achieve high levels of vaccination, if trust with the public was affected, it would have been important to increase policy transparency and improve information, education, and communication (IEC) activities.
Risk perception studies indicate that the public “will accept risk from voluntary activities that are roughly 1000 times as great as it would tolerate involuntary risks,” highlighting the importance to increase active demand for the vaccine [27] . The Ottawa and Bangkok charters for Health Promotion advocate for increases in health literacy as a means for improving public control over all modifiable determinants of health [29] . With increased health literacy, communities are often better able to determine what is best for their well-being, and advocate for programs like vaccination. For example, Friedman and Shepeard's study on HPV vaccine attitudes in the US found that though initially many participants did not know about HPV or risks associated with the vaccine, when empowered and informed, participants clarified their concerns and actively demanded the vaccine. As valued members, the participants also input their own ideas about how to best inform rather than alarm the public about the disease which was highly sensitive issue due to its high prevalence, nature as an STI, and carcinogenicity [30] . For example, some African Americans in the group recommended supplying the vaccine through private clinics with African American physicians rather than government health agencies due to the historical legacy of government distrust from the Tuskegee study, drawing parallels with “resistance” to OPV due to historically based distrust of the government amongst the Muslim community in India.
Most notably, increases in IEC activities have already improved delivery of vaccine with polio eradication in India. Mobilization of grassroots CMCs and BMCs to increase interpersonal communication about the benefits of OPV, assuage false fears such as the rumour that OPV causes sterility, and address other health grievances have dramatically reduced “resistance” to OPV [25] , [31] . Furthermore, UNICEF's Underserved Strategy and Social Mobilization Network (SMNet), whom this study was conducted with, had improved communication between the GPEI and local communities by holding educational skits and plays about polio, and recruiting grassroots stake holders such as religious clerics to advocate for vaccination [32] , [33] .
The proposed goal of these IEC activities is to provide accurate information and correct misunderstandings [34] . The SMNet and Underserved Strategy have been very successful and accomplished this goal by reducing “resistance” across the board. As the roots of “resistance” change, it is necessary to modify what is targeted. During the time of this study, the “resistance” shifted from sterilization rumours to fatigue and fear about vaccine policy and safety. If this situation was to arise in India again, or currently stands in the other polio endemic nations, it may be necessary to provide accurate information about vaccine policy and safety, empowering the public to make the right decisions for the health of their children.
For example, if the public had been informed of the strategy to eradicate P1 first, they would have been able to follow case type, vaccinate against P3 if they wanted, and note the progress of the program. It could have been widely publicized that P1 cases fell by 51% from 2008–2009 to increase confidence in the program [35] . Even less educated families would probably have understood that there were different types of polio, and thus upon hearing of, witnessing, or as seen in this study, experiencing the occurrence of other cases of AFP, understand that it may not have been due to OPV efficacy alone. Some may contend that this would become a huge and difficult task. Indeed it may since health literacy about serotypes is low in developed countries as well. However, since only polio remains the principal disease of eradication in a pseudo “opt-in” format with massive campaigns at a national level, adding this level of detail need not be ruled out.
If the public had been involved from the beginning, it is possible that some of the resistance may have been reduced. Families were especially frustrated with the top-down nature of the polio program since it was not their principle priority. With open sewers, diarrheal illnesses, and unpaved roads: their priority was development. During this and previous studies [1] , [2] , some families agreed to vaccinate only if roads were built and other medical services provided. A large part of the Underserved Strategy with the SMNet had to be dedicated to building this bridge in the end. Though the idea of “disease eradication” garners more attention and drives international funding, pursuing eradication without communicating with the public may have simple dragged the program out longer then it needed to be.
Though the introduction and wide success of the bivalent OPV solved most of the problems discussed here the monovalent strategy, drastically reducing both P1 and P3 with about the same serconversion rates as the monovalent vaccines [36] , it remains important to remember that the problems posed by lack of transparency likely persist and should be considered as policy shifts continue to occur.
An additional concern with the current strategy is that of medical ethics. Of the four medical principles, justice, beneficence, nonmaleficence, and autonomy, not informing the public of the small risks associated with OPV vaccination may impinge on the principle of autonomy: “giving patients the right to make their own choices” [37] . If patients are compelled to make a decision without access to information which could be provided, as seemed to be the case from the study where they were told that the vaccine is completely safe and has no side effects, it would prove problematic. The same issue would be the case with a dearth of information about the monovalent strategy.
To address this issue in future programs or countries still endemic with polio, increased transparency coupled with an adverse-effects compensation program could be considered. Many people had advocated for the introduction of inactivated polio vaccine (IPV) for eradication because it does not have the associated risk of VAPP like OPV [21] , [38] . However, the cost and difficulty in administering IPV, which will not be discussed here, makes this problematic. Rather, by increasing transparency about the risks associated with OPV, the public could be empowered, restoring the principle of autonomy. As there should be few cases of VAPP and VDPV, especially with the success of the bivalent vaccine, the compensation program may may provide a more equitable alternative.
Conclusions
A lack of transparency about the polio eradication program appeared to have contributed to “resistance” to vaccination in Aligarh in 2009. Families who had not been informed of the intensification of the program had come to doubt the vaccine's efficacy as polio cases occurred. This doubt seemed often exacerbated by the lack of transparency about the monovalent strategy to eradicate P1 as families had no way to differentiate polio serotypes. Many families in the study had even become fearful of the vaccine itself from what they perceived to be adverse events after being told there was no risk with vaccination. Though India has almost eradiated polio, the lessons learned here about the nature of social resistance should be considered to build and keep trust with the public in other polio-endemic regions and future eradication efforts.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
After publication of this article [ 1 ], concerns were raised that in Fig 5 , the adipose tissue panels for GPR41 and GPR43 appear to be derived from the same image.
10.1371/journal.pone.0239768.g001
Fig 5
Immunolocalization of GPR41 and GPR43 in spleen, ileum, colon and adipose tissue.
Sections were stained with the polyclonal human-specific GPR41 and GPR43 antibody (Santa Cruz, CA). The images were captured on each slide at 400x magnification under Olympus BX51 (Olympus, Japan). The cells marked by the arrows were GPR41 or GPR43 immunoactive. Bar = 20 μm.
The authors apologize that an error was made in generating Fig 5 such that the wrong image file was included for GPR41 adipose tissue. The original images for GPR41 and GPR43 adipose tissue are provided in S1 and S2 Files, and the authors provide here an updated version of Fig 5 including the correct data. The original images supporting other panels of Fig 5 are no longer available.
An Academic Editor reviewed the updated figure and underlying data and confirmed that they support the results and conclusions reported in the published article.
The authors apologize for the error in the published article.
With this Correction, the authors also provide underlying data supporting Fig 2 and Fig 3 of [ 1 ], in S3 and S4 Files.
Supporting information
S1 File
Original image for GPR41 adipose tissue results shown in the corrected version of Fig 5 .
(TIF)
S2 File
Original image supporting GPR43 adipose tissue results shown in Fig 5 .
(TIF)
S3 File
Raw quantitative data supporting Fig 2 and Fig 3.
(PZF)
S4 File
Raw quantitative data supporting Fig 3.
(PZF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Parkinson’s disease (PD), a chronic neurodegenerative disorder, is characterized by motor symptoms, such as tremors, bradykinesia, and rigidity, and a wide range of non-motor symptoms that affect many domains [ 1 ]. Similar to other progressive neurodegenerative disorders, intracellular protein deposition in patients with PD is a characteristic finding that is crucially attributed to a protein removal defect resulting from lysosomal dysfunction [ 2 ]. However, the underlying mechanism of PD remains poorly understood. PD is associated with various causes, including complex interactions between genetic and environmental factors. In addition, PD has an increasing global burden owing to the increasing older population and potential impact of environmental factors, which are expected to further increase in the future [ 3 ]. Therefore, identifying potentially modifiable risk factors for PD is indispensable.
Proton pump inhibitors (PPIs) are most commonly used to treat gastroesophageal reflux disease and ulcers and can pass through the blood-brain barrier (BBB) [ 4 ]. Their use contributes to the pathogenesis of neurodegenerative diseases by inhibiting lysosomal acidification through the inhibition of vacuolar proton pumps and by preventing the degradation ability of fibrillar amyloid-β (Aβ), an Aβ degradation product [ 5 ]. Therefore, PPI exposure may be a risk factor for various neurodegenerative diseases. Previous studies [ 6 – 8 ] have not demonstrated a relationship between PPI use and the risk of dementia, Alzheimer’s disease, and amyotrophic lateral sclerosis. Few epidemiological studies [ 9 – 11 ] have reported an association between PD and PPI use. However, these studies have some limitations. The dose-response relationship is unclear, and more prolonged effects are also not known. In practice, there are growing concerns reported regarding the potential abuse and use of PPIs for longer periods than recommended in clinical guidelines due to inappropriate indications. Additionally, identifying high-risk groups for PD among PPI users is important for primary prevention. Therefore, we aimed to examine the relevance of PPI use as a risk factor for PD and the dose–response relationship. Given the wide use of PPIs and that PD is one of the most serious diseases to date, the confirmation of these results is likely to be significant.
Materials and methods
Data source
Data were extracted from the population-based cohort of the Korean National Health Insurance Service (NHIS) database, which covers approximately 97% of the Korean population, is managed by the Korean government, and comprises > 95% of all national healthcare facilities. We used the 2002–2019 NHIS data in this study. Data were accessed on 4th December, 2020, for research purposes. The National Health Insurance recommends members to undergo a comprehensive health checkup at least once every 2 years, with changing recommendations based on each individual’s age. The Institutional Review Board of the author institute approved this study (approval number: EUMC 2020-07-028) and waived the requirement for obtaining informed consent because the database maintained de-identified and anonymous data of sampled individuals. This study also complied with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards.
Study design
A nested case-control study was designed with cases that had been diagnosed as PD between 2010 and 2019 (N = 200,777). Controls comprised participants with no PD diagnosis between 2010 and 2019. Under the regulation of the NHIS institute, the total control population should be reduced to four times that of the cases. Thus, the initial population of controls was randomly selected from the whole control cohort between the aforementioned period, as stratified by age and sex on the index date (N = 913,685). The index date was defined as the first date of PD diagnosis. After applying the exclusion criteria, cases (N = 43,405) and controls (N = 201,319) were matched in a ratio of 1:4 based on age, sex, body mass index (BMI), hypertension, and diabetes mellitus at 5 years before the index date. Finally, 31,326 matched cases and 125,304 matched controls were obtained ( Fig 1 ).
10.1371/journal.pone.0295981.g001
Fig 1
Flowchart of the study population.
*Among those without a PD diagnosis from 2010 to 2019, participants with the same age and sex at the index date as those with a PD diagnosis were randomly selected in quadruplicates. **The index date was defined as the first date of a PD diagnosis as the disease code recorded in the NHIS database. ***Variables were measured at 5 years before the index date. Abbreviations: BMI, body mass index; DM, diabetes mellitus; HTN, hypertension; NHIS, the Korean National Health Insurance Service; PD, Parkinson’s disease.
Identification of PD
Between 2010 and 2019, we defined patients with newly diagnosed PD as those meeting at least one of the following criteria:
Based on the International Classification of Diseases , 10 th Revision , Clinical Modification (ICD-10-CM) criteria, the G20 code was utilized at least once among the available diagnostic codes, including the primary and sub-diagnosis codes, and PD drugs were prescribed for ≥60 days ( S1 Table ) [ 12 ].
More than one diagnosis on the PD registration code (V124) in the Intractable Diseases Registry.
The first definition was formulated according to that in previous studies [ 13 – 15 ] and the expert opinion of a neurologist. PD drugs were also determined based on the expert opinion of a neurologist [ 16 ]. Among all drugs, we excluded those that can be used for indications other than PD, including amantadine, apomorphine, benzatropine, biperiden, bromocriptine, cabergoline, dihydroergocryptine, orphenadrine, piribedil, procyclidine, rotigotine, and trihexyphenidyl. Finally, we included the following PD drugs (in the format of Anatomical Therapeutic Chemical (ATC) classification system code [drug name]): N04BX02 (entacapone); N04BA02 and N04BA03 (levodopa combinations); N04BC02 (pergolide); N04BC05 (pramipexole); N04BD02 (rasagiline); N04BC04 (ropinirole); and N04BD01 (selegiline), which are specifically used for PD treatment ( S1 Table ) [ 17 ].
The second definition is very similar to the United Kingdom Parkinson’s Disease Society Brain Bank criteria for the V124 code used for clinical and research purposes [ 18 ]. We excluded patients diagnosed with PD before 2010 and those diagnosed with dementia (ICD-10 codes: F00, F01, F02, F03, G30, G31, and G32) or cerebrovascular disease (ICD-10 codes: I60-I69, G45-G46), which could lead to misclassification. Based on the definition of PD, we defined the index date of PD as the date of the first G20 diagnosis, the date of the first PD drug prescription (whichever came first), or the date of the first diagnosis of the V124 code. Patients with PD and controls (in a 1:4 ratio) were matched on the index date for age, sex, BMI, hypertension, and diabetes mellitus. Moreover, we calculated the Charlson Comorbidity Index (CCI) score based on the 5 years before the index date.
According to a study in Korea that investigated prevalence using the same definition as the one used in this study, the average prevalence of PD between 2012 and 2015 was 171.01 per 100,000 persons. The average age-standardized prevalence of PD according to the WHO 2000–2025 and US 2000 standard population between 2012 and 2015 was 114.13 and 176.21 per 100,000 persons, respectively [ 16 ]. Thus, the study showed a similar prevalence rate to existing worldwide data.
PPI exposure
PPI use was based on A02BC ATC codes (drug name) as follows: A02BC01 (omeprazole), A02BC02 (pantoprazole), A02BC03 (lansoprazole), A02BC04 (rabeprazole), A02BC05 (esomeprazole), A02BC06 (dexlansoprazole), and A02BX (ilaprazole) [ 19 ]. Information regarding all PPIs dispensed between January 1, 2010 and the index date was extracted from the Prescribed Drug Register. PPI use was defined as at least one PPI prescription during the study period. PPIs are not over-the-counter drugs in Korea; therefore, all drugs categorized as PPIs are registered in this database. The cumulative defined daily dose (cDDD) for PPIs was calculated as the total dispensed number of defined daily doses between January 1, 2010 and the index date. Therefore, we initially defined PPI use as a categorical variable with two groups (i.e., PPI users and non-users). PPI users were subsequently further subdivided into three groups based on the exposure amount and tertile distribution of cDDD as follows: (1) very low exposure (<180 DDDs); (2) low exposure (180–359 DDDs); and (3) high exposure (≥360 DDDs). We employed 1- to 3-year lag windows before the index date for sensitivity analyses.
Covariates for risk adjustment
The PD-related variables included diabetes mellitus [ 20 ], hypertension [ 21 ], dementia [ 22 ], stroke [ 23 ], smoking [ 24 ], alcohol use [ 25 ], and BMI [ 26 ]. Excluding the variable used for matching, the following four variables were included as potential confounders: calendar year of the index date, CCI, smoking status, and alcohol consumption. Smoking status and alcohol consumption were extracted from the records of health examinations and categorized as follows: smoking status, never smoker, ex-smoker, or current smoker; and alcohol consumption, none, rarely (1–2 times/week), or frequently (≥3 times/week). Additionally, disease status was assessed through the modified CCI proposed by Sundararajan et al., [ 27 ] calculated as presented in S2 Table .
Statistical analysis
Descriptive statistics are presented as means (standard deviations [SDs]) or medians (interquartile ranges) for continuous variables and frequency (percentage) for categorical variables. Group comparison was conducted using the two-sample t -test or Mann–Whitney U test for continuous variables based on the normality test results or chi-square or Fisher’s exact test for categorical variables, as appropriate. We analyzed the data under a case-control design to compare the use and dose of PPIs between patients with PD and healthy controls. Conditional logistic regression was conducted to determine the association between PPI exposure and PD with adjustment for covariates excluding the matching variables (i.e., calendar year of the index date, CCI, smoking status, and alcohol consumption). The odds ratios (ORs) and their 95% confidence intervals (CIs) represented the degree of association. Only PPI use that occurred within the window of 5 years before PD diagnosis was considered in the analysis, and the lag time was set at 1, 2, and 3 years before PD diagnosis for sensitivity analyses. A two-tailed p-value of <0.05 was considered statistically significant. All analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC, USA) and R version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria).
Results
Baseline characteristics
Table 1 summarizes the baseline characteristics of the included participants, who were well-matched between the groups. The cohort’s mean age (SD) was 67.65 (8.85) years, and 44.56% of patients were male. Regarding diseases considered as risk factors, a higher proportion of individuals with a CCI score of ≥3 points was observed in the PD group.
10.1371/journal.pone.0295981.t001
Table 1 Characteristics of the study population.
Variable
Matched cases
Matched controls
p-value
(N = 31,326)
(N = 125,304)
Age (years) at index date, mean±SD
67.72±8.92
67.63±8.83
0.134
Female, n (%)
17,366 (55.44%)
69,464 (55.44%)
1
BMI category (kg/m 2 ), n (%)
1
Underweight (<18.5)
830 (2.65%)
3,320 (2.65%)
Normal (18.5–24.9)
19,133 (61.08%)
76,536 (61.08%)
Overweight (25.0–29.9)
10,018 (31.98%)
40,072 (31.98%)
Obesity (≥30.0)
1,344 (4.29%)
5,376 (4.29%)
2018–2019
1,543 (4.93%)
5,465 (4.36%)
Comorbidities, n (%)
HTN
200 (0.64%)
800 (0.64%)
1
DM
68 (0.22%)
272 (0.22%)
1
Calendar year of the index date, n (%)
<0.001
2010–2011
5,948 (18.99%)
26,649 (21.27%)
2012–2013
7,063 (22.55%)
28,796 (22.98%)
2014–2015
8,401 (26.82%)
32,303 (25.78%)
2016–2017
8,370 (26.72%)
32,091 (25.61%)
2018–2019
1,543 (4.93%)
5,465 (4.36%)
CCI score group, n (%)
<0.001
0
9,318 (29.75%)
56,032 (44.72%)
1
9,393 (29.99%)
36,993 (29.52%)
2
6,274 (20.03%)
18,090 (14.44%)
≥3
6,340 (20.24%)
14,189 (11.32%)
Smoking status, n (%)
<0.001
Never
23,650 (75.50%)
87,961 (70.20%)
Ex-/current smoker
7,675 (24.50%)
37,343 (28.98%)
Alcohol consumption (times/week), n (%)
<0.001
None
25,280 (80.70%)
88,994 (71.02%)
1–2
4,058 (12.95%)
21,496 (17.16%)
≥3
1,987 (6.34%)
14,814 (11.82%)
Abbreviations: BMI, body mass index; CCI, Charlson Comorbidity Index; DM, diabetes mellitus; HTN, hypertension
PPI exposure and PD risk
Table 2 illustrates the odds of PPI exposure in PD. PPI users comprised 15,467 (49.4%) patients with PD and 55,407 (44.2%) healthy controls. PPI use was associated with an increased risk of PD (adjusted OR, 1.10; 95% CI, 1.07–1.13) after applying a 1-year lag period. We observed a significant positive dose-response relationship between the cDDDs of PPIs and PD development (p for trend <0.001), in which the highest dose-response effect was demonstrated among patients with a PPI exposure of ≥360 cDDD (adjusted OR, 1.36; 95% CI, 1.26–1.47). The results were similar after employing 2-year and 3-year lag windows.
10.1371/journal.pone.0295981.t002
Table 2 Association between PPI use and PD risk.
Category
Including a lag window of 1 year
Including a lag window of 2 years
Including a lag window of 3 years
Cases/controls
OR † (95% CI) *
Cases/controls
OR† (95% CI)
Cases/controls
OR † (95% CI)
Whole population (N = 156,629)
PPI non-user
15,858/69,897
1 (reference)
18,517/80,054
1.00 (reference)
21,817/91,973
1 (reference)
PPI user
15,467/55,407
1.10 (1.07–1.13) ***
12,808/45,250
1.11 (1.07–1.14) ***
9,508/33,331
1.09 (1.06–1.13) ***
By cDDD
0
15,858/69,897
1 (reference)
18,517/80,054
1.00 (reference)
21,817/91,973
1 (reference)
1–179
12,529/47,518
1.06 (1.03–1.10) ***
10,793/39,875
1.08 (1.05–1.11) ***
8,323/30,103
1.08 (1.04–1.11) ***
180–359
1,448/4,180
1.29 (1.20–1.39) ***
1,080/3,024
1.27 (1.16–1.38) ***
680/1,934
1.23 (1.10–1.37) ***
≥360
1,490/3,709
1.36 (1.26–1.47) ***
935/2,351
1.33 (1.21–1.47) ***
505/1,294
1.26 (1.11–1.43) ***
p for trend
<0.001
<0.001
<0.001
Male (n = 69,799)
PPI non-user
7,368/31,897
1.00 (reference)
8,523/36,254
1.00 (reference)
9,979/41,374
1 (reference)
PPI user
6,591/23,943
1.06 (1.01–1.11) *
5,436/19,586
1.07 (1.02–1.12) **
3,980/14,466
1.04 (0.99–1.10)
By cDDD
0
7,368/31,897
1.00 (reference)
8,523/36,254
1.00 (reference)
9,979/41,374
1 (reference)
1–179
2,513/10,163
1.03 (0.98–1.08)
2,243/8,781
1.05 (1.00–1.10)
1,758/6,938
1.03 (0.98–1.09)
180–359
1,972/7,491
1.24 (1.10–1.40) ***
1,678/6,231
1.19 (1.04–1.37) *
1,275/4,643
1.04 (0.88–1.24)
≥360
920/3,037
1.23 (1.08–1.39) *
701/2,349
1.23 (1.06–1.43) **
477/1,543
1.30 (1.06–1.59) *
p for trend
<0.001
<0.001
<0.001
Female (n = 86,830)
PPI non-user
8,490/38,000
1.00 (reference)
9,994/43,800
1.00 (reference)
11,838/50,599
1 (reference)
PPI user
8,876/31,464
1.13 (1.09–1.18) ***
7,372/25,664
1.14 (1.09–1.18) ***
5,528/18,865
1.13 (1.09–1.18) ***
By cDDD
0
8,490/38,000
1.00 (reference)
9,994/43,800
1.00 (reference)
11,838/50,599
1 (reference)
1–179
3,408/13,634
1.09 (1.05–1.14) ***
3,109/11,788
1.10 (1.06–1.15) ***
2,513/9,340
1.11 (1.06–1.16) ***
180–359
2,510/9,130
1.33 (1.20–1.46) ***
2,126/7,660
1.33 (1.19–1.49) ***
1,644/5,616
1.37 (1.19–1.57) ***
≥360
1,206/4,063
1.45 (1.32–1.61) ***
936/3,066
1.40 (1.24–1.59) ***
656/2,023
1.24 (1.06–1.46) **
p for trend
<0.001
<0.001
0.273
Age ≥50 years (n = 105,194)
PPI non-user
15,159/66,855
1.00 (reference)
17,722/76,713
1 (reference)
20,913/88,317
1 (reference)
PPI user
15,056/54,009
1.10 (1.07–1.13) ***
12,493/44,151
1.11 (1.07–1.14) ***
9,302/32,547
1.10 (1.06–1.13) ***
By cDDD
0
15,159/66,855
1.00 (reference)
17,722/76,713
1 (reference)
20,913/88,317
1 (reference)
1–179
5,675/22,965
1.06 (1.03–1.10) ***
5,159/19,882
1.08 (1.05–1.12) ***
4,138/15,764
1.08 (1.04–1.12) ***
180–359
4,377/16,210
1.29 (1.20–1.39) ***
3,724/13,588
1.27 (1.17–1.39) ***
2,872/10,045
1.23 (1.10–1.37) ***
≥360
2,095/7,006
1.36 (1.26–1.47) ***
1,614/5,352
1.34 (1.21–1.47) ***
1,119/3,529
1.26 (1.11–1.43) ***
p for trend
<0.001
<0.001
<0.001
Age <50 years (n = 151,079)
PPI non-user
699/3,042
1.00 (reference)
795/3,341
1 (reference)
904/3,656
1 (reference)
PPI user
411/1,398
1.09 (0.92–1.28)
315/1,099
1.00 (0.84–1.19)
206/784
0.94 (0.77–1.16)
By cDDD
0
699/3,042
1.00 (reference)
795/3,341
1 (reference)
904/3,656
1 (reference)
1–179
246/832
1.08 (0.92–1.28)
193/687
1.00 (0.84–1.19)
133/514
0.93 (0.75–1.14)
180–359
105/411
1.15 (0.60–2.21)
80/303
1.03 (0.50–2.09)
47/214
1.19 (0.49–2.85)
≥360
31/94
1.34 (0.49–3.62)
23/63
1.08 (0.27–4.32)
14/37
3.13 (0.26–38.13)
p for trend
0.005
0.137
0.418
Abbreviations: cDDD, cumulative defined daily dose CI, confidence interval; OR, odds ratio; PPI, proton pump inhibitor
*, p<0.05
**, p<0.01
***, p<0.001
† adjusted for the calendar year of the index date, Charlson Comorbidity Index score, smoking status, and alcohol consumption.
Subgroup analyses of the association between PPI use and PD risk
We also conducted subgroup analyses for the association between PPI use and PD risk. Subgroup analysis revealed that the PD risk with PPI use was higher in female than in male participants. PPI use increased the risk of PD in patients aged ≥50 years (OR, 1.10; 95% CI, 1.07–1.13); however, there was no significant risk in individuals aged <50 years (OR, 1.11; 95% CI, 0.93–1.31). PPI use was also associated with a significantly increased risk of PD in patients who did not consume alcohol rather than in those who consumed alcohol (1–2 times/week or ≥3 times/week). Further, PPI use was associated with a significantly increased risk of PD in individuals with a CCI score of ≥ 3 points compared with that in individuals with a CCI score of 0, 1, or 2 points as well as in non-smokers compared with that in ex- or current smokers ( Fig 2 and S3 Table ).
10.1371/journal.pone.0295981.g002
Fig 2
Adjusted odds ratios with 95% confidence intervals for Parkinson’s disease risk stratified by various clinical variables.
The OR and 95% CI were adjusted for the calendar year of the index date, CCI score, smoking status, and alcohol consumption. Abbreviations: BMI, body mass index; CCI, Charlson Comorbidity Index; CI, confidence interval; OR, odds ratio.
Discussion
In this nested case-control study, we evaluated the relationship between PPI use and PD risk using a Korean nationwide population-based dataset. Our results have demonstrated an association between PPI use and PD risk after applying a 2-year or 3-year lag window before diagnosis, with evidence of a dose-response relationship. Furthermore, older individuals (age ≥50 years) were more susceptible to the PPI-related risk of PD. Several hypotheses could explain our findings.
First, protopathic bias may exist because the agent of interest is used to treat the first symptom of an event. Patients with PD present with increased gastrointestinal dysfunction, which may have begun years before their PD diagnosis [ 24 ]. They may exhibit various symptoms, including nausea, abdominal bloating, constipation, early dyspepsia, upper abdominal pain, and weight loss, due to dysfunction of the PD-related autonomic and enteric nervous systems. Furthermore, abdominal symptoms of preclinical PD may increase hospitalization and clinic visits, leading to increased PPI use. Considering the characteristics of our patient group, a higher proportion of individuals with a CCI score of ≥1 points was observed in the PD group, and that group may have had increased PPI use. To reduce this bias, we applied various lag periods for the sensitivity analysis with adjustments for the CCI score. All analyses revealed a significant relationship between PPI use and PD risk.
The second hypothesis is related to the lipophilic properties of PPIs and their ability to cross the BBB. PPIs may affect Aβ metabolism, one of the pathological markers of Alzheimer’s disease. As PPIs can cross the BBB, Aβ degradation and deposition decrease and increase in the brain, respectively. Recent evidence has increasingly demonstrated the relevance of Aβ in PD and its possible role in related cognitive deficits [ 28 ]. PPI may also induce changes in the cholinergic system. A recent study [ 29 ] reported that it might limit the action of choline-acetyltransferase, which has an important role in the biosynthesis of cholinergic signaling substances, and damage to cholinergic interneurons caused by this action can lead to movement disorders, including PD.
Third, other possible mechanisms in the relationship between PPI use and PD risk could involve magnesium, vitamin B12, and iron deficiencies. Long-term PPI use may decrease the secretion of gastric acid, resulting in magnesium deficiency and decreased vitamin B12 binding protein levels, which may impair vitamin B12 and iron absorption [ 30 ]. A clinical study [ 31 ] reported that magnesium, via its neuroprotective action, reduces lipid peroxidation and regulates oxidative stress by inhibiting the generation of free radicals. Therefore, magnesium deficiency may increase calcium influx and cause neurotoxicity, thus leading to impaired neuronal function [ 31 ]. Decreased gastric acid secretion decreases acidity in the small intestine, leading to bacterial overgrowth and malabsorption [ 32 ]. Recent studies [ 33 , 34 ] have reported a relationship between vitamin B12 deficiency and PD risk. Moreover, patients with PD have lower vitamin B12 levels than healthy individuals; low vitamin B12 levels are associated with peripheral neuritis, cognitive impairment, and accelerated disease progression [ 34 ]. A study [ 35 ] that used iron-deficient anemic mouse models showed that iron deficiency can impair dopamine reuptake because tyrosine hydroxylase, which is iron-dependent, is important for dopamine synthesis.
Another possible mechanism is gut microbiota imbalance. PPIs can induce changes in gut microbiota composition; these changes have been confirmed in the stool samples of patients with PD compared with those of healthy controls [ 36 ]. Moreover, another study [ 37 ] demonstrated that the production of short-chain fatty acids, major metabolites of specific intestinal bacteria, decreased in patients with PD compared with that in controls, suggesting gut microbiota dysbiosis [ 37 ]. Consequently, this gut microbiota dysbiosis may be associated with pathophysiological changes in the gastrointestinal, enteric nervous, and central nervous systems.
Recent studies have not reported a clear relationship between PPI use and dementia, Alzheimer’s disease, and amyotrophic lateral sclerosis. However, we observed a significant relationship between PPI use and PD risk. Few studies have investigated the effect of PPI therapy on PD risk [ 9 – 11 ]. A study compared 3,026 patients with PD and 12,104 controls; comparison of current users (PPI usage time was 1 month before the index date) and past users (PPI usage time was 31–365 days before the index date), with ORs of 1.63 (95% CI, 1.44–1.84) and 1.12 (95% CI, 1.01–1.25), respectively, indicated a significant association between PPI use and PD risk. However, the aforementioned study only covered exposure to PPIs over a 1-year period; therefore, the long-term effects of PPIs on PD were not determined. In our study, we investigated the long-term effect by analyzing PPI use for 5 years and employing different lag periods. A case-control study [ 9 ] that compared patients with PD and controls (each group, N = 4,280) aged ≥65 years reported a significant relationship between PD and PPI use (OR, 1.15; 95% CI, 1.04–1.27). The results of the aforementioned study were also significantly associated with older age, which is consistent with our results. Similarly, a Danish study [ 10 ] that examined the relationship between PD and prior treatment for Helicobacter pylori infection showed an increased risk of PD among patients who were treated with PPIs alone based on a 0-year lag (OR, 1.26; 95% CI, 1.16–1.36) and 5-year lag (OR, 1. 23; 95% CI, 1.11–1.37) before PD diagnosis; the risk estimates were not modified by age stratification. However, in our study, subgroup analysis revealed age differences in the relationship between PPI use and PD risk. Furthermore, owing to differences in the clinical picture, progression, and response to drugs and based on the symptom onset period, PD can be classified as juvenile, young-onset (YOPD), and late-onset PD [ 38 ]. The onset age of YOPD is between 21 and 40–50 years [ 38 ]. We conducted a subgroup analysis with a 50-year-old cut-off value to investigate the effect of PPI use on YOPD. Our results show a significant relationship between PPI use and PD risk in patients aged ≥50 years but not in patients aged <50 years. Theoretically, genetic and environmental factors may be significantly involved in YOPD and sporadic PD, respectively. The onset age in PD is negatively associated with the risk of a genetic predisposition [ 39 ]; thus, PPI use is not associated with YOPD risk, which strongly involves genetic factors.
Furthermore, subgroup analyses were conducted to determine the relationship between PPI use and PD risk. Identifying high-risk groups for PD among PPI users is important for primary prevention ( Fig 2 and S3 Table ). As previously mentioned, PPI use significantly increased PD risk in patients aged ≥50 years; however, there was no increased risk in patients aged <50 years. In contrast, sex and BMI did not modify PD risk in PPI users. These results are similar to those of previous studies regarding the risk factors for PD. Many longitudinal studies identified no association between BMI and PD risk [ 40 , 41 ]; however, the lack of an association is inconclusive [ 26 ]. In our study, PPI use significantly increased PD risk in patients without a history of drinking or smoking. Similarly, PPI use was associated with a significantly increased PD risk in patients with a CCI score of ≥3 points compared with that in patients with a CCI score of 0, 1, or 2 points. Cigarette smoking is a well-known representative factor that plays a protective role in PD [ 42 ]. Previous studies evaluating the relationship between alcohol intake and PD risk have shown contrasting results [ 43 , 44 ]. However, recent studies reported that alcohol consumption at a light to moderate level is associated with a decreased risk of PD, whereas heavy drinking is not [ 45 , 46 ]. Therefore, the quantity of alcohol consumed is important for decreasing PD risk. Although measuring the exact amount of alcohol consumption by dividing alcohol consumption by the number of drinks per week is difficult, our study showed an increase in PD risk in non-drinkers. These results were consistent with those in previous studies. Biological protective components of alcohol may influence the risk of PD [ 47 – 50 ]. Therefore, drinking a modest amount of alcohol can be helpful and appears to have a more profound protective effect than a neurotoxic effect. SY Jung et al. also reported that PD risk was reduced in non-drinkers-turned-light drinkers, and light drinkers had an increased risk of developing PD when ceasing alcohol consumption [ 45 ]. Furthermore, they also reported that smoking and alcohol consumption may have a joint protective effect against PD; however, the optimal amount may differ by race and ethnicity. Further research is needed to elucidate the mechanism by which smoking and alcohol may interact in reducing future PD occurrence.
Our study has numerous strengths, including its large sample size and homogeneity of the study population. Furthermore, our study reflects the real-world pattern of PPI use using data from the NHIS database, which covers approximately 97% of the Korean population. Additionally, we adjusted for clinical variables that are risk factors for PD; risk behaviors, such as alcohol and smoking; and comorbidities, which were corrected using the CCI score. Moreover, we attempted to extensively exclude factors that could cause misclassification. Thus, our study provides supportive evidence for research to find possible mechanisms for the impact of PPIs on PD, which may offer insights into the pathogenesis of PD. Further research is also needed to evaluate groups at high risk for PD associated with PPI.
However, this study also has some limitations. First, although PD was defined according to the expert opinion of a neurologist and definitions from existing studies and cases with other codes that may have resulted in misclassification were excluded, we applied an operational definition of PD. Therefore, these diagnoses may be inaccurate and may lead to overdiagnosis and overestimation. Additionally, the inclusion of some Parkinsonism-like diseases may lead to bias in the results. Second, PPI use was determined by prescription claims; therefore, data regarding patient adherence to medication were unavailable. Third, we matched the patients with controls and adjusted for potential confounding factors; however, the possibility of prescription bias cannot be excluded in observational studies estimating long-term drug effects. Third, information on other confounders such as H2 blocker prescriptions and GERD that triggers PPI administration was not available for analysis in this study. Thus, future studies that examine the effect of PPI on PD risk with adjustments for those potential confounding factors would be necessary. Finally, the possibility of protopathic bias should be considered. Early PD diagnosis is necessary; however, at the time of PD diagnosis, it was difficult to determine whether PD was in the early or moderate stage. Accordingly, we applied various lag periods to reduce reverse causality bias. Furthermore, we conducted multiple regression analyses to adjust for potential confounding biases.
Conclusion
In conclusion, our findings revealed that PPI use was associated with an increased risk of PD. Additionally, a significant positive dose-response relationship existed between the cDDDs of PPIs and PD development, and older individuals (i.e., those aged ≥50 years) were more susceptible to the PPI-related risk of PD. Thus, greater caution may be required for older patients with comorbidities to reduce this risk. Nonetheless, our findings should be interpreted with caution, and future studies to validate and further characterize our findings are warranted.
Supporting information
S1 Table
Drug codes for proton pump inhibitors and Parkinson’s disease drugs.
(DOCX)
S2 Table
Modified Charlson comorbidity index based on the ICD-10 code.
(DOCX)
S3 Table
Risk of Parkinson’s disease in individuals with PPI exposure stratified by various clinical variables.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Approximately 70% of the world’s petroleum wells are situated in weakly consolidated reservoirs [ 1 ]. Consequently, many petroleum wells are susceptible to sand production that causes several problems, such as equipment damage and plugging, maintenance costs, and decline in reservoir recovery. Moreover, sand production can also lead to equipment erosion and safety, environmental, and health issues [ 2 – 5 ]. However, an early sand production prediction is highly recommended for the success of sand control management strategies [ 6 ].
There are three techniques used to predict sand production: numerical, analytical, and empirical methods. The numerical method includes the application of the finite element method, discrete element method, or finite difference method. However, the numerical methods are time-consuming and involve a complicated process. Furthermore, the input data required to obtain a numerical prediction (such as rock petrophysics, rock mechanics, and fluid properties) are challenging and laborious to find because of the need of some experimental data [ 7 ]. Similarly, the analytical methods have some drawbacks; they ignore stress anisotropy and assume symmetrical geometry and boundary conditions. Therefore, ignoring the main effect of stress anisotropy on sand, the method may not explain the sanding risk related to the orientation of the borehole. In general, assumptions or approximations are needed, making the models less reliable or accurate, even though they are complex [ 7 ].
On the other hand, empirical methods use well data and field observation to predict sand production. Sand prediction methods depend on field experiences to establish a correlation between sand production, well data, and field and operation parameters. In general, empirical methods are categorized into three types: one parameter, two parameters, and multiparameter correlations [ 6 ]. Tixier et al. [ 8 ] used acoustic log data to determine the shear modulus ratio to compressibility to obtain sand production. When the ratio is higher than 0.8 × 10 12 psi 2 , there is a lower probability of sand influx; and when the ratio is less than 0.7 × 10 12 psi 2 , there is a high probability of sand production [ 8 ]. Veeken et al. [ 9 ] applied a model with two parameters, which are the depletion reservoir pressure and drawdown pressure, as indicators for sand risk. Generally, increasing the number of parameters improves the accuracy of the sand prediction model [ 9 ].
Some models and correlations are available in the literature for the prediction of the critical total drawdown (CTD) that is used as an indicator of the onset of sand production. The CTD can be defined as the maximum difference between the reservoir pressure and the minimum well bottomhole flowing pressure that the formation can withstand without sand production. Some researchers used analytical models like Mohr Coulomb and modified Lade to predict the CTD; nevertheless, the models have some assumptions such as the formation rock mechanics properties are homogenous and isotropic [ 10 – 12 ]. Kanj and Abousleiman [ 13 ] used ANNs, feed-forward backpropagation network (BPN), and generalized regression neural network to predict the CTD using data of 31 wells from the Adriatic Sea. Multiple linear regression (MLR) and the genetic algorithm MLP (GA-MLR) were applied by Khamehchi et al. [ 6 ] to predict the CTD using data of 23 wells from the Adriatic Sea. However, these models are proven to have a lack of accuracy that reaches more than 20% error (AAPRE).
Numerous studies have used the fuzzy logic (FL) approach in petroleum engineering. Rezaee et al. [ 14 ] used petrophysical data and applied the FL tool to calculate shear wave velocity, which showed accurate predictions. In addition, Moradi et al. [ 15 ] used a FL approach to obtain the drilling rate. The FL model is proved to be more accurate than other models, such as Bourgoyne and Young models [ 15 ]. A FL tool was also developed to assist in selecting candidate wells for hydraulic fracturing treatment in a carbonate reservoir [ 16 ]. The FL model reduced the uncertainty that existed in the candidate well selection [ 16 ]. Ahmadi et al. [ 17 ] used the FL to calculate the breakthrough time of water coning in the fractured reservoirs. Akbarzadeh et al. [ 18 ] used a fuzzy model to predict conductivity; the fuzzy model was reported to be robust and accurate. Wang et al. [ 19 ] applied the FL to characterize reservoir heterogeneity and demonstrated that the model was accurate. The FL model was also used for forecasting petroleum economic parameters; the authors concluded that Mamdani type outperformed other models, such as autoregressive integrated moving average (ARIMA) [ 20 ]. Al-Jamimi and Saleh [ 21 ] used an FL tool to optimize the catalysts, and the FL model was successful in predicting catalyst performance. Artun and Kulga [ 22 ] used the FL to select candidate wells for refracturing in tight gas sand reservoirs. Karacan [ 23 ] used the FL model using 24 data points to estimate the recovery factors of miscible CO 2 . API, porosity, permeability, depth, hydrocarbon pore volumes (HCPV), net pay, initial pressure, well spacing, and S orw were included as features; the FL model was showed to be accurate [ 23 ].
This study aims to build a new robust and more accurate model for predicting the CTD by applying the FL. The developed model considers four parameters: total vertical depth (TVD), transit time (TT), cohesive strength (COH), and effective overburden vertical stress (EOVS). A trend analysis has been performed to investigate the accuracy of the physical behavior and trends of the model parameters. Furthermore, the performance of the model was compared with the most recent correlations.
Methodology
Data collection and description
This study has been performed in four phases: data collection and preparation, model development, trend analysis, and validation. A data set of 23 wells of the North Adriatic Sea was collected from the literature [ 24 ]. The data were split into two sections: for the first section, 70% of the data sets were allocated for developing the model, and for the second section, 30% of the data were used for verifying the model. Table 1 lists the data range and statistical analysis of the training and verification parameters.
10.1371/journal.pone.0250466.t001
Table 1 Data range and statistical analysis of the collected data for the developing FL model.
Parameter
TVD (m)
TT (micsec/ft)
COH (Mpa)
EOVS (Mpa)
CTD measured (Mpa)
Minimum
1070.000
85.000
0.539
10.885
0.314
Maximum
4548.000
170.000
5.217
80.709
43.973
Mean
2564.957
115.043
1.775
38.165
15.284
Median
2380.000
110.000
1.275
29.420
12.807
Range
3478.000
85.000
4.678
69.823
43.659
Skewness
0.187
0.940
1.234
0.398
0.600
Standard deviation
10.238
0.208
0.012
0.228
0.123
Fuzzy logic approach
Zadeh [ 25 ] invented a fuzzy set theory to handle data uncertainty. The benefit of using FL is that it considers the identification uncertainty present in any evaluation process in the developed model [ 26 ]. The FL can deal between zero and one, unlike the Boolean that can only take a zero or one. The fuzzy sets can provide gradual transitions from membership to non-membership [ 27 ]. The proposed fuzzy logic model offered high robust and reliable estimations and is thus well suited to other applications. The FL system is flexible and has a structure that can be modified. The FL methods can link human reasoning and concept formation through linguistic rules to obtain functions and control nonlinear systems. The FL can efficiently handle the complexity and uncertainty of the process with limited data [ 28 ]. The fuzzy logic can be applied with small data; hence they cannot occupy a huge memory space [ 29 ]. The fuzzy inference system contains five functional components, as illustrated in Fig 1 :
A fuzzification interface is used to convert crisp inputs to linguistic (fuzzy) variables applying membership functions.
A database is used to define membership functions.
A rule base contains several fuzzy if-then rules.
The IF of the rule describes a condition or assumption that is partly satisfied, whereas the THEN of the rule describes a conclusion or an action obtained when the conditions are hold true [ 26 ].
10.1371/journal.pone.0250466.g001
Fig 1
Fuzzy inference system for CTD model.
d The inference engine of the FL is a decision-making unit.
e The defuzzification interface is used to convert fuzzy outputs into crisp outputs [ 30 ]. The defuzzification can be conducted using some defuzzification methods such as the max or mean-max membership principles, the centroid method, and the weighted average method [ 31 ]. We used the weighted average method for this research.
The FL model in this study was developed by applying MATLAB R2020a. A membership function (MF) is known as the curve, which defines how each point in the input space can be designed to a membership value between (0–1) [ 32 ]. An MF can identify the fuzzy set by assigning a membership degree to each element [ 23 ]. Fuzziness is measured by using MFs as the fundamental constituents of the fuzzy set theory. The shape and type of MF should be accurately chosen because they impact the fuzzy inference system. Trapezoidal MFs were applied for the independent data because they show significantly enhanced outcomes compared to other MFs, whereas Gaussian MFs were used for the dependent data because they can be non-zero and smooth [ 33 ]. Gaussian MFs was used for this study. Table 2 presents the specifications of the FL MATLAB code used to obtain the CTD model.
10.1371/journal.pone.0250466.t002
Table 2 Specifications of the FL model.
Parameter
Description/value
Fuzzy structure
Sugeno-type
Initial FIS for training
genfis3
Membership function type
Gaussian MF
Output membership function
linear
The number of membership functions
4
The fuzzy rules
if x is A and y is B then z = f(x , y)
The fuzzification
Gaussian
The defuzzification
weighted average method
The number of clusters
4
Number of inputs
4
Number of outputs
1
Training epoch number
500
Radii
1.1585
Results and discussion
Two tests were performed to evaluate the proposed FL model. First, the FL approach was tested to show that it is robust and follows physical behavior trends by applying trend analysis. Second, the performance of the proposed FL approach was compared with the current correlations. Cross-plots and statistical error analyses, such as correlation coefficient (R), average percent relative error (APRE), average absolute percentage relative error (AAPRE), root mean square error (RMSE), and standard deviation (SD), were conducted.
Trend analysis
The trend analysis was performed to test the robustness of the model in the presence of uncertainty. The trend analysis is used to indicate the relationships between input and output variables in the model. The trend analysis defines errors in the models to show unexpected relationships between input and outputs, which highlights the need to display the reliability of the models. Furthermore, the trend analysis identifies and removes the unnecessary parts of the model structure [ 34 ]. Moreover, the trend analysis was used to identify significant connections among observations, model inputs, and predictions, guiding the development of robust models [ 35 ]. Therefore, the trend analysis is essential for this study.
The selected input parameter for investigation is varied between the minimum and maximum value, while other parameters are kept constant at their mean values [ 36 – 38 ]. Graphs are plotted for the input parameter values (x-axis) against the output CTD (y-axis) for the previous models and the FL model. Four input parameters, TVD, TT, COH, and EOVS, have been selected for the trend analysis.
Fig 2 shows the trend of TVD. Kanj and Abousleiman [ 13 ] correlation ( Fig 2 ) shows that the CTD was independent of TVD, because it is based only on the cohesive strength (COH). The TVD trend of the FL model obeys the trend of the existing correlations, as shown in Fig 3 . Ahad et al. [ 39 ] stated that older rocks can be more consolidated. On the other hand, shallow formations can be weakly consolidated [ 39 ]. Therefore, increasing the depth will increase the CTD.
10.1371/journal.pone.0250466.g002
Fig 2
TVD trend analysis of the FL model and previously published models.
10.1371/journal.pone.0250466.g003
Fig 3
TVD trend analysis of the FL model.
Fig 4 indicates that the TT is inversely proportional to the CTD, as illustrated by all the previous models except Kanj and Abousleiman [ 13 ] correlation, which demonstrates that the CTD is constant as they did not include the TT. As a result, Kanj and Abousleiman [ 13 ] correlation failed to represent the behavior accurately. The FL model also follows the trend of the existing correlations ( Fig 5 ), indicating the proper trend for the TT. The shorter TT implies that the sand is more consolidated [ 40 ]. Consequently, decreasing TT will increase the CTD.
10.1371/journal.pone.0250466.g004
Fig 4
TT trend analysis of the FL model and previously published models.
10.1371/journal.pone.0250466.g005
Fig 5
TT trend analysis of the proposed FL model.
Fig 6 indicates that the cohesive strength (COH) is directly proportional to the CTD. Kanj and Abousleiman [ 13 ] correlation followed the trend of existing correlations, but the CTD is negative (−2.57 MPa) when the COH is 0.539 MPa. Consequently, Kanj and Abousleiman [ 13 ] correlation has not proven a proper trend for the CTD correlation ( Fig 6 ). The FL model shows that the COH trendobeys the trend shown by the correlations in the literature where the COH is directly proportional to the CTD ( Fig 7 ). Hence, the FL model is successful in following the accurate trends. The formation failure decreases the strength of rock and causes sand production [ 41 ]. The cohesive strength increases the degree of cementation [ 42 ]. Increasing the cementation degree of sand grains can lead to a decrease in sand production. Thus, increasing the rock’s cohesive strength results in in rising the CTD.
10.1371/journal.pone.0250466.g006
Fig 6
COH trend analysis of the proposed FL model and previously published models.
10.1371/journal.pone.0250466.g007
Fig 7
COH trend analysis of the FL model.
The trend of the EOVS is illustrated in Fig 8 . The CTD follows an inverse relationship with EOVS. However, Kanj and Abousleiman [ 13 ] correlation displays a horizontal line, which indicates that their correlation does not consider the EOVS parameter. The trend expressed by the FL model is also shown to follow the trend of the previous correlations; Fig 9 indicates that it represents the proper trend for the EOVS. The overburden stress stays constant; however, when the pore pressure declines, the effective overburden stress must rise [ 42 ]. The critical drawdown pressure decreases with the decline in pore pressure [ 43 ]. Therefore, increasing EOVS decreases the CTD.
10.1371/journal.pone.0250466.g008
Fig 8
EOVS trend analysis of the FL model and previously published models.
10.1371/journal.pone.0250466.g009
Fig 9
EOVS trend analysis of the proposed FL model.
To summarize the trend analysis, all the input parameters (TVD, TT, COH, and EOVS) of the developed FL model can follow the correct trends, indicating the FL model’s reliability. Nevertheless, Kanj and Abousleiman [ 13 ] correlation trends fail to present the behavior correctly.
The comparison of FL model and current models
Cross-plotting analysis
The proposed FL model and current correlation cross-plots were presented. A 45° straight line is illustrated on the cross-plot of the measured and expected CTD values. The closer the plotted data points to the straight line, the higher the correlation or model’s accuracy.
Fig 10 illustrates the cross-plotting of the testing data set of the FL model, and Fig 11 illustrates the cross-plotting comparison of the FL model with the existing correlations. As shown in Figs 10 and 11 , the FL model represents the highest accuracy and can predict the CTD with the coefficient determination (R 2 ) of 0.9947.
10.1371/journal.pone.0250466.g010
Fig 10
Cross-plot of testing FL model.
10.1371/journal.pone.0250466.g011
Fig 11
Cross-plot comparison of the proposed FL model with the previously published models.
Statistical error analysis
The statistical error analysis is performed to verify the FL model’s accuracy and compare it against the current correlations. The statistical parameters used in this research are correlation coefficient (R), APRE, AAPRE, SD, RMSE, maximum absolute percent relative error ( E max .), and the minimum absolute percent relative error ( E min. ), as included in the S1 Appendix. The AAPRE and correlation coefficient (R) are used as indicators.
Fig 12 illustrates the AAPRE and correlation coefficient (R) comparison of the FL model with the existing models. As shown in Fig 12 , the proposed FL model has the lowest AAPRE of 8.647% and the highest correlation coefficient (R) of 0.9947. Khamehchi et al. [ 6 ] (GA-MLR) model shows the AAPRE (%) of 22.644% and correlation coefficient (R) of 0.9827, whereas Khamehchi et al. [ 6 ] (MLR) model shows the highest AAPRE of 30.485%.
10.1371/journal.pone.0250466.g012
Fig 12
Correlation coefficient (R) and AAPRE (%) comparison of the proposed FL model with the previously published models.
The published predictions of the performance correlations were compared with the proposed FL approach, as shown in Fig 13 . Statistical error analysis has been conducted to test the robustness of the proposed FL model. The FL model also has the lowest RMSE and SD compared to other models ( Fig 13 ). This comparison of all correlations and the FL model provides essential means for validating the performance of the proposed FL model. Investigation of these statistical error analyses indicates that the FL model outperforms all the existing correlations. The AAPRE and correlation coefficient (R) are taken as the primary indicators of accuracy in this study. Khamehchi et al. [ 6 ] (GA-MLR) correlation is ranked as the second correlation; it has an AAPRE of 22.644% and a correlation coefficient (R) of 0.9827. Khamehchi et al. [ 6 ] (GA-MLR) and (MLR) models show AAPRE of more than 20% ( Fig 13 ).
10.1371/journal.pone.0250466.g013
Fig 13
Comparison of the statistical parameters for the proposed FL model with other models.
(a) RMSE, SD, and R; (b) APRE, AAPRE, AAPRE max , and AAPRE min .
Conclusions
An FL model was developed for predicting the CTD in sand formations in oil and gas wells. Different techniques, such as trend analysis, cross-plotting, and statistical error analysis, were used to validate the model. The prediction outcomes were compared with the published models available in the literature. Based on the obtained results, the following conclusions are emphasized:
The FL model could accurately describe the proper trend of the CTD as a function of all the considered independent variables (i.e., TVD, TT, COH, and EOVS). The model is observed to follow the actual trend as expected from the physical relationship.
The FL model has provided the best CTD estimations as compared to other available correlations. The FL model presented the highest correlation coefficient of 0.9947, the lowest AAPRE of 8.647%, the lowest root mean squared error of 0.014, and the lowest SD of 0.082 compared to the published correlations. The model accuracy and reliability have further been enhanced by data randomization to ensure that each data set does not memorize the pattern and avoid generalization and model overfitting.
The FL model has shown an AAPRE of 8.647%, whereas the existing models have reported values higher than 20%.
Supporting information
S1 Appendix
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
As part of the Neolithic agricultural revolution, the domestication of cattle, which occurred about 8,000 years ago, changed the social and economical life of most human populations [1] and contributed to the gradual transition of hunter gatherers into farmers with permanent settlements. Although fully interfertile, we distinguish two taxa of domestic cattle - humpless taurines ( Bos taurus ) and humped zebuines ( Bos indicus ) [2] . Archeological evidence mostly osteometric and morphometric data early argued in favor of several separate cattle domestication events [3] which was supported by more recent genetic analyses based on mitochondrial DNA [4] , [5] and Y-chromosome polymorphisms [6] . However the number of domestication centres remains a source of intense debate and disagreement. Two main hypotheses have been formulated [2] : (i) two domestication events: a first major domestication event of taurine cattle in the Fertile Crescent ( i.e. between the Mediterranean sea and Iran) from the wild extinct aurochs Bos primigenius primigenius and a second separate one which lead to zebus in the Indus valley (including Rajasthan and present day Pakistan) from the wild extinct aurochs B. p. namadicus [7] and (ii) a total of three domestication events: the two previous ones and a third one leading to African taurines in Northeastern Africa from the wild extinct aurochs B. p. opisthonomous [3] , [8] . During the 3,000–4,000 years after domestication, cattle expansion had followed different and complex routes tightly related to the migration of early breeder populations and the spread of agriculture over Europe, Africa and Asia [3] . Nevertheless, the overall diffusion of populations was estimated to take place at a slow and continuous rate of around 1.1 km per year [9] . Throughout Europe, early breeders presumably spread from the Fertile Crescent towards North-West following two distinct routes [3] . One group of farmers progressed to the North along the Balkans' rivers (following the so-called Danubian route) establishing the Neolithic culture in Germany and the Netherlands approximately 6,500 BP. A second group migrated to the West via maritime routes across the Mediterranean Sea (following the so-called Mediterranean route) establishing the Neolithic culture in Italy 6,500 BP or in Spain and France 6,000 BP [3] . Nevertheless, a number of secondary livestock migrations might have accompanied human migrations in more recent historical times. Similarly, during these migration waves, some sporadic events of interbreeding between wild European aurochs ( B. p. primigenius ), which had been present until the Middle Age, and domestic stocks might have occurred to a substantial extent [2] , [4] , [10] . These complex origins of cattle associated with both natural and artificial selection gave rise to numerous different breeds displaying a broad phenotypic variety over a short period of time. In France, it is generally believed that some aspects of the Neolithic culture originated from central Europe and also via the Mediterranean route. A fine scale characterization of the genetic structure of French cattle breeds might thus be expected to display footprints of such migrations and provide in turn additional insights into the establishment of the Neolothic culture.
The recent advent of high-throughput and cost effective genotyping techniques makes it possible to provide a detailed genome wide assessment of the genetic structure and relationships among cattle populations. This might in turn allow to refine previous pioneering works usually performed on a small number of genetic markers ( e.g. [11] , [12] ). We present in this study a detailed analysis of cattle diversity based on 1,121 individuals sampled in 47 populations from different parts of the world (with a special focus on French cattle) genotyped for 44,706 autosomal SNPs. More precisely, the data set consisted of new genotypes for 296 individuals representing 14 French cattle breeds which were combined to those available in three previous published studies: i) 19 populations sampled by the Bovine Hapmap Consortium [13] and genotyped with the Illumina® BovineSNP50 chip assay [14] , ii) 11 African populations [15] and iii) 3 French dairy cattle breeds described in [16] . After characterizing SNP polymorphism in the different populations, we performed a detailed analysis of genetic structure at both the individual and population levels. This confirmed a clear partitioning of cattle diversity into distinct breeds. In addition, the overall pattern of differentiation among three main groups of populations (African taurine, European taurine and zebuine cattle) may provide some additional support for three distinct domestication centres. We further searched for spatial patterns of genetic diversity among 23 European populations, most of them being of French origin under the recently developed spatial Principal Component analysis (sPCA) framework [17] .
Results and Discussion
SNP data, polymorphism and Linkage Disequilibrium
We first performed a joint analysis based on SNP data generated for all 1,121 individuals representing 47 different populations (24 individuals per population on average) genotyped for 44,706 SNPs from this study and three previously published studies (see Materials and Methods and Table S1 ) to provide a global picture of cattle genetic diversity. As detailed in Table S1 and in agreement with previous studies [13] – [15] , SNP average heterozygosity was found higher in populations from European origin (from 0.2544 for JE2 to 0.3156 for PRP) compared to zebu cattle (from 0.1556 for GIR to 0.1945 for ZMA) and taurines from West Africa (from 0.1828 for LAG to 0.2240 for SOM) or East Africa (0.2432 for SHK). As previously discussed [14] , [15] , this trend might be directly related to the ascertainment bias introduced in the construction of the BovineSNP50 chip assay, SNPs being almost exclusively derived from sequences available in European cattle breeds. Using such data to infer genetic divergence among cattle might thus be done cautiously and is expected to bias the estimation of genetic divergence between more distantly related populations ( e.g. European and African taurines or zebus). Interestingly, populations of hybrid origin displayed generally higher levels of polymorphism than their population of origin as exemplified for i) the two synthetic breeds, BMA and SGT, which result from crosses between European taurines and zebus [13] ; ii) the West African hybrids (BOR and KUR) and to a lower extent West African zebus (ZBO and ZFU) which result from crosses between West African taurines and zebus and iii) the Moroccan breed OUL which has a probable hybrid origin between European and African taurine (see below).
As shown in Figure S1 and previously reported ( e.g. [18] ), the average within-population pairwise r 2 dropped quickly toward its asymptotic value when physical distances were above 200 kb. In our data set, SNP genome coverage was homogeneous with a very small proportion of inter-marker distances less than 20 kb ( Figure S2 , Table S2 and Materials and Methods ) and with 3.5 SNPs (from 0 to 9) per 200 kb on average. Thus most SNP pairs in this study displayed a level of within population Linkage Disequilibrium (LD) close to that observed between unlinked SNPs. We thus did not consider in the following any spatial dependencies among SNPs which might result from LD and subsequently carried out descriptive analyses to further assess the structure of genetic variability at both the individual (ignoring the information on the population of origin) and population levels.
Assessing the genetic structure at the individual level
We first carried out a principal component analysis (PCA) based on all available SNP information allowing to refine and extend previous reports [13] , [15] . In particular, from a worldwide perspective, the data from [14] and [15] were very complementary since the 19 Hapmap populations [14] (also analyzed with 37,470 SNPs in [13] ) contained only two populations from Africa (ND3 from Western Africa and SHK from Eastern Africa). On the other hand, 10 African populations (eight from Western Africa, one from Northern Africa and one from Madagascar) were surveyed by [15] but (pure) zebuine populations were lacking in this data set. As shown in Figure 1 , the first component which accounted for 10.17% of variation resulted in the separation of the underlying populations according to a zebuine/taurine gradient while the second one (accounting for 4.98%) could be interpreted as a European/African taurine gradient. The resulting 2-Dimensional global organization of cattle genetic diversity might thus be described as a triangle with apexes corresponding respectively to European taurines (EUT), West African taurines (WAT) and Zebus from Indian origin (ZEB). Following this representation, OUL individuals lay as expected on the EUT/WAT segment, BMA and SGT individuals on the EUT/ZEB segment (but closer to EUT) and West African hybrids (BOR and KUR) and zebus (ZBO and ZFU) on the WAT/ZEB segment. Similarly, SHK which is considered as a taurine population because humpless was positioned close to West African zebuine populations confirming previous reports [19] .
10.1371/journal.pone.0013038.g001 Figure 1
PCA results obtained with the whole data set (1,121 individuals, 44,706 SNPs).
Individuals are plotted according to their coordinates on the first two principal components. Ellipses characterize the dispersion of each breed around its center of gravity (assuming the cloud is a random sample distributed according to a bivariate gaussian distribution, the probability for an individual to be within the ellipse is 0.9).
Besides, the neighbor-joining (NJ) tree based on Allele Sharing Distances (ASD) unambiguously separated individuals according to their population of origin ( Figure 2 ) as confirmed when applying simple assignation tests (data not shown). As a consequence, at a higher hierarchical level, the three groups of populations corresponding to EUT, WAT and ZEB could also be clearly distinguished (respectively in upper, lower right and lower left Figure 2 ). In agreement with PCA results, individuals from West African hybrid populations (KUR, BOR) and West African zebus (ZFU and ZBO) branched in an intermediary position between WAT and ZEB; and OUL branched between EUT and WAT. Similarly, ZFU and ZBO were closer to ZMA and BMA and SGT branched within the EUT suggesting a lower influence of zebus than European taurines. In addition, among some of the closely related European cattle populations (similar breeds but different sample origin), BRU and BSW individuals, HOL and HO2 individuals and JER and JE2 individuals were almost indistinguishable suggesting that each of these different global populations might be considered as single populations as previously shown for the Holstein population [20] . However, a notable exception was observed for CHA and CHL individuals that were clearly separated.
10.1371/journal.pone.0013038.g002 Figure 2
Neighbor-Joining tree relating the 1,125 individuals (1,121 cattle and 4 american bisons).
The tree was constructed using allele sharing distances averaged over 44,706 SNPs. Edges are colored according to the individual breed of origin.
We finally performed a model-based unsupervised hierarchical clustering of the individuals using the program frappe [21] . As shown in Figure 3 (e.g. K = 3 and K = 4), results were in good agreement with above observations with a clear separation of EUT, WAT and ZEB. In addition, increasing the number of inferred clusters allowed to confirm the high admixture level and assumed origin of some populations (see above) such as West African hybrids (BOR and KUR) and zebus (ZBO and ZFU), SHK (which displayed similar characteristics as West African zebus [19] ), OUL and synthetic breeds. Interestingly, among WAT, LAG (representative of shorthorn African taurines) individuals could be clearly (from K = 6) separated from ND3, ND2 and ND1 (representative of longhorn African taurines). These two later populations displayed a low level of zebu admixture not detected previously [15] . Nevertheless, SOM and BAO which are West African shorthorn taurines displayed a high longhorn influence. Similarly, the African taurine influence of OUL seemed to be of longhorn origin. Among ZEB populations, individuals belonging to ZMA separated (K>6) from those belonging to the other three ones (GIR, BRM and NEL) which were imported from India to Brazil (GIR and NEL) or USA (BRM) about 200 years ago [2] . The zebu genetic fraction of the West African hybrids (BOR and KUR) and noticeably SHK seemed to equally originate from these two zebus cluster ( e.g. K = 10) while the infusion of zebus in SGT and BMA was of South American origin consistent with recent historical data [2] . Except for RMG and to a lower extent CHL (Charolais individuals sampled in the United Kingdom), EUT displayed no evidence of WAT or ZEB introgression (K = 6). The influence of WAT or ZEB in RMG (or other Italian breeds) has already been reported based on other kind of genetic data with i) the segregation of the T1 mtDNA haplogroup [10] or ii) the segregation of zebu associated microsatellite alleles [22] .
10.1371/journal.pone.0013038.g003 Figure 3
Unsupervised hierarchical clustering of the 1,121 individuals genotyped for 44,706 SNPs.
Results for an inferred number of clusters K varying from 2 to 6, K = 10 and K = 47 ( i.e. the number of breeds) are shown.
Finally, individuals belonging to some breeds such as HOL (and HO2), ANG (and RGU), HFD, BSW (and BRU) tended to be clearly assigned to a single cluster when K = 10. This latter trend was further confirmed for most populations (although not all) when increasing the number of inferred clusters toward the number of populations (K = 47).
Assessing the genetic structure at the population level
Overall, the different analyses performed at an individual scale suggested that the partitioning of cattle into distinct populations is relevant to characterize genetic diversity. This was expected, in particular when considering breeds originating from industrialized countries, and has already been reported for the 19 Hapmap populations [13] . Consistently, the F ST across populations was found equal to 0.190 with an average F IS almost null (−0.007) leading to an F IT of 0.185. Note that within all populations, F IS were also found close to zero (as close relationships among individuals were avoided during sampling) although moderately negative values (<−0.1) were observed for ND2 [15] and BLO ( Table S3 ). F ST computed for each pair of populations ( Table S4 ) ranged from 0.0044 (for HO2/HOL pair) to 0.4742 (for LAG/NEL pair) while within EUR populations, it ranged from 0.0044 (for HO2/HOL pair) to 0.2018 (for JE2/BLO pair).
We thus decided to extent the PCA described above by performing a so-called between-class PCA ( e.g. [23] ), classes being in our context identifiable to breeds. This latter analysis explicitly introduces population structure information in the PCA optimization criterion to find those axes that maximize the between-breed variance. Running both PCA and between-breed PCA allows under a model-free basis i) to compare the patterns of differentiation when performing the analyses at an individual level or at a population level and ii) to quantify the proportion of the total between-individual variance explained by the partitioning of genetic variability into breeds (between-breed variance). As shown in Figure S3 , results from the between-breed PCA were highly similar to those obtained with the PCA on individual genotypes, the first three axes explaining more than 51.4% of the genetic variability across populations. Note that the third axis might be interpreted as North/South gradient among EUT (see below). Moreover, the correlations between the first twelve eigenvectors from PCA and between-breed PCA were almost equal to 1 (in absolute value). Thus, analyses at both the individual (PCA) and breed (between breed PCA) levels revealed highly similar patterns of population differentiation. Finally, comparing variances ( i.e. sum of the eigenvalues) among the two analyses showed that genetic variability between populations explained 32.8% of the whole genetic variability (across individuals).
We finally constructed a neighbor-joining (NJ) tree based on Reynolds genetic distances relating the 47 cattle populations and including American Bison (OBB) as a rooting outgroup ( Figure 4 ). Given the amount of available information almost all nodes were highly reliable (node bootstrap values above 95%) and EUT (blue), WAT (green) and ZEB (orange) populations could be clearly separated. As expected, within WAT, longhorn taurines (ND1 and ND3) were separated from shorthorn taurines (BAO, SOM and LAG) although ND2, probably because of a higher zebu introgression (see Figure 3 ) branched immediately below the WAT node. Among African zebus and in agreement with above observations, ZMA was closer to zebus from Indian origin (NEL, GIR and BRM) than West African zebus (ZBO and ZFU) while SHK, BOR and KUR were in an intermediary position between WAT and ZEB. Likewise, SGT and BMA were in an intermediary position between ZEB and EUR although closer to EUR (see above).
10.1371/journal.pone.0013038.g004 Figure 4
Neighbor-Joining tree relating the 47 cattle populations and American bison (OBB) outgroup based on Reynolds genetic distances computed using allele frequencies at 44,706 SNPs.
Reliability of the nodes (percentage over 100 bootstrap samples) are indicated for each node.
Overall, these different analyses support a global partitioning of world-wide cattle diversity into EUT, WAT and ZEB. Although not excluding the possibility of strong founder effects, this might suggest three distinct domestication events [2] . This latter hypothesis is supported by genetic data based on mtDNA showing that the predominant haplogroup (namely T1) in Africa is absent in Europe and at a relatively low frequency in Anatolia and Middle East [10] , [24] . Similarly, non recombining part of the cattle Y chromosome allowed the identification of three main haplogroups referred to as Y1 and Y2 for taurines and Y3 for zebus. Most haplotypes identified in African taurines are assigned to haplogroup Y2 and are not present in other continents [6] , [25] . Interestingly, in the NJ tree ( Figure 4 ), the OBB root clearly isolated EUT from the WAT/ZEB group which disagrees with the assumption of a common domestication center for EUT and WAT. Nevertheless, such a topology might be sensitive to the ascertainment bias favoring SNP from European origin. For instance, Figure S4 shows the NJ tree resulting from the analysis of a subset of 27,527 SNPs of an origin ancestral to the ZEB, WAT and EUT breeds separation. These were indeed chosen to display a MAF>0.01 in at least i) two breeds among ZMA, GIR, BRM and NEL ii) two breeds among ND1, ND3, SOM, BAO and LAG and iii) two breeds among 24 EUR. The positioning of the OBB root separated ZEB from the lower order group formed by WAT and EUR in agreement with the NJ tree based on ASD distances between individuals on the complete data set ( Figure 2 ).
A focus on the genetic structure of European breeds
Within EUR, grouping of the breeds on the NJ tree ( Figure 4 ) was strikingly consistent with their geographical origin. Hence four main groups of breeds were found highly reliable (node bootstrap values equal to 100%) and lead to distinguish a) breeds from southwestern France (starting from the tips: AUB and SAL then BLO, GAS and LMS); b) breeds from Eastern French mountains (MON and ABO then TAR, VOS); c) breeds from the Channel Islands (JER and JE2 then GNS) and d) breeds from Northern European origin (ANG, RGU, HO2, HOL and NRC) together with the two French breeds MAN and PRP (see below). The Charolaise breeds (CHA and CHL) and the Brown Swiss breeds (BSW and BRU) branched with a less reliable node with groups a) and b) respectively. However, when considering tree resulting from the subset of 27,527 SNPs described above ( Figure S4 ), they branched after the merging of groups a) and b). The resulting node joining a) and b) then displayed a higher bootstrap value (equal to 89%) than the node joining a) (and CHA and CHL) and b) (and BSW and BRU) on the tree of Figure 4 (node bootstrap value equal to 76%). Similarly in this latter tree, the two Italian breeds (RMG and PMT) and OUL branched with the group merging a) and b). Although OUL was also found close to RMG (relatively to the other EUR ones) in PCA ( Figure 1 ), this positioning might be affected by ascertainment bias since these two breeds displayed substantial influence from African taurines and zebus respectively. Indeed and more consistently, in the tree of Figure S4 , OUL and RMG were clearly separated from other EUR breeds. Subsequently, as shown in Figure 4 , MAR (from northwestern France) was found at an intermediary position between the South European breeds (Italian breeds, groups a) and b)) and breeds from the Channel Islands (group c). This large resulting group finally branched with NOR and BPN (both originating northern than MAR) and group d). Finally, HFD was surprisingly the outgroup of other EUR breeds in Figure 4 . Yet and more expectedly, HFD branched (node bootstrap value equal to 71%) with other North European breeds (constituted by groups c and d defined above with BPN and NOR) in the tree of Figure S4 .
Overall, among European cattle, both NJ tree and PCA results suggested strong spatial patterns of genetic diversity. Hence, as pioneered by Cavalli-Sforza and collaborators for the reconstruction of the early history of human populations [26] , interpreting the structure of such genetic structure in the light of geographical data is expected to provide insights into the underlying history of cattle [12] . However PCA does not take explicitly into account spatial information while grouping of populations based on the NJ tree is sensitive, to some extent, to ascertainment bias. We thus further searched for spatial patterns of genetic diversity under the recently developed sPCA framework [17] , concentrating on French and other closer related European breeds.
Spatial patterns of genetic diversity in French Cattle breeds
PCA does not explicitly incorporate geographical information because the optimization criterion relies on the maximization of the genetic variance. Thus, PCA may fail to detect spatial structuring if this is not associated with the most pronounced genetic differentiation. Recently, after the works of [27] and [28] ; Jombart and collaborators [17] specifically developed a sPCA devoted to the analysis of allele frequency data, and showed that it performed better than PCA in retrieving simple spatial structures as well as more complex patterns among genotypes or populations. We thus used this approach to reveal the spatial patterns of genetic variation in French cattle breeds (in relation to other European ones) since they were particularly well represented in our data set. Hence, out of the 29 populations of European origin ( Table S1 ), we only considered 23 breeds, discarding RGU (which derived from ANG), BRU (which is from the same origin as BSW), HO2 (which is from the same origin as HOL), JER (which is from the same origin as JE2), CHL (which derived from CHA) and RMG due to its zebu influence (see above). Geographical breed locations ( Table S1 ) were summarized using a Gabriel neighboring graph which models the spatial structure of the breeds ( Figure S5 ).
As detailed in Table S5 , the first sPCA eigenvalue was strikingly large compared to the others and similar in magnitude to the first PCA eigenvalue. In addition, the genetic variance associated to the first PCA component was found only slightly higher (13% of the total variance) than the corresponding sPCA one (12% of the total variance). Correspondingly, spatial autocorrelation on the first axis, as measured by the Moran's I [29] was high in both analyses ( I 1 PCA = 0.73 and I 1 sPCA = 0.87). Hence the first axis in both PCA and sPCA unambiguously captures global spatial patterns while separating populations according to a North/South gradient ( Figure S6A ). Nevertheless, on subsequent PCA axes, spatial autocorrelation appeared very low ( I 2 PCA = 0.14 and I 3 PCA = 0.06) while the second and third sPCA axes displayed a Moran's I above 0.5 ( I 2 sPCA = 0.63 and I 3 sPCA = 0.80). This suggested that PCA might fail to identify relevant spatial patterns on these additional axes making it difficult to interpret the underlying variance in terms of geography. We thus focused subsequently on the first three sPCA axes ( Figure S6 ). Note that some axes, such as axis 22 ( I 22 sPCA = −0.63) displayed a relatively high negative spatial autocorrelation suggesting a strong local spatial pattern. This axis actually separated JE2 and GNS (data not shown) which are closely geographically related (Channel Islands) but most probably because of complete isolation of JER since the 18 th century [2] are clearly genetically distinct (see above).
The coordinates of each breed on the first three sPCA axes were synthesized on Figure 5 by means of colorplots [30] , [31] projected on the geographic map. Based on the different colors obtained (see also Figure S6D for a 3D representation), four groups of breeds showed high geographical consistency in good agreement with the NJ tree results ( Figure 4 ). The underlying four groups of colors were i) the dark green one which comprises 5 breeds (LMS, SAL, AUB, BLO and GAS) from central and southwestern France, ii) the light green one which comprises 4 breeds (JE2, GNS, NOR and BPN) from the Channel Islands and northwestern France iii) the blue one which comprises 6 breeds (VOS, MON, ABO, TAR, BSW, and PMT) from Eastern France and the Alps and iv) the brown red one which comprises 6 breeds (ANG, NRC, HFD, HOL, PRP, MAN) from Northern Europe origin. However, CHA and MAR remained difficult to assign to one of these four groups owing to their low scores on the first three sPCs.
10.1371/journal.pone.0013038.g005 Figure 5
Projection on a map of Europe of the colorplots synthesizing the breed coordinates on the three first sPCA principal components.
These plots can show up to three coordinates at the same time by translating each coordinate into a channel of color (Red, Green, and Blue). The obtained values are used to compose a color under the RGB system. The Danubian (solid line) and Mediterranean (dashed line) migration routes are also reproduced on the map [3] .
The grouping of French breeds is mostly in agreement with previous classification based on historical data, morphological characters (mostly craniometric and morphometric data), geographical proximity [32] and blood groups, transferrin and β-casein polymorphisms [33] . In addition, they are quite consistent geographically. Two notable exceptions are represented by the PRP and MAN which belong to the North European breed group (iv) although originating from an area closer to i) and ii) confirming results from a previous study based on microsatellite markers [34] . These inconsistencies between genetic and geographical data are actually expected since the PRP has been recently derived from the red Holstein and the Meuse-Rhin-Yssel breed from Germany. Similarly, introgression of British Durham during the 19 th century had been extensively reported in MAN. The grouping of the NOR with the Channel Islands and Northwestern France breeds (ii) used to be more controversial [33] although in agreement with early classification based on biochemical markers [35] . Finally, it should be noted that in our study the position of CHA differed according to the methods used. Consistently, some historical data established a connection between CHA and the South-Western France blonde breeds as in the tree of Figure 4 [33] while the most commonly accepted theory used to associate CHA with the Jurassic group (represented in our study by MON and VOS in our study) [32] . Overall, sPCA results might be related to archeological and historical data [3] . Indeed the geographical positions of the four identified groups are in agreement with the early postulated migration routes by which the Neolithic culture expanded towards France (see Introduction and [3] ). Hence, our groups iii) and iv) appeared closely related to the Danubian colonization route while groups i) and ii) might correspond to the Mediterranean colonization route ( Figure 5 ). [2] , [4] , [10] . However, adding more European populations to our combined data set remains of paramount importance to further demonstrate the influence of the postulated migration routes on the structure of French cattle populations.
More generally, including data for populations from other parts of the world ( e.g. Southern Europe, Northern Africa, India or Middle East) may provide additional useful insights to draw a more precise picture of the genetic history of cattle. As exemplified in the present study, such extension is straightforward because of the easy to share nature of SNP data and the widespread use and cost effectiveness of the Bovine SNP50 genotyping assay.
Materials and Methods
Ethics statement
No ethics statement was required for the collection of DNA samples. DNA was extracted either from commercial AI bull semen straws or from blood samples obtained from different veterinary practitioners visiting farms with the permission of the owners.
Genotyping data, quality control, marker selection and estimation of LD
For the purpose of this study 296 individuals belonging to 14 different French cattle breeds were genotyped on the Illumina BovineSNP50 chip assay [14] at the Centre National de Génotypage (CNG) platform (Evry, France) using standard procedures ( http://www.illumina.com ). Based on available pedigree information, every attempt was made to ensure that samples were typical of the breeds and to limit relationships. In addition, genotyping data were also included for 33 other breeds and 4 american bisons ( Bison bison ) [14] – [16] . For these latter studies and for the sake of homogeneity in sample size across the whole data set, the maximal number of individuals retained per breed was restricted to 31 individuals (trying to limit their relationships when pedigree information was available). In total, 1,121 individuals (from 14 to 31 per breed) were available for the different analyses ( Table S1 ).
Description of the origin of samples and genotyping data is detailed in Table 1. Among the 51,582 genotyped SNPs mapping to a bovine autosome on the Btau_4.0 bovine genome assembly [36] , 3,009 SNPs which were not genotyped for at least 75% of the individuals in at least one breed and 2,982 SNPs which were monomorphic in all breeds were discarded from further analyses. Notice that all the individuals were genotyped for at least 95% of the selected SNPs. Following [16] , an exact test for Hardy-Weinberg Equilibrium (HWE) [37] was further carried out within each breed separately on the 45,291 remaining SNPs. Based on the obtained p-values, q-values [38] were estimated for each SNP using the R package qvalue ( http://cran.r-project.org/web/packages/qvalue/index.html ). A total of 585 SNPs exhibiting q-value<0.05 in at least one breed were then discarded from further analysis. Thus, 44,706 SNPs were finally considered for the study leading to an average marker density of 1 SNP every 56.9 kb over the genome ( Table S2 ). Moreover, as shown in Figure S2 and detailed in Table S1 , the genome coverage was homogeneous with a median distance between consecutive SNPs equal to 40.4 kb. Few large gaps between SNPs were present since the 95th (99th) percentile of this distance was 146 kb (252 kb), the largest gap localized on BTA10 being 2 Mb long. Conversely, less than 0.5% of the distances between successive SNPs were shorter than 20 kb. In order to characterize the extent of LD, we computed the r 2 measure [39] between each marker pair within each breed separetely using Haploview 4.1 [40] .
Principal Component Analysis
PCA was carried out based on all available SNP information using the R packages ade4 [41] . Note that, as expected from the extent of LD, multivariate analyses using the SMARTPCA software package [42] which allows to perform correction for the extent of LD (by replacing individual SNP values with the residuals from a multivariate regression without intercept on the two preceding SNPs on the map, provided they are less than 200 kb apart) lead to almost identical results.
Spatial Principal Component Analysis
sPCA was carried out on a between-breed level using the R packages ade4 [41] and adegenet [30] . Briefly, while in PCA, the optimization criterion only deals with genetic variance (with the eigenvalue decomposition of X'X , where X is the matrix of allelic frequencies), sPCA aims at finding independent synthetic variables that maximize the product of the genetic variance and spatial autocorrelation measured by Moran's I [29] . This is accomplished by the eigenvalue decomposition of a matrix X'(L+L')X where L synthesizes spatial structure among populations via a neighboring graph (in our study a Gabriel neighboring graph was chosen) connecting the populations on the geographical map [17] , [43] to model spatial structure among breeds. Resulting eigenvalues can be either positive or negative reflecting respectively global or local spatial pattern. Finally, the overall spatial autocorrelation associated to each resulting sPCA principal component was quantified using the Moran's I [29] . For a thorough description of sPCA, interested reader should refer to [17] .
Neighbor-Joining trees construction
ASD were computed for each pair of individuals using all available SNP information by a simple counting algorithm: for a given pair of individuals i and j, ASD was defined as 1- x ij where x ij represents the proportion of alleles alike in state averaged over all genotyped SNPs. A neighbor-joining tree [44] was computed based on the resulting distance matrix using the R package APE [45] . Similarly a neighbor-joining tree based on the Reynolds genetic distances [46] between the different pairs of breeds was constructed using PHYLIP 3.65 package [47] . The reliability of each node was estimated from 100 random bootstrap resamplings of the data. The resulting dendrogram in Figure 4 was plotted using the program Dendroscope [48] .
Unsupervised Hierarchical Clustering of the individuals
Unsupervised hierarchical clustering of individuals based on SNP genotyping data was performed using the maximum likelihood method described in [21] which is implemented via an Expectation-Maximization algorithm in the program frappe . The program was allowed to run for 10,000 iterations, with pre-specified numbers of clusters varying from K = 2 to K = 47 (number of distinct populations). Convergence of the algorithm was empirically assessed by considering estimated cluster membership and data likelihood. Graphical displays of the results were done with the program Distruct [49] .
F-statistics
The global F- statistics F IT , F ST and F IS were estimated respectively in the form of F , θ and f [50] using the program GENEPOP 4.0 [51] . GENEPOP 4.0 was also used to estimate diversity for each locus and population both within individuals and among individuals within a population. The within breed F IS was derived from the average of these two quantities over all the SNPs. In order to evaluate the reliability of the F IS estimates we computed the mean and standard deviation over 10,000 samples of 5,000 randomly chosen SNPs.
Supporting Information
Figure S1
Decay of average pairwise r2 with inter-marker distance for the different populations.
(0.05 MB PDF)
Figure S2
Distribution of inter-SNP physical distances based on the Btau_4.0 bovine genome assembly ( http://genome.ucsc.edu/ ).
(0.00 MB PDF)
Figure S3
Between breed PCA for the 47 different bovine populations. Populations are plotted according to their coordinates on the first two (A) and first and third (B) principal components on the eigenanalysis.
(0.12 MB PDF)
Figure S4
Neighbor-Joining tree relating the 47 cattle populations and American bisons (OBB) outgroup based on Reynolds genetic distances computed using allele frequencies at 27,527 SNPs polymorphic (MAF>0.01) in at least two zebus, two WAT and two EUR breeds. Reliability of the nodes (percentage over 100 Bootstrap samples) are indicated for each node.
(0.05 MB PDF)
Figure S5
Gabriel neighboring graph modeling the spatial structure of breeds projected on the geographic map.
(0.03 MB PDF)
Figure S6
sPCA results. Projection of the breed coordinates on the first (A), second (B) and third (C) sPCA principal components onto the geographical map. The area of the square is proportional to the absolute value of the score while the color of the square (black or white) corresponds to its sign (positive or negative). D) 3D representation of the breed coordinates on the first three sPCA principal components (breed names are colored according to the synthetic score obtained in Figure 4 representation).
(0.10 MB PDF)
Table S1
Sample Description. The land of origin, country of sampling (N = North, S = South, SE = South East, SW = South West, NE = North East, NW = North West and M = Middle) and type of the populations (EUT = European Taurines, SYN = Synthetic breeds, NAT = North African Taurines, WAT = West African Taurines, WAH = West African Hybrids, WAZ = West African Zebus, EAZ = East African Taurines and ZEB = Zebus from Indian origin) are indicated, together with the number of individuals sampled. For each breed, marker polymorphism is summarized through average heterozygosity computed across the 44,706 SNPs considered in this study and the proportion of SNPs with a MAF above 0.05.
(0.03 MB XLS)
Table S2
SNP bovine genome coverage based on the Btau_4.0 bovine genome assembly ( http://genome.ucsc.edu/ ).
(0.01 MB XLS)
Table S3
Within population F IS .
(0.01 MB XLS)
Table S4
F ST for each pair of populations.
(0.05 MB XLS)
Table S5
Comparison of PCA and sPCA. For each analysis, eigenvalues, percentage of genetic variance explained and Moran's I spatial autocorrelation associated to the corresponding principal components are given.
(0.01 MB XLS)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Nearly 10 years after the first HIV cases were diagnosed in 1980, an inverse correlation was observed between HIV prevalence and male circumcision (MC) prevalence [ 1 ]. By 2007, randomized controlled trials in South Africa, Kenya, and Uganda had demonstrated that voluntary medical male circumcision (VMMC) conferred a 60% reduction in the risk of female-to-male HIV transmission [ 2 – 4 ]. The findings compelled the World Health Organization (WHO) and the Joint United Nations Programme on HIV and AIDS (UNAIDS) to support VMMC as a safe and effective method for HIV prevention. WHO and UNAIDS recommended the intervention be adopted and prioritized within a comprehensive HIV prevention package in 14 countries with high HIV prevalence and low MC prevalence [ 5 ]. Other preventive effects of MC are reduced risk of infection from ulcerative sexually transmitted infections [ 6 – 10 ].
According to the Tanzania HIV/AIDS and Malaria Indicator Survey 2011–12, HIV prevalence among males ages 15–49 is more than 50% higher among those who are uncircumcised (5.2%) than among those who are circumcised (3.3%). The prevalence of HIV among males in this age group also varies by urban (5.2%) versus rural (3.4%) populations, as well as by region: for example, Njombe (14.2%) is highly affected by HIV and Dodoma (3.7%) is less affected [ 11 ]. The prevalence of MC (71.4% nationally in 2012) varies by region (ranging from 27.5% to 99%), religion (ranging from 25% to 96.8%), and other socioeconomic factors. Men with secondary education or higher are more likely to be circumcised (87.5%) than men with no education (52%), as are men living in an urban location (94.2%) than those in a rural location (64.2%) and men in the highest wealth quintile (92.4%) than those in the lowest wealth quintile (64.8%) [ 11 ].
In 2008, a situational analysis was undertaken in six regions of Tanzania to inform the roll-out of a national VMMC program [ 12 ]. The results demonstrated high acceptance of MC across both traditionally circumcising and noncircumcising communities. Men and women primarily cited tradition, the reduced risk of acquiring sexually transmitted infections, and hygiene as reasons for accepting the practice. Among traditionally circumcising communities, MC is performed during childhood or adolescence and serves as an important rite of passage into adulthood. In noncircumcising communities, the preventive medical benefits are considered most important, and men are more likely to accept health facility-based circumcision than traditional circumcision [ 13 , 14 ].
In 2009, a VMMC program was piloted in three regions: Iringa, Kagera, and Mbeya. Together, the findings from the situational analysis and lessons learned from the pilot program shaped the National Strategy for Scaling Up Male Circumcision 2010–2015. This strategy proposed scale-up of VMMC for males ages 10–34 in seven regions and one district in which there was low circumcision prevalence and high HIV prevalence: Iringa, Kagera, Mbeya, Mwanza, Rukwa, Shinyanga, Tabora, and Rorya district, in Mara region. Since 2010, several of these priority regions have been further subdivided to add the new regions of Geita, Katavi, Njombe, and Simiyu (11 total priority regions and Rorya district, in Mara) [ 15 ]. The lower age bound for VMMC scale-up in the strategy reflected the preferred age at circumcision (5 to 10 years) reported in the situational analysis, and the potential platform for strengthening adolescent and reproductive health services; the upper age bound was based on evidence from the situational analysis indicating that males over the age of 35 are relatively less receptive to VMMC. The country targets were later disaggregated into a primary target (males ages 10–24) and a secondary target (males ages 25–34) [ 12 ].
From July 2010 to October 2014, more than one million VMMCs were performed in Tanzania through the national program [ 16 ]. Through another modeling exercise, we estimated these VMMCs will avert 17,000 HIV infections in the general population by 2025, even if the VMMC program were not continued after 2014 [ 17 ]. To achieve this number of VMMCs, the program has actively worked to overcome barriers in service delivery and communication [ 18 – 20 ]. The majority of those barriers disproportionately affected men over age 25, such as the preference for discrete and private services, preference for male service providers, the requirement for abstinence in the postsurgical period, loss of income, and fear of pain associated with postsurgical erections. Across multiple settings, VMMC was also perceived as shameful for older men and/or married men with children [ 18 , 21 , 22 ].
As a result, more than 70% of VMMC clients reached through the national program since its inception in 2010 have been between the ages of 10 and 19 [ 19 ]. The relative uptake of VMMC among adolescents ages 10–19 is higher than the relative age distribution of uncircumcised males would indicate [ 23 ]. The converse is true for males ages 25–49. Similarly, progress toward national VMMC goals varied considerably across regions. This led policymakers to investigate the relative impact and cost-effectiveness of circumcising males within different age groups and in different geographic settings.
In 2013, the government expressed interest in understanding the impact of investing in VMMC through the national program, and the added benefit of doing so in a geographically-targeted way. It also looked to measure the impact of attracting fewer clients over age 25 than expected. This analysis examines the cost and epidemiological consequences of reaching specific age groups and subnational regions as part of Tanzania’s National VMMC Program.
Methods
DMPPT 2.0 Model
The Decision Makers’ Program Planning Tool (DMPPT) 2.0 model is described in detail in a separate manuscript in this collection [ 24 ]. Briefly, DMPPT 2.0 is a simple compartmental model, implemented in Microsoft Excel 2010, designed to analyze the effects of age at circumcision on program impact and cost. The DMPPT 2.0 model tracks the number of circumcised males in newborns and in each five-year age group over time, taking into account age progression and mortality. The model calculates discounted VMMC program costs and HIV infections averted in the population in each year in a user-specified VMMC scale-up strategy, compared with a baseline scenario in which the MC prevalence remains the same as it was. The baseline scenario assumes that traditional or other circumcisions that produced the baseline MC prevalence continue at the same rate as before the VMMC program was initiated.
Tanzania data sources
A separate DMPPT 2.0 model was created for each of the 11 priority regions for VMMC scale-up (Geita, Iringa, Kagera, Katavi, Mbeya, Mwanza, Njombe, Rukwa, Shinyanga, Simiyu, and Tabora). The Mara region was not included, since VMMC scale-up within Mara is limited to a single district, and the required data were not available to populate the model at the district level. At the request of the country team, an additional model was created combining all 11 priority regions together but not including the rest of the country. National level estimates are not informative because of the high background prevalence of MC outside of the priority regions, but the country team wanted a single model that would collate the results for all of the regions where the VMMC program is being conducted.
All model inputs can be found in the supplemental materials ( S1 and S2 Appendices). The DMPPT 2.0 model is populated with population, mortality, and HIV incidence and prevalence projections from an external source. To create the regional DMPPT 2.0 files, we first created regional Spectrum/AIM files based on the validated 2013 National Spectrum/AIM [ 25 ] file obtained from UNAIDS, downloaded on January 13, 2014. Spectrum/AIM is a model that projects population size, mortality, and HIV prevalence and incidence based on data empirically collected from the country. HIV prevalence data from sentinel surveillance sites used in Spectrum AIM to project HIV incidence and prevalence were categorized into the 11 regions plus the remainder of the country. HIV prevalence curves were fit to the surveillance data within each region using the R-trend fitting method in the EPP3 module within Spectrum/AIM [ 25 ]. Sites from several newly formed regions were grouped for the curve-fitting, because historical Demographic and Health Survey data only existed for the region prior to being split. Grouping the sites also meant that each curve-fitting was based on more data, making it more robust. The sites for Geita, Mwanza, Shinyanga, and Simiyu were grouped for the curve-fitting. The sites for Iringa and Njombe were also grouped, as were the sites for Katavi and Rukwa. Population by age and year, mortality by age and year, and annual number of male births were exported from the regional Spectrum/AIM files into their respective regional DMPPT 2.0 files.
For HIV incidence, we used a single national Spectrum/AIM file in which the surveillance sites had been categorized into regions, and calculated the incidence for each region or combination of regions as described above. HIV incidence by age from a national Spectrum/Goals model [ 11 , 25 ] for Tanzania ( S2 Fig ) was multiplied by the ratio of the HIV incidence in each region (exported from the Spectrum/AIM file) to the national HIV incidence for each year between 2013 and 2020. For the years after 2021–2050, the regional to national incidence ratio for 2020 was used to scale the national incidence by age.
Numbers of VMMCs conducted in the country in each region in each year, disaggregated by age group, were extracted from the national health management information system on March 19, 2015. Male circumcision prevalence by age group in the model base year (2014) was based on the Tanzania HIV/AIDS and Malaria Indicator Survey 2011–12 [ 11 ]. The unit cost of VMMC used in the analysis was $83 USD (all subsequent references to currency are in 2014 U.S. dollars), based on [ 26 ] and validated by stakeholders in Tanzania. This is consistent with the 2014 costing study by Menon et al., which determined VMMC would cost $36–$128 depending on the region and service delivery model [ 27 ]. For ART we used a cost per patient year of treatment of $515 based on an international weighted average median price in 2011 of $145 for first- and second-line antiretroviral drugs [ 28 ]; $222 for average service delivery and monitoring costs [ 29 ]; plus an additional 40% for costs above the facility level for administration, logistics, training, planning, and so forth.
Analytical approaches
To examine the effect of client age on the impact of scaling up VMMC, we created a series of scenarios in which each scenario had a target of 80% MC prevalence for a single age group or combination of age groups, leaving the target for the other age groups at the same level as the baseline. We created one scenario for each individual five-year age group, in addition to several scenarios with 80% targets for combined age groups, such as 10- to 34-year-olds or 15- to 29-year-olds. In each scenario, MC coverage was scaled up between 2014 and 2018 by applying a linear interpolation to the baseline MC prevalence for each age group in 2013 and the target coverage in 2018. After 2018, the coverage for each age group was maintained at 80%. For each scenario, we compiled the decrease in HIV incidence in the scale-up scenario compared with the HIV incidence in the baseline scenario in each year of the model, and the total number of VMMCs required during the scale-up phase (2014–2018). The following model outputs for each scenario were measured over the 15-year period between 2014 and 2028, inclusive: the total number of HIV infections averted in the population (including secondary infections averted among females; see above), the number of VMMCs per HIV infection averted, and the total cost of the VMMC program. Costs, numbers of VMMCs (when calculating VMMC per HIV infection averted), and infections averted were all discounted at a rate of 3% per year. Uncertainty around the age distribution of HIV incidence was calculated as described in a separate manuscript in a separate manuscript in this collection [ 24 ], except that the incidence for each age group was varied by +/-15%, based on the estimated variance in incidence by age from the ALPHA Network [ 30 ] incidence rate ratios by age used in Spectrum/AIM.
To examine the differences in impact and cost-effectiveness across the different regions of Tanzania, we compiled the cost per HIV infection averted over the 15-year period 2014–2028, inclusive, from each regional DMMPT 2.0 file. The VMMC scale-up scenario used for this analysis was to scale up to 80% coverage among males ages 10–34. Uncertainty around the regional estimates was calculated as follows: The Spectrum/AIM Uncertainty Analysis tool was run on each individual regional Spectrum/AIM file described above. The median, lower 2.5%, and upper 97.5% bounds around the adult (ages 15–49) HIV incidence for the year 2020 were extracted from the Uncertainty tool. DMPPT 2.0 files for each region were created representing the lower and upper bounds of the HIV incidence by multiplying the incidence in the original file by the ratio of the lower to median or upper to median values. Cost per HIV infection averted extracted from the lower bound file for each region was used as the lower bound in the analysis, and likewise for the higher bound.
Results
We considered four metrics related to circumcision of different client age groups: immediacy of impact, magnitude of impact, cost-effectiveness, and number of VMMCs per HIV infection averted. In this analysis, immediacy of impact is measured by the relative reduction in HIV incidence over a 5-year period, magnitude of impact is measured by the relative reduction in HIV incidence over a 15-year period, cost-effectiveness is measured by the discounted cost per HIV infection averted over a 15-year period, and the number of VMMCs per HIV infection averted is measured in a 15-year period. All metrics are measured starting in the year 2014.
To look at immediacy of impact, we looked at the relative reduction in HIV incidence achieved after five years, by circumcising individual five-year age groups of clients compared with a baseline scenario that does not provide VMMCs ( Fig 1 ). The figure presents the reductions in HIV incidence in the general population over 2014–2050 that would be achieved by scenarios in which only clients in an indicated five-year age group are circumcised. The immediacy of impact (five years) is largely a function of the projected HIV incidence in the male subpopulation targeted. The highest HIV incidence among Tanzanian males is projected to occur among those in the 20- to 34-year-olds [ 24 ]; the most immediate reduction in HIV incidence is projected to result from VMMC scale-up among males in the 20- to 24-, 25- to 29-, and 30- to 34-year-olds.
10.1371/journal.pone.0153363.g001
Fig 1
Modeled relative reduction in HIV incidence by scaling up VMMC for individual age groups, compared with no scale-up of VMMC over baseline levels, 2014–2050.
(a) immediacy of impact (5 years). (b) magnitude of impact (15 years). The HIV incidence ratio represents the incidence in the scale-up scenario divided by the HIV incidence in a population where circumcision is not scaled-up over baseline levels. Each line represents the HIV incidence ratio under a scenario in which only the indicated five-year age group is circumcised. Marker a represents a five-year period from the base year (2014). Marker b represents a 15-year period from the base year.
Although in the short term (five years), circumcising males in the 20- to 24-, 25- to 29-, and 30- to 34-year-olds is projected to produce the greatest reduction in HIV incidence, over a 15-year period, circumcising younger age groups—those 10- to 14-, 15- to19-, and 20- to 24- years old—is projected to have the greatest impact on HIV incidence ( Fig 1 ). This is because males who are circumcised before sexual debut are protected for the longest proportion of their lifetime HIV exposure, and because the younger age groups are larger in population size.
The previous analyses examined scenarios in which only a single five-year age group at a time is circumcised. To assess the impact and cost-effectiveness of circumcising age groups that might be closer to actual implementation strategies, we examined several scenarios with 80% targets for combined age groups, such as 10- to 34-year-olds or 15- to 29-year-olds. Fig 2 considers the impact of targeting combined age groups for VMMC scale-up. The DMPPT 2.0 model tracks circumcised males as they transition into older age groups over time. Hence, the results in Fig 2 are not the aggregate of results derived from circumcising individual five-year age groups. This analysis shows that the greatest number of HIV infections averted is achieved with the broadest age group: males 10–49 years old. This is the age group with the largest number of clients, so it results in the greatest number of HIV infections averted. Circumcising males ages 15–49 results in slightly fewer HIV infections averted. If programmatic challenges make reaching males ages 35–49 impossible, 88% of HIV infections could still be averted by circumcising males ages 10–34.
10.1371/journal.pone.0153363.g002
Fig 2
HIV infections averted in scenarios scaling up VMMC among different client age groups.
The time period for measuring HIV infections averted was 15 years, 2014–2028, inclusive. Error bars represent uncertainty bounds.
Fig 3A and 3B compare the cost-effectiveness of circumcising five-year and broader age groups of clients. Circumcising males ages 20–24, 20–29, or 30–34 would achieve the lowest cost per infection averted, compared to circumcising males ages 10–49 ( Fig 3A ). In the combined age groups depicted in Fig 3B , the lowest cost per HIV infection averted would be achieved when circumcising males ages 15–34, but uncertainty bounds indicate that circumcising males ages 15–29 or 15–49 could be equally or more cost-effective. Although strategies that include males ages 10–14 lead to higher costs per HIV infection averted (less cost-effectiveness), historical program data indicate that males ages 10–14 constitute a large proportion of VMMC clients irrespective of targeted demand creation. By circumcising males ages 10–29 or 10–34, the program would achieve the same level of cost-effectiveness as circumcising males ages 10–49, given the uncertainty bounds. This analysis does not consider variations in cost across age groups due to differences in demand creation or service delivery.
10.1371/journal.pone.0153363.g003
Fig 3
a and b. Cost per HIV infection averted in scenarios scaling up VMMC among different client age groups. The time period for measuring HIV infections averted was 15 years, 2014–2028, inclusive. Error bars represent uncertainty bounds.
Table 1 summarizes the priority age group for each metric examined. VMMC scale-up to 80% MC coverage for males ages 20–24, 25–29, and 30–34 by 2018 would result in the fewest VMMC per HIV infection averted [ 24 ] and the most immediate impact on HIV incidence. This scenario would require 0.72 million VMMCs over 2014–2018. Over a 15-year period, the age groups that produce the greatest decrease in HIV incidence are 10- to 14-, 15- to 19-, and 20- to 24-year-olds, so these age groups are prioritized for magnitude of impact. The most cost-effective program, providing the lowest cost per HIV infection averted, would circumcise 80% of males ages 15–34 and would require 1.07 million VMMCs by 2018.
10.1371/journal.pone.0153363.t001
Table 1 Priority age groups and number of VMMCs required for each parameter in the model framework, Tanzania.
Parameter
Priority age group
No. VMMCs required to reach 80% MC coverage by 2018
VMMC/HIV Infection Averted
20–34
0.72 million
Immediacy of Impact
20–34
0.72 million
Magnitude of Impact
10–24
1.69 million
Cost-effectiveness
15–34
1.07 million
Country Age Targeting Strategy
10–34
2.16 million
In comparison, circumcising 80% of males ages 10–49 would require 2.57 million VMMCs over 2014–2018 and the current age strategy in Tanzania would require 2.16 million VMMCs over 2014–2018.
The Tanzania National VMMC Program is already prioritized geographically, by focusing on the 11 priority regions plus Rorya district, in Mara. We assessed whether further prioritization of the 11 priority regions would be advisable, by examining how the cost per HIV infection averted varied among the regions. Fig 4 compares the cost-effectiveness of scaling up VMMC within each of 11 priority regions. Cost-effectiveness varies across regions, due to regional differences in the projected HIV incidence. The model predicts that the most cost-effective VMMC programs will be in Njombe and Iringa, where the projected HIV incidence ranges from 0.6% to 1.1%. In the nine remaining regions, the projected HIV incidence ranges from 0.1% to 0.3% and cost-effectiveness of the VMMC programs therein is not distinguishable from the national estimate, considering uncertainty bounds. Despite the range in cost-effectiveness, the discounted HIV treatment costs averted by scaling up VMMC outweigh the projected discounted VMMC program costs in every region ( S2 Fig ). In other words, the VMMC program is cost saving in all 11 priority regions compared with HIV treatment costs averted.
10.1371/journal.pone.0153363.g004
Fig 4
Discounted cost per HIV infection averted by region, 2014–2028, given a scenario of scale-up to 80% of 10- to 34-year-olds.
In Tanzania, the national program has been successful in scaling up VMMC through 2014, but regions differ in their progress toward the 80% VMMC coverage target in 2018. In Fig 5A–5C , the bars 2013 and prior represent actual VMMCs conducted by the program. From 2014, the model projects the number of VMMCs needed, by age group, to scale up to 80% coverage among males ages 10–34 by 2018 and maintain 80% coverage thereafter. The trends suggest Mbeya can easily reach the 2018 targets if it continues to scale up VMMC. Geita would likely have difficulty reaching 80% coverage among males ages 10–34 by 2018 unless it significantly expands services. Iringa has nearly reached the target and is already transitioning into the maintenance phase.
10.1371/journal.pone.0153363.g005
Fig 5
a–c. Annual VMMCs required to scale up to 80% MC coverage among males ages 10–34.
Limitations
DMPPT 2.0 relies on available national and regional estimates of demographic and epidemiological data, and therefore modeled projections are subject to the assumptions and measurement errors of these inputs: baseline MC prevalence, unit costs, and projections of future HIV incidence. Future HIV incidence is especially uncertain, because it is dependent on many factors that are impossible to predict.
Cost assumptions were based on a fixed unit cost of $83 and therefore did not reflect differences in cost depending on client age, geographic location, implementation model, phase of scale-up, or other factors. In Tanzania, VMMC unit cost may be higher for men ages 20–34 and even higher for men ages 35 and older. Programmatic experience to date confirms these men would require more demand creation activities and health facility accommodations to increase privacy and adapt to lower volume settings. Higher unit costs would negatively affect the cost-effectiveness of the respective age groups. Total cost estimates for scale-up scenarios are therefore useful only for relative comparisons, and should not be used for budget projections.
Other factors transcending age and geography affect the implementation of the national VMMC program. In Tanzania, stigma associated with VMMC for married men with children, regardless of age, has been documented [ 21 , 22 ]. The model does not account for broader social, cultural, and logistical barriers to program acceptability and implementation. When selecting a scale-up strategy or target, Tanzania will need to consider the model results in the context of these and other influences, such as human resources, the political environment, and challenges to demand creation specific to certain geographic areas or subpopulations.
Discussion
The results of this analysis demonstrate that Tanzania can maximize HIV incidence reduction and cost-effectiveness over a 15-year period, by circumcising males ages 10–24 and 15–34, respectively. The DMPPT 2.0 modeling results also confirm that the VMMC program will be cost saving in the 11 priority regions, with programs in Njombe and Iringa regions being especially cost-effective. Given evidence that circumcision status may be associated with lower risk for certain sexually transmitted diseases in men and women, the HIV cost saving estimates reported here would underestimate the secondary benefits from reduced levels of sexually transmitted diseases [ 6 – 10 , 31 ].
Tanzania was unique in its response to the global call by WHO and UNAIDS in 2007 to scale up VMMC as an effective intervention for HIV prevention. At the time, global stakeholders had not recommended age-specific targeting for national programs. Rather, in 2011 UNAIDS released the Joint Strategic Action Framework to Accelerate the Scale-Up of VMMC for HIV Prevention in Eastern and Southern Africa , which aimed for priority VMMC countries to achieve 80% circumcision prevalence among males ages 15–49 by 2016 [ 32 ]. Contrary to such guidance, the Tanzania National VMMC Program launched in 2010 focused on males ages 10–34 in the VMMC priority regions. In the absence of modeling for subpopulation targeting scenarios, the Ministry of Health and Social Welfare relied on the thorough situational analysis and lessons learned from a pilot program to determine the target age group. The ministry based its decision on cultural preferences, the potential of VMMC programs to serve as a platform for strengthening adolescent and reproductive health services, and the high cost and minimal success of attracting older clients.
The DMPPT 2.0 results reinforced the target set forth in the National Strategy for Scaling-Up Male Circumcision 2010–2015. As a result of this analysis, Tanzanian policy makers and program implementers chose to continue to focus scale-up of VMMC on males ages 10–34 years, thereby maximizing impact and cost-effectiveness while acknowledging the programmatic realities related to demand among the younger and older age groups. Although the age target went unchanged, the new evidence base would profoundly impact VMMC program implementation going forward. Tanzania’s early deviation from international recommendations generated some hesitancy among implementing partners to focus efforts and resources on younger age groups. The findings from this analysis empowered the national program and implementing partners to adopt a focused approach to service delivery tailored to males in the 10- to 34-year-olds, and especially adolescents.
The Tanzania National VMMC Program is now looking toward the implementation of its second circumcision strategy, the VMMC Country Operational Plan 2014–2017 . The Country Operational Plan maintains its predecessor’s goal to achieve 80% circumcision prevalence among males ages 10–34. The refocus on adolescent VMMC services is reflected in the plan, which explicitly calls for heavy investments in “properly serving adolescents and their gatekeepers” [ 12 ]. The findings of this analysis also emphasize the importance of circumcision in younger populations to the long-term impact of VMMC on HIV incidence. As a result, the Ministry of Health and Social Welfare also defines three implementation phases of the Country Operational Plan: scale-up, catch-up, and sustainability. The scale-up and catch-up phases refer to the expansion of fixed sites and outreach schedules, while the sustainability phase entails a gradual transition toward VMMC for boys turning age 10 and early infant MC. By 2017, the country aims for all 12 regions to enter the sustainability phase [ 12 ].
The use of these results in Tanzania to inform ongoing implementation and the VMMC Country Operational Plan 2014–2017 exemplifies the Tanzania VMMC program’s commitment to applying evidence to improve policy, implementation, and the well-being of the Tanzanian population.
Supporting Information
S1 Appendix
Tanzania Model Inputs.
(DOCX)
S2 Appendix
Spectrum Inputs.
(XLSX)
S1 Fig
Age-specific male HIV incidence, 2013, from Spectrum/Goals model in Tanzania’s 11 priority regions.
(TIF)
S2 Fig
Comparison of VMMC program costs and HIV treatment costs averted, by region 2014–2028.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Background
The neonatal period is the period from birth to 28 days of life. It is within this period that infants are highly vulnerable to death [ 1 ]. Of newborn deaths in the first month of life, about a third of all neonatal deaths tend to occur on the day of birth and close to three quarters die in the first week of life [ 2 , 3 ]. Preterm birth complications, intrapartum-related complications, sepsis, congenital abnormalities, pneumonia, tetanus, and diarrhea were identified as the major causes of death in neonates [ 4 ].
Neonatal near miss (NNM) is a concept related to neonatal mortality where neonates survive either by chance or by the quality of care provided [ 5 ]. There is no standard definition or internationally agreed identification criteria for NNM cases, due to this it has been used inconsistently. Some kinds of literature defined it as a newborn who presented with a severe complication/s that occurred during pregnancy, birth, or within 28 days of extra-uterine life but survived [ 6 , 7 ]. While others are defined as a newborn who nearly died but survived having overcome serious complications during pregnancy, delivery, or within the first seven days of life [ 8 , 9 ]. However, the Latin American Centre for Perinatology (CLAP) from Pan American Health Organization prepared a standardized definition after reviewing different studies done on NNM as any newborn infant who exhibited pragmatic and/or management criteria and survived the first 27 days of life [ 10 ].
Globally 2.5 million neonatal deaths occur with an estimated neonatal mortality rate of 18 deaths per 1,000 live births [ 4 ]. Less than 1% of these deaths occur in developed countries [ 11 ]. While 98% of all neonatal deaths occur in developing countries, mostly at home, outside the formal health care system. Largely the deaths were related to infections, birth asphyxia and injuries, and consequences of prematurity, low birth weight, and congenital anomalies [ 12 ]. In Sub-Saharan Africa (SSA) neonatal mortality rate was 28 deaths per 1,000 live births and a child born in this region has 10 times more likely to die in the first month than a child born in a high-income country [ 4 ]. Of neonatal deaths at SSA 50% occurred in just five countries Ethiopia, Nigeria, DR Congo, Tanzania, and Uganda [ 11 ]. In addition to this, Ethiopia Mini Demographic Health Survey reported that the neonatal mortality rate remained increasing from 29/1000 LB to 30/1000 LB within the last four years (2015–2019) [ 13 ].
Irrespective of decreased neonatal mortality rate both in the developed and developing countries, the neonatal morbidity rate remains elevated [ 14 ]. A cross-sectional study done in Brazil showed the estimated number of survivors from NNM cases was seven times higher than the number of neonatal deaths, meaning, that for every neonatal death seven neonates were nearly died but survived [ 15 ]. Similarly, in Uganda, the NNM rate was two times higher than that of neonatal mortality rate [ 16 ]. Neonates who undergo severe complications might also develop long-term morbidity through effects on neurological and cognitive development and also has associations with chronic diseases such as diabetes, cardiovascular disease, and chronic lung disease as well as other major disabilities such as blindness or low vision and hearing loss in their late-life [ 12 , 17 ].
The magnitude of NNM was widely varied across studies because of the difference in criteria used. As studies that used only pragmatic criteria, the incidence of NNM varied between 21.4/1000 live births in Brazil [ 18 ] and 86.7/1000 live births in India [ 19 ]. Whereas, according to those studies that used both pragmatic and management criteria, the incidence of NNM ranged between 39.2/1000 live births [ 20 ] to 367/1000 live births [ 16 ].
The UN Agenda for Sustainable Development Goal (SDG) from 2016 to 2030 was to end preventable deaths of newborns and indicated that the neonatal mortality should be less than 12/1000 LB at the end of 2030 [ 21 ]. So identifying NNM cases and correcting contributing factors were of the utmost importance to get relevant controls for neonatal deaths, since many babies who die pass through a phase of organ dysfunction before dying and also to prevent long term consequences of severe neonatal morbidity [ 22 ].
Investigating NNM cases would aid in taking measures for further amendment of service delivery and programs. This research was therefore intended to determining the prevalence of NNM and to recognize associated factors since the prevalence of NNM is not well understood in Ethiopia. It also helps in recognizing the contributory factors of neonatal mortality and morbidity so that appropriate actions can be adopted at the community and health systems level.
Methods and materials
Study design and setting
A facility-based cross-sectional study design was conducted from April 1-30/2020 in Jimma zone, southwest Ethiopia. The zone has a surface area of 119,316 square kilometers. It has 18 woredas and 1 town administration with a total of 555 kebeles ( Kebele is the smallest administrative unit in Ethiopia ) of which 515 of them were rural and 40 were urban. The population projection of 2014/15 of the zone was 2,486,155. The zone has 3 general hospitals, 4 district hospitals, and 1 referral and teaching hospital.
The population of the study
All live birth neonates delivered at Jimma zone public hospitals were the source population for this study. The study population is comprised of alive neonates in selected hospitals that meet the eligibility criteria. Those mothers who gave birth at home and were critically ill during the data collection time were excluded from the study. Besides, mothers who had twins were also excluded.
Sample size determination
The study’s sample size was determined using a single population proportion formula. Initially, a sample size of 423 was calculated using the following parameters: a 50% prevalence of NNM (because no similar research had been conducted in Ethiopia), a 95% confidence level, 5% margins of error, and a 10% non-response rate. Since the total population is less than 10,000 (N = 534), we used the finite population correction formula. The total population of N = 534 was obtained by averaging the client flow trends for each of the selected hospitals in April and May over the previous three years.
n f = n i 1 + n i / N = 423 1 + 423 / 534 = 236
The overall sample size for the study was 260 after accounting for a 10% non-response rate.
Sampling technique
Of eight governmental hospitals found in the Jimma zone, four hospitals were selected randomly by lottery method. The sample size for each hospital was proportionally allocated by averaging the trends of the previous three years’ client flow for each of the selected hospitals in April and May. Neonates were included consecutively at discharge from the postnatal ward and NICU until the determined sample size was reached ( S1 Fig ).
Data collection techniques, tools, and personnel
The data from mothers was collected by using a pre-tested, interviewer-administered structured questionnaire which was adapted from relevant literature [ 9 , 10 , 22 – 26 ]. The tool has generally three parts involving maternal socio-demographic characteristics, reproductive and obstetric history, and medical history during pregnancy. As data collectors, four midwife nurses who have obstetric care experience (one per hospital) and who can speak the local language were recruited. Data was collected through face-to-face interviews with neonates’ mothers after the neonates were assured to be survived or at discharge and Maternal charts were reviewed for clarity of diagnosis. As supervisors, two public health professionals who have a bachelor’s degree have been recruited.
Near misses’ events were identified by data collectors from neonates’ medical records according to the criteria of CLAP [ 10 ].
Data quality management
Before the start of data collection, training was offered to data collectors for one day on the purpose of data collection, data collection techniques, the content of the questionnaires, and how to approach the respondents. The data collection tool for the maternal side was prepared in English and translated to the local language Afaan Oromo. Then re-translated back to English to verify the consistency. The pretest was done at Bedele general hospital by taking 13 (5%) of the total sample size before the actual data collection to assess instrument simplicity, flow, and consistency. A day-to-day follow-up during the data collection period was carried out by the principal investigator and supervisors. Every day the collected data was reviewed and cross-checked for completeness and relevance before data entry.
Data analysis
The data was coded and entered into Epi-Data version 4.2 and exported to statistical package for social science (SPSS) version 23 for analysis. Inconsistencies and missing values were checked by running frequencies. Descriptive statistics like frequency distributions, mean, and standard deviation were computed. The bivariable analysis was done primarily to check the association of each explanatory variable with the outcome variable (NNM). Explanatory variables with marginal associations (p-value <0.25) in the bivariable analysis were eligible for multivariable logistic regression analysis to identify significant predictors of NNM. Finally, adjusted odds ratios (AOR) with 95% CI were estimated to assess the strength and the direction of associations, and statistical significance was declared at a p-value < 0.05.
Variables of the study
Neonatal near miss
NNM was considered when the newborn faced at least one of the following proposed criteria (either pragmatic or management criteria) but survived. From pragmatic criteria: Birth weight < 1750g, gestational age < 33 weeks, 5th-minute Apgar score < 7 or from management criteria: parenteral therapeutic antibiotics up to 7 days and before 28 days of life; nasal continuous positive airway pressure; any intubation during the first 28 days of life; phototherapy within the first 24 hours of life; cardiopulmonary resuscitation; the use of vasoactive drugs, anticonvulsants, surfactants, blood products and steroids for refractory hypoglycemia, parenteral nutrition, any surgical procedure, Congenital malformation if considered as a near miss in other criteria’s [ 10 ].
Preterm birth
Birth at a gestational age of 28 weeks to less than 37 weeks.
Low birth weight
Defined as a birth weight of a live-born infant less than 2500g irrespective of gestational age.
Stillbirth
Defined as the birth of an infant that has died in the womb or during intrapartum after 28 weeks of gestation.
Apgar score
Score ranging from 0–10 based on a newborn’s tone, color, respiration, pulse rate, and responsiveness at 1, 5, and 10 minutes.
Birth interval
The duration between the current birth and the preceding birth.
Pre-eclampsia
Persistent systolic blood pressure of 160 mmHg or more or a diastolic blood pressure of 110 mmHg; and either proteinuria of 5 g or more in 24 hours; or oliguria of <400 ml in 24 hours; or HELLP syndrome or pulmonary edema without seizure of eclampsia and/or diagnosed as severe pre-eclampsia case by a physician.
Eclampsia
Generalized fits in a patient without previous history of epilepsy includes coma in pre-eclampsia and other causes of seizure were ruled out by a physician.
Ethical clearance and consent to participate
Ethical clearance was obtained from the Institutional Review Board of Jimma University, Institute of Health. Permission letter was taken from the department of nursing and midwifery and given to selected hospitals. For those aged 18 and over, written informed consent was obtained from study participants. Besides, after explaining the study goals and procedures, written informed consent was taken from a parent or guardian using normal disclosure processes for those participants less than 18 years of age. A specific ID number was allocated to preserve the anonymity of the questionnaire. The privacy and confidentiality of participants were guaranteed before data collection.
Results
Socio-demographic characteristics of the respondents
Of a total of 260 sampled respondents, 255 took part in the study and yielded a response rate of 98.1%. The majority of respondents (86.7%) belong to the 20–34 age group, with the mean (±SD) age of (25.5 ± 4.7) years. Almost all (99.6%) of the mothers were married. About four out of ten (40.4%) of mothers were Muslim by religion and more than half (52.5%) were Oromo in ethnicity. Nearly two-thirds (62.4%) were rural residents and one-third (33.3%) of them attended primary education ( Table 1 ).
10.1371/journal.pone.0251609.t001
Table 1 Socio-demographic characteristic of respondents in public hospitals of Jimma zone, southwest Ethiopia, 2020.
Variable Categories
Frequency
Percent
Age in years (n = 255)
15–19
21
8.2
20–34
221
86.7
35–49
13
5.1
Marital status (n = 255)
Married
254
99.6
Single
1
0.4
E thnicity (n = 255)
Oromo
134
52.5
Amhara
59
23.1
Dawuro
26
10.2
Gurage
25
9.8
Others *
11
4.3
Religion (n = 255)
Muslim
103
40.4
Orthodox
90
35.3
Protestant
61
23.9
Catholic
1
0.4
Maternal educational level (n = 255)
No formal education
56
22
Can read and write only
36
14.1
Primary (1–8)
85
33.3
Secondary (9–12)
47
18.4
College and above
31
12.2
Paternal educational level (n = 254)
No formal education
26
10.2
Read and write only
30
11.8
Primary (1–8)
72
28.3
Secondary (9–12)
74
29.1
College and above
52
20.5
Mother’s occupation (n = 255)
Housewife
174
68.2
Merchant
41
15.3
Government employee
28
11
Others **
12
4.6
Paternal occupation (n = 254)
Farmer
70
27.6
Merchant
94
37
Government employee
47
18.5
Daily laborer
6
2.4
Private employee
25
9.8
Other ***
12
4.7
Residence (n = 255)
Urban
96
37.6
Rural
159
62.4
Average monthly income (n = 255)
< = 2000
42
16.5
2001–3500
89
34.9
3501–5000
63
24.7
= >5001
61
23.9
* Tigrai, Yem, Kaffa
** students, daily laborer
***Unemployed, NGO
Obstetrics characteristics of the respondents
The majority of respondents (65.5%) were multiparous, and nearly two-thirds (63.5%) of mothers had at least one ANC visit during their most recent pregnancy. More than one-third (37.6%) of mothers gave birth with an interval of 24–48 months between the preceding and current birth and 194 (76.1%) of mothers gave birth through spontaneous vaginal delivery. Twenty- five (9.8%) and 10 (3.9%) of the respondents had experienced abortion and stillbirth, respectively ( Table 2 ).
10.1371/journal.pone.0251609.t002
Table 2 Obstetric characteristics of respondents in selected public hospitals of Jimma zone, southwest Ethiopia, 2020.
Variable Categories (n = 255)
Frequency
Percent
Gravidity
Primigravida
88
34.5
Multigravida > = 2
167
65.5
Parity
Primipara
90
34.3
Multipara 2–4
150
58.8
Grand multipara > = 5
15
5.9
History of stillbirth
Yes
10
3.9
No
245
96.1
History of abortion
Yes
25
9.8
No
230
90.2
History of preterm birth
Yes
6
2.4
No
249
97.6
History of neonatal death
Yes
10
3.9
No
245
96.1
Birth interval
<24
48
18.8
24–48
96
37.6
>48
21
8.2
Frequency of ANC follow up
No ANC visit
9
3.5
1–3 times
162
63.5
4 and above
84
33.0
Mode of delivery
SVD
194
76.1
Instrumental delivery
24
9.4
Cesarean delivery
37
14.5
Prolonged labor
Yes
36
14.1
No
219
85.9
Obstructed labor
Yes
28
11
No
227
89
Hypertension during pregnancy
Yes
22
8.6
No
233
91.4
Urinary tract infection
Yes
235
7.8
No
20
92.2
Premature rupture of membrane (PROM)
Yes
9
3.5
No
246
96.5
Antepartum hemorrhage (APH)
Yes
6
2.4
No
249
97.6
Newborn related characteristics
Of 255 selected neonates 152 (59.6%) of them were females and 216 (84.7%) of the neonates’ presentation during delivery was vertex.
The magnitude of neonatal near miss (NNM) conditions
The magnitude of neonatal near miss (NNM) in the study area was 26.7% (95%CI: 21.6%-32.5%). Of the management criteria, cardiopulmonary resuscitation (CPR) was the commonest service obtained by 22 (8.6%) of neonates with near miss conditions closely followed by the use of anticonvulsant 21 (8.2%). From pragmatic criteria, an APGAR score of less than 7 was the most common near-miss condition sustained by 13 (5.1%) of neonates. Unidentified criteria were the use of surfactants and vasoactive drugs ( Table 3 ).
10.1371/journal.pone.0251609.t003
Table 3 Distribution of neonatal near miss conditions among neonates delivered in selected public hospitals of Jimma zone, southwest Ethiopia, 2020 (n = 255).
Neonatal near miss (NNM) criteria
Frequency
Percent
Pragmatic criteria
APGAR score less than 7
13
5.1
Birth weight less than 1750g
10
3.9
Gestational age less than 33 weeks
8
3.1
Management criteria
Cardiopulmonary resuscitation
22
8.6
Use of anticonvulsant
21
8.2
Use of phototherapy in the first 24 hours
15
5.9
Use of intravenous antibiotic up to 7 days and before 28 days of life
13
5.1
Use of corticosteroid for the treatment of refractory hypoglycemia
11
4.3
Nasal continuous positive airway pressure (NCPAP)
10
3.9
Any surgical procedure
8
3.1
Congenital malformation
8
3.1
Transfusion of blood derivatives
6
2.4
Any intubation
6
2.4
Factors associated with NNM
In bivariable logistic regression analysis, ten variables namely; maternal age, mother’s level of education, father’s level of education, mode of delivery, obstructed labor, prolonged labor, hypertension, having urinary tract infection during pregnancy, fetal presentation at delivery, and sex of the newborn had shown association at p-value <0.25 and were a candidate for the multivariable logistic regression model. In multivariable logistic regression analysis hypertension during pregnancy, mode of delivery, prolonged labor, and non-vertex fetal presentation during delivery were identified as significant predictors of NNM.
For those mothers with hypertension during pregnancy, the odds of having NNM was 3.4 times higher than their counterparts [AOR: 3.4; 95%CI: 1.32–8.88]. Being an NNM has been significantly associated with obstetric complications like obstructed labor during the last delivery. In contrast to those women with normal labor, the likelihood of NNM was just about 3 times higher among women with obstructed labor [AOR: 2.95; 95%CI: 1.32–6.45]. As a factor influencing the occurrence of the NNM condition, a fetal presentation was also identified. Compared to those with vertex presentation, neonates that had a non-vertex presentation were 4.6 times more likely to sustain a near-miss event [AOR: 4.61; 95%CI: 2.16–9.84]. On the other hand, in contrast to those mothers who gave birth through spontaneous vaginal delivery, the probability of having an NNM was 3.3 times higher among those mothers who gave birth by cesarean section [AOR: 3.3; 95% CI: 1.48–7.45] ( Table 4 ).
10.1371/journal.pone.0251609.t004
Table 4 Factors associated with NNM in selected public hospitals of Jimma zone, southwest Ethiopia, 2020 (n = 255).
Variable Categories
NNM
COR 95% CI
AOR 95% CI
No (%)
Yes (%)
Age
15–19
13 (61.9)
8 (38.1)
1
1
20–34
163 (73.8)
58 (26.2)
0.578 (0.228–1.466)
0.779 (0.271–2.238)
35–49
11 (84.6)
2 (15.4)
0.295 (0.052–1.692)
0.207 (0.029–1.494)
Mother’s educational level
No formal education
36 (64.3)
20 (35.7)
1
1
Can read and write only
28 (77.8)
8 (22.2)
0.514 (0.197–1.339)
0.565 (0.177–1.802)
Primary (1–8)
64 (75.3)
21 (24.7)
0.591 (0.283–1.233)
0.656 (0.257–1.676)
Secondary (9–12)
34 (72.3)
13 (27.7)
0.688 (0.297–1.596)
0.845 (0.277–2.578)
College and above
25 (80.7)
6 (19.3)
0.432 (0.152–1.229)
0.364 (0.082–1.614)
Paternal educational level
No formal education
15 (57.7)
11 (42.3)
1
1
Can read and write only
22 (73.3)
8 (26.7)
0.496 (0.161–1.524)
0.484 (0.135–1.706)
Primary (1–8)
51 (70.8)
21 (29.2)
0.561 (0.222–1.422)
0.692 (0.237–1.898)
Secondary (9–12)
59 (79.7)
15 (20.3)
0.347 (0.142–1.960)
0.331 (0.113–1.971)
College and above
40 (76.9)
12 (23.1)
0.409 (0.149–1.124)
0.511 (0.168–1.551)
Mode of delivery
SVD
151 (77.8)
43 (22.2)
1
1
Instrumental delivery
16 (66.7)
8 (33.3)
1.756 (0.704–4.379)
1.99 (0.74–5.35)
Cesarean section
20 (54.1)
17 (45.9)
2.985 (1.439–6.194)
3.326 (1.485–7.451) *
Obstructed labor
No
173 (76.2)
54 (23.8)
1
1
Yes
14 (50.0)
14 (50.0)
3.204 (1.438–7.139)
0.630 (0.108–3.683)
Fetal presentation
Vertex
169 (78.2)
47 (21.8)
1
1
Non-vertex
18 (46.1)
21 (53.9)
4.195 (2.067–8.513)
4.614 (2.163–9.84) *
Prolonged labor
No
168 (76.7)
51 (23.83)
1
1
Yes
19 (52.8)
17 (47.2)
2.947 (1.427–6.089)
2.959 (1.318–6.595) *
Hypertension during pregnancy
No
176 (75.5)
57 (24.5)
1
1
Yes
11 (50.0)
11 (50.0)
3.088 (1.271–7.500)
3.421 (1.318–8.881) *
Urinary tract infection (UTI)
No
175 (74.5)
60 (25.5)
1
1
Yes
12 (60.0)
8 (40.0)
1.944 (0.758–4.985)
1.356 (0.435–4.228)
Sex of the newborn
Female
116 (76.3)
36 (23.7)
1
1
Male
71 (68.9)
32 (31.1)
0.689 (0.393–1.206)
0.546 (0.290–1.025)
Key: 1: Reference category; AOR = Adjusted odds ratio, COR = Crude odds ratio, *Statically significant at p-value<0.05*
Discussion
This study was conducted to determine the magnitude of NNM and associated factors at selected public hospitals in Jimma zone, southwest Ethiopia. The finding of this study shows that the magnitude of NNM was 26.7% with 95% CI: (21.6%-32.5%). This finding is comparable with the finding of the studies that were conducted in Brazilian university hospitals 30.37% [ 7 ] and northeastern Brazil 22% [ 6 ].
However, the current finding is lower when compared to the study done in Uganda which was 36.7% [ 16 ]. This might be due to a study carried out in Uganda was among mothers with serious obstetric complications, and these complications during pregnancy, labor, and delivery could lead to life-threatening conditions in neonates and place them in the NNM.
On the other hand, the prevalence of NNM in the current study was greater than the finding of the studies conducted in Brazil that reported prevalence of NNM from 3.3% to 8.6% [ 18 , 24 , 26 , 28 ] and in India 8.76% [ 27 ]. These differences might be due to differences in socio-economic characteristics of the study population, health care delivery system (technologies, early detection of problems). Furthermore, the discrepancy may also be attributable to the criteria used in the identification of NNM cases in which studies conducted in the southeast and northeast Brazil and India used only pragmatic criteria to identify NNM [ 18 , 27 , 28 ] (whereas both pragmatic and management criteria were used in the current study.
Hypertension increased the odds of NNM by three times as compared to those mothers who had no history of hypertension during pregnancy. This result is in line with the finding of studies done in Brazil [ 25 , 28 ]. This might be due to hypertension during pregnancy may cause complications to fetuses during intrauterine life like intrauterine growth restriction and in extrauterine life such as preterm delivery which is more likely to be LBW and also causes birth asphyxia [ 29 ].
In this study, non-vertex fetal presentation during delivery had found to increase the chance of developing NNM. Similarly, the study conducted in Gamo Gofa, Ethiopia, found non-vertex fetal presentation was the determinant factor of NNM [ 23 ]. This might be since malpresentation during pregnancy and labor has a high risk of birth asphyxia, birth trauma, and other complications and also lead to obstructed and prolonged labor which can result in different complications to the newborn [ 30 ].
Obstetric complications during labor and delivery were also showed a significant association with NNM. This study revealed that the odds of NNM was 3 times higher in mothers with obstructed labor when compared to mothers with normal labor. This might be because abnormal progress of labor that diminishes uteroplacental blood flow can cause fetal distress, fetal hypoxia, and other complications that predispose neonates to life-threatening conditions [ 31 ].
In this study cesarean mode of deliveries was associated with an increased risk of NNM. This is in line with the finding of a study done in Southern Ethiopia [ 23 ] and studies done in Brazil [ 15 , 20 , 25 ]. Various studies also found that cesarean delivery was associated with an increased risk of APGAR score less than 7 at the 5th minute, preterm birth, low birth weight, neonatal resuscitation, and admission to neonatal intensive care units (NICU), all of which collectively increased the tendency of becoming a near miss [ 32 – 34 ].
The strength of this study was that validated and standardized Neonatal Near Miss identification criteria were used to reduce misclassification and also cross-checked maternal medical records to mitigate recall bias and enhance its validity. The study, however, has some limitations as the neonates were only sampled from hospitals, which may lead to underestimation of the prevalence of NNM as mothers who deliver at home and low-level health facilities (facilities without NICU) were not included in the study because it is difficult to obtain information on the condition of newborns at birth like; APGAR scores at 5 th minute, birth weight and gestational age for those neonates delivered at home.
Conclusion
The magnitude of NNM in the study area was found to be high compared to most studies. In this study hypertension during pregnancy, prolonged labor, cesarean mode of delivery, and non-vertex fetal presentation during delivery were significantly associated with being a near miss. Therefore, concerted efforts are needed from local health planners and health care providers to improve maternal health care services especially in early identification of the complications and taking appropriate management. And also further research is needed to identify other factors by using other study designs.
Supporting information
S1 Fig
Schematic representation of the sampling procedure followed to get study participants in Jimma zone, Southwest Ethiopia, 2020.
(TIF)
S1 Dataset
The raw data supporting the findings of this article.
(SAV)
S1 Questionnaire
Data collection tool for the study.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Systemic sclerosis (SSc) is an autoimmune disease characterized by microvascular dysfunction, activation of the immune system and tissue fibrosis. Pathogenesis of SSc is complex and poorly understood and it has been suggested that a genetic predisposition might contribute to the development of the disease together with environmental agents, such as viruses or chemical agents, which could activate both cellular and humoral immunity [ 1 , 2 ]. According to the current understanding, immune system leads to vascular injury with either release of pro-inflammatory cytokine or production of auto-antibodies that damage endothelial cells (ECs), resulting in promoted fibroblast proliferation [ 3 – 6 ].
So far the contribution of complement system to the pathogenesis of SSc has not been deeply investigated, most likely because in clinical practice the main plasma complement proteins (C3 and C4) are usually within the reference range. Nevertheless, hypocomplementemia has been described in SSc patients with more severe disease [ 7 , 8 ], while high plasma levels of complement activation products have been correlated with clinical severity of SSc [ 9 – 12 ]. Recently, Batal et al. found that small vessel C4d score was higher in SSc patients with renal crisis compared with normotensive controls and that this score correlated with increased risk of unrecovered renal function [ 13 ]. Furthermore, Arason outlined a deficiency of complement-dependent prevention of immune precipitation in SSc [ 14 ] and Sprott et al. documented presence of the C5b-9 complex and C5a receptor in microvessels of SSc skin sections both in early and in late phases of the disease [ 15 ].
It is conceivable that activation of complement system in SSc might be due to immune complexes [ 16 , 17 ], but inadequate protection of the EC surface might also be involved. In fact, ECs located at the interface between blood and tissues are natural targets of complement attack. The classical functions of complement, such as opsonization, recruitment of inflammatory cells, target cell lysis, immune complex clearance, and its capability to influence many other pathways, such as coagulation cascade and angiogenesis, seem to be pivotal for the integrity of ECs [ 18 ]. In normal conditions, complement attack is tightly regulated by fluid-phase and surface-bound regulatory proteins which allow adequate immune surveillance while ensuring protection of host cells [ 19 ]. In different vascular diseases, overtly activated or poorly controlled complement activation not only promotes EC damage and apoptosis, but also enhances the expression of vascular cell adhesion molecules and amplifies the local immune response [ 20 ].
Factor H (FH) is the main fluid-phase regulator of the alternative complement pathway (AP). It acts on C3, the central component of the complement cascade by accelerating decay of C3 convertase and acting as a cofactor of factor I (FI) in the inactivation of C3b. This plasma regulator also contributes to human tissue protection allowing complement activation only to foreign targets or altered self cells [ 21 , 22 ]. In our previous study, we documented high FH levels in sera of SSc and Sclerodermatous Graft Versus Host Disease (ScGVHD) patients, but only in SSc subjects we found a defective capacity of FH to protect cellular surface from complement mediated damage in in vitro experiments [ 23 ].
On human ECs, other complement regulators participate in cell protection from activation of both AP and classical complement pathway (CP). The group of membrane-bound complement regulators include the membrane cofactor protein (MCP or CD46), which is a cofactor of FI in the proteolytic inactivation of C3b and C4b, and the decay accelerating factor (DAF or CD55), which accelerates the breakdown of C3- and C5-convertases [ 24 – 26 ]. Recently, Venneker et al. demonstrated an impaired expression of MCP and DAF in endothelium of the lesional and non-lesional skin of SSc patients and in the skin of patients with morphea, in comparison to healthy controls and subjects affected by other autoimmune diseases, suggesting that a defective endothelial protection might be mediated by reduced expression of the complement regulatory proteins [ 27 , 28 ].
Since the mechanisms involved in SSc pathogenesis are still under investigation, we focused our attention on local complement activation and regulation, using skin biopsies as an observational window of the EC damage related to SSc.
Materials and Methods
Patient selection
The initial study population consisted of 71 SSc patients and 29 age- and sex-matched healthy volunteers (H). All SSc patients fulfilled the American College of Rheumatology criteria for the classification of SSc. Distinction between limited cutaneous SSc (lcSSc) and diffuse cutaneous SSc (dcSSc) was made according to the criteria of LeRoy et al [ 29 ]. Exclusion criteria for the SSc patients were co-morbidities associated with complement activation due to a potential confounding effect [ 30 – 32 ]. On this basis ten patients were excluded, so 25 subjects with dcSSc and 36 subjects with lcSSc were finally enrolled in the study.
All the enrolled patients underwent a detailed clinical examination and laboratory evaluation, including analysis of antinuclear antibodies by indirect immunofluorescence on HEp-2 cells, anti-ENA antibodies by an ELISA method, CRP determination and standard direct and indirect Coombs tests. Skin involvement was assessed modified Rodnan skin score (mRSS) [ 33 ] by evaluating dermal thickening in seventeen anatomic sites, using a score from 0 to 3 (where 0 indicates normal). Skin lesions were subdivided on the basis of skin score into mild (mRSS <14), moderate (mRSS between 15 and 29), high (mRSS between 30 and 39), and severe (mRSS ≥ 40) [ 34 ]. Moreover, all the patients underwent the following investigations: electrocardiogram, pulmonary function test with diffusing capacity for carbon monoxide adjusted to haemoglobin (DLCO), Doppler echocardiogram, and chest high-resolution computed tomography. Nailfold videocapillaroscopy (NVC) was assessed in 54 patients (25 with dcSSc and 29 with lcSSclcSSc) by a unique operator, who was unaware of the aims of the study. The microvascular alterations were classified into 3 different patterns: early, active, and late [ 35 ].
Archival skin punch biopsies of 8 patients with SSc (4 lesional SSc skins from wrists and 4 non-lesional SSc skins from backs), were processed and compared with those of 4 patients with ScGVHD (4 lesional GVHD skins from wrists, leg or forearm) and 8 healthy individuals (4 non-lesional H skins from wrists and 4 non-lesional H skins from backs). The institutional review board of the Verona Hospital approved the protocol (CE 1183/1570) and all patients and healthy controls provided written informed consent before participating in the study.
Sample collection
Venous blood was drawn into 10 ml BD Vacutainer tubes and allowed to clot at room temperature for 1 hour. Venous blood (10 ml) was also collected in pre-cooled tubes containing 0.015 M sodium citrate and centrifuged immediately at 4°C. Serum or plasma was separated from cells by centrifugation at 3000 × g for 15 min at 4°C followed by an additional similar centrifugation in order to remove cellular debris. Serum and plasma samples were then aliquoted in 1.5 mL Eppendorf tubes and stored at -80°C until use.
Enzyme-linked immunosorbent assay (ELISA)
Concentrations of FH in serum and SC5b-9 in plasma were assessed using the Human Complement Factor H ELISA kit (Hycult Biotech, Uden, The Netherlands) and the MicroVue C5a Plus EIA (QUIDEL), respectively, according to the manufacturers’ instructions. Serum samples were assayed after a 1:8000 dilution, while plasma samples were assayed using a 1:10 dilution.
Concentrations were calculated using standard curves generated with specific standards provided by the manufacturers. Optical density was measured by microtitre plate reader at 450 nm. Each sample was measured in duplicate.
FH-dependent hemolysis assay
The hemolysis test was performed as previously described [ 36 ]. Briefly, 100 μl of each serum was diluted in 400 μl of alternative pathway activating buffer (AP buffer: 2.5 mM barbital, 1.5 mM sodium barbital, 100 mM NaCl, 7mM MgCl 2 , 10 mM EGTA pH 7.2–7.4). A duplicate of each sample was prepared in the same buffer plus 50 mM EDTA and was used as blank. 200 μl of sheep erythrocytes (1x10 8 cells/ml in AP buffer) were added to both samples and blanks. The mixtures were incubated at 37°C under mixing. After 15 min, samples were transferred to 0°C and the reaction was stopped with 1 ml of stop solution (2.5 mM barbital, 1.5 mM sodium barbital, 144 mM NaCl, 2 mM EDTA, pH 7.2–7.4). A sample with 400 μl of AP buffer and 200 μl of sheep erythrocytes, incubated in the same conditions but stopped with 1 ml of stop solution plus 0.1% Triton X-100 was used as the “control of total lysis”.
The mixtures were centrifuged at 2600 × g for 15 min and hemolysis was determined by measuring the absorbance at 414 nm in the supernatants. The percentage of lysis in each sample was calculated as percentage of the absorbance of the sample divided by that of the “control of total lysis” (OD at 414 nm). Samples were considered positive when percentage of lysis was higher than 12.5%; spontaneous lysis of sheep RBC in the presence of normal human serum was 0.1–12.15%.
Western blotting
Proteins were extracted from skin specimens by Tri-Reagent method (Sigma) and quantified using BCA assay (Pierce Company). Sixty micrograms of proteins were subjected to electrophoresis in SDS-(12%) polyacrylamide gel under reducing conditions and then blotted to a nitrocellulose membrane using the Mini Trans-Blot Cell (Bio-Rad). The membranes were incubated with blocking buffer (3% w/v skimmed milk in Tris Buffered Saline and 0.1% v/v Tween-20) and then probed with 1:5,000 primary rabbit anti-human MCP antibody (Abcam), under shaking for 1 hour at 22°C. After the washing and the incubation with 1:15,000 horseradish peroxidase-conjugated anti-rabbit antibody (Abcam) for 1 hour at 22°C, the immunocomplexes were detected by chemiluminescence using the ECL Plus Western Blotting Detection Reagents (GE Healthcare Life Sciences). We developed the images on Kodak BioMax XAR Films in the darkroom, adjusting the exposure time depending on the intensity of the protein bands. Blots were stripped by adding a Stripping Acid Solution (50 mM glycine, 1% w/v SDS, 1% w/v Tween-20; pH 2.2 with HCl), shaking for 40 min at 37°C and reincubated with mouse monoclonal anti-human ß-tubulin antibodies (1:1,000 dilution; Abcam) to confirm the equal sample loading of the gels and the efficiency in electrophoretic transfer. Densitometric analysis of the bands was performed using Quantity One software (Biorad).
C5b-9 and MCP immunofluorescence assay
Paraffin-embedded tissue blocks were cut into 2–3 μm sections and mounted on adhesive microscope glass slides. After deparaffinization and rehydration the antigen retrieval was performed in pre-warmed citrate buffer (pH 6 temp. 95°C) for 30 minutes. Sections were cooled to room temperature and incubated with a protein blocking serum-free solution for 15 minutes at 22°C to block non-specific binding.
For immunofluorescence staining, sections were separately incubated with three different antibodies: a monoclonal mouse anti-human CD31 antibody (Dako, JC70A, 1:50 dilution; a marker for vascular endothelium), a monoclonal rabbit anti-human CD46 antibody (Abcam, ab108307, 1:500 dilution), and a monoclonal murine anti-human SC5b-9 antibody (QUIDEL, A239, 1:250 dilution). Slides were incubated 30 minutes at pH 6 with the corresponding Alexa 546-conjugated antibody (anti-mouse or anti-rabbit; INVITROGEN Molecular Probe, diluted 1:1000). Reduction of the autofluorescence background was obtained by the incubation with Sudan Black B 0.1% (Sigma-Aldrich). Nuclei were stained with Prolong Gold antifade reagent with DAPI (INVITROGEN Molecular Probe). Slides were analysed by a Olympus BX61 microscope.
Genetic analysis
DNA was extracted from the blood buffy coat by automated Blood DNA purification kit on the Maxwell 16 instrument (Promega), according to the manufacturer’s instructions. DNA preparates were stored at -80°C until the analyses were performed.
DNA samples of six SSc patients, which resulted positive to the hemolysis test, were sent to Secugen Diagnostic (Madrid, Spain) for the genetic analyses. Exons and promoter regions of genes for FH, FI and MCP were sequenced and compared to the published sequences in Ensemble, NCBI, and aHUS databases. Genotypes and haplotypes for common polymorphisms (SNPs) in these genes were also analyzed. The two polymorphic variants of MCP promoter region (-366A>G, rs2796268 and -652A>G, rs2796267), of those identified thereby, were assessed with TaqMan allelic discrimination assays designed on demand (Applied Biosystems, Foster City, CA, USA) on a 7500 Real Time PCR instrument (Applied Biosystems).
Statistical analysis
Differences in FH concentrations was evaluated by Kruskal Wallis test, as this variable presented a highly skewed distribution according to Skewness—Kurtosis test. The Mann Witney test was performed, adjusting for multiple comparisons by Bonferroni correction, if the Kruskal Wallis test result was significant. The same analyses were performed for C5b-9 and CD46 vessel ratios.
The differences in categorical variables, including allele frequencies of each SNP, were evaluated by the Fisher’s exact test. Significance was set at p <0.05.
Results
Demographic and clinical data of SSc patients
The main clinical characteristics of the SSc patients studied are listed in Table 1 . Twenty SSc patients presented anti-Scl70 antibodies, thirty-one subjects were positive for anti-centromere antibodies, and ten patients presented antinuclear antibodies. The mRSS score documented the presence of mild disease in 73.8% and of moderate disease in 24.6% of the cases. One patient only had high skin involvement. Evidence of interstitial lung disease was found in seventeen patients (27.9%).
10.1371/journal.pone.0114856.t001
Table 1 Clinical data of SSc patients.
Age (years)
61.6±13.1
Sex §
men
11 (18.0%)
women
50 (82.0%)
Disease pattern §
dcSSc
25 (41.0%)
lcSSc
36 (59.0%)
Autoantibody pattern §
Anti-Scl70
20 (32.8%)
ACA
31 (50.8%)
ANA
10 (16.4%)
mRSS §
mild
45 (73.8%)
moderate
15 (24.6%)
severe
1 (1.6%)
NCV pattern §
early
12 (19.7%)
active
19 (31.1%)
late
23 (37.7%)
missing
7 (11.5%)
Pulmonary fibrosis §
present
17 (27.9%)
absent
44 (72.1%)
Anti-Scl70 = anti-Scl70 antibodies; ACA = anticentromere antibodies; ANA = antinuclear antibodies; mRSS = modified Rodnan Skin Score; NVC = nailfold videocapillaroscopy.
§ values expressed as absolute number and percentages.
At videocapillaroscopic analysis, twelve subjects showed early microvascular alteration pattern, while nineteen and twenty-three patients had respectively active and late nailfold microvascular damage ( Table 1 ).
In SSc, FH levels are increased in presence of normal soluble C5b-9 complement complex
Previously, we have reported that FH is increased in serum of the SSc patients [ 23 ]. Here, we evaluated C5b-9 plasma concentrations as an index of systemic complement activation. We found that C5b-9 plasma concentrations were similar in the SSc patients and healthy controls (dcSSc 134 ng/ml, IQR 93–203; lcSSc141 ng/ml, IQR 89–202; H 124 ng/ml, IQR 82–159; p = 0.49); whereas, the serum levels of FH were higher in SSc patients with both diffuse and limited subsets (dcSSc patients 126 μg/ml, IQR 114–150, p = 0.0025; lcSSc patients 124 μg/ml, IQR 108–152; p = 0.0054), compared to healthy controls (108 μg/ml, IQR 93–120 μg/ml), according to the Bonferroni correction.
No correlation was observed between serum FH concentrations and CRP levels in the samples from the SSc patients (p = 0.93; data not shown), although CRP had been reported to correlate with FH in patients with age-related macular degeneration (AMD) [ 37 ].
The function of FH is impaired in SSc patients
The presence of normal values of C5b-9 in the peripheral circulation of SSc patients with increased levels of FH, made us to consider local activation of complement. To study the regulatory activity of FH on cell surfaces, we carried out a hemolysis assay in the presence of Mg-EGTA, which selectively blocks CP and lectin pathway of complement (LP). As shown in Table 2 , the FH-dependent hemolysis test revealed that 40% and 16% samples of the dcSSc and lcSSc groups were positive, respectively, whereas healthy controls were all negative (p<0.001. In addition, we found that the serum hemolytic activity was higher in SSc patients with high levels of FH (p = 0.012) ( Table 2 ).
10.1371/journal.pone.0114856.t002
Table 2 Serum hemolytic activity in healthy subjects and SSc patients.
Healthy controls
SSc patients
Hemolysis test
Negative [n = 29]
Positive [n = 0]
Negative [n = 45]
Positive [n = 16]
FH serum level, mean (IQR) [μg/ml]
108 (93–120)
-
122 (103–139)
151 (122–186)
To exclude possible additional factors that might interfere with the hemolysis test (i.e. antibodies against sheep red cells), we carried out the direct and the indirect Coombs tests, which resulted negative in all the SSc patients studied (data not shown).
Serum hemolytic activity in SSc is mediated by the alternative and classical complement pathways
Since the local complement activation could be mediated by any of the three complement pathways, we evaluated the serum hemolytic activity of SSc patients both in the presence and absence of Mg-EGTA. Interestingly, the hemolysis in the absence of Mg-EGTA was higher than in presence of Mg-EGTA ( Fig. 1 ), suggesting a possible concert action of both AP and CP/LP in the complement dysregulation on the cell surface.
10.1371/journal.pone.0114856.g001
Fig 1
Serum hemolytic activity is increased in SSc patients.
Sera from three different SSc patients (SSc1-SSc3) and three different healthy subjects (H1–H3) were tested both in the presence and absence of Mg-EGTA. Mg-EGTA was used to block the classical and lectin pathways of complement. Data are expressed as percentage of full RBC lysis.
The endothelium of SSc patients is not protected from activated complement
Considering that the vascular endothelial bed is involved in the early phase of SSc, we investigated whether the ECs of SSc patients might be exposed to complement mediated damage, as previously suggested by Sprott [ 15 ]. To study this we compared archival skin biopsies of 8 SSc patients showing positive hemolysis test to those of 4 subjects with ScGVHD and 8 normal individuals. As shown in Fig. 2 , C5b-9 was detected in vessels of SSc and ScGVHD skin, while microvasculature of healthy subjects resulted completely negative (median proportion of positive vessels—H: 0.000; lesional SSc: 0.087; non-lesional SSc: 0.033; ScGVHD: 0.044). Pairwise comparisons after Kruskal Wallis test documented a significant difference between the samples from healthy skin and those from lesional SSc (p = 0.0002; Bonferroni-adjusted level for significance = 0.0042) ( Fig. 2 ).
10.1371/journal.pone.0114856.g002
Fig 2
Detection of the complement membrane attack complex (or C5b-9) on skin endothelial cells of SSc patients.
Skin biopsies from 8 SSc patients (4 lesional SSc skin and 4 non-lesional SSc skin), 8 healthy subjects and 4 patients with ScGVHD were immunostained for C5b-9 specific antibody (orange). Anti-CD31 antibody (green) was used as marker of endothelial cells and DAPI to stain the nuclei of the cells. Data of the analysed fields ( n = 20 for each slide) are expressed as a ratio between C5b-9 positive vessels and total number of vessels.
This finding supported the local C5b-9 deposition around blood vessels of SSc skin. Thus, we looked for the expression of MCP, which is normally expressed on ECs and acts as a local complement regulator. The immunofluorescence staining revealed significantly reduced amount of MCP in the endothelium of lesional SSc skin than in the healthy controls (median proportion of positive vessels—H: 0.35; lesional SSc: 0.12; non-lesional SSc: 0.25; ScGVHD: 0.22; p value of pairwise comparisons H vs lesional SSc = 0.0004; Bonferroni-adjusted level for significance = 0.0042) ( Fig. 3 ). Immunoblot analysis provided additional evidence of low expression of MCP in SSc skin sections (median proportion of CD46 H: 0.570; SSc: 0.258; p = 0.03 vs H) ( Fig. 4 ).
10.1371/journal.pone.0114856.g003
Fig 3
The local complement regulator MCP (CD46) is reduced on vascular endothelial cells of SSc skin biopsies.
Skin sections from 8 SSc patients (4 lesional SSc skin and 4 non-lesional SSc skin), 12 controls (8 healthy subjects and 4 ScGVHD patients) were immunostained with anti-CD46 (green) specific antibody. Anti-CD31 antibody (orange) was used as a marker of endothelial cells. Data of the analysed fields ( n = 20 for each slide) are expressed as ratio between MCP (CD46) positive vessels and total number of vessels.
10.1371/journal.pone.0114856.g004
Fig 4
Immunoblot analysis of MCP (CD46) of skin biopsies from healthy controls and SSc patients.
Protein extract from skin biopsies of 8 SSc patients (4 lesional SSc skin and 4 non-lesional SSc skin) and 8 healthy subjects were analyzed by immunoblot using a specific antibody. Results of a representative experiment of four are shown. Data are expressed as an optical density ratio of MCP (CD46) and β-tubulin.
SSc population has higher prevalence of SNPs polymorphic variants in MCP promoter region
The sequencing analysis of FH, FI and MCP genes revealed one heterozygous mutation in FI gene (c.1534+5G>T) in one of the six analyzed patients (patient n. 1), while no mutations were found in FH and MCP genes. Homozygosis for non-prevailing SNPs of FH gene were found in three subjects (patients n. 2, n. 3 and n. 6), instead homozygosis for uncommon SNPs of MCP gene were detected in four patients (patient n. 1, n. 2, n. 3 and n. 4), as shown in S1 Table . Only patient n. 5 did not present mutations or non-prevailing SNPs in the analyzed genes.
Since expression of MCP was different in SSc and healthy skin, we next analyzed prevalence of the two allelic variants in the promoter region of the MCP gene (-366A>G and -652A>G SNPs) in all the SSc patients and healthy subjects. These polymorphisms have been reported to correlate with the disease severity of the atypical hemolytic uremic syndrome (aHUS) [ 38 ].
We found differences in the distribution of both the MCP polymorphisms between the SSc patients and healthy controls despite the small size of our sample. In particular, in SSc the allele frequency reached the statistical significance for -366G (0.42 vs 0.24; p = 0.021), whereas for -652G the analysis were close to significance (0.42 vs 0.27; p = 0.071) compared to the controls ( Table 3 ).
10.1371/journal.pone.0114856.t003
Table 3 Allele frequencies of two SNPs in MCP gene promoter.
SNPs
NCBI id
SSc patients
Healthy controls
P -values
Genotypes
Allele frequencies
Genotypes
Allele frequencies
1/1
1/2
2/2
1
2
1/1
1/2
2/2
1
2
-366A>G
rs2796268
23
36
12
0.58
0.42
16
9
2
0.76
0.24
0.021
-652A>G
rs2796267
22
38
11
0.58
0.42
15
9
3
0.72
0.28
0.071
The P- value is the result of a two-sided Fisher exact test for the comparison of the allelic frequencies of each SNP between the SSc and control groups.
Discussion
Here, we propose the endothelium-bound membrane attack complex of complement (MAC or C5b-9) as a promising marker of active vascular damage in SSc despite its normal plasma levels. Previous studies in other autoimmune diseases have found similar discrepancy between plasma levels of complement activation products and local complement activation. In SLE , the activation of complement system has been reported in different organs, such as lung and kidney, without changes in serum levels of C3 and C4 [ 39 ]. In rheumatoid arthritis the presence of complement activation fragments in joint fluid and the deposition of C5b-9 in synovial tissue are common findings, too, although the plasma level of C5b-9 may not be altered [ 40 ].
In this study, we used erythrocytes as a model of cellular surface sharing common characteristics with other cell types. The serum hemolytic activity of SSc patients does not, however, represent well complement activation and regulation on endothelium since red blood cells are devoid of MCP. Thus, we studied SSc skin biopsies as an observational window of endothelial damage in SSc. We found that the complement system is locally activated, as documented by the abnormal deposition of MAC on the endothelium and, concordant with this, we found reduced MCP expression on vascular endothelial surface. This finding is in agreement with that previously reported by Venneker on SSc skin [ 27 ]. Based on these results we propose that the lack of the fine local regulation of complement activation on vascular endothelium might promote complement activation leading to sublethal MAC depositions which are known to cause EC apoptosis. This could explain at least partially the initiation or propagation of tissue fibrosis in SSc.
In other disorders characterized by abnormal complement regulation (i.e. aHUS and AMD), it has been shown that mutations or uncommon SNPs of complement regulatory proteins, such as FH and MCP, generate a state of haploinsufficiency unable to prevent complement mediated tissue damage [ 41 – 44 ]. As some SNPs are organized in specific haplotype blocks within the regulator of complement activation gene cluster in human chromosome 1q32, we examined two SNPs in the MCP promoter region (-366 A>G, -652 A>G) showing a strong linkage disequilibrium in the region of the MCP gene. It is of interest to note that the polymorphic variants of these two SNPs have been related to a 25% lower transcriptional activity of the gene promoter and have been linked to enhanced severity of aHUS disease [ 38 ]. In our SSc patients, the minor variants of the two SNPs were more usual than in healthy controls, suggesting a possible role of these SNPs in the severity of SSc disease. It remains to be studied if the SNP -366A>G in the MCP gene could be used as a predictive marker for more severe or progressive disease.
Beside the clinical similarities between SSc and ScGVHD, we did not observe the prevalence of the same polymorphic variants in the promoter region of MCP gene in ScGVHD patients studied (data not shown). Moreover, the observation that in skin biopsies of ScGVHD patients the local amount of C5b-9 was increased without significant differences in the MCP expression, supports the hypothesis of additional mechanisms involved in complement activation on skin endothelium in ScGVHD. The important role of complement system in transplant rejection reactions is supported by the encouraging results with the anti-C5 monoclonal antibody (Eculizumab) that blocks the activation of the terminal complement cascade and formation of MAC [ 45 , 46 ].
We propose that several mechanisms known to be involved in SSc pathogenesis might be affected by locally activated complement caused by impaired fine-tuning of this powerful innate immune defense ( Fig. 5 ). In fact, different pathways are likely to contribute to vascular dysfunction processes in SSc, such as direct vascular damage, pro-inflammatory response and coagulation cascade activation. Studies in different diseases have shown functional connections between activated complement molecules and these pathways [ 18 , 47 – 49 ].
10.1371/journal.pone.0114856.g005
Fig 5
Schematic diagram of a possible role of activated complement in the pathogenesis of SSc.
Beside its conventional role in the innate immunity, recent evidence suggests that complement system modulates the acquired immunity and regulates the coagulation cascade activation. Locally, activated complement products reduce neoangiogenesis and promote tissue fibrosis.
In aHUS dysregulated complement activation is clearly causing the endothelial cell injury, hemolysis and microvascular thrombosis. In addition, in models of multifactorial disease (e.g. antiphospholipid syndrome) a partial or complete loss of function of complement regulators might play a relevant role in the pathogenesis, contributing to more severe organ damages and clinical complications. The revision of the literature on these disorders confirms that pharmacological treatment with Eculizumab might ameliorate clinical manifestations in severe cases [ 50 , 51 ].
Future studies need to be carried out to better characterize the role of complement system on vascular damage in SSc and to verify in a large number of SSc patients the possible beneficial effects of a pharmacological treatment with inhibitors of complement system.
Supporting Information
S1 Table
SNPs of FH and MCP genes.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Glioblastoma multiforme (GBM; grade IV glioma) is the most prevalent and malignant form of primary brain tumor in adults [ 1 , 2 ]. Median survival is a mere 15 months despite radiotherapy, surgical resection, and chemotherapeutic interventions [ 1 ]. GBM tumors are especially difficult to treat, since surgical resection invariably leaves behind glioblastoma stem-like cells (GSCs), which are highly invasive tumor cells uniquely resistant to standard therapies.
GSCs are a population of GBM cells that play a major role in the particularly aggressive nature of GBM tumors and share traits with neural stem cells (NSCs), including self-renewal and multipotency [ 3 ]. Remarkably, transplantation of only 100 GSCs into the mouse forebrain is sufficient to form a glioma tumor [ 4 ]. Several features of GSCs contribute to GBM malignancy following initial tumor formation, including rapid proliferation and highly diffuse invasion throughout the brain [ 5 ]. Additionally, standard chemotherapeutic agents, which eradicate the majority of GBM cells, have a reduced effect on GSCs, and surviving GSCs contribute to tumor recurrence, a hallmark of GBM [ 5 , 6 ]. These features underscore the necessity for development of novel therapeutic candidates that precisely target GSCs and halt uncontrolled growth and invasion.
Ion channels passively conduct ions down their electrochemical gradient in response to external stimuli, whereas ion transporters use energy to pump ions across their concentration gradients [ 7 , 8 ]. Ion channels and pumps are responsible for conducting electrical currents in all nerve, muscle, and cardiac cells, however, they also play vital roles outside of regulating electrical excitability in both normal and cancerous cells. It is increasingly being understood that dysregulated ion channels and pumps are implicated in multiple processes in various cancers [ 9 ], including regulation of the cell cycle [ 10 ], migration [ 11 ], and apoptosis [ 12 ]. Promisingly, inhibitors to various ion channels have been demonstrated to hinder tumor formation and growth [ 13 , 14 ].
Ion channels and transporters are likewise implicated in GBM tumor growth and malignancy [ 15 – 17 ]. Genomic analysis reveals that genes involved in passing or transporting Na + , K + , and Ca 2+ are among the most frequently mutated functional groups in GBM affecting 90% of the GBM samples studied [ 18 , 19 ]. Functionally, ion channels and pumps influence both GBM cell migration and proliferation. For instance, dysregulated K + and Cl - channels regulate osmotic drive allowing for cell shape and volume changes that promote glioma cell migration [ 20 ], and Ca 2+ -activated K + (BK) channels control glioma cell growth [ 21 ]. However, little is known about the expression and functional relevance of ion channels in the stem cell population despite their central importance to GBM tumor initiation and progression.
We propose that dysregulation of ion channel expression is central to the abnormal growth and migratory properties that drive GSC malignancy. Therefore, a greater understanding of the ion channels operating in GSCs may reveal novel, therapeutically relevant mechanisms to target GSCs. To assess the expression pattern of ion channels that may contribute to glioma malignancy, we analyzed an RNA sequencing database of 20 patient-derived GSC isolates and 5 neural cell type controls. We identified a unique set of druggable ion channels enriched in GSCs that were associated with distinct gene mutation signatures and poor patient survival outcomes. Pharmacological blockade and genetic knockdown of these channels impaired GSC viability. Identification of GSC-enriched ion channels and the mechanism by which they drive GSC malignancy could identify novel therapeutics to inhibit GSC-driven tumor growth and improve patient outcomes.
Results
Expression of ion channels, transporters, and gap junctions in GSCs
To profile the enrichment patterns of ion channel genes in GSCs, we compared transcriptomic data for 20 human GSC isolates to that of 3 human NSC lines and 2 normal human astrocyte (NHA) cell lines. We used a comprehensive list of 266 (7 out of the original 273 were not available in our dataset) druggable human ion channel genes [ 22 ] (guidetopharmacology.org) and 152 human ion transporter genes (broadinstitute.org/gsea/msigdb, GO:0015075). The strategy of comparing GSCs to NSCs/NHAs was used as a first approximation to enrich for genes specific to malignant stem cell phenotypes and not shared by non-transformed neural progenitors or astrocytic glia. Gene expression was calculated as CPM (counts per million) values for this data analysis. Note that, as opposed to CPM, other reports referred to later used FPKM (fragments per kilobase of exon per million reads mapped) values in a similar way to quantify and report differential gene expression. Additional details are provided for: cell lines ( S1 Table ), RNA-seq methods ( Materials and methods ), and ion channel gene set ( S2 Table ). S1 Fig summarizes the experimental design used throughout this study.
Using a simple fold-change approach, we found differential expression (≥|2| log 2 fold change) in 56 out of 251 ion channel-related genes (15 genes were excluded due to zero values in denominator) when comparing GSCs to NSCs/NHAs. Of these 56 genes, 44 were GSC-enriched (≥2 log 2 fold change; Fig 1A , red points), and 12 were NSC/NHA-enriched (≤-2 log 2 fold change; Fig 1A , blue points). Since fold change differences can skew contributions from lowly expressed genes, Gene Set Enrichment Analysis (GSEA; details in Materials and methods ) was used to further analyze differential expression patterns. Using a Signal2Noise metric, which accounts for both mean absolute levels and variance within the classes, we found that 107 out of 266 ion channels highly contributed to the enrichment of ion channels in GSCs compared to control cells. These genes were rank-ordered according to their GSEA enrichment score, and the top 40 differentially expressed ion channel-related genes were selected for further study. Twenty-five of these met our threshold expression criteria (mean GSC CPM ≥1) and formed the basis of a GSC-enriched ion channel gene set (hereby referred to as “IGCs”) used for all remaining analyses in this study ( Fig 1B ; see S1 Fig for summary of selection criteria). Notably, these genes represented a diversity of ion channel types. Scatter plots of CPM values for selected highly-ranked IGCs demonstrated the significant differences in expression between GSC and NSC/NHA cell lines ( Fig 1C ). This was validated in multiple GSC lines compared to the NSC line CB660 using real-time quantitative PCR (RT-qPCR) for all IGCs tested except P2RX4 , for which expression differences were not robust ( Fig 1D ). In summary, we found a high frequency (~40%) of ion channel genes associated with GSCs compared to control NSCs/NHAs, and ~10% overall formed the basis of a GSC-enriched ion channel gene set.
10.1371/journal.pone.0172884.g001
Fig 1
RNA sequencing identifies ion channels enriched in GSCs.
A. Mean GSC CPM values plotted against mean control (NSC and NHA) CPM values for all ion channel-related genes. Red and blue points represent genes that are differentially expressed ≥|2| log 2 fold change in each subclass. B. Heat map of the most differentially enriched ion channels in GSCs compared to control NSCs and NHAs by GSEA analysis. Each column represents log 2 fold-change values (compared to averaged values across NSCs/NHAs) from a distinct cell isolate after averaging triplicate CPM values. Ion channels for which average GSC CPM values were <1 were not included, and individual CPM values of 0 were replaced with 0.01. C. CPM values for six of the most differentially enriched genes shown in panel (B). Bars, mean ± SEM. Mann-Whitney test; *p<0.05, **p<0.01, ns = not significant. D. Real-time qPCR analysis of a selected number of IGCs in several GSC isolates. C T values were normalized to ACTB (β-Actin) C T values; ddC T values relative to NSC-CB660 are shown. Bars, mean ± stdev. N = 3. GSC, Glioblastoma stem-like cells; NSC, human fetal neural stem cells including c-myc immortalized from cortex (NSC-CX) and brainstem (NSC-VM); NHA, normal human astrocytes including RasV12 transformed (NHA-RAS).
By contrast, ion transporters, which also contribute to cancer progression in multiple cell types [ 23 ], were infrequently enriched in GSCs. Of the 152 ion transporters assessed, only 1 (~0.7%) met our criteria to be considered differentially enriched (average CPM values ≥1 and ≥2 log 2 fold change enriched in GSCs vs. NSCs/NHAs; data not shown). Gap junction proteins also modulate electrical properties of cells and have a recently demonstrated role in promoting malignant phenotypes in GBM [ 24 ]. Gap junction-forming connexins, which were included in the overall set of 266 ion channel genes, were enriched in some GSC isolates ( S2 Fig ), but their low absolute expression levels failed to meet our cutoff for further analysis. In summary, when comparing GSCs to other neural cell types, differential expression of ion channel genes is markedly more prevalent than that of ion transporter and connexin genes. Based on these observations, our subsequent analyses focused on the potential importance of this selected cohort of 25 IGCs (see S4 Table for summary of IGCs).
IGC expression in normal neural cells
The therapeutic relevance of IGCs as drug targets in GBM may be limited by their expression in other normal tissues and neural cell types not included in the primary enrichment strategy described above. Therefore, we examined their expression in public databases of normal tissues and other neural cell types. Among the top IGCs of interest, the majority (16/25) were appreciably expressed in brain compared to other tissues (gtexportal.org; data not shown), 6/25 were more highly expressed in one or multiple other tissue types compared to brain, and 3/25 were not tissue-specific. To examine IGC expression in brain-specific cell subpopulations, we analyzed expression of IGCs in an RNA-seq database of isolated cortical human neural cell types, including astrocytes (fetal, adult, and reactive), neurons, oligodendrocytes, microglia, and endothelial cells [ 25 ] ( S3 Fig ). Out of our 25 previously identified IGCs, 1 was not available in this database, and 3 were expressed at levels below the arbitrary threshold used for selection of the IGC gene set (mean FPKM ≥1 in at least one cell type). Of the 21 remaining IGCs, 10 were specifically enriched in neurons ( GRIA2 , GRIA3 , GABRB3 , GABRA3 , KCNB1 , KCNA3 , SCN8A , GRIN2B , HCN1 , and GABRG3 ), 5 in astrocytes ( KCNJ16 , CLCN6 , TRPM3 , GRIK3 , and CNGA3 ), and 1 in microglia ( P2RX4 ), while the remaining 5 ( P2RX7 , KCNK10 , HCN3 , GABRQ , and CHRN2B ) were expressed across multiple cell types. As noted above, 3 of the genes in the IGC set of 25 ( KCNC3 , SCN11A , and GRIK4 ) were expressed at very low levels (mean subclass FPKM <1) in all normal neural cell classes ( S3 Fig , “Low abundance”), which may suggest that these IGCs are specific to cancerous cell types. This analysis identified non-cancerous tissue- and cell type-specific IGC expression patterns that could be important considerations for selecting IGCs as potential therapeutic targets for GBM. The varied patterns in IGC expression suggest that each IGC may possess a unique profile of systemic and neural toxicity that must be considered when targeting IGCs.
IGC expression is associated with GBM molecular classification
To better understand the potential clinical relevance of IGC expression, we characterized correlations between IGC expression and GBM molecular subtypes (Classical, Mesenchymal, Neural, or Proneural) [ 26 ], as well as driver genetic mutations in the GSCs. Molecular subtype classifications were previously assigned to GSC isolates ( S1 Table ) [ 27 ]. Unsupervised hierarchical clustering was performed based on CPM values of the entire set of 266 ion channels. Based on ion channel expression alone, GSC isolates of the same molecular subtype generally clustered together ( Fig 2A ), suggesting that ion channel expression patterns segregate with molecular subtypes.
10.1371/journal.pone.0172884.g002
Fig 2
GSC-enriched ion channels are associated with GBM molecular features.
A. Hierarchical clustering of GSCs by molecular subtype based on CPM values of all 266 ion channels. B. Correlation of IGCs with gene mutation signatures of critical GBM signaling pathways (see S1 Table for details). Bars, mean ± min/max, IQR. Two-tailed Mann-Whitney test; *p<0.05, **p<0.01.
We next investigated the association of IGCs with well-characterized gene mutations in three critical GBM signaling pathways: RTK/RAS/PI(3)K, p53, and RB [ 28 ]. Exome sequencing was carried out on GSCs to reveal gene mutations in these pathways ( S1 Table ), and direct associations between IGC expression and gene mutations were identified ( Fig 2B ). The majority of associations were between IGCs and mutations in the RTK/RAS/PI(3)K pathway ( EGFR , PI3KCA , NF1 , MET ), which can regulate proliferation and survival [ 28 ]. Fewer associations were observed in the RB pathway ( RB1 , CDKN2A/B ), responsible for regulating G1/S progression, and the p53 pathway ( TP53 , CDKN2A/B ), which regulates senescence and apoptosis. These preliminary findings suggest associations between IGC expression and clinically relevant GBM subtypes as well as potential functional interactions between specific IGCs and mutation-driven oncogenic signaling pathways that may prove to have prognostic or therapeutic value. However, further prospective studies with larger sample sizes and routine standardized molecular analyses are required for validation.
IGC expression predicts GBM patient survival
To test the prognostic significance of selected IGCs, we determined the correlation between their expression levels and patient survival using The Cancer Genome Atlas (TCGA) human GBM microarray database (n = 525; https://tcga-data.nci.nih.gov/tcga/ ). While the lack of paired normal samples precluded a determination of IGC enrichment in GBM versus normal brain tissue, analysis of IGC expression did reveal correlations with patient outcomes. Survival of patient cohorts was quantitated using Kaplan-Meier analysis and stratified by high (top 10%) or low (bottom 10%) IGC expression for the top 25 IGCs. Using this approach, four IGCs were found to have significant associations with survival. High expression of CNGA3 , TRPM3 , and P2RX4 was associated with significantly shorter median survival times, while high expression of GABRG3 predicted longer survival times ( Fig 3 ). The divergent associations between IGC expression and survival suggest the possibility that individual IGCs may function to either promote or inhibit malignancy. Since TCGA expression data derives from bulk GBM tissue samples, further studies are warranted in GSC models to determine whether associations between IGC expression and GBM malignancy can be attributed to GSC-specific function.
10.1371/journal.pone.0172884.g003
Fig 3
IGCs are associated with poor clinical outcomes.
Expression levels of IGCs were identified from 525 TCGA bulk GBM microarray expression samples. Samples for which expression levels were highest and lowest (top and bottom 10%) were then compared for time to death. Median days to death for each group reported on graph.
IGCs are enriched in distinct GBM tumor regions
To determine whether IGCs are regionally expressed in unique histological domains of GBM tumors with functional and clinical relevance, we examined IGC expression patterns in the Ivy Glioblastoma Atlas Project (Ivy GAP) database [ 29 ]. The Ivy GAP database dissects specific anatomic tumor regions that include the leading edge (LE) (with few tumor cells), infiltrating tumor (IT), cellular (central solid) tumor (CT), necrotic zones (PAN/PZ), and vascular regions (HBV/MVP). Overall, 13 of 25 IGCs were either not expressed or detected at low levels (<1 FPKM across all anatomic regions). When considering relative differences in expression by location, the greatest number of IGCs (18/25) were enriched in the LE and IT compartments compared to other regions ( Fig 4A ; ≥2 fold-change of mean LE/IT vs. mean all other regions). This was reflective of a larger trend, whereby half of all ion channels (131/273) were ≥2-fold higher at the tumor edge (LE/IT). Since tumor cell densities are negligible in LE and low in IT, IGC enrichment in LE/IT likely reflects IGC expression in normal neural cell types (neurons, microglia, etc.), which were not included in the original RNA-seq screen to identify IGCs.
10.1371/journal.pone.0172884.g004
Fig 4
IGCs are enriched in distinct GBM tumor regions.
A. Heat map of expression levels of the top 25 IGCs in various GBM tumor regions from the Ivy GAP RNA-seq database. B. IGCs associated with particular molecular subtypes in CT samples. Colored headers indicate predominant subtype associated with that ion channel. Bars, mean ± SEM. Kruskal-Wallis test across molecular subtypes. C. Survival curve for CNGA3 within CT samples stratified high/low by median FPKM value (3.908). Log-rank (Mantel-Cox) test. Median days to death for each group reported on graph. Leading Edge (LE), Infiltrating Tumor (IT), Cellular Tumor (CT), Perinecrotic zone (PZ), Pseudopalisading cells around necrosis (PAN), Hyperplastic blood vessels in cellular tumor (HBV), Microvascular Proliferation (MVP).
We reasoned that increased absolute expression levels in different histological regions of GBM could also identify IGCs with potential clinical and functional relevance. Overall, 9/25 IGCs were consistently expressed at ≥5 FPKM in at least one anatomic region ( Fig 4A ). Nearly all of the samples from the LE/IT and CT regions demonstrated this level of expression for the nine IGCs, while lower and more variable expression was observed in samples from necrotic (PZ/PAN) and vascular (HBV/MVP) features. For example, in the HBV/MVP vascular regions, five of these nine IGCs ( GRIA2 , KCNJ16 , GRIA3 , P2RX7 , and P2RX4 ) were expressed at appreciable levels (≥5 FPKM) in a majority of samples. IGC expression in the vascular compartment is consistent with the known role of GBM vasculature to provide a supportive niche for GSCs [ 30 , 31 ]. However, since GSCs are a minority cell population and IGC expression may overlap with other cell types (see S3 Fig ), the specificity of IGC expression for GSC localization requires additional detailed studies.
We next examined how IGC expression levels in the Ivy GAP database correlated with molecular subtypes and survival. We chose to study this in the solid CT region only, since this region forms the bulk of GBM tumors and comprised the largest number of samples for any of the anatomic subsets. Among all 25 IGCs studied that were expressed at appreciable levels (average FPKM across CT >1), five were significantly associated with particular molecular subtypes in CT ( P2RX4 , GRIA3 , GABRA3 , CHRNB2 , and GABRB3 ; Fig 4B ). We also examined whether IGC expression within the CT region was associated with prolonged or reduced survival in these patients. Kaplan-Meier plots were generated for IGCs with median CT FPKM levels >1 (9/25) that compared survival times in patient samples with high (above median FPKM) or low (below median FPKM) IGC levels. Samples with low levels of CNGA3 were associated with significantly reduced survival rates ( Fig 4C ). Significant differences in survival rates were not observed between molecular subtypes overall (data not shown), indicating that the survival association with CNGA3 is not an artifact of the disproportionate number of subtype-specific samples with high CNGA3 expression.
Pharmacological blockade of ion channels restricts GSC viability
Thus far, we have identified individual GSC-enriched ion channels, which are members of larger ion channel families with overlapping functions. However, the majority of available drugs target the broader functionality of these families rather than specific channels. Therefore, to explore the potential therapeutic relevance of targeting IGCs, we assessed GSC enrichment of classes of functionally related ion channels. We reasoned that this approach would mitigate the potential limitations imposed by ion channel redundancy. We found that 12 out of 22 ion channel families were enriched more than 2-fold in GSCs compared to NSC/NHA controls, and all 12 of these families also contained a high proportion of IGCs (at least 1/5 of members enriched >2-fold) ( Fig 5A , orange points). The most highly enriched families (>5-fold) were epithelial Na + channels and GABA A receptors. Fold-change values, however, are likely to be skewed by contributions from rare, yet highly enriched ion channels. Therefore, it is also important to consider the proportion of family members enriched, as we have here; the ionotropic glutamate receptor family had the highest proportion of enriched members (11/18). We also examined regional patterns of ion channel family expression from the Ivy GAP dataset ( Fig 5B ; S4 Fig ). A large proportion of IP3 receptors, Ca 2+ -activated K + channels, and inwardly-rectifying K + channels were highly enriched across all tumor regions. Once again, several ion channel classes were highly enriched at the tumor leading edge with decreasing levels of expression towards HBV/MVP areas.
10.1371/journal.pone.0172884.g005
Fig 5
Ion channel blockade reduces GSC viability.
A. Summary of functionally-related ion channel families enriched in GSCs (orange) compared to controls. Y-axis, mean log 2 fold change (GSC vs. NSC/NHA) values for each ion channel family; x-axis, proportion of ion channel family members > 2-fold change (GSC vs. NSC/NHA); bubble size corresponds to average GSC CPM value. Genes were excluded if average NSC/NHA values were zero. Families were considered if they contained more than one member. B. Proportion of ion channel family members expressed at >1 FPKM within distinct GBM regional compartments as revealed by Ivy GAP analysis (only families with at least three members are shown). C. Antagonists for top IGC families and selected specific ion channel blockers were applied to GSC-0827, GSC-0131, and NSC-CB660 in media at indicated concentrations. MTT viability assay was performed at 72 hours. Fluorescence arbitrary units (AU) were averaged across triplicates and normalized to control media conditions (dashed line at 1.0). Bars, mean ± SEM. N = 3–4. Repeated measures two-way ANOVA with Dunnett’s multiple comparison test compared to NSC-CB660. n.s., not significant.
To test whether IGC-related families are functionally relevant to the malignant properties of GSCs, we examined whether ion channel blockers inhibit GSC growth in vitro. Pharmacological blockade has advantages over genetic knockdown/loss-of-function studies, since the action of similarly functioning ion channels can be blocked while avoiding compensation by alternative channels [ 32 ]. An MTT viability assay was performed for large-scale, rapid viability assessment. Several drugs were tested based on the enrichment of their channel targets in the GSC RNA-seq dataset ( S4 Table ; see Materials and methods for drug details). GSC lines -0827 and -0131 were selected for in vitro pharmacological evaluation of viability due to their rapid growth rates in vitro, tumorigenic capacity in vivo [ 33 ], and extensive molecular characterization [ 27 ]. Compounds were added at increasing concentrations to cell isolates, and viability was measured 72 hours later. Increasing doses of TTX, TEA, 4-AP, CPP, CNQX, ω-Conotoxin MVIIC, and CdCl 2 dramatically reduced cell viability across all lines, while K + enhanced viability in a dose-dependent manner ( Fig 5C ). GSC lines -0827 and -0131 were more sensitive to these effects than NSC-CB660 for all compounds tested. Neither gabazine nor picrotoxin (GABA A receptor blockers) affected GSC or NSC viability. These results demonstrate the functional relevance of K + channels, voltage-gated Na + channels, voltage-gated Ca 2+ channels, voltage-gated Cl - channels, and ionotropic glutamate receptors to GSC viability as predicted by GSC enrichment of IGC families or individual members.
SCN8A , KCNB1 , or GRIA3 knockdown reduces GSC viability
To establish the potential functional importance of specific IGCs, as opposed to broader classes of ion channels as tested above, we quantified the effects of siRNA-mediated gene expression knockdown of three IGCs, SCN8A , KCNB1 , and GRIA3 . These IGCs were selected because they were among the most highly implicated in GSC enrichment and blocking their activity with broadly-acting drugs reduced GSC viability. Three unique siRNAs for each IGC were transiently transfected into GSC lines that had high expression of the corresponding IGC ( SCN8A , GSC-0827; KCNB1 , GSC-G19; GRIA3 , GSC-0827). Robust knockdown was achieved for all nine siRNA candidates ( Fig 6A ). The most effective siRNA candidate for each IGC, along with a scrambled siRNA control (siScr), was tested for its effects on viability, and all siRNAs demonstrated a dose response for GSC growth inhibition 72 hours after transfection ( Fig 6B ). At the highest dose (20 pmol per well), growth of each GSC line was inhibited 55–62% compared to siScr controls. By comparison, the NSC line CB660, which had low baseline levels of SCN8A , KCNB1 , and GRIA3 by RT-qPCR (data not shown), was inhibited only 0–16% at the highest siRNA dose. These results suggest that in addition to the broad pharmacological targeting of ion channel classes in GSCs, inhibition of specific IGCs may also have therapeutic relevance.
10.1371/journal.pone.0172884.g006
Fig 6
siRNA-mediated knockdown of selected IGCs impairs GSC viability.
A. Real-time qPCR analysis of IGC expression levels 24 hours after transfection of scrambled negative control siRNA (siScr) and individual siRNA candidates (1, 2, 3) for each IGC. C T values were normalized to ACTB (β-Actin) C T values; ddC T from siScr. Bars, mean ± stdev. N = 3. One-way ANOVA with Dunnet’s post-test compared to siScr, *p<0.05, **p<0.01, ***p<0.001, ****p<0.0001. B. CellTiter-Glo viability assay 72 hours following siRNA-mediated knockdown of IGCs. GSC viability decreased in response to increasing amounts of siRNA (2, 5, and 20 pmol). Fluorescence arbitrary units (AU) were normalized to siScr levels. Bars, mean ± SEM. N = 3–4. Two-way ANOVA with repeated measures; SCN8A : cell type p = 0.0001, dose p<0.0001, interaction p<0.0001; KCNB1 : cell type p = 0.0042, dose p<0.0001, interaction p<0.0001; GRIA3 : cell type p = 0.0011, dose p<0.0001, interaction p<0.0001.
Discussion
The inevitable treatment failures of GBM, the most malignant and common adult human brain tumor, are largely established by the phenotypes of its resident cancer stem cells or GSCs. By virtue of their unique propensities for self-renewal, multipotent differentiation, invasiveness, and treatment resistance, GSCs drive GBM formation and progression [ 4 ]. Therefore, a better understanding of the mechanisms that regulate GSC intrinsic physiology is expected to yield clinical benefits and is lacking in our current understanding. Ion channels are increasingly recognized as regulators of the malignant phenotypes of cancer cells [ 9 ], but have not yet been leveraged as therapeutic targets in GSCs. This is due in large part to an incomplete understanding of their GSC-specific expression patterns compared with non-cancerous neural cells. Here, we identified GSC-enriched ion channels that correlated with molecular and clinical features of GBM. Furthermore, pharmacologic blockade or genetic knockdown of IGCs differentially inhibited GSC growth compared to that of normal NSCs. Together, these data strongly support potential functional and therapeutic roles for IGCs in GBM.
This study identified a set of 25 ion channels that were highly expressed by GSCs compared to normal neural cells and were representative of many different ion channel subfamilies. This was consistent with a recent RNA-seq study that reported a collection of 18 ion channel genes identified as a molecular signature of glioma [ 34 ]. These genes were dysregulated in high grade glioma, associated with poorer survival, and representative of many different types of ion channel classes; however, there was little overlap of individual ion channels between their study and ours, likely because our study focused on ion channels unique to glioma stem cells rather than bulk tumor cells. When our study results are compiled ( S5 Fig ), several IGC candidates emerge as consistently implicated in contributing to GSC malignancy, among which are SCN8A , KCNB1 , and GRIA3 . One of the most reliably involved IGCs in our study is SCN8A (Nav1.6), a voltage-gated Na + channel. Voltage-gated Na + channel isoforms are aberrantly expressed in cancer cells contributing to metastatic behaviors [ 35 , 36 ], and Na + channel mutations have been estimated in at least 90% of GBM samples [ 19 ], suggesting that they play a role in GBM malignancy. Several classes of K + channels were also implicated in this study, including inwardly-rectifying, voltage-gated, and two-pore-domain K + channels. These classes have been well-studied in cancer [ 37 ] and tied to GBM malignancy [ 38 ]. One candidate of particular interest is the voltage-gated K + channel KCNB1 (Kv2.1), which was expressed at high levels in GSCs and contributed to reduced GSC viability when knocked-down. mRNA knockdown of all three of these IGCs, SCN8A , KCNB1 , and GRIA3 , reduced GSC viability in our study, suggesting their importance to GSC malignancy and therapeutic potential.
One of the goals of this work was to identify ion channels that are uniquely expressed by GSCs and avoid those expressed by other cell types; the rationale of this strategy was that increased selectivity of expression in GSCs might translate to less toxic off-target effects in the therapeutic clinical setting. Along with NSCs/NHAs, we also observed differences in GSC gene expression compared to bulk GBM samples and other normal neural cell types. However, we found that many IGCs were expressed in several other neural cell populations. Across ion channel families, the majority of ionotropic glutamate and glycine receptors were enriched in both GSCs and bulk GBM samples, while 5-HT3 receptors, TRP channels, CatSper/two-pore channels, and ryanodine receptors were consistently low in both populations. GSC-specific families included GABA A receptors, epithelial Na + , voltage-gated Ca 2+ , and two-pore domain K + channels. At the individual level, several IGCs that were highly enriched in GSCs had very low or absent expression in bulk GBM tumor samples (e.g. GRIK4 , SCN8A , KCNC3 , ASIC3 , and HCN3 ), suggesting that these candidates may be the most therapeutically tractable for specifically targeting GSCs within the tumor. Alternatively, the low expression may indicate that GSCs grown in vitro may not reflect true expression levels in vivo or that IGC expression may be obscured by the low frequency of GSCs in bulk tumor. Several of these candidates, KCNC3 , SCN11A , and GRIK4 , also failed to show appreciable expression in any normal neural cell populations ( S3 Fig ), which may be advantageous for precise clinical targeting of GSCs.
IGCs were also associated with distinct clinical features in this study, including GBM subtype and critical oncogenic genomic mutations. Across all 266 ion channels, there was a trend for GSCs to stratify by molecular subtype. This suggests that ion channels generally correlate with clinical features, although batch effects in GSC source origination could account for some of these effects. In line with this, IGC expression was also associated with altered clinical outcomes. This is consistent with other reports; for instance, Na + channel mutations in GBM tumors have been correlated to poorer survival outcomes [ 19 ]. When considering the extremes of expression in the TCGA database, higher IGC expression of three genes ( CNGA3 , TRPM3 , and P2RX4 ) correlated with decreased survival while higher expression of one ( GABRG3 ) correlated with increased survival. By contrast, when considering expression restricted to the solid tumor (CT) region in the Ivy GAP database, high CNGA3 was associated with increased, rather than decreased, survival as noted in TCGA. This discrepancy may reflect the complexity of ion channel functions in tumor and tumor-associated stromal cells, as well as differences in sample composition between databases and bias introduced by restricting the Ivy GAP analysis to the CT region.
A challenging problem of solid tumor biology is understanding the regional and cell type-specific variations in expression and function. We observed interesting patterns in IGC expression in GBM anatomical regions within the Ivy GAP database. Although we expected to find IGCs in the putative stem cell niche associated with the vasculature, many IGCs were highly expressed at the leading edge of the tumor instead (including many additional ion channels not shown). This likely reflects contamination from surrounding normal neural cells expressing these ICs but could be due to the presence of stem cells residing at the tumor edge and/or ion channels that are preferentially upregulated at the tumor edge for communication with the normal brain surround. Much more work will be needed to understand the regional and cell-type heterogeneity of IGCs within the tumor.
One of the major findings from this work was that both pharmacological blockade and RNA knockdown of IGCs inhibited GSC growth, which may implicate new pathways for therapeutic targeting of GBM. There is a clear and dire need for novel molecular targets in this arena. Ion channel and transporter targeting drugs are currently used to treat a variety of clinical conditions and represented over 13% of FDA-approved drugs in 2006—the second largest class of existing drugs [ 39 ]. A variety of ion channel blockers have been used to target various cancers in pre-clinical animal models [ 13 , 14 ]. Compellingly, anti-epileptic drugs that target voltage-gated Na + channels inhibit metastatic behaviors in several cancer cell types [ 36 ], suggesting that these channels may be practical therapeutic targets in GBM. Clinically, targeting ion channels within brain tumors has many challenges associated with it. First, there is considerable functional redundancy among ion channel classes, which may mean that broad channel antagonists are needed to meaningfully impact malignant phenotypes. However, the lack of specificity of these drugs may result in profound off-target effects and deleterious side effects. Ion channels also play many crucial functions in normal surrounding neural cells, and so methods for precise targeting of ion channel-harboring brain tumor cells is needed. We have made attempts in this study to distinguish ion channels enriched specifically in GSCs compared to normal neural cell types, yet many identified IGCs are expressed in normal neural cell types. Despite these caveats, our results indicate that GBM research may be poised for novel ion channel drug discovery, and our findings offer a tractable starting point.
There were several limitations of this study. First, we compared ion channels expressed in GSCs to NSC and NHA lines in vitro to understand GSC-specific ion channel expression. While this is a good foundation for understanding GSC-specific enrichment, many control neural cell types were not included in this analysis. Lowly abundant genes were also ignored in our analyses, however it is not known how absolute CPM expression levels translate to biological processes, and it is conceivable that low CPM levels are still meaningful. We also examined expression patterns in isolated GSC lines maintained over several passages, however, GSCs normally interact with other cell types in the tumor microenvironment, which are likely to influence ion channel expression and function. Ion channel dysregulation was also monitored at the mRNA level, however, other cellular processes could modulate ion channel function and contribute to GSC malignancy. Two well-studied examples of dysregulated, post-translational cellular processes in glioma include the mislocalization of ion channels [ 40 ] and changes to ion channel sensitivity [ 41 ]. Epigenetic dysregulation is likely to play a role in glioma malignancy as well, since aberrant DNA methylation has been linked to multiple cancers [ 42 , 43 ]. Additionally, ion channel expression can oscillate with phases of the cell cycle, and bulk sequencing would miss these dynamic expression changes. Finally, the collection of GSC isolates studied here may not capture the full spectrum of GBM and GSC heterogeneity, which could impact IGC profiles and their associations with specific GBM subtypes. Therefore, building off of our current paradigm, we propose that future studies should aim to increase GSC and normal neural sample complexity. Although challenging, this could be accomplished in time through consolidation of existing RNA-seq databases or acquisition and analysis of additional samples.
IGCs may regulate GSC malignancy through several potential mechanisms. Several studies have shown that ion channels can regulate cell cycle dynamics [ 10 ], migration [ 11 ], apoptosis [ 12 ], and vascularization [ 44 ] contributing to cancer progression. The expression levels of some ion channels are regulated in tune with the cell cycle [ 45 ], and voltage-gated K + channels, in particular, are known to exhibit cell-cycle-dependent fluctuations in expression or activity in non-cancerous [ 46 ] and cancerous [ 45 ] cells, which contribute to cell cycle checkpoint regulation. This supports our finding that increasing concentrations of extracellular K + increase GSC and NSC viability, presumably through enhanced proliferation. Furthermore, ion channel blockers can inhibit cell cycle progression, arrest aberrantly cycling cancer cells, and inhibit tumor formation [ 9 , 45 ]. Ion channels can also regulate the resting membrane potential (V m ) of cells, a process that is deregulated in cancer cells [ 47 , 48 ]. V m is regulated with the cell cycle in non-cancerous and cancerous cells, and experimental reversal of V m at these stages can arrest or stimulate the cell cycle [ 49 , 50 ]. Thus, the aberrant ion channel dysregulation in GBM and GSCs observed in this study could modulate V m contributing to tumor progression and malignancy, and these electrophysiological changes should be examined in future studies.
There is a substantial lack of understanding of how GBM functions within the context of the normal brain environment. We propose that the neural environment and electrophysiological mechanisms play a vital role in GBM oncogenesis and maintenance. Ion channels play a crucial role in these mechanisms in multiple cancers, and increasing evidence suggests that they contribute to malignancy in GBM as well, specifically within the most malignant tumor subpopulation, the GSCs. Future studies will need to parse out the role that these ion channels play in electrophysiological interactions with the surrounding neural environment. We have identified ion channel candidates that are enriched in GSCs and are linked to GBM prognosis. Many questions still need to be addressed to understand the contribution of ion channels to GSC biology. Although connexins and transporters did not play a significant role in these findings, other molecular players, such as metabotropic receptors, should be examined in this context. Additionally, it will be critical to understand how these channels contribute to malignancy and the downstream pathways involved, particularly V m -related and non-current passing mechanisms. Furthermore, ion channels are known to interact with migratory, angiogenic, and apoptotic factors, which should be explored further in this setting. While choosing candidate ion channels to target based solely on expression data is informative, it will be vital to carry out functional screening assays in the future to find meaningful functional outcomes. Nevertheless, the broad enrichment of ion channel types supports the notion that electrical activity within the tumor microenvironment may regulate GBM malignancy, and much more research is needed to understand the potential reciprocal interactions between GSCs and the neural surround. Despite the study limitations described, our findings offer a starting point for exploring hypotheses of novel ion channel-based drug targeting of GSCs in vitro and in preclinical models.
Materials and methods
GSC culture
Human GSCs were previously isolated from resected stage IV glioma tumors ( S1 Table ) [ 33 , 51 – 56 ]. Non-tumor neural cell lines included: human fetal cortical neural stem cells (NSC-CB660), v-myc immortalized brainstem NSCs (NSC-VM; ReNcell, EMD Millipore), c-myc immortalized cortical NSCs (NSC-CX; ReNcell, EMD Millipore), normal human astrocytes (NHA; StemCell Technologies), and Ras-V12 infected NHAs (NHA-RAS) [ 57 ] ( S1 Table ). For RT-qPCR, MTT viability, and siRNA knockdown assays, cells were newly thawed from frozen stocks and maintained in culture over multiple passages using previously published protocols [ 3 , 55 ]. Cells were grown as adherent cultures on flasks coated with Natural Mouse Laminin (10 ng/ml; Thermo Fisher Scientific, #23017–015) in Human NeuroCult NS-A Proliferation Kit (StemCell Technologies, #05751), Heparin sodium salt (2 mg/ml, Sigma-Aldrich, #H3149), Antibiotic-Antimycotic (Thermo Fisher Scientific, #15240–062; 10,000 units/ml of penicillin, 10,000 μg/ml of streptomycin, and 25 μg/ml of Fungizone Antimycotic), supplemented with human recombinant EGF (10 ng/μl; Peprotech, #AF-100-15) and bFGF (10 ng/μl, Stemgent, #03–0002). GSCs were maintained at low passage number and passaged at 80–90% confluence (approximately every 3–4 days) with StemPro Accutase Cell Dissociation Reagent (Thermo Fisher Scientific, #A11105-01); cells were seeded at 4–5×10 5 cells per 25 cm 2 flask. To assess overall similarity between samples, cell isolates were authenticated by exome or RNA sequencing analysis, which included non-supervised clustering, principle component analysis, and differential expression analysis using DESeq2 [ 58 ].
RNA sequencing and molecular subtyping
RNA sequencing was carried out as previously described [ 27 ]. Briefly, sequencing was performed using an Illumina HiSeq 2000 in Rapid Run mode and employed a paired-end, 50 base read length (PE50) sequencing strategy. RNA-seq reads were aligned to the UCSC hg19 assembly using Tophat2 [ 59 ] and counted for gene associations against the UCSC genes database with HTSeq [ 60 ]. All data was combined and normalized using a trimmed mean of M-values (TMM) method from the R package, edgeR [ 61 ]. Sequencing data can be accessed at Sequence Read Archive SRP092795 and NCBI Gene Expression Omnibus under GSE89623 . Molecular classifications were determined according to a previous report [ 27 ]. For downstream analyses, replicates (2–3 per cell isolate) were averaged to assign a single CPM value to each sample. GSEA analysis [ 62 , 63 ] was performed using a Signal2Noise metric for ranking genes. GSC isolates were classified by molecular subtype according to gene expression signatures produced by The Cancer Genome Atlas [ 26 , 64 ]. Our isolates were clustered using 770 of these genes using a Manhattan distance complete-linkage method, and centroids were computed as the median expression of each gene across the core TCGA samples [ 26 ]. Each GSC sample replicate was compared against the centroids using Single Sample Predictor (SSP) method [ 65 ]. In addition, samples were assigned to GBM subtypes by maximizing the Spearman rank based correlation between expression of new samples and GBM subtype centroids. Each replicate was assigned separately and then the consensus was used to assign a final classification. For hierarchical clustering, the clustergram function in the bioinformatics toolbox of MATLAB (v. R2015b, MathWorks) was run using the euclidean distance metric and unweighted average distance linkage method.
Exome sequencing
Exome sequencing and preprocessing were performed at the Genome Core Facility of Mount Sinai School of Medicine. Whole genome amplified was used for exome sequencing. Whole-exome capture libraries were constructed using ligation of Illumina adaptors. Each captured library was then loaded onto the HiSeq 2500 sequencing platform. Exome sequence preprocessing and analysis were performed using standard pipelines recommended by the Genome Analysis Toolkit (GATK) [ 66 ]. Three GSC cell lines were aligned independently. For each sample, the reads were aligned to NCBI build 37 (hg19) human reference sequence using BWA ( http://bio-bwa.sourceforge.net ) [ 67 ], and duplicated were marked using Picard ( http://broadinstitute.github.io/picard/ ). Local realignment around indels and base recalibration process were performed ending in an analysis-ready BAM file for each cell line. Mutation detection and annotation were performed at the Genome Core Facility of Mount Sinai School of Medicine as follows. For each sample, GATK was used to detect all variants that differed from a reference genome. Variants identified were annotated using the snpEff software [ 68 ]. The variants were filtered in four steps according to a previous study [ 69 ]. First, the variants with low allelic fraction were excluded. The allelic fraction was calculated for each detected variant per cell line as a fraction of reads that supported an alternative allele (e.g. different from the reference) among reads overlapping the position. Only reads with allelic fractions above 0.25 were used in the downstream analysis. Additionally, the variants that were detected as common germline variants were excluded. Variants for which the global allele frequency (GAF) in dbSNP138 or allele frequency in the NHLBI Exome Sequencing Project ( http://evs.gs.washington.edu/EVS , data release ESP2500) was higher than 0.1% were excluded from further analysis. Furthermore, variants detected in a panel of 278 whole exomes sequenced at the Broad Institute as part of the 1000 Genomes Project were excluded from further analysis. Finally, the variants with low quality (e.g. insufficient read depth and insufficient genotype quality) were filtered with the variant quality score tools. We selected high-confident mutations by their annotation obtained from snpEff. We filtered silent mutations and extracted high and moderate impact of mutations, including non-synonymous, nonsense, frame shift, and codon insertion/deletion mutations. Exome sequencing data can be accessed at NCBI Sequence Read Archive SRP09879 under BioProject PRJNA369688 .
TCGA and Ivy GAP GBM databases
TCGA level 3 GBM data for U133A microarrays (539 samples), along with the corresponding clinical data, was downloaded from the TCGA Data Portal ( https://tcga-data.nci.nih.gov/tcga/ ). The data was imported into an R environment. Samples for which there was no clinical data relating to time of death were excluded resulting in 525 total samples. Expression levels for the gene of interest were then pulled, and samples were sorted low to high by expression level; the top and bottom expressing samples were identified using the quantile function (0.10 and 0.90, respectively). Kaplan-Meier curves were generated using the survfit function in the survival library of R comparing the low and high cohorts. For Ivy GAP analyses, FPKM values and sample information were downloaded from the Ivy GAP website (glioblastoma.alleninstitute.org). FPKM values were averaged across replicates within regions to generate single values for each sample within a particular region.
RNA isolation and real-time qPCR
RNA was extracted from cell cultures every two weeks during the exponential growth phase; RNA was isolated from 0.5–1.0×10 6 cells. Total RNA was prepared using RNeasy Mini Kit with DNase I (Qiagen). RNA concentration and quality (A260/A280) was measured using a NanoDrop 1000 Spectrophotometer (Thermo Scientific). cDNA was synthesized from 2 μg of RNA using iScript Reverse Transcription Supermix for RT-qPCR (Bio-Rad). Quantitative real-time PCR was performed with iTaq Universal SYBR Green Supermix (Bio-Rad) on an Applied Biosystems 7300 Real Time PCR System. Reactions were performed in triplicate and values normalized to ACTB (β-Actin). RT-qPCR primer sequences were designed to span exon-exon boundaries and are listed in S3 Table .
Drugs
Drugs used for this study included: tetrodotoxin (TTX, voltage-gated Na + channel blocker, Tocris, #1078), tetraethylammonium hydroxide (TEA, non-inactivating K + channel blocker, Sigma-Aldrich, #T6393), 4-Aminopyridine (4-AP, transient/A-type K + channel blocker, Sigma-Aldrich, #275875), 3-((±)2-carboxypiperazin-4yl)propyl-1-phosphate (CPP, NMDA receptor antagonist, Tocris, #0173), 6-Cyano-7-nitroquinoxaline-2,3-dione (CNQX, AMPA/kainate receptor antagonist, Alomone Labs, #C-140, diluted in DMSO), ω-Conotoxin MVIIC (N-, P/Q-type Ca 2+ channel blocker, Alomone Labs, #C-150), cadmium chloride (CdCl 2 , non-selective voltage-gated Ca 2+ channel blocker, Sigma-Aldrich, #C3141), Gabazine (GABA A receptor blocker, Tocris, #1262), Picrotoxin (PTX, GABA A receptor blocker, Sigma-Aldrich, #P1675, diluted in DMSO). KCl was added at indicated concentrations to growth media, which had undisclosed levels of K + .
MTT viability assay
Cells were harvested and seeded on laminin-coated 96-well plates at a density of 10 4 cells per well. The following day, cells were attached and compounds were added to the media at indicated concentrations. Untreated conditions received DMSO or water, as appropriate. Seventy-two hours after compounds were added, the Vybrant MTT Cell Proliferation Assay Kit (Thermo Fisher Scientific, #V13154) was performed according to manufacturer’s instructions. 10 μl per well of 12 mM stock MTT compound (dissolved in sterile PBS) was added to wells and incubated for 4 hours at 37°C. Viable cells reduced MTT into purple formazan crystals, which was solubilized in a solution of SDS-0.1 M HCl added to wells for 4 hours at 37°C. Absorbance at 570 nm was read at 37°C using a SpectraMax 190 Gemini Microplate Reader. Each condition was run in triplicate. Fluorescence arbitrary units (AU) were subtracted from background levels, averaged across triplicate wells, and normalized to untreated wells.
siRNA-mediated knockdown
Human DsiRNA kits (Integrated DNA Technologies) consisting of three predesigned DsiRNAs against selected IGCs ( GRIA3 , KCNB1 , and SCN8A ) were selected along with a Scrambled Negative Control DsiRNA. Cells were plated on laminin-coated tissue culture plates for CellTiter-Glo Luminescent Cell Viability (Promega; 96-well) or RT-qPCR (6-well) assays 24 hours before transfection. Cells were 30–40% confluent at the time of transfection. For RT-qPCR experiments, 100 pmol RNA was transfected per well, and 2, 5, or 20 pmol RNA was transfected per well for the CellTiter-Glo assay. Antibiotics and antimycotics were removed from media, and transfection was carried out with Lipofectamine RNAiMAX Transfection Reagent (Thermo Fisher Scientific) and Opti-MEM Reduced Serum Medium (Thermo Fisher Scientific) according to the manufacturer’s instructions. Negative controls for all experiments included non-transfection wells (not shown) and wells transfected with scrambled DsiRNA. RNA isolation and RT-qPCR was carried out 24 hours after transfection, and the CellTiter-Glo assay was performed 72 hours after transfection according to methods outlined by the manufacturer.
Statistics
Statistics were performed using GraphPad Prism 6, R version 3.2.3, or MATLAB (R2015b, MathWorks). Comparisons between two groups were tested for significance with the Mann-Whitney test. For multiple comparisons, the Kruskal-Wallis test was used to compare across molecular subtypes. Kaplan-Meier plots were created using the R package “survival” with the survfit function, and the log-rank (Mantel-Cox) test was used to test for significance of the Ivy GAP data. Two-way ANOVA with repeated measures was used for data with multiple drug concentrations (MTT and CellTiter-Glo assays); runs with missing values were excluded from statistical testing but included on plots.
Supporting information
S1 Table
Table of cell isolates used in this study.
Only gene mutations in RTK/RAS/PI(3)K, p53, and RB pathways in which the mutation was present in at least three samples were included.
(EPS)
S2 Table
RNA sequencing CPM values for all genes and ion channel gene set.
Ion channel gene list derived from guidetopharmacology.org [ 22 ].
(XLSX)
S3 Table
Real-time qPCR primer sequences.
(EPS)
S4 Table
GSC-enriched ion channel details.
Top 25 IGCs identified by differential enrichment analysis grouped by functional relatedness.
(EPS)
S1 Fig
Schematic summarizing study’s design.
(EPS)
S2 Fig
Gap junction protein expression values.
Heat map (unranked) of log 2 fold change values of all connexins and pannexins in GSCs compared to control NSCs/NHAs. Each column represents log 2 fold change values (compared to averaged values across NSCs/NHAs) from a distinct cell isolate after averaging triplicate CPM values.
(EPS)
S3 Fig
Expression pattern of IGCs in non-tumor neural cell types.
Expression levels of top 25 IGC candidates were examined in several human non-tumor neural cell types from RNA sequencing data described by Zhang et al. [ 25 ] and downloaded from http://www.brainrnaseq.org . IGCs were specifically enriched in neurons, astrocytes, microglia, or across multiple classes. Three IGCs are listed that had low abundance (mean expression in at least one cell type was not ≥1 FPKM). Bars, mean FPKM values ± SEM. GBM-A, GBM/peri-tumor astrocytes; SH-A, sclerotic hippocampi astrocytes; F-A, fetal astrocytes; M-A, mature astrocytes; N, neurons; O, oligodendrocytes; M, microglia; E, endothelial cells; WC, whole cortex.
(EPS)
S4 Fig
Expression heat map of ion channel families by tumor region.
FPKM values of ion channels partitioned by ion channel family and GBM tumor region (Ivy GAP RNA-seq database).
(EPS)
S5 Fig
Summary of results relating to IGCs reported in this study.
Top 25 IGCs and associated analyses. R/R/P, RTK/RAS/PI(3)K pathway; TP53, TP53 pathway; RB, RB pathway.
(EPS)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Healthcare decision-making for persons with multiple chronic conditions (MCCs) is difficult [ 1 – 5 ]. The focus on managing individual conditions fails to account for interactions among multiple conditions and their treatments, leading to uncertain benefit and potential harm [ 3 – 6 ]. Evidence to guide care is often lacking because individuals with MCCs are excluded from most clinical trials [ 7 – 8 ]. Even trials that include these individuals address disease-specific outcomes or survival, not always the outcomes most valued by older adults with MCCs [ 2 ]. The number and complexity of patient-related activities such as medications, testing, health visits, and self-monitoring tasks, are increasingly burdensome [ 2 , 9 – 12 ]. Older adults with multiple chronic conditions, when faced with tradeoffs that require difficult choices, vary in their health outcome goals and in their preferences for the healthcare they are willing and able to receive [ 12 – 15 ].
There is consensus that healthcare should be, “ respectful of and responsive to individual patient preferences , needs , and values and ensuring that patient values guide all clinical decisions ” [ 16 ]. Communication strategies facilitate patient preferences—and priorities—based decision-making for persons with serious illness or near the end-of-life [ 17 – 20 ]. However, methods for ascertaining the health priorities of older adults with multiple conditions who are not near the end-of life remain lacking, as do reliable approaches for clinicians wishing to align decision-making and care with these priorities [ 1 , 2 , 12 , 13 ].
To fill these gaps, we launched the Patient Priorities Care initiative with input from clinicians, patients, caregivers, healthcare system representatives, health information technology and redesign experts, and payers [ 21 ]. This diverse group of stakeholders agreed that identifying and aligning decision-making with each patient’s health priorities was the approach that best addressed the uncertainty and treatment burden inherent in the care of multiple conditions while honoring the directive to ensure, “… that patient values guide all clinical decisions .” [ 15 , 21 ]. Based on this input, we designed a prototype, outlined in Fig 1 , that aligns decision-making and care with each patient’s health priorities, namely their health outcome goals and healthcare preferences (see Fig 1 for definitions) [ 21 , 22 ]. As described previously, Patient Priorities Care is a continuous process that begins when patients—and family members or friends when desired by patients—identify specific, actionable, realistic, and reliable health priorities facilitated by a member (e.g. nurse, advanced practice nurse, social worker) of the healthcare team ( Step 1 in Fig 1 ) [ 23 ]. These health priorities are transmitted to clinicians who use them in their communication and decision-making with patients and other clinicians ( Steps 2 and 3 ). We recently described the feasibility of implementing Patient Priorities Care in practice [ 24 ].
10.1371/journal.pone.0218249.g001
Fig 1
Steps in Patient Priorities Care.
Effective and feasible strategies for translating disease-specific into priorities-aligned care options and addressing the challenges inherent in patient health priorities aligned decision-making are essential to clinician participation in this approach. To help identify these strategies, we sought insights into how clinicians might link patients’ priorities to decision-making ( Steps 2 and 3 ). The first aim was to describe challenges the clinicians faced in trying to align clinical decisions with patients’ health priorities. The second aim was to identify strategies that emerged to help clinicians overcome challenges to providing patient priorities-aligned care.
Methods
Design
We used a qualitative participant observation design in which investigators observed, participated in, and interpreted discussions with clinicians [ 25 ].
Setting
This work involved the primary care and cardiology practices in Connecticut participating in development and testing of Patient Priorities Care [ 23 , 24 ].
Participants
The Patient Priorities implementation team included four geriatricians, a general internist expert in clinician training, an expert in behavioral medicine, the clinical champion from the primary care practice, two priorities facilitators (an Advanced Practice Nurse (APN) and a care manager employed by the primary care practice who helped patients identify their health priorities), and two project managers. The practicing clinicians who participated in the pilot included the ten primary care providers (three APNs, one physician assistant, and five physicians) who worked at the pilot primary care practice and the five cardiologists who provided most of the cardiology care for patients from the pilot primary care practice. All primary care clinicians and cardiologists providing care at these practices participated in the pilot. The primary care practice is the largest clinical site of a multi-site primary care group practice providing care to nearly 15% of people in Connecticut. We included cardiologists because they are responsible for a large amount of the specialty care received by older adults with MCCs. The selection and recruitment of the pilot practice and the participating clinicians were described previously [ 24 ]. The clinicians received modest stipends for participating in Patient Priorities Care.
The primary care clinicians and cardiologists were introduced to the concepts of Patient Priorities Care through an introductory webinar; they then participated in two case-based face-to-face training sessions in September and October 2016 [ 24 ]. During these training sessions they role played patient-clinician and clinician-clinician scenarios involving commonly encountered decisional issues for older adults with multiple chronic conditions. From November 2016-February 2017, geriatrician members of the Patient Priorities team (MT, CB, GO) met monthly with the PCPs and every other month with the cardiologists to discuss the workflow involved in Patient Priorities Care, challenges encountered, and decisions or changes in care that occurred from knowing patients’ health priorities.
Exposure
Following the initial training and experience in providing Patient Priorities Care, the participating clinicians and the Patient Priorities implementation team participated in 21 group discussions between March 2017- March 2018. Participants received a summary of the health conditions, medications, recent procedures, functional status, and health priorities template of 1–2 patients participating in the pilot selected by the priorities facilitators (example of a template in S1 File ). The template included the patient’s health values, health outcome goals, healthcare preferences, self-perception of their health trajectory, and Specific Ask (i.e. the health or healthcare issue patients most wanted to focus on to help achieve their most desired activity) [ 23 ]. Each session began by reviewing the template. Because there was no a priori approach for aligning individual patient’s priorities with clinical decisions, the group engaged in emergent learning, defined as learning which arises from conversations and interactions among the participants [ 26 ]. Following collaborative learning techniques, everyone was encouraged to share their experiences and suggestions concerning patient priorities-aligned decision-making for the selected patient and other similar patients [ 27 – 29 ]. The group discussed challenges, posed possible solutions, and worked together to determine how best to align care options with the selected patient’s specific health priorities.
Most discussions were recorded and transcribed by a professional transcription service; detailed notes were available for the remaining sessions. This study was approved by the Human Investigations Committee at Yale University.
Analysis
Congruent with participant observation design, the members of Patient Priorities implementation team with expertise in clinician training, behavioral medicine, and care of older adults with MMCs, participated both in the discussions with clinicians and in interpreting the discussions from which the challenges and strategies were identified [ 25 ]. Following a qualitative analytical design, two members of the Patient Priorities implementation team (DS and LD) independently reviewed and interpreted the content of the transcripts and notes from the facilitated discussions to identify decisional challenges to patient priorities care alignment and strategies to potentially address these challenges. Using the constant comparative method, they continuously compared and categorized the data to identify, and progressively refine, emerging challenges until no new decisional challenges were found [ 30 , 31 ]. They then consolidated, refined, and agreed on an initial set of decisional challenges [ 30 , 31 ]. These initial challenges were reviewed and modified by all members of the Patient Priorities implementation team, described above, until there was consensus that they reflected the clinical scenarios and discussions. The same process was used to identify strategies to address these challenges and to help clinicians align decision-making with each patient’s health priorities. We followed the COREQ standards for reporting qualitative research findings [ 32 , 33 ].
Results
The Patient Priorities implementation team and clinicians reviewed and discussed 35 patient scenarios over the 21 sessions. These sessions included 10 face-to-face (five with PCPs; four with cardiologists; and one with both groups) and 11 telephonic (six with PCPs, two with cardiologists; and three with PCPs and cardiologists) case-based group discussions. The patients discussed ranged from 67–98 (median, 78) years old; 75% were female; all were Caucasian. All had at least five chronic conditions. The number of active problems listed in patients’ EHR ranged from 7–66 (median, 19). Patients received from 5–16 (median, 10) prescription medications. Descriptions of the 35 patients discussed are displayed in S1 Table . Sociodemographic and function information, chronic conditions, active problems, and medications were ascertained from their EHRs. Health outcome goals and healthcare preferences were ascertained from their health priorities identification process.
Several challenges and strategies were identified from review and interpretation of the discussion sessions. Rationales were also identified for the strategies as were approaches for implementing the strategies.
Decisional challenges ( Table 1 )
10.1371/journal.pone.0218249.t001
Table 1 Challenges in aligning clinical decisions with patients’ health priorities among older adults with multiple chronic conditions.
Challenges
Representative Quote
Uncertainty, complexity, and multiplicity make decision-making difficult
No obvious best decision; not knowing where to start as there was so much going on
“This is a complicated lady , she’s got so many comorbidities . There’s too many parts of this puzzle that we don’t have enough information about it to make smart recommendations . ” (PCP about Patient # 6) “… His diabetes is one thing , but what contributions his aortic stenosis and/or medications may be having ? He’s on a lot of high blood pressure medications…” “He’s got like 30 things wrong and he’s on a ton of medications . ” (Cardiologist about patient # 15 for whom lightheadedness impedes his goal to go to casino and grocery shop)
Often no single identifiable or remediable symptom
“He has symptoms that can be related to a lot of different things…” (PCP about patient # 28 who complains of fatigue and lightheadedness) “…our discussion revolved around these different problems , which are difficult to tease out . How much of her fatigue and shortness of breath is [due to] obesity , how much is DJD , how much depression plays into it… .? ” (PCP about Patient # 34)
Differing perspectives on what matters most
Patient prioritizes current discomfort or treatment burden; clinician prioritizes risk of future event
“I think that if we can explain to patients the reason why they’re on Coumadin . The risk of stroke versus risk of bleeding … strokes , even at 92 , are oftentimes in the setting of afib debilitating . ” (Cardiologist about patient # 2)
Clinicians vary in their views of the relative importance of information and which treatments are most likely to help patient
“He wants to leave her on the Coumadin… I approached from “what matters most to her” and he quoted studies to me… we can’t get any consensus …” (PCP about patient # 2) Cardiologist‘s perspective concerning patient # 2 is noted above.
Difficulty switching from disease guidelines to patients’ priorities as the focus of decision-making (even knowing patients’ priorities)
Uncertainty whether treatment benefits reported in disease guidelines apply to this population
“We have established a bunch of fairly specific , rigid guidelines for the care of patients…I don’t know if it helps ” (PCP about decision-making with patients with MCCs) “I’m not sure how many 90-year-olds were in the studies… . I would venture that there probably is a little more uncertainty in a 90-year-old …” (Discussion concerning Patient # 2)
Revert to disease guideline-based decision-making despite knowing patients’ priorities.
“… . He needs all of his heart medicines . He has class one guideline evidence supporting his medications . ” (Cardiologist about patient # 8) “I didn’t talk specifically about her goals . My goal is to fix her blood pressure . ” (PCP about patient # 143)
The three categories of challenges discussed during the sessions were the uncertainty, complexity, and multiplicity of conditions and treatments; differing perspectives on what mattered most; and the difficulty switching from disease guidelines to patients’ priorities as the focus of decision-making. Other challenges mentioned by a participant, but not discussed during the sessions, included outcome goals that were too vague or not actionable; a disconnect between the outcome goals patients desire and what they are willing to do to achieve these outcomes; and the impossibility for health priorities to fit every potential medical decision. Only the three categories of challenges that were discussed among the participants are included in Table 1 .
The first category of challenges centered around the uncertainty, complexity, and multiplicity of health conditions and treatments—inherent to the presence of MCCs—that make decision-making difficult even with knowledge of individuals’ health priorities. Clinicians commented on not knowing where to start and not having an obvious best decision, “There’s a lot of variables to account for . He’s got like 30 things wrong and he’s on a ton of medications . ” Dealing with multiple symptoms in the face of multiple conditions and treatments—what caused them, how to eliminate them, and which symptoms were most limiting—was mentioned often as a source of uncertainty and complexity.
Differences in perspectives of what mattered most between patients and clinicians or among clinicians was the second category of challenges. Clinicians occasionally expressed concern if patients considered current harms and burdens of care more important than prevention of future bad events, feeling that some patients did not appreciate the importance of these events, “I think that if we can explain to patients the reason why they’re on Coumadin , and perhaps some patients like numbers . I don’t know , whatever works for that patient …” Clinicians caring for the same patient may differ in what treatments best helped patients achieve their outcome goals as suggested by one clinician, “ I talked with [another clinician caring for the patient] a nd he wants to leave her on the Coumadin . I don’t think she is going to have any benefit . I approached from ‘what matters most to her’ and he quoted studies to me . ”
The third challenge category observed was the tendency to revert to disease-based decision-making. Clinicians recognized that the need to follow disease guidelines at times impeded making patient priorities-aligned decisions. As one PCP noted, “ We have established a bunch of fairly specific , rigid guidelines for the care of patients…”
Strategies for implementing patient priorities-aligned decision-making ( Table 2 )
10.1371/journal.pone.0218249.t002
Table 2 Strategies for implementing patient priorities-aligned decision-making.
Rationale for the Strategy
Tips and Scripts for Using the Strategy
Strategy 1: Start with one actionable thing that matters most to the patient
In the absence of an obvious best decision and in the presence of uncertainty, where else would you start except with what matter most to the patient? Adherence is likely to improve if you begin with what matters most to the patient.
The “Specific Ask” helps start patient priorities-based communication and decision-making. “I want less back pain and dizziness so that I can : keep living at home and do more with my husband around the house” (Patient # 457). Use the response to the “Specific Ask” in decision-making.
Strategy 2: Conduct serial trials of starting, stopping, or continuing therapies. Measure effect on patient’s health priorities
In the face of uncertainty, serial trials, measuring success (or failure) against attainment of health priorities, helps clinicians titrate care to maximize benefit and reduce burden.
Acknowledge that there is no single right answer. Establish timelines and use patient goals as metrics of success or failure. “We can’t always be sure what will work best for each person , …We can see if it [the change] helps over the next two months . If not , we will work together to try different things . ”
Strategy 3: Function over symptoms (Focus on achieving activities—health outcome goals—rather than eliminating symptoms)
Focusing on how symptoms are interfering with meaningful activities may be more productive than trying to eliminate symptoms. This is because, as noted in challenges, it is often uncertain what is causing the bothersome symptom. Also, it is often not possible to eliminate the symptom completely. In these situations, linking treatments to the patient’s specific goal activity can guide decision-making and be effective. Focusing on activity often paradoxically improves symptoms and is a good metric for tracking whether a treatment change is working.
Focus on achieving patient’s desired activity, “If you were in less pain (less tired , SOB) , what would you be doing more of ? ” Acknowledge the uncertainty and that serial trials are often needed. “There are several possibilities… we can’t be sure what is causing (symptom) . A good place to start is (proposed change) . We’ll see if it helps you (desired activity) . ” “The nice thing is you have something to guide against- is the lightheadedness sufficiently resolving to do these activities . ” (PCP about patient # 423)
Strategy 4: Priorities-based communication (Use patient’s health outcome goals and healthcare preferences to discuss care).
Focusing communication on patients’ priorities encourages decision-making based on these priorities. Adherence is likely to improve if recommendations are tied to meaningful outcomes for patients.
Link recommendations to goals and care preferences, “I know you don’t want procedures but the valve procedure may help the tiredness that keeps you from walking your dog and may help decrease your medications . You said those things are important . ”
Strategy 5: Negotiate a shared decision (when there are differences in perspectives)
Individuals may have different perspectives and use different information to make decisions. There is no one best answer for patients with multiple conditions and variable priorities. In the face of uncertainty, patients’ priorities are the obvious unifying target of decision-making.
Agree on information that informs decision (i.e. patient’s priorities, intervention burden, all chronic conditions, life situation, health trajectory). When patient-clinician differ, present estimates of benefits and harms, expressed in the context of patient’s priorities. Be realistic about absolute benefit (often modest). Accept that older adults appropriately may value current health over future events. When clinicians differ, use collaborative negotiations, brainstorm alternatives. and agree on compromise solution, “We could try a three-month trial off Coumadin… does she get to her dining room ? (Cardiologist and PCP about patient # 39)
The first strategy that emerged to address these challenges was to start with what was most important to each patient. Prioritizing one actionable thing that is most important to the patient helped simplify decision-making. This operationalizes patient priorities into one focused “Specific Ask (One thing): “ The one thing about my healthcare I most want to focus on is X so that I can do (desired activity) more often or more easily .” The behavioral medicine expert (LD) noted that starting with patients’ priorities helps engage patients as active partners in their care while encouraging clinicians to focus on what was important to the patient. The Patient Priorities implementation team felt the Specific Ask served as a means of linking patients’ priorities with decision-making by aligning the outcome patients most desired with the health or healthcare issue they considered the key barrier to achieving the outcome.
Conducting serial trials was a second strategy identified. Participants agreed that serial trials of starting, stopping, or continuing various interventions was the most practical strategy for ongoing decision-making given the complexity and uncertainty involved. Success or failure of the trials should be defined by whether patients achieved their health outcome goals or healthcare preferences, “There are several things that we could do…We will work together to try different things if that is ok with you . ”
The third strategy was focusing on achieving desired activities (patient’s health outcome goals) rather than eliminating symptoms. The uncertainty of causality and low likelihood of complete alleviation of symptoms often makes the achievement of activities a more successful clinical strategy. Increased participation in meaningful activities is also a motivator for patients and a good metric for tracking whether a treatment change is working, “If you were in less pain (less dizzy , not so tired , weren’t so depressed) , what would you be doing more of ? ”
Basing communication between patients and clinicians and among clinicians on patients’ priorities was a fourth strategy that evolved. Reminders that patient’s health priorities, rather than disease guidelines alone, should be the focus of communication helped move decision-making to patient’s priorities by considering the potential benefits, harms, or burdens of diagnostic and treatment options within the context of these priorities. “ Would you be willing to try the CPAP for a week and see if it helps with your fatigue ? If it helps with your fatigue , you may be able to walk more with your wife . ”
The fifth strategy focused on arriving at shared decisions based on patients’ priorities when there were differences in perspectives between clinicians and patients or among clinicians. Steps that contribute to shared decisions consistent with patients’ priorities include agreeing on the information that informs the decision; realistically estimating treatment benefit; using collaborative negotiations including brainstorming alternatives; and accepting patients’ decisions if they understand the benefits and harms ( Table 2 ).
The implementation team concurred that the identified challenges could be addressed by various combinations of the strategies. Conversely, each strategy addresses one or more challenge.
Discussion
Through discussions among clinicians and the Patient Priorities implementation team, we identified challenges to aligning clinical decisions with patients’ health priorities as well as feasible strategies for translating patients’ health priorities into decisions and care. The challenges can be addressed by several of the strategies and, in turn, each strategy addresses one or more challenge. For example, all five strategies facilitate decision-making in the face of the uncertainty and complexity inherent in the care of older adults with multiple coexisting conditions and variable priorities. Similarly, all the strategies help get patients and their clinicians on the same page when they start with differing perspectives. Clinicians facing difficulty switching from disease guideline—to patients’ health priorities—aligned decision-making can start with one actionable thing that matters most to the patients and conduct serial trials using patient’s health outcome goals and healthcare preferences in communications with patients and other clinicians. Taken together and applied in appropriate situations, the decisional strategies can guide implementation of patient-centered care, particularly for persons with MCCs [ 16 ].
Other approaches for helping clinicians focus on achieving patients’ goals and preferences have been developed [ 15 , 17 – 20 , 34 – 37 ]. These approaches target specific health problems or patient populations [ 15 , 17 – 20 , 34 , 35 ], or provide limited guidance in how to align decision-making with patients’ goals and preferences [ 36 , 37 ]. Patient Priorities Care builds on this earlier work to encompass all persons with multiple conditions and provide guidance in how to link all types of available healthcare to patients’ priorities.
A strength of the current work is its foundation in empiric observation and collaborative, emergent learning [ 25 – 29 ]. Problem-solving evolved from discussion of actual clinical scenarios, with their inherent complexities and nuances, ensuring our strategies and tactics were based on situations as they occur in practice. Both strategies and challenges were identified from multiple perspectives including primary care providers, specialty clinicians, and experts in clinician training and patient and clinician behavior. The strategies that emerged were those considered effective and feasible from all these perspectives. Some of the strategies recommended were based on experience from other fields. The focus on functional activities rather than solely relief of symptoms, for example, comes from the pain management field [ 38 , 39 ]. Negotiating a shared decision when there are differences in perspectives is a collaborative negotiating technique in the business field and is also an underpinning of collaborative and shared decision-making in healthcare [ 40 , 41 ]. Starting with one actionable thing that matters most and conducting serial trials were practical recommendations for addressing the uncertainty and complexity of care involving older adults with multiple conditions.
Prior to starting the pilot, the participating clinicians raised concerns about urgent decisions not covered by patients’ health priorities and about patients’ identifying unrealistic goals given their health status and trajectory. However, during the pilot, the clinicians concurred that the process the patient went through in setting health priorities helped frame big decision discussions, such as one patient with heart failure facing a decision about placement of an implantable cardioverter defibrillator. Because identifying realistic, achievable goals was part of the process, unrealistic goals were rare. One example was a 90+ year old woman with advanced heart failure who wanted to play nine holes of golf, walking the course. Using collaborative negotiations and priorities-based communication, the patient agreed to previously declined testing to better identify potential treatments that might help her achieve her priorities. She also agreed to a more realistic goal of using a golf cart and playing fewer holes.
Of necessity, the strategies are general and must be adapted for each clinical situation. As experience with this approach to decision-making increases, additional challenges and strategies may emerge for aligning healthcare with patients’ health priorities. Nevertheless, it was reassuring to observe that a few strategies could be used to address wide-ranging issues faced by patients with multiple conditions and by the clinicians caring for them.
There are additional limitations to this study. The challenges and strategies reflect observation of a single group of clinicians in a single setting involving a discrete number of patients. While valid within the context of the population and setting of the current study, our findings require replication in additional settings with diverse groups of patients. The inclusion of other specialties who care for older adults may reveal additional challenges and strategies. We have not yet pursued widespread uptake of these strategies nor have we determined the effect such strategies and tactics would have on patient outcomes. Evidence supporting the benefits of current diagnostic, preventive, therapeutic and other interventions for individual health priorities remains lacking. As research focuses directly on the outcomes that matter to persons with multiple conditions, this evidence will increasingly emerge.
Identifying feasible strategies suggest that patient priorities-aligned decision-making may be possible in clinical practice. Indeed, clinicians noted using some of these strategies before participating in this pilot. The current work suggests that implementing these strategies systematically and explicitly may offer an approach to decision-making in the face of complexity and uncertainty for persons with multiple conditions, perhaps helping to reduce intervention burden and increase likelihood of achieving patients’ most desired outcome goals.
The Patient Priorities Care initiative is continuing. In parallel to our work with clinicians, we are guiding patients in how to identify and communicate their health priorities [ 23 ]. Ongoing work also includes evaluating the effects of identifying patients’ priorities and using them in decision-making on patient, clinician, and health system outcomes. Once we have further developed and tested the Patient Priorities Care approach, including the strategies identified in this study, we will engage additional clinical practices and sites to assess acceptance and uptake. The eventual goal is the development and dissemination of an approach to decision-making and care that maximizes benefits that matter to individuals with multiple chronic conditions while minimizing harm and intervention burden.
IRB APPROVAL : Yale Institutional Review Board (IRB) approved this study prior to data collection. The IRB approved verbal consent to participate in this study. All patients provided verbal consent to participate, share the data collected in the study with the investigators and publish without identifying information. Several of the authors, the Yale Chief HIPAA officer, and the Executive Director at the participating primary care practice reviewed the information included in the Supporting Information to ensure individuals could not be identified based on personalized details.
Supporting information
S1 File
Example of completed health priorities template.
(DOCX)
S1 Table
Characteristics, health priorities (health outcome goals and healthcare preferences) of patients selected for the facilitated discussion sessions*.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
All archaeologists know that bifaces are chipped stone tools worked on two opposing faces separated by retouched edges. The category “biface” includes finished tools—points of various defined types and/or other functional classes like knives—but also unfinished specimens that span a range of reduction from slightly modified flake blanks to nearly finished tools. This paper concerns the process of biface reduction. It evaluates two views of the process’s nature, as continuous or as segmented into real, technically determined, stages [ 1 ].
In studies of biface reduction, no consistent vocabulary of relevant terms has evolved. Here I follow usage that traces back to Muto ([ 2 ], 109), in which a “blank” is an unmodified object, a cobble or large flake, of size and form suitable for reducing to a tool. A “preform” is a blank that has undergone partial modification toward a finished form. Thus, a blank is any piece of stone that can be reduced to a tool such that “The shape or form of the final product is not disclosed in the blank” ([ 3 ], 42) except in the broadest sense. A preform “is an unfinished, unused form of the proposed artifact. It is larger than, and without the refinement of, the completed tool…thick, with deep bulbar scars, has irregular edges, and no means of hafting” ([ 3 ], 85). Finished tools are the result of the final transformation of preforms.
The study of preform assemblages can reveal the production process that begins with blanks and concludes with finished tools, thereby tracing “the evolution of the biface from a raw-material blank to a refined finished product” ([ 4 ], 180). Raw cobbles rarely were thin for their plan area, and flake blanks may only have approximated this form. Most bifaces were used to pierce or cut, so their working part usually was a tip and one or more sharp, acute-angled edges. But those working edges were only part of a larger whole that had to be grasped in the hand or secured in a haft. Generally, then, bifaces were thin for their length, width and area, which served the dual purposes of producing those tips and sharp edges and providing the heft and size for prehension to use the edges effectively. Many also were modified near the base to facilitate hafting. To reach desired sizes and forms, the biface reduction process involved a series of controlled flake removals that disproportionately decreased thickness relative to length and width, while also creating two opposing faces that converged on first crudely and then finely sinuous edges. The progress, therefore, from raw cobble or blank through preform to finished tool involved allometric reduction via thinning, edge-formation and generally finer approximation to the desired final form and size.
Reconstructing reduction sequences—the patterned ways that cores and preforms were reduced to tools—is a common goal of North American lithic analysis. Holmes [ 5 ] originated the reduction concept in the context of the Trenton Gravels debate. Following him, other archaeologists today view reduction as a sequence of essentially discrete stages [ 1 , 6 ]. Ironically, though, Holmes himself thought of stages as merely heuristic devices at levels from the evolution of culture ([ 7 ], 248–249) to the production of stone tools [ 5 ], i.e. as descriptive conveniences, not fixed empirical states [ 8 ].
Whatever Holmes’s view, Callahan’s Paleoindian stage model legitimately is celebrated both for the great detail of its text and the beauty of its tool drawings. As Bennett put it, “Callahan’s interpretation of the Clovis manufacture method has been the basis of all subsequent conceptual models of tool production applied to Paleoindian groups” ([ 9 ] 29) in America North and South [ 10 ], as well as to later prehistoric industries. Callahan viewed reduction largely in sequential terms, although like Holmes he spoke also of continua ([ 1 ], 3; [ 11 ], 515 captured the ambivalent character of Callahan’s views on stages and continua). Callahan established the vocabulary that endures today in Paleoindian studies and elsewhere: North American archaeologists routinely speak of, for instance, “Stage 2” or “Stage 3” bifaces. (Callahan [ 1 ] used Roman numerals to label stages, but most archaeologists now use Arabic numerals, as Callahan later did [ 12 ]. Villa et al. ([ 13 ], 446) called their equivalent units “phases,” which seemed “less formal” to them.) Doing so, they employ Callahan’s concepts and terms and engage his assumptions about the reduction process.
The concept of stage seems intuitively reasonable, particularly from the perspective of replicators whose purposes and procedures can be organized by changes in hammer and in goals as the emerging biface approaches its desired form. For instance, Apel ([ 14 ], 129), an ardent if less nuanced advocate than Callahan, perceived stages as real entities that reflect knapper intent and sequential changes in “technique and method,” i.e., as real, well defined, and unambiguous. Bleed [ 15 ] also advocated stages, to some extent equating them with any spatial discontinuity in reduction sequences rather than with discrete technical steps or processes, all based on refit chains comprising small parts of unknown representativeness of larger flake assemblages. Yet the stage concept can embed assumptions. Bleed, for instance, assumed that continuous reduction must be unbroken in time, beginning to end occurring in one episode in one place. It also assumes the validity of stages, whose number and nature vary between replicators and analysts. One unsurprising result is considerable differences in stage assignment. Also starting from Holmes and increasing in frequency as biface-reduction modeling has expanded in recent decades a view of reduction has arisen as a continuous process better suited to analysis in quantitative terms.
To ask which of stage or continuum approaches is valid is, to some extent, a contrived question because both can be useful. Yet it is not contrived to consider which approach is best. Stages are at least analytical constructs, and may or may not be real. The stage concept possesses value as an organizing device for description and perhaps comparison between industries and reduction sequences, but in some respects is inherently ambiguous and may be needlessly restrictive. The continuum alternative also is an analytical construct, possessing its own strengths and weaknesses. Neither may be sufficient alone for comprehensive study of reduction sequences. But the stage concept is of longer standing, being firmly established in thought and practice by the 1970s to the virtual exclusion, until recently, of alternatives. Even as it increasingly is questioned, the stage view’s dominance justifies focus upon it and its assumptions [ 16 ], followed by one approach to continuous analysis of biface preform data.
The stage approach
The stage concept assumes that stages are valid and replicable. Stages are valid if they capture or correspond to legitimate patterns of association between variables, such that all members of a stage share at least some essential qualities that separate them from other stages. The qualities may be particular technological traits, ideal values of dimensions or ratios, or at least discrete, non-overlapping ranges of metric values. If what are defined as discrete stages show weak patterns of variable association, or continuous ranges of variation and significant overlap between successive stages, then they are merely subdivisions of a complex continuum of biface reduction. In that case, the stage concept’s validity is in doubt.
Stages are replicable if all archaeologists who contemplate the same specimens would define the same number of stages that possessed the same characteristics and would assign the specimens to the same stages. If what are defined as distinct stages differ in number, nature or type between analysts, or if the same set of specimens is apportioned among defined types to significantly different degree by different analysts, then the stage concept’s replicability is in doubt.
Whatever other technological reduction sequences may reveal about the nature or reality of stages in technical or spatial terms [ 15 ], focus here is upon biface production. This is not necessarily distinct from core reduction, because cores can be progressively transformed into bifaces in the course of reduction ([ 17 ], 115). Yet the process should be similar whether cores are reduced to bifaces in the process of producing usable flakes or bifaces are the sole objective.
Callahan ([ 1 ], Table 10) defined stages in this process by combinations of discrete and continuous variables. The former included cross-section form, flake-scar intervals, size and form of flake scars (and variation in both), extension of flakes to or across the longitudinal midline (“nature of opposing flake scar contact,” expressed as the proportion of flakes that extended to or past the midline), regularity of plan form, and platform preparation. Qualitative variables or at least criteria also included some that could only be inferred, not observed (e.g., hammer type, knapper emphasis upon edge, face, or outline, knapper concentration or attention-to-task, work pace, “degree of trim”). Interval- and continuous-scale variables included the number of flake removals, the ratio of width to thickness, edge angle, and the ratio of stage weight to final weight. In detailed description of his experimental replicas, Callahan ([ 1 ], Figs. 1–73) reported not only width-thickness ratio, average edge angle, and weight, but also length, width and thickness.
Callahan’s approach proved undeniably popular (e.g., [ 4 ], 180–181, Table 7.7; [ 9 ], 82–87; [ 10 ]; [ 11 ], 515–6; [ 14 ], Fig 2.3; [ 18 ], 26–65; [ 19 ], Table 5.23; [ 20 ], 202–214; [ 21 ], 82–110)). To Callahan Stage 1 was the blank, whether flake, core or cobble (cf. Sharrock[[ 22 ], 43] who described Stage 1 as the “flake blank” [[ 22 ], 43]] but to judge from description and illustrations [[ 22 ], Figs. 23–24] was an early-stage worked preform). Stage 2 was the interval in which initial edging of the flake blank creates the preform’s perimeter and, broadly, its plan form. In Stage 3, the preform was thinned across most or all of both faces, flakes extend across the longitudinal midline, moderately convex surfaces are formed on both faces, and edge sinuosity was reduced, all in the process of removing areas of disproportionate thickness. At this stage failure was fairly common, either by transverse fracture, plunging ( outrepassé ) terminations often called “overshots” in the Clovis literature ([[ 23 ], 68], although overshooting may be error more than design ([ 24 ], 53], and occur in later prehistoric industries as well [e.g. [ 17 ], 118]) or abandonment (e.g., remnant crowns defined by step fractures that failed to terminate normally and therefore failed to remove excess thickness). Stage 4 involved secondary thinning, yielding preforms with relatively flat faces, regular cross-sections and moderately sinuous edges. Stages 5 and higher involved largely haft (including fluted in Paleoindian industries) and perhaps tip modification.
Even Clovis-biface reduction models that differed from Callahan’s in details of fluting and finishing were descriptively similar or identical through Stage 4 ([ 19 ], Table 5.23; [ 20 ], Table 29, Fig 47). Callahan reported data only for specimens in Stages 1–4. Accordingly, handbooks of lithic analysis use Callahan’s scheme up through its Stage 4 ([ 4 ], 180; [ 25 ], 100) explicitly followed Callahan, although Andrefsky’s stage criteria did not always match Callahan’s—and the question of stages in biface reduction largely involves blanks (Stage 1) and Callahan Stages 2–4.
Table 1 summarizes Callahan’s stage-assignment scheme. Note the variability among them, and the complex nature of some of these variables (e.g., nature and size of flake scars, outline regularity) used to describe them. Perhaps as a result, in Callahan’s detailed illustrations of preforms as they progressed through his stages most discrete variables were not coded, measured, or otherwise reported. For instance, Callahan’s ([ 1 ], 68, 92, 118) Stage 2–4 exemplars were coded by degree of success, hammer type, and problems encountered—all difficult if possible at all to code in archaeological specimens—not the attributes listed in his scheme. Table 1 also summarizes biface-reduction schemes and preform stages drawn from other sources. Sharrock’s predates the Callahan model but resembles it in many salient respects; others were influenced or guided by Callahan’s approach. Muñiz’s ([ 17 ], Table 7.2) somewhat similar model emphasizes outline form, presence/absence of cortex, edging and faceting; it includes only thickness among dimensions. Huckell ([ 26 ], 191–194) also devised a similar approach that emphasized extent of flaking across the preform surface, which he considered a diagnostic Clovis characteristic (cf. [ 17 ], 122). Villa et al. ([ 13 ], 447, Table 3) defined four successive phases in production of South Africa Still Bay bifaces (excluding resharpening of finished specimens) that resembled Callahan’s stages and were defined in part by variables similar to those listed in Table 1 .
10.1371/journal.pone.0170947.t001
Table 1 Callahan’s stage scheme and selected variants. Sources: [ 1 ], Fig 1–63; [ 4 ], 180–184, Table 7.7; [ 11 ], 516–518, Table 1; [ 18 ], 43–65, Appendix A; [ 20 ], 201–213, Appendices A-E; [ 21 ], 82–121, Figs 13–27, Tables 12–14; [ 22 ], 40-46b.
Callahan
Sanders
Morrow
Sharrock
Andrefsky
Dickens
Hill
1979
1983
1996
1966
1998
2005
2013
STAGE 1
cortex
present
length
50–300 mm
w/t 1
2
plan form
elongate
edge sinuosity
high
edge angle
cross-section
facets
6-10/face
STAGE 2
cortex
present
absent
present
length
60–112 mm
ca. 115 mm
w/t 1
2.0–3.0
2.0–3.0
3
2.0–4.0
2.0–3.1
2.0–3.0
plan form
irregular
elongate
irregular
edge sinuosity
lesser
high?
edge angle
55–75°
40–110°
50–80°
55–75°
cross-section
lenticular-irreg.
lenticular-irreg.
lenticular-hexagonal-irregular
facets
12–24; variable,
none to
10-20/face
widely spaced,
midline
widely spaced
not to midline
Some across midline
not to midline
STAGE 3
cortex
absent
rare
length
73–137 mm
100–250 mm
ca. 104 mm
w/t
3.0–4.0
2.6–5.5
5.0–8.0
3.0–4.0
2.1–3.5
3.0–4.0
plan form
semi-regular
ovate
elongate
semi-regular
edge sinuosity
fine
edge angle
40–60°
40–70°
40–50°
40–60°
cross-section
regular
relatively thick
lenticular
Facets
6–12; variable,
more; more
closely spaced,
regular
across midline
pattern
across midline
across midline
across midline
STAGE 4
Cortex
absent
length
50–100 mm
ca. 105 mm
w/t
4.0–5.0
4.0 2
10
4.1–6.0
4.1
4.0–5.0
plan form
regular
elongate
lanceolate
regular
edge sinuosity
fine
edge angle
25–45°
35–75°
25–45°
25–45°
cross-section
flat, lenticular
flat, lenticular
Facets
12–24; regular;
closely spaced,
across midline
STAGE 5
Cortex
absent
length or size
25–50 mm
w/t
4.0–6.0+
4
4.1–6.0
4.7
plan form
elongate
edge sinuosity
fine
edge angle
40–50°
25–45°
cross-section
Facets
1 width/thickness
2 [ 20 ], 210) reported mean length of 103 mm, width of 43 mm, thickness of 10.7 mm. 43/10.7 = 4.02.
Although there is broad agreement between schemes and fairly close agreement in some particulars, clearly these heuristic models are not entirely compatible, nor do all include the same set of measures. Some models reported dimensions by stage and employed width/thickness ratios, while others reported no size dimensions at all. Some emphasized cross-section or scar patterns; others ignored these attributes. Where it concerns flaking, some schemes referred to amount, others to pattern, still others to whether primary thinning flakes reached midlines or not. Even where specific measures like width/thickness were used, schemes differed. Andrefsky and Sharrock, for instance, used substantially different edge-angle ranges and width/thickness ratios by stage and Sanders’s Stage-2 edge-angle range was considerably wider than others’. Whatever their differences in width/thickness ratio, most sources reported progressive increase in the ratio as thickness declined disproportionately in the reduction process. Yet elsewhere Bradley ([ 27 ], Fig 17.7) reported increasing then declining ratios, such that early-stage preforms were more similar to finished points than were later-stage preforms. Archaeologists agree that there are stages to preform reduction, but apparently not always on the number or defining criteria of those stages. Clearly, there is some difference in which metric or discrete attributes are important and, in the former’s case, which values of them are important.
As above, Callahan’s stages were defined by detailed sets of discrete and continuous technological and flaking variables. Because most such variables were not reported for illustrated specimens in the original source, stage assignments using that source can produce ambiguous results. Only repeated blind tests in which several analysts independently assigned preforms to stages can determine the replicability of Callahan’s scheme. Until then, archaeologists must use their own judgment, applied to a fairly wide range of production preforms that vary not only in degree or stage of reduction but also in toolstone, flake- or core-blank size and form, and size and form of the desired end product. In the face of multiple causality Hill ([ 11 ], 515) concluded that “refined flintknapping, such as biface reduction, involved a complex decision-making process, requiring an understanding of many complex, interacting variables.” Stage models may not be the only or always the best method to determine the nature and pattern of complex variation in preform assemblages.
Stage or continuum?
Study of any complex process, certainly of biface reduction, can engage ambiguity. To Bennett, Callahan’s approach “can be readily applied to any bifacial tool technology” ([ 9 ], 81). Yet Johnson suspected that he was “not the only one who has tried to use Callahan’s… comprehensive replicative study…only to be frustrated by the nine stage typology. It is practically like working with raw data” ([ 28 ], 159). Descriptive detail and comprehensiveness are not flaws but may act against clarity in analysis and results. Johnson legitimately questioned the workability of Callahan’s system and suggested the risk of inconsistent application of its own criteria to its subjects. Sanders found “varying degrees of Stage IV reduction” of preforms ([ 21 ], 111) at the Adams quarry/workshop, suggesting considerable variation within this “stage.” Even his analysis, which followed Callahan’s approach closely, ([ 21 ], 83,99,111) reported mean length, width and thickness values for Adams site preforms by stage that did not match Callahan’s values for the same stages. Similarly, to Amick “stage classification remains a subjective, qualitative assessment” ([ 29 ], 140).
These criticisms echo in more recent studies. Jones ([ 30 ], 23, 164) rejected stage models in part for ignoring toolstone variation that cross-cut defined stages. To Miller and Smallwood “the use of stage designations relied heavily on the lithic analyst’s subjective assessment” ([ 31 ], 31). Prasciunas found that “discrete stages in the reduction process…can be fairly arbitrary and are defined differently by different researchers” ([ 32 ], 38). The ambiguity that inheres in application of Callahan’s system is captured nicely in Dickens’s [ 18 ] comparison of the Adams assemblage [ 21 ] to his own Gault specimens. Finding significant differences in metric and other attributes in Stage 5 Adams and Gault preforms, Dickens suggested that the former “may actually be late Stage IV or very early Stage V bifaces, as opposed to a more developed and recognizable Stage V” ([ 18 ], 68).
Such views are not confined to Paleoindian studies. Despite following Callahan’s system as closely as possible and arguing for at least the partial empirical reality of stages, Beck et al. found “stage assignment…somewhat arbitrary” ([ 33 ], 494), and reported disagreement between experienced analysts in assignments. Like Beck et al., my own assignment of preforms to Callahan stages was consistent internally but not necessarily consistent with others’ ([ 34 ], 555). In Polynesian adze studies, there exists a confusing diversity in stage models resulting in a “lack of consensus in clearly defining stages along the blank-preform continuum” ([ 35 ], 362; see also [ 36 ], 10).
Accordingly, some archaeologists have questioned the validity and replicability of “stages.” As opposed to stages, even before Callahan’s influential study Muto spoke of the “‘blank-preform-product’” ([ 2 ], 109) continuum, and Collins ([ 6 ], 16–17) viewed reduction as a linear and therefore continuous process that was merely “convenient” to divide into stages. Other archaeologists demonstrated continuous variation in the reduction process, perhaps ironically by study of flake debris rather than bifaces or other reduction products themselves. Ingbar et al. ([ 37 ]; see also [ 38 – 40 ]), for instance, showed that reduction as reflected in the character of flake assemlages is better understood as a continuum, not a sequence of discrete stages. They accomplished this by demonstrating how flakes that were ordered by removal from cores varied not so much by stage—all in this or that stage being very similar or identical in variables that defined the stage—but by continuous degree. Even limited and incomplete refit chains from archaeological cores could be modeled in continuous terms ([ 16 ], 325–330; [ 41 ]). These flake-debris studies suggested that discrete reduction stages may exist but that they must be demonstrated, not assumed, and that the continuous nature of reduction can be modeled mathematically.
Yet continuous modeling of the reduction process came only recently to the study of production bifaces. Despite adhering to the stage view, Julig ([ 19 ], 148, Fig 5.53) plotted preform edge angle against width:thickness ratio to show a fairly strong and continuous inverse relationship that did not reveal discrete clusters corresponding to stages. To Morrow and Fiedel, “bifacially flaked artifacts represent…a continuum of all stages of fluted point production” ([ 42 ], 126), and their plot ([ 42 ], Fig 7.3) of width versus thickness of Anzick specimens showed an unbroken continuum, not clusters of values that might correspond to discrete stages. Similarly, the Murray Springs Clovis assemblage comprised a “single reduction continuum” ([ 26 ], 193). Muñiz ([ 17 ], Table 7.2) spoke of reduction in “continuum stages.” More generally in Clovis assemblages, finished bifaces “resulted from a continuous reduction process” ([ 23 ], 78). To Waters et al. at Gault “biface reduction occurs along a continuum, [so] it is difficult to assign bifaces unambiguously to specific stages” ([ 43 ], 84). Santarone recognized the need for some purposes of “imposition of categories on a reduction continuum” ([ 44 ], 21) but cautioned against concluding that imposed stages were real. Although he continued to speak of “stages” for comparison to earlier scholarship, Muñiz considered biface reduction “as occurring along a continuum” ([ 17 ], 115), the variables he coded and measured permitting him to classify bifaces and bifacial cores “based on overall morphology along a continuum of reduction” (ibid).
Consistent with these views, archaeologists have begun to develop continuous measures suitable to track variation in the reduction process. Possibly the first, Johnson’s [ 45 ] Thinning Index, is considered below. Carper ([ 46 ], 136) used a symmetry index to model continuous variation across Callahan preform stages. Wilson and Andrefsky’s ([ 47 ], 96–97) “ridge-count reduction index,” designed specifically for analysis of production bifaces rather than finished tools, and Shipton’s ([ 48 ], 151; see also ‘[ 49 ]) scar-density index expressed faceting as a function of implement size and degree of reduction. Archer and Braun’s ([ 50 ], 207) analysis of Acheulian bifaces, corroborated by experimental specimens, yielded a single dimension of continuous variation that corresponded to degree of reduction measured by number of flake removals.
One of the most significant recent studies, of the Clovis biface assemblage at Topper, questioned the validity of stage definitions and replicability of stage assignments, and argued instead that “biface production was technologically a continuous process” ([ 31 ], 29). Callahan and other stage systems yielded ambiguous results when applied to Topper bifaces. Instead, Miller and Smallwood found that continuous variables, notably a measure of average number of thinning flakes per edge that they called the Flaking Index ([ 31 ], 31)(FI henceforth), better characterized the reduction process than did assignment of bifaces to stages. Complex variation in Topper bifaces could be modeled in continuous measures; any stages defined did not pattern as discrete sets but instead conformed to continuous trends ([ 31 ], 33–34). Compared to their FI, Miller and Smallwood concluded that “traditional stage models are ill equipped for describing the variability within the biface assemblage” ([ 31 ], 36).
In recent years, then, archaeologists have expressed serious reservations about the validity of stage models. Interval- and ratio-scale measures reveal continuous dimensions inherent in the reduction process that the stage concept accommodates poorly. The diversity of views and serious doubts expressed about the stage concept, accordingly, justify its close examination.
Materials and methods
Testing the stage approach in continuous data
If stages exist, in Holmes’s or anyone else’s sense, analysis can reveal them. It should not, however, assume their existence. Emerging critiques of the stage approach, and the strong possibility that continuous variables can reveal nature and degree of patterning otherwise difficult to perceive, justify exploration of several related lines of continuum-based analysis. This examination is possible only because Callahan ([ 1 ], Figs. 1–63) reported weight and the basic dimensions of length, width and thickness (all in cm, converted here to mm) by reduction stages. Again, although he defined from six to nine stages, Stage 1 is the blank and later stages correspond to relatively fine details of haft-element formation, so only Stages 2–4 are considered here.
No single variable or criterion can capture nearly the full range of complex variation in biface-preforms as they progress from blank to finished production. If continuous variables reveal patterns of variation consistent with stages, to that extent a stage approach is validated. If, conversely, analysis reveals patterns of inherently continuous variation then to that extent a continuum approach is supported.
Here, Callahan’s length, width, thickness and weight serve as gross size measures and two ratios serve as measures jointly of shape and degree of reduction. One is the simple ratio of width to thickness, which figured in Callahan’s original stage approach. The other is the Johnson Thinning Index [ 45 ] devised, as in the quarry studies that partly inspired this analysis [ 34 ], to measure progressive allometric change in preforms from blank to finished biface.
Johnson ([ 45 ], 13) defined JTI as the ratio of weight to plan area:
J T I ( g m / c m 2 ) = w e i g h t / p l a n a r e a
for each preform. JTI declines in value as reduction advances, because weight declines at a faster rate than plan area in the process of thinning biface preforms to completion. JTI and weight are not independent, obviously, because weight is the numerator in the ratio that yields JTI. Weight is a direct measure of size but only indirectly a measure of reduction because preforms of different weight can be at the same stage or position in reduction, and those of similar weight at different stages or positions in the reduction continuum, depending upon their starting weights and other factors. JTI is a derived reduction measure that simultaneously takes account of weight and plan area and, in the process, captures the allometry—change in shape with change in size—inherent in the reduction process.
Johnson ([ 45 ], 18) estimated plan area manually from a two-dimensional scan of each preform and polar-coordinate vectors. Amick ([ 29 ], 142) used two-dimensional scans but then a compensating polar planimeter to measure area. Like Amick, Beck et al. ([ 33 ], 495) used two-dimensional preform scans, but measured area by overlaying an orthogonal two-dimensional grid graduated in 5-mm increments on the scanned images. Thus, three different studies used different methods to calculate plan area for use in the JTI.
Here I estimate JTI (Johnson 1981:13) in a fourth way, calculated from area measured differently. The method, following Douglass et al. ([ 51 ], 518, where the expression is marred by a copy-editing error; see also [ 52 ]), models preforms as general ellipsoids whose surface areas are given from their main dimensions as:
S u r f a c e A r e a = [ ( a p b p + a p c p + b p c p ) / 3 ] 1 / p .
where p = ln(3)/ln(2), and a, b, and c are length, width and thickness, respectively. In [ 52 ] the expression above is multiplied by 2π to account for the three-dimensional surface area of ellipsoids. Because Johnson used two-dimensional surface area, I omitted the coefficient. In Callahan data, resulting values correlated strongly with a crude measure of plan area obtained as length multiplied by width (r = .61, p<.01), the Thomsen measure always was less than the length-width product as it must be except unless preforms were perfectly rectangular in plan form. Also, the least-squares regression intercept approached 0 and slope of the regression line of Thomsen surface area upon the product of length and width was about 0.8, i.e. <1, as it must be if the Thomsen measure varies over a narrower range than and rises with less than unit increase with the length-width product. Weight also correlated very strongly with the Thomsen measure (r = .89 p<.01). When weight was divided by the Thomsen estimate, resulting JTI values scaled similarly to values that Johnson ([ 45 ], Fig 2.6) reported.
Weight and JTI are sufficiently correlated that analysis of both may seem redundant. But all Callahan preforms analyzed were intact, whereas many archaeological preforms are broken. Weight is not a direct measure of original size in broken specimens. However, JTI is a valid measure of both intact and broken tools ([ 45 ], 18), as is FI ([ 31 ], 31). JTI scales weight to the plan area of a specimen. For instance, an intact preform that weighs, say, 10 g and measures, say, 5 cm 2 gives a JTI of 10/5 = 2.0. If the preform were broken into two equal halves, each fragment’s JTI would be the same, 5/2.5 = 2.0.
Seeking stages in the distribution of continuous variables
Here I examine variation in weight and JTI by seeking gaps or modes in their continuous distributions. Then I calculate differences within and between stages in mean weight. Finally, I consider the fidelity of Callahan’s weight/thickness ratio to his stage model.
To Verrey, “Callahan’s staging hypothesis will be reinforced if there are discrete clusters of values” ([ 53 ], 4) for continuous variables like weight and reduction measures like JTI. In Callahan’s dataset, the weight distribution had a single major mode and no conspicuous gaps or secondary modes that might correspond to stages ( Fig 1 ). The distribution was right-skewed with a very small mode at the upper extreme. But even that quasi-mode is ambiguous in distinguishing stages; of its five specimens it, three were assigned to Stage 2, two to Stage 3. Most preforms in both stages did not fall in it. Distributions by stage showed extensive overlap. Most gaps that might distinguish subsets of preforms occurred within Stage 2, not between it and other stages; such gaps may be nothing more than the product of small samples. Modes in Stages 3 and 4 were very similar in location.
10.1371/journal.pone.0170947.g001
Fig 1
Callahan Stage 2–4 preforms frequency distribution by weight: a. all preforms; b. by stage.
The JTI distribution also was right-skewed ( Fig 2 ); again, the highly ambiguous and very small mode at the upper end of the distribution was composed of three Stage-2 and two Stage-3 preforms, so again did not represent a discrete stage. Again, distributions by Callahan stage were similar, at least for Stages 2 and 3. Although the Stage-3 distribution included two apparent modes, one of them coincided in JTI range with the single Stage-4 mode, so preforms in the same JTI range were allocated to two different stages. On balance, no discrete modes or clusters of values were evident in the weight or JTI distributions.
10.1371/journal.pone.0170947.g002
Fig 2
Callahan Stage 2–4 preforms frequency distribution by JTI: a. all preforms; b. by stage.
When defined by combinations of categorical and continuous variables, Callahan’s stages may well be valid. But the replicability of his approach, and stages’ distinctiveness and internal integrity are undemonstrated. Mean variables by stage certainly differed (for weight, F = 13.7, p <.01; all pairwise least-significant differences [LSD] p <.03; for JTI, F = 20.7, p <.01; pairwise LSD comparisons were significant except between Stages 1 and 2)( Fig 3 ), no surprise since stages were defined partly by size, which these continuous variables measured. But differences in mean dimensions seem more an artifact of analysis than a property of stages themselves. Height in adult males is a continuous variable best portrayed in continuous terms. Nothing about the distribution of height in a sample of men prevents defining types or stages of height (e.g., short, medium and tall) as arbitrary intervals of the range that surely would differ in mean value as Callahan stages differ in mean size. But unless specified on valid grounds boundaries between such types are arbitrary, variation as great within as between them. There are no types of heights, but continuous variation in the variable “height.” Any types defined may impart some descriptive convenience but are not faithful to the nature of height variation.
10.1371/journal.pone.0170947.g003
Fig 3
Callahan Stage 2–4 preforms boxplots by stage: a. weight; b. JTI.
If reduction stages are valid, then they differ among themselves as groups but specimens within a stage do not differ appreciably. This proposition can be tested in Callahan data, for instance by subdividing each stage into two equal halves (e.g. dividing Stage 2 preforms into Substage 2A, its heaviest half, and 2B, its lightest half) by weight. If stages are valid, then there should be differences between them. There are in Callahan data. But weight also differed within stages, i.e. between equal but arbitrary subdivisions of each stage (F = 28.1, p <.01; all pairwise LSD p <.05 except between noncontiguous Stages 2A and 4B [p = .42] and 3A and 4A [p = .26], which also demonstrates the complex, not deterministic, relationship between preform size and degree of reduction). Weight differed between successive substages as it did between stages, because both arbitrarily divide a continuum of variation. As with “types” of height, metric differences within Callahan’s stages were as great as those between them. Elsewhere, discriminant analysis using stages misclassified about 30% of Callahan Stage 2–4 specimens ([ 16 ], 321–323). Defined stages are resistant to statistical analysis.
As above, Callahan’s ([ 1 ], Table 10) stage scheme was based partly on what are reported as critical values of the width-thickness ratio. Viewing these values, arguably, as strict criteria, eight of Callahan’s 12 Stage 2 bifaces (including two broken ones not otherwise included in analysis here) fell beneath or above that stage’s range, nine of 18 Stage 3 preforms outside of that stage’s range, and that stage’s range, and 19 or 34 Stage 4 specimens outside of that stage’s range ( Table 2 ). Thus, 38 of 64 specimens yielded width-thickness ratios that fell outside their stage’s defined range. Even if stage boundaries were broadened (for Stage 2 2.0±0.25 to 3.0±0.25, for Stage 3 3.0±0.25 to 4.0±0.25, for Stage 4 4.0±0.25 to 5.0±0.25), 28 of 64 preforms still would fall outside the specified range.
10.1371/journal.pone.0170947.t002
Table 2 Callahan preforms by stage within or beyond the stage’s width-thickness range.
Width-thickness
n within
STAGE
Range
n<minimum
range
n>maximum
% w/in range
2
2.0–3.0
1
4
7
33
3
3.0–4.0
3
9
6
50
4
4.0–5.0
9
13
12
38.3
Preform analysis: Continuous variation
Distributions of individual continuous variables and ratios did not reveal empirical gaps or modes that might signify discrete stages, nor did analysis of the width/thickness ratio suggest complete validity of the stage model as conceived by Callahan. It is worth studying bivariate patterns of variation as well, either by assigned stage or across entire preform datasets. For this purpose, both the variables plotted and the form and nature of any relationship between them are continuous.
Besides ratios and weight, continuous variables that measure preform size or form include linear dimensions (minimally length, width and thickess) and weight. Here, I use a composite measure of linear dimensions produced by the ordination method principal-components analysis (PCA) of length, width and thickness. The purpose is not detailed examination of correlation among the three variables. This would discover merely the obvious, because PCA of linear dimensions typically yields a first component on which all major dimensions load positively, interpretable as a gestalt size measure (e.g., [ 54 ], 188). Instead, the purposes are first to reduce dimensions to that single size measure and second to determine if there are other significant components of variation (i.e. components whose eigenvalues>1) besides overall size.
PCA and regression
PCA was conducted in SPSS using the correlation matrix to minimize scale differences between linear dimensions. It produced a single significant component, PC1, that accounted for 63.3% of variance and on which all three dimensions loaded strongly and positively. As above, PC1 can be interpreted as a size measure on which all three dimensions show high positive loadings ( Table 3 ). (See [ 50 ], 207 and [ 55 ], 142 for similar results and interpretations, using different sets of variables.) Again, it is no surprise that weight (r = .97, p<.01) and JTI (p = .83, p<.01) both correlated strongly with PC1 (Figs 4 and 5 ). Yet it is significant that they varied with PC1 such that lower PC1 scores and therefore smaller size indicated lower weight and JTI values. These results corroborate the interpretation of continuous PC1 as a size dimension.
10.1371/journal.pone.0170947.t003
Table 3 Summary of PCA of Callahan Stage 2–4 Data.
cumulative
Variable
Loadings
Component
eigenvalue
%variance
%variance
length
width
thickness
1
1.900
63.3
63.3
0.797
0.826
0.764
2
0.611
20.4
83.7
3
0.490
16.3
100.0
10.1371/journal.pone.0170947.g004
Fig 4
Callahan Stage 2–4 preforms, weight vs. PC1: a. actual data; b. hypothetical distribution if weight covaried with size measured by PC1.
10.1371/journal.pone.0170947.g005
Fig 5
Callahan Stage 2–4 preforms, JTI vs. PC1.
Figs 4 and 5 are ordered so that weight and JTI on the one hand, and PC1 on the other both increase from the origin. Because reduction, trivially, reduces preform size, progress in the reduction continuum is read from upper right to lower left in the figures. In this context, however, the plots have three other significant aspects. First, both variables exhibited continuous distributions, not discrete clusters of observations. Continuous variables can form discrete or quasi-discrete clusters; these did not. Quadratic regression nicely described the relationship between weight and PC1 (r 2 = .92), linear regression only slightly less so (r 2 = .86) despite being more robust because it requires fewer parameters ( Fig 6 ). Linear regression includes two parameters: B0, an estimate of the constant or y-intercept, and B1, an estimate of the slope of the regression line that measures the rate of change in the dependent variable with unit change in the independent. Regression is a continuous model of variation and relationship as useful in analysis of biface reduction as it was to flake debris (e.g,. [ 37 ]). Jointly in size and weight, Callahan stages comprise a continuum of variation.
10.1371/journal.pone.0170947.g006
Fig 6
Callahan Stage 2–4 preforms, quadratic (dashed line) and Linear (solid line) Models of weight vs. PC1.
Second, preforms did not sort neatly by stage in the weight:PC1 plot. Rather, there was considerable overlap among stages. Although the largest specimens were Stage-2 preforms, other Stage-2 specimens distributed across roughly three-quarters of the range in both variables and several smaller than many Stage-3 and some Stage-4 preforms. Similarly, Stage-3 preforms were almost freely interspersed on scatters. Stage-4 specimens occupied narrower ranges, but they too were distributed across roughly one-half to two-thirds of both variable’s ranges. (See [[ 46 ], Fig 8]] and [[ 31 ], Fig 3.3]] for similar plots with considerable overlap between “stages” in other preform continuous variables.) Had stages strictly segmented the continuous distributions of weight and PC1, the pattern shown in Fig 4b would be found. Jointly in size and weight, Callahan stages somewhat arbitrarily parse the continuum of variation evident in preforms.
Third, following Johnson’s (1981) logic, later stages or segments of preform finishing involve edging and thinning more than the general and size/mass reduction of earlier stages. Similarly, Smallwood’s ([ 56 ], 155–162; [ 57 ], 703) study of production-stage bifaces documented proportionally greater reduction in thickness than width in middle and late segments of the reduction continuum. No more than the progressive resharpening of finished and extensively used tools (e.g. [ 58 ]) is reduction during the production process a matter of proportional decrease such that later “stages” are scale models of earlier ones; instead it is allometric, involving change in shape and proportion with change (in stone tools, necessarily reduction) in size. Particularly significant is that regression’s B1 coefficient measures the slope, and therefore rate, of covariation between weight or JTI and PC1 ( Fig 7 ). Following this logic, closer examination of these relationships is warranted.
10.1371/journal.pone.0170947.g007
Fig 7
Callahan Stage 2–4 preform variables vs. PC1 by stage: a. weight; b. JTI.
Covariation of weight and JTI with PC1
Table 4 reports specimens by assigned stage in Callahan’s replications. Least-squares regression of weight and JTI upon PC1 was conducted separately by defined stage. Unless otherwise indicated, in this and all subsequent analysis regression results were significant (i.e. p≤.05) and slope coefficients declined by stage. Differences in regression coefficients were evaluated by Student’s t using the method of Statistical Consulting Group [ 59 ]( Table 5 ). Slope differences therefore are relevant for their magnitude and, being consistent in direction, t values are reported without sign.
10.1371/journal.pone.0170947.t004
Table 4 Preforms by Callahan Stages 2–4 in study datasets.
Count
by
Stage
Dataset
2
3
4
Sources
Callahan replications
10
18
34
[ 1 ]
Pelegrin & Chauchat
10
10
8
[ 62 ]
Adams
23
6
0
[ 21 ]
Thunderbird
20
0
0
[ 53 ]
Gault
8
8
0
[ 18 , 43 ]
10.1371/journal.pone.0170947.t005
Table 5 Summary of test of similarity in B1 coefficients between Callahan stages.
WEIGHT
WEIGHT
WEIGHT
Callahan
Stage 2
Callahan
Stage 3
Dataset
stage
B1
t
p
t
p
Callahan
2
73.6
---
---
Callahan
3
57.9
2 . 08
0 . 05
---
---
Callahan
4
46.8
5 . 92
< . 01
2 . 55
0 . 01
JTI
JTI
JTI
Callahan
Stage 2
Callahan
Stage 3
Dataset
stage
B1
t
p
t
p
Callahan
2
0.39
---
---
Callahan
3
0.32
0.74
0.46
---
---
Callahan
4
0.1
4 . 13
< . 01
3 . 72
0 . 01
Regression results for Callahan’s Stages 2–4 data, with standard errors of estimates in parentheses, are ( Fig 7b ):
Callahan Stage 2 weight = 101.3 + 73.6 * PC 1
(6.3) (5.8)
Callahan Stage 3 weight = 110.1 + 57.9 * PC 1
(3.8) (4.4)
Callahan Stage 4 weight = 99.8 + 46.8 * PC 1
(1.7) (2.1)
Weight patterns clearly, and slope coefficients differ significantly ( Table 5 ), with PC1 by stage. This finding implicates reduction allometry, “stages” patterning differently such that weight decreases with size (or decreases during reduction, to read Fig 7b from right to left) at lower rates in successive “stages”. That is, weight declines with size at ever-declining rate by “stage,” i.e., arbitrary subdivision of the reduction continuum. This is no trivial statistical observation. Rather, the slope of the regression of weight upon PC1 becomes a ratio-scale estimate of degree of reduction in ranges or segments of the reduction continuum, an estimate that complements and expands ordinal-scale stage classification. Jointly in size and weight, continuous analysis of Callahan stages captures the allometry that governs biface reduction, and B1 serves as the allometry coefficient. Useful concepts like “tempo” of reduction ([ 57 ], 695, 703–705; [ 60 ], [ 61 ], 653;) measure the same process although as reported they reduce continuous variation to ranges of values that correspond to assigned preform stages.
Results for JTI are somewhat ambiguous. As above, JTI is not independent of weight (r s = .90 p<.01). Nor, obviously, is it independent of PC1 which summarizes the scores of the same major dimensions that form JTI’s denominator. Regression of JTI upon PC1 by stage patterned consistently ( Fig 7b ):
Callahan Stage 2 JTI = 1.43 + 0.38 * PC 1
(.11) (.09)
Callahan Stage 3 JTI = 1.49 + 0.27 * PC 1
(.05) (.05)
Callahan Stage 4 JTI = 1.24 + 0.07 * PC 1.
(.03) (.03)
However, slope coefficients did not differ significantly between Stages 2 and 3 ( Table 5 ). Moreover, for Stage 4 the scatter was diffuse, the resulting slope extremely low; although results are significant, they do not account for most variation (r = .44, r 2 = .19, p = .01). JTI imperfectly models the reduction process in these data.
z-score transformations
Callahan and other datasets to which it is compared below scale somewhat differently in the range of original variables, which complicates direct comparison of their regression results. One reason is that Callahan data span Stages 2–4, other datasets only two stages, usually 2 and 3. The range of variation therefore is somewhat greater in Callahan’s than other datasets. To minimize scale differences that can affect PCA, data were transformed to standard scores (“z-scores”) for each dataset separately from others. z-scores rescale raw data to units of each variable’s standard deviation, their sign indicating that they are above (positive) or below (negative) the mean. Standardized weight is denoted z-weight. Because PCA proceeds by standardizing variables, component eigenvalues and variable loadings are identical to PCA of original variables, and the single resulting significant component is denoted z-PC1. JTI is a ratio between two original variables, thus a derived variable. JTI may be standardized as was weight, but interpretation of a standardized ratio is unclear. Therefore, JTI was not standardized, so patterned and scaled with z-PC1 exactly as it did with PC1.
As in original data, PCA of Callahan data yielded a single significant component on which z-length, z-width and z-thickness all loaded strongly and positively, so again is a general size component. Callahan’s Stages 2–4 data patterned consistently as follows ( Fig 8 ):
Callahan Stage 2 z − weight = −0.14 + 1.38 * z − PC 1
(.20) (.16)
Callahan Stage 3 z − weight = 0.13 + 0.94 * z − PC 1
(.10) (.10)
Callahan Stage 4 z − weight = −0.15 + 0.69 * z − PC 1
(.10) (.04)
10.1371/journal.pone.0170947.g008
Fig 8
Callahan Stage 2–4 preforms, z-weight vs. z-PC1 by stage.
Such that regression intercepts approached or fell within one standard error of 0 and slope declined with stage as reduction progressed. All B1 coefficients differed significantly from one another (for Stages 2 and 3, t = 2.20 p = .04; for Stages 3 and 4, t = 2.61, p = .01). Trivially, Callahan z-weight against z-PC1 scaled differently from original variables; more significantly, transformed variables patterned as did original ones and yielded equally significant regression results. As above, because JTI was not standardized and PC1 and z-PC1 are identical, there is no need to recalculate the regression of JTI upon z-PC1. Also because JTI was not standardized, here and in subsequent analyses its regression intercepts considerably exceed 0.
At first glance, results might be thought to support the existence of stages because slope coefficients differed significantly between stages in regression of z-weight upon z-PC1. But preceding analysis of univariate distributions and ratios and of bivariate plots suggested instead that statistical difference may owe to stages’ arbitrary parsing of a reduction continuum. As in earlier analysis, therefore, the validity of stages must be tested in multivariate data.
Subdividing each Callahan stage as in preceding analysis is questionable, as it would involve multivariate analysis of very small data subsets. Instead, if reduction is continuous then overlapping subsets of stages should pattern in regression of z-weight upon z-PC1 score such that B1 should differ between the combinations and in value fall between the B1 values of individual stages. To test these expectations, Callahan preforms were combined into two overlapping subsets, Stages 2 and 3, and Stages 3 and 4. z-scores were calculated separately for each subset and then ordinated by PCA. As expected, B1 coefficients of the combined Callahan subsets patterned consistently, a higher value and slope characterizing combined Stages 2 and 3 ( Fig 9 ). Slope coefficients differed significantly from one another (for combined Stages 2–3, B1 = 1.06; for combined Stages 3–4, B = 10.83; t = 2.57, p <.01), despite the two subsets sharing 18 Stage-3 specimens, and coefficients fell between the B1 values of the respective individual stages. Significant difference between stages or stage groups can occur even when such stage groups share many specimens. This conclusion implicates meaningful variation within and between “stages” that only a continuous approach to biface production can reveal.
10.1371/journal.pone.0170947.g009
Fig 9
Callahan Stage 2–4 preforms, z-weight vs. z-PC1 by combined stages 2–3 and 3–4.
Other datasets
Callahan’s [ 1 ] data by stage were not linked to specific bifaces such that each specimen can be traced through the reduction continuum. Lacking this degree of experimental control, it is unclear if pattern in continuous data reflects each preform’s reduction trajectory or is the result of data aggregation. Datasets in which the same specimens can be followed through successive reduction stages should be examined for possibly similar patterning.
For instance, Pelegrin and Chauchat ([ 62 ], Tables 1–2) reported all relevant variables for experimental replicas of Paiján points, a late Pleistocene type found on the central Andean coast. Although the finished point is stemmed, their replicated specimens remain unstemmed preforms through the stages analyzed here. Pelegrin and Chauchat ([ 62 ], 370) defined sequential Stages 1–4 that followed the flake blank ( soporte inicial ). Thus, their flake blank is Stage 0 whereas Callahan’s blank is Stage 1. More importantly, individual specimens can be followed across all stages. If analysis of Pelegrin and Chauchat data resembles Callahan’s then patterning in the latter is corroborated and the approach is worth applying to archaeological data.
Pelegrin and Chauchat made 13 preforms, 10 of whose dimensions they reported across three stages, only six of them to the fourth stage ([ 62 ], Table 2, where data from subdivided Stage 2 are from Stage 2b); they reported dimensions for two additional specimens at their fourth stage, for stage totals of 10, 10 and eight preforms. Fig 10 plots weight against stage for Pelegrin and Chauchat data; because the two largest specimens are disproportionately large at Stage 1, Fig 10b omits them to more clearly show patterning in other specimens (although they were retained for all analysis reported below). All preforms followed the same pattern. As in Callahan data, PCA of Pelegrin and Chauchat’s ([ 62 ], Tables 1–2) Stages 2–4 data yielded one significant component on which z-length, z-width and z-thickness all loaded strongly. z-weight plotted against z-PC1 by specimen ( Fig 11 ) yielded slope coefficients that ranged from 0.32 to 1.68. (Each regression involved only three data points, which for most analytical purposes are too few to lend much confidence to results. Instead, regression serves merely to corroborate patterning in Callahan data using a dataset where each preform’s dimensions and weight can be traced across “stages.”) Callahan Stage 2–4 values fall comfortably within this range, which suggests robust patterning between datasets. JTI was not calculated for Pelegrin and Chauchat specimens.
10.1371/journal.pone.0170947.g010
Fig 10
Pelegrin and Chauchat preforms, weight by stage: a. all specimens; b. two largest specimens omitted.
10.1371/journal.pone.0170947.g011
Fig 11
Pelegrin and Chauchat preforms, z-weight vs. z-PC1 by stage.
For further comparison to Callahan’s data, therefore, analysis can include Chauchat and Pelegrin’s as well as relevant North American Paleoindian datasets ( Table 4 ). The latter include:
Adams. Sanders [ 21 ] analyzed the Adams Paleoindian assemblage from western Kentucky. Adams was a Clovis workshop adjacent to outcrops of Ste. Genevieve chert. Sanders closely followed Callahan’s stage scheme, at least through the Stage-4 data analyzed here ([ 21 ], 82–121), and reported length, width, thickness and weight of Stage 2–4 preforms ([ 21 ], Tables 10–12). Unfortunately, no Stage 4 preforms and only six of more than 20 Stage 3 preforms were intact, so Sanders’s data are mostly from Stage 2.
Thunderbird. Verrey [ 53 ] described a late Paleoindian cache, apparently of local jasper, from Feature 17 at Thunderbird. He considered “all but 3” ([ 53 ], 4) to be Callahan Stage-2 bifaces and reported 23 specimens including two biface/scrapers and one large lateral fragment, as well as metric data on cache specimens (R. Verrey, personal communication 18 November 1985). I assume that the latter three items were not Callahan Stage 2 preforms, so confine analysis to the other 20 preforms as Callahan Stage 2 specimens.
Gault. As above, Dickens [ 18 ] assigned Gault preforms to Callahan stages. For 16 specimens that fall in Callahan Stages 2–4, those assignments can be cross-referenced to Waters et al.’s ([ 43 ], Table 20) data.
All datasets appear in S1 Table . Other sources consulted could not be included. Despite its extensive documentation [ 23 , 30 , 63 , 64 ], Anzick’s involved collections history prevented its inclusion. Several cataloguing systems complicated the process of reconciling specimens and their metric data between sources. Although the Anzick preform assemblage is large, relatively few retained all dimensions necessary for analysis here. Kilby could not weigh specimens, so estimated weight as a function of major dimensions, reasonable but not strictly equivalent to the direct measurement of weight in other assemblages. Jones ([ 30 ], 164) considered some Anzick specimens to be finished and used tools, not preforms. Morrow’s ([ 20 ], Appendices A-E) highly detailed dataset did not include weight. Bamforth ([ 24 ], Table 4.2) reported length, width, thickness and several ratios but not weight or stage assignment for Mahaffy preforms. Bement ([ 60 ],Table 5.1) reported all relevant dimensions and weight for JS cache preforms, but not stage assignments in Callahan’s or other schemes, nor did Jennings ([ 61 ]) report specimen dimensions or Callahan stage assignments. Muñiz ([ 17 ], Table 7.1) also reported dimensions but not weight for CW cache preforms. Although advocating a continuum-based view, he also devised a classification system different from Callahan’s, based on the treatment of preforms as cores. Muñiz assigned all 11 CW preforms to a single stage, his “late-stage bifacial core” ([ 17 ], 116). Huckell [ 65 ] did not report either dimensions or weight for the Beach cache. As useful as these sources are, none provides data that are directly comparable to Callahan’s or others’ used here.
Analyzing combined data
z-weight
As in Callahan data, PCA analysis of each separate source yielded a single significant component on which all dimensions loaded positively so is interpretable as gestalt size, all regressions by stage were significant and, where relevant, successive stages had lower slope coefficients. Adams and Gault Stage 2–3 preforms showed similar patterns to Callahan data (Figs 12 and 13 ; Table 6 ). Overall, therefore, patterning was robust and clear: as reduction progressed, z-weight declined with z-PC1 at ever-declining rate. Again, this pattern was revealed only in continuous data, and itself is a continuous relationship that exclusive focus on qualitative variation would not reveal.
10.1371/journal.pone.0170947.g012
Fig 12
Adams Preforms, z-weight (a) and JTI (b) vs. z-PC1.
10.1371/journal.pone.0170947.g013
Fig 13
Gault Preforms, z-weight (a) and JTI (b) vs. z-PC1.
10.1371/journal.pone.0170947.t006
Table 6 Summary of test of similarity in B1 coefficients between Callahan stages and other data sources. Italics and underscores indicate t and p values significant @ .05, also shaded for ease of reference.
z-weight
z-weight
z-weight
Callahan
Stage 2
Callahan
Stage 3
Callahan
Stage 4
Dataset
stage
B1
t
p
t
p
t
p
Callahan
2
1.37
---
---
Callahan
3
0.94
2.20 _
0.04 _
---
---
Callahan
4
0.69
4.33 _
0.00 _
2.61 _
0.01 _
---
---
Adams
2
1.02
2.23 _
0.03 _
0.63
0.54
4.13 _
0.00 _
Adams
3
0.69
3.57 _
0.04 _
1.39
0.17
0.20
0.84
Thunderbird
2
0.97
2.77 _
0.01 _
1.04
0.30
1.49
0.11
Gault
2
0.89
2.58 _
0.02 _
0.32
0.75
2.49 _
0.02 _
Gault
3
0.54
2.33 _
0.04 _
1.03
0.32
0.78
0.44
P&C 1
2
1.30
0.21
0.84
1.47
0.15
3.72 _
0.00 _
P&C
3
0.62
4.86 _
0.01 _
2.22 _
0.04 _
1.06
0.30
P&C
4
0.18
4.35 _
0.01 _
2.49 _
0.02 _
3.62 _
0.00 _
JTI
JTI
JTI
Callahan
Stage 2
Callahan
Stage 3
Callahan
Stage 4
Dataset
stage
B1
t
p
t
p
t
p
Callahan
2
0.38
---
---
Callahan
3
0.27
0.23
0.82
---
---
Callahan
4
0.07
4.33 _
0.00 _
0.43
0.67
---
---
Adams
2
0.39
1.46
0.16
1.62
0.11
1.93
0.10
Adams
3
0.18
5.73 _
0.00 _
8.12 _
0.00 _
8.42 _
0.00 _
Thunderbird
2
0.49
0.92
0.37
2.06 _
0.05 _
5.43 _
0.00 _
Gault
2
0.35
0.15
0.88
0.60
0.55
2.98 _
0.01 _
Gault
3
0.35
0.06
0.95
0.34
0.74
1.52
0.14
Yet pairwise comparison of datasets did not always pattern consistently. Regression slopes for Callahan Stage 3 and Adams Stage 2 were nearly identical, as were slopes for Callahan Stage 4 and Adams Stage 3 ( Table 6 ). That is, Callahan Stage 3 patterned in slope as did Adams Stage 2. Callahan Stage 2 had a higher slope than any other regression, and Callahan and Adams data did not correspond directly by stage, Adams’ values instead being offset by one stage relative to Callahan’s. Generalizing from this pair, Callahan’s Stage-2 slope coefficient was considerably higher than most other datasets’ (Pelegrin and Chauchet’s Stage 2 excepted) and values for its successive stages often approximated those for other datasets’ next-lowest stage. Possibly Callahan data differed in some unrecognized way from other sources, or the large Callahan sample and its wider range of technological and size variation, roughly encompassed by its three stages compared to one or two in other sources, produced the difference. Alternatively, ambiguity in Callahan stage assignments accounts for the difference.
Whatever the cause, this difference suggests unexplained variation between datasets. To interrogate this variation, the Callahan dataset’s wide range of “stage” assignments and variation justifies its examination first. Separate PCA and variable standardization were performed on Callahan Stages 2–3 and Stages 3–4 combined in Callahan data alone. That is, only preforms in Stages 2–3 were submitted for z-score PCA analysis, then separately only preforms from Stages 3–4 were analyzed. This treatment approximates the two-“stage” range of other datasets. If the separate subdivisions of Callahan data pattern and scale with reduction as do other datasets, then differences between datasets owe to Callahan’s wider range of data variation. In both Callahan subsets, z-weight was regressed against z-PC1 by stage. In Callahan’s complete dataset, three slope coefficients ranged from 0.69 to 1.37 ( Table 7 ). In the subsets of combined Stages 2–3 and combined Stages 3–4, four slope coefficients together ranged from 0.79 to 1.23, a modest reduction. In each paired subset, the two successive stages differed significantly in slope (in the Stage 2–3 subset, t = 2.17, p = .04; in the Stage 3–4 subset, t = 2.03, p = .05), as they did in the complete Callahan dataset. However, slope coefficients did not differ significantly by stage between Callahan’s complete dataset and paired-stage subsets ( Table 7 ). Variation in Callahan’s data is reduced only modestly by this treatment.
10.1371/journal.pone.0170947.t007
Table 7 Summary of test of similarity in B1 coefficients between original Callahan data and separate Callahan combinations of Stages 2–3 and 3–4.
Stage
Callahan
Stage 2
Callahan
Stage 3
Callahan
Stage 4
Subset
stage
B1
t
p
t
p
t
p
none
2
1.37
none
3
0.94
none
4
0.69
2&3
2
1.23
0.65
0.52
2&3
3
0.84
0.70
0.49
3&4
3
1.01
0.43
0.67
3&4
4
0.79
1.58
0.12
Congruence in slope by stage was inconsistent in separate analysis of other datasets, although Callahan and Pelegrin and Chauchat Stage 2 sets were similar, as was significant difference between stages. For instance, Adams Stage 2 had a significantly higher slope than did Adams Stage 3 ( Table 8 ); although Gault Stage-2 preforms had a higher slope than did its Stage-3 ones, that difference was not significant. Also, although Adams Stage 2 and Stage 3 differed significantly in slope, Adams Stage 2 and Gault Stage 3 did not, despite the greater absolute difference in slope coefficients.
10.1371/journal.pone.0170947.t008
Table 8 Summary of test of similarity in B1 coefficients among data sources, excluding Callahan. Italics and underscores indicate t and p values significant @ .05, also shaded for ease of reference.
z-weight
z-weight
z-weight
z-weight
z-weight
Adams 2
Adams 3
Thunderbird
Gault 2
Gault 3
Dataset
stage
B1
t
p
t
p
t
p
t
p
t
p
Adams
2
1.02
--
--
Adams
3
0.69
2.36 _
0.03 _
--
--
Thunderbird
2
0.97
0.42
0.68
2.73 _
0.01 _
--
--
Gault
2
0.89
1.04
0.31
1.59
0.14
0.77
0.45
--
--
Gault
3
0.54
1.55
0.13
0.75
0.47
1.80
0.09
1.23
0.24
--
--
P&C 1
2
1.30
1.37
0.18
1.65
0.12
2.23 _
0.04 _
1.36
0.19
1.10
0.29
P&C
3
0.62
3.55 _
0.01 _
0.86
0.41
2.40 _
0.02 _
3.57 _
0.02 _
0.42
0.68
P&C
4
0.18
3.52 _
0.01 _
12.12 _
0.00 _
4.29 _
0.00 _
3.55 _
0.00 _
1.99
0.07
JTI
JTI
JTI
JTI
JTI
Adams 2
Adams 3
Thunderbird
Anzick
Gault 2
Dataset
stage
B1
t
p
t
p
t
p
t
p
t
p
Adams
2
0.39
--
--
Adams
3
0.18
1.78 _
0.09 _
--
--
Thunderbird
2
0.49
1.69
0.10
0.83
0.42
--
--
Gault
2
0.35
0.36
0.72
0.74
0.48
0.68
0.51
0.87
0.40
--
--
Gault
3
0.35
0.12
0.91
0.57
0.58
0.39
0.71
0.74
0.47
0.03
0.98
1 Pelegrin & Chauchat
More robust patterning might be found by pooling data by assigned stage from all datasets. This treatment has the added virtue of attenuating the Callahan dataset’s wide range of variation. In all datasets combined, results are clear ( Fig 14a ):
Combined Stage 2 z − weight = 0.01 + 1.05 * z − PC 1 Eq. 1
(.06) (.05)
Combined Stage 3 z − weight = 0.02 + 0.82 * z − PC 1 Eq. 2
(.06) (.06)
Combined Stage 4 z − weight = −0.14 + 0.64 * z − PC 1 Eq. 3
(.04) (.04)
10.1371/journal.pone.0170947.g014
Fig 14
Combined dataset, z-weight (a) and JTI (b) vs. z-PC1.
Results are significant for several reasons. First, Stage 2 preforms patterned in z-weight against z-PC1 essentially at unit rate (i.e., B1≈1), successive “stages” at proportionally-lower rate. That is, early segments of the reduction continuum involve unit reduction in weight and volume, while later segments progressively decline in weight reduction relative to volume reduction. Accordingly, B1 serves as an allometric coefficient that declines steadily as reduction advances. Second, in the combined dataset—though not necessarily in individual ones, as separate analysis above demonstrated—“stages” apportion this allometric variation fairly proportionally, B1 coefficients declining by nearly equal magnitude by stage. Third, all stages’ B0 coefficients →0, although Stage 4’s is slightly but significantly negative. Therefore no constant, either positive or negative, is needed to account for variation between stages. Magnitude of decline of z-weight with z-PC1 was roughly constant in subsequent assigned stages, as successive B1 values were evenly spaced. Finally, all successive stage pairs differed significantly in B1 (between stages 2 and 3, t = 2.42 p = .02; between stages 3 and 4, t = 2.44 p = .02).
Unlike analysis of separate datasets, that is, pattern and scale both were consistent in combined data. B1 consistently declined as reduction advanced, and rate of decline consistently declined by near-constant magnitude across stages. Whatever the differences in pattern and scale by original sources, combined data are consistent in both respects. Differences in stage assignment between datasets effectively cancel out in combined analysis, where more robust results are consistent not only in pattern but also scale.
JTI
The JTI index patterned with PC1 somewhat more ambiguously (again, JTI was not computed for Pelegrin and Chauchat replicas). No JTI slope coefficients were statistically similar to their corresponding stages in Callahan’s own data ( Table 6 ). Gault Stages 2 and 3 were identical in slope but different in intercept or location. Excluding Callahan data, the only near-significant differences in JTI slope upon z-PC1 were between Adams Stage 2 and its Stage 3 ( Table 8 ). Otherwise, JTI slope coefficients did not differ significantly although in all cases Stage 2 values always were ≥ Stage 3 values, and Stage 2 values ranged from 0.35 upward, Stage 3 from 0.14 to 0.35.
In all datasets combined, results were as follows ( Fig 14b ):
Combined Stage 2 JTI = 1.81 + 0.38 * z − PC 1
(.06) (.06)
Combined Stage 3 JTI = 1.59 + 0.22 * z − PC 1
(.05) (.05)
Combined Stage 4 JTI = 1.24 + 0.07 * z − PC 1
(.03) (.04)
Difference in Stage-2 and -3 slopes approached but did not reach significance (t = 1.75, p = .08) but Stage-3 and -4 slopes differed significantly (t = 2.25, p = .03).
Discussion
Continuous measures z-weight and JTI patterned with overall preform size as given by z-PC1. At successive Callahan “stages,” most regression slopes against z-PC1 declined. That much is consistent. However, individual datasets did not scale similarly because the same Callahan “stages” sometimes yielded significantly different z-weight slope coefficients. That much is ambiguous. In particular, Callahan’s own dataset differed significantly in z-weight slope coefficient from most others, which in turn scaled nearer to one another.
Why the scale ambiguity between datasets? One possible reason is that the technological criteria by which Callahan stages are defined are independent of size and continuous measures of reduction like weight and JTI. Another is that stage assignments and the criteria on which they are based are not replicable between analysts. As Dickens [ 18 ] suggested, one archaeologist’s Stage 2 may be another’s Stage 3. Complex patterns of both discrete and continuous variation are imperfectly captured by “stage” assignment, and when using stages datasets may not scale consistently even if they pattern clearly.
Among other things, the scale differences between individual datasets suggest that none, including Callahan’s, is an ideal master sequence against which future data should be compared. Fortunately, results of combined analysis of all datasets by stage assignment were consistent in both pattern and scale and, because Eq. 1–3 B1 values were fairly evenly spaced, implicate constantly declining rate of decline of z-weight upon z-PC1. Results support a continuous approach to both the reduction process and its analysis. They also suggest that the ambiguity in stage assignment noted above is offset in larger, combined datasets. To this extent, combined datasets like the final one analyzed here are better reference datasets for future studies. In general, JTI results were similar but statistically less robust. Further research is needed to determine how well JTI performs in reduction analysis.
Based on this analysis, the slope coefficient from regression of z-weight upon z-PC1 conveys useful information. Calculated easily in SPSS or other common platforms for any preform assemblage that reports length, width, thickness and weight, it can aid future studies in several ways. For instance, possible stage assignments in other preform assemblages, particularly relative homogeneous ones like caches, can be tested by analyzing them as here; assignments are validated if they approximate the regression results of Eq. 1–3. For combined datasets, that is, results scale the various stage assignments found in the literature to a single reduction continuum. If datasets analyzed here are representative--a plausible tentative hypothesis considering the range of toolstones, contexts, and empirical and experimental sources encompassed--they comprise a reasonable continuum of biface-production variation between original blanks and nearly finished tools, by which ambiguous, ordinal-scale stage assignments can be more precisely calibrated. Even more useful is to abjure qualitative stage designations and instead analyze relatively homogeneous subsets of larger preform assemblages.
Such assemblages (e.g., from caches), can be analyzed as here. Resulting regression coefficients then can be used as continuous measures of degree of reduction. Coefficients would be particularly useful in detailed models of biface reduction across the landscape, including behavioral-ecology ones (e.g. [ 33 – 34 , 66 ]), thereby resolving apparently qualitative production stages or steps to the continuous process that underlies them. For instance, instead of assigning, say, two preform assemblages to Callahan Stage 2 and a third to Stage 3, analysis could demonstrate that one lies at the point where z-weight’s B1 coefficient upon z-PC1 is, say, 1.07, a second at 0.94 and a third at 0.61. In this way, all three can be situated at different, relatively precisely calibrated segments of the reduction continuum modeled by Eq. 1–3, and measure biface reduction across landscapes at scales commensurate with continuous behavioral models (e.g. [ 34 , 67 ]). Similarly, preform assemblages that can be subdivided by context, toolstone or technological criteria can be compared and contrasted for the segments or positions in reduction continua that regression analysis indicates.
The stage concept remains an ambiguous descriptive and perhaps analytical heuristic until a sufficient set of valid, replicable and usually discrete technological criteria can be specified. They must be valid in the sense that the theory of brittle fracture and/or detailed replication experiments demonstrate the criteria’s contribution to the patterning in data by which stages emerge. They must be replicable in the sense that all analysts who apply them to the same experimental or empirical preforms would, as a result, make the same stage assignments.
At the same time, continuous reduction measures like weight, JTI, FI and others (e.g., 47–49) should be recorded on the same replicated specimens. One promising area of research is scanning methods that might make efficient and rigorous the measurement of cross-section variables like area and symmetry and faceting variables like number, pattern and size variation in facets (e.g. [ 68 – 71 ]. If such measures covary strongly with levels or states of qualitative technological criteria and if they contribute to the definition of relatively cohesive groups of specimens distinguished from others by considerable distance in bivariate or multivariate space, then they corroborate the value of the stage approach. Measures may not covary strongly with levels or states of qualitative criteria but instead pattern independently of them, as significant covariation between weight and JTI on the one hand, and size as measured by PC1 on the other. In that case they corroborate the independent value of continuous measures and cast doubt on the validity of stages. If so, we should use controlled replications to define entirely continuous reduction models that efficiently summarize multivariate patterns in dimensions, weight, gestalt size as measured by PC1, JTI and other reduction measures.
Until then, the most conservative approach is dual. For comparison to earlier studies, and for somewhat ambiguous descriptive convenience, the interpretive construct “stage” may be justified. Yet it is essential to report measures like size, weight, and JTI and to analyze continuous variation in them. At least crudely, the two approaches can be compared and possibly cross-validated by compiling the proportion of preforms assigned to Callahan stages and also plotting z-weight or JTI against z-PC1. Within assigned “stages,” regression slope coefficients might approach values found here for the comparable “stages” either in Callahan or other datasets. Across them, pattern in slope coefficients should approximate those found here. As proportions shift progressively toward more advanced reduction, slope of regression of z-weight or JTI upon z-PC1 should decline. This possibility can be tested in the comparison of preform assemblages from quarries to workshops or residences [ 34 ], between time periods, or otherwise across space as the reduction continuum can be segmented. As much as possible, such analyses should control for toolstone, time-space, and industrial differences.
Even if stage and continuum approaches are complementary, the latter possesses some unique advantages. By definition, continuum models document fine-scale variation in and between assemblages. Such variation can owe to comparable adaptive or technological variation, or chronological or cultural change. Continuous variation also can be interrogated for sometimes fine-grained conformity to continuous-scale behavioral models [ 34 , 67 , 72 ] in ways that coarser stage models cannot.
Although preforms in the reduction process were the focus of this study, of course finished bifaces continued to be reduced in use for resharpening and repair. Archaeologists have devised a range of measures at the level of individual specimens to track the allometric variation produced by these practices [ 47 , 54 , 58 , 73 – 76 ], as well as variation that is not allometric [ 77 ]. It is worth examining assemblage-level variation in finished points by such measures to complement the assemblage-level variation in reduction preforms documented here, mindful that different measures can be suitable for preforms and finished points.
Conclusion
None of this is to criticize biface-production stage approaches generally, nor to suggest that reduction is continuous in all salient respects. Callahan’s ([ 1 ], Table 10) model includes categorical variables that may pattern with assigned stage. Unfortunately, some are difficult to replicate or of unknown relevance (e.g., regularity of outline, “degree of concentration during fabrication,” “degree of trim,” “nature of reduction emphasis” which in Hill’s [[ 11 ], Table 1; see also [ 19 ], Table 5–23] approach signify relative emphasis upon edge, thinning, and outline, however determined).
Callahan’s and other traditional approaches to the reconstruction of reduction sequences are highly detailed. Yet stage models are not always valid and replicable, nor do they control a considerable dimension of continuous variation. There is robust patterning in the relationship between continuous preform variables of weight, JTI and linear dimensions that would not emerge from a stage approach [ 16 ]. Complete analyses of any reduction sequence may be qualitative to some extent but also must be quantitative, for both individual specimens and entire assemblages. This paper suggests some statistical methods for continuous data, possibly useful in reduction-sequence analysis, that complement detailed technological description.
Supporting information
S1 Table
Stage assignment, dimensions and weight for study preforms.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Background
Populations of South Asian origin are known to be at particularly high risk of type 2 diabetes mellitus and complications, and are thus an important target for active screening and prevention [ 1 – 5 ]. A recent study has calculated that a substantial benefit can potentially be realised by specifically targeting South Asian populations for active screening and prevention [ 4 ].
The potential benefit efficiency/effectiveness of such an approach may in a large part depend on the uptake [ 6 ]. Furthermore, studies have shown that the uptake may be influenced by the method used for screening, with less invasive methods associated with a higher uptake [ 7 – 12 ]. The oral glucose tolerance test (OGTT) was for years considered the criterion standard for diagnosis of type 2 diabetes, and several studies have used this method for screening, sometimes after pre-selection based on fasting plasma glucose or risk scores [ 4 , 13 – 16 ]. Nevertheless, the concern has been expressed that the decision to commit participants to an arduous OGTT negatively affected the uptake of screening, and that more people would be tested and diagnosed if a more convenient test had been used [ 4 , 6 ].
In recent years, recommendations have been updated to include glycated hemoglobin (HbA1c) as a diagnostic option [ 17 – 19 ]. Because Hba1c can be determined with a single blood sample, it has practical advantages and is less burdensome than the OGTT, and may therefore be associated with a higher participation in screening and ultimately a higher yield of screening. If that is indeed the case, use of Hba1c could potentially have an important impact on efficiency of screening in the South Asian origin population. However, the possible effect on the uptake and yield of screening has not been evaluated in this high-risk population [ 4 – 5 , 20 ].
Therefore, the main aim of our study was to investigate the difference in the uptake, defined by the response to the invitation and participation in the screening, between 18–60 year old South Asian Surinamese men and women offered screening by means of an HbA1c measurement and those offered screening by means of an OGTT. We evaluated whether differences were consistent across age and sex groups. Moreover, we analysed whether a different subset of the population was reached, by analysing whether the characteristics of the screened participants differed according to the screening method. Finally, we estimated the yield of by comparing the percentage of cases of type 2 diabetes mellitus and prediabetes in our population identified if HbA1c as compared to the OGTT had been used.
Materials and Methods
Study population
We analysed data on South Asian Surinamese men and women, aged 18–60 years, who were invited to participate in the screening that took place to identify potential participants to be invited for the DH!AAN study, a randomized controlled trial of a lifestyle intervention for the prevention of type 2 diabetes mellitus in the Netherlands (Dutch Trial Register: NTR1499; [ 21 – 22 ]).
The term South Asian Surinamese is used to refer to people with South Asian ancestral origin, and their offspring who migrated to the Netherlands via Suriname. The South Asian Surinamese are the descendants of the indentured labourers from North India—Uttar Pradesh, Uttaranchal and West Bihar- between 1873 and 1917. The two large migration waves, around 1975 and 1980, of South Asian Surinamese to the Netherlands were mainly due to the political situation in Suriname [ 23 ].
Recruitment strategy
We selected 10,420 South Asian Surinamese, aged 18–60 years, from 48 general practice lists in The Hague by means of name analysis. The researcher, the physician or the practice nurse, and a trained research assistant of South Asian Surinamese origin analysed the names. People known to have type 2 diabetes mellitus and pregnant women were excluded.
The general practitioner sent each potential participant an invitation letter with a reply card that could be returned if further contact was unwanted. In DH!AAN, all people invited and screened between May 18 th , 2009 and April 18 th 2010 were offered an OGTT (n = 8408). To account for the possible effect of start-up problems on the recruitment, we excluded those invited before June 18 th 2009. We also excluded those invited after December 31 st 2009 to reduce the overflow from the OGTT group to the HbA1c measurement period. Thus, the OGTT group consisted of n = 3173 invitees. People invited from April 19 th 2010 to November 11 th , 2010 were offered screening by means of an HbA1c measurement (n = 2012, HbA1c group). We are able to make this comparison between methods due to an abrupt change in the study protocol [ 21 ]. Due to the shorter duration of a screening with a single measurement, a greater number of people could be screened within the available time.
The invitations for both groups were similar in content and form, except for the measurement offered. The OGTT group was informed that ‘blood sugar’ would be determined, after which a ‘sugar drink test’ would be performed that consisted of consumption of a ‘sugar drink’ and a ‘measurement of the blood sugars after 2 hours’. Invitees were also informed that, while they waited, their weight, height, waist circumference and blood pressure were measured. The HbA1c group was informed that their weight, height, blood pressure and ‘blood sugar’ would be determined. Screening in both groups took place from Monday to Saturday, after an overnight fast from 10 p.m.
Invitees who had not responded to the invitation within 2 weeks received a written reminder inviting the recipient to make an appointment. The reminder also said that if there was no response (no appointment or reply card), the invitee would be contacted by telephone. The study team phoned those who had not responded within 1 week at least 3 times. If no telephone number was available (24.6% of participants) or potential participants could not be reached, we sent a second written reminder.
All people who made an appointment for screening (‘response’) received a letter of confirmation. In addition, a text message was sent the day before the screening to remind the participants of their appointment.
The materials (e.g. using the colours of the Surinamese flag in the logo of the study, adjusting the risk information in the invitation to the population) and the recruitment strategy used, were based on the approach tested in a previous pilot study [ 22 ].
Data collection screening
People who attended the screening (‘participation’) were requested to fill out a brief questionnaire. We collected data on generation (first/second), education level (primary or lower, secondary, lower vocational, higher vocational or more), paid work (yes/no), known cardiovascular risk (previous diagnosis of high blood pressure, previous diagnosis of high cholesterol or experienced complaints related to heart disease; yes/no), and family history of type 2 diabetes mellitus (yes/no).
Trained research staff carried out a physical examination using a standardized protocol. The participants were weighed in light clothing on a Seca mechanical scale to the nearest 500 g (SECA gmbh & co, Hamburg, Germany). Height was recorded to the nearest 0.01 m on a Seca portable stadiometer. The anthropometric measurements were obtained twice, and the means were used for analysis. From weight and height we calculated the body mass index (BMI; weight in kilograms/height in meters 2 ). Blood pressure (Omron M5-1; Omron Healthcare Europe BV, Hoofddorp, the Netherlands) was measured in the seated position around the non-dominant arm supported at heart level. At most, five measurements were taken. We calculated the mean of the first two measurements with less than 5 mm Hg difference in both systolic and diastolic blood pressure [ 24 ]. Hypertension was defined as a systolic blood pressure ≥ 140 mm Hg, or a diastolic blood pressure ≥ 90 mm Hg, or a self-reported previous diagnosis of hypertension.
Finally, the participants in both groups were asked to give a fasting blood sample for the measurement of HbA1c and fasting plasma glucose and, in the OGTT group only, to undergo an OGTT (glucose load 75 g). The laboratory methods used have been described previously [ 25 ]. Type 2 diabetes mellitus and prediabetes were subsequently classified [ 17 ]:
according to HbA1c: HbA1c ≥ 48 mmol/mol (≥ 6.5%) as type 2 diabetes, and 39 ≤ HbA1c <48 mmol/mol (5.7 to 6.5%) as prediabetes.
according to fasting plasma glucose: a fasting plasma glucose of 126 mg/dl (7.0 mmol/l) or more as type 2 diabetes mellitus and fasting plasma glucose of 100–125 mg/dl (5.6–6.9 mmol/l) as prediabetes.
according to the OGTT: type 2 diabetes mellitus by a fasting plasma glucose of 126 mg/dl (7.0 mmol/l) or more and/or 2-h plasma glucose of 200 mg/dl (11.1 mmol/l) or more, and prediabetes (prediabetes) by fasting plasma glucose of 100–125 mg/dl (5.6–6.9 mmol/l) and /or a 2-h glucose post load of 140–199 mg/dl (7.8–11.1 mmol/l).
Statistical analysis
We calculated the response and participation rates for the HbA1c group and the OGTT group, in the total population and stratified by age and sex. In addition, we described the reasons quoted for non-response. Differences between groups in response, participation and reasons for non-response were assessed with Pearson Chi square tests. We also used logistic regression analysis to calculate the age and sex adjusted odds ratios (with corresponding 95%-confidence intervals) for response and participation in the HbA1c versus the OGTT group. Because we expected variation across general practices in the response [ 26 ], we used two-level regression models for these analyses. In these models, participants (level 1) were nested within general practice (level 2). To allow for dependencies between participants registered with the same practice, a random intercept (level 2) was incorporated into the model. In our tables, we report the estimates for the associations derived from these analyses. We also tested for interaction by age or sex by including an interaction term for age* type of invitation and age* type of invitation into the models. A p-value lower than 0.05 for the F test was considered indicative of interaction. Subsequently, we described the differences in the prevalence of characteristics of participants in the HbA1c group and the OGTT group. We tested differences between the groups with Pearson chi-square tests or analysis of variance. We then used two-level regression models to calculate the odds ratio for belonging to the OGTT versus HbA1c group, adjusted for sex and age. We also report the p-values derived from the F-tests for each of the fixed effects.
Finally, we calculated the percentage of new cases of type 2 diabetes mellitus and prediabetes identified per strategy. We estimating the percentage that would have been detected if specified strategy had been used, assuming a similar risk among participants and non-participants. For instance, the percentage of cases detected for the HbA1c strategy was calculated as: the participation rate in HbA1c group times the estimated prevalence defined by HbA1c, divided by the total prevalence as defined by HbA1c and OGTT combined. We defined a range by repeating the calculations with the lower bound and, then, upper boud of the confidence interval estimates for participation and prevalence.
The analyses were performed using the SAS package, version 9.3 (SAS Institute Inc., Cary, USA.). A p-value <0.05 was considered statistically significant. The data underlying our analyses have been made available in ( S1 Dataset . Data on response and participation in screening DHIAAN).
Ethical approval
The Institutional Review Board of the Academic Medical Centre of the University of Amsterdam approved the study. The study was carried out conform the Declaration of Helsinki. All participants provided oral and written informed consent.
Results
Of those invited, 48,8% were men. The mean age of invitees was 37.3 (37.0–37.6); 29.3% were 18–29 years old, 40.2% 30–44 years and 30.3% 45 years or older.
We observed a slightly higher response and participation rate in the HbA1c group than in the OGTT group ( Table 1 ). The response rate in the HbA1C group was 29.2% (CI 27.6–30.8) versus 24.6% (CI 23.1–26.1) in the OGTT group (p = 0.0002), and the corresponding participation rate 23.9% (CI 22.0–25.8) in the HbA1c group and 19.3% (CI 18.8–21.6) in the OGTT group (p<0.0001). This pattern of differences between the HbA1c and OGTT groups was also observed among men and women and across age groups ( Table 1 ). In both the OGTT group and the HbA1c group, the most frequently cited reason for non-response was ‘no time’ or ‘not interested’, followed by ‘not eligible’ ( Table 2 ). The differences in reasons for non-response between the groups were small and not significant (p = 0.26).
10.1371/journal.pone.0136734.t001
Table 1 Differences in response and participation between the HbA1c group and the OGTT group. OGTT group = people offered screening by means of an oral glucose tolerance test; HbA1c group = people offered screening by means of a glycated hemoglobin measurement; Adjusted OR = odds ratio for response or participation, adjusted for age and sex; CI = confidence interval.
HbA1c group,N = 2012
OGTT group,N = 3173
% responders or participants
% responders or participants
Adjusted OR(95%-CI)
P-valueinteract
Response
29.2
24.6 a
1.30 (1.00–1.68) b
- By sex
Men
24.7
20.1
1.35 (1.00–1.83)
0.59
Women
33.7
28.9
1.26 (0.95–1.67)
- By agegroup
18–29 years
20.9
15.2
1.54 (1.09–2.17)
0.05
30–44 years
28.0
26.8
1.08 (0.80–1.46)
45–60 years
39.1
30.7
1.48 (1.08–2.02)
Participation
23.9
19.3 a
1.30 (1.01–1.69) b
- By sex
Men
20.4
15.8
1.36 (1.01–1.84)
0.64
Women
27.4
22.5
1.27 (0.96–1.68)
- By agegroup
18–29 years
14.8
9.9
1.58 (1.09–2.29)
0.28
30–44 years
22.8
19.9
1.17 (0.87–1.58)
45–60 years
34.4
27.3
1.38 (1.01–1.86)
a P-value Pearson chi square overall response p = 0.0002 and overall participation p<0.0001;
b P-value F test overall response p = 0.049 and overall participation p = 0.039; P-value interact = p-value from F test of interaction by agegroup or sex; Response = response rate, that is percentage of invitees who made an appointment; participation = participation rate, that is percentage of invitees who were screened; The presented response and participation rates do not account for non-eligibility. The estimated rates accounting for eligibility, by excluding people who were not eligible from the selected population, would be higher.
10.1371/journal.pone.0136734.t002
Table 2 Reasons for non-response in the HbA1c group and the OGTT group. OGTT group = people offered screening by means of an oral glucose tolerance test; HbA1c group = people offered screening by means of a glycated hemoglobin measurement; Not eligible = outside age range, longterm illness, currently pregnant, already participating in other research project(s), moved away from The Hague; Unknown = no contact established or no reason provided during telephone contact or on reply card; OGTT = oral glucose tolerance test.
HbA1c group, N = 1424
OGTT group, N = 2392
N
%
N
%
Not eligible
255
17.9
370
15.5 a
Time or interest
312
21.9
529
22.1
- No time , too busy
93
6 . 5
290
12 . 1
- No interest or priority
219
15 . 4
239
10 . 0
Language problems
3
0.2
6
0.3
Unknown
854
60.0
1487
62.2
a P = 0.26 for the Pearson Chi square test for differences in the main categories of reasons for non-response between the OGTT and HbA1c groups.
The age and sex adjusted OR for response in the HbA1c group versus the OGTT group was 1.30 (95%-CI 1.00–1.68). A similar difference between the HbA1c group and the OGTT group was observed for the participation (adjusted OR = 1.30 (95%-CI 1.01–1.69). We did not find evidence for a different association between men and women and between younger and older age groups ( Table 1 ).
Further analysis revealed that the screened population in the HbA1c group was similar to the screened population in OGTT group with regard to age, sex, marital status, level of education, having paid work and known cardiovascular risk, but those in the HbA1c group were less likely to belong to the first generation (p = 0.02) or to have a family history of diabetes than participants in the OGTT group (p = 0.03). After accounting for general practice variation and adjustment for sex and age, the odds of belonging to the HbA1c or OGTT group did not differ significantly between for any of these characteristics ( Table 3 ). The prevalence of overweight, hypertension and the prevalence of type 2 diabetes mellitus and prediabetes defined by fasting plasma glucose or HbA1c were also similar between groups ( Table 3 ).
10.1371/journal.pone.0136734.t003
Table 3 Characteristics of participants in the HbA1c group and the OGTT group. OGTT group = people offered screening by means of an OGTT; HbA1c group = people offered screening by means of a glycated hemoglobin measurement; OGTT = oral glucose tolerance test; Adjusted OR = odds ratio for belonging to the OGTT versus HbA1c group, adjusted for age and sex; CI = confidence interval; CVD = cardiovascular disease; T2DM = Type 2 diabetes; Ref = reference group.
HbA1c group, N = 481
OGTT group, N = 611
Mean (sd) or % within group
Mean (sd) or % within group a
Adjusted OR(95%-CI)
P-value
Age
in years
41.3 (0.52)
41.4 (0.42)
0.91 (0.14–5.86)
0.94
Sex
(men)
42.4
39.6
1.00 (0.92–1.09)
0.92
Marital status
(married)
37.6
34.2
0.87 (0.13–5.98)
0.89
Education
Primary or less
16.5
14.7
0.84 (0.03–22.90)
0.999
Secondary
13.6
11.7
0.83 (0.05–12.81)
Lower vocational
53.5
56.8
0.76 (0.02–24.79)
≥ Higher vocational
16.5
16.8
Ref
Paid work
(yes)
71.5
74.1
0.94 (0.12–7.11)
0.95
Generation
(first)
78.3
83.6 a
1.77 (0.09–35.48)
0.71
Selfreported known CVD risk
(yes)
37.6
37.2
0.94 (0.13–6.84)
0.95
Family history T2DM
(yes)
68.4
74.8 a
1.10 (0.15–8.26)
0.92
Body mass index
in kg/m2
26.1 (0.21)
26.1 (0.18)
0.99 (0.81–1.22)
0.95
Hypertension
(yes)
40.0
37.0
0.90 (0.13–6.44)
0.92
Status defined by Hba1c
T2DM
3.3
3.3 b
1.12 (0.01–172.36)
0.996
Prediabetes
33.2
37.1
1.09 (0.14–8.52)
Normal
63.5
59.6
Ref
Status defined by fasting plasma glucose
T2DM
1.7
2.0 b
1.10 (0.00–855.10)
0.999
prediabetes
18.5
16.8
0.96 (0.07–13.21)
Normal
79.8
81.2
Ref
Status defined by OGTT
T2DM
-
3.5 b
-
-
prediabetes
21.0
Normal
75.5
a P-values for the univariate differences between groups (Pearson Chi square or ANOVA) are age 0.90, sex 0.35, marital status 0.25, education 0.82, paid work 0.33, generation 0.03, self-reported known CVD risk 0.90, family history T2DM 0.02, body mass index 0.86, hypertension 0.32, Status defined by Hba1c 0.56, Status defined by fasting plasma glucose 0.66;
b Calculated from data on n = 453 people. We excluded 159 people who did not have a full OGTT measurement: n = 13 invitees refused the 2-hour measurement and n = 145 invitees did not have 2-hour measurement because their appointment was scheduled after April 19 th 2010, i.e. after the switch to HbA1c as the standard method.
Finally, the percentage of diabetes cases identified in our population appeared to be independent of the strategy used ( Table 4 ). However, the HbA1c strategy did identify more prediabetes cases than the OGTT strategy.
10.1371/journal.pone.0136734.t004
Table 4 Estimation of the percentage of the total of cases with type 2 diabetes and prediabetes in the population detected in a 18–60 year old South Asian population in The Hague. Overall prevalence = prevalence based on combined OGTT and HbA1c measurement. OGTT = oral glucose tolerance test; HbA1c = glycated hemoglobin measurement; CI = 95%- confidence interval.
Type 2 diabetes
Prediabetes
Overall prevalence a , b
5 . 1
42 . 6
If invited for HbA1c
a) Participation rate (CI)
a) 23.9 (22.0–25.8)
a) 23.9 (22.0–25.8)
b) Prevalence method (CI) b
b) 3.3 (1.7–4.9)
b) 37.1 (32.8–41.4)
c) Percentage cases detected (range) c
c) 15.5 (7.4–24.7)
c) 20.8 (17.0–25.1)
If invited for OGTT
a) Participation rate (CI)
a) 19.3 (18.8–21.6)
a) 19.3 (18.8–21.6)
b) Prevalence method (CI) b
b) 3.5 (1.8–5.2)
b) 21.0 (17.2–24.8)
c) Percentage cases detected (range) c
c) 13.9 (6.7–22.0)
c) 9.9 (7.6–12.6)
a Prevalence in total study population, based on diagnosis according to either HbA1c or OGTT,
b Calculated from data on n = 453 people with a full OGTT and HbA1c measurement,
c Percentage cases detected is calculated as (participation rate x prevalence method)/overall prevalence. To calculate the lower limit of the range we used the lower bound of the confidence interval estimate for participation and prevalence. For the upper limit, we used the upper bounds.
Discussion
Main findings
Among men and women and across age groups, we found a higher response and participation among those invited for screening by means of an HbA1c measurement than among those invited for a screening consisting of an OGTT. Although we found that participants in the HbA1c group were less likely to have a family history of diabetes that participants in the OGTT group, no differences were found in other characteristics. The estimated yield, in terms of the percentage of all cases identified if all invitees had been offered screening by means of HbA1c differed from the estimated yield of the OGTT strategy for prediabetes, but not for type 2 diabetes mellitus.
Discussion of the main findings
Uptake of screening
The response and participation rates, regardless of the recruitment strategy, were higher in our study than the rates in two recent studies among South Asian origin populations selected from general practices in the UK [ 4 , 27 ]. This may be related to the more intensive recruitment strategy used in our study as compared to those studies. As compared to screening studies among European populations, the uptake is only slightly lower in our study [ 4 , 6 – 12 ]. This lower uptake was expected as a lower participation is often observed in studies among migrant populations in industrialised countries [ 4 , 28 – 30 ].
Differences in uptake in HbA1c vs. OGTT group
The higher uptake for HbA1c is in line with the assumption that a more burdensome test is associated with a lower response [ 4 , 20 ]. It is also in line with the observation from the ADDITION study that participation appeared lower among those offered an OGTT screening than among those offered other methods [ 7 ]. However, direct comparisons could not be made in that study as the different methods were used in different populations in different settings. Our findings, however, indicate that the absolute difference in uptake between both methods in this population is relatively small. Interestingly, the burden was not substantially more frequently quoted (reason declined ‘no time or interest’) among non-responders in the OGTT group than in the HbA1c group. Unfortunately, this self-reported reason may not be an adequate reflection of the invitees’ most important considerations for declining participation. Previous studies have demonstrated that participation in preventive programs is influenced by many other factors, such as perceived risk [ 30 – 33 ]. However, perceived risk and other possible factors that nay influence participation in screening were not available for non-participants in our study.
Selective participation
One important question is whether the two methods appeal to or reach different subsets of the population, as selection of a subgroup at a higher or lower risk of type 2 diabetes mellitus may affect the efficiency of screening. A relatively higher participation of people with a family history of diabetes in the OGTT group would fit the finding that people with a higher risk perception, related to having a family member with type 2 diabetes, may be more motivated to participate in prevention programmes than others [ 32 – 34 ] However, the differences in the odds of belonging to the HbA1c or OGTT group were not significant after adjustment for age and sex and accounting for general practice variations. Moreover, we did not find differences in the prevalence of type 2 diabetes, prediabetes and other metabolic outcomes between the groups.
Yield of screening
Ideally, a screening method efficiently identifies people with previously undiagnosed type 2 diabetes mellitus and, in light of the potential effectiveness of early lifestyle intervention [ 35 ], people at risk of type 2 diabetes (e.g. those with prediabetes). Based on the higher uptake, an invitation for screening by means of an Hba1c measurement would seem a better strategy than screening by means of an OGTT. However, the yield for 2 diabetes appeared similar for both strategies. This difference between uptake and yield is likely related to the estimated prevalence of type 2 diabetes, and the overlap of measures in our population. We have previously shown that Hba1c may not detect all new cases in our population that would have been identified if the OGTT had been used and vice versa [ 25 ].
While the estimated yield was similar for type 2 diabetes, the yield did differ between strategies for prediabetes. This difference between outcomes may be related to the greater correlation between both measures among people with diabetes than in those without diabetes [ 36 ]. In particular, because we chose the American Diabetes Association criteria to classify ‘prediabetes’, which use a broader range to define impaired glucose regulation as compared to other criteria [ 17 , 18 ]. The broader criteria result in a markedly higher proportion of people only identified with HbA1c as having prediabetes [ 37 ]. This is proportionally reflected in our calculation of the yield. However, as the uptake and prevalence in other populations or settings may be different, replication of our findings is warranted before a final conclusion is drawn.
Limitations
We made a comparison between potential invitees who were invited for a screening including an OGTT and people who were not. This ‘natural experiment’ was the result of an abrupt change in the study protocol [ 21 ]. However, a non-randomised comparison has some potential drawbacks. For instance, the difference in participation in the two groups may have been affected by changes in external circumstances. However, the recruitment took place in a relatively short period during which there were no major changes in local circumstances, such as local screening policies. Thus, it is unlikely that this explains the differences between the groups.
Another potential problem may be that the results were affected by baseline differences in characteristics that we did not measure. In addition, the uptake may be affected by characteristics of the general practices that participants were registered with [ 26 ]. We attempted to account for this variation by using multilevel analyses.
In our study, we asked all potential participants to fast prior to the screening appointment. However, fasting is not required for an HbA1c measurement [ 17 – 19 ]. The difference in uptake with the OGTT group might have been larger if we had dropped the requirement to fast for the HbA1c group.
Finally, we assumed a similar prevalence among people who were not screened than among people screened in our calculations of the yield, while this may not be the case [ 38 ]. This may have affected the absolute estimates of the yield if people with or without disease were more or less inclined to respond to an invitation for an OGTT measurement than for an HbA1c measurement. However, the comparison of the characteristics of screened participants did not show differences in metabolic profile between groups. Therefore, we expect that the effect on our results is small.
Conclusions
An invitation for screening with HbA1c was associated with a slightly higher uptake of screening in population of South Asian origin than an invitation for an OGTT, but the methods did not appeal to a substantially different subset of the population. The difference between strategies in the yield of screening for prediabetes suggests that the OGTT may be a less efficient than HbA1c for programs aimed at the identification of people at risk of diabetes. For type 2 diabetes, the yield was similar for HbA1c and the OGTT. This suggests that either method may be chosen for type 2 diabetes screening. However, although the yield is an important consideration, the final choice in practice should also be determined by factors that we did not record in our study, such as cost, organisational aspects and patients’ experience.
Supporting Information
S1 Dataset
Data on response and participation in screening DHIAAN.
A SAS 9.3 dataset with the data underlying this manuscript.
(SAS7BDAT)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Lung cancer is the leading cause of cancer death in developed countries. Due to the lack of efficient treatment for advanced disease, the prognosis of lung cancer is still poor, with less than 15% surviving 5 years after diagnosis [ 1 ]. Adjacent invasion and distant metastasis are the major causes of cancer-related death [ 2 ]. Therefore a search for inhibitors for cancer cell invasion and migration ability could reveal a new therapy for cancer treatment. Although herbs have been employed in the treatment of cancers for thousands of years, they remain a very important source of biologically active products. The aim of this study was to identify potential therapeutic agents to improve the survival of patients with lung cancer metastasis.
Lichens are symbiotic organisms that produce a large number of bioactive substances over 800 [ 3 ], comprising many classes of compounds: amino acid derivatives, sugar alcohols, aliphatic acids, γ-, δ- and macrocyclic lactones, monocyclic aromatic compounds, quinones, chromones, xanthones, dibenzofuranes, depsides, depsidones, depsones, terpenoids, steroids, carotenoids [ 4 ] and diphenyl ethers [ 5 , 6 ]. Slowly growing organisms in low-resource habitats produce higher levels of defense chemicals [ 7 ]. Therefore, lichens are a source of unique chemical agents of which some have already been proved to be effective against various cancer in vitro models [ 8 ]. Here, the current study examined the inhibitory activity of seven lichen species collected from the Romanian Carpathian Mountains against migration and invasion ability of human lung cancer cells and further investigated the possible molecular mechanisms underlying their anti-metastatic activity to identify potential compounds for novel anti-metastasis agents.
Material and Methods
Preparation of lichen extracts
Lichen specimens used in this study, collected from Romania in 2011, were identified at the Korean Lichen Research Institute (KoLRI), Sunchon National University, Korea. Briefly, thalli of lichen were collected from Romania in 2011 during the field trip in the National Park Călimani (47°07'28.6"N, 25°13'34.8"E) and the Natural Park Bucegi (45°20'21.7"N, 25°27'41.4"E) organized by Dr. Crişan at Babeş-Bolyai University, Cluj-Napoca, Romania [ 9 ]. The permit to collect lichen specimens from those locations was issued by the Administration of the National Park Călimani and the Administration of the Natural Park Bucegi, with the approval of the Commission for Protection of Natural Monuments (Romanian Academy). The field studies did not involve any endangered or protected species. The duplicates were deposited into the Korean Lichen and Allied Bioresource Center (KOLABIC) in the Korean Lichen Research Institute (KoLRI), Sunchon National University, Korea. The dried thalli of the lichens were extracted with acetone at room temperature for 48 h. The acetone extracts were then filtered and dried in rotary vacuum evaporator at 45°C. The dry extracts were dissolved in dimethylsulfoxide (DMSO) as 5 mg/ml concentration (1000×) for all experiments. Seven Romanian lichen species and their voucher specimen numbers used in this study were listed in Table 1 .
10.1371/journal.pone.0146575.t001
Table 1 Seven Romanian lichen species used in this study.
Collection No.
Family
Lichen species
Known lichen substances
Reference
RO11025
Parmeliaceae
Alectoria samentosa
(-)-Usnic acid, Physodic acid, 8’-O-ethyl-P-alectoronic acid, Alectosarmentin
[ 13 ]
RO11045
Parmeliaceae
Flavocetraria nivalis
(±)- Usnic acid, Isousnic acid, Divaricatic acid, p -Hydroxybenzoic acid, Vanillic acid
[ 14 – 17 ]
RO11084
Parmeliaceae
Alectoria ochroleuca
Diffractaic acid, (-)-Usnic acid, Isousnic acid, Friedelin, Barbatic acid
[ 16 , 18 ]
RO11111
Parmeliaceae
Bryoria capillaris
Alectorialic acid, Barbatolic acid
[ 19 ]
RO11166
Parmeliaceae
Hypogymnia physodes
Atronorin, Physodalic acid, Protocetraric acid, Physodic acid
[ 18 , 19 ]
RO11176
Parmeliaceae
Usnea florida
(+)-Usnic acid, Barbatic acid, Salazinic acid, Norstic acid, β-Orcinol depsidones, (±)-Thamnolic acid, Stictic acid
[ 18 , 20 ]
RO11209
Parmeliaceae
Evernia divaricata
Divaricatic acid
[ 19 ]
High performance liquid chromatography (HPLC) analysis of lichen material
Acetone extract of lichen thalli at a concentration of 5 mg/ml were subjected to high performance liquid chromatography (HPLC) analyses (LC-20A; Shimadzu, Kyoto, Japan) on a YMC-Pack ODS-A (150 × 3.9 mm I.D.) reversed-phase column containing fully end-capped C18 material (particle size, 5 μm; pore size, 12 nm). Elution was performed at a flow rate of 1 ml/min under the following conditions before subsequent injection: column temperature, 40°C; solvent system, methanol: water: phosphoric acid (80: 20: 1, v/v/v). Analyses were monitored by a photodiode array detector (SPD-M20A; Shimadzu) with a range of 190~800 nm throughout the HPLC run. Observed peaks were scanned between 190 and 400 nm. The standard used for salazinic acid (t R = 2.27 ± 0.2 min) was isolated from lichen Lobaria pulmonaria . Usnic acid used in our study was purchased from Sigma-Aldrich (St. Louis, USA) (329967-5G). Voucher specimens were deposited in the herbarium of the Lichen & Allied Bioresource Centre at the Korean Lichen Research Institute, Sunchon National University, South Korea.
Liquid chromatography-mass spectrometry (LC-MS) and optical rotation analaysis of lichen material
LC—MS spectra were recorded on a spectrometer with an electrospray ionization source using Agilent 6460 triple Quadrupole LC/MS. The values of optical rotation were measured at 25°C using Jasco P-1010 polarimeter with a sodium lamp, and described as follows: [α]D, T (c (g/100 mL), solvent). Specific rotation of pure (+)-usnic acid (Sigma-Aldrich, St. Louis, USA) is a physical property of at a given wavelength and temperature and can be looked up in literature.
Cell culture
The human lung cancer cells including A549, H460, H1650, and H1975 were cultured in RPMI 1640 culture medium supplemented with 10% fetal bovine serum, 1% Penicillin-Streptomycin solution under a humidified 5% CO 2 atmosphere at 37°C.
Wound healing assay
A549 cells were plated at a density of 2.5 × 10 5 cells/well on 6-well tissue culture plates (Corning, New York, USA) and grown overnight to confluence. Monolayer cells were scratched with a pipette tip to create a wound. The cells were then washed twice with serum-free RPMI 1640 to remove floating cells and incubated in RPMI1640 culture medium supplemented with 2% FBS with 5 μg/ml of the lichen extract or 5 μM usnic acid. Photographs of cells were taken at 0, 24, 48, and 72 h after wounding to measure the width of the wound. For each sample, an average of five wound assays was taken to determine the average rate of migration at a given concentration of acetone extract or usnic acid. Experiments were repeated at least three times.
Invasion assay
Invasion assays were performed in transwell chambers (Corning, New York, USA) with 8μm pore size polycarbonate membrane coated with 1% gelatin. Cells were plated at 2 × 10 5 cells/well in RPMI1640 containing 0.2% bovine serum albumin in the upper compartment of the chamber with or without 5 μg/ml crude lichen extracts. Then RPMI1640 medium with 10 μg/ml fibronectin was added to the lower chamber to serve as a chemotactic agent. After 48-h incubation, the cells in the upper chamber were fixed with Diff Quik kit (Sysmex, Kobe, Japan). Then the cells inside the chamber were mechanically removed from the membrane with a cotton swab, and the cells adhering to the under-side of the membrane were stained and counted under light microscope (5 fields per chamber). Each invasion assay was repeated in three independent experiments. The results are expressed as the mean number of cells migrating per high-power field.
Reporter assay
HEK293T cells were plated into 24-well plates 12 h before transfection. After transfection of the TOPFLASH or AP-1 reporter plasmid with the respective activator, β-catenin or KITENIN, cells were treated with usnic acid for 48 h and then analyzed using a Dual-Luciferase ® reporter assay system (Promega, Madison, WI, USA). The Renilla luciferase reporter plasmid (pRL-TK) was used as the internal control for the transfection efficiency. The experiments were performed in triplicate, and at least three results from independent experiments were included in the analysis. Fold changes were calculated using values normalized to Renilla luciferase activity.
Quantitative real-time PCR
The quantitative real-time PCR was performed as described previously [ 9 ]. Briefly, total RNA was isolated from human lung cancer cells by using RNAiso Plus (TaKaRa, Otsu, Shiga 520–2193, Japan) according to the manufacturer’s instructions. Total RNA (1 μg) from each group of treated cells was converted to cDNA using a M-MLV reverse Transcriptase kit (Invitrogen, Carlsbad, USA) and SYBR green (Enzynomics, Seoul, Korea). The primers used for real-time PCR were Cyclin D1 (forward) 5’-ccgtccatgcggaagatc-3’ and (reverse) 5’-gaagacctcctcctcgcact-3’ ; c-myc (forward) 5’-aatgaaaaggcccccaaggtagttatcc-3’ and (reverse) 5’-gtcgtttccgcaacaagtcctcttc-3’ ; CD44 (forward) 5’-tgccgctttgcaggtgtat-3’ and (reverse) 5’-ggcctccgtccgagaga-3’ ; GAPDH (forward) 5’-atcaccatcttccaggagcga-3’ and (reverse) 5’-agttgtcatggatgaccttggc-3’ . Real-time PCR reaction and analysis were performed using CFX (Bio-Rad, Hercules, USA).
Affinity Precipitation of Cellular GTPases
The cellular Rac1 and Cdc42 activities were determined using GST-RBD/PBD as previously described [ 10 , 11 ]. Briefly, the cells were lysed in lysis buffer (50 mM Tris, pH 7.4, 1% Triton X-100, 0.5% sodium deoxycholate, 0.1% SDS, 500 mM NaCl, 10 mM MgCl 2 , and protease inhibitor mixture). The lysates were incubated with GST- RBD/PBD beads at RT for 1 h. The beads were then washed four times with washing buffer (50 mM Tris, pH 7.4, 1% Triton X-100, 150 mM NaCl, 10 mM MgCl 2 , and protease inhibitor mixture). The bound Rac1 and Cdc42 proteins were detected by immunoblotting using a monoclonal antibody against Rac1 (MILLIPORE 05–389) and Cdc42 (SANTA CRUZ SC-87). The relative activity of each GTPase was determined by quantifying each band of GTP-bound GTPase and the total amount of GTPase using Multi-Gauge 3.0, and the values of the GTP-bound bands were normalized to the value of the total amount. All results were determined using three different exposures from at least three independent experiments.
Results
Inhibition of A549 cell motility by lichen extracts
Cell migration plays a crucial role during cancer metastasis. To find the anti-migratory lichen secondary metabolite on human lung cancer cells, wound healing assay was performed among seven acetone extracts of Romanian lichens listed in Table 1 . Lichens produce unique polyketide secondary metabolites including depsides, depsidones, dibenzofurans, and depsones; and these hydrophobic compounds were normally extracted with acetone [ 12 ]. As shown in Fig 1A , Alectoria samentosa , Flavocetraria nivalis , Alectoria ochroleuca , and Usnea florida inhibited A549 cell migration at a concentration of 5 μg/ml. The length between the edges of the wound at 72 h with these candidates were significantly wider than that in the DMSO-treated group or the non-active sample ( Bryoria capillaris ). In particular, F . nivalis showed more than 60% inhibitory activity compared to control ( Fig 1A and 1B ).
10.1371/journal.pone.0146575.g001
Fig 1
Inhibition of A549 cell motility by acetone extracts of lichens.
(A–B) Quantitative analysis of migration assay of A549 cells treated with 5 μg/ml of acetone extracts of Alectoria samentosa , Flavocetraria nivalis , Alectoria ochroleuca , Bryoria capillaris , Hypogymnia physodes , Usnea florida and Evernia divaricata (A), and representative images of migration assay of A549 cells treated with the extracts of A . samentosa , F . nivalis , A . ochroleuca , U . florida and B . capillaris (B). (C-D) Invasion assay of A549 cells treated with 5 μg/ml of acetone extracts of A . samentosa , F . nivalis , A . ochroleuca , U . florida and B . capillaris (C), and quantitative analysis of invaded cell numbers in each group (D). Representative images were shown from three independent experiments, n = 3. Data represent mean ± S.E.M. (standard error of the mean). ***p<0.001; NS, no significant difference compared to 0.01% DMSO-treated A549 cells.
The effects of acetone extracts of A . samentosa , F . nivalis , A . ochroleuca , and U . florida on A549 cell invasion were then determined using transwell chamber invasion assay. As a result, the lichen extracts significantly decreased the invaded cell numbers by as much as 50% compared with DMSO or B . capillaries (negative control) ( Fig 1C and 1D ). These findings demonstrated that acetone extracts of A . samentosa , F . nivalis , A . ochroleuca , and U . florida inhibited both migration and invasion ability of A549 lung cancer cells.
Usnic acid is the active inhibitor from the lichen species
To identify the components of the acetone extract of lichen, A . samentosa , F . nivalis , A . ochroleuca , and U . florida extracts were run on HPLC. As shown in Fig 2A , usnic acid was identified as the main compound in all four of these candidates after comparison with the internal standard of purified (+)-usnic acid (Sigma-Aldrich, St. Louis, USA), and these were consistent with previous report ( Table 1 ) [ 13 – 20 ]. Identity of usnic acid and their optical status were analyzed by LC-MS analysis and optical activity analysis, respectively ( S1 File ). The %intensity of peak for the usnic acid in the candidate lichens at a concentration of 5 mg/ml was obtained by comparing to that of peak for pure 5 mg/ml usnic acid. It is worth noting that usnic acid content is highest in F . nivalis extract, which may explain its potent inhibitory effect on cell migration ( Fig 2A ). It was speculated that (-)-usnic acid has similar or more potent inhibitory activity on cell motility as acetone extract of A . samentosa and A . ochroleuca which are known to have (-)-usnic acid as their subcomponent [ 13 , 16 , 18 ] showed similar or more potent inhibitory activity on migration and invasion, respectively, than (+)-usnic acid containing U . florida ( Fig 1 ). In our previous report, we showed that acetone extract of lichen F . cucullata and its component, usnic acid, inhibited tumorigenicity and motility of cancer cells [ 9 ]. In accordance with this, (+)-usnic acid at concentration of 5 μM significantly inhibited the migration and invasion of A549 cells ( Fig 2 ). At this concentration, usnic acid did not show cytotoxicity and/or inhibit cell proliferation (IC 50 value of usnic acid on A549 cells = 65.3 ± 0.65 μM) [ 9 ]. As shown in Fig 2B–2E , inhibitory activity of (+)-usnic acid at 5 μM is as high as 50% at 72 h treatment for migration and is around 40% at 48 h treatment for invasion. To further examine the inhibitory activity of (+)-usnic acid on the other lung cancer cells, invasion assay was performed using H1650, and H1975 cells. As a result, (+)-usnic acid treatment significantly decreased invaded cell number of H1650 and H1975 cells at a concentration of 5 μM ( Fig 3A ). The quantitative analysis revealed that inhibition was as high as 40% in both cells compared with vehicle-treated cells ( Fig 3B ). Together, the results demonstrated that (+)-usnic acid has inhibitory activity against cell motility of human lung cancer cells.
10.1371/journal.pone.0146575.g002
Fig 2
Identification of lichen secondary metabolite from candidate lichens.
(A) High performance liquid chromatography (HPLC) analysis of lichen acetone extracts. The %intensity of peak for the usnic acid in the extract at a concentration of 5 mg/ml was obtained by comparing to that of peak for pure 5 mg/ml usnic acid. (B–C) Migration assay of A549 cells treated with 5 μM of (+)-usnic acid (B), and quantitative analysis of wound length (C). (D–E) Invasion assay of A549 cells treated with 5 μM of (+)-usnic acid (D), and quantitative analysis of invaded cell numbers in each group (E). Representative images are shown from three independent experiments, n = 3. Data represent mean ± S.E.M. (standard error of the mean). ***p<0.001; NS, no significant difference compared to 0.01% DMSO-treated A549 cells.
10.1371/journal.pone.0146575.g003
Fig 3
(+)-Usnic acid inhibits invasion of H1650 and H1975 human lung cancer cell.
(A-B) Invasion assay of H1650, and H1975 cells treated with 5 μM of (+)-usnic acid (A), and quantitative analysis of invaded cell numbers in each cell line (B). Representative images are shown from three independent experiments, n = 3. Data represent mean ± S.E.M. (standard error of the mean). **p<0.01; ***p<0.001; NS, no significant difference compared to 0.01% DMSO-treated A549 cells.
(+)-Usnic acid decreases β-catenin-mediated TOPFLASH activity and KITENIN-mediated AP-1 activity
To investigate underlying mechanisms for the inhibitory activity of (+)-usnic acid, we performed TOPFLASH and AP-1 reporter assays to assess whether (+)-usnic acid can modulate β-catenin-mediated and/or KITENIN-mediated signaling activity. As shown in Fig 4A , (+)-usnic acid significantly decreased TOPFLASH activity by 18% at 1 μM and 37% at 10 μM. As for AP-1 activity, dose-dependent decreases were observed from 0.5 μM, and the decrease was significant, as much as 50% at 10 μM. Moreover, (+)-usnic acid also significantly decreased EGF-activated KITENIN-mediated AP-1 activity by 32% at 1 μM and 41% at 10 μM ( Fig 4B ). To check whether the level of downstream target genes of β-catenin/LEF and c-jun/AP-1 were affected by (+)-usnic acid treatment, quantitative real-time PCR analysis was performed. As shown in Fig 5 , relative expression levels of CD44, cyclin D1, and c-myc were significantly decreased by (+)-usnic acid treatment in lung cancer cells to different extents ( Fig 5A–5D ). These results suggest that (+)-usnic acid shows inhibitory activity against cell motility through the modulation of β-catenin-mediated and KITENIN-mediated signaling activity in lung cancer cells.
10.1371/journal.pone.0146575.g004
Fig 4
(+)-Usnic acid decreases β-catenin-mediated TOPFLASH activity and KITENIN-mediated AP-1 activity.
(A) β-Catenin-mediated transcriptional activity of TOPFLASH promoter was decreased by (+)-usnic acid treatment. HEK 293T cells were transfected with β-catenin and TOPFLASH reporter plasmid. After 12 h transfection, cells were treated with (+)-usnic acid for 48h. (B) KITENIN-mediated transcriptional activity of AP1 promoter was decreased by (+)-usnic acid treatment. The HEK 293T cells were transfected with KITENIN and AP-1 reporter plasmid. After 12 h transfection, cells were treated with (+)-usnic acid for 48h with or without EGF. Experiments were performed in at least three independent cultures, n = 3. Data represent mean ± S.E.M. (standard error of the mean). *p<0.05; **p<0.01; ***p<0.001; NS, no significant difference compared to 0.01% DMSO-treated HEK 293T cells.
10.1371/journal.pone.0146575.g005
Fig 5
(+)-Usnic acid decreases mRNA level of downstream target genes of β-catenin/LEF and c-jun/AP-1.
(A-D) Quantitative analysis of the mRNA level of CD44, c-myc, and Cyclin D1 in A549 (A), H1650 (B), H1975 (C), and H460 (D) cells treated with 5 μM of (+)-usnic acid. Data represent mean ± S.E.M. (standard error of the mean), n = 3. *p<0.05; ***p<0.001; NS, no significant difference when compared to the 0.01% DMSO-treated group in each cell line.
(+)-Usnic acid decreases GTP-Rac1 and -RhoA level
The activities of Rac1 and Cdc42 are involved in mesenchymal mode of migration [ 21 , 22 ]. To determine whether (+)-usnic acid can affect the activities of these proteins in A549 cells, GST pull-down assays were performed using GST-PBD (p21-binding domain). As shown in Fig 6 , (+)-usnic acid treatment significantly decreased the level of GTP-Rac1 by 22% compared to vehicle-treated cells ( Fig 6A ). However, no significant change in the level of GTP-Cdc42 was observed by (+)-usnic acid treatment ( Fig 6B ). RhoA promotes junctional formation, apical constriction, and reduces adhesion and cell spreading [ 23 , 24 ]. To determine whether (+)-usnic acid can affect the activity of RhoA in A549 cells, GST pull-down assays were performed using GST-RBD (Rho-binding domain). As shown in Fig 6C , (+)-usnic acid treatment significantly decreased the level of GTP-RhoA by 40% compared to vehicle-treated cells. Taken together, these results suggest that (+)-usnic acid inhibits cell motility through the regulation of Rho GTPases.
10.1371/journal.pone.0146575.g006
Fig 6
Regulation of RhoGTPases activity by (+)-usnic acid.
(A-C) The levels of GTP-bound Rac1, Cdc42 and RhoA were measured in A549 cells treated with 5 μM of (+)-usnic acid. GTP-Rac1 and -Cdc42 were measured using GST-PBD, and GTP-RhoA was measured using GST-RBD. The total amounts of RhoA, Rac1, and Cdc42 were also shown. The relative activities of Rac1 (A), Cdc42 (B), and RhoA (C) were determined as described in Materials and Methods. The data represent the mean ± SEM (standard error of the mean), n = 3. **p<0.01; ***p<0.001; NS, no significant difference compared to 0.01% DMSO-treated A549 cells.
(+)-Usnic acid shows additive inhibitory activity with cetuximab
Cetuximab (Erbitux, C225), monoclonal antibody to epidermal growth factor receptor (EGFR) is used as one of anti-EGFR agents for the treatment of metastatic colon and lung cancer. To examine whether (+)-usnic acid has therapeutic relevancy with cetuximab, invasion assay was performed with various concentration of cetuximab and/or (+)-usnic acid. As shown in Fig 7 , inhibitory activity of cetuximab was ~30% at 1 μg/ml and ~40% at 10 μg/ml on A549 cells, and treatment of 5 μM (+)-usnic acid showed similar inhibitory activity with 10 μg/ml of cetuximab (~40% inhibition) on these cells. Interestingly, higher inhibitory activity for cell invasion was observed when the cells were treated with 1 μg/ml of cetuximab together with 5 μM of (+)-usnic acid ( Fig 7 ). These results suggest that (+)-usnic acid not only can inhibit lung cancer cell motility by alone but also can potentiate the therapeutic activity of cetuximab. As usnic acid decreased KITENIN-mediated AP-1 activity (Figs 4 and 5 ) and KITENIN/ErbB4-mediated downstream signal of EGF plays one of the molecular basis for conferring resistance to anti-EGFR agents [ 25 ], these results suggest that usnic acid may have potential beneficial activity in overcoming the limited clinical efficacy of anti-EGFR therapy.
10.1371/journal.pone.0146575.g007
Fig 7
(+)-Usnic acid shows additive inhibitory activity with cetuximab.
(A-B) Invasion assay of A549 cells treated with 5 μM of (+)-usnic acid and/or various concentration of cetuximab (A), and quantitative analysis of invaded cell numbers in each group (B). Representative images are shown from three independent experiments, n = 3. Data represent mean ± S.E.M. (standard error of the mean). *p<0.05; **p<0.01; ***p<0.001; NS, no significant difference between indicated group.
Discussion
Cancer cell acquires biological capabilities including resisting cell death, sustaining proliferative signaling, evading growth suppressors, activating invasion and metastasis, and so forth in developing from early to late stages [ 26 , 27 ], and targeting either of these acquirements can be grouped as ‘anticancer’. In this regard, our observations that (+)-usnic acid inhibits migration and invasion ability in lung cancer cells are novel in anticancer activity of (+)-usnic acid. In addition, our results demonstrated that (+)-usnic acid have specific mechanisms of action for their anticancer activity, and these are quite different from those of previous literature showing cytotoxicity, one of mechanism of actions for the anticancer activity of (+)-usnic acid in various cancer cells.
Hepatotoxicity of usnic acid may restrict their potential medicinal use in cancer therapeutics [ 28 ]. However, most of hepatotoxicity in human was observed when high dose of usnic acid was orally administrated as a dietary supplement for the purpose of weight loss [ 29 – 31 ]. In cancer therapeutics, hepatotoxicity can be avoided by adjusting dosage, formulation and/or route of medication. For example, da Silva Santos et al. showed that nano-encapsulation of usnic acid enable to maintain and improve antitumor activity and considerably reduce the hepatotoxicity [ 32 ]. Furthermore, it has been shown that supplementation of anti-oxidant, i.e. vitamin E, together with usnic acid could greatly reduce the usnic acid-induced hepatotoxicity in primary cultured mouse hepatocytes [ 33 ]. It should also be noted that some of the papers showing hepatotoxicity were carried on HepG2 cells, originated from hepatocellular carcinoma tissue of 15 years male adolescent [ 34 , 35 ]. In our previous paper [ 9 ], we demonstrated that usnic acid somehow has selective cytotoxicity in cancer cells when compared to normal cells. In alternative point of views, severe cytotoxicity of usnic acid on HepG2 cells reflects selective and specific anticancer activity of usnic acid on liver cancer in tested concentrations.
In our previous report, it was shown that the cytotoxicity of usnic acid is specific to cancer cells such as HT29 (colorectal cancer cells; IC 50 = 95.2 ± 0.85 μM), AGS (gastric cancer cell; IC 50 = 15.01 ± 0.52 μM), A549 (lung cancer cell; IC 50 = 65.3 ± 0.65 μM), and CWR22Rv-1 (prostate cancer cells; IC 50 = 24.1 ± 0.63 μM), while non-cancer cells such as MDCK (Mardin-Darby canine kidney; IC 50 = 133.04 ± 3.5 μM), RIE (rat intestinal epithelial cells; IC 50 = 126 ± 4.25 μM), NIH 3T3 (mouse embryonic fibroblast; IC 50 = 164.2 ± 3.7 μM), and HaCaT (human keratinocyte; IC 50 = 185.7 ± 4.8 μM) cells, were not severely damaged [ 9 ]. Given that the average circulating blood volume for mice is 72 mL/kg [ 36 ] and the molecular weight of usnic acid is 344, LD 50 value of usnic acid (mouse-oral; 838 mg/kg) in MSDS sheet of usnic acid can be calculated to 33.8 mM. As inhibitory activity of usnic acid in inhibiting lung cancer cell motility is observed at concentration of 5 μM, our results indicated that usnic acid would be used for anti-metastasis agent with little toxicity at working concentrations.
Usnic acid has a low degree of aqueous solubility, which is an obstacle in the drug development as it is likely to result in poor bioavailability. To avoid this problem, various approaches can be made for the enhancement of solubility with its own compound by particle size reduction, crystal engineering, salt formation, solid dispersion, use of surfactant, complexation, and so forth [ 37 ]. The selection of the enhancement method should be determined to each drug in required dosage form characteristics. For usnic acid, recent article reported that potassium salt of usnic acid (potassium usnate) has 100% solubility without losing its biological activity at 10 μg/ml (approximately 26 μM) [ 38 ]. Together with the facts that (+) and (-) enantiomeric forms of usnic acid showed moderate to strong biological activities [ 39 , 40 ], further study is required to evaluate the potential medicinal usage of usnic acid in anticancer therapy in various cancers.
Supporting Information
S1 File
LC-MS Analysis (Figure A in S1 File) and Optical activity Analysis (Figure B in S1 File) of samples used in this study.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Defects in enzymatic pathways of liver are associated with altered metabolism leading to hypoglycemia ± hepatomegaly and/or liver disease in hepatic forms of glycogen storage disorder (GSD) [ 1 ]. Based on affected enzyme and its relative expression in the liver, kidney, skeletal muscle, or heart, the clinical manifestations of GSDs vary from one disorder to the other [ 2 ]. Type I glycogen storage disease (GSD I) is the commonest most autosomal recessive form that typically presents in early infancy [ 3 ]. The catalytic subunit of microsomal glucose-6-phosphatase (G-6-Pase; E.C. 3.1.3.9) plays a pivotal role in glycogenolysis and gluconeogenesis catalyzing the last step of both metabolic pathways. Its deficiency leads to glycogen storage disease type Ia (GSD Ia; Von Gierke Disease; MIM #232200), which is usually characterized by hepatomegaly, hypoglycemia, lactic acidemia, hyperuricemia, hyperlipidemia. Untreated patients may have a cushingoid appearance, failure to thrive, an enlarged liver, protuberant abdomen, and delayed motor development [ 4 , 5 ]. Cerebral damage resulting from recurrent hypoglycemic episodes may lead to abnormal cognitive development [ 6 ]. The demonstration of a reduced G-6-Pase activity measured in a fresh liver biopsy specimen is still considered the gold standard for verification of the clinical diagnosis. However, in 1993 the gene ( G6PC ; GDB 231927) spanning 12.5 kb on chromosome 17 and consisting of 5 exons coding for the enzyme was cloned [ 7 ]. The protein encoded by this gene contains 357 amino acids and is an endoplasmic reticulum (ER) membrane associated protein containing the ER retention signal, and possesses six putative membrane spanning segments [ 8 ]. To date 146 sequence variations have been identified in G6PC gene related to GSD 1a, which have been documented from various countries (The Human Gene Mutation Database (HGMD1). Available at: http://www.hgmd.cf.ac.uk/ac/index.php , Accessed: 20July 2022).
Owing to recessive inheritance pattern of GSD Ia, its incidence is higher in populations with customary consanguineous marriages like Pakistan, necessitating comprehensive clinical and genetic studies on this disease from our local population. However there are only few reports from Pakistan focusing on the biochemical findings and clinical manifestations of the disease in children with GSD 1a, and a report on the disease-causing variation identified in a Pakistani case, described that disease causing variants may not be comprehensive, and there may be additional mutations yet to be identified for Pakistani patients with GSD Ia [ 9 , 10 ]. To understand molecular basis of GSD Ia in Pakistani population for potential therapeutic targets there is a need of extensive research regarding contributing genetic and environmental risk factors of GSD Ia. Therefore, the present study was aimed to check clinical heterogeneity among forty Pakistani GSD Ia patients presented at two tertiary care hospitals. We aimed to investigate the variants in G6PC gene among 20 cases based on willingness to participate in molecular analysis, to identify the underlying molecular defects leading to GSD1a phenotype.
Methods
The study was carried out at Molecular Biology Lab, Quaid-i-Azam University (QAU), Islamabad, Pakistan after approval from Bioethical Committee of Faculty of Biological Sciences, QAU. Patients affected with GSD-1a, diagnosed at Neurology and Gastroenterology departments of Children Hospital Lahore (CHL) and Pakistan Institute of Medical Sciences (PIMS), Islamabad were recruited in this prospective study. Blood samples of patients and available unaffected family members were collected along with the family history and clinical data after informed written consent. Forty patients (35 with parental cousin marriages) were diagnosed clinically for enlarged liver and kidneys, growth retardation and short stature, abnormal levels of glucose, lactate, uric acid, triglycerides, and cholesterol. Frequencies of observed clinical features and mean values of diagnostic tests were calculated using SPSS 21.0. Blood samples were stored in EDTA containing tubes and DNA extraction was performed using phenol chloroform method [ 11 ]. Primers of all 5 exons of G6PC gene (ENST00000253801, NM_000151) were designed by using Primer3 software ( http://bioinfo.ut.ee/primer3-0.4.0/ ) ( Table 1 ). The PCR reactions were carried out using protocol described by Gul et al ., 2022 [ 12 ]. The amplified PCR products were loaded on the 1.5% agarose gel along with 1 kb size ladder to evaluate product size. The purification was done by using DNA purification Kit (Wiz Bio Solutions, Seongnam, Korea) and purified products were sent for commercial Sanger’s sequencing. The sequenced data was analyzed by using Sequencher 5.4.6 software. Pathogenicity prediction for each variant was done by various bioinformatics tools named Mutalyser ( https://mutalyzer.nl/ ), MutationTaster ( http://www.mutationtaster.org/ ), PROVEAN ( http://provean.jcvi.org/index.php), Polyphen-2 ( http://genetics.bwh.harvard.edu/pph2/ ), Mutation assessor ( http://mutationassessor.org/r3/ ), SIFT ( http://sift.jcvi.org/ ), HOPE ( https://www3.cmbi.umcn.nl/hope/method ), Varsome ( https://varsome.com/ ), CADD ( https://cadd.gs.washington.edu/snv ), HSF (Human Splice Site Finder) software version 3.0 ( www.umd.be/HSF3/ ) to determine effects of sequence variations on exonic splicing signals. American College of Medical Genetics and Genomics (ACMG) classification is used for variant classification.
10.1371/journal.pone.0288965.t001
Table 1 Primers used for amplification of exons 1, 2, 3, 4 and 5 of G6PC gene.
Exon No.
Primer Sequence (5’ ➝ 3’)
Product Size (bp)
1
F
TTGAGTCCAAAGATCAGGGC
483 bp
R
TGAATAGCCTGGGGAAAGCA
2
F
CCACCCAGTTCTCCCTTCTA
519 bp
R
CTTTCTCAGGACACAGCGCT
3
F
GGTAGATGGGTGGATAGGGG
289 bp
R
AGAATACGTGGTGTGTCAGC
4
F
AAAATTCCACTGAGAGCACCT
358 bp
R
ACCCACAGAAATGCTAACAGT
5a
F
GCAGAACGGATGGCATGTCA
385 bp
R
AGCTCTCCCTGTACATGCTG
5b
F
GTGGACTCTGGAGAAAGCCC
524 bp
R
GACCCTCCAATCTGCCATCC
Results
A total of forty patients diagnosed with GSD Ia, included in this study showed diverse clinical symptoms (see S1 Table ). Frequency of disease was observed more in males as out of 40 enrolled cases 26 (65%) were males and 14 (35%) patients were females. Parental consanguinity was observed in 35 (87.5%) cases with 23 (57.5%) showing family history of disease. Out of forty enrolled cases, 23 (57.5%) patients died within 6 months because of disease severity. Hepatomegaly was observed in all cases (100%), however hepatic adenomas were present in 2 (5%) cases. 10 (25%) cases had a history of seizures. Epistaxis was observed in 7 (17.5%) cases. Delayed motor development was observed in 5 (12.5%) however 3 (7.5%) showed cushingoid appearance. Osteopenia was seen in 9 (22.5%) cases and 15 (37.5%) had inflammatory bowel disease (see S1 Table ). The mean values of diagnostic tests performed for GSD Ia cases are shown in Table 2 .
10.1371/journal.pone.0288965.t002
Table 2 Age and the mean values of diagnostic tests performed for GSD Ia cases included in this study.
Continuous variables
Mean± Std. Deviation
Age
7.913±4.9625
Hypoglycemia (mg/dL)
58.500±6.5984
High microalbuminuria
166.725±105.1239
Hyperuricemia
6.215±.9989
Proteinuria
192.225±53.1994
Hypertriglyceridemia
279.075±59.9912
Lactic Acidosis
5.728±.8249
Std = Standard
For molecular analysis blood samples of 20 cases were collected based on patient’s willingness to participate in genetic analysis as well as blood transfusion records. Upon sequencing of coding exons, their flanking intronic regions and 3’ as well as 5’ untranscribed regions (UTRs) of G6PC (chromosome:GRCh38:17:42900197:42913969:1) gene in 20 patients, overall, 21 variants were detected ( Table 3 ). Out of 21 identified variants there were 8 novel disease-causing variants (5 homozygous and 3 heterozygous) as predicted by mutation taster in the coding regions of exon 1, 2 and 5b of G6PC gene and 13 polymorphisms (10 homozygous and 3 heterozygous). The identified disease-causing variants were neither found in 1000G nor in Exome Aggregation Consortium” ExAC–composed of 60,706 unrelated individuals, and the Online Archive of Brazilian Mutations.
10.1371/journal.pone.0288965.t003
Table 3 List of identified variants in G6PC gene in GSD 1a cases identified in this study.
HGVS
Patient ID
Exon
Zygosity
Polyphen 2 Prediction (Score)
ACMG Classification
Functional domain
SIFT (P/S)
Novelity/ID
G6PC(NM_00151.4):c.49_50insT
2
E1
Homo
NA
Likely pathogenic
5’UTR
N/A
N
G6PC(NM_00151.4):c.71A>C(p.Gln24Pro)
11
E1
Hetero
B (0)
Likely benign
CDS
T 0.32
N
G6PC(NM_00151.4):c.109G>C(p.Ala37Pro)
11
E1
Homo
PD (0)
Uncertain significance
CDS
APF 0.05
N
G6PC(NM_00151.4):c.133G>C(p.Val145Leu)
5
E1
Hetero
B (0.079)
Uncertain significance
CDS
T 0.53
N
G6PC(NM_00151.4):c.205G>A(p.Asp69Asn)
5
E1
Homo
PD (1)
Uncertain significance
CDS
APF 0.02
N
G6PC(NM_00151.4):c.244C>A(p.Gln82Lys)
9
E2
Homo
PD (0.907)
Uncertain significance
CDS
T 0.21
N
G6PC(NM_00151.4):c.322A>C(p.Thr108Pro)
4
E2
Homo
PD (1)
Likely pathogenic
CDS
D 0.02
N
G6PC(NM_00151.4):c.322A>C(p.Cys284Tyr)
19
E5
Hetero
PD (0.771)
Uncertain significance
CDS
T 0.08
N
NC_000017.9:g.38301348_38301349insT
2
E1
Homo
NA
Likely benign
5’UTR
N/A
N
NC_000017.9:g.38304430delA
10,18
E2
Homo
NA
Uncertain significance
I
N/A
N
NC_000017.9:g.38304606delG>T
7,20
E2
Hetero
NA
Benign
I
N/A
N
NC_000017.9:g.38304772T>C
1,2,4,18,20
E2
Homo
NA
Benign
I
N/A
rs2593595
NC_000017.9:g.38308244_38308245insA
9
E3
Homo
NA
Benign
I
N/A
N
NC_000017.9:g.38308101T>A
3
E3
Homo
NA
Benign
CDS
N/A
N
NC_000017.9:g.38315015T>C
9,3,2
E4
Hetero
NA
Benign
I
N/A
rs161622
NC_000017.9:g.38309808_38309809insA
2,8
E4
Homo
NA
Uncertain significance
I
N/A
N
NC_000017.9:g.38309815_38309816insC
6
E4
Homo
NA
Uncertain significance
I
N/A
N
NC_000017.9:g.38309809T>G
2
E4
Homo
NA
Benign
I
N/A
N
NC_000017.9:g.38316992T>C
7,10,18,20
E5
Hetero
NA
Likely benign
3’UTR
N/A
rs2229611
NC_000017.9:g.38311697G>A
9
E5
Homo
NA
Benign
CDS
N/A
N
NC_000017.9:g.38312116A>C
1
E5
Homo
NA
Benign
3’UTR
N/A
N
PD: Probably damaging, B: Benign, HGVS: Human Genome Variation Society, N/A: Not applicable, CDS: Coding sequence, I: Intron, N: Novel, 3’UTR: 3’ Untranslated region. APF: Affected protein function, D: Damaging, T: Tolerant.
Five novel homozygous disease causing variants include G6PC (NM_000151.4):c.109G>C (p.Ala37Pro (patient ID 11), c.205G>A (p.Asp69Asn) (patient ID 5), c.49_50insT (patient ID 2) in exon 1 of G6PC gene; c.244C>A (p.Gln82Lys) (patient ID 9) and c.322A>C(p.Thr108Pro) (patient ID 4) in exon 2 of G6PC gene ( Fig 1A–1E ). Three novel identified heterozygous disease causing variants include c.71A>C (p.Gln24Pro) (patient ID 11), c.133G>C (p.Val45Leu) (patient ID 5) in exon 1 of G6PC gene and c.322A>C (p.Cys284Tyr) (patient ID 19) in exon 5 of G6PC gene ( Fig 2A–2C ) ( Table 3 ). All variants predicted to be disease causing by mutation taster were predicted to be damaging by Polyphen2 ( Table 3 ). These variants were not found in ExAC. Hope analysis showed that the original wild-type residue and newly introduced mutant residue differ in properties for each novel missense variant playing role in disease pathogenicity.
10.1371/journal.pone.0288965.g001
Fig 1
Electropherograms of homozygous disease-causing variants identified in G6PC gene.
A . Electropherogram showing c.109G>C (p.Ala37Pro) in patient 11 in exon 1 of G6PC gene. B . Electropherogram showing c.205G>A (p.Asp69Asn) in patient ID 5 in exon 1 of G6PC gene. C . Electropherogram showing c.244C>A (p.Gln82Lys) in patient ID 9 in exon 2 of G6PC gene. D . Electropherogram showing c.322A>C (p.Thr108Pro) in patient ID 4 in exon 2 of G6PC gene. E . Electropherogram showing c.49_50insT identified in patient ID 2 in exon 1 of G6PC gene.
10.1371/journal.pone.0288965.g002
Fig 2
Electropherograms showing heterozygous disease-causing variants identified in G6PC gene.
A . Electropherogram showing c.71A>C (p.Gln24Pro)identified in exon 1 of G6PC gene. B . Electropherogram showing c.133G>C (p.Val45Leu) identified in exon 1 of G6PC gene. C . Electropherogram showing c.322A>C (p.Cys284Tyr) identified in exon 5b of G6PC gene.
Among three known polymorphisms identified in this study, a homozygous variant g.38304772T>C ( rs2593595 ) was identified in 25% cases and two heterozygous variants i.e., g.38315015T>C ( rs161622 ) and g.38316992T>C (rs2229611) were identified in 15% and 20% of cases respectively ( Fig 3A–3C , Table 3 ). The novel polymorphisms identified in GSD Ia cases include g.38304430delA, g.38308244_38308245insA, g.38309808_38309809insA, g.38309815_38309816insC, g.38308101T>A, g.38311697G>A, g.38312116A>C and g.38309809T>G in homozygous states whereas variant i.e., g.38304606delG>T was found in heterozygous state ( Fig 4A–4J , Table 3 ).
10.1371/journal.pone.0288965.g003
Fig 3
Electropherograms showing reported polymorphisms in G6PC gene identified in this study.
A- g.38315015T>C identified in heterozygous state ( rs161622 ), B. g.38304772T>C identified in homozygous state (rs2593595 ) C. g.38316992T>C identified in heterozygous state ( rs2229611 ).
10.1371/journal.pone.0288965.g004
Fig 4
Electropherograms showing novel polymorphisms identified in G6PC gene in this study.
A . g. g.38304430delA (homozygous). B . g.38304606delG>T (heterozygous), C . g.38301348_38301349insT (homozygous). D . g.38308244_38308245insA (homozygous). E . g.38309808_38309809insA (homozygous) F . g.38309815_38309816insC (homozygous), G. g.38309809T>G g.8481T>G (homozygous). H . g.38308101T>A (homozygous). I g.38311697G>A (homozygous) J . g.38312116A>C (homozygous).
Discussion
Glucose 6-phosphatase ( G6PC ) enzyme catalyzes the hydrolysis of glucose-6-phosphate (G6P) to produce inorganic phosphate and glucose in liver and kidney cells. The glucose produced is then transported out of the cell to contribute in maintenance of the blood glucose level even during starvation [ 13 ]. Disease causing variants in the G6PC gene result in defective glucose-6-phosphatase activity causing storage of glycogen in liver and kidney cells leading to glycogen storage disease type Ia (GSD Ia). Although there have been some studies on the incidence of GSD Ia in Pakistan, there is still a need for more comprehensive epidemiological data of the disease. This would include data on the incidence, prevalence, and distribution of the disease across different regions and populations in Pakistan. GSD Ia is a rare but serious metabolic condition that runs in families as an autosomal recessive disorder [ 8 , 14 ]. There have been some studies on the genetic basis of GSD Ia from Pakistan, there is still much to be learned about the specific mutations and genetic variants that are most commonly associated with the disease in this population. This knowledge could help to design more effective diagnostic and treatment approaches. Buildup of glycogen in the liver and kidneys, cause progressive hepatomegaly and nephromegaly. Hypercholesterolemia, hypertriglyceridemia, hyperuricemia, and lactic acidemia are all metabolic implications of elevated cytoplasmic G6P levels [ 15 ]. Analysis of G6PC pathogenic variants is required to acquire differential clinical diagnosis [ 16 , 17 ]. To date, various pathogenic variants have been reported in the G6PC gene worldwide, including missense (the most prevalent form), nonsense, insertion/deletion and splice site variants, but there is no significant data from Pakistan [ 14 ]. Characterization of G6PC and identification of disease-causing variants in this gene provide a DNA-based tool to diagnose patients clinically suspected for GSD Ia. Moreover, disease causing variant analysis of a family at risk of conceiving offspring with GSD Ia offers genetic counseling. Furthermore for affected individuals, early diagnosis allows the employment of adequate metabolic control strategies and treatments to prevent complications and thus increases the quality of life [ 18 ]. Autosomal recessive disorders are prevalent in Pakistan because of the high rate of consanguineous marriages [ 19 ]. Incidence of hepatic glycogenesis is unknown for Pakistan, and there have been no in-depth disease causing variant investigations of GSDs to date.
Current study focuses on clinical and genetic analysis of GSD 1a cases from Pakistan. Clinical analysis was performed on 40 GSD 1a cases, with presenting symptoms of seizures, irritability and increased respiratory rate caused by hypoglycemia and hyperlactacidaemia as well as hepatomegaly was predominantly present (see S1 Table ).
The study identified that males (65%) were more affected as compared to females (35%). In current study the ratio of males identified with disease was more than the females. The disease being autosomal, is not linked with any of the gender. So, it could be due to underdiagnosis. 70% of the patients belonged to the age group 0–10 years in which the disease appeared during early ages. 87.5% of cases in this study belonged to consanguineous families. 57.5% patients could not survive disease severity due to lack of proper disease management resulting in complications including severity of hypoglycemic events leading to high mortality rate as reported previously by Ai et al., 2020 [ 20 ]. Seizures and delayed motor development were observed in 25% and 12.5% cases respectively, as observed in previous studies [ 21 , 22 ]. Three patients were observed with cushingoid appearance and osteopenia was found in 9 cases (see S1 Table ). Cushingoid appearance, delayed motor development and osteopenia is attributed to untreated GSD Ia [ 23 , 24 ]. These complications are attributed to hypothalamic-pituitary-adrenal (HPA) axis stimulation due to chronic hypoglycemic stress causing elevated glucocorticoid secretion [ 25 ]. In our study cohort, inflammatory bowel disease (IBD) was present in 15 patients (37.5%) (see S1 Table ), that is higher as studies have reported an occasional presence of IBD in GSD Ia cases [ 26 , 27 ]. Disease causing variant analysis of GSD Ia in 20 selected cases identified four novel missense variants i.e., c.109G>C(p.Ala37Pro), c.205G>A(p.Asp69Asn), c.244C>A(p.Gln82Lys) and c.322A>C(p.Thr108Pro) and an insertion c.49_50insT in homozygous condition. All these variants were predicted to be disease causing according to Polyphen 2 and SIFT prediction. In addition, we also identified three missense variants including c.71A>C (p.Gln24Pro), c.133G>C(p.Val45Leu) and c.322A>C(p.Cys284Tyr)in heterozygous conditions in GSD Ia affected cases. All these disease-causing variants have not been reported yet. Variants identified in coding sequence of exon 1 are c.71A>C (g.151A>C) with protein change p.Gln24Pro, c.109G>C (g.189G>C) with protein change p.Ala37Pro, c.133G>C (g.213G>C) and c.205G>A (g.285G>A) with amino acid changes p.Val45Leu and p.Asp69Asn respectively. An insertion disease causing variant also identified in 5’UTRc.49_50insT. The variant c.322A>C (g.3226A>C) and c.244C>A (g.3148C>A) identified in coding sequence of exon 2 are associated with protein changes p.Thr108Pro (replacement of a hydrophilic with hydrophobic amino acid) and p.Gln82Lys (replacement of acidic with basic amino acid) respectively and were predicted to be damaging with Polyphen score of 1 and 0.907. A missense damaging variant c.322A>C (g.10407G>A) with amino acid change i.e., p.Cys284Tyr (replacement of sulfur containing amino acid with aromatic amino acid) is identified in coding sequence of exon 5 which is predicted to be damaging with Polyphen score of 0.771 and SIFT score of 0.08. Sequencing of exon 1 of patient ID 5 and patient ID 11 identified homozygous i.e. c.205G>A (p.Asp69Asn), c.109G>C(p.Ala37Pro) and heterozygous i.e., c.133G>C(p.Val45Leu), c.71A>C(p.Gln24Pro) disease causing variants in both patients, however sequencing identified no polymorphism in these two cases. Identified homozygous variants in both patients were predicted to be probably damaging according to Polyphen 2 prediction, while both heterozygous variants were benign highlighting the need to sequence remaining non coding and regulatory sequences of gene to identify the molecular genetic defects underlying the disease phenotype. For validation of all variants, sequencing with both forward and reverse primers was done. The previously reported hotspot pathogenic variant i.e., p.Arg83Cys that was detected in 50% of alleles in French and Tunisian patients, 80% of Sicilian and 100% of alleles in Ashkenazi Jewish patients is not identified in this study cohort [ 28 ]. One of the possible explanation is small sample size of present study and a need of future large cohort studies that might identify such previously known hotspot variants from our Pakistani population.
Glucose 6 phosphatase is anchored in ER membrane by nine transmembrane helix structures, the amino terminus lies in the lumen of ER while carboxy terminal in cellular cytoplasm [ 29 ] and all of our identified missense disease-causing variants are detected in the transmembrane helix structures of G6PC . Shieh and Angaroni., 2003 have suggested that the majority of helical missense variants cause decreased stability of G6PC protein compared to the wild-type enzyme [ 30 ]. Although the functional studies could not be performed to confirm damages caused by pathogenic variants at protein level but the bioinformatic analysis demonstrated the pathogenic statuses of identified variants. Hence identification of homozygous missense disease-causing variants in five cases and heterozygous variants in three cases provide the molecular genetic basis of clinical manifestations of the GSD Ia in these patients.
In fourteen cases showing clinical symptoms of GSD Ia, no disease-causing variants in coding regions was found, which highlights the ratio of G6PC disease causing variants in our study to be 30%, however there is still need of further molecular studies since due to overlapping clinical presentations of glycogen storage disease types. Identification of variants in this study provided additional tool for genetic counselling.
Despite many advances at molecular genetics level, there are still a number of inconsistencies in GSD 1a that remained unresolved, i.e., the etiology of renal and liver disease in GSD-Ia remains unclear, phenotypic heterogeneity and the lack of a stringent genotype-phenotype in GSD-Ia. It is necessary to conduct extensive genetic studies in the local population due to a high suspected incidence of disease, a lack of molecular genetic data, the clinical heterogeneity of GSD with challenging disease diagnosis,the high mortality and economic burden of end-stage disease treatment, such as liver transplantation. Early genetic diagnosis of affected individuals and their asymptomatic family members will be aided by these tests, which will enable regular follow-ups to enhance patient management and genetic counselling. Research for genotype-phenotype correlation of local GSD Ia patients is required to help our health care providers in better understanding of GSD’s clinical presentation, minimizing the risk of morbidity and mortality due to late diagnosis in the future. Common laboratory findings of GSD Ia and Ib are hypoglycemia, hyperlipidemia, hyperuricemia, and lactic acidemia. GSD type Ib, shows systemic infections like stomatitis, Crohn-like enteritis because of neutropenia, neutrophil, and monocyte dysfunction [ 21 ], these symptoms were not observed in enrolled cases in this study. 2-step diagnostic procedure without liver biopsy to confirm the clinical diagnosis of GSD Ia: the plasma biotinidase assay followed by the molecular analysis of the G6Pase gene are needed to avoid the misdiagnosis. Although inflammatory bowel disease is a characteristic of GSD 1b, but based on other symptoms and based on the diagnosis by the expert physicians G6PC gene was screened for variants in selected patients. Patients might be misdiagnosed due to phenotypic overlap, so screening of SLC3A4 gene in the cases is our future consideration which was not possible now due to limited funding. There have been some studies on the clinical features and complications of GSD Ia in Pakistan, there is still a need for more research on the impact of the disease on patients’ quality of life. This could help inform more patient-centered approaches to care and improve outcomes for individuals with GSD Ia in Pakistan. Overall, there is a need for more research on GSD Ia in Pakistan, particularly in terms of epidemiology, genetics, treatment, and patient outcomes. Addressing these research gaps could help improve our understanding of the disease and lead to cost effective prevention and management strategies.
Supporting information
S1 Table
Frequency of demographic and clinical variants in 40 enrolled cases.
(DOCX)
S2 Table
Unidentified individuals data.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Creating a novel anatomical atlas requires a technique showing the organ in a different perspective (e.g. detailedness, staining, and tissue-fidelity), or that the applied method results in enhanced image quality compared to a previous atlas. In order to validate the need for cryosectioning of a canine brain to create a new comparative image series, we provide an overview of the main techniques that make possible considering the requirements mentioned above. There are several ways to visualize macro- or microanatomical structures in anatomy: conventional preparations and sections can be made shortly post mortem on a fresh cadaver, or previously fixed with a fixative agent [ 1 ], creating macerated bones and skeletons [ 2 – 4 ], or corrosion casting [ 5 – 7 ]. The result of the tissue preparation can be captured in photographs or videos, or these procedures can be combined with the different imaging methods (like CT or MRI), so that the selected region or specimen can be digitized for 2- and 3-dimensional analysis. These techniques are all suitable methods to visualize the different systems of the body, and they can be grouped by direct/indirect methods and tissue maintaining/tissue destructive methods. Direct imaging means that the original tissue can be seen either in its real color (e.g. with endoscopy) or dyed using histological staining [ 8 ]. Indirect imaging, such as computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET) and single-photon emission computed tomography (SPECT) provides a computer-generated picture (most commonly a grayscale image), and during post-processing an artificial color is added for differentiation based on tissue properties. Examples include X-ray attenuation (CT), or spin and precessing of the proton and water diffusion (MR) [ 9 ]. Indirect imaging methods are tissue maintainers, but have some limitations to show the real environment (mostly due to small spatial resolution). The central nervous system is enclosed into a bony capsule, and in the case of larger animals and humans the head cannot be sectioned together with the skull due to the hardness of the bone—although the present methods make it possible to produce histological sections as large as a complete human brain [ 10 , 11 ]. Plastination of the slices is a good method for tissue preservation and selective staining is possible [ 12 , 13 ], but the specimen’s color is partly altered during the procedure and shrinkage could occur [ 14 – 16 ].
If one wants to study the brain in situ , then different imaging methods or sectioning of the entire head are required. Diagnostic imaging techniques create post-processed indirect grayscale images and the quality of the CT and MR imaging depends on spatial resolution, signal to noise ratio (SNR) and various artifacts [ 17 – 20 ]. In order to create true-color macro-anatomical sections, which involve the entire neurocranium, there are two main possibilities: slicing the object into layers with a saw/macrotome blade, or to mill the selected volume stepwise and record the resulting surfaces with a camera. In the first case the slices could be handled individually, and their average size (thickness) could vary from centimetres to millimetres. These slices can be preserved through fixation, staining and/or plastination. In contrast, milling removes a layer from the volume’s surface (tissue destructive method), and the consecutive photographs always record the upcoming polished surface. With this procedure layer thickness depends only on the applied milling technique and its precision, ranging from millimetres to micrometres. This method is called cryosectioning, or cryomacrotomisation [ 21 – 23 ]. There have been several human studies in the cryosectioning field [ 24 – 26 ], and an initiative by the National Library of Medicine in 1996, the Visible Human Project (carried out in association with the University of Colorado Center for Human Simulation), used the cryomacrotomisation to visualize an entire male human body [ 21 ]. During the past decades similar projects were made in China (Chinese Visible Human, Virtual Chinese Human projects) [ 27 , 28 ], in South-Korea (Visible Korean Human) [ 22 , 29 ], and by others who also used this technique [ 10 ]. Cryomacrotomisation of smaller animals, such as mice and rats has also been performed [ 30 , 31 ].
Cryomacrotomisation focuses on showing the macro-anatomical structures of the body, however, mapping the brain on an ultrastructural level and showing its functional connectivity requires other approaches. Below an overview is given regarding integrative projects that develop detailed brain maps on the cellular level and thus complete the macro-anatomical atlases. Several projects study the central nervous system through detailed micro-anatomical reconstructions and computer simulations. The main projects are: (a) the Blue Brain Project ( https://bluebrain.epfl.ch ); (b) the Human Brain Project ( https://www.humanbrainproject.eu ); (c) the Human Connectome Project ( http://www.humanconnectomeproject.org ); (d) the SpiNNaker ( http://apt.cs.manchester.ac.uk/projects/SpiNNaker ); (e) the BRAIN Initiative ( http://www.braininitiative.org ); (f) the Brain/MINDS project ( https://brainminds.jp/en ); (g) and the China Brain Project [ 32 ]. Some of these initiatives created structural brain maps with extremely high resolution, e.g. the BigBrain project produced a 3-dimensional reconstruction of the human brain from 7404 histological sections with 20 μm slice thickness, enabling the visibility of cells (in plane resolution of 10 μm, 2400 dpi and 16-bit color depth) [ 11 ]. This atlas also comprised the MRI dataset of the same brain. Another project, in the framework of the Allen Institute for Brain Science, created a publicly available multimodal gene expression atlas [ 33 , 34 ]. They presented MRI and DWI images together with 1356 high resolution (1 μm per pixel) histological sections marking 862 individual structures (including both macroscopic structures like gyri and sulci and also making cytoarchitectural parcellations connected to Brodmann areas). Recent technologies, like optical coherence tomography (OCT) or light sheet fluorescence microscopy (LSFM) are capable of showing the biological tissues without destroying them. Tissue maps produced by these technologies serve as good diagnostic tools, for example the OCT is used in ophthalmology [ 35 ], cardiology [ 36 ] and brain research [ 37 ], whilst LSFM visualizes tissues with subcellular resolution [ 38 ].
The cryosectioning of an entire dog was first performed in 1999 [ 39 , 40 ], then a study on a whole body of a one-year-old female Beagle in 2014 [ 23 ], and a one-year-old short-hair cat in 2018 [ 41 ] have been published. We performed our study in order to further develop the cryosectioning technique and to provide the base for a high resolution multimodal comparative canine brain atlas for research, aid graduate and postgraduate trainings (e.g. not only with comparative images, but also with three-dimensional models) and to provide an aid in neurosurgical intervention planning, and to show what kind of major improvements can be achieved compared to the previous studies. First we made in - and ex vivo CT and MR imaging on a brain of a two-year-old female Beagle dog with adequate T1- and T2-weighted sequences. To be able to contrast the imaging techniques 1112 consecutive, high resolution, full-color images were created from the same animal’s brain in its original position within the neurocranium including the orbits during cryosectioning. This made possible the precise comparison and analysis of the individual structures in the identical coordinate system.
Materials and methods
Subject
In order to ensure that our results were comparable with other studies that show normal anatomical variations [ 23 , 42 ], and to be in accordance with previous widely accepted laboratory studies, we used the Beagle breed as a basic model animal for our study in accordance with the 3R-principle [ 43 ]. The subject, a healthy, two-year-old female Beagle dog, weighted 13.5 kg, and was vaccinated and treated against parasites according to the standard veterinary program. The research was performed in accordance with the international recommendations [ 44 ]. The animal was obtained from an official research company (National Research Institute for Radiobiology and Radiohygiene, Department of Radiobiology, Division of Animal Experiments and Experimental Animal House), which had the right to keep laboratory animals according to EU regulations and welfare criteria, in a standard kennel environment. The diagnostic imaging and euthanasia were performed on the same day upon receiving the animal, so no separate housing was necessary. The in vivo neuroimaging procedures were carried out under general anesthesia and all efforts were made to minimize discomfort for the animal. All husbandry and experimental procedures were approved by the Institutional Ethics Committee and the Hungarian Directorate for Food Chain Safety and Animal Health (PEI/001/956-4/2013).
Imaging protocol
The MR imaging was obtained using a 3T Magnetom TIM Trio whole-body MRI scanner (Siemens AG, Erlangen, Germany) with a 12-channel phased array head coil. Under the same anesthetic episode and immediately following the MRI, CT scans were obtained using a Siemens Somatom Perspective 128 slices CT (Siemens AG, Berlin and München 2013). The animal was placed in dorsal recumbency during the MR and CT procedures. Transverse slices (which in human terminology are used as axial slices) were oriented perpendicular to the defined axis of the brain (which was set through the rostral and caudal commissures). In order to avoid any unintentional movements during general anesthesia, and to ensure that both the MR and CT examination took place with the animal in the same position (both ante and post mortem), the dog was placed into a double plastic tube, which fixed the position of the body and the head separately. A system was created using plastic screws to hold the head fixed at the zygomatic arch on both sides and at the nuchal region. The holding device was checked both for radiopacity and MR compatibility to avoid any artefacts during scanning. Anesthesia premedication was performed using 6 ml of Propofol intravenously (Fresenius Kabi Deutschland, Propofol 1% MCT/LCT Fresenius emulsion for injection, 10 mg/ml), via the cephalic vein. After intubation inhalational anesthesia with 2.5 volume concentration (%V/V) isoflurane (Isoflurane USP, Abbots) was performed.
First the MR scanning was accomplished, during which the following sequences were obtained: T2-weighted sequence in sagittal plane (TR = 6000 ms, TE = 90 ms, slice thickness = 2 mm, FOV = 96x160 mm 2 and a 192x320 matrix with voxel size of 0.5x0.5x2 mm), T2-weighted imaging in transverse plane (TR = 10342 ms, TE = 90 ms, slice thickness = 2 mm, FOV = 96x160 mm 2 and a 192x320 matrix with voxel size of 0.5x0.5x2 mm), 3D T2-SPACE sequence in transverse plane (TR = 1000 ms, TE = 89 ms, slice thickness = 0.5 mm, FOV = 160x160 mm 2 and a 324x320 matrix with voxel size of 0.5x0.5x0.5 mm), T1-weighted imaging in transverse plane (TR = 300 ms, TE = 2.8 ms, slice thickness = 3 mm, FOV = 160x160 mm 2 and a 320x320 matrix with voxel size of 0.5x0.5x3 mm). After the native scans, dynamic contrast-enhanced MR angiography was also performed by giving 3 ml gadobutrol (Gadovist, Bayer Schering Pharma AG, Berlin, Germany), followed by 10 ml saline (0.9% NaCl) intravenously through the cephalic vein. During the angiographic examination a coronal T1-weighted sequence was obtained using the following imaging parameters: TR = 2.9 ms, TE = 1.1 ms, slice thickness = 1 mm, FOV = 270x360 mm 2 and a 288x521 matrix with voxel size of 0.93x0.7x1 mm). Following the MR imaging the animal was transferred in the same position to the CT unit, then CT scanning was carried out (130 kV, 30 mAs, slice thickness = 1 mm, pitch = 0.5, spiral scanning mode). The head and the cervical region up to the third cervical vertebra were scanned. For reconstruction, a soft tissue specific (J30s kernel) reconstruction and a bone specific (H70s kernel) with voxel size 1x1x1 mm was used.
After the in vivo imaging, the animal was euthanized while still under general anesthesia by administering a 1.5 ml mixture (T-61 injection) intravenously (containing 300 mg embutramide, 75 mg mebezonium iodide, and 7.5 mg tetracaine hydrochloride). Two hours later—while the animal was still in dorsal recumbency–both common carotid arteries and external jugular veins were exposed minimally-invasively near to the thoracic inlet, and cannulas were placed into the vessels. The amount of 18–18 ml red colored polyurethane resin (VytaTlex-10, urethane rubber) was injected through the common carotids in order to fill the arterial system of the head. Two and a half hours post mortem the same MR protocol (except MR angiography) was performed on the cadaver to provide the possibility for ante and post mortem comparison and to check the brain for the resin-filling. Following the repeated MR scanning the dog was placed into a -80°C deep freezer (without applying cryo-protective agents, based on studies with the same decision [ 23 , 39 ]), to minimize further lytic process taking place after death, and to prepare the tissues for the cryosectioning.
Blocking and embedding
The 3D-reconstructions were made from the acquired DICOM-images using Thermo Scientific Amira for Life Sciences 6.0 software (FEI Visualization Sciences Group, http://www.fei.com ). Using the in- and ex vivo CT and MR images we precisely defined the transverse plane of milling, and the boundaries of the frozen neurocranium block were determined ( Fig 1 ). The horizontal extension went from the level of the infraorbital foramen to the cranial vertebral incisure of the third cervical vertebra (a total of 150.19 mm). The vertical extension went from the highest point of the calvaria till the level of the upper fourth premolar teeth (a total of 80.06 mm). Based on the acquired 3D-models, a head block was made ( S1 Fig ), which was embedded into a thermoregulated box containing water-gelatin mixture (according to a previous study [ 22 ], but without adding methylene blue dye).
10.1371/journal.pone.0213458.g001
Fig 1
Localizing the boundaries of the head block.
Based on the CT-scans, three-dimensional reconstructions were made to determine the borders of the head block (red rectangle). Lateral (A) and dorsal (B) views.
Cryomacrotomisation
The cryomacrotomisation was performed with a Kondia NCT B-640 precision milling machine ( S2 Fig ). The cut surface was cleaned with 10% isopropyl-alcohol, cooled before each sectioning with a CO 2 cryo-gun (Linde AG, Germany) and grinded dry ice (3 mm pellets). The embedding box was repeatedly filled with dry ice to achieve sufficient mantle cooling around the block. Liquid nitrogen was also used to cool the block. In order to capture the individual surfaces, we used a Nikon D800 camera and an AF-S VR Micro-Nikkor 105 mm f/2.8G IF-ED lens with polarized filters and color-checkers (ISO-100, focus distance 105 mm, exposition time 1/200 sec, aperture size 3.1, F-stop f/8, with X-Rite ColorChecker passport and polarizing filters). Photos were recorded in 24-bit color depth and 300 dpi RAW pictures, the dimension of each picture was 7360x4912 pixels, which was set to the extent of the neurocranium block. Overall, we recorded 1112 images.
Image registration
The RGB images were imported into Amira (creating a single volume where voxel size was 19.5x19.5x100 μm), then DICOM images from in vivo CT and MR imaging were also brought into the same “Project View” space. Possible shifts between adjacent color images were corrected with the “Align Slices” module, then the whole volume was resampled and JPEG images were exported in all the main orthogonal (“xy”, “xz”, and “yz”) directions. CT images were aligned with the MR images with the “Register Images” module, then the whole RGB volume from the cryomacrotomisation was also fit to the MRI series, thus all the three volumes (MR, CT and cryosectioned RGB images) were in the same coordinate system. Registration steps included rigid and non-rigid transformations. The rigid transformations were rotation and translation, and the non-rigid transformations were isoscaling, anisoscaling and shearing. The metrics used in the Amira during the registration included ‘Correlation’ and ‘Normalized Mutual Information’. The primary data of the registration was the cryosectioned image volume, and the overlay data on it was the bone kernel defined CT-volume. The registration was done automatically by the software. After aligning the centers of the primary and overlay data, first a rigid, then a non-rigid transformation was applied. In the Multiplanar viewer the overlay of the two image sets could be checked in the three main orthogonal planes. Fitting of the osseous structures, which were clearly visible both on the cryosectioned and CT images, was used to check the proper alignment. It also included the inspection of the area of the lamina cribrosa, the frontal sinus, external sagittal crest, external occipital protuberance, tympanic bulla, and the placement of the basioccipital-, basi-. and presphenoid bones. After the registration was successfully completed between the two dataset, the CT series was resampled (with the ‘Resample transformed image’ module). Afterwards, the aligned CT-series was used as a primary data, and the MR-series for the overlay data, and as described previously, rigid and non-rigid transformations were applied to properly align the two datasets. After the inspection of the result, the MR-image volume was also resampled, and thus all the three image volumes were aligned with each other. Using the “Slice” module the same orthogonal views could be set on the three imaging modalities, making them directly comparable in the same position according to the global coordinate system.
Image segmentation and 3D-modeling
Using Adobe Photoshop CS3, two sets of grayscale images were created from the original RGB-series with selective filtering and enhancing the structures to filter out the arteries and veins for the semi-automatic segmentation ( Fig 2 ). In order to selectively filter the arteries and veins, Photoshop action-files were created; these performed a pre-programmed action-series on multiple images. In the case of the arteries, a ‘High pass filter’ was used to enhance the contrast, the ‘red’ channel was extracted to a separate layer from the RGB image, and selective color reducing carried out on the red channel. Subsequently, a black and white conversion was applied by removing the cyan tones, then contrast was enhanced and brightness was slightly reduced with the ‘brightness and contrast’ module. Finally, the RGB image was converted and saved in a grayscale mode to make the segmentation with Amira possible. In order to select the veins (which were already dark due to the post mortem highly deoxygenated blood they contained, which also made their distinction easier), ‘color balance’ and ‘selective color’ modules were used to increase the tone of the veins. Afterwards a black and white conversion was applied, and the contrast was increased in the ‘brightness and contrast’ module, following a decrease in the gamma correction. Finally, grayscale conversion and saving of the image file was recorded in the action file. After production of the completed two action-sets, the automatic process was started to convert all the 1112 cryosectioned images. The JPEG-image series were imported into Amira, and then the same 19.5x19.5x100 μm sized volumes were generated and aligned with the original RGB-volume. For each grayscale volume an “Edit New Label Field” module was generated, and a manually controlled semi-automatic segmentation was performed in the “Segmentation” area. The brain, bones, arteries and veins were labelled separately. Using the “Generate Surface” module, 3D stereolithography (STL) models were created from each label field. Smoothing and refinement of the STL-models were carried out in Amira and Autodesk MeshMixer (freeware, http://www.meshmixer.com ).
10.1371/journal.pone.0213458.g002
Fig 2
Subtracting the vessels from a cryosectioned image: Effects of the applied filters on the same slice.
Original color image (A), and after selective filtering for the arteries (B) and veins (C).
Results
Using high resolution photography, the brain and the surrounding structures were visualized with high acuity and detailedness, as it is shown in examples zooming on the ethmoturbinates, intrinsic lingual muscles and structures of the middle cranial fossa (Figs 3 and 4 ).
10.1371/journal.pone.0213458.g003
Fig 3
Transverse section at the level of the orbit (A) . Blue frames on image (A) show the zoomed regions (B and C). Close-up views show the detailedness of the ethmoturbinates in the nasal cavity (B) and the intrinsic muscles of the tongue (C).
10.1371/journal.pone.0213458.g004
Fig 4
Transverse section at the mid-thalamic level (A) . Blue frame on image (A) shows the zoomed region (B). Close-up view (B) shows the detailedness of the vessels (like sinus cavernosus, internal carotid arteries and branches of circulus arteriosus). https://doi.org/10.6084/m9.figshare.c.4365566.v1 .
The known major challenges when cryosectioning the brain could occur during the cooling process. Since the brain contains a large amount of water (approx. 70%) and freezing takes a long time, artefacts can occur. One of this is that rapid cooling of the tissues during the embedding procedure (using liquid nitrogen or dry ice) can result in soft tissue expansion. We could observe this effect during the process, as a small part of the splenial gyrus was moved underneath the tentorium cerebelli membranaceum, and the pyramid of the vermis moved towards the foramen magnum. The subarachnoid space and the encephalic ventricles were also smaller compared to those seen on the MRI. On the other hand, if the cooling process of the block is not properly performed and it takes a longer time to freeze the block, then discoloration of the gray matter and surface artefacts (“frozen lines”) can be seen during the cryosectioning, as we found in our previous study on a cadaveric brain ( Fig 5 ). Another important issue during cryosectioning may occur when fibrous tissues are found in the working area (e.g. tendons, or thicker perimysium and epimysium). These tend to become frayed or fibrous because, if they are not properly cooled prior to sectioning, they are not cut sufficiently by the device ( Fig 6A ). This phenomenon has also been observed by other researchers [ 21 , 23 ]. If this happens, then a manual intervention is required, e.g. using a scalpel or scissor to trim these fibers. However, if the upcoming layer before milling is appropriately cooled, this ensures optimal results in the cryosectioning ( Fig 6B ). During the cryomacrotomisation of the brain of the current study we did not see frozen lines on the surfaces, which confirmed that the cooling of the neurocranium block was appropriate, no signs of discoloration occurred, and the fiber-formation was continuously controlled in order to remove possible filaments. The gray matter hence appears in its original reddish color, and even though no tissue staining was applied, the boundary between the gray and white matter and the outline of the major subcortical nuclei can be clearly identified ( Fig 7 ).
10.1371/journal.pone.0213458.g005
Fig 5
Transverse section of a dog brain showing discoloration and frozen line artefacts.
10.1371/journal.pone.0213458.g006
Fig 6
Fiber-formation: Demonstrating the importance of proper cooling.
Arrows show the fibrous-tendinous tissue of the temporal muscle attaching on the coronoid process of the mandible (A). After the treatment with dry ice and liquid nitrogen these fibers harden enough to eliminate them during the upcoming sectionings (B).
10.1371/journal.pone.0213458.g007
Fig 7
Transverse section at the level of the rostral commissure (A) and at the level of lateral geniculate body (B).
(1) Cingulate gyrus. (2) Corpus callosum. (3) Caudate nucleus. (4) Claustrum. (5) Putamen. (6) Globus pallidus. (7) Septal area. (8) Rostral commissure. (9) Optic tract. (10) Hippocampus. (11) Lateral geniculate nucleus.
The relative small (100 μm) slice interval made it possible to reconstruct the other orthogonal (sagittal and dorsal) planes without losing the detailedness of the individual structures ( Fig 8 ). This means that even on a higher magnification the structures of the computer-reconstructed slices appear as sharp, detailed and uninterrupted, as if the cryosectioning had occurred along that plane. Registering the CT and MR images to the cryosectioned volume resulted in a nearly perfect comparison between the different imaging modalities. Any possible biases from the original position are the result of the mild expansion of the tissues during the freezing process. Images from various sites show this good multimodal comparability (Figs 9 and 10 ).
10.1371/journal.pone.0213458.g008
Fig 8
Reconstruction based on the original transverse cryosectioned images.
(A) Transverse plane (original image). (B) Sagittal plane (reconstructed image). (C) Dorsal plane (reconstructed image). (D) Three-dimensional composite picture of the orthogonal views.
10.1371/journal.pone.0213458.g009
Fig 9
The three co-registered imaging modalities in a perspective view.
MR is in the transverse plane (yellow frame), CT is in the sagittal plane (blue frame), cryosectioned image is in the dorsal plane (orange frame).
10.1371/journal.pone.0213458.g010
Fig 10
Comparing the result of different imaging modalities at the same level.
(A) At the level of the olfactory peduncle. (B) At the level of the hypophysis. (C) At the level of the cerebellum. (1) Cryosectioned image. (2) CT with brain window. (3) T2-weighted MRI. (4) CT with bone window.
As arteries and veins were segmented from the cryosectioned image volume, separate 3D-models were generated and transferred into a common space, showing the difference between highly detailed segmentation and focusing only to the main branches ( Fig 11 ). This can be best seen at the area of the lamina cribrosa, where the caudal septal nasal arteries pass into the nasal cavity: on the B1 part of Fig 11 only the main contour of the ethmoidal fossa is highlighted by the larger vessels, but on B2 the small nasal branches are also visible. As all the image volumes (cryosectioned images, MR and CT series) and the surface models were registered in the same coordinate system there were several visualization possibilities. They could be shown in any association with each other, thus allowing easier identification of the vessels on the grayscale cryosectioned images ( Fig 12 ), or by determining the position of the arteries and veins (according to their 3D-models) the tracking on the MR and CT imaging could be more straightforward (Figs 13 and 14 ).
10.1371/journal.pone.0213458.g011
Fig 11
Arteries (in red) and veins (in blue) on the 3D-model with different detailedness.
(A) Lateral view. (B) Dorsal view. (1) Less detailed 3D-model. (2) Highly detailed 3D-model.
10.1371/journal.pone.0213458.g012
Fig 12
Transverse section at the level of the frontal lobe.
Composite image with the grayscale-converted cryosectioned image and the crossing vessels from the 3D-models (arteries with red color and veins with blue color).
10.1371/journal.pone.0213458.g013
Fig 13
The different imaging modalities together with the vascular 3D-models.
Composite image with the cryosectioned layer (dorsal plane, orange frame), the T2-weighted MRI (transverse plane, yellow frame), and the 3D-models of the major arteries and veins. Left rostro-lateral view.
10.1371/journal.pone.0213458.g014
Fig 14
The different imaging modalities together with the vascular 3D-models.
Composite image with the cryosectioned layer (reconstructed in dorsal plane, orange frame), the T2-weighted MRI (transverse plane, yellow frame), CT-scan (sagittal plane, blue frame) and the 3D-models of the major arteries and veins. Left caudo-lateral view.
Finally, to give further perspectives for the possible use of this kind of dataset, the interactive use of the softwares allowed us to dynamically examine the subtracted volume rendered 3D-models (cut in any arbitrary plane) in order to show only the brain and its supplying vessels from different point of views ( Fig 15 ). By setting a relative starting point (choosing any point inside the volume), all the three main imaging modalities can be visualized in different planes with the 3D-model of the skull ( Fig 16 ). With the latter method, immediate comparison of the neurocranial structures is possible by changing the slice number and switching between the planes and imaging modalities.
10.1371/journal.pone.0213458.g015
Fig 15
Surface model of the vessels, and the extracted brain model.
Models as a whole (A), and sectioned in the transverse plane (B) and the dorsal plane (C). Arteries with red, veins with blue color.
10.1371/journal.pone.0213458.g016
Fig 16
Surface model of the skull with the three main imaging modalities presented in the different orthogonal planes.
Cryosectioned image is on the dorsal (“xy”) plane. MR image is on the transverse (“xz”) plane. CT image is on the sagittal (“yz”) plane.
Discussion
One of the main advantages of cryosectioning is that any region of the body can be studied, regardless of the tissue composition including thick bones that can be easily cut. Therefore, high resolution RGB-images can be obtained, and if the layer thickness is small enough, multiplanar reconstructions (MPR) can be made in any arbitrary plane by using specially designed softwares (e.g. Amira or 3DSlicer). When structural imaging techniques (MR, CT) are associated, the volumes can be registered together, thus multimodal atlases can be easily generated in the same coordinate system. As ante mortem functional studies, tractography of neural pathways (Diffusion Tensor Imaging, DTI), PET (Positron-Emission Tomography) and SPECT (Single Photon Emission Computed Tomography) imaging could also be done prior cryosectioning. The segmentation of individual structures or systems makes it possible to create three-dimensional models beneficial to several fields, including education (e.g. understanding the relationship between the 2D images series and the 3D-structure), research (comparative neuroanatomy) and clinical work (reference atlas for studying the conservative brains regions during surgical planning). As in the current study the brain was not removed from the neurocranial cavity, which is usually required to allow histological or macroscopic slicing [ 42 , 45 ], the cryosectioned structures can be seen in their original position (depending on the occurrence of freezing expansion or shrinkage). For example, the vessels around the brain and the nerves, which are entering and leaving the skull through different foramina can be traced around the osseous structures and the central nervous system. As no fixative agent was used in our study, and only the arteries were filled with red polyurethane (without producing extravasates), all the tissues can be seen in their original color. This was also confirmed by using a color checker passport, which helps with camera and image calibration during the post-processing procedure of the cryosectioned images.
The time required for the entire process can be divided into three main parts: the pre-sectioning, sectioning and post-sectioning work. (1) The pre-sectioning workflow includes the implementation of the diagnostic imaging techniques (e.g. CT and MRI), freezing the body and blocking the region of interest, constructing the embedding device, establishing the angle of cut in accordance with the imaging techniques orthogonal planes (so the milled surface could be parallel with them) and embedding the block. (2) The cryosectioning part comprises the preparation of the milling device, the cooling-, camera- and computer systems, and taking the images. (3) The post-sectioning part covers the photography post-processing (e.g. setting the proper color balance based on the color checker, equalizing shadows, and generating JPEG or TIFF images from the raw CR2 image series), and depending on the aim of the study, the image analysis or 3D-modelling (which consists of the volume-registration, segmentation, and 3D-modelling work at the required level of detailedness). The time of workflow is affected by the number of people who contribute to the research (e.g. in the case of multiple species, organizing a parallel work to progress is advised), the stops during the cryosectioning (in our case the milling procedure of the canine neurocranium took approximately 40 hours). The pre-sectioning work required two weeks, photo-conversion of the images at the post-sectioning process took three days. The post-sectioning image analysis and 3D-modelling work solely depends on the research goals (e.g. number of modelled structures and their detailedness), and the number of people working simultaneously on a project.
Known technological limitations are mostly dependent on the milling device, and the cryogen process itself. There is no technical limitation regarding the species, as any tissue can be sectioned by the milling device, which was originally designed to work with metals (thus, even sites with surgical implants could be cryosectioned). If the block is too large it can be divided into smaller parts, as it was done during the Visible Human Project. Imbalanced cooling could cause artefacts, which include discoloration, surface frozen lines or expansion/shrinkage of the tissues, as described before. The 100 μm layer thickness can be still reduced, as the milling device is capable of working with a 1 μm precision with metals. However, the expansion (due to the heat originating from the milling) and the refreezing (after capturing the surface, prior to the next sectioning), could cause not only temperature-fluctuations, but uneven freezing and discoloration of the surface (based on tissue quality) can also occur. Thus, going below 20–30 μm can lead to challenges with the cryomacrotomisation process, if no prior tissue fixative and/or cryo-protective agent is used. Due to the fact that the block is destroyed during the milling procedure (the device “carving down” the block layer by layer), the sectioned tissue cannot be used for other purposes after the cryomacrotomisation. Theoretically, the resolution achievable can be nearly unlimited, as with a high quality DSLR camera sensor one could not only capture the overall surface, but also the surface could be subdivided into four, twenty or hundreds of smaller regions to be captured individually. Thus, one surface can be merged from several regions, so the resolution of the final image (even with the same camera) will increase. Consequently, the pixel per centimetre ratio could be increased, and is only limited by the actual intention of the researcher. During the post-processing work the limits of the segmentation depends on: (a) the original image resolution (the smaller the pixel to μm ratio, the smaller the vessel that can be identified); (b) the contrast between the vessels and their environment (thus the brightest and the most non-tissue colored resins should be used); (c) the efficiency of the vascular filling with the polymer used; and finally (d) the computer’s capacity, as the hardware should be capable of dealing with larger datasets (e. g. tens of gigabytes) during the computational process.
Compared to other projects that aimed to section dogs [ 23 , 40 ], and a cat [ 41 ] we obtained a higher resolution and detailed RGB image series from the region of interest (ROI): (a) we focused entirely on cryosectioning the neurocranium with the brain, whilst other authors sectioned and captured entire cadavers, therefore our ROI for the head was smaller; (b) our slice interval was 0.1 mm compared to 0.2 mm [ 23 , 41 ] or 1 mm [ 40 ]; (c) the resolution of our images were higher, 7360x4912 pixels compared to 5616x3328 pixels [ 23 , 41 ] or 1928x1459 pixels [ 40 ]; (d) the pixel size was 19.5x19.5 μm compared to 100x100 μm [ 23 , 41 ] or 180x180 μm [ 40 ]; (e) we only used coloring agent for the arteries, which did not affect the other tissues compared to the formalin-fixation method used previously [ 40 ]; (f) discoloration of the brain or frozen lines inside did not occur compared to other studies [ 41 ].
When comparing the current work with canine brain and head atlases, some show the histological aspect of the brain, thus an ex situ slicing was made and tissue staining was applied [ 42 , 46 – 50 ]. During our study we did not use any tissue staining, except for coloring the arterial system, so beside that all the tissues show their original color, and still the outline of the gray matter and position of the major subcortical nuclei can be distinguished (as seen in Fig 7 ). We wanted to show the brain in its in situ position, where the surrounding vessels and the cranial nerves, the skull and associated structures can be traced. Other studies have made structural imaging of the canine head and brain: from just a few sections of the canine head and brain for demonstrational purposes with MRI [ 51 – 55 ], or with CT [ 56 , 57 ], ultrasonography [ 58 , 59 ], to detail anatomical marking on the MRI images [ 47 , 60 , 61 ], using both MR and CT imaging of the brain [ 62 ], and the creation of different MRI brain templates from breed-averaged data [ 63 , 64 ]. The brain was removed to section without performing previous diagnostic imaging in two studies. In the first, 4 mm thick transverse and sagittal plastinated slices were made from the brain [ 65 ]. In the second study, surface photographs and 18 transverse slices were created following formalin fixation [ 45 ]. The advantage of making a cryosectioned series from a single individual who underwent prior diagnostic imaging is that these images are directly comparable to each other. Due to the small slice thickness and the fact that the image volume can be reconstructed in any arbitrary plane, it could also serve as a visual aid to assist in the interpretation of previous studies if the proper slice is chosen from the Beagle brain series. Studies that made both diagnostic imaging and tissue sectioning of the canine head and brain include: (a) head CT with in situ cryosectioning of the entire head, with a slice thickness of 8 mm [ 66 ]; (b) CT made with an ex situ brain slicing, creating 17 transverse sections with a 5 mm interval [ 67 ]; (c) in situ formalin-fixed sectioning with 9 transverse head sections (from which four included the brain), with two transverse and one sagittal MR images [ 68 ]; (d) ex situ formalin-fixed and stained sectioning with previous CT-imaging, which resulted in 9 transverse sections of the brain [ 69 ]; (e) ex situ , formalin fixed and histologically stained brain sections with previous MR-imaging, producing 12 transverse sections (thickness were 3 to 10 millimetres), and three of the brain slices were correlated with the MR images [ 70 ]; (f) ex situ formalin-fixed brain slicing with prior MR and CT imaging, where 9 transverse, 4 dorsal and 3 sagittal images were compared between the three methods [ 71 ]; (g) in situ formalin-fixed sectioning of the entire head, with previous CT and MR imaging, creating 18 transverse sections of the entire head, which comprised 8 transverse sections where the brain could be seen [ 72 ]. When comparing our dataset to these, the advantages and novelties of the current cryomacrotomisation study are: (a) we used only one dog to obtain all the diagnostic and cryosectioned images; (b) we performed both CT- and MR scanning; (c) we repeated the MR imaging post mortem to have the basis for comparing ante and post mortem changes; (d) we did not use formalin fixation or dyes in order to show the original tissue color; (e) there was no need for the removal of the brain for sectioning, or decalcification of the bones; (f) focused the entire cryosectioning process and the photography only on the neurocranium; (g) used image capturing with a high level of detailedness (24-bit color depth, 300 dpi, 7360x4912 pixels and a pixel size of 19.5x19.5 μm); (h) obtained sectioning interval was only 100 μm; (i) originally 1112 sections of the neurocranium were made in the transverse plane, but due to small voxel size with the software-based volume-rendering, detailed images in any other orthogonal or oblique plane could be created (as Fig 8 shows), or resample with the required slice interval; (j) the registration of the cryosectioned volume with the MR and CT-data could provide directly comparable images ( Fig 10 ); (k) the segmentation of anatomical structures and 3-dimensional reconstruction can be carried out, used on their own, or can be merged with the 2-dimensional imaging data; (l) after normalization to another image volume, the results are comparable with previous studies, and subsequent imaging surveys could be integrated (e.g. matching with a cryosectioned brain when cryo-protection is used).
Development in conventional image analysis and 3D-image fusion is essential as diagnostic imaging methods are gaining more importance in small animal veterinary medicine [ 73 , 74 ] and in fMRI studies [ 75 ]. The brain of the dog as a species was previously examined with CT and MRI in several journal publications during the last decades [ 51 , 52 , 60 , 76 , 77 ], it was compared in textbooks [ 78 – 80 ], and histological [ 42 ] and diagnostic imaging atlases [ 61 – 64 , 72 , 81 ]. Furthermore, due to similarities in development, aging and comorbidity between humans and dogs, investigations have been recently focused on the canine central nervous system [ 63 , 82 – 85 ]. Veterinary educational modules are also published recently to help graduate learning [ 86 , 87 ]. Compared to those, one of the advantages of our methodology in 3D-modeling is that the structures which could be created are more realistic due to the detailedness and surface morphology because of the original high-resolution cryosectioned images. The knowledge gained by making this study was also successfully used while planning of a transsphenoidal brain surgery in the case of a dog with pituitary adenocarcinoma [ 88 ]. These researches and interests highlight the importance of these type of studies, that allow direct comparison with the most up to date diagnostic imaging methods and histological atlases.
Conclusions
The improved method utilized for cryomacrotomisation in the current study has proved to be successful to serve as a reliable base for a comparative, multimodal brain imaging atlas. As the cryosectioning procedure is not equivalent with a histological study, but it represents a macro-anatomical cross-sectional approach, the desirable resolution for future studies depends on the aim of the study: we believe that the resolution we used is enough for reliable comparison with the CT and MRI series, and for the segmentation of major anatomical structures. When investigating smaller structures and thinner layers (below 20–30 micrometres) we recommend the standard histological procedures. This is the first study in dogs that has visualized the brain using comparative imaging modalities resulting in excellent quality and detailedness (high resolution images with 0.1 mm layer thickness). Possibilities also include the application of 3D-modeling and 3D-printing to enhance learning in graduate and postgraduate studies. In the results section we showed the main advantages of the improved cryosectioning technique, and gave several examples through multiplanar reconstructions how it can be an aid in the comparative imaging by merging the diagnostic imaging modalities (CT, MR) with cryosectioned images and three-dimensional models. We also plan to use cryo-protective agent before freezing the block in future works. Due to the detailedness of the images, the image sets are suitable for selective structure extraction and exact segmentation, as was shown with the skull, brain, arteries and the veins. Thus, the resulting models can be exported, zoomed into and studied in any perspective, which would be an excellent support to accompany textbooks on the subject matter. We are also considering the fact of inter-subject variability would be good to complete these studies results, thus obtaining cryosectioned image series from dogs with dolichocephalic, mesocephalic and brachycephalic skull types are advantageous. Finally, the cryosectioning technique used, provides a unique tool for examining any parts of the body, no matter the hardness of the tissues. As a result, future works can use this technique in order to provide helpful material for educational, scientific and medical purposes.
Supporting information
S1 Fig
The embedding box (A) and the frozen head block inside the holder (B).
(TIF)
S2 Fig
The cryomacrotomisation process.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The adult corneal epithelium is a constantly renewing tissue and it is widely accepted that, during normal homeostasis, it is maintained by a stem cell population in the basal limbal region that proliferates slowly unless stimulated by injury [1] , [2] . These limbal epithelial stem cells (LESCs) give rise to fast-dividing transient (or transit) amplifying cells (TACs), which migrate centripetally in the basal layer of the corneal epithelium [3] , [4] , [5] . Here they proliferate for a limited time before undergoing a final division, whereupon both daughter cells usually detach from the basement membrane, move vertically (apically) through the suprabasal layers, becoming terminally differentiated and are eventually shed from the most superficial layer [6] , [7] .
The absence of reliable markers, able to distinguish adult stem cell populations from early TACs in the corneal epithelium, means that various indirect methods have been used to deduce that the basal limbal epithelium is the niche for corneal epithelial stem cells. Two threads of information from mouse studies have been important: the demonstration of centripetal migration of corneal keratinocytes from the limbus towards the central cornea [4] , [5] and the identification of putative stem cells as slow cycling ‘label-retaining cells’ (LRCs).
Early studies revealed that a characteristic feature of epithelial stem cells is that they divide relatively infrequently [8] and a widely held hypothesis is that stem cells are generally slow cycling during normal homeostasis but they can be induced to proliferate faster after injury. Dividing cells can be labelled by incorporating a label into the DNA (e.g. bromodeoxyuridine, BrdU) and to ensure slow cycling cells are labelled, the animals can be exposed to the label for a prolonged period. This is followed by an extended chase period, which dilutes the label more quickly in more rapidly dividing cells so revealing slow-cycling putative stem cells by their ability to retain the label. In the wild-type (WT) ocular surface, LRCs are found in the basal layer of the conjunctival and limbal epithelia, whereas the corneal epithelium is usually devoid of such slow-cycling cells [2] , [6] , [9] , [10] , [11] , [12] , [13] .
Human aniridia is an inherited eye disease caused by heterozygosity for a defective PAX6 gene. The phenotype involves developmental eye abnormalities, including a reduced or absent iris, [14] , [15] , [16] , [17] , and postnatal corneal deterioration known as aniridic keratopathy or aniridia-related keratopathy (ARK) [18] , [19] , [20] . The mouse Pax6 Sey-Neu mutant allele is considered to be a Pax6 − null allele and heterozygous Pax6 +/Sey-Neu (here abbreviated to Pax6 +/− ) mice have small eyes, hypoplastic irides and a range of other ocular abnormalities, including corneal deterioration, so provide a good model for human PAX6 +/− aniridia and ARK [21] . Some mouse Pax6 +/− corneal abnormalities arise during development (e.g. the corneal epithelium is already thinner than normal by embryonic day 18.5 (E18.5) [21] ) whereas other abnormalities arise during adulthood (e.g. blood vessels invade the corneal stroma, goblet cells accumulate in the corneal epithelium and centripetal epithelial cell movement is disrupted) [21] , [22] .
The corneal epithelial deterioration seen in PAX6 +/− humans and Pax6 +/− mice can be considered to represent abnormal tissue homeostasis. From a purely quantitative perspective, tissue homeostasis can be defined as the maintenance of an approximately constant cell number in a renewing tissue and so involves a balance of cell production and cell loss. However, to conserve full tissue functionality it is critical that lost cells are replaced by cells of the appropriate phenotype. Thus, tissue homeostasis can be considered to have both quantitative and qualitative aspects. Normal corneal epithelial homeostasis requires maintenance of an adequate number of cells of the phenotype required to ensure the cornea is able to maintain full transparency and an adequate barrier function. The accumulation of goblet cells that occurs during the deterioration of the corneal epithelia in PAX6 +/− humans and Pax6 +/− mice implies that homeostasis is qualitatively abnormal. This could arise either because corneal epithelial cell differentiation is abnormal or because conjunctival or other cells encroach onto the corneal surface to compensate for a numerical deficiency in corneal epithelial cells caused by reduced cell production and/or excessive cell loss.
It has been proposed that LESC deficiency is the principal cause of deterioration of the ocular surface in aniridia patients [20] , [23] . Indirect evidence from analysis of coherent clonal lineages in adult mouse X-inactivation mosaics, transgenic mosaics and chimeras predicts that LESCs are also either numerically deficient and/or functionally defective in heterozygous Pax6 +/− mice [22] and older WT mice [24] , [25] , [26] . These mice exhibit radial stripes, which are thought to represent clonally related populations of corneal keratinocytes migrating from a LESC population at the periphery. Pax6 +/− X-inactivation mosaic corneas had fewer radial stripes than WT mosaic corneas at 15 weeks [22] . This was assumed to imply either that LESCs were reduced in number or qualitatively defective but this could also be explained by an alternative hypothesis. The diagrams in Fig. 1 represent production of radial corneal stripes after LESC activation in the WT ocular surface ( Fig. 1A–D ) and illustrate two hypotheses, which could account for why Pax6 +/− mosaic corneas have fewer, wider stripes. Hypothesis 1 ( Fig. 1E–H ) proposes that fewer Pax6 +/− LESCs are initially specified during embryonic development (or some LESCs fail to survive), so there is a stem cell deficiency. Hypothesis 2 ( Fig. 1I–L ) proposes that the extent of cell mixing during development is less in the Pax6 +/− ocular surface than in WT, so the mosaicism is more coarse-grained (with larger coherent clones) and clonally related LESCs are more likely to be adjacent to each other. It is important to evaluate whether cell mixing is reduced during development (hypothesis 2) before drawing the conclusion that the mosaic analysis predicts that LESC function is impaired in Pax6 +/− mice (hypothesis 1).
10.1371/journal.pone.0071117.g001 Figure 1
Alternative hypotheses to explain the reduction in corrected stripe numbers in Pax6 +/− mosaic corneas.
(A–D) Normal development: The developing mosaic surface ectoderm (A) comprises two genetically marked cell populations (shown as blue and yellow hexagons) and the future corneal epithelium is shown as a disk with the limbus at its periphery. LESCs are specified (shown as stars in B) from a pool of cells in the surface ectoderm and become active postnatally (C). Centripetal movement of daughter TACs forms radial stripe patterns in the adult corneal epithelium (D). (E–H) Hypothesis 1– reduced Pax6 +/− LESC numbers: If fewer LESCs are specified in the Pax6 +/− ocular surface (F) then individual clones of corneal epithelial cells produced by each LESC will colonise a larger sector of the cornea, forming fewer, wider stripes (H). A similar result is expected if normal numbers of LESCs are specified but some fail to survive. (I–L) Hypothesis 2– reduced cell mixing during Pax6 +/− development: If there is less mixing of the two genetically marked cell populations in the Pax6 +/− mosaic surface ectoderm during development, cells will be grouped into larger coherent clones to form a coarse-grained mosaic pattern (I). There is a higher probability that two adjacent stem cells belong to the same population (e.g. adjacent blue stem cells in J) so wider stripes are produced. Although, in this case, the distribution of LESCs around the circumference will be non-random, the distribution of LESC clones should still be random. The corrected stripe number described in the text is expected to be proportional to the number of LESC clones but the number of LESCs per clone may vary as shown (e.g. compare H and L).
Other factors also probably contribute to abnormal corneal homeostasis in Pax6 +/− heterozygotes. Two observations suggest that cell turnover is more rapid in the corneal epithelium. First, the frequency of BrdU labelled cells has been reported to be higher in Pax6 +/− mice than WT mice, either for the whole corneal epithelium [27] or specifically the basal layer [28] . Second, ex vivo evidence suggests that the Pax6 +/− corneal epithelium is more fragile than normal [27] . This is consistent with observed abnormalities in cell adhesion molecules, junctional complex proteins and actin-based cytoskeletal structures plus the reduced expression of keratin 12 (K12) [27] , [28] , [29] . K12 is a specific marker of corneal epithelial differentiation that is regulated by Pax6 [30] , [31] , [32] , so its reduced expression in Pax6 +/− corneal epithelia suggests that differentiation of the Pax6 +/− corneal epithelium is abnormal [27] , [28] . Down-regulation of K12 also probably contributes to Pax6 +/− epithelial fragility, as reported for K12 knockout mice [33] . This increased fragility prompted the proposal that the Pax6 +/− cornea is in a chronic wound-healing state [29] , [34] . These observations all predict that Pax6 +/− corneal epithelial cell loss would be greater than normal but this has not been investigated directly.
The aims of the present study were: (1) to compare the rate of movement of cells from the basal corneal epithelium to the suprabasal layers in Pax6 +/− and WT mice, as this is the irreversible first step of cell loss; (2) to re-evaluate the prediction from studies of mosaic mice that Pax6 +/− mice have fewer LESCs by testing whether the coarser mosaic patterns could be accounted for by reduced cell mixing during development of the Pax6 +/− ocular surface, so producing larger coherent clones (hypothesis 2 shown in Fig. 1I–L ); (3) to compare the number of label-retaining cells (LRCs) in the limbus of Pax6 +/− and WT mice at two ages (15 and 30 weeks), as a means of comparing relative LESC numbers, to test whether Pax6 +/− mice have fewer LESCs than WT (hypothesis 1 shown in Fig. 1E–H ) and whether LESC numbers decline with age in WT mice.
Materials and Methods
Ethics Statement
All the animal work in this study was approved by the University of Edinburgh Ethical Review Committee (application PL21-06) and performed in accordance with UK Home Office regulations under project license number PPL 60/3635. All surgery was performed under general anaesthesia and all efforts were made to minimise suffering.
Mice
Heterozygous Pax6 +/Sey-Neu ( Pax6 +/− ) and Pax6 +/+ , wild-type (WT) littermates were maintained on a congenic CBA/Ca genetic background. They were distinguished by eye size and their genotypes confirmed by PCR [35] . Transgenic H253 mice [36] , which ubiquitously express the X-linked Tg(Hmgcr-LacZ)H253Sest, nLacZ transgene (abbreviated to XLacZ ), were maintained on a mixed genetic background (predominantly C57BL/6 and CBA/Ca). Hemizygous males and females are designated respectively XLacZ Tg/Y and XLacZ Tg/− . Both Pax6 +/− , XLacZ Tg/− and WT, XLacZ Tg/− X-inactivation mosaics were produced from Pax6 +/− female × XLacZ Tg/Y male crosses. Groups of mice were compared at 15 and 30 weeks because previous comparisons of WT X-inactivation mosaics showed that an age-related decline in corrected corneal epithelial stripe numbers could be detected between these ages [22] , [24] , [25] , [37] . We, therefore, compared both corrected corneal epithelial stripe numbers and the numbers of limbal label-retaining cells in Pax6 +/− and WT X-inactivation mosaics at 15 and 30 weeks, as described later. Differences between these ages are considered to be age-related differences but not effects of “old age” as laboratory mice can live to about 2 years [38] . Mice were bred and maintained in animal facilities at the University of Edinburgh and killed by cervical dislocation following inhalation of gaseous anaesthetic.
BrdU Treatment
For acute BrdU labelling of the adult cornea, 12-week old mice were given single intraperitoneal (i.p.) injections of BrdU (10 mg BrdU/ml saline; 0.2 ml/mouse) at 10.00 am. Mice were killed after 4 hours or 1, 3, 7 or 14 days and samples collected. For BrdU pulse-chase identification of label-retaining cells (LRCs), 0.1 ml Alzet mini-osmotic pumps (model 1007D, Charles River, UK Ltd.), containing BrdU solution (0.1 ml; 50 mg BrdU/ml in saline) were surgically implanted, sub-cutaneously, under Isoflurane anaesthesia, into 15- or 30-week old mice to provide 0.5 µl BrdU solution/h. The pumps were removed after 7 days and mice were killed 10 weeks later and samples collected.
BrdU Immunohistochemistry
Eyes were fixed in 4% paraformaldehyde for 24 h. and embedded in paraffin wax after dehydration in a graded ethanol series. Longitudinal sections were cut at 7 µm thickness, mounted on Polysine slides (VWR laboratory supplies), deparaffinised in Histoclear and processed for BrdU immunoshistochemistry with a 3-3′diaminobenzidine (DAB) endpoint and lightly counterstained with haematoxylin, as described previously [28] . However, for antigen retrieval, sections were immersed in citrate buffer (10 mM sodium citrate pH 6.0) in a covered Coplin jar in a 90°C water bath. Control slides were incubated with normal serum without the primary antibody but otherwise treated in the same way.
Analysis of Cell Proliferation and Loss
After BrdU immunohistochemistry, calibrated digital images of mid-sections of the corneal epithelium were captured with a Nikon Coolpix 995 digital camera on a Zeiss Axioplan 2 compound microscope. Every basal corneal epithelial cell was scored as positive or negative, across the corneal diameter in a section from the middle of the eye. These data were divided into six equal sized groups according to corneal region, so the % BrdU-labelled nuclei in the basal layer (basal layer labelling index, LI) could be determined separately for two peripheral (P), intermediate (I) and central (C) regions which were then combined to provide one mean value per eye for the P, I and C regions. The % BrdU-positive cells in all the suprabasal layers (suprabasal LI) were scored in an equivalent way. The Pax6 +/− corneal epithelium has fewer cell layers than WT and this could exaggerate or mask differences in suprabasal LI. To take account of this, an adjusted % BrdU-positive cells in the suprabasal layers (adjusted suprabasal LI) was calculated by expressing the number of BrdU-positive cells in the suprabasal layers as a percentage of the total cell number in the underlying basal layer rather than the total cell number in the suprabasal layers.
Whole-mount Cornea Immunofluorescence
Whole-mount corneal immunofluorescence was adapted from a published method [39] . Eyes were dissected, washed with PBS, fixed in 4 methanol : 1 DMSO and rehydrated through a graded methanol series to PBS. Corneal buttons were dissected, incubated for 2 h. in blocking serum, comprising PBS-T (PBS containing 0.02% Tween20), 1% BSA and 1% goat serum. They were then incubated with primary antibody overnight at 4°C, washed (PBS-T; 5 h.), treated with blocking serum and incubated with secondary antibodies overnight at 4°C. Finally, corneas were washed (PBS-T; 4 h.), counterstained with TO-PRO3 iodide (Invitrogen) containing RNaseA, flattened with radial cuts and mounted on slides with MoWiol 4–88 (Calbiochem) containing 2.5% DABCO (1,4-diazobicyclooctane). For LRC analysis, each cornea was flat-mounted with 8 radial cuts, creating 8 corneal sectors.
The primary antibodies used were rat anti-BrdU (Abcam), rabbit anti-K19 (LifeSpan Biosciences) and rat anti-CD31 (BD Pharmingen). The secondary antibodies were goat Alexa 488 Anti-Rat IgG (Invitrogen), goat Alexa 488 Anti-Rabbit IgG (Invitrogen) and goat Alexa 568 Anti-Rat IgG (Invitrogen). For double immunofluorescent staining, antibodies were applied sequentially and incubated overnight at 4°C. To image K19 and CD31 immunofluorescence, overlapping Z-stack images were acquired with a Leica TCS NT confocal microscope and exported to ImageJ ( http://rsb.info.nih.gov/ij/ ). Composite images were constructed using the MosaicJ plug-in ( http://bigwww.epfl.ch/thevenaz/mosaicj ). BrdU immunofluorescence analysis is described below.
Analysis of BrdU Label-retaining Cells in Whole Mount Corneas
Eight Z-stack images, comprising 5–6 optical sections at 2–3 µm intervals that included the limbus, were acquired per cornea (1 for each of 8 sectors), in a clockwise direction with a Zeiss LSM510 CLSM confocal microscope using a ×20 lens. Images (with the limbus in the middle) were acquired at 1024×1024 pixel resolution. Individual photographs were compiled using Zeiss LSM Image Browser software (version 4.2.0.121) and projected to a single image based on maximum intensities. ImageJ was used for thresholding, to include particle sizes of 150–5000 voxels (3–15 µm diameter nuclei). A calibrated 340×200 µm sampling box was superimposed on the image across the width of the limbal region in each sector and BrdU-positive nuclei were counted. Fluorescent spots outside the size range for BrdU-positive nuclei were not included in the analysis.
Counts of BrdU-positive nuclei were expressed as the mean number of LRCs per sampling box (one sampling box per sector). To correct for differences in corneal diameters, the mean number of LRCs per sampling box was multiplied by the corneal circumference (estimated from 3 diameter measurements of the flattened cornea) and divided by the length of the sampling box (340 µm) to provide an estimate of the number of LRCs in a 200 µm wide ring around the circumference.
To reduce the effect of variability among individual immunofluorescence experiments, the mean number of LRCs per sampling box was also expressed as a normalised ‘LRC index per sampling box’. This was calculated separately for each immunofluorescence experiment that yielded results for at least one eye from each of the four groups compared. First, to accommodate experiments with unequal number of eyes in each group, a normalised total number of LRCs was calculated for each experiment as the sum of the mean number of LRCs per sampling box for each of the four groups. The ‘LRC index per sampling box’ was then calculated separately for each of the four groups as the number of LRCs per sampling box, expressed as the percentage of the normalised total number of LRCs per sampling box (for all four groups) for each experiment. The sum of the LRC indices for the four groups is, therefore, equal to 100% in each experiment. The estimated number of LRCs per circumference was similarly expressed as the ‘LRC index per circumference’.
The confocal corneal images were also used to compare cell-packing densities in the four groups, using 3 regions from each of 3 corneas per group. Nuclei were counted manually in a 150×150 µm sampling box that was superimposed on the image by an independent person using Adobe Photoshop 7.0.
β-galactosidase Histochemistry and Analysis of Mosaic Patterns in Histological Sections
Whole eyes from XLacZ Tg/− , X-inactivation mosaic mice were stained with X-gal as described elsewhere [24] , [25] . For analysis of patch sizes at 3–10 weeks, the stained eyes were processed for routine paraffin wax histology and sections were counterstained in neutral red and eosin and mounted in DPX under coverslips. Cells in the basal layer of the mid-section of the corneal epithelium were counted and scored as β-gal-positive (stained blue) or β-gal-negative (red counterstain) across the diameter of the cornea by microscopy using a ×40 objective and the numbers of cells in alternate β-gal-positive and β-gal-negative patches were recorded. The number of patches, mean patch length and median patch length were recorded separately for the β-gal-positive and β-gal-negative cell populations and the % β-gal-positive cells was calculated. The total numbers of basal cells and patches across the corneal diameter were also recorded.
Analysis of Stripe Numbers in the Corneal Epithelium of Adult X-inactivation Mosaics
Whole adult eyes were stained with X-gal and the radial stripe patterns in the corneal epithelia were analysed semi-automatically from digital photographs showing the entire cornea, using the ImageJ plugin ‘Clonal Tools’ in batch mode as described previously to provide a ‘corrected stripe number’ [37] , [40] . This corrects for the probability that stripes would contain multiple adjacent β-gal-positive corneal epithelial clones as described previously [24] , [25] . The corrected stripe number was also divided by the circumference of each cornea measured to allow for differences in eye sizes.
Statistical Analysis
GraphPad Prism software was used for statistical tests as indicated in the text and figure legends. The choice of parametric or non-parametric tests was guided, in part, by normality tests. In some cases, data (or the data plus 1, to accommodate zero values) were log-transformed for analysis with parametric statistical tests. The error bars in the figures are 95% confidence intervals (CI).
Results
Cell Proliferation in the Basal Corneal Epithelium in WT and Pax6 +/− Mice
The 12-week, WT corneal epithelium comprised 4–6 layers and the mid sections contained 424±10.1 basal cells (mean ±95% CI) and 669±21.9 broader and flatter suprabasal cells. Pax6 +/− corneas were smaller in diameter (305±10.6 basal cells) and had fewer suprabasal layers (307±10.7 suprabasal cells across the diameter). As expected, BrdU-positive cells were almost entirely confined to the basal layer four hours after labelling ( Fig. 2 ). The percentage of BrdU-positive cells in the basal corneal epithelial cell layer (basal labelling index, LI) increased slightly (but not significantly) between 4 and 24 hours in both Pax6 +/− and WT corneas and then declined ( Fig. 2E ). There were also some significant differences in basal BrdU labelling among corneal epithelial regions, most notably at 1 day ( Fig. 3A–C and legend). Although the basal LI was higher in Pax6 +/− than WT eyes at both 4 and 24 hours after BrdU injection ( Fig. 2E ) as reported previously [28] , the genotype differences did not reach statistical significance in the present experiment either when compared in whole corneal epithelium ( Fig. 2E ) or individual regions ( Fig. 3A–C ).
10.1371/journal.pone.0071117.g002 Figure 2
Acute BrdU labelling of WT and Pax6 +/− corneal epithelia.
( A, B ) BrdU immunohistochemistry 4 hours after BrdU injection of (A) WT and (B) heterozygous Pax6 +/− mice. ( C, D) BrdU immunohistochemistry 24 hours after BrdU injection of (C) WT and (D) Pax6 +/− mice. BrdU-positive nuclei in the corneal epithelium appear dark. ( E–G ) The mean (±95% CI) BrdU labelling indices for mid-sections are shown for chase periods of 4 hours (4 h) to 14 days (14 d). ( E ) BrdU basal labelling index (BrdU positive basal cells as a percentage of total basal cells). ( F ) BrdU suprabasal labelling index (BrdU positive suprabasal cells as a percentage of total suprabasal cells). ( G ) Adjusted suprabasal BrdU labelling index (BrdU positive suprabasal cells as a percentage of total basal cells). Results for 2-way analyses of variance (ANOVAs) for log transformed data are shown. Where genotype differences were significant overall, pairwise comparisons were made between genotypes for each time point using Bonferroni post-hoc tests (significant differences are shown by asterisks). Separate 1-way ANOVAs and Bonferroni post-hoc tests for each genotype showed that the frequencies of BrdU-positive cells increased in the suprabasal layers from 4 h to 3 days ( P <0.001 for both WT and Pax6 +/− in E & F) and then declined from 3 to 14 days ( P <0.001 for both WT and Pax6 +/− in E & F). Abbreviations: LI, labelling index; NS, not significant; ** P <0.01; *** P <0.001; **** P <0.0001. 6–12 eyes per group as shown within or above the bars.
10.1371/journal.pone.0071117.g003 Figure 3
Distributions of BrdU-positive cells in different regions of WT and Pax6 +/− corneal epithelia.
The mean (±95% CI) BrdU labelling indices are shown separately for the peripheral (P), intermediate (I) and central (C) regions of the cornea for chase periods of 4 hours (A, D & G), 1 day (B, E & H) and 3 days (C, F & I). ( A–C ) BrdU basal labelling index (BrdU positive basal cells as a percentage of total basal cells). ( D–F ) BrdU suprabasal labelling index (BrdU positive suprabasal cells as a percentage of total suprabasal cells). ( G–I ) Adjusted suprabasal BrdU labelling index (BrdU positive suprabasal cells as a percentage of total basal cells). In most cases statistical comparisons were made by 2-way analyses of variance (ANOVAs) of log-transformed data followed by pairwise Bonferroni post-hoc tests and separate linear regression analyses for WT and Pax6 +/− genotypes. Non-parametric Kruskal-Wallis (KW) tests followed by pairwise Dunn’s multiple comparison tests were used for D and G because there were many zero values and the log-transformed data were not normally distributed. Significant differences for the 2-way ANOVAs and linear regressions (or Kruskal-Wallis tests) are shown in each panel. The only two significant pairwise post-hoc tests between genotypes are shown by asterisks over the two bars compared (central region in E and H). The post-hoc tests between regions are not shown on the histograms. For WT corneas, only the post-hoc test between regions P vs. C in B was significant ( P <0.05). For Pax6 +/− corneas, post-hoc tests between pairs of regions were significant for P vs. C in A ( P <0.05), B ( P <0.05), E ( P <0.001) and H ( P <0.001), and for P vs. I in B ( P <0.05) and H ( P <0.05). Abbreviations: LI, labelling index; WT, wild-type; NS, not significant; * P <0.05; ** P <0.01; *** P <0.001; **** P <0.0001. 6–12 eyes per group as shown in Fig. 2E–G.
Basal Corneal Epithelial Cells Move More Quickly to the Suprabasal Layers in Pax6 +/− Mice
As Pax6 +/− corneal epithelia had fewer suprabasal layers than WT, an ‘adjusted suprabasal LI’ was calculated as well as the suprabasal LI, to compare the percentage of BrdU-positive cells in the suprabasal layers (see Materials and Methods). The frequencies of BrdU-positive cells increased significantly in both WT and Pax6 +/− suprabasal layers during the first three days and then declined ( Fig. 2F,G ). The initial increase in suprabasal BrdU labelling occurred more rapidly in Pax6 +/− corneas and the suprabasal LI was significantly higher in Pax6 +/− than WT eyes at 1 day ( Fig. 2F ), suggesting that more BrdU-labelled cells had moved from the basal layer to the suprabasal layers in Pax6 +/− corneas. The adjusted suprabasal LI showed a similar trend at 1 day but the difference between genotypes was not significant ( Fig. 2G ). However, at this time the percentage of labelled cells in the suprabasal layers was significantly higher in Pax6 +/− corneas, with or without adjustment for fewer suprabasal layers in Pax6 +/− corneas ( Table 1 ). Thus, the greater accumulation of labelled cells in Pax6 +/− suprabasal layers is not simply a consequence of more labelled cells in the basal layer or fewer suprabasal layers and implies that cells move to the suprabasal layers more frequently in Pax6 +/− corneas.
10.1371/journal.pone.0071117.t001 Table 1
Percentage of all the labelled cells that are in the suprabasal layers 24 hours after BrdU injection.
Calculation of % of labelled cellsthat are in the suprabasal layers
Mean Percentage ± SEM (N)
P -value
Wild-type
Pax6 +/−
SLI ×100/(SLI+BLI)
18.23±1.64 (7)
40.03±3.93 (10)
P = 0.0003
ASLI ×100/(ASLI+BLI
26.20±1.93 (7)
38.01±3.72 (10)
P = 0.0144
Abbreviations: ASLI = adjusted suprabasal labelling index; BLI = basal labelling index; SLI = suprabasal labelling index. P -value is for t -test with Welch’s correction for unequal variances.
Pax6 +/− Basal Cells Move More Quickly to the Suprabasal Layers in the Central Corneal Epithelium
Differences in suprabasal BrdU LI between regions and genotypes were only significant at 1 day after BrdU injection ( Fig. 3D–F and legend). Regional differences were confined to Pax6 +/− corneas, which showed a much higher suprabasal LI in the central cornea than the periphery ( P <0.001; Fig. 3E and legend). This was confirmed by linear regression analysis among regions, which was significant for Pax6 +/− ( P <0.01) but not WT corneas. At this time the suprabasal LI in the central cornea was significantly higher for Pax6 +/− than WT ( P <0.001; Fig. 3E ). The results indicate that movement of basal cells to the suprabasal layers occurs more rapidly in Pax6 +/− than WT corneas and in Pax6 +/− corneas it occurs most rapidly in the central region. The adjusted suprabasal LIs followed a similar trend to the unadjusted suprabasal LIs ( Fig. 3G–I ). It varied significantly with region and Pax6 genotype at 1 day ( Fig. 3H and legend) and again genotype differences were significant only for the central region at 1 day ( P <0.05; Fig. 3H ).
As the basal to suprabasal cell movement is irreversible and is the first step in a sequence of events that culminates in cell loss from the corneal epithelial surface, the results shown in Fig. 3E,H imply that cell loss is greater from corneas of Pax6 +/− than WT mice and that this difference is most pronounced in the central region of the cornea.
Mosaic Patterns are more Coarse-grained in the Adult Pax6 +/− Corneal Epithelium than WT
As noted in the Introduction a previous study, showing that corrected stripe numbers were lower in Pax6 +/− , XLacZ Tg/− than WT, XLacZ Tg/− mosaic corneas at 15 weeks [22] , was used to predict that LESCs were reduced in number or qualitatively defective in Pax6 +/− eyes. We confirmed these trends, using a semi-automated quantification method [40] . The striping patterns in Pax6 +/− , XLacZ +/− mosaics were often more disrupted than in WT, with fewer larger patches that frequently did not extend radially from the limbus to the centre of the cornea, suggesting that cell movement was abnormal ( Fig. 4A,B ). The semi-automated analysis ( Fig. 4D,E ) confirmed previous results obtained by manual quantification [22] . In WT XLacZ +/− corneas, the corrected stripe numbers declined between 15 and 30 weeks whereas, in Pax6 +/− , XLacZ +/− corneas, corrected stripe numbers were significantly reduced at 15 weeks but did not decline further between 15 and 30 weeks ( Fig. 4D ). These differences remained significant when the corrected stripe numbers were expressed per mm of corneal circumference to correct for the smaller size of Pax6 +/− eyes ( Fig. 4E ). These results show that the mosaic pattern is more coarse-grained (fewer and wider stripes) in Pax6 +/− than WT mosaic corneas and that in WT mosaic corneas the pattern becomes coarser between 15 and 30 weeks. The difference between Pax6 +/− and WT mosaic patterns at 15 weeks could be explained by either reduced LESC numbers (hypotheses 1 in Fig. 1E–H ) or reduced cell mixing during development (hypothesis 2 in Fig. 1I–L ).
10.1371/journal.pone.0071117.g004 Figure 4
Analysis of mosaic patterns in the corneal epithelium of adult and 3–10 week old WT and Pax6 +/− X-inactivation mosaics.
( A ) Adult WT XLacZ +/− mosaic eye showing characteristic radial stripes in the corneal epithelium. ( B ) Adult Pax6 +/− XLacZ +/− mosaic eye showing disrupted patterns with larger, more irregular patches. ( C ) Section through cornea of a 3-week old XLacZ +/− mosaic mouse eye showing β-gal positive and β-gal negative patches that were measured in the basal corneal epithelium. ( D, E ) Corrected stripe numbers per circumference (D) and corrected stripe numbers per mm of corneal circumference (E) for corneal epithelia from adult WT and Pax6 +/− , X-inactivation mosaics at 15 and 30 weeks. Statistical comparisons were made by 2-way analysis of variance (ANOVA) followed by pairwise Bonferroni post-hoc tests. ( F, G ) Corrected mean patch lengths (in cell numbers across the corneal diameter of the basal corneal epithelium) for the β-gal positive cell population (F) and median patch lengths for the minor cell population (G) in WT and Pax6 +/− mosaics at 3 weeks. Statistical comparisons were made by Student’s t -test with Welch’s correction (F) and Mann-Whitney U-test (G). ( H, I ) Mean numbers of cells (H) and patches (I) across the corneal diameter of the basal corneal epithelium in WT and Pax6 +/− mosaics at three ages. Statistical comparisons were made by 2-way ANOVA followed by pairwise Bonferroni post-hoc comparisons of genotypes and separate linear regression analyses of ages for WT and Pax6 +/− genotypes. Abbreviations: WT, wild-type; NS, not significant; * P <0.05; ** P <0.01; *** P <0.001; **** P <0.0001. In D and E, corrected stripe numbers were calculated for both eyes and the mean value for each mouse was used for analysis. In F-I, the central histological section of one eye per mouse was analysed. The numbers of eyes per group is shown within the bars. Error bars are 95% CI. Scale bars: A,B = 1 mm; C = 100 µm.
Mosaic Patterns are not more Coarse-grained Pax6 +/− Corneal Epithelium than WT at 3 Weeks
To test whether cell mixing is reduced during development of the ocular surface in Pax6 +/− mice (hypothesis 2 in Fig. 1I–L ), we measured β-gal-positive patch lengths in the basal corneal epithelium in histological sections of Pax6 +/− and WT X-inactivation mosaic eyes ( Figs. 4C,F,G ), to compare cell mixing at three weeks, which is before the stripes emerge [24] , [25] . The extent of cell mixing during development and growth of a mosaic corneal epithelium will affect the sizes of the coherent clones of contiguous β-gal-positive cells before the radial stripes emerge. The size of each β-gal-positive patch observed will depend on both the size of the constituent coherent clones and the number of adjacent coherent clones in the patch. For a one-dimensional string of β-gal-positive and β-gal-negative basal corneal epithelial cells in a tissue section, the mean number of β-gal-positive coherent clones per patch of β-gal-positive cells can be estimated as 1/(1–p), where p is the proportion of β-gal-positive cells [41] , [42] , [43] . The observed mean β-gal-positive patch length (mean cells per patch) was corrected by dividing it by 1/(1–p), to derive the ‘corrected mean patch length’, which is an estimate of the mean coherent clone length [43] , [44] .
The corrected mean patch length (calculated for β-gal positive patches) and the uncorrected median patch length calculated for the minority cell population [45] were used to compare the extent of cell mixing in the corneal epithelia of different groups of X-inactivation mosaics at 3 weeks (before stripes emerge). Both were smaller in Pax6 +/− , XLacZ +/− than WT, XLacZ +/− corneas at 3 weeks and this difference was statistically significant for the corrected mean patch lengths ( Fig. 4F ). Thus, before the emergence of stripes, the mosaic pattern is not more coarse-grained (with larger coherent clones) in the Pax6 +/− corneal epithelium than in WT, as would be predicted if the difference in adult mosaic stripe patterns was attributable to differences in cell mixing during development (hypothesis 2 in Fig. 1I–L ). Indeed, cells appeared to be more finely mixed, which may reflect the reported reduction in cell-cell adhesion between Pax6 +/− cells or between Pax6 +/− cells and the stroma [27] . We therefore rejected the reduced cell mixing hypothesis (hypothesis 2 in Fig. 1I–K ) and investigated whether adult Pax6 +/− corneal epithelia are maintained by fewer LESC clones than WT (hypothesis 1 in Fig. 1E–H ).
Activation of LESCs in the Pax6 +/− Ocular Surface
Analysis of patch lengths in mosaic corneas also provides a means of investigating the transition from patches to stripes. Fig. 4H,I shows that, between 3 and 10 weeks, the mean basal cell number across the corneal diameter increases with age for WT, XLacZ +/− corneas whereas the mean patch number decreases, implying that the patches increase in length over this period. This is consistent with observations on whole mount preparations showing that stripes emerge from the periphery at or before 5 weeks as a consequence of LESC activation and replace the randomly orientated patches formed during development [24] , [25] . Similar trends were found for Pax6 +/− , XLacZ +/− corneas ( Fig. 4H,I ), although the decrease in patch numbers was less pronounced and not significant by linear regression analysis. Nevertheless, the significant increase in cell number across the Pax6 +/− corneal epithelial diameter, coupled with a trend for a decrease in patch numbers, suggests that LESCs are probably also activated in Pax6 +/− eyes at this time and that they produce clones of cells that can extend radially across the corneal diameter.
Identifying the Limbus in WT and Pax6 +/− Whole-mount Preparations
Before testing whether limbal-LRCs were reduced in Pax6 +/− eyes (predicted by hypothesis 1 in Fig. 1E–H ), we characterised the boundary between the limbus and cornea in WT and Pax6 +/− eyes by immunostaining for the blood vessel marker CD31 and keratin 19 (K19), which is present in the limbal and conjunctival epithelia but not the corneal epithelium [46] . Double K19/CD31 immunofluorescence of WT eyes ( Fig. 5A,B ) confirmed that both markers were restricted to the conjunctival and limbal epithelia and did not extend into the corneal epithelium, so identifying a clear corneo-limbal boundary. However, in Pax6 +/− eyes, both K19 positive cells and, as previously reported [21] , CD31-positive blood vessels extended into the cornea in Pax6 +/− mice ( Fig. 5C,D ). Using TO-PRO3 as a nuclear counterstain, we were unable to identify nuclear morphological differences between limbal and corneal epithelial cells reported for whole-mount preparations counterstained with DAPI [10] . However, the K19 boundary in WT mice coincided with a morphological crease or in-folding that formed a ridge, which was apparent in nearly all the whole-mount WT specimens but less clear in Pax6 +/− eyes (data not shown). This infolding has been described previously for WT mouse eyes [47] , [48] and we used it as a marker to identify the limbus for analysis of BrdU label-retaining cells (LRCs).
10.1371/journal.pone.0071117.g005 Figure 5
Identification of the corneo-limbal boundary in the WT and Pax6 +/− mouse ocular surface epithelium.
CD31 (red) and keratin 19 (green) double immunofluorescence staining in the ocular surface of ( A, B ) WT and ( C, D ) Pax6 +/− mice. Images (B) and (D) are higher magnifications of the areas outlined in (A) and (C) respectively. Both CD31-positive blood vessels and keratin 19-positive epithelial cells are restricted to the conjunctiva and limbus in WT eyes but they both extend into the corneal epithelium in Pax6 +/− mice. Abbreviations: L: Limbus; Co: Cornea; Cj: Conjunctiva. Red immunofluorescence: CD31; green: keratin 19; blue: TO-PRO3 iodide nuclear counterstain. Scale bars are 1 mm (A,C) and 0.1 mm (B,D).
Limbal Label-retaining Cell Numbers are not Reduced in Pax6 +/− Eyes or in Older WT Eyes
Having rejected reduced cell mixing during development of the Pax6 +/− ocular surface (hypothesis 2 in Fig. 1I–K ) as an explanation of the reduced stripe numbers in adult Pax6 +/− mosaic corneas, we wanted to test whether the adult Pax6 +/− corneal epithelium is maintained by fewer LESCs than normal (hypothesis 1 in Fig. 1E–H ). Results shown in Fig. 4D,E and previous work [22] , [24] , [25] also predicted a decline in LESC clone numbers in WT mice between 15 and 30 weeks. As most LRCs are thought to be putative stem cells, we compared LRC numbers in WT and Pax6 +/− at both these ages. We exposed 15- and 30-week old WT and Pax6 +/− mice to BrdU for 7 days and identified LRCs in the limbal region of eight corneal sectors ( Fig. 6A ). LRCs were identified by confocal microscopy and counted in a 340×200 µm sampling box superimposed over the limbal region in each sector ( Fig. 6B ) for each of the four groups of mice ( Fig. 6C–F ).
10.1371/journal.pone.0071117.g006 Figure 6
Identification of label-retaining cells in the limbal region of the WT and Pax6 +/− ocular surface epithelium.
( A ) Diagram showing the 8 radial cuts in a corneal button, shaded to represent the cornea (lightest), limbus (intermediate) and conjunctiva (darkest). The rectangles show the location of the sampling boxes (one per sector but not to scale). ( B ) Rectangular 340×200 µm sampling box (yellow outline), used to count LRCs, superimposed on the limbal region of an image of BrdU-labelled nuclei (red) counterstained with TO-PRO3 iodide (blue) in a whole mount flattened corneal button with associated conjunctival tissue. ( C–F ) Examples of BrdU label-retaining cells in the limbal region after 1-week BrdU exposure and 10 week chase period in ( C ) 15-week old WT, ( D ) 15-week old Pax6 +/− , ( E ) 30-week old WT and ( F ) 30-week old Pax6 +/− . Pixel resolution: 1024×1024. Abbreviations: Cj: Conjunctiva; L: Limbus; Co: Cornea. Red immunofluorescence: BrdU; Blue: TO-PRO3 iodide counterstain. Scale bars are 100 µm.
The mean number of LRCs per sampling box in the limbal region did not differ significantly between Pax6 +/− and WT at either 15 or 30-weeks of age and did not differ between ages for either genotype ( Fig. 7A ). As Pax6 +/− eyes are smaller than WT eyes we also estimated the number of LRCs within a 200 µm wide ring (the sampling box width) around the whole limbal circumference. Again, there were no significant differences among the four groups by 2-way ANOVA ( Fig. 7B ).
10.1371/journal.pone.0071117.g007 Figure 7
Comparison of label-retaining cell numbers between WT and Pax6 +/− limbal and corneal epithelia.
( A ) Label-retaining cells (LRCs) per 340×200 µm sampling area. ( B ) LRCs per 200 µm wide ring around the limbal circumference. ( C ) LRC index per sampling area. ( D ) LRC index per limbal circumference. See text for explanation of LRC index. ( E ) LRCs that were within the cornea rather than the limbal area (more central than the sampling box shown in Fig. 6B). Results were compared by 2-way ANOVA followed by Bonferroni post-hoc tests (A–D) or Kruskal-Wallis test followed by Dunn’s multiple comparison test (E). Significant pairwise differences for genotypes or ages are shown by asterisks: * P <0.05, *** P <0.001, **** P <0.0001, NS, not significant. Error bars are 95% confidence intervals. The number of eyes per group is shown within each bar.
To reduce the effect of variation among individual immunofluorescence experiments, we also calculated an “LRC index” for each eye as described in the Materials and Methods. This normalised the results for each experiment and expressed them as a percentage value, such that the sum of the mean LRC indices for each of the four groups compared in each experiment was equal to 100%. The LRC index was significantly higher for Pax6 +/− than WT eyes at 30 weeks when expressed per unit area but not at 15 weeks ( Fig. 7C ). However, when expressed as the LRC index per limbal circumference, the difference at 30 weeks was no longer significant ( Fig. 7D ). This implies that LRCs were significantly more frequent per unit area of limbus in Pax6 +/− eyes than WT eyes at 30 weeks but because Pax6 +/− eyes are smaller the total estimated number of limbal LRCs per eye is not significantly greater. None of the comparisons shown in Fig. 7A–D revealed a significant difference between ages for either genotype.
LRC cells were counted per unit area without counting BrdU-negative cells so we also compared the total cell packing densities in the ocular surface of the four groups. For technical reasons, the packing density was evaluated in the central cornea rather than the limbus. The mean cell numbers (±95% CI) per 150×150 µm sized sampling box in 3 fields from 3 eyes per group were estimated as 279±41 cells for 15-week WT, 207±45 for 15-week Pax6 +/− , 266±117 for 30-week WT and 252±47 for 30-week Pax6 +/− . A 2-way ANOVA revealed no significant differences between genotypes or ages. Thus, there was no evidence that the greater LRC index per unit area seen in Pax6 +/− eyes compared to WT at 30 weeks ( Fig. 7C ) was a consequence of greater cell packing density in the Pax6 +/− ocular surface in the older age group. Therefore, the only difference between Pax6 +/− and WT genotypes observed, suggests that the number of limbal LRCs may be higher in some Pax6 +/− mice, not lower as predicted if LESC numbers were reduced (hypothesis 1 in Fig. 1E–H ).
LRCs also Occur in Pax6 +/− Corneas and are Associated with Blood Vessels
During the analysis of LRCs in the Pax6 +/− limbus it was noted that there were also some LRCs in the Pax6 +/− cornea itself (more central than the sampling box). The images used for counting limbal LRCs were also used to count corneal LRCs so this analysis was restricted to the peripheral areas of the cornea. A few corneal LRCs were identified and counted in WT eyes but, if the limbal sampling box was imprecisely positioned over the limbus, these could have been limbal LRCs. Significantly more corneal LRCs were identified in Pax6 +/− corneas than in WT corneas at both ages ( Fig. 7E ).
The identification of LRCs in the corneal epithelium of Pax6 +/− eyes in conjunction with the observation that blood vessels invade the corneal epithelium in Pax6 +/− mice ( Fig. 5C,D ) prompted a study to localise both LRCs and blood vessels in the same samples. WT and Pax6 +/− mice were prepared for LRC analysis at 15 weeks (1-week BrdU; 10-week chase) and eyes were immunostained for BrdU and CD31 blood vessels. The distribution of LRCs appeared very similar to the blood vessels ( Fig. 8 ). In WT eyes, LRCs were restricted to the limbus ( Fig. 8A ) but in Pax6 +/− eyes they were present in both the limbus and cornea ( Fig. 8B ) and usually associated with blood vessels ( Fig. 8C,D ). The fluorophores that were used did not provide a suitable colour combination to determine whether BrdU staining (yellow) co-localised with CD31-positive blood vessels (red). However, this observation suggests that the BrdU-LRCs in Pax6 +/− eyes are not all LESCs.
10.1371/journal.pone.0071117.g008 Figure 8
Association of BrdU label-retaining cells with blood vessels in Pax6 +/− corneas.
Double immunofluorescent detection of BrdU (yellow) and CD31-positive blood vessels (red) in WT and Pax6 +/− mice. ( A ) Flattened WT cornea showing BrdU-positive, label-retaining cells (LRCs) and CD31-positive blood vessels in the limbus. ( B ) Flattened Pax6 +/− cornea demonstrating CD31-positive blood vessels extending from the limbus into the cornea and BrdU LRCs in the cornea. ( C, D ) Montages of flattened Pax6 +/− corneas showing CD31-positive blood vessels and BrdU LRCs even in the central cornea. Arrows show blood vessels with adjacent BrdU-positive cells (LRCs). For demonstration purposes the counterstain channel was deactivated. Abbreviations: L: Limbus; Co: Cornea; Yellow immunofluorescence: BrdU. Red immunofluorescence: CD31; Blue: TO-PRO3 iodide counterstain. Scale bars are 100 µm.
Discussion
Cell Turnover in the Wild-type Mouse Corneal Epithelium
The decline in BrdU in the WT basal corneal epithelium after 24 hours, as cells moved vertically (apically) to the suprabasal layers, was followed by a decline in labelled suprabasal cells, after 3 days. This showed that most cells labelled in the basal layer were lost from the corneal epithelium by day 14 and implies that the corneal epithelium ‘turnover time’ is ≤14 days. However, most BrdU-positive cells will survive for less than this, once they leave the basal layer, because some will remain in the basal layer for several cell generations. The decline of labelled suprabasal cell numbers between 3 and 7 days suggests that some basal cells are shed from the surface within 7 days of labelling. This is similar to previous turnover time estimates of 6–7 days in mice [49] , 3 1 / 2 –4 days, 12.3 days or 2 weeks in rats [49] , [50] , [51] and 9–10 days or >14 days in rabbits [52] , [53] .
The turnover time is shorter than the time required for LESCs to replace the corneal epithelium. This has been termed the ‘renewal time’ [52] and estimated as 7 weeks for adult mice, [5] and 9–12 months in rabbits [52] .
For WT corneas, there was little evidence of regional differences in cell proliferation or movement to the suprabasal layers. Results of previous studies are inconsistent [50] , [53] , [54] , [55] , [56] , [57] . Some differences among studies may reflect species or technical differences and others may be explained by circadian rhythms that affect corneal epithelial proliferation in mice, rats and rabbits [55] , [56] , [58] and may affect proliferation differently at the corneal periphery and centre [55] .
Increased Cell Turnover in the Pax6 +/− Mouse Corneal Epithelium
A previous BrdU study showed significantly higher labelling in the basal Pax6 +/− corneal epithelium compared to WT [28] . Our results showed a similar trend but it was not significant. The present study is the first to compare the distributions of BrdU-labelled cells separately in the basal and suprabasal layers of Pax6 +/− and WT corneas. This demonstrated that relatively more labelled cells move from the basal to the suprabasal layers within the first 24 hours of labelling in Pax6 +/− than WT corneas and the difference is greatest in the central cornea. As this vertical cell movement is irreversible and is the beginning of cell loss, we conclude that greater corneal epithelial cell loss occurs in Pax6 +/− than WT corneas. Overall, this shows that Pax6 +/− corneal epithelial cell turnover is faster because basal cells both proliferate more frequently [28] and are more readily lost from the cornea than WT basal cells. Furthermore, the greater cell loss from the Pax6 +/− central corneal epithelium may be because the Pax6 +/− cornea is more fragile and the central region is most vulnerable to injury.
Limbal Epithelial Stem Cell Function in Older WT Mice
We used LRCs to identify putative LESCs but this is not specific for stem cells. If both actively proliferating and relatively quiescent stem cell populations exist, as they do in some other tissues, only the quiescent ones will be identified as LRCs [59] , [60] , [61] and only those that divide during the labelling period will be included. Furthermore, any somatic cells that divide during the labelling period but subsequently undergo terminal differentiation before diluting the label significantly will also be identified as LRCs.
The presence of LRCs in the limbus but not the central cornea in WT mice is consistent with the conventional view that the adult corneal epithelium is maintained by slow-cycling LESCs, some of which are identified as label-retaining cells (LRCs). However, the quantitative LRC results failed to support two predictions prompted by analysis of radial mosaic corneal epithelial stripe patterns. There was no evidence that LRC numbers declined between 15 and 30 weeks in WT mice and no evidence that Pax6 +/− mice had fewer LRCs than WT mice at 15 weeks. The mosaic analysis compares corrected stripe numbers, which relates to the numbers of active coherent clones of stem cells (comprising 1 or more stem cells) rather than the numbers of individual stem cells. Thus, stripe numbers depend both on numbers of stem cell clones (≤ number of stem cells) and the ability of the stem cells to function (generate a stripe of TACs in the corneal epithelium). Different results may be obtained for the mosaic and LRC analyses for two reasons. First, LESC function may change in a way that affects stripe numbers and LRC numbers differently. Second, LESC function may not change with age but stripe numbers could decline for an unrelated reason.
In the first case, several age-related changes could affect LESCs and cause stripe numbers to change independently of LRC numbers. (a) LESCs in older mice may continue to divide, and be detectable as LRCs, but more may be defective and unable to establish long-lived clones of corneal epithelial cells, identifiable as stripes in mosaic eyes. (b) Age-related changes in LESC cell cycle kinetics may cause longer or more frequent periods of quiescence, which reduce stripe production, but this might not reduce LRC numbers if, for example, reduced BrdU labelling was balanced by increased BrdU retention. (c) If the limbus harbours separate populations of actively proliferating stem cells (which produce most of the stripes) and more quiescent stem cells (detectable as LRCs) the age-related change could be restricted to stripe numbers if only the more active stem cell population decreased with age. (d) Both stripe numbers and LRCs produced by LESCs might decline with age but if the cell cycle slowed in some other cell types these might also be identified as LRCs and the overall LRC number would not decline in parallel with the stripe numbers.
In the second case, an age-related decline in corrected stripe number might reflect a decline in numbers of active LESC clones without any reduction in the numbers of active stem cells or their function if stochastic neutral drift in LESC populations eliminated some clones and expanded others. This has been suggested as an explanation of the coarsening of mosaic patterns derived from stem cell clones in the mouse testis [62] and mouse intestine [63] , [64] , [65] and could also apply to the corneal stripes as discussed elsewhere [66] .
Limbal Epithelial Stem Cell Function in Pax6 +/− Mice
There was no evidence to support the prediction from the stripe analysis that Pax6 +/− mice had fewer LRCs than WT mice at 15 weeks. Moreover, the limbal LRC index per unit area was significantly higher (not lower) for Pax6 +/− than WT eyes at 30 weeks ( Fig. 7C ). There were also significant numbers of LRCs within the cornea itself ( Fig. 7E ). These results are consistent with two possibilities.
First, Pax6 +/− eyes could have at least as many slow-cycling stem cells in the limbal epithelium as WT mice plus an additional population of slow-cycling stem cells in the corneal epithelium itself. The presence of LRCs in the corneal epithelium has also been reported for Dstn corn1/corn1 mice [67] . In this case there was no evidence of centripetal cell movement and the authors suggested that the mutant corneal epithelium was maintained by stem cells within the cornea itself. Cell movement may be abnormal in Pax6 +/− corneas [68] , [69] but there is no evidence it has ceased. Therefore, if the Pax6 +/− corneal LRCs represent stem cells, they could have arisen within the cornea or moved into the cornea from the limbus or conjunctiva. However, our LRCs study provides no evidence for stem cells within the corneal epithelia of WT mice as proposed to explain the results of limbal transplantation experiments [70] .
If the Pax6 +/− corneal LRCs were stem cells, their apparent association with blood vessels ( Fig. 8 ) is consistent with other reports of close relationships between putative stem cells, identified as LRCs, and blood vessels in different tissues [71] , [72] , [73] , [74] , [75] . The trend for Pax6 +/− eyes to have more LRCs than WT, particularly at 30 weeks ( Fig. 7 ) also raises the possibility that an increasing demand for cells to maintain the corneal epithelium in this model of chronic corneal wound healing produces higher numbers of LESCs in Pax6 +/− heterozygotes (e.g. by symmetrical LESC division) so Pax6 +/− LESC numbers might actually increase with age. However, the increase in LRCs between 15 and 30 weeks was not statistically significant and there is no other evidence for an age-related increase in Pax6 +/− LESCs.
The second possibility is that some of the LRCs in the Pax6 +/− limbus, and perhaps all of those in the cornea, are other types of cells that divided during the BrdU exposure period but then dropped out of the cell cycle. The close association between corneal LRCs and blood vessels in Pax6 +/− corneas suggests that some LRCs may actually be vascular endothelial cells or their associated pericytes, immediately underlying the epithelium, which proliferated during the period of BrdU exposure and then terminally differentiated. Moreover, the LRC index per unit area was only significantly greater in Pax6 +/− than WT eyes in the older age group ( Fig. 7C ), which is when more Pax6 +/− corneas are likely to show neovascularisation. In addition, some Pax6 +/− corneal LRCs could be inflammatory cells that phagocytosed labelled cells, as reported for activated macrophages [76] , [77] . We, therefore, conclude that our limbal LRC analysis is unlikely to only identify stem cells in the Pax6 +/− ocular surface so does not accurately represent the relationship between Pax6 +/− and WT LESC numbers. The LRC approach has not provided an adequate means of testing whether Pax6 +/− mice have a LESC deficiency, so alternative approaches are required.
Abnormal Maintenance of the Corneal Epithelium in Pax6 +/− Mice
Although our results do not demonstrate whether LESCs are depleted or defective in Pax6 +/− mice, they have provided further insights about the abnormal maintenance of the Pax6 +/− corneal epithelium. We confirmed that the Pax6 +/− corneo-limbal boundary is indistinct, which is consistent with the widely-accepted notion that conjunctiva intrudes into the mouse Pax6 +/− and human PAX6 +/− cornea because there is a LESC deficiency. However, this is not proof of LESC deficiency and other explanations remain possible [28] . For example, a recent study shows that when goblet cells are present in the corneal epithelium they need not originate from the conjunctiva [78] .
Our evidence that cell loss is greater than normal in the Pax6 +/− corneal epithelium is consistent with previous evidence for greater corneal epithelial fragility, attributed to abnormal expression of molecules important for cell adhesion [27] , [28] , [29] . In the WT corneal epithelium, cell loss may be driven by a combination of factors, possibly including basal cell overcrowding, as reported for the Drosophila notum [79] . In the Pax6 +/− corneal epithelium, it seems likely that the reverse applies, such that greater epithelial fragility causes increased cell loss, which drives increased production of TACs.
To ensure the corneal epithelium is maintained at a uniform thickness, cell production, movement and loss must be co-ordinated and balanced. For WT corneas this balance was originally described in terms of the X, Y, Z hypothesis [80] . Updated as the limbal stem cell hypothesis, this can be restated as Y SC +X TAC = Z L [66] where Y SC denotes production of basal corneal epithelial cells by LESCs, X TAC denotes the proliferation of corneal epithelial TACs, and Z L denotes epithelial cell loss from the corneal surface. In the Pax6 +/− corneal epithelium, cell loss is high and may exceed cell production. TAC proliferation is also high [28] but we still do not know whether TAC production by LESCs is low, so maintenance of the Pax6 +/− corneal epithelium may be summarised provisionally as Y SC (?) +X TAC high = Z L high if corneal homeostasis is in equilibrium or Y SC (?) +X TAC high <Z L high if corneal epithelial cell numbers decline (excluding other encroaching cell types). Corneal epithelial TACs might be stimulated to be more proliferative in response to greater cell loss from a fragile corneal epithelium and/or to compensate for deficient or defective LESCs. This could involve decreasing cell cycle times, increasing the number of TAC divisions beyond the normal maximum or changing the balance of cell divisions so more TACs fulfil their maximum proliferative potential, as may occur in response to wounding [6] .
The corneal epithelium of Pax6 +/− mice is already thin by E18.5, just before birth [21] , implying that this is a developmental abnormality. If homeostasis is initially quantitatively normal, the corneal epithelium might be maintained adequately for some time, although it will be thinner than in WT mice. However, in adult Pax6 +/− mice, it seems likely that increased corneal epithelial cell loss exceeds the cell production capacity and causes corneal homeostasis to become unstable, resulting in progressive corneal deterioration. This is supported by morphological evidence of cells sloughing off an irregular, vesiculated superficial Pax6 +/− epithelial layer [21] . In this study the number of epithelial cell layers was reduced from 4–5 cell layers in WT to 2–4 in Pax6 +/Sey-Neu heterozygotes and a similar difference was seen in the present investigation (compare WT in Fig. 2A,C with Pax6 +/− Fig. 2B,D ). However, an even greater reduction (from 8–10 cell layers in WT to 1–7 layers) was reported for both Pax6 +/Sey-Neu and Pax6 +/Sey-Dey heterozygotes on a different genetic background [27] . The extreme cases where the corneal epithelium was reduced to a single layer strongly suggest that corneal homeostasis had become quantitatively destabilised.
Qualitative corneal epithelial deterioration is likely, even if additional encroaching cells maintain total cell numbers adequately, because these will not be phenotypically identical to corneal epithelial stem cells and so corneal function will be compromised. A shortfall in corneal epithelial cell production may arise simply because the increased cell loss from the Pax6 +/− corneal epithelium exceeds the renewal capacity of even a normal complement of LESCs. We suggest that hallmarks of corneal deterioration, including ‘conjunctivilisation’ and the appearance of goblet cells within the corneal epithelium, that are often taken as evidence of LESC-deficiency, might also occur in the absence of LESC deficiency if corneal homeostasis is destabilised by excessive cell loss.
There is also evidence that reduced Pax6 levels in the human PAX6 +/− corneal epithelium is accompanied by a down-regulation of K12 [81] . This parallels the situation described in the Introduction for Pax6 +/− mice and indicates that corneal epithelial differentiation is abnormal. Reduced K12 in PAX6 +/− aniridia also probably increases epithelial fragility, so contributing to the recurrent corneal erosions that are characteristic of human aniridia-related keratopathy [82] . This increased epithelial fragility suggests that cell loss is likely to be abnormally high in people with PAX6 +/− aniridia as well as heterozygous Pax6 +/− mice. These observations indicate that the effect of Pax6 depletion on the human corneal epithelium is not restricted to the LESCs so the abnormal phenotype associated with ARK is probably not mediated entirely via LESC-deficiency.
Conclusion
Quantitative LRC experiments did not support the prediction that LESCs decline with age in WT mice and failed to test adequately whether Pax6 +/− mice have a LESC-deficiency. Several possible explanations are discussed that require further investigation. Although it remains unclear whether Pax6 +/− mice have LESC-deficiency, our BrdU-labelling analysis implied that epithelial cell turnover is faster than normal and this is probably driven by increased cell loss. Excessive cell loss might destabilise corneal homeostasis and cause Pax6 +/− corneal epithelial deterioration, even in the absence of LESC-deficiency. This also raises the possibility that excessive cell loss could also play a causal role in human aniridia-related keratopathy.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Microbial surveys based on 16S rRNA gene amplicon sequencing are an important tool in environmental and biomedical research [ 1 – 2 ]. Microbial community structure can provide valuable insights not only into the workings of natural ecosystems, but increasingly into the relationship between the human host and its bacterial colonizers. Rapid progress in DNA sequencing technology has provided ever-increasing outputs coupled with lowered costs, facilitating an explosion in amplicon sequencing studies [ 3 ]. Unfortunately, these studies are vulnerable to potential biases introduced along the workflow and there is a lack of consensus regarding best practices [ 4 – 5 ]. This paper aims to provide researchers (e.g. ecologists, microbiologists, biomedical researchers) with a overview of the strengths and weaknesses of six of the most popular current bioinformatic pipelines for 16S rRNA gene amplicon sequencing. While this selection is not a comprehensive set, it includes some of the most used (QIIME [ 6 ], MOTHUR [ 4 ], and USEARCH [ 5 ]) as well as more recent options (DADA2 [ 6 ] and Qiime2-Deblur [ 7 – 8 ]). Three of these pipelines cluster sequences at (typically) 97% identity into Operational Taxonomical Units (OTUs): QIIME-uclust, MOTHUR and USEARCH-UPARSE. The other three (Qiime2-Deblur, DADA2, and USEARCH-UNOISE3) attempt to reconstruct the exact biological sequences present in the sample, so-called Amplicon Sequence Variants (ASVs) [ 9 ]. ASVs are referred to by other authors as “zero noise OTUs” [ 10 ] or “sub-OTUs” [ 7 ].
The pipelines benchmarked here may perform better than reported in this paper if their parameters are customly tuned for an individual dataset. However, we believe that the vast majority of users employ either default or author-recommended settings. We therefore aimed to compare pipelines under these typical conditions in order to match the most plausible use scenarios. We examined the effect of different quality filtering steps (for QIIME-uclust, Qiime2-Deblur, and DADA2) and of different clustering algorithms and cutoffs (for MOTHUR), as we deemed these to be the most likely pipeline variations users might attempt. While other benchmarking studies have been published, they relied only on simulated (synthetic) reads [ 11 ] or on very small data sets [ 12 ].
In this paper, pipelines were compared using a mock sample sequenced repeatedly over multiple sequencing runs as well as a large (N = 2170 individuals) fecal sample dataset from the “Healthy Life in an Urban Setting” (HELIUS) multi-ethnic study [ 13 – 14 ]. We examined the specificity and sensitivity of each workflow (e.g. number of spurious OTUs/ASVs produced), the quantitative agreement between the inferred relative abundances, as well as any pipeline-specific effects on downstream alpha-diversity measures.
Material and methods
Datasets
Mock community
Genomic DNA from the Microbial Mock Community B (Even, Low concentration), v5.1L (Catalog no. HM-782D, obtained through BEI Resources, NIAID, NIH as part of the Human Microbiome Project) was sequenced in three separate runs. Details of mock composition are included in S1 Table . The mock contains DNA from 20 bacterial strains in equimolar (Even) ribosomal RNA operon counts (100000 copies per organism per μL). Two of the strains ( Bacteriodes vulgatus and Clostridium beijerinckii ) have multiple sequence variants in the V4 region of the 16S rRNA gene. B . vulgatus has three variants (in a 5:1:1 ratio), whereas C . beijerinckii has two variants (in a 13:1 ratio). The 16S rRNA sequences of Staphylococcus aureus and Staphylococcus epidermidis are identical in the V4 region. Therefore, the mock contains a total of 22 variants (ASVs) of the 16S gene in the V4 region. These sequences correspond to 19 OTUs when clustered at 97% identity. The mock community was sequenced three times in different sequencing runs. The mock raw sequence data is publicly available ( https://github.com/andreiprodan/mock-sequences ).
HELIUS fecal samples dataset
A total of 2170 fecal samples obtained from adult individuals from six ethnic groups in Amsterdam, the Netherlands (the HELIUS study) were sequenced. Cohort information and detailed sample collection and processing protocols have been previously described [ 13 – 14 ]. The HELIUS fecal sample dataset contained 177.08 million paired-end reads obtained from 17 individual sequencing runs. All raw sequencing data from this dataset is available on the European Genome-phenome Archive repository (accession no. EGAD00001004106).
Library preparation and sequencing
Library preparation and sequencing was performed at the Wallenberg Laboratory (Sahlgrenska University of Gothenburg, Sweden). Total genomic DNA was extracted from a 150 mg fecal sample aliquot using a repeated bead beating method as previously described [ 15 ]. Fecal microbiome composition was profiled by sequencing the V4 region of the 16S rRNA gene on an Illumina MiSeq instrument (Illumina RTA v1.17.28; MCS v2.5) with 515F and 806R primers designed for dual indexing [ 16 ] and the V2 Illumina kit (2x250 bp paired-end reads). 16S rRNA genes from each sample were amplified in duplicate reactions in volumes of 25 μL containing 1x Five Prime Hot Master Mix (5 PRIME GmbH), 200 nM of each primer, 0.4 mg/ml BSA, 5% DMSO and 20 ng of genomic DNA. PCR was carried out under the following conditions: initial denaturation for 47 min at 94°C, followed by 25 cycles of denaturation for 45 sec at 94°C, annealing for 60 sec at 52°C and elongation for 90 sec at 72°C, and a final elongation step for 10 min at 72°C. Duplicates were combined, purified with the NucleoSpin Gel and PCR Clean-up kit (Macherey-Nagel) and quantified using the Quant-iT PicoGreen dsDNA kit (Invitrogen). Purified PCR products were diluted to 10 ng/μL and pooled in equal amounts. The pooled amplicons were purified again using Ampure magnetic purification beads (Agencourt) to remove short amplification products. Libraries for sequencing were prepared by mixing the pooled amplicons with PhiX control DNA purchased from Illumina. The input DNA had a concentration of 3 pM and contained 15% PhiX and resulted in the generation of about 700K clusters/mm 2 and an overall percentage of bases with quality score higher than 30 (Q30) higher than 70%.
Pipelines and parameters
Six different pipelines were included in this comparison: QIIME (v.1.9.1) [ 17 ], MOTHUR (v.1.39.5) [ 4 ], DADA2 (1.7.0) [ 6 ], Qiime2 (v.2017.6.0)-Deblur [ 7 , 8 ], USEARCH (v.10.0.240)-UPARSE [ 18 ], and USEARCH (v.10.0.240)-UNOISE3 [ 10 ].
Paired-end reads merging and quality filtering
In MOTHUR, the merging and quality filtering of reads (“screening”) is an integral part of the pipeline, not easily performed outside MOTHUR. We therefore used MOTHUR only with its internal read merging and filtering. In DADA2—in contrast to all other pipelines—denoising is always implemented separately on the forward and the reverse reads, with resulting ASVs merged at the end of the workflow. External merging / filtering is therefore not applicable to DADA2, while it can be integrated with relative ease into QIIME and Qiime2 workflows. We therefore implemented one workflow in QIIME-uclust and one workflow in Qiime2-Deblur where reads were merged and filtered externally using USEARCH. The author of USEARCH explicitly advises against using the default USEARCH read merging parameters for reads with a long overlap (e.g. MiSeq 2 x 250bp V4) and argues in favor of maximizing the proportion of reads that survive the merging step. We therefore benchmarked several settings for the merging step. Based on the outcomes ( Fig 1 ), we chose 30 max. allowed differences in the overlapping region (“maxdiffs”) for the merging step (using the “fastq_mergepairs” command) and max. 1 expected errors (“fastq_maxee”) as a quality filter threshold (using the “fastq_filter” command). The rationale was to use permissive parameters in the merging step in order to fully exploit the error correction made possible by the overlapping of the reads (e.g. a lower Q-score sequencing error in the reverse read can be rectified using the higher Q-score correct base call from the forward read). Strict thresholds (i.e. low “maxdiffs”) discard read pairs where mismatches due to an error on one read might have been easily corrected using the complementary read. As Fig 1 shows, more relaxed merging parameters resulted in around 10% more of total raw reads (82.9% compared to 73.5%) passing the quality filter. These merging / filtering parameters were used in the USEARCH-UPARSE, USEARCH-UNOISE3, QIIME-uclust (e30.ee1), and Qiime2-Deblur (e30.ee1) flows. Expected error-based read quality filtering is described in detail in Edgar et al. 2015 [ 19 ].
10.1371/journal.pone.0227434.g001
Fig 1
Effect of different USEARCH paired-end read merging parameters (“maxdiffs”).
QIIME-uclust
In the typical QIIME-uclust workflow, forward and reverse reads are merged using the “multiple_join_paired_ends.py” script. Subsequently, quality control and demultiplexing are performed simultaneously using the “multiple_split_libraries_fastq.py” script, which truncates the reads if more than r (default 3) consecutive bases do not have a Q-score higher than q (default 3). Reads are discarded if, after trimming, the read length drops to less than p (default 0.75) of initial length. By default, no ambiguous bases (“N”) are allowed (default is 0). OTU clustering was performed using the “pick_open_reference_otus.py” script, with all default parameters. This script implements the latest QIIME open reference OTU clustering [ 20 ]. In brief, it performs closed reference clustering against the Greengenes (v.13.8) 97% OTU database, using UCLUST v.1.2.22q [ 5 ]; reads that do not map in this first step are subsampled (default proportion of subsampling = 0.001) and used as new centroids for a de novo OTU clustering step. Remaining unmapped reads are subsequently closed-reference clustered against these de novo OTUs. Finally, another step of de novo clustering is performed on the remaining unmapped reads.
Three different QIIME-uclust workflows were run using different merging and quality control parameters. One QIIME-uclust flow used all default parameters (“QIIME-uclust (default)”), while another was run with the q parameter set to 19 instead of default 3 (“QIIME-default (Q20)”). For the 3rd flow the read merging and quality control steps were performed outside QIIME, using USEARCH with the same parameters used in the USEARCH flows: max. 30 allowed mismatches in the overlapping region for merging, max. 1 allowed expected error per merged read for filtering (“QIIME-uclust (e30.ee1)”).
MOTHUR
In MOTHUR, paired-end reads are merged using the “make.contigs” command. This aligns the forward and reverse reads and, if a position in the overlapping region has different base calls in the forward read versus the reverse read, compares the forward read Q-score and the reverse read Q-score for that position. If one of the two Q-scores is at least “deltaq” points higher than the other (default deltaq is 6), then the merged read will use the respective base call. Otherwise, if the difference between the forward and the reverse base call Q-scores is less than deltaq, the base at that position is re-labelled as ambiguous (“N”}). Quality filtering is implemented with the”screen.seqs” command, which (by default) removes all merged reads containing ambiguous bases. Sequences are then deduplicated, aligned to a database (SILVA v.128), and pre-clustered (i.e. sequences with less than 2 base differences (“diffs”) from a more abundant sequence are merged with the more abundant sequence). Chimeric reads are removed (“chimera.vsearch”) and remaining uniques are clustered into OTUs.
Two different clustering algorithms were used in MOTHUR: Opticlust (the current default) and DGC (Distance-based Greedy Clustering).
Opticlust [ 21 ] is a distance-based algorithm and therefore requires a distance matrix to be constructed between all unique sequences. The size of this matrix scales with the square of the number of sequences and can thus become problematic with large datasets (since the matrix must fit into available RAM memory). This issue can be side-stepped with the “split.abund” command, removing sequences with extremely low abundance after the pre-clustering step. In this study, a cutoff of 3 was used for the MOTHUR-Opticlust flow, keeping only sequences with more than 3 counts in the entire dataset (“MOTHUR (Opticlust.3)”).
DGC uses Vsearch [ 22 ] (an open source alternative to USEARCH) in order to perform greedy clustering which does not require a distance matrix. Three different MOTHUR pipelines were run with DGC clustering, applying different cutoffs to the “split.abund” command after the preclustering step: “MOTHUR (DGC.0)” (cutoff 0, i.e. not removing any sequences), “MOTHUR (DGC.1)” (cutoff 1, i.e. only removing singletons, sequences that appear only once in the entire dataset), and “MOTHUR (DGC.3)” (cutoff 3, i.e. removing unique sequences with at most 3 counts in the entire dataset).
DADA2
DADA2 (Divisive Amplicon Denoising Algorithm 2) uses a parametric model to infer true biological sequences from reads. The model relies on input read abundances (true reads are likely to be more abundant) and distances (less abundant reads only a few base-differences away from a more abundant sequence are likely error-derived). Base Q-scores are used to calculate a substitution model, estimating a probability for each possible base substitution (e.g. A replacing G, G replacing T, etc). Based on this substitution model and on the input reads abundances and reciprocal distances, DADA2 uses a probability threshold to decide whether to assign counts from a less abundant, “error-derived” read to a more abundant, “true” sequence. DADA2 was run as an R script (in R v.3.4) using its R package (dada2 v.1.7).
In DADA2, reads are quality-filtered using the “filterAndTrim” function. Error rates are subsequently learned from a set of subsampled reads (i.e. 1 million random reads). Error rates are estimated separately for each sequencing run, since different runs may have different error profiles. Reads are then deduplicated and ASVs are inferred. Uniquely, DADA2 retains a summary of the quality scores associated with each unique sequence (the average of the positional qualities from deduplicated reads). These quality scores are subsequently used to perform ASV inference. Also, unlike all the other pipelines, DADA2 denoises the forward and the reverse reads independently. ASVs from the forward and reverse flows are only merged at the end of the workflow prior to the removal of chimeric ASVs, using “removeChimeraDenovo”.
Two DADA2 pipeline flows were run using different quality score filters. Both flows truncated reads by removing the last 10 bases from the forward reads and the last 40 bases from the reverse reads, as well as truncating / removing reads with ambiguous bases. The 1st DADA2 flow, “DADA2 (ee2)”, filtered (trimmed) reads to max. 2 allowed expected errors per read, while the 2nd, “DADA2 (no filter)”, did not use an expected error-based filter. Removing or truncating reads with ambiguous bases is mandatory in DADA2.
Qiime2-Deblur
Qiime2 [ 8 ] is the successor platform to QIIME. It incorporates several plugins, including DADA2 and Deblur. We used Qiime2 in combination with its Deblur plugin [ 7 ], following the flow for paired-end reads from the Qiime2 website ( https://qiime2.org/ ). Raw reads were imported into a Qiime2 artifact before merging paired-end reads and quality filtering. An artifact is a Qiime2-specific file format which holds data as well as metadata, provenance, and version information. Reads were then denoised using the “deblur denoise-16S” command, trimming reads at a length of 250 bases.
Deblur compares sequence-to-sequence Hamming distances to an upper-bound error profile combined with a greedy algorithm [ 7 ]. Sequences are sorted by abundance, then the number of predicted error-derived reads is subtracted from the counts of neighboring reads based on their read-to-read Hamming distance. Any sequence whose abundance drops to 0 during this process is removed. The Deblur algorithm is applied to each sample independently.
Three different Qiime2-Deblur flows were run. In the first, “Qiime2-Deblur (default), merging and trimming were performed inside Qiime2 using the “vsearch” plugin [ 22 ] (for merging) and the “quality-filter q-score” plugin (for filtering) with default options”. Similar to the quality filtering in QIIME, the Qiime2 quality-filter truncates reads if more than 3 consecutive bases do not have a Q-score higher than 3 and discards them if post-trimming read length is less than 0.75 of initial length. No ambiguous bases are allowed. The second Qiime2 flow, “Qiime2-Deblur (Q20)”, used the same parameters as the first, but with the Q-score threshold (“p-min-quality”) set to 20 In the third Qiime2 flow, “Qiime2-Deblur (e30.ee1)”, read merging and quality control were performed outside Qiime2, using USEARCH with the same parameters used in the USEARCH flows.
USEARCH-UPARSE and USEARCH-UNOISE3
Both the UPARSE [ 18 ] and the UNOISE3 [ 10 ] pipelines are implemented in USEARCH [ 5 ]. The merging and filtering (covered in the “Paired-end reads merging and quality filtering” subsection) as well as deduplicating (“fastx_uniques” command) are therefore identical for UPARSE and UNOISE3 and only need to be performed once, before the pipelines branch off into OTU-level clustering (with UPARSE) or ASV-level denoising (with UNOISE3). Indeed, the author of USEARCH advises that UPARSE and UNOISE3 should be performed together ( https://drive5.com/usearch/manual/faq_uparse_or_unoise.html ).
UPARSE uses the UPARSE-REF greedy algorithm to infer errors using the concept of parsimony. In brief, UPARSE-REF aims to explain a given input sequence starting from sequences in a database, using the fewest possible number of events (i.e. PCR or sequencing errors). It constructs a model sequence using one or more sequences from the database (i.g. a single sequence representing a non-chimeric amplicon, or multiple concatenated segments representing a chimeric amplicon). Different penalty scores are given for chimeric crossover and for mismatches and the model sequence with the lowest total score is chosen as the true OTU sequence. Sequences are ranked in decreasing order of abundance, discarding singletons. Each input sequence is then compared to the current OTU set and to the maximum parsimony model sequence constructed using UPARSE-REF. If the model sequence is more than 97% identical to an existing OTU, the sequence is assigned to the respective OTU. It the model sequence is chimeric, the sequence is discarded. Finally, if the model is less than 97% identical to any existing OTU, the sequence is added to the existing OTU set. After all OTUs are found, all merged reads (including those dropped during quality filtering) are mapped against the OTUs to construct the OTU table.
The UNOISE3 flow ranks sequences in decreasing order of abundance, discarding sequences with less than 8 counts (the default min. abundance default threshold). The ASV set is initially empty. A model is then applied ( Eq 1 ) to each input sequence in order to test whether its abundance (a M ) is sufficiently large compared to the abundance of its closest sequence (a C ) at Levenshtein distance d . The default value of the α parameter is 2. If Eq ( 1 ) holds, the input sequence becomes a new ASV; else, the input sequence it assigned to the nearest existing ASV. The final set of ASVs undergoes chimera filtering using the UCHIME2 [ 23 ] algorithm in de novo mode. Similar to UPARSE, the final step in the UNOISE3 flow is to map all merged reads to non-chimeric ASVs in order to construct the ASV table.
a M / a C ≤ 1 / 2 α d + 1
(1)
Both the UPARSE clustering (the “cluster_otus” command) and the UNOISE3 denoising (the “unoise3” command) steps were executed with all default settings. In both flows, all samples in a data set were processed together, rather than individually, in order to achieve optimal sensitivity.
Data analysis
All OTU/ASV tables produced by the pipelines were converted into phyloseq objects using the “phyloseq” package [ 24 ](v.1.24.2). Alpha diversity measures (richness, Chao1, Shannon index, inverse Simpson index) were calculated using the “estimate_richness” function from “phyloseq”. OTUs/ASVs were classified as “Exact” (perfect match to a true sequence in the mock community), “One-off” (at 1 Hamming distance away from a true sequence), or “Other” (at more than 1 Hamming distance from a true sequence). Sequence-to-sequence Hamming distances were calculated using the “stringdist” R package (v.0.9.5.1). “One-off” and “Other” were together labeled as “Spurious”.
Plots were constructed in R [ 25 ] (v.3.4) using the “ggplot2”[ 26 ] (v.3.1.0), “corrplot”[ 27 ] (v.0.84), and “VennDiagram”[ 28 ] (v.1.6.20) packages.
Results and discussion
Mock community
Sensitivity and specificity
An overview of Exact, One-Off, and Spurious OTUs/ASVs produced by the different pipelines using reads from the three mock sequencing runs is shown in Table 1 . The three mock sample runs had 36464, 84054, and 146653 paired-end reads, respectively. For each OTU/ASV sequence produced by the different flows, Hamming distances to the closest true sequence ( Fig 2 ) and to the closest other OTU/ASV in the respective flow ( Fig 3 ) were plotted as function of ASV/OTU abundance.
10.1371/journal.pone.0227434.g002
Fig 2
Hamming distance (no. of base differences) from each ASV/OTU sequence to the closest true sequence present in the mock community.
10.1371/journal.pone.0227434.g003
Fig 3
Hamming distance from each ASV/OTU sequence to the closest other ASV/OTU sequence.
Dashed line marks the Hamming distance = 7 threshold, corresponding to the 97% identity threshold for OTUs in V4 16S rRNA gene amplicons. Blue ellipses highlight ASVs that are only 1 Hamming distance away from each other.
10.1371/journal.pone.0227434.t001
Table 1 Sensitivity and specificity over three mock sequencing runs. Values are reported as mean (standard deviation).
Pipeline workflow
Exact
One-Off
Spurious
OTU-level
QIIME-uclust
QIIME-uclust (default)
19 (0) a
134 (27)
412 (236)
QIIME-uclust (e30.ee1)
19 (0) a
133 (31)
341 (198)
QIIME-uclust (Q20)
19 (0) a
132 (26)
400 (232)
MOTHUR
MOTHUR (DGC.0)
19 (0)
none
48 (14)
MOTHUR (DGC.1)
19 (0)
none
24 (8)
MOTHUR (DGC.3)
19 (0)
none
5 (1)
MOTHUR (Opticlust.3)
19 (0)
none
9 (4)
UPARSE
USEARCH-UPARSE
19 (0)
none
13 (7)
ASV-level
DADA2
DADA2 (ee2)
21.7 (0.6) b
none
6 (4)
DADA2 (no filter)
21.7 (0.6) b
none
5 (4)
Qiime2-Deblur
Qiime2-Deblur (default)
19 (0)
none
none
Qiime2-Deblur (e30.ee1)
19 (0)
none
none
Qiime2-Deblur (Q20)
19 (0)
none
none
UNOISE3
USEARCH-UNOISE3
21 (0) c
none
none
a QIIME-uclust erroneously produced separate OTUs for the two C . beijerinckii sequence variants, even though they have only 1 bp difference. It did not detect P . acnes in one of the three mock runs.
b DADA2 did not find the lower copy number C . beijerinckii variant in one of the three mock runs.
c USEARCH-UNOISE3 could not differentiate the two C . beijerinckii variants (13:1 copy number ratio).
DADA2 showed the best sensitivity, detecting all 22 true ASVs present in the mock and was was the only pipeline able to differentiate sequences at single-base resolution even at high abundance ratios (e.g. the 13:1 ratio between the two C . beijenrickii variants). It only missed the low-abundance C . beijenrickii variant ASV in the sequencing run with the lowest number of raw reads. USEARCH-UNOISE3 was also capable of single-base resolution, but was limited (due to the default setting of α parameter, see Eq 1 , Methods) to single-base difference variants present at a no more than 8:1 abundance ratio. Thus, it was able to differentiate between the three B . vulgatus ASVs (ratio 5:1:1), but not between the two C . beijenrickii variants (ratio 13:1) ( Fig 3 ). In effect, USEARCH-UNOISE3 (with default parameters) does not detect ASVs at 1 Hamming distance from another sequence that is >8 times more abundant.
USEARCH-UNOISE3 and Qiime2-Deblur were the only two pipelines to show perfect specificity on the mock sample sequencing data, producing no spurious OTUs/ASVs. Although showing the best sensitivity, DADA2 flows did produce some spurious ASVs. USEARCH-UPARSE and MOTHUR also produces some spurious OTUs. In MOTHUR, there was a visible effect of the cutoff applied to very low abundance sequences prior to clustering. While MOTHUR (DGC.0) and MOTHUR (DGC.1) both produced more spurious OTUs compared to USEARCH-UPARSE, the higher cutoff MOTHUR workflows (DGC.3 and Opticlust.3) produced fewer spurious features and were the OTU-level flows with the best specificity. In our analysis, all MOTHUR flows generated far fewer OTUs compared to QIIME-uclust, in contrast to a similar benchmark performed by the DADA2 authors [ 6 ]. It should be noted that USEARCH-UPARSE (by default) automatically removes singletons in the OTU clustering step.
Both DADA2 and USEARCH-UNOISE3 showed high accuracy in the quantification of abundance ratios in the case of 1-base-pair-difference ASVs ( Table 2 ), yielding values close to the true ratios. Qiime2-Deblur did not differentiate any of the mock ASVs that were only 1 Hamming distance apart and thus did not demonstrate single-base resolution in this analysis.
10.1371/journal.pone.0227434.t002
Table 2 Inferred ratios of 16S rRNA gene variants. Expected ratios (based on known copy numbers of the respective 16S rRNA gene variants) are shown in bold. USEARCH-UNOISE3 could not differentiate the two C . beijerinckii variants. Qiime2-Deblur could not differentiate any of the variants.
B . vulgatus variants
C . beijenrijkii variants
Expected ratio:
V1:V2 (5:1)
V1:V3 (5:1)
V2:V3 (1:1)
V1:V2 (13:1)
DADA2
5.60
5.23
0.94
14.28
USEARCH-UNOISE3
5.31
4.90
0.93
NA
Qiime2-Deblur
NA
NA
NA
NA
Extremely low abundance spurious OTUs/AVSs can be filtered out with relative ease in downstream analysis steps and may therefore have marginal impact on end results. However, QIIME-uclust flows assigned more than 12% of total counts to spurious OTUs, compared to at most 0.17% in other pipelines ( Table 3 ). All three QIIME-uclust flows produced hundreds of spurious OTUs (around 25 times more than the number of true sequences in the mock sample), orders of magnitude more compared to other pipelines. While the flow using the external (and more stringent) quality control produced fewer spurious OTUs compared to the other two flows, the improvement was relatively small (around 10%).
10.1371/journal.pone.0227434.t003
Table 3 Proportion of counts assigned to either true or spurious OTUs/ASVs.
Pipeline
Counts in Exact ASVs/OTUs [%]
Counts in Spurious ASVs/OTUs [%]
QIIME-uclust
87.77
12.21
MOTHUR
99.83
0.17
USEARCH-UPARSE
99.84
0.16
DADA2
99.88
0.12
Qiime2-Deblur
100
none
USEARCH-UNOISE3
100
none
Pipeline- and parameter-dependent biases
Biases affecting the inferred sample composition (systematic under- or over-estimation of certain taxa) pose a problem for amplicon sequencing bioinformatic pipelines, particularly if influenced by factors that can vary between samples or sequencing runs (e.g. read sequencing quality). We observed one such bias in the QIIME-uclust output ( Fig 4A ). While most workflows yielded very similar relative abundance values, all QIIME-uclust flows severely under-estimated the abundance of three OTUs (corresponding to Neisseria meningitis , Pseudomonas aeruginosa , and Rhodobacter sphaeroides ). The bias was caused by QIIME-uclust assigning a large proportion of the counts of these true OTUs to other, spurious OTUs. This effect was independent of quality filtering parameters (i.e. it was observed in all three QIIME-uclust flows) and is likely intrinsic to the closed-reference OTU clustering specific to QIIME-uclust.
10.1371/journal.pone.0227434.g004
Fig 4
Inferred mock community composition.
A) Comparison of QIIME-uclust vs. other pipelines. B) Comparison of DADA (no filter) vs. DADA2 (ee2). OTUs/ASVs whose abundance was under-estimated are indicated with arrows.
Another bias was induced in DADA2 ( Fig 4B ) by quality filtering. While the DADA2 (no filter) flow gave results in line with that of other pipelines ( Fig 4A ), the DADA2 (ee2) flow under-estimated the relative abundance of three ASVs ( Lactobacillus gasseri , Streptococcus agalactiae , and Streptococcus pneumoniae ). This bias was caused by preferential filtering (exclusion) of reads from these ASVs in the quality filtering step. While it is widely known that Illumina sequencing error rates are position-dependent (i.e. error rates tend to increase towards the end of the read), it is often neglected that they may also be affected by underlying sequence patterns [ 29 ]. Particular patterns of bases may result in much higher base call error rates than would be expected. Examples of such patterns are “GGC” triplets or inverted repeats (more than 8 bases long) located upstream of the respective position [ 29 ]. Thus, if a particular ASV sequence happens to contains such a pattern, application of a quality filter will exclude its reads preferentially before the denoising step. The V4 region of the 16S rRNA gene contains 8 instances of the “GGC” pattern for L . gasseri , 7 for S . agalactiae , and 9 for S . pneumoniae , though other patterns likely contribute to the effect. In practice, this presents an issue only for DADA2 (in the case of paired-end sequencing) since all other pipelines merge paired-end reads before clustering/denoising. In these other flows, the errors at the position where the pattern is present are corrected using information from the complementary read in the pair. Considering that the additional quality filter did not improve the specificity of the DADA2 pipeline ( Table 2 , Fig 2 ) while introducing a significant bias in the output ( Fig 4B ), we advise against it.
HELIUS fecal sample dataset
Conversion of reads to counts
Large throughput is desirable to improve detection of low abundance taxa and to maximize the chance than samples with a lower number of sequencing reads will yield sufficient counts to be included in downstream analyses. In this study, we observed a tendency of Qiime2-Deblur to output far fewer counts than other pipelines ( Fig 5 ). While other workflows converted more than 70% of reads form the mock community into counts (with highest conversion rate for USEARCH-UPARSE and USEARCH-UNOISE), Qiime2-Deblur flows converted less than 50%.
10.1371/journal.pone.0227434.g005
Fig 5
Raw reads conversion to final counts.
Quantitative comparison of pipeline outputs
Agreement between the sample composition profiles produced by different pipeline flows was generally high (as measured by the median Spearman's ρ correlation across all OTUs) ( Fig 6 ). For this comparison, DADA2, USEARCH-UNOISE3, and Qiime2-Deblur ASVs were clustered into 97% OTUs in order to be comparable to output from OTU-level pipelines. Different quality filtering parameters (tested in QIIME-uclust and Qiime2-Deblur) or clustering algorithm and cutoffs (tested in MOTHUR) had negligible effect on the inferred composition. The exception was DADA2, for which additional quality filtering shifted the composition profile. While different flows of the same pipeline were clearly grouped together when using hierarchical clustering ( Fig 6 ), DADA2 (no filter) clustered next to USEARCH-UPARSE and USEARCH-UNOISE, while DADA2 (ee2) clustered together with the MOTHUR flows.
10.1371/journal.pone.0227434.g006
Fig 6
Spearman's rho correlation averaged across all samples of the HELIUS fecal sample dataset (N = 2170).
A) Actual values. B) Values scaled to range between 0 and 1. Hierarchical clustering was applied to both rows and columns in order to group pipelines based on the degree of correlation of their outputs.
Table 4 shows read tracking for the different workflows as well as the total numbers of OTUs/ASVs produced from the HELIUS fecal sample dataset. Consistent with results from the mock community analysis, QIIME-uclust flows produced very large numbers of OTUs (around 200000). More stringent read quality filtering only reducing this number by approx. 25%. All QIIME-uclust flows produced an order of magnitude more OTUs compared to any other OTU-level workflow. Based on mock community results, the vast majority of these OTUs are expected to be spurious. USEARCH-UPARSE and both MOTHUR flows using cutoff 3 (Opticlust and DGC) produced a similar number of OTUs (ranging from around 4000 to 5500 OTUs) suggesting that this is the probable range for the number of true OTUs in this dataset. In contrast, QIIME-uclust produced between 150000 and 200000 OTUs. While the cutoff parameter had a large effect on the number or OTUs produced by MOTHUR, there was little to no effect of different quality filtering parameters on the number of ASVs produced by DADA2 and Qiime2-Deblur.
10.1371/journal.pone.0227434.t004
Table 4 Read tracking information and OTU/ASV outputs for the pipeline flows applied to the HELIUS data.
Pipeline Flow
Merged [% of raw]
Filtered [% of raw]
Clustered / Denoised [% of raw]
Conversion to OTUs/ASVs [% of filtered]
Conversion to OTUs/ASVs [% of raw]
Total no. of OTUs/ASVs
Non-singleton OTUs / ASVs
Singleton OTUs / ASV
OTU-level
QIIME-uclust (default)
86.78
86.78
86.78
99.57
86.41
201735
201735
0
QIIME-uclust (Q20)
86.78
86.41
86.41
99.66
86.12
195377
195377
0
QIIME-uclust (e30.ee1)
93.01
82.90
82.90
99.91
82.82
150752
150752
0
MOTHUR (DGC.0)
70.48
66.37
94.17
66.37
23347
16161
7186
MOTHUR (DGC.1)
70.48
66.39
94.19
66.39
12822
12640
182
MOTHUR (DGC.3)
70.48
65.47
92.89
65.47
4022
3832
190
MOTHUR (Opticlust.3)
70.48
65.47
92.89
65.47
5302
5053
249
USEARCH-UPARSE
93.01
82.90
82.90
96.60
89.85
5559
5557
0
ASV-level
DADA2 (ee2)
73.82
71.26
95.00
70.13
26763
26763
0
DADA2 (no filter)
98.12
90.11
90.45
88.75
24469
24469
0
Qiime2-Deblur (default)
74.53
74.53
74.53
51.50
38.39
11120
11120
0
Qiime2-Deblur (Q20)
74.53
74.53
74.53
51.50
38.38
11120
11120
0
Qiime2-Deblur (e30.ee1)
93.01
82.90
82.90
51.04
42.31
11735
11735
0
USEARCH-UNOISE3
93.01
82.90
82.90
97.48
90.67
7659
7519
140
Qiime2-Deblur produced far fewer counts than other pipelines ( Fig 5 , Table 4 ), while QIIME-uclust, USEARCH-UPARSE, and USEARCH-UNOISE3 had conversion rates of more than 90% of initial raw reads. The low conversion rate of Qiime2-Deblur flows is due to the “count substraction”-based algorithm of Deblur [ 7 ], which removes more than 50% of the (filtered reads) counts entering the denoising step ( Table 4 ). The proportion of chimeric reads removed by the different pipelines was very similar, averaging around 1% of raw read counts.
In the HELIUS fecal sample dataset analysis there was a 3.5-fold difference between the highest number of ASVs produced by a pipeline (around 25000, in DADA2) and the lowest number (more than 7500, in USEARCH-UNOISE3). Qiime2-Deblur produced around 11000 ASVs ( Table 4 ). The representative sequence of an OTU may vary depending on the nature of the ( de novo ) clustering algorithm and is influenced by other sequences present, particularly in complex samples. However, by definition, ASV sequences are exact representations of biological sequences and are therefore directly comparable between workflows. We compared ASV sequences from each of the three ASV-level pipelines, identified perfect matches, and constructed Venn diagrams showing the overlap between the different outputs ( Fig 7 ). Only around 4000 ASVs were found by all three pipelines ( Fig 7A ). Around 9000 (around 37% of total) ASVs produced by DADA2 were not found by either USEARCH-UNOISE3 or by Qiime2-Deblur. This mirrors findings from the mock community analysis, where DADA2 showed both the best sensitivity and the highest propensity for spurious ASVs among the three ASV-level pipelines. USEARCH-UNOISE3 and Qiime2-Deblur ach produced more than 2000 ASVs not found by other pipelines. These differences were mostly associated with low-abundance ASVs. To illustrate, 95.5% of counts produced by DADA2 were assigned to ASVs found by USEARCH-UNOISE3, while 98% of counts produced by USEARCH-UNOISE3 were assigned to ASVs found by DADA2. ASVs shared by USEARCH-UNOISE3 and Qiime2-Deblur accounted for around 98% of total counts. Thus, while the number of non-consensus ASVs produced by the ASV-level pipelines was large, it accounted for only 2% to 4% of the total number of counts.
10.1371/journal.pone.0227434.g007
Fig 7
Venn diagram showing the overlap between the ASVs produced by three denoising pipelines from the HELIUS fecal sample data (N = 2170).
Workflows shown are DADA2 (no filter), Qiime2-Deblur (e30.ee1), and USEARCH-UNOISE3 . A) ASVs remaining after rarefaction to 10 000 counts. B) Filtered ASVs (mean relative abundance of at least 0.002% of rarefied counts).
The creation of spurious OTUs/ASVs is a known issue for 16S rRNA bioinformatic pipelines. Some authors recommend applying a minimum relative abundance threshold filter in order to remove OTUs/ASVs with extremely low abundance that have a higher probability of being spurious. When a 0.002% minimum relative abundance filter was applied to the ASVs tables, most non-consensus ASVs were removed ( Fig 7B ). From each pipeline, 1396 ASVs (DADA2), 1294 ASVs (UNOISE3), and 1296 ASVs (Qiime2-Deblur) passed the filter, of which 1023 ASVs were found by all three pipelines. Thus, around 26% of filtered DADA ASVs, 21% of filtered UNOISE3 ASVs, and 21% of filtered Qiime2-Deblur ASVs were found by at most two of the three pipelines (i.e. were non-consensus ASVs). An analysis of closest matches between the (filtered) ASVs produced by the different pipelines showed that while DADA2 and Qiime2-Deblur non-consensus ASVs were generally 1 base away from the closest UNOISE3 ASVs, the most common closest match for UNOISE3 non-consensus ASVs was at 3 or 4 bases distance ( S1 Fig ). Thus, pipeline-specific biases remained after the application of a typical low-abundance filter. Moreover, the filter also removed around 75% of the 4029 consensus ASVs. While abundance-based filters may remove some of the spurious ASVs, they will also remove many true low-abundance biological features.
There was a significant effect of the pipeline on downstream alpha-diversity measures ( Fig 8 ). Ground-truth ASV-level data should always yield higher alpha-diversity than OTU-level data. However, two types of errors can bias perceived alpha-diversity. First, as observed in QIIME-uclust workflows, massive numbers of spurious OTUs can greatly inflate perceived alpha-diversity. Spurious OTUs are responsible for QIIME-uclust yielding much higher alpha-diversity values than all others pipelines, including all ASV-level workflows. The relatively steep downward slope observed for QIIME-uclust when plotting richness as function of rarefaction level or of abundance-based OTU filtering (Figs 8A and 9A ) is indicative of large numbers of very low-abundance OTUs, most of which are likely spurious (Figs 2 and 3 ). The propensity of QIIME-uclust to generate spurious OTUs and inflate alpha-diversity measures has been previously reported by other authors [ 6 , 30 ]. Second, as observed for QIIME2-Deblur, an ASV-level pipeline can fail to distinguish very closely related true biological sequences and clump them together into a single ASV. This will artificially decrease perceived alpha-diversity compared to higher sensitivity ASV-level pipelines, and is the reason why Qiime2-Deblur yielded lower alpha-diversity values compared to DADA2 and USEARCH-UNOISE3 ( Fig 8B ). Pipeline-induced biases (e.g. inflation of sample richness and diversity) cannot be fully addressed using filters ( Fig 9 ). We applied a wide range of filters to the OTU/ASV tables and observed that inter-pipeline differences in alpha-diversity measures remain after the application of typical filters (i.e. 0.002% to 0.005% [ 31 ] of relative abundance).
10.1371/journal.pone.0227434.g008
Fig 8
Alpha-diversity measures at different rarefaction levels.
Values shown are averages across all samples in the HELIUS fecal sample dataset. A) Sample richness (no. of OTUs/ASVs per individual sample). B) Shannon index. Only one workflow from each pipeline is shown: DADA2 (no filter), QIIME-uclust (e30.ee1), Qiime2-Deblur (e30.ee1) and MOTHUR (DGC.1).
10.1371/journal.pone.0227434.g009
Fig 9
Alpha-diversity measures after downstream filtering of very low-abundance OTUs/ASVs.
X-axis shows the no. of counts that an OTU/ASV must reach (in the entire dataset) in order to be retained. All OTU/ASV tables rarefied to 10000 counts / sample prior to filtering. Values shown are averaged across all samples in the HELIUS fecal sample dataset. A) Sample richness. B) Shannon index. The blue vertical bar marks the filter threshold corresponding to 0.002% of rarefied counts.
Conclusion
Large differences in sensitivity and specificity were observed between different pipelines. DADA2 showed the best sensitivity and resolution (followed by USEARCH-UNOISE3) at the cost of producing higher number of spurious ASVs compared to USEARCH-UNOISE3 and Qiime2-Deblur. USEARCH-UPARSE and MOTHUR produced similar numbers of OTUs, especially when a cutoff value was used in MOTHUR to remove singletons or extremely low abundance sequences before clustering. QIIME-uclust workflows produced huge numbers of spurious OTUs as well as inflated alpha-diversity measures, regardless of quality filtering parameters. Current QIIME users may consider switching to other pipelines. Indeed, the authors of QIIME have stopped supporting the platform since 1st January 2018 and are encouraging users to switch over to Qiime2. Biological conclusions based on alpha-diversity measures obtained from QIIME-uclust pipelines may warrant revisiting or confirmation other pipelines. ASV-level workflows offer superior resolution compared to OTU-level, and in this study showed better specificity and lower spurious sequence rates. Moreover, ASV-level pipelines allow for easier inter-study integration of biological features, as ASVs have intrinsic biological meaning, independent of reference database or study context [ 9 ].
We found DADA2 to be the best choice for studies requiring the highest possible biological resolution (e.g. studies focused on differentiating closely related strains). However, USEARCH-UNOISE3 showed arguably the best overall performance, combining high sensitivity with excellent specificity.
Current advances in sequencing technology and bioinformatic pipelines offer new opportunities for ecologists, microbiologists and biomedical scientists. This paper aimed to guide researchers in their choice of the pipeline most suited for their goal while pointing out some of the associated pitfalls and limitations.
Supporting information
S1 Fig
Levenshtein distance from the the ASVs of each pipeline (DADA2, USEARCH-UNOISE3, and Qiime2-Deblur) to the closest ASV in another pipeline's ASV output.
Data is shown for the rarefied ASV tables, filtered using a minimum relative abundance threshold (0.002%). For the Levenshtein distance calculation, DADA2 and UNOISE3 ASVs were trimmed to 250 bp to match the length of Qiime2-Deblur ASVs (which are trimmed to 250 bp in the pipeline flow).
(TIF)
S1 Table
Description of mock sample composition.
Accession number, 16S rRNA gene copy number, and number of 16S rRNA gene amplicon variants are given for each mock community member.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The urinary tract is prone to a number of adverse conditions which impact its functioning. One prominent urinary tract disease is urolithiasis, which is the third most frequent urological affliction in humans [ 1 ]. Stone disease affects 2–20% population worldwide [ 2 ] with a prevalence rate of 15% in India [ 3 ]. The occurrence of CaOx kidney stone is the net effect of a panoply of factors which manifests themselves both within the cells as well as the environment in which the cells are present. The composition of the urinary fluid in terms of various ions, the propensity of these ions to form crystals and their further growth, as well as the presence of macromolecules in the fluid, are some examples of extracellular factors which occur in the tubular lumen and pelvis of the kidney. Cellular events/triggers which take place in the renal epithelial and interstitial cells, comprise of the interaction of oxalate and/or CaOx crystals on renal epithelial cells and how these cells respond to high oxalate and/or CaOx crystals [ 4 ].
Under physiological condition, most of the CaOx crystallizes within urinary tract and is then freely excreted in urine. However, if the crystals are retained within the kidney they have the propensity to grow and develop into stones, which in due course of time lead to injury to the renal epithelial cells and further create a site for the formation of a stone nidus [ 5 ]. Under pathological conditions, exposure to high concentrations of oxalate ions and/or CaOx crystals results in toxicity to renal cells. This damage to the renal cells induces alterations in cell surface properties thus, unmasking attachment site for adhesion and/or internalization of crystals by renal epithelial cells [ 6 , 7 ]. The interaction between renal epithelial cells and oxalate and/or CaOx crystals modifies renal cellular functions as well as the extracellular environment, leading to crystal retention and thus, plays a significant role in CaOx stone formation [ 4 ].
Treatment and prevention strategies for urolithiasis include combination of surgical procedures, medications and dietary manipulations. Despite the technical advancements for stone removal, problems of recurrence persist [ 8 ]. Toxicity, side effects and high cost are the major factors associated with the use of modern synthetic drugs and surgical treatments. To overcome these problems, development of phytotherapeutic agents which can function as an alternate therapy to treat kidney stones is a very attractive option [ 9 ]. Various medicinal plants, which are a part of traditional medicine since ancient times, have been utilized as therapeutic remedies as they have been reported to possess antilithiatic activity [ 10 ].
The current focus of various pharmaceutical industries is on developing plant- protein based, therapeutic drugs. These phytoproteins could be produced on a large scale and modified by recombinant DNA technology [ 11 ] to enhance effectiveness, reduce immunogenicity, allergenic properties as well as toxicity, so as to be safer for use [ 12 ]. Most of the plant based antilithiatic proteins identified till date are anionic in nature, and this is due to the presence of acidic amino acids and/or calcium binding domains (EF Hand motifs) [ 13 ]. These acidic amino acids bind to the calcium thus preventing its interaction with oxalate and adherence to renal cells [ 14 ].
With this background the present study was designed to evaluate the bioactivity of anionic anti-calcifying proteins of dried bark T . arjuna , on oxalate injured renal epithelial cells, so as to establish a scientific foundation for the anti urolithiatic potential of T . arjuna . This plant which belongs to the family Combretaceae, is widely used in the traditional system of medicine for various disorders of the cardiovascular system. It is also known to possess antioxidants, anti-inflammatory and mild diuretic activity [ 15 ]. Studies have also shown that T . arjuna inhibits in vitro CaOx crystallization and crystal adhesion to renal epithelial cells [ 16 – 18 ]. This study was undertaken to isolate, purify and characterize antilithiatic proteins from the dried bark of Terminalia arjuna and assess their influence on different stages of CaOx stone formation. In this study, we present in vitro evidence for the presence of four anionic antilithiatic proteins from the bark of Terminalia arjuna which could play a crucial role in inhibiting CaOx formation.
Materials and Methods
Plant
The dried bark of Terminalia arjuna was purchased from Natural Remedies Pvt. Ltd., Bangalore, India. A collection voucher specimen is available at the company.
CaOx crystallization assay to measure inhibitory activity
Inhibitory activity against CaOx crystal nucleation and aggregation was measured using a time-course measurement of optical density as described previously by Hess et al. [ 19 ] with some modifications. Stock solutions of 10.0mM calcium chloride (CaCl 2 ) and 1.0 mM sodium oxalate (Na 2 C 2 O 4 ), containing 200 mM sodium chloride (NaCl) and 10 mM sodium acetate, were adjusted to pH 5.7. For crystallization experiments, the solutions were warmed up to 37°C followed by addition of 1.5 mL of the CaCl 2 solution and 1.5 mL of the Na 2 C 2 O 4 solution in a cuvette to achieve final assay concentrations of 5.0 mM for calcium and 0.5 mM for oxalate, respectively. In the cuvette, the final solutions were stirred continuously and maintained at 37°C and optical density at 620 nm (OD 620 ) was recorded after every 60 seconds over 40 minutes. Experiments with protein samples (100 μL) were extended to 60 minutes due to lower rate of nucleation and aggregation. All crystallization experiments were performed thrice in triplicate. Slopes of nucleation ( S N ) and aggregation ( S A ) phases were calculated using linear regression analysis. Using the slopes, percentage inhibitory activity of protein sample was calculated as (1-(Tsi/Tsc)) x100, where Tsc was the turbidity slope of the control and Tsi the turbidity slope in the presence of the inhibitor.
CaOx crystal growth assay to measure inhibitory activity
CaOx crystal growth inhibitory activity was measured using the seeded solution-depletion assay described previously by Nakagawa et al. [ 20 ] with some modifications. Briefly, aqueous solution 4 mM of CaCl 2 , 4 mM Na 2 C 2 O 4 and 10 mM Tris-HCl containing 90 mM NaCl (pH 7.2) were prepared and equilibrated at 37°C. Stone slurry (1.5 mg/mL) was prepared in 50 mM sodium acetate buffer (pH 5.7) and 30 μL of stone slurry was added to a solution containing 1 mM CaCl 2 and 1 mM Na 2 C 2 O 4 . The reaction of CaCl 2 and Na 2 C 2 O 4 with crystal seed led to deposition of CaOx crystals on the crystal seed surfaces, reflecting the loss of oxalate due to CaOx crystal growth that is detectable by spectrophotometry at λ214 nm. When protein sample (10 μL) is added into this solution, the consumption of free oxalate to form CaOx crystals will decrease if the sample inhibits CaOx crystal growth. The absorbance was monitored after every 60 seconds for 20 minutes at 214 nm. Rate of depletion of free oxalate was calculated using the baseline value and the value after 60 seconds for 20 minutes, with or without protein sample. The percentage inhibitory activity was calculated as ((C-S)/C) x100, where C is the rate of reduction of free oxalate without a protein sample and S is the rate of reduction of free oxalate with a protein sample.
Protein extraction and purification from the dried bark of Terminalia arjuna
The extraction and purification of proteins from the dried bark of T . arjuna was conducted as described in [ 21 ] with slight modifications. The dried bark of T . arjuna was ground to fine powder. To obtain whole protein extract, 100 grams of bark powder was extracted with extraction buffer (50 mM Tris-Cl buffer (pH 7.4), containing 0.25 M NaCl, 1 mM Phenylmethylsulfonyl fluoride, 0.01% sodium azide and 5% Polyvinylpyrrolidone). The slurry was stirred continuously for 24 hours at 4°C. After 24 hours, the slurry was centrifuged at 10,000 g for 20 minutes at 4°C. The supernatant was collected and stored at -20°C for further experimentation. This supernatant was referred to as the whole protein extract of Terminalia arjuna . Whole protein extract (WPE) of Terminalia arjuna was separated into <3 kDa and >3 kDa fractions and dialyzed against 10 mM Tris-Cl buffer at pH 7.4 by centrifugation with the help of Amicon Ultra-4 centrifugal separating tubes (Millipore) of 3 kDa cut off molecular weight. Thus, two fractions <3 kDa and >3 kDa were obtained. Whole protein extract, <3 kDa and >3 kDa fractions of T . arjuna were assessed for CaOx crystallization and crystal growth inhibitory activity. The >3 kDa fraction exhibited significant inhibitory activity and was loaded on to strong anion exchanger Q Sepharose (GE Healthcare) packed in a column (XK 16/20) equilibrated with 10 mM Tris-Cl buffer (pH 7.4). Bound proteins were eluted by using a linear concentration gradient of NaCl (0–1 M) in 10 mM Tris-Cl buffer (pH 7.4) at a flow rate of 0.5 mL/min. Fractions under each peak were pooled, dialyzed against 10 mM Tris-Cl buffer (pH 7.4) and their inhibitory bioactivity towards CaOx crystal nucleation, aggregation and growth was studied. All the peaks obtained were concentrated and loaded one by one on a Bio gel® P-100 gel molecular sieve column (XK 16/70) equilibrated and eluted with the 10 mM Tris-Cl buffer (pH 7.4) at a flow rate of 0.2 mL/min. The eluted fractions under each peak were pooled to study their activity w.r.t. CaOx crystal nucleation, aggregation and growth as well as on oxalate induced injury to NRK-52E and MDCK cells. The homogeneity of purified proteins was analyzed by Native-PAGE and reverse phase HPLC. Total protein concentration was determined by Bradford Assay at each step of purification using Bovine Serum Albumin (BSA) as a standard [ 22 ].
Native-PAGE
The purified proteins obtained from molecular sieve chromatography were reconstituted in non-reducing sample buffer and analyzed by native page using 1 mm thick, 12% resolving gel and 5% stacking gel without using SDS or 2-mercaptoethanol (reducing agent) with a Mini-Protean III apparatus (Bio-Rad) at 100 V [ 23 , 24 ]. The gels were stained using silver staining.
Homogeneity of purified proteins by High Performance Liquid Chromatography
Homogeneity of purified proteins (A1, B1, B2, C1) was determined by performing reverse phase HPLC using Waters Spherisorb® C18 (5 μm, 4.6 X 250 mm) column with solvent A (0.1% TFA in water) and solvent B (100% acetonitrile containing 0.1% TFA) in a linear gradient of acetonitrile (20–70%) over a period of 50 minutes at a flow rate of 1 mL/min. The sample injection volume was 20 μL and column was washed with solvent A and brought to 20% acetonitrile in 5 minutes. The protein peak was detected at 280 nm using Waters 2996 photodiode array detector [ 25 ].
Trypsin digestion and peptide mass fingerprinting
The identified bands after analysis were excised and diced into small pieces (1 mm) followed by de-staining using 1:1 (v/v) of potassium ferricyanide and sodium thiosulfate for 10 minutes, and this process was repeated 3–4 times until they become translucent white. They were dehydrated using 100% acetonitrile and Speedvac to complete dryness. The gel pieces were rehydrated with 1.5 mg/mL DTT and incubated for an hour. After incubation, the DTT solution was removed, and further incubated with 10 mg/mL of iodoacetamide for 45 minutes. The supernatant was removed and the gel was dehydrated with 100% acetonitrile for 10 minutes and Speedvac till completely dry. 12 ng/μL trypsin solution was added and incubated overnight at 37°C. The peptides were extracted with 1:1 5% formic acid/H 2 O and final extraction with 1:1 5% formic acid/acetonitrile was performed thrice, followed by Speedvac to complete dryness. The dried peptide mix was suspended in 50% trifluoroacetic acid (TFA) containing 0.1% acetonitrile (ACN) buffer. The peptides obtained were mixed with α-cyano-4-hydroxycinnamic acid (HCCA) matrix in 1:1 ratio and resulting 2 μL was spotted on the MALDI TOF/TOF ULTRAFLEX III instrument and further analysis was done with FLEX ANALYSIS SOFTWARE for obtaining the Peptide Mass Fingerprint.
Protein identification and domain prediction
The mass over charge ratio obtained in the peptide mass fingerprinting were submitted for MASCOT search in MASCOT search engine ( https://www.matrixscience.com ) using SwissProt database for the identification of the proteins. The search parameters used were monoisotopic, oxidized at methionine residues as a variable modification and carbamidomethylated at cysteine residues as a fixed modification. The search was done using a Viridiplantae taxonomy, only one missed tryptic cleavage and peptide mass tolerance of 120 parts per million. The presence of the active domains was found using the online tool, ScanProsite and the amino acid sequence of the hit obtained from MASCOT search was used as an input to search for the presence of active domain.
Preparation of protein samples for cell line studies
The >3 kDa fraction and purified proteins obtained after molecular sieve chromatography were dialyzed against distilled water by centrifugation through Amicon Ultra-4 centrifugal separating tubes (Millipore) of 3 kDa cut off molecular weight and then lyophilized. The lyophilized >3 kDa fraction and purified proteins were reconstituted in serum free DMEM and filtered by 0.22 μm syringe filter. Filtered solution of 1.44 mM Tris-Cl buffer (pH 7.4) was used as a solvent system.
Cell culture
Experimental studies were done using in vitro models of Normal rat epithelial derived renal tubular epithelial (NRK-52E) and Madin-Darby Canine Kidney (MDCK) cell lines obtained from NCCS (National Centre for Cell Science), Pune, India. Both the cell lines were maintained in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% Fetal Bovine Serum (FBS) and 1% Penicillin (100 units/mL)-Streptomycin (10,000 μg/mL). Cells were cultured in 25 cm 2 and 75 cm 2 tissue-culture treated flasks at 37°C under 5% CO 2 in humidified chambers [ 26 ]. The cells were grown to 80% confluency for subsequent subculture in preparation for experiments.
Exposure to oxalate
A stock solution of 10 mM sodium oxalate was prepared and diluted to 2 mM in serum-free medium. NRK-52E cells and MDCK cells were exposed to 2 mM sodium oxalate in the absence and presence of different concentrations of >3 kDa fraction and purified proteins (A1, B1, B2, C1) for a period of 48 hours [ 27 , 28 ]. Cystone drug at a concentration of 10 μg/mL was used as a positive control.
MTT assay
1x10 4 cells were seeded into each well of a 96-well microplate and incubated at 37°C and 5% CO 2 in humidified chambers. At 80% confluency, cells were treated with 2 mM sodium oxalate in the absence and presence of different concentrations (4 μg/mL, 6 μg/mL, 8 μg/mL, 10 μg/mL) of >3 kDa fraction and purified proteins (A1, B1, B2, C1) for 48 hours at 37°C. At the end of the treatment, 25 μL of MTT reagent (final concentration of 0.5 mg/mL) was added to each well and incubated for 4 hours at 37°C. After incubation, medium was replaced with 200 μL DMSO (100%) and kept still at room temperature for 15–20 minutes. After gentle mixing, absorbance values were determined at a 570 nm test wavelength and a 630 nm reference wavelength to evaluate the cell viability using a microplate reader (Model 680, Bio-Rad) [ 29 ].
CaOx crystal adhesion
Cells were seeded on sterile glass coverslips placed in a 6-well plate at a density of 2x10 5 cells/coverslip and cultured at 37°C and 5% CO 2 in humidified chambers to achieve 80% confluency. After incubation, cells were treated with 2 mM sodium oxalate in the absence and presence of >3 kDa fraction and purified proteins (A1, B1, B2, C1) at a concentration of 10 μg/mL for 48 hours at 37°C. After the treatment, medium was removed and cells were washed twice with 1X PBS followed by fixation of cells with 4% paraformaldehyde for 30 minutes. After incubation, cells were again washed twice with 1X PBS and then observed under phase contrast and polarization upright microscope (BX53, Olympus Corporation, Japan) at a magnification of 20X to study cell-crystal interactions [ 30 ].
Hoechst 33258 staining
Cells were seeded on sterile glass coverslips placed in a 6-well plate at a density of 2x10 5 cells/coverslip and cultured at 37°C and 5% CO 2 in humidified chambers to achieve 80% confluence. After incubation, cells were treated with 2 mM sodium oxalate in the absence and presence of >3 kDa fraction and purified proteins (A1, B1, B2, C1) at a concentration of 10 μg/mL for 48 hours at 37°C. At the end of treatment, medium was removed and cells were washed twice with 1X PBS followed by fixation of cells with 4% paraformaldehyde for 30 minutes. After incubation, cells were washed twice with 1X PBS, stained with 5 μg/mL of Hoechst 33258 dye for 10 minutes at room temperature in the dark and again washed twice with 1X PBS. Stained nuclei were observed under fluorescence upright microscope (BX53, Olympus Corporation, Japan) at a magnification of 20X [ 31 ].
Annexin V/ propidium iodide staining
6x10 5 cells were seeded into 60 mm dishes and incubated at 37°C and 5% CO 2 in humidified chambers. At 80% confluency, cells were treated with 2 mM sodium oxalate in the absence and presence of >3 kDa fraction and purified proteins (A1, B1, B2, C1) at a concentration of 10 μg/mL for 48 hours at 37°C. After treatment, the subsequent procedure followed was in accordance to the instructions of BD Pharminogen TM FITC ANNEXIN V Apoptosis Detection Kit 1 (catalogue no. 556547), where, cell suspension and cells from monolayer were pooled together. The cells were washed with cold 1X PBS twice. The pellet was resuspended in 100 μL of 1X binding buffer followed by addition of 2 μL of FITC Annexin V and 2 μL of Propidium iodide (PI). The cells were gently vortexed and incubated for 15 minutes at room temperature in the dark. 400 μL of 1X binding buffer was added to each group and the cells analyzed by flow cytometry (BD Accuri C6, BD Biosciences).
Detection of active caspase-3
6x10 5 cells were seeded into 60 mm dishes and incubated at 37°C and 5% CO 2 in humidified chambers. At 80% confluency, cells were treated with 2 mM sodium oxalate in the absence and presence of >3 kDa fraction and purified proteins (A1, B1, B2, C1) at a concentration of 10 μg/mL for 48 hours at 37°C. At the end of the treatment, the subsequent procedure followed was in accordance to the instructions of BD Pharminogen TM FITC Active Caspase-3 Apoptosis Kit (catalogue no. 550480), where, cell suspension and cells from monolayer were pooled together. Cells were washed with cold PBS twice and resuspended in BD Cytofix/Cytoperm solution. After incubation on ice for 20 minutes, BD Cytofix/Cytoperm solution was discarded. The cells were washed twice with 1X BD Perm/Wash buffer at room temperature. The cells were then resuspended in the 1X BD Perm/Wash buffer plus 10 μL of antibody and incubated for 30 minutes at room temperature. After incubation, the cells were washed with 0.5 mL of 1X BD Perm/Wash buffer and then resuspended in 0.5 mL of 1X BD Perm/Wash buffer for analysis by flow cytometry (BD Accuri C6, BD Biosciences).
Statistical analysis
Statistical procedures were performed with GraphPad Prism software version 6.01. The statistically different groups were identified by one-way analysis of variance (ANOVA), followed by Dunnet’s multiple comparisons test. Results were expressed as the mean ± SD. A p-value of <0.05 was considered significant. All the experiments were performed three times, each time in triplicate.
Results
CaOx crystal nucleation, aggregation and growth inhibitory activity of whole protein extract, <3 kDa and >3 kDa fractions
Inhibitory activity of whole protein extract, <3 kDa fraction and >3 kDa fraction was studied individually against CaOx crystal nucleation, aggregation and growth kinetics and inhibitory activity wherein a dose dependent effect was seen. Although the whole protein extract and <3 kDa fraction showed inhibitory activity towards CaOx crystal nucleation ( Fig 1A ), aggregation ( Fig 1B ) and growth assay ( Fig 1C ) system, however, the >3 kDa fraction exhibited maximum inhibitory activity ( Fig 1A, 1B and 1C ) and was taken further for bioactivity guided purification.
10.1371/journal.pone.0162600.g001
Fig 1
CaOx crystallization inhibitory activity of WPE, < 3kDa and > 3kDa fractions.
Effect of WPE, <3 kD and >3 kDa fraction on calcium oxalate crystallization. (A)Nucleation. (B)Aggregation fraction. (C)Growth. Data are mean ± S.D of three independent observations. The statistically different groups were identified by one-way analysis of variance (ANOVA), followed by Dunnet’s multiple comparisons test. * p < 0.05 vs control, ** p < 0.01 vs control, *** p < 0.001 vs control and ‘ns’ represents not significant.
Purification of antilithiatic proteins
In order to purify the antilithiatic proteins from the > 3 kDa fraction of T . arjuna a multi- step protocol involving anion exchange and gel filtration chromatography, followed by Native-PAGE was undertaken ( S1 Fig ). On subjecting the > 3 kDa fraction to an anion exchanger Q Sepharose column, successive fractions under each peak with increasing gradient of NaCl were collected, pooled and named P1 to P3 ( Fig 2 ). These eluted peaks were dialyzed against 10 mM Tris-Cl buffer (pH 7.4) and it was seen that peaks P1, P2 and P3 exhibited significant inhibitory activity against CaOx crystal nucleation, aggregation and growth kinetics ( Fig 3 ). The order of inhibitory activity was peak P2 > peak P3 >peak P1 at a concentration of 10 μg/mL.
10.1371/journal.pone.0162600.g002
Fig 2
Anion exchange chromatography profile of >3 kDa fraction.
Elution profile of >3 kDa fraction of T . arjuna loaded on Q Sepharose anion exchanger, with a linear gradient of NaCl (0–1 M).
10.1371/journal.pone.0162600.g003
Fig 3
CaOx crystallization inhibitory activity of peaks P1, P2 and P3.
Percentage inhibitory activity of peaks of anion exchange chromatography on nucleation, aggregation and growth of CaOx crystals. Data are mean ± S.D of three independent observations. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. * p < 0.05 vs control, ** p < 0.01 vs control, *** p < 0.001 vs control.
Anionic protein separation was followed by molecular separation of proteins by gel filtration chromatography. Peaks P1, P2 and P3 were individually loaded (Figs 4 , 5 and 6 ) on a Bio gel ® P-100 gel packed column (XK 16/70) and proteins were eluted isocratically with 10 mM Tris-Cl buffer (pH 7.4). On subjecting peak 1, peak 2 and peak 3 to gel filtration chromatography the subsequent fractions obtained were pooled and named as A1 to A4, B1 to B3 and C1 to C2, respectively. These peaks were then assessed for their inhibitory activity and it was seen that peaks A1, B1, B2, C1 exhibited the maximum inhibitory potential against CaOx crystal nucleation, aggregation and growth kinetics ( Fig 7 ). We observed appreciable inhibition with 10 μg/ml in the > 3kDa fraction which basically contained a number of proteins, and following the purification procedure, the specific activity of these proteins towards inhibition of nucleation, aggregation and growth was more pronounced.
10.1371/journal.pone.0162600.g004
Fig 4
Gel filtration chromatography profile of peak P1.
Chromatogram of peak P1 subjected to Bio gel ® P-100 gel packed column and proteins eluted in an isocratic 10 mM Tris-Cl buffer (pH 7.4).
10.1371/journal.pone.0162600.g005
Fig 5
Gel filtration chromatography profile of peak P2.
Chromatogram of peak P2 subjected to Bio gel ® P-100 gel packed column and proteins eluted in an isocratic 10 mM Tris-Cl buffer (pH 7.4).
10.1371/journal.pone.0162600.g006
Fig 6
Gel filtration chromatography profile of peak P3.
Chromatogram of peak P3 subjected to Bio gel ® P-100 gel packed column and proteins eluted in an isocratic 10 mM Tris-Cl buffer (pH 7.4).
10.1371/journal.pone.0162600.g007
Fig 7
CaOx crystallization inhibitory activity of eluted purified proteins.
Percentage inhibitory activity of most active peaks obtained from gel filtration chromatography on nucleation, aggregation and growth of CaOx crystals. Data are mean ± S.D of three independent observations. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. * p < 0.05 vs control, ** p < 0.01 vs control, *** p < 0.001 vs control.
The purified peaks A1, B1, B2, C1 exhibiting marked inhibitory activity were subjected to Native-PAGE ( Fig 8 ) wherein single bands of MW ~190 kDa, ~130 kDa, ~90 kDa and ~90 kDa, respectively, were seen. Homogeneity of purified proteins was further verified using RP-HPLC, wherein single peaks were observed ( Fig 9 ).
10.1371/journal.pone.0162600.g008
Fig 8
Native-PAGE analysis.
Native-PAGE of most active peaks with maximum inhibitory activity obtained from gel filtration chromatography. The short red horizontal line indicates the position of the purified protein band. Lane 1: Broad range marker (Biorad), Lane 2: A1, Lane 3: B1, Lane 4: B2 and Lane 5: C1.
10.1371/journal.pone.0162600.g009
Fig 9
Homogeneity of purified proteins by HPLC analysis.
Reverse phase HPLC analysis of purified protein (A) A1. (B) B1. (C) B2. (D) C1.
MALDI-TOF Mass spectrometric analysis and identification of purified proteins
The purified proteins obtained were subjected to MALDI-TOF MS and MASCOT search engine analysis. The mass over charge ratio data obtained from the MALDI-TOF of the peaks A1, B1, B2, C1 matched significantly with Nuclear pore anchor, DEAD Box ATP-dependent RNA helicase 45, Lon protease homolog 1 and Heat shock protein 90–3, respectively ( Table 1 ). The amino acid sequence of these respective proteins obtained from MASCOT search was used as an input to search for the presence of active domain using ScanProsite ( Table 1 ).
10.1371/journal.pone.0162600.t001
Table 1 Summary of proteins identified and characterized using the Mascot search engine and ScanProsite.
Purified protein peaks
Significant protein match from MASCOT search engine
Matching Score
Sequence Coverage
Molecular Weight (Da)
pI
Active Domains/Motifs from ScanProsite
A1
Nuclear pore anchor
50
15
237,637
5.01
• Lysine rich region • Glutamic acid rich region
B1
DEAD Box ATP-dependent RNA helicase 45
33
14
105,223
5.14
• Glutamic acid rich region • Aspartic acid rich region • Alanine rich region
B2
Lon protease homolog 1
49
22
104,379
5.42
• Cell attachment sequence i.e. RGD (Arg-Gly-Asp)
C1
Heat shock protein 90–3
27
11
80,287
4.95
• Glutamic acid rich region
Protective effect of purified proteins against oxalate-induced renal epithelial cell injury (MTT assay)
The protective effect of >3 kDa fraction ( Fig 10 ) and purified proteins (A1, B1, B2, C1) ( Fig 11 ) of Terminalia arjuna against oxalate induced injury to NRK-52E cells and MDCK cells was evaluated after 48 hours of treatment, by measuring the reduction of yellow tetrazolium to purple formazan crystals. Cells treated with serum free medium were taken as the untreated control group. On exposing the renal epithelial cells to either the solvent system (1.44 mM Tris-Cl buffer pH 7.4) or >3 kDa fraction (10 μg/mL) or the purified proteins (10 μg/mL), no significant effect on the cell viability was observed. The dose of 10 μg/ml was chosen on the basis of studies carried out earlier in our lab using
10.1371/journal.pone.0162600.g010
Fig 10
Protective effect of >3 kDa fraction on oxalate injured renal cells (MTT assay).
(A) Assessment of viability of oxalate injured NRK-52E cells in the presence of >3 kDa fraction of T . arjuna by MTT assay. Data are mean ± S.D of three independent observations. ns: not significant. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. ‘*’ represents P values versus untreated cells (control) and ‘#’ represents P values versus oxalate treated cells, where * p< 0.05 vs control, ** p < 0.005 vs control, *** p < 0.001 vs control, # p < 0.05 vs oxalate, ## p < 0.01 vs oxalate. (B) Assessment of viability of oxalate injured MDCK cells in the presence of >3 kDa fraction of T . arjuna by MTT assay. Data are mean ± S.D of three independent observations. ns: not significant. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. ‘*’ represents P values versus untreated cells (control) and ‘#’ represents P values versus oxalate injured cells, where * p< 0.05 vs control, ** p < 0.005 vs control, *** p < 0.001 vs control, # p < 0.05 vs oxalate, ## p < 0.005 vs oxalate.
10.1371/journal.pone.0162600.g011
Fig 11
Protective effect of proteins on oxalate injured renal cells (MTT assay).
(A) Effect of purified proteins of T . arjuna on oxalate injured NRK-52E cell viability assessed by MTT assay. Data are mean ± S.D of three independent observations. ns: not significant. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. ** p< 0.01 vs control, *** p < 0.001 vs control and # p < 0.05 vs oxalate, ## p < 0.01vs oxalate. (B) The effect of purified proteins of T . arjuna on oxalate injured MDCK cell viability assessed by MTT assay. Data are mean ± S.D of three independent observations. ns: not significant. All treatment groups were simultaneously compared via one-way ANOVA using Dunnett’s multiple comparisons test. ‘*’ represents P values versus untreated cells (control) and ‘#’ represents P values versus oxalate injured cells l, ** p< 0.01vs control, *** p < 0.001vs control; and # p < 0.05 vs oxalate, ## p < 0.01vs oxalate.
> 3kDa and protein fractions isolated from other plants wherein, we had seen protection against oxalate induced cell injury. Based on our previous findings we tested various concentrations of > 3kDa and protein fractions of Terminalia arjuna , however, at concentrations beyond 10ug/ml protection against oxalate injury was lost, resulting in a decrease in cell viability (data not shown). This cell killing could be attributed to an increase in the concentration of salts in the protein fractions, as the extraction procedure was carried out in Tris-Cl buffer.
Marked cell death was seen on exposure to 2 mM oxalate wherein, there was a sharp decline in viability from 100% in untreated cells to 23.13 ± 1.64% (p<0.001) in NRK-52E cells and 24.36 ± 0.97% (p<0.001) in MDCK cells, respectively. When the oxalate injured cells were co-treated for 48 hours with either the >3 kDa fraction or the purified proteins, the viability significantly increased in a concentration dependent manner. Cystone at a concentration of 10 μg/mL was used as a positive control and which was also shown to protect the cells from oxalate induced injury.
Loss of CaOx crystal adherence to renal tubular epithelial cells by purified proteins
To see the manner in which oxalate injures the renal epithelial cells, CaOx crystal adhesion to renal cell surface of was studies in both the NRK-52E cells ( Fig 12 ) and MDCK cells ( Fig 13 ) 48 hours after treatment. Cells incubated with serum free defined medium were considered as untreated cells i.e. control group. In both the cell lines, cells treated with the solvent system, >3 kDa fraction and purified proteins (A1, B1, B2, C1) showed healthy cellular morphology, indicating that there was no adverse effect to the cells (data not shown). Upon exposure of NRK-52E cells and MDCK cells to oxalate, a marked change in the morphology of the cells was apparent with a decrease in size, increase in the granularity and lesser number of cells attached to the dish with a number of dead cells floating in the medium. In addition, COM crystals were observed microscopically, when the renal cells were treated with 2mM sodium oxalate, and these crystals remained adhered tightly to the cells even after several washes using PBS. The adhesion of the crystals led to cellular damage and eventually cell death, which was reflected by a concomitant decrease in the number of viable cells, as compared to untreated cells. The treatment of oxalate injured NRK-52E and MDCK cells with 10 μg/mL of >3 kDa fraction or purified proteins (A1, B1, B2, C1) disrupted the interaction between cells and crystals, showing loss of crystal adherence to cells, which could be visualized from the polarization micrographs, in which very few crystals could be seen as compared to the oxalate injured cells. On observing the morphology of the cells, they appeared to be similar in morphology to the control cells and further was an increase in the number of viable cells attached in the culture dish. Treatment with Cystone also exhibited similar cytoprotective effects.
10.1371/journal.pone.0162600.g012
Fig 12
Decreased CaOx crystal adhesion to NRK-52E cells by proteins.
Effect of >3 kDa fraction and purified proteins of T . arjuna on CaOx crystal adherence to oxalate injured NRK-52E cells, visualized under polarization and phase contrast at magnification 20X and scale bar 100 microns.
10.1371/journal.pone.0162600.g013
Fig 13
Decreased CaOx crystal adhesion to MDCK cells by proteins.
Effect of >3 kDa fraction and purified proteins of T . arjuna on CaOx crystal adherence to oxalate injured MDCK cells, visualized under polarization and phase contrast at magnification 20X and scale bar 100 microns.
Hoechst staining of oxalate injured renal tubular epithelial cells to detect apoptosis
To ascertain the nuclear changes, NRK-52E cells ( Fig 14 ) and MDCK cells ( Fig 15 ) were stained with Hoechst 33258 dye. The untreated NRK-52E and MDCK nuclei reflected a healthy morphology with intact chromatin. Treatment of cells with the solvent system, >3 kDa fraction and purified proteins (A1, B1, B2, C1) in the absence of any oxalate damage caused no harmful effects to cells w.r.t. control and nuclei showed intact chromatin (data not shown). However, when both the cell lines were exposed to oxalate, marked changes were observed w.r.t. untreated control group. The nuclear changes reflected signs fragmentation (inset) and were brightly stained, indicative of early apoptotic changes. The effect of 10 μg/mL of >3 kDa fraction and purified proteins (A1, B1, B2, C1) on oxalate treated NRK-52E cells and MDCK cells was assessed and it was apparent that treatment protected the cells as nuclear morphology was similar to the untreated cells. Similar changes were seen upon treatment with Cystone, which also lead to more viable cells as compared to oxalate injured cells.
10.1371/journal.pone.0162600.g014
Fig 14
Detection of oxalate induced apoptosis in NRK-52E cells by Hoechst staining.
Effect of >3 kDa fraction and purified proteins of T . arjuna on oxalate induced apoptosis in NRK-52E cells, visualized under fluorescence microscopy at magnification 20X; scale bar 100 microns.
10.1371/journal.pone.0162600.g015
Fig 15
Detection of oxalate induced apoptosis in MDCK cells by Hoechst staining.
Effect of >3 kDa fraction and purified proteins of T . arjuna on oxalate induced apoptosis in MDCK cells, visualized under fluorescence microscopy at magnification 20X; scale bar 100 microns.
Reduction of oxalate-induced apoptosis in renal tubular epithelial cell by purified proteins
Oxalate-induced apoptosis was assessed after 48 hours in NRK-52E cells ( Fig 16 ) and MDCK cells ( Fig 17 ) using Annexin V/PI staining by flow cytometry. There was no significant apoptosis in the untreated NRK-52E cells and MDCK cells as the maximum number of cells were present in the lower left hand quadrant indicative of a live cell population. Treatment of cells with either the solvent system, or >3 kDa fraction or purified proteins (A1, B1, B2, C1) showed a similar trend (data not shown). Following treatment of the NRK-52E and MDCK cells with oxalate, the percent cell death (early apoptosis, lower right hand quadrant) increased from 0.2% in control to 58% in NRK-52E cells and 64% in MDCK cells. The number of late apoptotic cells (top right hand quadrant) also increased in comparison to the respective controls to 13.2% in NRK-52E and 8.3% in MDCK cells. Treatment with 10 μg/mL of >3 kDa fraction or purified proteins (A1, B1, B2, C1) on oxalate injured NRK-52E and MDCK cells significantly reduced the number of apoptotic cells in comparison to the oxalate treatment alone to 35.9% and 32.2%, 13.9%, 22.2% and 26% in NRK-52E cells and to 32.4% and 31%, 17.2%, 16.3% and 22.3% in MDCK cells, respectively. These results, alluded to the fact that >3 kDa fraction and purified proteins protected the cells from oxalate induced apoptosis. The addition of Cystone to oxalate treated renal cells also improved cell viability in comparison to oxalate injured cells from 28.2% to 40.5% in NRK-52E cells and from 26.7% to 42% in MDCK cells, respectively.
10.1371/journal.pone.0162600.g016
Fig 16
Attenuation of apoptotic cell death in NRK-52E cells by T . arjuna proteins.
Flow cytometry analysis showing the effect of >3 kDa and purified proteins of T . arjuna on oxalate induced apoptosis in NRK-52E cells, visualized by AnnexinV/PI staining.
10.1371/journal.pone.0162600.g017
Fig 17
Attenuation of apoptotic cell death in MDCK cells by T . arjuna proteins.
Flow cytometry analysis showing the effect of >3 kDa and purified proteins of T . arjuna on oxalate induced apoptosis in MDCK cells, visualized by Annexin V/PI staining.
Oxalate-induced apoptosis in NRK-52E cells ( Fig 18 ) and MDCK cells ( Fig 19 ) was further evaluated by anti-active Caspase-3 antibody, which preferentially stains the cells undergoing apoptosis. When both the cell lines were treated with oxalate for a period of 48 hours, the cells undergoing apoptosis increased from 20.9% in control to 75.7% in NRK-52E cells and from 20.1% in control to 78.4% in MDCK cells, respectively. The treatment of 10 μg/mL of >3 kDa fraction or purified proteins A1, B1, B2 and C1 on oxalate injured cells significantly reduced the number of cells expressing active caspase-3 enzyme to 56.5% and 55.4%, 41%, 39.7% and 51.1% in NRK-52E cells and to 56.6% and 54.4%, 37.4%, 38.6% and 48.6% in MDCK cells, respectively, indicating that >3 kDa fraction and purified proteins diminished oxalate induced apoptosis to the renal cells. Although cystone treatment on oxalate injured renal cells did reduce the level of apoptosis to 64.9% in NRK-52E cells and to 64.1% in MDCK cells, respectively, however, the effect was less as compared to the injured cells treated with the purified proteins.
10.1371/journal.pone.0162600.g018
Fig 18
Diminution of Active Caspase-3 in oxalate injured NRK-52E cells.
Flow cytometry analysis showing the effect of >3 kDa and purified proteins of T . arjuna on oxalate induced apoptosis in NRK-52E cells, visualized by Anti-Active Caspase-3 antibody staining. M1 depicts the percentage of viable cells within the margins of marker M1; M2 depicts the percentage of cells undergoing apoptosis within the margins of marker M2(shown in blue).
10.1371/journal.pone.0162600.g019
Fig 19
Diminution of Active Caspase-3 in oxalate injured MDCK cells.
Flow cytometry analysis showing the effect of >3 kDa and purified proteins of T . arjuna on oxalate induced apoptosis in MDCK cells, visualized by Anti-Active Caspase-3 antibody staining. M1 depicts the percentage of viable cells within the margins of marker M1; M2 depicts the percentage of cells undergoing apoptosis within the margins of marker M2(shown in blue).
Discussion
Studies using cultured renal cell lines have put forth strong evidence that hyperoxaluria is directly responsible for causing injury to renal epithelial cells, thereby putting into motion the process for the development of kidney stones [ 32 ]. In vitro studies have suggested that exposure of renal epithelial cells to oxalate and/or CaOx crystals causes an increase in expression of immediate early genes (c- myc , Egr-1 , c- jun and nur-77 ) [ 33 – 35 ] and production of urinary macromolecules (Tamm-Horsfall protein, Osteopontin, Prothrombin fragment-1, Bikunin and inter-α-inhibitor, α 1 -Microglobulin, CD44, Calgranulin, Heparan sulfate, Osteonectin, Fibronectin, Matrix Gla Protein), modulating CaOx crystallization and crystal retention in the kidneys, as an adaptive responses of cells to oxalate [ 36 ]. Urinary Trefoil Factor 1 (TFF1) is an inhibitor of CaOx crystal growth, as it contains 4C-terminal glutamic residues that binds with calcium ions and prevents CaOx crystal growth [ 37 ]. Fibronectin, containing a tripeptide cell attachment sequence RGD, interacts with cells, thus inhibiting endocytosis of CaOx crystals, thereby protecting renal cells from oxalate induced injury [ 38 ]. Osteopontin, also contains this tripeptide sequence and a calcium binding site [ 39 ]. This protein has also shown to inhibit nucleation, aggregation, growth and cellular attachment of CaOx crystals [ 40 , 41 ]. Several other inhibitory urinary crystallization modulators are produced in kidneys and isolated from the kidney stones and urine [ 42 ].
Oxalate is a metabolite excreted by the kidney [ 43 , 44 ] that produces oxidant stress [ 32 ] and death of renal cells [ 45 ] at pathophysiological concentrations. High levels of oxalate exposure to renal cells may be responsible for varied morbidities seen in the kidneys, including CaOx nephrolithiasis [ 46 , 47 ]. For cell line based in vitro studies, the foremost step was to create a hyperoxaluric condition and this was achieved by exposing renal epithelial cells to 2mM sodium oxalate. The rationale behind this being that a salt of oxalate such as sodium oxalate, dissociates in liquid and in the presence of calcium ions (which are present in the body fluids or cell culture medium) forms insoluble calcium oxalate [ 48 ] which then can perpetuate the damage. According to the published research, addition of 1mM oxalate to cell-free media did not result in the formation of calcium oxalate crystals [ 28 ] and did not affect cell viability of renal cells [ 49 ]. Thus, this concentration was considered as the metastable limit [ 28 ] and it was further seen that in order to form CaOx crystals, the concentration of sodium oxalate required in DMEM [containing 1.8 mM Ca 2+ ] needed to be greater than 0.5 mM [ 50 ]. Hence in our study, the concentration of oxalate ions and the duration for which the renal cells were exposed to these ions was in accordance to results of earlier reports and akin to environment inside the kidney as it is known that the urinary oxalate concentration changes as it traverses the nephron and is 0.22 mM in the excreted urine of non-stone formers, 0.44 mM in mild hyperoxaluria and 1.5 mM in conditions of primary hyperoxaluria [ 51 ]. Till date, several in vitro studies have been conducted in which the renal epithelial cells have been exposed to varying concentrations ranging from 0.1–4 mM oxalate ions [ 28 , 49 , 52 , 53 ].
In the present study, we identified and characterized four anionic proteins, namely, Nuclear pore anchor (A1), DEAD Box ATP-dependent RNA helicase 45 (B1), Lon protease homolog 1 (B2) and Heat shock protein 90–3 (C1), from the dried bark of T . arjuna which have the capability to inhibit CaOx crystallization and crystal growth and also have anti-apoptotic potential, thereby leading to cell survival. The anionic nature of the proteins is in line with documented research which states that proteins which are capable of inhibiting CaOx crystallization are anionic.
Nuclear pore anchor (NUA), is a 237 kDa protein and is conserved across plants and animals. NUA is localized to the inner surface of the nuclear envelope and is a component of the nuclear pore complex (NPC), which is large multiprotein complex consisting of more than 20 proteins and acts as a channel for exchange of macromolecules between the cytoplasm and the nucleus [ 54 ]. Around 67 percent of all the oxalate which is taken up by cells ends up in the nuclei of the cells of the liver and kidney and 40 percent of this is localized in the nuclear membrane. The presence of an oxalate binding protein in the NPC lends support to our study and that of others who have reported an increase in oxalate binding in experimental hyperoxaluria [ 55 ]. It has been further reported that increased expression of oxalate binding nuclear proteins leads to more than 50% increase in oxalate binding, seen in experimentally induced hyperoxaluria [ 56 ]. The presence of Nuclear pore anchor (NUA) in the T . arjuna extract having the capability of binding to the oxalate would thereby protect the renal cells from injury. Our results add proof to this theory and we have shown that this protein (A1), reduced the injury and apoptosis caused by oxalate to renal epithelial cells (NRK-52E and MDCK). The mechanism by which it could be leading to the attenuation of damage could be attributed to its composition. The NUA protein has a lysine rich region which could be involved in oxalate binding, thereby inhibiting the formation of COM crystals and reducing the injury to renal cells, by preventing interaction of oxalate to calcium and renal epithelial cells. In addition to lysine rich region, the presence of acidic polyglutamic acid residues may bind to calcium ions and thus, prevent the adhesion of COM crystals to the epithelial cell surface. Since COM crystals adhere to negatively charged cell surface molecules these could be inhibited by GAG, nephrocalcin, uropontin and various macromolecular components containing polyglutamic acid, polyaspartic acid, as well as citrate [ 14 ].
The second protein having antilithiatic potential purified from the T . arjuna extract showed sequence similarity to DEAD Box ATP-dependent RNA helicase 45. DEAD-box proteins are the largest and most characterized family of RNA helicases and contain a core of ~ 400 amino acids comprising seven to nine conserved motifs. The name is derived from its motif II that includes the sequence D-E-A-D (Asp-Glu-Ala-Asp) [ 57 ]. The cells are exposed to various stimuli which impose stress that can impact cell survival. Data suggests that DEAD/H RNA helicases have the ability to form stress granules, which promote cell survival by coordinating the stress signals [ 58 ]. These stress granules act as antioxidants by activating G3BP1 (GTPase-activating protein SH3 domain binding protein 1) and USP10 (ubiquitin-specific protease 10) in response to stress and protect cells from ROS-induced apoptosis [ 59 ]. The cell survival function of stress granules could be associated with suppression of ROS production during stress [ 59 ]. Since this protein is rich in glutamic acid and aspartic acid amino acids, this protein could bind calcium ions and thus, prevent the adhesion of COM crystals to the epithelial cell surface [ 14 ]. Our results have pointed to the ability of this protein (B1) which we had isolated and purified from T . arjuna extract, to protect the renal epithelial cells exposed to oxalate injury as evidenced by the various assays which showed an increase in cell viability and decreased levels of apoptotic cell death.
The third protein identified from the T . arjuna extract was Lon protease homolog 1, which is an ATP dependent mitochondrial protease [ 60 ] involved in the degradation of damaged and oxidized proteins of the mitochondrial matrix [ 61 ]. A common motif seen in various proteins which inhibit CaOx crystallization is the presence of the Arg-Gly-Asp (RGD) motif, examples being, nephrocalcin, THP, OPN, etc [ 62 ]. RGD cell attachment sequence was originally identified in fibronectin that binds to fibronectin receptor, integrin α 5 β 1 [ 63 ]. It is known that the CaOx crystals bind to various anionic components present on the cell membrane. In addition, proteins such as nephrocalcin have the ability to bind the CaOx crystals at specific sites and thereby reduce their growth kinetics, as well as alter their morphology. Such a strategy could be followed by other proteins as well [ 14 , 64 ]. Therefore, it is very probable that the protein that we have identified, Lon protease homolog 1, may have a role similar to nephrocalcin. By binding via its RGD motif to the integrins present on the renal epithelial cells, this protein isolated from T . arjuna , could inhibit direct binding of CaOx crystals to cell surface and hence prevent the key step in stone formation i.e. crystal- cell interaction. Heparin, transforming growth factor-β2 (TGF-β2), and the tetrapeptide arginine-lycine-aspartic acid-serine (RGDS) negatively regulate the CaOx crystal endocytosis by interacting with cell adhesion sites on renal cell surface [ 38 ]. Studies by Lieske and colleagues have demonstrated diminution of crystal adhesion and internalization in BSC-1 cells upon treatment with RGDS, or fibronectin protein, containing this cell attachment sequence [ 38 ]. The degree of CaOx crystal deposition was inhibited by 60–80% in the cyclic RGD pretreated MDCK cells [ 65 ]. The presence of this protein could therefore act by inhibiting the binding of CaOx crystals to the renal cell surface and rescue cells from damage. In our present study we observed the cells treated with this protein indeed exhibited lower levels of apoptotic cell death owing to oxalate injury.
The fourth protein which we identified from the extract was, Heat shock protein 90–3, which is a molecular chaperone possessing antiapoptotic activity. It is a well-established fact that Heat-shock proteins (HSPs) are molecular chaperones which are induced by sub lethal cellular stresses, including temperature elevation, hypoxia and oxidative damage [ 66 ]. Hsp90 plays a key role in the refolding of denatured or damaged proteins and also in protein transport [ 67 ]. Hsp90 inhibits apoptosis and promotes cell survival by activating tumor necrosis factor-α which recruits receptor interacting protein (RIP) at the TNF receptor-1 to maintain NFκB activity [ 68 ]. Heat shock protein 90–3 also contains glutamic rich region which may aid in the binding of calcium ions and thus, prevent the adhesion of COM crystals to the epithelial cell surface [ 14 ]. In the light of these facts the attenuation of damage seen in the cells treated with this protein (B3) could be attributed to these functions.
The working hypothesis that was confirmed in this study is that, as per literature, oxalate and/or COM crystals induce oxidative stress that contributes to renal tubular epithelial cell injury, followed by death either by apoptosis and/or necrosis [ 69 , 70 ]. Most of the cells undergo apoptotic cell death and few cells die due to necrosis, and this may be in response to alterations in mitochondrial function that are characterized by initial increase in free-radical production, followed by a dissipation of the mitochondrial membrane potential and release of pro-apoptotic factors [ 71 ]. Oxalate damaged cells provide altered cell surface properties that act as a site for crystal adhesion [ 6 ] and cellular debris for crystal nucleation and growth [ 32 ]. In order to treat and prevent kidney stones, it is important to 1) target the interaction of free oxalate ions with the calcium ions and inhibit the formation of CaOx crystals and 2) avert the crystal retention to renal epithelial cells thus promoting cell survival.
Conclusion
In this study we identified 4 novel proteins namely, Nuclear pore anchor, DEAD Box ATP-dependent RNA helicase 45, Lon protease homolog 1 and Heat shock protein 90–3, as anionic inhibitors of CaOx crystallization from the bark of Terminalia arjuna . We have put forth evidence that these anionic proteins possessed antilithiatic activity in terms of inhibition of CaOx crystallization and crystal growth kinetics. Further, these proteins protected the renal epithelial cells NRK-52E and MDCK from oxalate induced injury. A putative mechanism of action could be attributed to the anionic nature since, these proteins contain either polyglutamic acid, polyaspartic acid, polylysine rich regions and/or RGD sequence and these may be the factors which contribute to their antilithiatic activity. In addition, some of these proteins have anti-apoptotic activity and stimulate cell survival by inhibiting the pro-apoptotic factors. Our study points to the therapeutic value of the proteins present in Terminalia arjuna which may play a vital role in inhibiting the CaOx crystallization and open up exciting avenues to study therapeutic proteins from plants for the treatment of urolithiasis.
Supporting Information
S1 Fig
Flow Chart depicting procedure followed for obtaining the most potent proteins from Terminalia arjuna in terms of inhibition of CaOx crystallization.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Cardiac glycosides (CGs) are a group of plant-derived compounds that have been used for many years in traditional medicine and that are currently used in treatment of cardiac failure and atrial fibrillation. In parallel to this use, CGs have also received attention as potential drugs in the treatment of various malignant diseases. Epidemiological observations suggest that patients on digitalis medication diagnosed with breast cancer in general present with lower proliferating tumours of smaller size, and subsequently better prognosis than control groups [1] , [2] , [3] , [4] , and that a high concentration of digitoxin could reduce the risk of developing leukaemia, lymphoma or urogenital cancer [5] . In vitro experiments have shown that CGs can induce cell death in several cell lines derived from solid cancers [6] , [7] as well as in leukaemic cell lines [8] , [9] , [10] , [11] .
In the myocardium, CGs bind reversibly to the α-subunit of Na + /K + ATPase, leading to a rise in intracellular sodium levels, which then results in an increase of calcium ions in the myocytes. The mechanism of the cytotoxic effects of CGs on tumour cells has been a subject of many studies but it largely remains unanswered. Binding to Na + /K + ATPase is not only a way of regulating ion pumps in the cell membrane but it can also activate several signalling pathways in the cell. For example, calcium-dependent activation of caspases and other hydrolytic enzymes [7] , [12] , [13] , generation of reactive oxygen species (ROS) [14] , topoisomerase inhibition [15] , interference with signal transduction pathways (e.g. Src-mediated phosphorylation of epidermal growth factor receptor (EGFR) and induction of the cell cycle inhibitor p21 Cip1 [16] have all been associated with the anti-tumour effects of CGs.
Digoxin-Like Immunoreactive Factors (DLIFs, also termed Digitalis-Like Compounds (DLCs)) are endogenous steroids identified in human tissues and they are identical or similar to plant and amphibian steroids. The DLIFs are believed to be synthesized in the adrenal gland, and to affect ion transport via Na + /K + ATPase. Binding of these compounds to Na + /K + ATPase may activate changes in intracellular Ca 2+ homeostasis and in specific gene expression [17] , and may be associated with the development of malignancies [18] . Interestingly, DLIFs selectively induce apoptosis in a human acute T-cell lymphoblastic leukaemia cell line but not in the myelogenous leukaemia cell line K-562 or healthy human peripheral blood mononuclear cells (PBMCs) [19] .
Despite years of interest in these effects and numerous studies in vitro and in animals, it has not yet been possible to utilize the anti-cancerous potential of CGs clinically. Recently, a very discouraging report on this issue was published, suggesting general inhibition of protein synthesis as the main mechanism of the anti-cancerous effects of CGs, and species differences of a magnitude sufficient to explain the results of most preclinical studies [20] .
During a routine screening programme carried out in vitro we observed that some samples of acute leukaemia were extremely sensitive to the cytotoxic effects of digitoxin, thus prompting further investigation. Hence this study was undertaken to categorize the activity of some CGs in primary cultures from patients with various leukaemic diagnoses, and to determine if general protein inhibition is the dominant mechanism of action, and if a therapeutic index in vitro exists.
Materials and Methods
Patient Samples and Cell Lines
Cryopreserved cells from bone marrow or peripheral blood from adult patients with B-precursor or T-acute lymphoblastic leukaemia (ALL), acute myeloid leukaemia (AML) and chronic lymphocytic leukaemia (CLL) were used in the study. Peripheral blood mononuclear cells (PBMCs) from healthy donors were used as controls.
Informed consent was obtained from all patients (verbal until 2006 and written thereafter) to save diagnostic samples in a biobank to be used for scientific research. The informed consent was verbal until 2006 (in accordance with the approval from the Ethics Committee) and the Ethics Committee did not demand that the consent should be documented at this time (until 2006). Since 2006 written consent has been obtained. Sampling for drug sensitivity testing was approved by the local Ethics Committee in Uppsala (Regionala etikprövningsnämnden i Uppsala, Sweden, approval number Dnr 21/93). The samples from the biobank used in the study were coded but labelled with diagnosis.
The T-lymphoblast-like cell line CEM/VBL 100 (CCRF-CEM) [21] was kindly donated by professor W.T. Beck, St. Jude's Children's Research Hospital, USA and the B-precursor Philadelphia-positive cell line SUP-B15 [22] was obtained from Deutsche Sammlung von Mikroorganismen und Zellkulturen GmbH, Germany. The cells were grown in complete medium (CCRF-CEM in RPMI 1640 and SUP-B15 in McCoy's 5A) and split twice weekly. The colorectal adenocarcinoma cell lines Hct116, HT29 and CC20 were used as comparator cell lines as regards effects on protein synthesis.
Chemicals and Reagents
The CGs digitoxin, digoxin and ouabain were purchased from Sigma (Sigma-Aldrich, Stockholm, Sweden). The compounds were dissolved in DMSO and further dilution was in PBS. The drugs were tested at five concentrations (ten-fold dilution steps) ranging from 100 µM to 0.01 µM or 10 µM to 0.001 µM. The maximum concentration of DMSO did not exceed 1% in the cell cultures. The serially diluted stock solutions were transferred to 384-well microtitre plates (NUNC, Roskilde, Denmark) and control wells were filled with PBS only (5 µl/well).
Measurement of Cytotoxic Activity
The cytotoxic activity of the CGs was measured by using a fluorometric microculture cytotoxicity assay (FMCA) as previously described [23] , [24] . The method is based on measurement of the fluorescence derived by hydrolysis of fluorescein diacetate (FDA) to fluorescein by cells with intact plasma membranes. Cell suspensions were seeded into drug-prepared 384-well microtitre plates. Wells with medium only served as blanks. The plates were incubated at 37°C for 72 hours; thereafter FMCAs were performed using an automated Optimized Robot for Chemical Analysis (Orca; Beckman Coulter Fullerton, CA) programmed through SAMI software (Beckman Coulter). The plates were washed in physiological buffer and FDA added. After 50 min of incubation at 37°C, fluorescence was measured at 485/520 nm using a Fluostar Optima microplate reader (BMG Technologies, Germany). The fluorescence measured is proportional to the number of living cells in each well.
Because a 30–40 nM plasma concentration of digitoxin can be maintained in patients for many days, additional experiments with extended incubation time (6 days) was performed. Leukaemic patient cells and the CCRF-CEM cell line was treated with therapeutically achievable concentrations (10, 30 and 50 nM) of digitoxin in 96-well microtiter plates. To simulate a situation with daily dosing, 50% of the medium (including drug to treated cells) was exchanged every day.
Quality control was evaluated for each test by demanding a signal/blank ratio of >10 and a coefficient of variation in controls and blanks of <30%. The proportion of leukaemic cells in primary cultures should exceed 70% on days zero or three when examined morphologically.
Measurements of protein and nucleic acid synthesis inhibition
Effects on DNA and protein synthesis were monitored in Cytostar-T® plates (available in the “ In Situ mRNA Cytostar-T® assay” kit, Amersham International, Buckinghamshire, UK) using 14 C-labelled thymidine and leucine. A Cytostar-T® plate is a 96-well microtitre plate with scintillants molded into the transparent polystyrene bottom. When labelled substrate is absorbed into the intracellular compartment of the cells at the bottom of the wells, the radioisotope is brought into proximity with the scintillant, thereby generating a detectable signal. Free radiolabelled substrate in the supernatant is unable to stimulate the scintillant [25] , [26] .
CCRF-CEM and SUP-B15 cells were suspended in fresh medium containing 14 C-thymidine (111 nCi/ml; for DNA experiments) or 14 C-leucine (222 nCi/ml; for protein experiments), yielding final radioactivity in the wells of 20 and 40 nCi, respectively. Cell suspension (50×10 3 cells in 180 µl) was added to each well; blank wells had isotope-containing medium only. Drugs (digoxin and digitoxin at final concentrations of 10 µM to 1 nM) and PBS in test and control wells were added in duplicate (20 µl per well) 2 hours after cell seeding, when the measured radioactivity in cell-containing wells was at least double compared with blank wells. Radioactivity was measured with a computer-controlled Wallac 1450 MicroBeta® trilux liquid scintillation counter (Wallac OY., Turku, Finland) immediately after addition of the cell suspension and at different time points up to 72 hours. Between measurements, the plates were stored in an incubator at 37°C. During measurement, the plates were covered with a plate sealer to inhibit microbiological contamination.
Results
The cytotoxic activities of digitoxin and ouabain were studied in primary leukaemic cells: T-ALL (n = digitoxin 4; ouabain 7), B-precursor ALL (n = 10; 6), AML (n = 11;11), CLL (n = 9; 6) and PBMCs (n = 4; 8) using the FMCA. Similar tests on the activity of digoxin, digitoxin and ouabain were performed in the leukaemia cell lines CCRF-CEM and SUP-B15. All tests were carried out in triplicate.
The primary T- and B-precursor ALL cells were significantly more sensitive to digitoxin than CLL cells (Mann–Whitney test, p = 0.02 and 0.006 respectively) and PBMCs (p = 0.02 and 0.005 respectively) ( Figure 1A ). The median IC 50 value regarding digitoxin was 0.07 µM for T-ALL and 0.06 µM for B-precursor ALL cells, compared with 0.44 µM for PBMCs ( Table 1 ). For the AML cells a scattered distribution regarding IC 50 values was observed. As regards ouabain, the IC 50 value was significantly lower (Mann–Whitney test, p = 0.02) for the T-ALL cells than for CLL cells but otherwise no significant differences were observed in the different leukaemic cells or the PBMCs ( Figure 1B and Table 1 ).
10.1371/journal.pone.0015718.g001
Figure 1
Cytotoxic IC 50 values.
Cytotoxic IC 50 values (µM) for digitoxin (A) and oubain (B) in primary cultures of human leukaemia cells (T-ALL, B-precursor ALL, AML and CLL) and PBMCs.
10.1371/journal.pone.0015718.t001 Table 1
IC 50 (µM) regarding the effects of digitoxin and ouabain on primary leukaemic cells and PBMCs.
IC 50%
T-ALL
B-precursor ALL
AML
CLL
PBMCs
Digitoxin
Median (range)
0.07 (0.02–0.09)
0.06 (0.02–0.23)
0.12 (0.01–1)
0.28 (0.03–0.92)
0.44 (0.36–0.68)
Ouabain
Median (range)
0.05 (0.004–0.08)
0.06 (0.05–0.34)
0.04 (0.03–10 * )
0.12 (0.06–0.28)
0.09 (0.03–0.19)
*Never reached IC 50%.
At 0.1 µM the B-precursor and T-ALL cells were significantly more sensitive to digitoxin than the CLL cells and PBMCs (Mann–Whitney test, B-precursor ALL vs. CLL p = 0.0003, B-precursor ALL vs. PBMCs p = 0.002, T-ALL vs. CLL, p = 0.004 and T-ALL vs. PBMCs, p = 0.03). In addition, the B-precursor ALL cells were more sensitive than the AML cells (p = 0.02) ( Figure 2A ). With extended exposure (6 days) both T- and B-precursor ALL cells appeared sensitive at clinically achievable concentrations ( Figure 2B ).
10.1371/journal.pone.0015718.g002
Figure 2
Cytotoxic effects of digitoxin.
A: Cytotoxic effects of digitoxin (0.1 µM) expressed as Survival Index % ( vs. untreated control cells) in the various leukaemia types (T-ALL, B precursor ALL, AML and CLL) and PBMC using the standard 72 h assay. B: Effects of therapeutically achievable digitoxin concentrations against leukaemic cells from patients and CCRF-CEM cell line over six days with daily medium change.
The cell line SUP-B15 was highly sensitive to all tested CGs, with IC 50 values as follows: digitoxin 0.002 µM, ouabain 0.004 µM and digoxin 0.03 µM. The T-lymphoblast-like cell line CCRF-CEM showed effects similar to those in the primary ALL cells, with IC 50 values of 0.04 µM for ouabain, 0.12 µM for digitoxin and 0.22 µM for digoxin. In both cell lines there was a tendency towards a lower sensitivity to digoxin than to the other two CGs.
Both digoxin and digitoxin inhibited DNA as well as protein synthesis in CCRF-CEM and SUP-B15 cells. The effects of the specific inhibitors aphidicolin and cycloheximide, used as positive controls, were strong and immediate. As presented in Figure 3 (digitoxin only; digoxin showed similar results), the effects were, however, only observed at relatively high concentrations, and at 100 nM, surprisingly, no significant effects were detected during the 24-h observation period. At the highest concentration tested (1.0 µM), well exceeding the IC 50 value for cytotoxicity, the effects on DNA and protein synthesis were similar in time and magnitude, i.e. the cells tended to “shut down” in an expected manner due to the toxic insult, presumably related to severe ionic imbalance. To put these results into perspective, a comparator cell line, the adenocarcinoma cell line Hct116was also analysed. As the Hct116 cell line is more tolerant to the cytotoxic effects of glycosides, slightly higher CG concentrations were used. Figure 4 shows the effects of various digitoxin concentrations on DNA (A) and protein synthesis (B). In contrast to the results in the leukaemia cell lines, protein synthesis in Hct116 cells was decreased more efficiently (i.e. at concentrations corresponding to cytotoxic activity) and at an earlier time point than DNA synthesis. With 1 µM digitoxin, protein synthesis in Hct116 cells was effectively inhibited at early time points (significant from 6 hours), while DNA synthesis did not appear to slow down until 24 h. This concentration is comparable to the cytotoxic IC 50 value for Hct116 cells measured in the FMCA (0.71 µM, Figure 4C ). The effects in two other colorectal adenocarcinoma cell lines (HT29 and CC20) were similar (not shown). The leukaemic SUP-B15 cell line was approximately 500 times more sensitive to the cytotoxic effects of digitoxin, with an IC 50 value of 1.5 nM ( Figure 4C ). Despite this, exposure of SUP-B15 cells to 10 nM digitoxin had no effect on protein synthesis up to 24 h ( Figures 3D and 4D ), but it effectively reduced viability at 72 h ( Figure 4D ). Results in the leukaemic CCRF-CEM cell line were similar (not shown).
10.1371/journal.pone.0015718.g003
Figure 3
Effects of digitoxin on DNA synthesis in leukaemic cell lines.
Effects of digitoxin exposure on DNA synthesis (i.e. thymidine incorporation) in CCRF-CEM (A) and SUP-B15 (B) cells. The DNA polymerase inhibitor aphidicolin (15 µM) was used as a positive control. Effects of digitoxin exposure on protein synthesis (i.e. leucine incorporation) in CCRF-CEM (C) and SUP-B15 (D) cells. The ribosomal inhibitor cycloheximide (36 µM) was used as a positive control.
10.1371/journal.pone.0015718.g004
Figure 4
Comparison between colorectal and leukaemic cell lines.
For comparative purposes the effects of digitoxin on DNA and protein synthesis (A and B respectively) were also monitored in the colorectal adenocarcinoma cell line Hct116. Figure 4C shows differences in cytotoxic activity of digitoxin in the Hct116 cell line and the leukaemic cell lines (CCRF-CEM and SUP-B15) used in the present study. In Figure 4D the effects on survival (at 72 h, measured by FMCA) and protein synthesis at 24 hours at digitoxin concentrations slightly exceeding the IC 50 values in HT29 and SUP-B15 cells are shown.
Discussion
In this study we have identified primary B-precursor and T-ALL cells as being particularly susceptible to the cytotoxic effects of CGs, being significantly more sensitive than CLL cells and PBMCs. Digitoxin was the most potent of the CGs tested, and the measured IC 50 value was comparable with therapeutic serum concentrations, especially if exposure time was extended to 6 days. The therapeutic concentration range for CGs in clinical use is narrow – for digitoxin 15–40 nmol/l (10–30 ng/ml) [27] . This, of course, raises the question of whether or not there could be a place for CGs in the treatment of acute leukaemia, despite negative results in solid tumours, for example. In ALL maintenance treatment with moderate doses of 6-mercaptopurin and methotrexate for a period of approximately two years has a documented effect on the risk of relapse [28] . The intensity of the treatment is limited by haematological and/or liver toxicity. New drugs with different toxicity spectra could clearly be of substantial benefit in this setting. Therefore, digitoxin as an addition to conventional maintenance chemotherapy in ALL could be proposed as a future clinical study, but only after careful preclinical investigation of possible drug-drug interactions. For example, previous studies have indicated that digitoxin at concentrations commonly found in the plasma of cardiac patients, significantly reduced etoposide and idarubicin-induced topoisomerase II cleavable complexes in K562 leukemia cells [29] .
The cytotoxic effect of digitoxin on both primary ALL cells as well as in the extremely sensitive cell line SUP-B15S indicates that a mechanism of inducing cell death other than inhibition of sodium- and potassium-activated adenosine triphosphatase (Na + /K + ATPase) may be possible. Numerous effects of cardiac glycosides in cancer cell lines ultimately leading to apoptosis have been demonstrated previously. Most of these effects are probably mediated through the main target enzyme [30] . However, interpretation of experiments involving established cell lines calls for caution because of possible genetic drift and altered properties. This is illustrated by the extreme sensitivity towards CGs in the cell line SUP-B15, while results in the CEM cell line were similar to those in the primary leukaemic cells.
Despite years of interest in these effects and numerous studies in vitro and in animals, it has not yet been possible to utilize the anti-cancer potential of CGs clinically. Several preclinical studies have demonstrated these compounds to be involved in selective control of human proliferation, which support the use of CGs to treat malignancies [30] , [31] . Different glycosides with different profile regarding cellular effects and cardiotoxicity have been developed, and several clinical trials have been performed. For example, UNBS-1450, a semisynthetic cardenolide, with good preclinical activity against NSCLC has entered a phase I clinical trial in Belgium [32] . A phase I trial of Anvirzel™, an aqueous extract from Nerium Oleander , in patients with refractory solid tumors has been reported [33] . Studies of the addition of digoxin to combination chemotherapy and immunotherapy in patients with advanced malignant melanoma have also been initiated [34] , the study has not been finally reported.
There are several possible explanations for this slow progression into clinical practise; one is of course the possibility of in vitro biases, and/or species differences. For example, it has been demonstrated that oleandrin activates MAPK and JNK and also induces expression of FasL, leading to apoptosis in human, but not in murine cells (28). This difference has also been detected at the cellular membrane level, as oleandrin altered its fluidity, inhibiting Na + /K + ATPase activity, and increasing intracellular free Ca 2+ levels, followed by calcineurin activity only in human, but not in murine cells. The results suggested that murine plasma membranes might be different from human membranes, which interact with oleandrin, disturbing the Na + /K + ATPase pump and resulting in calcification, followed by induction of Ca 2+ -dependent cellular responses such as apoptosis [35] . In line with these findings, Perne et al. recently published experimental results suggesting that general inhibition of protein synthesis is the main mechanism of the anti-cancerous effects of CGs in human cells, and that physiological species differences may explain the previously observed sensitivity of human cancer cells in mouse xenograft experiments [20] . It was proposed that inhibition of protein synthesis is directly related to effects on the Na + /K + ATPase pump. Indeed, this protein synthesis-inhibiting mechanism is supported by observations in the colorectal adenocarcinoma cell line Hct116 ( Figure 4 ). With 1 µM digitoxin, protein synthesis is effectively inhibited at early time points (significant from 6 hours), leading to decreased cell numbers at 72 hours. It might be concluded that Hct116 cells die because of an inability to synthesize vital proteins. The results from the leukaemia cell lines appear to be in sharp contrast to this – protein synthesis levels at time points up to 24 h are unaffected by concentrations having severe effects on cellular viability. Thus, it may be concluded that these cells stop synthesizing their proteins because they are dying, rather than the opposite.
Conclusions
Cardiac glycosides have been used for treatment of congestive heart failure and atrial fibrillation for over a century. During the last twenty years this class of compounds (e.g. digitoxin and oubain) has received much attention as regards their potential as anti-cancer agents, based on epidemiological as well as experimental findings. Despite great efforts the mechanism behind the anti-cancerous effects has been difficult to determine, as have the potential benefits in the clinic. The recently published study by Perne et al. [20] shed some light on this issue, but also strongly discouraged further clinical development of CGs, and derivatives thereof, for use in this therapeutic area. In this study we describe how some subsets of leukaemic cells from patients express high sensitivity towards CGs, particularly digitoxin, and at concentrations that may be achieved in the clinic. Furthermore, it was shown that the suggested mechanism of protein inhibition is valid in tumour cell lines with moderate or high resistance to the cytotoxic effects of CGs, but probably not in highly sensitive cancer cell lines. It is thus suggested that further investigation may be focused on diagnoses like T- and B-precursor ALL.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
The authors wish to add the following to the Acknowledgements section: "The authors acknowledge prior work on whole-cell micropipette aspiration and robotic deposition of single cells into microfabricated wells [1]. Anis et al. reported a robotic system for the pick-place positioning of single cells into silica wells, using a micromanipulation system. Based on previous work [1] [2], we developed a system and systematically investigated two modes of operation: whole- and partial-cell aspiration.
The positioning of a micropipette for whole-cell aspiration was automated in [1]. However, cell position inside the micropipette was not controlled after cell aspiration. Our experiments show that uncontrolled cell position inside the micropipette could cause cell deposition failure/repeatability because it is unknown when the cell would come out of the micropipette. Additionally, our experiments prove that the partial-cell aspiration technique is faster and achieves a higher success rate compared to the whole-cell aspiration technique.
We also demonstrate in this work the broad applicability of whole- and partial-cell aspiration techniques in enabling single cell culture on complex microfabricated surfaces, by positioning multiple cell types on several microfabricated cell culture substrates. Finally, we demonstrate the use of this serial manipulation method as an augmentative technology for existing parallel approaches to single cell positioning. This combination of technologies maintains the high-throughput positioning of cells obtained with other methods, while significantly improving their accuracy and specificity.
[1] Y. H. Anis, M. R. Holl, and D. R. Meldrum, "Automated Selection and Placement of Single Cells Using Vision-Based Feedback Controls," IEEE Transactions on Automation Science and Engineering, doi: 10.1109/TASE.2009.2035709.
[2] X. Y. Liu, Y. F. Wang, and Y. Sun, "Cell contour tracking and data synchronization for real-time, high-accuracy micropipette aspiration," IEEE Transactions on Automation Science and Engineering, Vol. 6, pp. 536-543, 2009."
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
It has long been known that endothelial cells are rounded and randomly oriented when cultured under static conditions, but become elongated and aligned with the flow when exposed to a unidirectional shear stress [1] . Non-reversing pulsatile shear has a greater elongatory effect than steady shear of the same mean magnitude, reversing pulsatile shear has a lesser effect and oscillatory shear has no effect at all [2] . Endothelial cell border and nuclear elongation have been used to assess local shear stresses occurring in vivo and hence to explore the relation between shear and susceptibility to atherosclerosis [3] , [4] . It has been proposed that more rounded cells, found in areas prone to plaque formation and predicted to be exposed to low shear stress, are dysfunctional and thus much has been done to investigate the physical properties of cells from both susceptible and protected regions of the vasculature and identify the differences between them. Initial studies focused on assessment of cell nuclear elongation and alignment in vivo [5] whilst later work went on to focus on cell morphology both in vivo [6] , [7] and in vitro [1] , [8] , [9] . An increasing body of work suggests that cell morphology change is driven by cytoskeletal rearrangement [10] , [11] , [12] though the full mechanisms behind the mechanotransduction of shear stress remain unclear.
One area of focus in the evaluation of cell structure is measurement of local cell mechanobiology. It is known that on exposure to shear stress endothelial cell movement is reduced whilst polymerisation of filamentous actin is increased and actin stress fibres form in alignment with the direction of flow. These stress fibres are thought to cause a fall in the compliance of the endothelial cell membrane [13] . Endothelial cell membrane compliance has been measured by a number of methods. Such methods include those exerting a pulling force such as use of optical tweezers, whereby a bead is trapped in position over a cell and manipulated to make contact with the cell and then repositioned pulling on the cell membrane, magnetic rheology, in which an integrin coated bead is bound to a cell and manipulated by magnets [14] and micropippete aspiration [13] whereby suction pressure is used to pull the cell surface up into a glass tube. Though these methods allow for application of a range of forces they are limited in their ability to study cells in their native context, often requiring cells to be in suspension and require direct and potentially damaging contact with the cell. Other methods such as atomic force microscopy where the cell is probed with a cantilever exert a force pushing into the cell [15] but again this level of contact could be potentially damaging giving rise to artefacts potentially influencing results. The majority of the above methods are also limited in their spatial or temporal abilities and can only record one feature at a time. With the development of Scanning Ion Conductance Microscopy (SICM), such restrictions can be avoided. By allowing scanning and mechanical testing with minimal invasiveness, SICM is ideal for the investigation of living cells, enabling accurate recording of cell properties in their native context and when grown in vitro [16] , [17] .
We recently characterised a model to subject endothelial cells to different patterns of shear stress [18] . By placing circular culture wells on an orbital shaker, cells at the edge of the well are exposed to shear that fluctuates in magnitude and has a preferred (tangential) orientation whilst cells in the centre of the well are exposed to shear of a uniform magnitude (with the same mean value as at the edge) that fluctuates between radial and tangential orientations. We showed that cells at the edge are more aligned than those at the centre [18] . Here we use this model in conjunction with SICM to investigate whether porcine aortic endothelial cells (PAEC) differ in elongation and compliance depending upon the shear profile they experience. We additionally investigated these properties in the inner and outer curvature of the ascending porcine aorta.
Materials and Methods
Isolation and culture of porcine aortic endothelial cells (PAEC)
Cells were isolated from the descending thoracic aortas of Landrace cross pigs (Fresh Tissue Supplies, East Sussex, UK) by collagenase digestion as described in the method of Bogle et al [19] . These aortas were collected in Hank's Balance Salt Solution (HBSS) supplemented with antibiotics (All Sigma UK), delivered within 24 hours of animal slaughter and stored at 4°C until use. All cells used in experiments were at passage 2.
Dissection of porcine aortas
Porcine hearts with attached aortas were obtained directly from the abattoir (Cheale Meats, Essex, UK) within 4 hours of animal slaughter in antibiotic supplemented HBSS. Aortas were inspected for damage before use. The aorta was first removed from the heart before excision of squares of tissue, approximately 2.5 cm×2.5 cm in area, from regions of interest defined by anatomical landmarks. The ‘outer curvature’ section was taken approximately 0.5 cm proximal to the origin of the bracheocephalic artery and the inner curvature from a region approximately opposite this ( Figure 1 ).
10.1371/journal.pone.0031228.g001
Figure 1
Image of a porcine heart and aorta illustrating the regions examined in this study.
A is a section of the outer curvature, and B the inner curvature, of the ascending proximal aorta.
The tissue samples were stripped of excess connective tissue and washed in PBS before they were pinned to wax patches in dishes containing L15 medium without phenol red (Invitrogen, UK) for SICM.
Scanning Ion conductance microscopy (SICM)
The SICM method has been described previously [17] . Briefly, the SICM probe consists of a glass nanopipette filled with electrolyte. The ion current will decrease the closer the pipette is brought to the sample surface until a predetermined level is reached and a feedback control stops the pipette from moving further. A plot of pipette position creates an image of the cell surface. To improve the reliability of the method we recently introduced hopping mode SICM [20] .
Monolayers of PAEC grown with or without shear stress were scanned by SICM (Ionscope). Micropipettes for scanning were pulled using a Laser puller (Sutter Instruments, P-2000); they had a diameter of 500 nm and a resistance of 25 MΩ. Topographic images were obtained by scanning an 80 µm×80 µm or 64 µm×64 µm area in hopping mode. Cultured cells and aortic segments were scanned for a maximum of 2 hours.
Compliance measurements
Using the principle described before [17] , SICM was used to determine mechanical properties of cells by subjecting them to locally produced dynamic pressure (see figure 2 ).
10.1371/journal.pone.0031228.g002
Figure 2
Schematic depiction of the SICM setup in a typical cell compliance experiment.
To calculate the compliance of single cells as result of local dynamic pressure, the resistance of the pipette (R), the induced pressure and the resulting displacement of the cell surface were determined. Calculations were performed using the following mathematical formula: where C = compliance d = displacement of the membrane (µm); P = pressure (kPa); R = actual pipette resistance (Ω); Ri = ideal pipette resistance (Ω).
The displacement measurement, d, was obtained at maximal applied P, which was usually in the range 30–40 kPa. Compliance measurements were calculated only where there was demonstrable recovery of the cell membrane to its original height.
Calculation of index of elongation and data inclusion
Measurements of elongation were obtained using ScanIC image (Ionscope, UK). Only complete cells with clear borders were included in the analysis. Images were excluded where the endothelial cell surface appeared damaged in any way. Length measurements were taken at the longest point of the cell and width values from the broadest part of the cell. Length was divided by width to obtain the index of elongation ( Figure 3 A , third image).
10.1371/journal.pone.0031228.g003
Figure 3
Representative SICM images from different parts of aorta and endothelial cells after shear stress.
(A) Representative topographical SICM images of 80×80 µm regions of the inner and outer curvature regions of the ascending pig aorta and of PAEC cultured under static (control) or sheared conditions. The greyscale represents the height of the cells. Edge = a region near the edge of the well thought to experience directional pulsatile shear and centre = a region near the centre of the well thought to experience non-directional steady shear (B) Distribution of IE under all conditions studied. (Upper graphs- inner and outer parts of aorta; Bottom graphs-cell culture-static centre, shear centre and shear edge).
Results
SICM images revealed that PAEC grown under static conditions and in the centre of a culture well under shear had polygonal morphology whereas PAEC from the edge of the sheared well had an elongated and aligned appearance. Endothelial cells lining the outer curvature of the ascending porcine aorta were also elongated and aligned with each other, whilst cells from the inner curvature were more rounded and had no tendency for alignment ( Figure 3A ).
Cell morphology was quantified using the Index of Elongation (IE), a calculated length-to-width ratio generated by ScanIC Image. Cells at the sheared well edges had an average IE of 2.02, exceeding the average IE of 1.4 for cells grown in the centres of the same wells. Cells of the outer curvature of the pig aorta had an average IE of 2.61, exceeding that of PAEC grown under all conditions ( Figure 3 ). The cells of the inner curvature of the aorta had an average ratio of 1.7. There was increased variability in alignment in cells from the native aorta as compared to cultured cells as might be expected based on differences between vessel geometries. There was a tendency for PAEC from the outer curvature to have a greater IE than PAEC from the inner curvature, many of which appeared only partially elongated or round ( Figure 3B ).
The SICM probe is a nanopipette that can produce a jet of buffer, and this allowed measurement of cell compliance by local application of a dynamic pressure ( Figure 2 ). The compliance of cells in the outer curvature of the ascending aorta (0.012±0.002 µm/kPa) was close to the compliance of PAEC cultured at the sheared well edge (0.016±0.003 µm/kPa). The compliance of cells in the inner curvature was higher (0.027±0.008 µm/kPa) and approached that of PAEC cultured in the shear well centre (0.038±0.006 µm/kPa). The compliance of cells from all regions of the well cultured under static conditions (0.035±0.006 µm/kPa) was nearly identical to those of PAEC cultured in the shear well centre ( Figure 4A ).
10.1371/journal.pone.0031228.g004
Figure 4
Compliance measurements from different parts of aorta and endothelial cells after shear stress.
(A); (B) Compliance of cells of the inner and outer curvatures of the ascending aorta and of cultured PAEC. Data were analysed by T-test. P≤0.05 was deemed significant and is indicated by *. (C) Relation between shape and compliance of cells in the aorta (linear correlation coefficient R = −0.56) and PAEC (linear correlation coefficient R = −0.66).
Figure 4B shows the relationship between index of elongation and compliance for endothelial cells from the in vitro shear stress model and the native aorta. For the range of compliances occupied by cells of both types, there was a remarkable agreement in the slope of the regression lines. Some compliance values measured from the least elongated cells in vitro substantially exceeded any compliances measured in the intact aorta; a large spread was seen in compliance values for these cells.
Discussion
Scanning ion conductance microscopy (SICM) was invented by Paul Hansma in the 1980s [21] and was later adopted to image and analyse the surface topography of live cells [20] , [22] , [23] . This was possible because SICM uses a nanopipette as a scanning probe which can be utilised to image cell surface structures with nanometre resolution without touching the cell [23] . Our laboratory further developed this system, introducing a plethora of associated methods [24] , [25] , [26] . Amongst other methods we developed SICM for measuring cell shape and volume. Recently, we have improved on the reliability and accuracy of SICM by introducing hopping probe ion conductance microscopy (HPICM) [20] . This allows us to obtain nanoscale resolution in highly convoluted live cell samples, such as whole tissue preparations e.g. the organ of Corti, without compromising the scan speed.
Endothelial cell elongation [2] and compliance [13] have both previously been shown to depend on applied shear stress in vitro. Here we used a recently developed cell culture model in conjunction with Scanning Ion Conductance Microscopy to measure elongation and compliance of individual cells exposed to pulsatile, oriented shear or non-pulsatile, non-oriented shear. We found a strong relation between the two cell properties with elongated cells showing reduced compliance. This in in agreement with previous studies carried out using AFM on cells that had experienced an acute application of shear stress [27] .
SICM also enabled us to make measurements of endothelial cell compliance from different regions of the aorta. We chose the inner and outer curvature of the ascending aorta; in mice these regions are atheroprone and atheroprotected, respectively [3] , and experience different patterns of shear [28] , although these properties may not hold for larger species. Differences between the regions of the aorta were similar to those found between the different culture conditions, and there was a similar relation between compliance and elongation. This new application of SICM presents exciting opportunities for future study.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Acute lower respiratory tract infections (ALRI) are responsible for the death of approximately 160,000 neonates and over 760,000 infants annually [ 1 ].Respiratory syncytial virus (RSV) and rhinovirus (RV) are the most frequent causes of acute respiratory infections in children [ 2 – 4 ].Primary RSV infections may lead to severe bronchiolitis and pneumonia [ 5 ]. RV usually causes rhinopharyngitis[ 6 ], but may also cause illness of the lower respiratory tract and asthma exacerbations, and has been associated with severe ALRI in young children [ 7 , 8 ]. A few studies have investigated the viral profile in children attending the pediatric emergency room (ER) with ALRI and its association with disease severity [ 9 – 12 ], but finding vary. Moreover, the role of viral co-infections in illness severity has been controversial [ 10 , 12 , 13 – 15 ]. In young children who presented to an outpatient clinic or ER in Japan with acute respiratory illness, RSV, RV, parainfluenza viruses (PIV) and human metapneumovirus (HMPV) were the most prevalent viruses. Similarly to what was found in a previous study in Brazil [ 4 ], detection of RSV, alone or in co-detections, was associated with increased disease severity in children in Japan [ 9 ]. In Seattle, USA, RV was the most frequent virus detected in children less than three years old presenting to the ER with ALRI, and 52% of them required hospital admission [ 10 ]. In addition, RV viral load and co-detection with RSV was associated with more severe disease in that study [ 10 ]. In children younger than two years old who presented to a pediatric ER with ALRI in Malaysia, RSV and RV were the most frequently detected viruses and, although RSV was associated with a history of wheezing, virus detection was not associated with need for hospitalization [ 11 ]. In The Netherlands, the most frequent viruses detected in children presenting to the ER or outpatient clinic with ALRI were RSV, RV and human coronavirus (HCoV), and RSV detection correlated with longer duration of oxygen therapy [ 12 ]. However, there was no association between the virus species and hospital length of stay, and virus co-detections were not associated with disease outcome [ 12 ].In the present study, a comprehensive polymerase-chain-reaction(PCR) panel of primers and probes was used to detect respiratory viruses, enabling virus identification to the species level, in respiratory samples collected from children younger than three years with ALRI seen at the pediatric ER and subsequently admitted to hospital. Association of different respiratory viruses with disease severity, defined as need for admission to the pediatric intensive care unit (PICU), was assessed.
Material and methods
Patients and samples
This was a prospective cohort study conducted in a tertiary-care university hospital in Brazil. The study was approved by the Research Ethics Committee of the Clinical Hospital, School of Medicine of Ribeirão Preto, University of São Paulo, a tertiary-care facility in the city of Ribeirão Preto, state of Sao Paulo, Brazil. A written informed consent was obtained from parents/guardians (protocol number HCRP A07-020). All consecutive children younger than three years attending the pediatric emergency room with ALRI who were admitted to hospital from June 1 st , 2008 to May 31 st , 2009were eligible for the study. ALRI was defined by the presence of cough, tachypnea, respiratory distress with prolonged expiratory time, and wheezing or crackles on auscultation. Patients with a diagnosis of bacterial pneumonia as indicated by clinical presentation and chest X-ray findings or a positive blood culture were excluded from analysis.
Demographic, clinical and outcome data were collected from patients’ health records. Need for PICU admission was considered the main indicator of disease severity. Nasopharyngeal aspirates were collected from patients within the first 48 hours of hospitalization, as previously described [ 16 ] and they were split into aliquots: two 250μL aliquots mixed with 750μL of TRIzol (Invitrogen, Thermo Fisher Scientific, MA, USA) to DNA/RNA extraction and two 500μL backup aliquots mixed 1:1 with 500μL of viral transport medium, which consists of minimal essential medium with Eagle’s salts plus 20% fetal bovine serum, 15% glycerol and 1% antibiotic-antimycotic solution (GIBco, Thermo). All backup aliquots were stored at -70°C until analysis.
Rapid screening for RSV
Nasopharyngeal aspirates were routinely screened for RSV, by eithera rapid chromatographic immunoassay for RSV antigen (Directigen EZ RSV Test, Becton Dickinson and Company, Franklin Lakes, NJ, USA) or indirect immunofluorescence (IF) assay. Directigen EZ RSV was performed on 250μL of nasopharyngeal aspirates following the manufacturer’s protocol. Indirect IF assay was performed with RSV-specific monoclonal antibody (MAb 858–4; Millipore, MA, USA) diluted 1:100 in phosphate-buffered saline (GIBco, Thermo), revealed with Alexa Fluor 488-labeled donkey anti-mouse IgG (Life Technologies, Carlsbad, CA, USA) diluted 1:200 in phosphate-buffered saline. Slide preparation and IF protocols are available at http://dx.doi.org/10.17504/protocols.io.w8rfhv6 and http://dx.doi.org/10.17504/protocols.io.w8ufhww .
Detection of RNA and DNA respiratory viruses by real-time PCR
TRIzol aliquots from each sample were used to do DNA and RNA extractions. Total RNA was extracted following manufacturer’s protocol with some adaptations( http://dx.doi.org/10.17504/protocols.io.w8vfhw6 ), and DNA-enriched fractions were used for DNA purification using the Wizard Genomic DNA purification kit (Promega, Madison, WI, USA) as per manufacturer’s protocol. One microgram of total RNA was used in a reverse transcription reaction carried out with “Multiscribe reverse transcriptase” (Applied Biosystems, Thermo), primed with random hexamers and following manufacturer’s protocol. This random cDNA was used in TaqMan real-time PCR assays to detect RSV-A and B; HMPV A and B; human influenza viruses (FLU) A and B; and PIV 1 and 3. Human beta-actin housekeeping gene was used as internal controls in all assays. All PCR assays were performed using only one set of primers-probe per reaction, except for RSV and HMPV detection, which were tested in duplex format. Human bocavirus (HBoV) and human adenovirus (HAdV) were detected by qPCR in a single format assay ( http://dx.doi.org/10.17504/protocols.io.w8yfhxw ). HCoV detection was performed by a nested RT-PCR method that uses primers targeting the RNA-dependent RNA polymerasegene [ 17 ]. Rhinovirus (RV) detection wasperformed by a two-step PCR method ( http://dx.doi.org/10.17504/protocols.io.w83fhyn ) [ 18 , 19 ].
All PCR reactions were performed on a Thermocycler 7300 (Applied Biosystems, Thermo), using published primers and probes sequences [ 20 ]. All real-time PCR plates included appropriate negative controls matched to every step of the testing of nasopharyngeal aspirates. Negative controls were total RNA or DNA extracted from uninfected Hela cells and ultrapure water treated in the same way as clinical samples.
Statistical analysis
The analysis was made using SAS 9.4 (SAS/STAT User’s Guide, Version 9.4, Cary, NC: SAS Institute Inc., 2013). Data were expressed as median (range) or number (%). Patients were grouped according to the need for PICU admission. Continuous variables between groups were compared by Mann-Whitney U test and categorical variables, by Fisher’s exact test. Simple and multiple log-binomial regression models were constructed to assess associations of virus type with need for PICU admission. Relative risks (RR) and 95% confidence intervals (95%CI) were obtained after adjusting log-binomial regression models. Initially, simple log-binomial regression models were fitted, resulting in crude relative risks. Subsequently, the adjustment of multiple log-binomial regression models considering age, prematurity, the presence of an underlying disease and congenital heart disease as covariates, resulted in adjusted relative risks [ 21 ]. A 5% significance level was considered in all analysis.
Results
Over the study period, nasopharyngeal aspirates were collected from 279 children seen at the pediatric emergency room with ALRI. Thestudy population comprised 236 patients who had at least one respiratory virus detected in respiratory specimens. All patients were hospitalized. Twenty-five percent of patients (n = 60) had comorbidities. The most common underlying diseases were neurological impairment (n = 17), congenital anomalies (n = 13) and chronic lung disease (n = 11). Forty-seven patients (19.9%) were admitted to the PICU; 25 of them (53.2%) received invasive mechanical ventilation for a median time of 7 days (range 1 to 99 days). Length of PICU stay ranged from 1 to 254 days (median 9.5 days). Demographic data were not significantly different between patients admitted to the PICU compared with those who did not need PICU admission. However, use of systemic antibiotics and the presence of underlying diseases and congenital heart disease were more frequent in patients admitted to the PICU, and they also had a longer hospital length of stay, Table 1 .
10.1371/journal.pone.0217744.t001
Table 1 Demographic and clinical data.
Characteristic
All (n = 236)
PICU admission (n = 47)
No PICU admission (n = 189)
Age (months)
5.2 (0.2–35)
3.6 (0.2–35)
5.9 (0.3–34)
Age < 6 months
125 (53)
29 (61.7)
96 (50.8)
Weight (kg)
6 (2–21)
5.2 (2.1–14)
6.2 (2–21)
Male gender
89 (38)
18 (38.3)
71 (37.6)
Prematurity
43 (18.2)
7 (14.9)
36 (19)
Underlying disease
60 (25)
21 (44.7)
39 (20.6) *
Congenital heart disease
23 (9.7)
9 (19)
14 (7.4) *
Use of systemic antibiotics
153 (64.8)
37 (78.7)
116 (61.4) *
Length of hospital stay (days)
8 (1–254)
18 (3–254)
7 (1–210) *
Data are expressed as median (range) or n (%). PICU, pediatric intensive care unit.
*P < 0.05 for comparison between PICU admission and No PICU admission groups
The most frequently detected virus was RV (85.6%), followed by RSV (59.8%), HBoV (23.7%), HMPV (17.8%), HCoV (11.4%), HAdV (10.6%), PIV (10.2%) and FLU (8.5%). Co-detections were found in 182 (78%) patients, S1 Table .
The results of multiple log-binomial regression analyses showed thatthe detection of HCoV alone (adjusted relative risk (RR) 2.18; 95% CI 1.15–4.15) or in co-infection with RV-C (adjusted RR 2.37; 95% CI 1.23–4.58) was independently associated with PICU admission, S2 Table .
Eight patients (3.4%) died. Their median age was 3.5 months (range 2.3–9.3 months); six (75%) were female. Six patients (75%) who died had comorbidities:hydrocephalus (n = 3),Pompe disease (n = 1), Down syndrome (n = 1) and pulmonary hypoplasia (n = 1). Three patients had a single virus type detected in their respiratory samples (RV-C in two patients, and RV-A in one patient), two patients had dual viral co-detection (RV-C + RSV-B and RV-C + HCoV OC43), and three patients had triple viral co-detection (RSV-A + FLU-A + HBoV; RV-C + RSV-A + HMPV-A; and RV-C + HMPV-A + FLU-B). The causes of death were respiratory insufficiency (n = 4), cardiogenic shock (n = 2) and septic shock (n = 2).Of note, the patient who had the longest hospital stay (254 days) and also the longest duration of mechanical ventilation (99 days) had pulmonary hypoplasia, a severe underlying condition, and he ultimately died of sepsis.
Discussion
In this study, viruses were detected in 85% of young children attending the emergency room with ALRI and subsequently admitted to the hospital. RV, especially RV-C, was the most frequently detected virus, followed by RSV. Twenty percent of patients were admitted to the PICU. Comorbidities and congenital heart disease were more frequent in patients who were admitted to the PICU. In addition, the detection of HCoV alone or in co-infection with RV-C was independently associated with PICU admission.
Similarly to our data, RV was also the most commonly detected virus in children younger than 3 years presenting to the Seattle Children’s Hospital pediatric ER with a symptomatic respiratory tract infection, and the majority of them had lower respiratory tract infection and required hospitalization [ 10 ]. RV was also the most prevalent virus detected in children aged two weeks to 5 years admitted to a hospital with ALRI in Germany, exceeding the frequency of RSV [ 22 ]. Moreover, RV-C has been reported as the most frequent RV species and has been associated with severe disease in children less than 3 years old [ 8 , 22 ].
Viral co-detection was found in approximately three-quarters of our study population, with RV/RSV as the most frequent duo. However, we observed that viral co-detection per se was not a risk factor for PICU admission. In keeping with this, a recent systematic review and meta-analysis showed that respiratory viral coinfection was not associated with need for hospitalization, intensive care admission or length of stay in children [ 15 ]. In addition, the number of detected viruses has not been associated with illness severity [ 14 ]. Indeed, surprisingly, an inverse correlation between the clinical severity score and the number of viruses detected has been observed [ 23 ]. Nevertheless, a previous study performed in Southeast Brazil showed that coinfections, especially involving RSV, were associated with increased severity [ 13 ]. In the present study, RSV in co-detection with other viruses was not associated with increased disease severity.
HCoV was detected in 11% of children admitted to hospital with ALRI in our study, with a predominance of OC43 (40.7%) and 229E (33%)types. Although human coronaviruses most frequently cause common colds, they may cause severe respiratory diseases, such as severe acute respiratory syndrome [ 24 ]. We observed that one-third of children with HCoV infectionand 41% of patients withHCoV in co-infection withRV-C were admitted to the PICU. Furthermore, we found that the detection of HCoV alone or in co-infection with RV-C was an independent risk factor for PICU admission.Similar to our results, HCoV was detected in 8.2% of hospitalized children aged 3.2 ± 3.9 years with respiratory tract infection in New York, OC43 was the most prevalent type (40.1%), and 11% of patients with HCoVinfectionneeded PICU admission. Additionally, the presence of chronic complex underlying conditions, including cardiovascular, genetic and respiratory diseases was associated with increased disease severity [ 25 ], which is corroborated by our data.
Almost two-thirds of our children were treated with systemic antibiotics. This high frequency of antibiotic use is similar to that reported in developed countries [ 26 , 27 ]. Empiric antimicrobial therapy, driven by unfavorable clinical conditions in the face of probable bacterial infections, is likely to be initiated in the absence of immediate laboratory confirmation of virus detection. In the present study, the only rapid diagnostic test for respiratory virus availableearly at hospital admission was the RSV rapid antigen detection test. Therefore, the development of clinically relevant rapid tests for respiratory viruses is needed and shall help antimicrobialstewardship programs.
Since the present study was done in a single health center, this raises concern that epidemiologic data may not be generalizable. While this limitation is acknowledged, the results obtained are similar to those reported from Europe and North America [ 8 , 22 ].Another concern could be the assumption of disease etiology based on virus detection atone-time point. However, this is a practical diagnostic approach worldwide and all patients enrolled in the study had signs and symptoms of ALRI requiring hospitalization concomitantly with virus detection. It should also be noted that, although the diagnosis of bacterial pneumonia, as indicated by clinical presentation, chest X-ray findings or positive blood culture, was an exclusion criterion, it is difficult to discriminate viral and bacterial pneumonia in cases with negative bacterial cultures, because clinical presentation and chest radiography findings may overlap.
In conclusion, the detection of HCoV alone or in co-infection with RV-C was independently associated with PICU admission in young children hospitalized for ALRI.Rapid and reliable diagnostic tests for respiratory viruses associated with disease severity should be widely available to improve patient management and optimize healthcare resources.
Supporting information
S1 Table
Viruses detected in respiratory specimens of all patients (n = 236).
(DOCX)
S2 Table
Viruses detected in respiratory samples and risk for admission to the pediatric intensive care unit (PICU).
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Temperature drives development of ectotherms. Within an ecologically relevant thermal range, the entire development from egg to adult is faster at higher temperatures because they enhance the development rate through increased metabolic rate [ 1 ]. Moreover, previous studies found that temperature does not affect the proportion of time spent in any given non-diapausing developmental stage, at least within the thermal range for which the rate of development increases linearly with temperature ( Fig 1a ). This concept is called ‘equiproportional development’ (EPD) in the copepod literature, where it was first described by Corket and McLaren [ 2 ] and formalized by Corket [ 3 ], and ‘developmental rate isomorphy’ (DRI) in the insect literature, where it was first proposed by van Rijn et al . [ 4 ] and formalized by Jarošík et al . [ 5 ]. Although EPD has a longer history, considers a wider thermal range beyond the near linear part of the response in development rate and has been extended to situations when food is limiting [ 6 – 8 ], we primarily use the term DRI as our focus in this paper is on temperature-dependent development in insects under conditions of food satiation.
10.1371/journal.pone.0129341.g001
Fig 1
Possible mechanisms leading to DRI violation.
Left column shows development rates of three hypothetical life stages, early (blue dotted line), intermediate (orange dashed line) and late (green line), and the overall rate of development (thick grey); grey rectangles highlight the thermal range ( T 1 , T 2 ) in which the overall rate of development increases almost linearly with temperature; time units are omitted. Right column shows the corresponding relative duration of each stage in the ( T 1 , T 2 ) range. (a) DRI; (b) violation of DRI due to shifted, stage-dependent temperature optima; (c) violation of DRI due to progressively limiting, temperature-dependent metabolic scope for growth in individual stages; (d) violation of DRI due to limiting metabolic scope for growth in the intermediate stage. Even more severe DRI violation outside the ( T 1 , T 2 ) range in panels (b)–(d) is omitted for clarity. See text for further details.
DRI has important consequences for the effect of temperature on individual development. For example, DRI implies that lower developmental threshold (LDT) temperature, at which the development should be completely arrested, is equal for all stages in a species’ ontogeny [ 5 , 9 , 10 ]. DRI appears to be common in insects and mites: in the largest study to date, Jarošík et al . [ 5 ] found DRI in approximately two thirds of more than 420 populations belonging to almost 350 species, and reported that data from most remaining populations violated the DRI concept only to a small extent. Additional studies suggested that relative duration of a particular developmental stage may be a fundamental life history invariant that is independent not only of temperature, but also of host plants, geographical origin of the population, and other factors [ 5 , 9 ], and reported that DRI prevails in other vertebrate and invertebrate ectotherms beyond arthropods [ 11 ]. This contrasts with the lack of a clear pattern in temperature dependence of the ratio of total copepodid to total naupliar duration in papers dealing with EPD in copepods [ 6 ].
Widespread validity of DRI would have important consequences for pest monitoring and forecasting [ 11 ] as well as for the presumed effects of climate change on populations. If all ectotherms more or less complied with DRI, models predicting the impact of climate change on ontogeny and biotic interactions could be simplified because all individuals of a given species would respond similarly to changes in temperature, at least when other factors such as food availability would not constrain their development and growth rates. However, as different species respond markedly differently to temperature [ 12 ], it is equally plausible that similar differences may arise during ontogeny and lead to DRI violation.
So how strong is the evidence for the ubiquity of DRI? Analysis of marine pelagic copepod data based on a linear mixed-effect model found that development rates of eggs, nauplii and copepodites scale differently with temperature when all data are pooled and species, study, sex, and stage are treated as random effects [ 8 ]. This study has provided a strong indirect support to the idea that DRI may not hold in some groups. All other statistical methods currently used to evaluate DRI are based on analyses of data for individual species. They have their strengths but also potentially serious caveats discussed below.
Most importantly, proportions of time spent in different instars represent compositional data, for which dedicated statistical methods exist [ 13 ] but have not been used in DRI or EPD studies. To address this issue, we introduce a new method of DRI analysis that takes into account the inherent proportional structure of the data. The underlying Dirichlet distribution is a continuous multivariate probability distribution generalizing the beta distribution for n ≥ 2 variables. The so-called common parameterization of a Dirichlet distribution consists of a vector α = ( α 1 , …, α n ) of positive real numbers, where α 0 = ∑ i = 1 n α i provides a ‘precision’ parameter (higher α 0 means less variation around the expected value) and α i / α 0 specifies the mean value of the i -th variable, such as relative duration of the i -th developmental stage; see [ 14 , 15 ] for details. Dirichlet regression directly evaluates DRI for individual-level data coming from a Dirichlet distribution: it is analogous to linear regression because it estimates the link between any explanatory variables such as temperature and the common parameterization of the distribution.
As a case study, we apply Dirichlet regression to individually resolved data on temperature-dependent development of four terrestrial and five (semi)aquatic insect species. Although a full review of possible causes of DRI violation is outside the scope of our paper, we argue that mechanistic insights [ 8 ] can challenge the presumed widespread validity of DRI with alternative testable hypotheses. Here we use the analyses to propose and examine two previously untested mechanisms that could explain the violation of DRI in non-extreme situations, i.e., the thermal range within which the total development rate scales linearly with temperature. This approach is conservative and allows for direct comparison with previous DRI studies, although it could neglect informative DRI violations at suboptimal thermal conditions.
The first such mechanism derives from classical life history theory [ 16 – 18 ]: in temperate regions, different instars of univoltine species may encounter predictably seasonal environment with different temperatures favouring life histories with non-isomorphic rates of development. More precisely, relatively slower development of late developmental stages at lower temperatures could be adaptive if thermal preference curves of individual instars could evolve to track the seasonal trends in temperature by shifting their maxima ( Fig 1b ). The second mechanism follows recent advances in metabolic ecology: with increasing temperatures, larger individuals can have a progressively narrowing scope for growth due to different allometric scaling of energy intake and expenditure [ 19 , 20 ]. Consequently, larger individuals would develop relatively slowly at higher temperatures. This could be either the latest developmental stage if all stages are actively feeding ( Fig 1c ), or the intermediate stage if the data are resolved into eggs, larvae and pupae ( Fig 1d ), because the pupal stage does not feed and draws energy from the reserves accumulated by the larva. Last but not least, the amount of dissolved oxygen declines strongly with water temperature. Stage-specific development rates of aquatic and terrestrial ectotherms may thus respond differently to temperature, similar to the differences in intraspecific patterns of temperature-size responses and latitude-size clones [ 21 ].
Overview of DRI Analyses
Currently used methods: their strengths and caveats
Five statistical methods have been used in DRI analyses of individual species data [ 10 ]. They either examine the relationship between developmental stage and LDT temperature or ask whether the proportions of time spent in each developmental stage depend on temperature. However, each of them might suffer from potential statistical artifacts.
Two previously used methods rely on the fact that DRI holds if and only if the LDTs of all studied developmental stages are equal. One of them (Method 1 in [ 10 ]) estimates mean and standard error of LDT values from the presumed linear relationship between temperature T and the development rate r (time -1 ) in the given stage,
r = a + b T
(1)
LDT is estimated from eq (1) by the intercept (LDT = – a / b ) of the regression line of development rate on temperature; [ 22 ] gave an approximate formula for the standard error of the LDT estimate, but a rigorous procedure to test for equality among two or more LDT values in different developmental stages is not available [ 10 ]. Another method (Method 2 in [ 10 ]) therefore multiplies eq (1) by d / b , where d = r -1 is the developmental time in a given stage [ 23 ]. Rearranging the terms leads to an estimate of the slope of a new linear regression,
d T = SET + LDT d
(2)
which links developmental time d to the sum of degree-days d T. The intercept SET = 1/ b represents the sum of effective temperatures, equal to the number of day-degrees above the LDT required to complete a particular developmental stage. ANCOVA is used to test if the slopes of the relationship, i.e. the LDT values, are equal across all developmental stages [ 10 ]. However, the (random) values of d introduce a hidden correlation between the explanatory and explained variable when individual-level data are used (see Fig 1 in [ 10 ] for an example) and the explanatory variable ( d ) no longer corresponds to a directly manipulated variable, i.e. ANCOVA assumptions are violated [ 24 ].
Shi et al . [ 25 ] developed another method (Method 5 in [ 10 ]) based on the rotation of regression lines to test the independence of LDT on temperature with a Chow test [ 26 ]. This is sufficient for two regression lines. To compare LDTs of more than two developmental stages with this method, Kuang et al . [ 10 ] proposed multiple pairwise comparisons of all regression lines. Increased risk of type I error in such analyses, neglected in [ 10 ], can be amended with procedures controlling the family-wise error rate (such as Bonferroni or Holm correction) or the false discovery rate, but a full consensus on their use among ecologists seems to be lacking [ 27 ].
The most widely used DRI test was introduced in [ 5 ]. It uses angular transformation of the proportional data ( arcsin p i , where p i is the relative time spent in stage i ). Linear regression of these transformed data against temperature (Method 3 in [ 10 ]) or ANOVA (Method 4 in [ 10 ]) is used to assess if relative developmental time of each stage changes with temperature. Factors such as sex, photoperiod, food quality or geographic origin can be included as additional explanatory variables in the linear regression or an ANCOVA analysis [ 5 ]. This method is usually applied to mean values of p i reported in experiments because many papers, especially older ones, do not report individual-level data. Kuang et al . [ 10 ] suggested that individual-level values of p i can be transformed and assessed in a similar way as the mean values. This approach would, however, leave out the individual identity, i.e. the fact that the sum of p i , j equals 1 for each individual j .
However, the use of angular transformation in the analysis of ecological data is now considered outdated. In addition to the general objections raised in [ 28 ], Methods 3 and 4 often suffer from another previously unreported flaw. The values of p usually fall in the (0.2, 0.8) range, in which arcsin p is nearly linear ( arcsin p ≈ p + 0.2854 ) and, as the sum of all relative durations of k developmental stages equals 1, the sum of transformed proportions remains close to 1+0.285 k . This issue is particularly relevant for datasets with 2–4 developmental stages of which none dominates in duration. Such transformed data then do not pass the assumption of independent data required for linear regression, ANOVA and ANCOVA. Moreover, results in [ 5 ] have been inadvertently loaded with two additional issues that might introduce bias both in favour of or against DRI ( S1 Text ). The net result of these biases cannot be assessed without a detailed reanalysis of all published data on DRI, which is beyond the scope of this study.
Last but not least, Method 3 used for population mean values in is not suitable for data obtained at two temperatures. Most older studies commonly reported only mean developmental times of each stage at each temperature (i.e. n = 1 for each developmental stage and temperature). It is not possible to assess DRI from such data at two temperatures because linear regression and ANOVA yield perfect fit unless there are replicated measurements across another factor (see [ 5 ] for examples). Datasets with two temperatures have therefore been out of reach for DRI studies. However, if individual-level data are available and there is compelling evidence that the two experimental temperatures lie within the range in which the development rate scales linearly with temperature, there is no reason to abandon an appropriate method such as Dirichlet regression. Test of DRI then simply evaluates if a Dirichlet distribution with temperature-independent parameters fits data measured at the two temperatures.
The above comparison leads us to the proposition that Dirichlet regression is currently the most appropriate and versatile method to analyse DRI data. Its main limitation is the need of individually resolved data, unavailable for many older studies. However, the availability of underlying ecological data has improved over the past decade owing to more widespread data sharing and appearance of new sharing infrastructure and tools [ 29 ], making it more and more possible and desirable to replace the standard publication of summary data and statistics by more detailed, individual-level data in new studies. The method developed here does not a priori require large datasets and can be used to simultaneously compare models based on various assumptions concerning the effects of temperature and other factors on individual ontogeny. Using data on nine individually reared insect species as a case study, we show that some of the previous conclusions may have been burdened by statistical artifacts. This suggests that the entire concept of DRI should be critically re-evaluated.
Case study: Dirichlet regression
We apply Dirichlet regression to five species of aquatic and semiaquatic insects ( Cloeon dipterum (Linnaeus, 1761), Microvelia reticulata (Burmeister, 1835), Velia caprai (Tamanini, 1947), Notonecta glauca (Linnaeus, 1758), and Acilius canaliculatus (Nicolai, 1822)) and four species of terrestrial insects ( Amara communis (Panzer, 1797), Gastrophysa viridula (De Geer, 1775), Leptinotarsa decemlineata Say, 1824, and Loxostege sticticalis (Linnaeus, 1761)) for which we had individually resolved developmental data. We refer to all species by their generic name in the following text and Appendices. Among the (semi)aquatic species, only Cloeon is truly aquatic in the sense of relying on dissolved oxygen for respiration; all developmental stages of all other species included in this study breathe atmospheric oxygen. All species were reared individually or in groups in the laboratory at 2–6 different temperatures ( S1 Table ). Larvae of all species were fed ad libitum on a daily basis (see S2 Text for details). No permits were required to collect the individuals in the field and carry out the experiments in agreement with relevant national legislations.
Further details of the experimental protocols varied among species due to their different life histories and environmental requirements. In brief, field-collected adults of terrestrial species were kept in glass or plastic vials and checked for eggs once or twice a day. Overwintered females of Microvelia and Notonecta laid eggs in the laboratory aquaria, which were randomly placed into one of the experimental temperatures but egg developmental time was not monitored. Attempts to obtain sufficiently large numbers of eggs in the laboratory failed for Velia , Acilius and Cloeon , and early instar larvae collected in the field were instead used in the experiments (see S2 Text for details).
Developmental time (in days) of all larval instars was recorded for Microvelia (L1–L5, i.e., all feeding stages); data on L2–L5 for Velia and Notonecta (i.e., a subset of feeding stages), and data on L2, L3 and pupa for Acilius , in which L1–L3 are the feeding stages and pupa is non-feeding. Because Cloeon like other mayflies does not have a fixed number of preimaginal developmental stages, we divided its development into a stage containing 1–5 instars from the start of the experiment to the moult into the pre-final instar (hereafter called “early stage”), the pre-final instar, and the final instar before the subimago emerged (see S2 Text for our reasoning supporting this decision); all these stages are feeding. In experiments on terrestrial species, individuals were checked daily or twice a day depending on the species and experimental temperature. Only the hatching of larvae from eggs, pupation and emergence of adults were monitored. We thus collected data on the duration of the egg, larval and pupal stage; only the larval stage is feeding. Because all our data are individually resolved (see S2 Table for the raw data) and Dirichlet regression is applied directly to proportional stage durations without the need to convert them to stage-specific development rates (as in, e.g. [ 8 ]), we do not report summary statistics such as time to 50% moult. Total development rate was calculated as the reciprocal value of the total duration of instars included in the analysis of each dataset.
Analyses of DRI based in individual-level data
Relative duration of a given developmental stage, i.e. the time spent in that stage divided by the complete developmental time (in days), was calculated for all individuals and stages included in the respective analysis. For Gastrophysa , Leptinotarsa and Loxostege , we analyzed both individual data and mean stage durations for data aggregated by egg clutches or rearing groups. The latter analysis avoids potential pseudoreplication issues and is more appropriate if growth and molting of individuals that develop together is highly synchronized. This was clearly true for leaf beetle eggs that always hatch in synchrony (D. Kutcherov, unpublished data). Previously developed DRI analyses [ 5 , 10 , 22 ] required the temperatures to be within a range in which the development rate increases linearly with temperature; their major motivation was to obtain a meaningful LDT value. Although our approach does not require such a restriction, we followed those analyses for the sake of direct comparison of the different methods and used linear regression (details not shown) to constrain the data to temperatures for which the criterion was met. This requirement excluded 0–3 temperatures from each dataset; final datasets included 2–4 temperatures in each species.
We compared a suite of Dirichlet regressions differing in the choice of predictors. We used the “common” parameterization in the Dirichlet regression, i.e. we modelled all parameters α = ( α 1 , …, α n ) independently [ 14 ], and always included the same set of predictors for all parameters. The simplest model (referred to as const in Results and Tables 1 and 2 ) corresponding to DRI assumed that the proportions of time spent in individual instars are constant and independent of other factors. Another model ( f ) assumed that the proportions are affected by an additional factor (sex: Acilius , Amara , Cloeon ; experimental photoperiod: Amara , Gastrophysa , Leptinotarsa ; geographic origin of the population: Loxostege ) but not by temperature. Model f therefore corresponds to DRI that differs between sexes, populations, or depends on photoperiod as a factor affecting the rate of development [ 5 ]. For Amara , we first investigated a model with both sex and photoperiod as two explanatory factors. Inclusion of sex did not significantly improve the fit, and we thus pooled all individuals in the subsequent analyses (details not shown).
10.1371/journal.pone.0129341.t001
Table 1 Summary of ΔAIC c values for Dirichlet regression models for individual species of aquatic and semiaquatic insects.
ΔAIC c
Species
Stages 1 ( n )
N T
N F
Factor f
Note 2
const
f
t
t F
f+t
f+t F
f*t
f*t F
Acilius
L2, L3, pupa (3)
3
2
sex
all temperatures
81.6
85.7
2.1
1.1
1.1
0
6.9
11.6
Cloeon
early, pre-final, final (3)
2
2
sex
all temperatures
49.4
27.6
27.5
-
3.3
-
0
-
Microvelia
L1–L5 (5)
3
2
sex
19–25°C
13.3
21.6
9.6
0
19.0
13.3
-
-
Microvelia
L1–L5 (5)
3
2
sex
17–21°C
44.7
53.1
53.1
0
63.7
14.4
-
-
Notonecta
L1–L5 (5)
3
2
sex
all temperatures
31.4
44.2
7.0
0
28.5
32.9
-
-
Velia
L2–L5 (4)
3
2
sex
12–19°C
86.4
90.4
27.8
0
40.6
14.8
-
-
1 Stages, temperatures and additional factors: n = number of stages; N T = number of temperatures; N F = number of additional factors.
2 Note: Range of temperatures included in the analyses, based on linearity of the relationship between temperature and development rate. Models including the interaction of the additional factor f and temperature (models f*t and/or f*t F ) are given when both the factor and temperature significantly improve the fit. Values of ΔAIC c ≤ 2 for each species and dataset given in bold. See the main text for model abbreviations.
10.1371/journal.pone.0129341.t002
Table 2 Summary of ΔAIC c values for Dirichlet regression models for individual species of terrestrial insects.
ΔAIC c
species
Stages 1 ( n )
N T
N F
Factor f
Note 2
const
f
t
t F
t + t 2
f+t
f+t + t 2
f+t F
f * t
f *( t + t 2 )
f * t F
Amara
E, L, P (3)
4
2
PP
all temperatures
347
28.5
228
235
-
0
3.8
0.2
4.9
-
-
Gastrophysa
E, L, P (3)
3
2
PP
I, 18–22°C
469
444
132
94.8
-
79.1
-
37.5
46.4
-
0
Gastrophysa
E, L, P (3)
3
2
PP
C, 18–22°C
23.6
24.6
1.1
5.6
-
0
-
4.1
0.9
-
12.3
Leptinotarsa
E, L, P (3)
4
3
PP
I, all temperatures
655
392
394
311
317
84.3
9.4
2.6
91.8
-
0
Leptinotarsa
E, L, P (3)
3
3
PP
I, 21–27°C
541
298
307
296
-
17.4
-
0
15
-
2.0
Leptinotarsa
E, L, P (3)
4
3
PP
C, all temperatures
168
101
94.7
78.3
78.9
13.4
0
1.5
25.6
24.3
28.1
Leptinotarsa
E, L, P (3)
3
3
PP
C, 21–27°C
123
68.2
67.7
61.7
-
4.9
-
0
12.6
-
13.9
Loxostege
E, L, P (3)
3
3
O
I, 18–24°C
215
53.6
205
205
-
33.9
-
29.3
33.4
-
0
Loxostege
E, L, P (3)
3
3
O
C, 21–27°C
15.4
0
13.5
20.3
-
0.3
-
9.9
6.1
-
66.3
1 Stages and additional factors. Stage abbreviations: E = egg, L = larva, P = pupa; factors: PP = photoperiod, O = geographic origin of population.
2 Range of temperatures included in the analyses, based on linearity of the relationship between temperature and development rate; I = individual-based data, C = mean clutch values.
Other details as in Table 1 .
The remaining models included the effect of temperature. The simplest one ( t ) assumed a linear effect of temperature on the parameters α of the underlying Dirichlet distribution. This translates into a nearly, but not perfectly, linear relationship between temperature and relative duration of a given developmental stage. We thus calculated the amount of DRI violation V D as the average difference between relative developmental times p ^ predicted for two successive temperatures and divided by the temperature difference,
V D = 1 N T − 1 ∑ i = 1 N T − 1 p ^ ( T i + 1 ) − p ^ i ( T i ) T i + 1 − T i
(3)
and compared it to the measure of DRI violation V A introduced in [ 5 ]; see eqn (A2) in S1 Text . We also investigated the possibility of independent effects of each temperature ( t F ), i.e. we treated temperature as a factor for data with three or more different temperatures, and the possibility of a unimodal dependence on temperature described by a second order polynomial (model t+t 2 ) for data with four different temperatures. Finally, we included models that consider joint additive effect of the additional factor and temperature (continuous temperature: f+t and f+t+t 2 temperature as factor: f+t F ) or their interaction on model parameters (continuous temperature: f*t and f* ( t+t 2 ), temperature as factor: f*t F ). The flexibility of Dirichlet regression models would allow for mutually independent effects of temperature or additional factors on individual model parameters; see [ 14 ] for details. We did not explore this possibility as we had no a priori hypotheses on which we could base such models. Instead, we assumed that temperature or the additional factor affected all model parameters equally.
For each dataset, we compared all models using Akaike information criterion with correction for small sample sizes (AIC c ). We chose the model with the lowest AIC c value as the most appropriate description of the underlying relationship; models for which the difference of the AIC c value from the lowest value is at most 2 and hence their evidence ratio does not deviate too strongly from unity also provide good fit to the data [ 30 ]. Comparing the AIC c values of models const and t F (or AIC c values of model f and f+t F ) is analogous to the deletion tests used in [ 5 ] to determine if the data are consistent with DRI.
Comparison of Dirichlet regression and standard ANCOVA analyses of DRI
To illustrate the differences between Dirichlet regression and previously used methods, we also carried out the ANCOVA analysis of mean values of the transformed proportional data following the procedure outlined in [ 5 ], i.e. Method 3 in [ 10 ]. We used the same datasets as in Dirichlet regression and compared the measures of DRI violation V D and V A across species. We also examined the dependence of DRI violation measure V D on ontogeny. To do so, we rescaled individual stages of pre-adult ontogeny of each species between 0 and 1 using equidistant intervals with the resolution depending on that of the data: we use 0 for egg, 0.5 for the combined larval stages and 1 for pupa of the four terrestrial species, the values of 0.25, 0.5, 0.75 and 1 for L1–L3 larvae and pupa of Acilius and the values of 0.2, 0.4, 0.6, 0.8 and 1 for L1–L5 larvae of Heteroptera. Finally, we assigned the values of 0.5, 0.75 and 1 to early, pre-final and final instars of Cloeon . These values are somewhat arbitrary but different values did not qualitatively change our conclusions (results not shown). More than one dataset could be analyzed for some species (either the linear dependence of development rate on temperature could be achieved for different, partly non-overlapping temperature intervals, or we wanted to highlight differences between individual- and clutch-based data). To avoid pseudoreplication, we averaged the values of DRI violation resulting from these analyses for each stage of each species across all datasets and levels of a factor.
All analyses were implemented in R software version 3.1.0 [ 31 ]; Dirichlet regression was implemented in the DirichletReg package v 0.6–0 [ 14 ] and graphs drawn in the ggplot2 package [ 32 ]. An example of the Dirichlet regression analysis is provided in S3 Text .
Results
DRI was violated in all species based on the results from Dirichlet regression except Loxostege , in which the best model was consistent with DRI for the clutch-aggregated data but suggested DRI violation for the individual-level data (Tables 1 and 2 ). Models including temperature as a factor ( t F ) were clearly favoured in the sense of having the lowest AIC c scores in three species: Velia , Microvelia and Notonecta . We found no support for sex-specific development in these species ( Table 1 ).
Inclusion of an additional factor, sometimes in interaction with temperature, led to a significantly improved model in all remaining species and datasets: sex in Cloeon and Acilius (although the differences between male and female Acilius were minor), photoperiod in Amara , Gastrophysa and Leptinotarsa , and geographic origin in Loxostege ( Fig 2 and Tables 1 and 2 ). Other model variants involving the factor and temperature usually provided a similarly good fit of the data (bold values in Tables 1 and 2 ) and their predictions were similar to the best-fitting model (details not shown).
10.1371/journal.pone.0129341.g002
Fig 2
Examples of results of Dirichlet regression for (a) Acilius and (b) Loxostege .
Acilius : best fitting model ( f+t F , grey solid line) compared to DRI (model const , black dashed lines); Loxostege : dataset with average clutch data, 21–27°C and different population origin (B = Buryatia, K = Krasnodar, H = Hebei), best fitting DRI model ( f , black dashed lines) compared to model f+t (black solid lines). Box and whisker plots of raw data: horizontal line = median, box = first to third quartiles, line = data within 1.5 times the interquartile range; dots = outliers. Non-feeding stages labelled with asterisk.
Although DRI violation detected by Dirichlet regression varied among the species, we observed a common pattern suggesting that the relative duration of intermediate developmental stages increased with temperature at the expense of shortened early instars and the last pre-adult instars ( Fig 3a ). This pattern did not differ significantly between aquatic and terrestrial species. The best model describing the DRI violation detected by Dirichlet regression V D as a function of relative developmental stage S and habitat included only a quadratic dependence on S , V D = -0.24 + 1.73 S– 1.69 S 2 ( F 2,29 = 3.61, P = 0.040, adj. r 2 = 0.14), and the linear and quadratic coefficients were both significantly different from zero ( P = 0.02 and 0.01, respectively). DRI violation was strongest in Cloeon and Acilius , in which the relative development rate of the intermediate stages included in the experiment (pre-final instar in Cloeon and L3 in Acilius ) increased by more than 1%.(°C) -1 . Removing these two species did not qualitatively change the results but increased the proportion of explained variance ( V D = -0.15 + 1.38 S– 1.38 S 2 ( F 2,18 = 6.12, P = 0.009, adj. r 2 = 0.34). The averaged slope of DRI violation did not exceed 0.5%.(°C) -1 in all other species and stages ( Fig 3a ).
10.1371/journal.pone.0129341.g003
Fig 3
DRI violation in the nine insect species studied.
(a) Average stage-specific DRI violation from Dirichlet regression. Stages on a relative scale (0 = egg, 1 = last pre-adult stage). Curve ± 95% confidence interval = best model describing the dependence of DRI violation V D on the relative developmental stage S (see main text for details). (b) Comparison of average stage-specific DRI violation from Dirichlet regression and standard ANCOVA analysis as described in Jarošík et al. (2002). Line ± 95% confidence interval = regression of V A on V D (see main text for details). Symbols represent values averaged across all levels of a factor and all datasets for each species included in Tables 1 and 2 ; small amount of horizontal jitter added to all data. Green fill = feeding stages of aquatic species, orange fill = non-feeding stages of aquatic species, grey fill = feeding stages of terrestrial species, no fill = non-feeding stages of terrestrial species; circles = Coleoptera, squares = Cloeon , diamond = Heteroptera, triangles = Loxostege .
ANCOVA analysis of mean values of the transformed proportional data following the procedure outlined in [ 5 ] revealed DRI violation in only four species ( Amara , Gastrophysa , Leptinotarsa and Acilius ). Moreover, it did not find a significant effect of the additional factor on relative developmental times in Acilius and Gastrophysa detected by Dirichlet regression ( S3 Table ). Finally, the best model describing the back-transformed values of DRI violation V A detected by the ANCOVA analysis as a function of relative developmental stage S and habitat included only the intercept, V A = -1.7·10 −4 ; the result was qualitatively identical when we dropped data for Acilius that had the highest leverage. This means that the standard DRI analysis could not detect variation in the temperature dependence of development rates during ontogeny. Magnitude of V A did not exceed 0.15%.(°C) -1 and was typically two orders of magnitude lower than V D (linear regression model: V A = -0.00019 +0.0082 V D; F 1,27 = 22.4, P < 10 −4 , adj. r 2 = 0.43; Fig 3b ) for reasons explained in S1 Text .
Discussion
Developmental rate isomorphy, akin to equiproportional development, is an important life history concept. Its widespread validity would greatly simplify efforts to understand the responses of individuals and populations to the anticipated climate change, because it implies that species respond uniformly to changing temperature during ontogeny [ 5 ]. Moreover, the ontogeny of species that comply with DRI could be conveniently characterized by only two numbers, the lower developmental threshold (LDT) below which the development should stop, and the slope relating development rate to temperature [ 5 ]. Previous studies found that DRI holds in many ectotherms and that violations of DRI are minor [ 5 , 11 ]. On the other hand, studies on EPD indicate that the concept is often violated in copepods [ 6 , 8 ].
Our results suggest that universal validity of DRI is unlikely. Most importantly, multiple characteristics of individuals that may alter the rates of growth and development change during ontogeny (e.g., ontogenetic diet and niche shifts [ 33 , 34 ]). Recent meta-analyses and experiments show that the upper thermal limit of growth of different species decreases with their body size [ 19 , 20 ]. If the underlying interspecific allometries also apply within some species, DRI will not hold for them unless the changes in growth rates are perfectly matched by changes in development rates. In our case study, DRI was violated in at least eight of the nine species when we used individual-level data. Contrary to our results, previous analyses of data on Leptinotarsa and Gastrophysa mostly supported DRI ( Leptinotarsa : [ 5 ]; Gastrophysa : [ 35 ], but see [ 36 ]).
Results of our case study point towards three limitations of previous DRI studies. First, the different results may arise from differences in statistical methods and data resolution explained above. Second, most previous studies have grouped all larval instars together. We show that using individual instars could detect DRI violation at higher resolution. Third, previous coverage of taxa might have been naturally biased. Although Jarošík et al . [ 11 ] extended the coverage of DRI to other invertebrate and vertebrate ectotherms (echinoderms, annelids, fish and anurans), the main bulk of DRI evidence remains rooted in insect studies. A large proportion of data come from economically important terrestrial species, usually pests and disease vectors and their predators; see [ 5 ]. Data on aquatic and semiaquatic insects are scarce: they are usually of low economic importance and do not include typical model species used for laboratory studies. On the other hand, different thermal conditions and temperature-dependent aerobic scopes in terrestrial and aquatic environments may lead to environment-specific patterns of growth and development [ 37 , 38 ] and ultimately to different temperature-size relationships [ 21 ].
Neither is the current DRI concept fully suitable to characterize species with non-linear dependence of development rate on temperature [ 8 ] and populations in suboptimal habitats, in which variable food limitation over ontogeny may further alter temperature dependence of development rates [ 7 ]. Future studies of DRI should therefore include both terrestrial and (semi)aquatic species, relax the constraint on linear relationship between development rate and temperature, and consider limiting food conditions. To this end, joint analyses of data from EPD and DRI studies would be particularly useful.
Violation of DRI: Possible causes and consequences
To identify which environmental factors and life history traits are responsible for the violation of DRI, individually resolved data on different generations in multivoltine populations and comparisons of related species with different voltinism or inhabiting environments with different thermal amplitudes would be particularly useful. We found no differences between the overall pattern of DRI violation in terrestrial and aquatic insects included in our study. This is unexpected, given that aquatic insects should experience smaller temperature fluctuations due to the higher thermal capacity of water and might perceive the environment as more predictable than terrestrial insects. We could not test if different aerobic scope in aquatic and terrestrial environments drives the patterns of DRI violations because our study included only one truly aquatic species ( Cloeon ) that does not breathe atmospheric oxygen.
The main pattern of DRI violation found in our analyses was shared among species: relative duration of the early and the last preimaginal stages decreased with temperature, whereas intermediate instars tended to last relatively longer. This result is consistent with our second hypothesis based on the narrowing metabolic scope of growth at higher temperatures, which affects the actively foraging larval stage but not the eggs or pupae ( Fig 1d ). This explanation is not satisfactory only for the Heteroptera, in which even the last larval instar actively forages for food. However, the patterns of DRI violation in the Heteroptera were generally variable ( Fig 3a ) and other currently unknown factors and mechanisms may be responsible for the result.
Seasonal constraints on development and variation in photoperiod can also lead to DRI violation. For example, absolute durations of egg and pupal stage were not affected by photoperiod in our data on Amara and Leptinotarsa , but the rate of larval development changed with day length and consequently changed the relative duration of all three stages. Other published data (e.g. for the damselfly Lestes eurinus , [ 39 ]) also suggest DRI violation under late-season conditions, presumably as some late-instar individuals accelerate growth to emerge before the end of the season and to avoid overwintering. The resulting pattern is qualitatively identical with our findings; its proximate cause includes a joint effect of temperature and day length on development rates (e.g. [ 40 ]).
Conclusions
We conclude that the DRI and EPD concepts developed for different taxa should be unified and possible patterns and causes of DRI/EPD violation critically re-evaluated. As we have illustrated, a fruitful approach could utilize the concept of stage- or size-specific thermal performance curves. Contrary to most previous studies, our experiments indicated that DRI is often violated in ectotherms and that this violation can be substantial. Using individually resolved data, we found DRI violation in insects that are both terrestrial and (semi)aquatic, predatory and herbivorous, and hemi- and holometabolous. We therefore suggest that modern statistical methods applied to individual-level data should be employed in DRI analyses whenever possible. Dirichlet regression used in this paper provides a highly flexible instrument that explicitly deals with the constraints imposed by proportional data that inherently arise in the study of DRI/EPD.
Supporting Information
S1 Table
Summary of experimental data used in the analyses.
(DOC)
S2 Table
Raw individual-level data on all nine species used in the analyses.
(XLS)
S3 Table
Results of standard DRI analysis.
(DOC)
S1 Text
Properties of the measure of DRI violation proposed in Jarošík et al. (2002).
(DOC)
S2 Text
Study organisms and experimental details.
(DOC)
S3 Text
Example of DRI analysis using Dirichlet regression.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Despite years of focused researches and advancements in therapies, acute respiratory distress syndrome (ARDS) remains a fatal disease with a mortality rate of 40–46% [ 1 , 2 ]. The new Berlin definition of ARDS guides physicians to the best treatment options based on illness severity and proposes ECMO as a valuable therapeutic option for patients with severe ARDS (PiO 2 /FiO 2 below 100) to take over lung function and minimize ventilator-induced lung injury when conventional support fails [ 3 ].
ECMO is one of several terms used for an extracorporeal circuit that directly oxygenates and removes carbon dioxide from the blood. ECMO was first used successfully on an adult patient in 1971 [ 4 ]. Early studies in adults did not demonstrate a survival benefit from ECMO for severe acute respiratory failure and showed a mortality rate of approximately 60–80% [ 5 – 8 ]. However, technological advances and continued experience, particularly involving the use of ECMO for ARDS during the 2009 influenza A (H1N1) pandemic, generated widespread interest in ECMO techniques and increased the survival rate of ARDS patients treated with ECMO to 56%-70% between 2006 and 2010 [ 9 – 11 ]. Paden et al. showed that in the United States from 1996 to 2006, the use of ECMO remained steady at approximately 100 cases per year; however, in 2009, the use of ECMO dramatically increased to 400 cases per year [ 12 ]. Because this significant increase in the use of ECMO, which requires highly specialized staff and equipment, may increase resource utilization and hospital costs, early identification of mortality risk factors is needed [ 13 ]. However, outcome predictors for ARDS patients treated with ECMO remain unclear, and studies evaluating the mortality rate of severe ARDS adult patients undergoing ECMO based on the new Berlin definition of ARDS are scarce. The role and proper use of ECMO for ARDS patients have not been definitively established, despite the completion of the conventional ventilatory support vs extracorporeal membrane oxygenation for severe adult respiratory failure (CESAR) trial [ 14 ].
Therefore, the aim of our study was to review the use of ECMO in severe ARDS patients based on the new Berlin definition of ARDS. We analyzed the epidemiological characteristics, clinical features and predictors of survival among severe ARDS patients treated with ECMO in our center.
Materials and Methods
Ethics statement
This study was approved by the Institutional Research Ethics Board of the First Affiliated Hospital of Guangzhou Medical University (Permit No 2016–03), which waived the need for informed consent for the retrospective collection of demographic, physiological and hospital outcome data based on Chinese legislation. All patient records/data were anonymized and de-identified prior to analysis.
Study design and patients
The First Affiliated Hospital of Guangzhou Medical University is a specialized acute-care university hospital. We reviewed the ECMO database, which identified all patients treated with ECMO between July 2009 and December 2014. We included all adult patients with a confirmed diagnosis of ARDS that was considered potentially reversible by the treating clinician based on the new Berlin definition of ARDS. Patients under 18 years of age were excluded from the study. ECMO therapy was indicated if patients exhibited a partial pressure of arterial oxygen (PaO 2 )/fraction of inspired oxygen (FiO 2 ) ratio below 80 mmHg for at least 2 h with FiO 2 of 1.0 and positive end expiratory pressure (PEEP)>5 cmH 2 O (We titrated PEEP according to ARDS network before ECMO support) or respiratory acidosis according to pH<7.20 despite the implementation of a lung-protective ventilation strategy (plateau pressure<30 cmH 2 O, tidal volume of 6 ml/kg).
Data collection
Basic information was collected from our institution’s ECMO database for all patients. The following retrospective data were obtained: demographic data, such as age, sex, height and weight; primary diagnosis for ECMO implementation; chest radiographs; respiratory and hemodynamic parameters; ventilator settings; blood gas values; lactate levels; and APACHE II score, which was calculated according to data obtained prior to the initiation of ECMO. We evaluated the chest radiographs of patients to define pre-ECMO barotrauma events. Radiographic evidence of barotrauma included pneumothorax, pneumomediastinum, pneumatoceles, or subcutaneous emphysema. Because barotrauma in mechanically ventilated patients often presents as acute, life-threatening hypoxemia or hypotension, we reasoned that not all cases of barotrauma would have previously been documented on chest radiography before chest tube placement. As a consequence, we chose an inclusive definition of barotrauma that would capture most events [ 15 ]. We also collected data concerning ECMO management, including ECMO mode; cannulation; duration of ECMO support and complications; mechanical ventilation time before ECMO initiation and after ECMO support; interventions such as renal replacement therapy; dates of hospitalization, discharge from the hospital; and cause of death. The primary outcome for this study was the hospital mortality rate. The secondary end-point was the outcome of weaning from ECMO. Successfully weaned patients were defined as those who remained alive within 48 hours after weaning from ECMO.
ECMO management
ECMO was initiated in our intensive care unit (ICU), which contains an established system with which to implement ECMO when needed. Percutaneous cannulation is performed by our ECMO team under general anesthesia in the ICU. The standard ECMO configuration for support of hypoxemic respiratory failure was veno-venous (femoro-jugular) ECMO. We used centrifugal pumps (Bioline, Maquet, Hirrlingen, Germany) at a flow rate of 3–5 L/min in all patients. Circuits were heparin-coated and composed of Quadrox PLS oxygenators (Bioline, Maquet, Hirrlingen, Germany) with HU 35 heater units (Maquet, Hirrlingen, Germany). Two circuit connectors were available between the pre- and post-oxygenators to provide renal replacement therapy via the ECMO circuit if required. Anticoagulation was maintained using continuous intravenous unfractionated heparin by targeting an activated clotting time of 160–180 s. The ventilator settings during ECMO were as follows: pressure control mode, PEEP 10–12 cmH 2 O, pressure above PEEP 12–15 cmH 2 O, plateau pressure 25–28 cmH 2 O, respiratory rate 16–18 breaths/min, and FiO 2 adjusted to obtain an arterial O 2 saturation of 90–95%; however, FiO 2 was set to 1.0 on the oxygenator. 25/38 (65.8%) of the patients were curarized during the mechanical ventilation before the initiation of ECMO and all patients did not receive neuromuscular blocking agents during the ECMO support. ECMO was continued until lung recovery or until irreversible multiorgan failure leading to death. Patients were weaned from ECMO when the following criteria were met: after stopping gas flow to ECMO, a PaO 2 /FiO 2 above 150 mmHg with PEEP<12 cmH 2 O, plateau pressure below 30 cmH 2 O and tidal volume of 5–7 ml/kg.
Statistical analyses
Continuous variables are expressed as means ± standard deviation (SD) or medians with interquartile range, and categorical variables are expressed as percentages. Student’s t-test was applied to compare the means of continuous variables for normally distributed data; otherwise, the Mann-Whitney U test was employed. Categorical data were tested using the x 2 test. Prognostic variables for mortality were analyzed by using the univariate logistic regression analyses, and variables with p <0.05 were used in multivariate logistic regression analyses with stepwise selection and the results were reported as odds ratios (ORs) and 95% confidence intervals (CIs). P<0.05 was considered statistically significant. Statistical analysis was performed using SPSS 12.0 software (SPSS, Inc., Chicago, IL, USA).
Results
Patient characteristics
During the study period, 43 patients with severe respiratory failure received ECMO treatment, and 38 of these patients were analyzed. 1 patient under 18 years of age, 3 patients who did not have severe ARDS according to the new Berlin definition of ARDS, and 1 patient who experienced cardiac arrest were excluded from the study ( Fig 1 ). As listed in Tables 1 and 2 , 32 patients were male, and the mean age was 51.3±13.2 years (range 27–74 years). The average age of the successful ECMO group was lower than that of the unsuccessful ECMO group (49.9±13.7 vs 57.3±10.1 years, P<0.05), and the mean body mass index (BMI) was similar between the two groups (25.62±2.42 vs 26.36±2.95 kg/m 2 , P>0.05). Furthermore, 65% (13/20) of patients without underlying lung disease were successfully weaned from ECMO (P<0.05), but 9 patients who had a history of lung fibrosis died during ECMO. It was noted that of the 9 patients with interstitial lung disease which included 3 patients who were suspected to suffer from connective tissue disease involving the lung based on clinical diagnoses and 4 patients who were considered to suffer from idiopathic pulmonary fibrosis (IPF) and were awaiting lung transplantation, and 2 were diagnosed with interstitial lung disease of unknown origin. In addition, of the patients receiving ECMO, 91.67% (11/12) of those with pulmonary barotrauma prior to ECMO died while receiving ECMO. The main cause of ARDS was documented as infectious diseases (32/38), with 26 (81.25%) patients presenting with bacterial pneumonia, 3 with virus infection, and 3 with fungal pneumonia. The mean APACHE II score on first admission to the ICU was 21.34±3.98. The average APACHE II score of the non-survival group was higher than that of the survival group (22.18±3.49 vs 19.4±3.2, P<0.05).
10.1371/journal.pone.0158061.g001
Fig 1
Flowchart and outcome of patients included in the study.
ARDS, acute respiratory distress syndrome; ECMO, extracorporeal membrane oxygenation.
10.1371/journal.pone.0158061.t001
Table 1 Baseline characteristics of the patients.
Characteristic
Total (n = 38)
Age (years)
51.39±13.27
Sex (male/female)
32 (84.2%)/6 (15.7%)
BMI (kg/m 2 )
25.73± 2.84
Underlying lung disease
Healthy lungs
18 (47.3%)
COPD
6 (15.7%)
Lung fibrosis
9 (23.7%)
IPF
4 (10.5%)
CTD-relative fibrosis
3 (7.9%)
Unknown
2 (5.3%)
Lung carcinoma
2 (5.3%)
Post-lung transplantation
5 (13.2%)
Barotrauma
12 (31.6%)
Co-morbidities
Hypertension
7 (18.4%)
Diabetes mellitus
5 (13.2%)
Renal insufficiency
19 (50.0%)
Infections
32 (84.2%)
Bacterial infection
26 (68.4%)
Viral pneumonia
3 (7.9%)
Fungal pneumonia
3 (7.9%)
APACHE II score
21.34±3.98
Continuous variables presented as means + SD, and categorical data are presented as numbers (%). BMI, body mass index; ARDS, acute respiratory distress syndrome; APACHE II, Acute Physiology and Chronic Health Evaluation II; CMV, cytomegalovirus; IPF, Idiopathic Pulmonary Fibrosis; CTD, Connective Tissue Disease.
10.1371/journal.pone.0158061.t002
Table 2 Baseline characteristics of patients who were successfully vs unsuccessfully weaned from ECMO and of survivors vs non- survivors.
Characteristic
Successful(n = 20)
Unsuccessful(n = 18)
p- value
Survivors (n = 16)
Non-survivors(n = 22)
p- value
Age (years)
49.9±13.7
57.3±10.1
0.039
50.±11.3
56.6±9.6
0.038
Sex (male/female)
16/4
16/2
12/4
20/2
BMI (kg/m 2 )
25.6±2.4
26.3±2.9
0.398
25.5±2.5
26.2±2.7
0.39
Underlying lung disease
Healthy lungs
13 (65.0%)
5 (27.8%)
0.024
11 (68.8%)
7 (31.8%)
0.027
COPD
4 (20.0%)
2 (11.1%)
0.384
2 (12.5%)
4 (18.2%)
0.498
Lung fibrosis
0
9 (50.0%)
<0.001
0
9 (40.9%)
<0.001
IPF
0
4 (22.2%)
0.041
0
4 (18.2%)
0.124
CTD-relative fibrosis
0
3 (16.7%)
0.158
0
3 (13.6%)
0.183
Unknown
0
2 (11.1%)
0.474
0
2 (65.0%)
0.329
Lung carcinoma
0
2 (11.1%)
0.218
0
2 (9.1%)
0.329
Post-lung transplantation
4 (20.0%)
1 (5.6%)
0.126
4 (25.0%)
1 (4.5%)
0.088
Barotrauma
1 (5.0%)
11 (61.1%)
<0.001
1 (6.3%)
11 (50.0%)
<0.001
Co-morbidities
Hypertension (n)
5 (25.0%)
2 (11.1%)
0.249
5 (31.3%)
2 (9.1%)
0.108
Diabetes mellitus, % (n)
2 (10.0%)
3 (16.7%)
0.50
2 (12.5%)
3 (13.6%)
0.918
Renal insufficiency
8 (40.0%)
11 (61.1%)
0.165
6 (37.5%)
13 (59.1%)
0.325
Infections
15 (75.0%)
17 (94.4%)
0.234
13(81.3%)
19 (86.4%)
0.682
Bacterial infection
9 (45.0%)
17 (94.4%)
0.001
8 (50.0%)
18 (81.8%)
0.020
Viral pneumonia
3 (15.0%)
0
0.135
3 (18.8%)
0
0.066
Fungal pneumonia
3 (15.0%)
0
0.135
2 (12.5%)
1 (4.5%)
0.562
APACHE II score
19.7±4.079
23.17±3.015
0.006
19.4±3.2
22.18±3.49
0.019
PH
7.28±0.14
7.36±0.09
0.054
7.29±0.14
7.33±0.10
0.391
PaCO 2 (mmHg)
65.76±26.78
57.12±18.44
0.263
57.87±13.9
64.49±28.2
0.394
Continuous variables presented as means + SD, and categorical data are presented as numbers (%). BMI, body mass index; ARDS, acute respiratory distress syndrome; APACHE II, Acute Physiology and Chronic Health Evaluation II; CMV, cytomegalovirus; IPF, Idiopathic Pulmonary Fibrosis; CTD, Connective Tissue Disease.
Respiratory characteristics and ventilation variables before and after ECMO
As shown in Table 3 , before ECMO, patients had severe respiratory failure despite advanced mechanical ventilator support, with a mean PaO 2 /FiO 2 of 70.32±18.71 mmHg and a PEEP of 13.47±1.33 cmH 2 O. Notably, as shown in Table 4 , early improvement of PaO 2 /FiO 2 was significantly greater in ECMO survivors than in non-survivors after ECMO initiation (142.7±54.10 vs 107.4±23.36 mmHg, p<0.05) despite similar ventilator parameters and ECMO settings.
10.1371/journal.pone.0158061.t003
Table 3 Respiratory and ventilation characteristics of those who were successfully vs unsuccessfully weaned from ECMO in the 4 h before and during ECMO.
Characteristic
Total (n = 38)
Successful (n = 20)
Unsuccessful (n = 18)
p -value
Ventilation parameters
PaO 2 /FiO 2
Before
70.32±18.71
71.07±17.92
76.20±20.67
0.288
After
119.8±43.12
135.5±51.39 * p
102.4±27.49
0.048
PH
Before
7.32±0.126
7.28±0.14
7.36±0.096
0.054
After
7.46±0.078
7.44±0.09
7.48±0.064
0.199
PaCO 2 (mmHg)
Before
61.70±23.31
65.76±26.78
57.12±18.44
0.263
After
39.3±7.02
39.84±7.42
38.70±6.71
0.626
PEEP (cmH 2 O)
Before
13.47±1.33
13.10±1.29
13.89±1.28
0.067
After
10.71±1.41
10.35±1.53
11.11±1.18
0.098
Tidal volume (ml)
Before
411.6±26.05
412.5±28.81
410.6±23.38
0.822
After
290.8±22.94
294.0±26.64
287.2±18.09
0.370
Plateau pressure
Before
27.32±1.165
27.55±1.98
28.39±1.24
0.132
After
24.63±1.78
25.05±1.90
26.00±1.08
0.071
Respiratory rate
Before
27.69±5.18
26.85±4.59
29.06±5.50
0.130
After
20.15±4.77
20.05±4.25 * p
20.50±5.43
0.68
Heart rate
Before
110.5±18.4
113.2±20.07
107.7±16.56
0.372
After
93.14±11.52
93.79±10.9 * p
92.78±12.14
0.791
MAP
Before
69.76±4.53
70.47±5.19
68.98±3.65
0.318
After
73.88±5.49
73.60±5.83
74.20±5.22
0.741
Other parameters
WBC count, 10 9 /L
Before
14.46±5.67
13.19±6.03
15.46±4.58
0.27
After
14.52±5.25
13.06±4.86
16.22±4.99
0.088
HCT
Before
30.14±5.73
30.35±6.62
29.89±4.52
0.806
After
28.73±3.75
28.50±3.79
28.94±3.68
0.285
Hemoglobin
Before
103±21.0
101.3±24.26
105.4±17.11
0.550
After
98.00±12.42
98.95±15.27
96.94±8.52
0.626
Lactate, mmol/L
Before
2.65 (0.78–10.65)
1.9 (0.78–5.99)
3.1 (1.68–10.65)
0.022
After
2.15 (1.1–12)
2.12 (1.1–4.9)
2.10 (1.60–12)
0.429
Renal replacement therapy
12
5
7
0.45
Duration of ventilation before ECMO (days)
6.41±7.58
3.87±4.64
8.94±9.21
0.036
Duration of ECMO (days)
11.13±14.64
6.59±5.56
16.47±19.74
0.039
Duration of ventilation after ECMO (days)
16.84±15.59
16.73±17.17
The data are expressed as n (%), medians (interquartile ranges), or means ± SD;
*p<0.05.
PaCO 2 , Partial pressure of carbon dioxide; PEEP, positive expiratory end pressure; MAP, mean arterial pressure; WBC, white blood cell; HCT, hematocrit.
10.1371/journal.pone.0158061.t004
Table 4 Respiratory and ventilation characteristics of survivors and non-survivors in the 4 h before and during ECMO.
Characteristic
Survivors (n = 16)
Non-survivors(n = 22)
p -value
Ventilation parameters
PaO 2 /FiO 2
Before
68.94±19.26
77.87±16.82
0.137
After
142.7±54.10 *
107.4±23.36
0.011
PH
Before
7.29±0.146
7.33±0.10
0.391
After
7.43±0.07
7.47±0.074
0.083
PaCO 2 (mmHg)
Before
57.87±13.91
64.49±28.28
0.394
After
39.22±7.84
39.36±6.54
0.952
PEEP (cmH 2 O)
Before
13.19±1.377
13.68±1.29
0.26
After
10.5±1.51
10.86±1.36
0.44
Tidal volume (ml)
Before
412.5±26.2
410.9±26.53
0.86
After
291.9±24.01
290±22.28
0.81
Plateau pressure
Before
27.80±2.13
28.05±1.36
0.68
After
25.38±1.89
25.45±1.62
0.89
Respiratory rate
Before
27.63±4.70
28.09±5.47
0.79
After
20.05±4.60 *
20.09±5.01
0.80
Heart rate
Before
113.8±17.6
108.3±18.61
0.37
After
94.5±8.88 *
92.41±12.78
0.58
MAP
Before
70.91±4.69
68.92±4.32
0.18
After
74.29±6.13
73.59±5.09
0.70
Other parameters
WBC count, 10 9 /L
Before
13.07±5.47
15.02±5.63
0.33
After
13.42±5.36
15.08±4.76
0.36
HCT
Before
29.25±6.52
30.77±4.99
0.42
After
28.13±4.06
29.41±3.44
0.41
Hemoglobin
Before
95.94±21.94
108.6±19.04
0.07
After
96.69±14.24
98.95±11.16
0.58
Lactate, mmol/L
Before
1.9 (0.78–5.99)
2.80 (1.1–10.65)
0.09
After
2.12 (1.4–4.9)
2.10 (1.1–12)
0.58
Renal replacement therapy
3
9
0.18
Duration of ventilation before ECMO (days)
3.08±3.89
8.59±8.69
0.023
Duration of ECMO (days)
5.73±4.91
15.57±17.88
0.040
Duration of ventilation after ECMO (days)
21.09±17.89
The data are expressed as n (%), medians [interquartile ranges], or means ± SD;
*p<0.05.
PaCO 2 , Partial pressure of carbon dioxide; PEEP, positive expiratory end pressure; MAP, mean arterial pressure; WBC, white blood cell; HCT, hematocrit.
Most patients were mechanically ventilated for fewer than 7 days prior to the initiation of ECMO. The median time between intubation and ECMO cannulation was 6.41 (0.4–28) days. As shown in Table 4 , compared with the surviving patients treated with ECMO, the non-survivors receiving ECMO experienced much longer mechanical ventilation durations before ECMO treatment (3.08±3.89 vs 8.59±8.69, p<0.05). In addition, the non-survivors displayed higher lactate levels [2.80 (1.1–10.65) vs 1.9 (0.78–5.99) mmol/L, P = 0.09]. During ECMO therapy, other characteristics were not significantly different between survivors and non-survivors. The median (range) duration of ECMO therapy was 5.73±4.91 days in survivors and 15.57±17.88 days in non-survivors. The duration of ECMO support was 56 days in one patient awaiting lung transplantation. In addition, the duration of mechanical ventilation after ECMO was 21.09±17.89 days in survivors. There was no significant difference in the creatinine or hemoglobin level or in the WBC count before ECMO between survivors and non-survivors. The duration of ECMO was longer in non-survivors than in survivors.
Outcomes of patients and predictors of mortality
A total of 20 patients (52.63%) were successfully weaned from ECMO, and 16 patients (42.11%) survived to hospital discharge. The complications and outcomes of patients treated with ECMO are listed in Table 5 . With respect to complications, hemorrhagic events occurred during ECMO in 16 patients (42.11%): gastrointestinal hemorrhage occurred in 9 patients, intrapulmonary hemorrhage occurred in 5 patients, intracerebral hemorrhage occurred in 1 patients and retroperitoneal hematoma in occurred in 1 patients; 1 of these patients required surgical treatment. Multiple organ failure associated with intractable respiratory failure was the most common cause of death; 18.18% of the patients died of severe infection, and 1 patient (4.5%) died of hemorrhagic complications.
10.1371/journal.pone.0158061.t005
Table 5 Outcomes of patients on ECMO support according to survival status.
Complications
Total (n = 38)
Survivors (n = 16)
Non-survivors(n = 22)
p -value
Major hemorrhagic complications
16
4
12
0.067
Gastrointestinal hemorrhage
9
5
4
0.45
Intrapulmonary hemorrhage
5
2
3
0.654
Intracerebral hemorrhage
1
0
1
0.579
Retroperitoneal hematoma
1
0
1
0.579
Deep venous thrombosis
4
1
3
0.433
Pulmonary embolism
3
0
3
0.183
Post-ECMO infection
5
1
4
0.286
Post-ECMO barotrauma
7
4
3
0.317
Acute kidney injury
11
4
7
0.466
Cause of death
Multi-organ failure
12
12
Irreversible respiratory failure
4
4
Severe infection
4
4
Intracerebral hemorrhage
1
1
Others
1
1
We further determined the relationship between patient characteristics and hospital mortality. Univariate analysis ( Table 6 ) identified 4 variables as statistically significant prognostic factors for hospital mortality: age, duration of ventilation before ECMO, barotrauma pre-ECMO and underlying lung disease. We then included these 4 significant risk factors identified from univariate analysis in multivariate logistic analysis and Multivariate analysis ( Table 6 ) showed that barotrauma pre-ECMO and underlying lung disease were significant and independent risk factors for hospital mortality, whereas Age (p = 0.567) and Duration of ventilation before ECMO (p = 0.117) were not significant.
10.1371/journal.pone.0158061.t006
Table 6 Multivariate logistic regression analysis: independent predictors of mortality.
Univariate logistic regression
Multivariate logistic regression
Variable
OR
p
OR
95% CI
p
Age
4.444
0.034
1.892
0.213–168.22
0.567
Duration of ventilation before ECMO
1.190
0.044
1.232
0.949–1.599
0.117
ECMO duration
1.117
0.111
APACHE II score before ECMO
1.203
0.082
PaO 2 /FiO 2 before ECMO
1.013
0.420
Lactate before ECMO
1.304
0.243
WBC before ECMO
1.073
0.322
Hemoglobin before ECMO
1.0320
0.074
Barotrauma pre-ECMO
26.25
0.004
34.176
2.193–532.497
0.012
Underlying lung disease
14.733
0.001
12.213
1.220–122.242
0.033
OR: odds ratio
Discussion
ECMO has been controversial because of its association with serious complications and poor outcomes over the last several years; however, advances in extracorporeal technology have renewed interest based on accumulating new evidence [ 16 ]. To date, most studies examining the rates and causes of death among critically ill patients, such as those with severe myocardial dysfunction and life-threatening respiratory failure, focused on the time point after the initiation of ECMO. In a large multicenter database of 1,473 adult patients supported with ECMO during respiratory failure, the rate of survival to hospital discharge was 50% [ 17 ]. The results of our study show that the mortality rate of adult patients suffering from severe ARDS undergoing ECMO is 57.89%. Studies evaluating the mortality rate of severe ARDS adult patients undergoing ECMO based on the new Berlin definition of ARDS are scarce. Several authors have reported the successful use of ECMO on patients with influenza A (H1N1) virus-induced ARDS infection-associated severe respiratory failure in Australian and New Zealand ICUs, resulting in a mortality rate varying from 21 to 33% [ 10 , 18 ]. In the CESAR trial, 63% of patients treated with ECMO survived, and this result demonstrated that ECMO produced favorable outcomes [ 14 ]. Most studies have shown that application of ECMO results in encouraging survival rates.
This is the first article to review the use of ECMO on severe ARDS patients admitted to the ICU in China based on the new Berlin definition of ARDS and to identify predictors of mortality among severe ARDS patients supported with ECMO. Compared with previous studies, the present study reported a higher mortality rate. Several factors may have accounted for the higher mortality rate at our center. First and most importantly, the present study reviewed the use of ECMO on severe ARDS patients according to the new Berlin definition of ARDS. A PiO 2 /FiO 2 >100 is an exclusion criterion for ECMO treatment; therefore, some ARDS patients with a PiO 2 /FiO 2 >100 and high carbon dioxide levels were excluded from this study. The patients enrolled in our study who were supported with ECMO may have been very sick and may have had refractory hypoxia [ 3 ]. Additionally, our center is considered as a regional reference center for the treatment of the most severe cases, and other centers would likely not use ECMO at the same rate on such high-risk patients; this difference could explain the high ARDS severity observed in our patients [ 19 ]. Second, compared with other studies, the present study reported a longer mean duration of ventilation before ECMO, and this result likely contributed to the high mortality rate. As reported previously, the ventilation time before ECMO is related to the risk of mortality: the longer the ventilation time, the higher the mortality rate [ 20 ]. Finally, most patients enrolled in this study had pneumonia associated with severe sepsis, and many of these patients required renal replacement therapy during ECMO support; these conditions may also have contributed to the high mortality rate observed in this study [ 21 ]. The mortality rate is most likely strongly infiuenced by characteristics that vary among centers [ 22 ].
Another objective of this study was to identify early prognostic factors of successful weaning from ECMO and mortality among severe ARDS patients treated with ECMO to help clinicians decide whether to treat patients with ECMO. We found that age, duration of ventilation before ECMO, underlying lung disease, and barotrauma prior to ECMO affected the hospital mortality rate of ARDS patients treated with ECMO. Furthermore, we observed that the average age of the survival group was lower than that of the non-survival group. Specifically, 53.8% of the patients less than 50 years old had a markedly good prognosis without severe disability, whereas most patients over 50 years of age did not survive when treated with ECMO. This result suggests that ECMO should be considered for young patients, even if they have other contraindications, and this conclusion is in agreement with many other studies [ 19 – 21 , 23 ]. Based on the present results, in addition to age, the duration of mechanical ventilation before ECMO was associated with mortality. According to the established protocol [ 12 ], a 7-day duration of ventilation is the cut-off point: a ventilation time>7 days is an exclusion criterion for ECMO treatment. Notably, in our study, most of the patients with barotrauma before ECMO were not successfully weaned from ECMO. This observation suggests that barotrauma prior to ECMO is associated with death, and this relationship between barotrauma and mechanical ventilation may explain why the duration of mechanical ventilation before ECMO was also associated with poor prognosis. However, in our study, 3 patients with ventilation times before ECMO of 9, 15, and 17 days were successfully weaned from ECMO and survived to hospital discharge. This result suggests that ECMO support could be initiated on patients who had been mechanically ventilated for more than 7 days if there was no barotrauma prior to ECMO. It may be appropriate to focus on the implementation of a protective ventilation strategy before ECMO. Further studies are necessary to clearly address this issue. In addition, in our study population, there were 9 ARDS patients with lung fibrosis, including fibrosis in the primary and secondary stages, such as connective tissue diseases ; ECMO failed in these patients. This result suggests that physicians should select appropriate candidates for ECMO among severe ARDS patients. Patients with potentially irreversible underlying lung diseases such as connective tissue diseases should not be recommended for ECMO unless they are awaiting a lung transplant. Although the severity of hypoxemia before ECMO was not different between ECMO survivors and non-survivors, we observed a greater improvement in PaO 2 among survivors than among non-survivors after ECMO initiation, and this improvement was associated with the implementation of protective ventilation.
Complications of ECMO, such as bleeding, remain a clinically significant issue [ 24 ]. Hemorrhagic events occurred in 16 patients (42.11%) undergoing ECMO in our study. Non-survivors displayed a higher rate of complications, including hemorrhage, deep venous thrombosis, pulmonary embolism, infection and renal complications. Most patients experienced at least one complication attributed to ECMO, e.g., brain death, cerebral infarction and seizures, thromboembolism, and circuit clots [ 4 ]. Other centers have found similarly high complication rates [ 19 – 21 , 23 ]. Notably, one patient who suffered from intrapulmonary hemorrhage was successfully treated via endoscopic hemostasis and survived to discharge. Nevertheless, the most common causes of death among ARDS patients are related to multiple organ failure associated with intractable respiratory failure and severe sepsis. As shown previously, the lactate levels of the non-survival group were higher than those of the survival group; this result suggests that ECMO may be less useful for severe ARDS patients with septic shock, greater cardiac output, and impaired peripheral oxygen extraction when their pulmonary gas-exchange capacity is severely impaired [ 13 , 25 , 26 ]. Whether ECMO is an appropriate therapy for septic adults is still controversial [ 27 ].
Importantly, ECMO assistance was successfully used as a bridge to lung transplantation in one patient with pulmonary fibrosis in our study. The patient survived to discharge without disability and had returned to work by the end of the study. Our center has begun using ECMO for critically ill adults as a bridge to lung transplantation, even though the currently available data are limited. However, the use of ECMO while awaiting lung transplantation has been very promising in some patients with reversible lung disease [ 28 , 29 ].
Another important point to discuss is the ECMO team. It is essential to recognize the crucial nature of the collaborative effort of physicians, respiratory therapists and nurses to manage patients using complex technology safely and effectively in the ICU [ 30 ]. ECMO should be performed at centers with high case volumes, established protocols, and clinicians who are experienced in its use [ 31 ].
Several limitations of this study should be acknowledged. First, this was a retrospective study performed at a single medical center, and this design limits the generalizability of these findings. Second, the number of included patients was small, and we failed to evaluate the long-term outcome of our patients, particularly in relation to the degree of pulmonary dysfunction and quality of life; this lack of data limited the conclusions that could be drawn. Future work in this area, where possible, should include a larger number of subjects and a control group not receiving ECMO from the same population. One strength of the present study is that it only includes patients with severe ARDS according to the New Berlin definition. Previous studies have frequently mixed patients with ARDS and those with cardiogenic shock, but these diseases are likely to affect different populations and to have different prognostic factors [ 32 ].
In conclusion, our results revealed a hospital mortality rate of 57.89% among severe ARDS patients. Our data also demonstrated that advanced age, a long duration of ventilation before ECMO, underlying lung disease and barotrauma prior to ECMO affected the mortality rate of ARDS patients being treated with ECMO. Barotrauma prior to ECMO and underlying lung disease were independent prognostic factors for survival to hospital discharge among ARDS patients treated with ECMO based on multivariate analysis. These results might help physicians select appropriate candidates for ECMO among severe ARDS patients. Accordingly, our findings indicate that ECMO can be used as an alternative therapy for severe ARDS patients when conventional ventilation fails. However, additional studies should be conducted to further define the indications for ECMO use in severe ARDS patients.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Dementia is the sixth-leading cause of death and a major leading cause of disability in the United States. [ 1 , 2 ] It is the only disease among the top 10 causes of death that has no effective treatment or prevention.[ 3 ] Over the past 20 years, research has linked several cardiovascular risk factors (CVRF) to higher risks of dementia, Alzheimer’s disease, and vascular dementia.[ 4 – 7 ] For many CVRF, the risk for later cognitive decline depends on the timing of risk factor evaluation. Numerous studies indicate that dementia and cognitive dysfunction are more strongly associated with CVRF measured in midlife[ 6 , 8 , 9 ] than with CVRF assessed later in life.[ 10 – 13 ]
Better understanding of cardiac function in midlife may offer novel insights for dementia prevention. The QT interval on electrocardiogram represents the length of time required for the process of ventricular depolarization and repolarization, with longer QT interval indicating slower repolarization. Although clinically relevant prolonged QT interval is a known risk factor for major cardiovascular events like stroke,[ 14 , 15 ] few studies have examined the potential link between ventricular repolarization and cognitive outcomes. All studies focused on cross-sectional associations with mild cognitive impairment (MCI) or overt dementia.[ 16 – 19 ] For example, QT dispersion, a measure of short-term variability in QT interval, was associated with worse cognitive performance in patients with MCI in small cross-sectional studies comparing individuals with normal cognition, MCI, and dementia.[ 16 , 18 ] Conversely, QT interval was not cross-sectionally associated with cognitive performance in 839 very-old adults from the Chicago Health and Aging Project (mean age of 81years).[ 20 ]
We are unaware of any longitudinal studies of the association between prolonged QT interval and cognition. Despite the evidence that midlife CVRFs are often most relevant to late-life cognitive health, the longitudinal association of ventricular repolarization at midlife and late-life cognitive decline has not yet been considered. Therefore, we aimed to examine whether prolonged QT interval in midlife was predictive of late-life cognitive decline over 25 years of follow-up.
Material and methods
Participants
The Honolulu-Asia Aging Study (HAAS) began in 1991 and has been described in detail previously.[ 21 ] HAAS extends the Honolulu Heart Program (HHP), a prospective study of heart disease and stroke in Japanese-American men born between 1900 and 1919, who lived on Oahu in 1965. Exams 1 (1965/68), 2 (1967/70), and 3 (1971/74) were conducted in midlife as a part of the original HHP study. Beginning twenty years after Exam 3, Exams 4 (1991/93), 5 (1994/96), 6 (1997/99), and 7 (1999/2000) were conducted in late life as part of HAAS. Of the original HHP cohort of 8,006 men, 2,511were enrolled in HAAS and assessed longitudinally for cognitive performance beginning at Exam 4 in 1991/93 ( Fig 1 ). HAAS was approved by the Kuakini Medical Center Institutional Review Board and by the Human Research committees, and participants were informed about the study and signed informed consent forms.
10.1371/journal.pone.0229519.g001
Fig 1
Flowchart of the study participants from the Honolulu-Asia Aging study.
QT interval measurements, adjustment for ventricular rate, and identification of incident prolonged QT interval
The QT interval, which is comprised of the QRS complex, the ST segment, and T wave, is a widely-used measure of ventricular repolarization.[ 22 ] We used a correction formula that considers QT interval variation by ventricular rate.[ 23 , 24 ] Specifically, we used ECG data from Exam 1 to estimate the coefficient ß 1 in the following linear regression model: E (QT interval) = ß 0 + ß 1 (RR interval), where the RR interval is the average time elapsed in seconds between ventricular beats (60 / ventricular rate). We then used the estimated coefficient (ß 1 = 0.158) to calculate a ventricular rate-corrected QT interval: QT adj = QT + 0.158(1-RR). We also explored linear regression models in which the ß 1 coefficient was adjusted for age or allowed to vary with age, but found these enhancements did not meaningfully change the coefficient.
As our exposure of interest, we identified incident prolonged QT interval at Exam 2 or 3, rather than prevalent prolonged QT interval at Exam 1. Incident prolonged QT interval represents a change from faster to slower ventricular repolarization over the course of a few years, suggesting a person may be transitioning away from their normal ventricular function. Also, identifying incident prolonged QT interval at Exam 2 or 3 as our exposure of interest allowed us to use variables measured at Exam 1 to control for confounding via inverse probability weighting in the statistical analysis (see below). To identify participants who experienced incident prolonged QT interval, we first identified the 75 th percentile of rate-corrected QT interval at Exam 1, which was 407 milliseconds. Next, we excluded participants whose rate-corrected QT interval was above this value at Exam 1, considering them to have prevalent prolonged QT interval at Exam 1, and thus be ineligible for developing incident prolonged QT interval. Among the remaining participants, we then identified those whose rate-corrected QT interval was above 407 milliseconds at Exam 2 or 3 as having incident prolonged QT interval ( Fig 1 ).
Cognitive Abilities Screening Instrument (CASI)
CASI is a 40-item test of global cognitive function with scores ranging from 0 to 100. Higher scores indicate better cognitive performance.[ 25 , 26 ] Inventory items were designed to be comparable to the Hasegawa Dementia Screening Scale,[ 27 ] the Mini Mental State Examination (MMSE),[ 28 ] and the Modified MMSE.[ 29 ] One additional item indexing judgment ability was also included in the inventory. To ensure equivalence across English and Japanese languages, the CASI was developed in parallel in English and Japanese at three workshops in which items were scrutinized for cultural equivalence, back-translated, and pilot-tested. CASI scores obtained in HAAS and in a purely Japanese sample were analyzed using item-response theory to further validate the cross-cultural sensitivity of CASI; no evidence of salient differential item functioning due to language of testing was found.[ 25 ] For analysis, the CASI was scored using item response theory, to measure change over time on a consistent metric;[ 30 ] the mean CASI-IRT score across Exams 4–7 was approximately normal, with mean 0 and standard deviation of 1.
Statistical analysis
We estimated the association of incident prolonged QT interval at Exam 2 or 3 with trajectories of cognitive function from Exam 4 to Exam 7 using a marginal structural model.[ 31 ] This model uses weighting to address potential bias due to confounding and additionally implements weighting to address the potential for informative missingness related to loss to follow-up.[ 31 , 32 ] To address confounding, we reweight persons in the observed sample using inverse probability of exposure weights (IPEW) to eliminate associations between potential confounders and incident prolonged QT. Thus, under standard assumptions, IPEW addresses confounding by making the exposure statistically independent of the confounders. Similarly, to reduce bias due to selective attrition, we reweight persons in the observed sample who are similar to those who drop out of the sample using inverse probability of attrition weights (IPAW). Thus, under standard assumptions, IPAW addresses informative censoring by making attrition independent of predictors of attrition, including ascertained exposure and outcome.
To establish our sample for analysis of midlife incident prolonged QT interval in relation late life to cognitive decline, we first excluded participants who did not have an electrocardiogram (ECG) at Exam 1 (n = 570), had prevalent prolonged QT interval at Exam 1 (n = 1,844), were lost to follow-up after Exam 1 (n = 317), or did not have ECG at Exam 2 or 3 (n = 538), leaving 4,737 participants ( Fig 1 ). Data from these 4,737 persons were used to derive IPEW to account for confounding and IPAW to account for attrition from Exam 2 to Exam 4. After calculating those weights, we then further excluded participants who died or were lost to follow-up between Exam 2 and Exam 4 (n = 2,208) or had no CASI data at Exam 4 (n = 18), leaving 2,511 participants who completed CASI at Exam 4. Data from these 2,511 persons were used to derive IPAW to account for attrition from Exam 4 to Exam 7 ( Fig 1 ).
We estimated stabilized IPEW using logistic regression models on a dataset including one line for each participant at risk of experiencing incident prolonged QT interval at either Exam 2 or Exam 3. Covariates included in the denominator model were potential confounders of the association between prolonged QT interval and cognition ( S1 Table ). Covariates included in all numerator models included the subset of baseline or time-invariant covariates included in the denominator model. Stabilized IPEW were derived for each participant at each time point using standard formulae, and IPEW at each visit for each participant were then multiplied together to provide a final stabilized IPEW for each participant.[ 31 , 32 ]
Similarly, we estimated stabilized IPAW to address attrition from Exam 2 through Exam 7 using logistic regression models. Because cognitive data from CASI (our outcome measure) was available only from Exam 4 onward, we modeled attrition from Exam 2 to 4 and from Exam 4 to 7 separately. Furthermore, given the expectation of two separate attrition processes—attrition due to death and attrition due to non-death drop-out—we also used separate models to account for these two processes, with models for attrition due to non-death drop out conditional on remaining alive.[ 33 , 34 ]. Details of variables included in each model are available in S1 Table . As with the IPEW models, time-invariant and baseline covariates were included in the numerator models used to derive stabilized IPAW.
Stabilized weights for each process were derived for each participant at each time point, and were then multiplied together to provide a final set of stabilized IPAW or IPEW weights, with unique weights assigned to each participant at each time point. Finally, our stabilized IPEW and IPAW were multiplied together to create a unique weight for each participant contributing to the final analysis at each time point from Exams 4 to 7. Extreme weights were truncated at the 99 th and 1 st percentiles for use in primary analyses, as use of truncation often provides a good balance between bias and efficiency.[ 35 ]
For our primary analysis (Model 1), we estimated the association of incident prolonged QT interval on cognitive trajectories using weighted linear regression models estimated using generalized estimating equations with an independence covariance matrix.[ 36 ] We adjusted this analysis for covariates included in the models to estimate the numerators of the stabilized IPAW or IPEW, as well as important predictors of cognition: baseline age, time in study, an age by time in study interaction, generation, the presence of any APOEε 4 alleles, education, occupation (Exam 1), height (Exam 2), chest depth (Exam 1), alcohol use (Exam 1), physical activity level (Exam 1), and history of hypertension (Exam 2). In secondary analyses, we considered the sensitivity of our findings to our modeling choices. Specifically, we compared our primary analyses to analyses using non-truncated stabilized weights (Model 2), or only stabilized IPEW weights (Model 3), complete omission of weighting, i.e., adjustment only for baseline/time-invariant covariates (Model 4), and omission of multiple imputation, i.e., analyzing only participants who had complete data on all covariates (Model 5).
Missing data in covariates used to create the weights requires either censoring of persons at the time of first missing data or implementation of methods to address missingness. Given that many participants were missing data on at least one of the covariates included in models used to create our weights, we implemented multiple imputation by chained equations (MICE) to address missing covariate data.[ 37 ] We imputed five replicate datasets after a burn-in of 10 iterations. Derivation of the weights and estimation of the MSM occurred separately within each of our five imputed datasets. Reported findings combined estimates from each of the five imputed datasets using standard methods.[ 38 ]
Results
Of the 4,737 participants at Exam 2 (midlife), 2,511 had follow-up at Exam 4 (late life), when cognitive performance was first evaluated ( Fig 1 ). These 2,511 were our analytic sample, weighted to represent the 4,737 who met our eligibility criteria. Their mean IRT-CASI score at Exam 4 was 0.4±0.9. Their mean corrected QT interval had been 395.1±19.3 milliseconds (range 318–475) at Exam 2, and 1,076 participants (42.9%) had incident prolonged QT interval at either Exam 2 or 3. Participant characteristics at Exam 2 by incident prolonged QT interval at Exam 2 or 3 are shown in Table 1 . The mean of the final stabilized weights applied to the Visit 4 data from the 2,511 participants in our analytic sample was 0.99 (range: 0.46, 1.86). Additional details about the weights are provided in the Supplemental Appendix ( S2 , S3 , and S4 Tables).
10.1371/journal.pone.0229519.t001
Table 1 Participant characteristics at Exam 2, stratified by incident prolonged QT interval at midlife (Exam 2 or 3).
Total (n = 2,511)
Prolonged QT (n = 1,076)
No Prolonged QT (n = 1,435)
p-value a
Age (years), mean (SD)
54.5 (4.4)
54.7 (4.6)
54.3 (4.4)
0.053
Generation, n (%)
0.691
Issei
149 (6%)
68 (6%)
81 (6%)
Kibei
226 (9%)
93 (9%)
133 (9%)
Nisei
2,136 (85%)
915 (85%)
1,221 (85%)
Education, n (%)
0.217
None or primary
449 (18%)
210 (20%)
239 (17%)
Intermediate
682 (27%)
295 (27%)
387 (27%)
High School
821 (33%)
332 (31%)
489 (34%)
Technical School
278 (11%)
104 (10%)
174 (12%)
University
281 (11%)
135 (13%)
146 (10%)
Clerical, sales, professional or managerial job b , n (%)
787 (31%)
368 (34%)
419 (29%)
0.007
Hypertension diagnosis, n (%)
225 (9%)
107 (10%)
118 (8%)
0.135
Alcohol (ounces/ month), mean (SD)
11.9 (20.3)
12.0 (19.4)
11.8 (20.9)
0.797
Height (cm), mean (SD)
164 (6)
164 (6)
163 (6)
0.001
Chest depth (cm) b , mean (SD)
19.2 (1.8)
19.3 (1.9)
19.1 (1.8)
0.123
Physical Activity Index (midlife) b , mean (SD)
32.9 (4.7)
32.8 (4.6)
33.0 (4.8)
0.169
Presence of at least one APOE-4 allele, n (%)
462 (19%)
218 (21%)
244 (18%)
0.038
a T-tests for continuous variables, chi-squared for categorical variables, and Wilcoxon’s rank sum test for education.
b Evaluated at Exam 1
Prolonged QT interval in midlife (Exam 2 or 3) was not associated with IRT-CASI score at Exam 4 (20 years after Exam 3) or with subsequent decline in IRT-CASI score over subsequent exams spanning the next 10 years ( Table 2 ). In our primary analysis (Model 1), assuming reference level for all covariates, participants without midlife prolonged QT interval declined on IRT-CASI score by an average of 0.09 points per year (95% CI: 0.08 to 0.09 points per year). Participants with midlife prolonged QT interval were not significantly different in the magnitude of subsequent decline in IRT-CASI score over time (estimated difference of -0.002 points of decline per year; 95% CI: -0.013 to 0.010 points; P = 0.79). In a series of sensitivity analyses in which we applied different modeling strategies ( Table 2 , Models 2–5), we obtained results very similar to those from our primary analysis.
10.1371/journal.pone.0229519.t002
Table 2 Association of elevated QT interval in midlife with item response theory-adjusted Cognitive Abilities Screening Instrument (IRT-CASI) score later in life.
Model and parameter
Estimated IRT-CASI score
95% CI
P value
Model 1
Study time (years)
-0.09
(-0.09, -0.08)
<0.0001
Elevated QT
0.04
(-0.28, 0.35)
0.81
Elevated QT × study time
-0.002
(-0.013, 0.010)
0.79
Model 2
Study time, y
-0.10
(-0.11, -0.09)
<0.0001
Elevated QT
-0.06
(-0.46, 0.34)
0.76
Elevated QT × study time
0.003
(-0.013, 0.018)
0.71
Model 3
Study time, y
-0.07
(-0.08, -0.06)
<0.0001
Elevated QT
0.06
(-0.25, 0.39)
0.68
Elevated QT × study time
-0.003
(-0.015, 0.009)
0.62
Model 4
Study time, y
-0.07
(-0.08, -0.06)
<0.0001
Elevated QT
0.05
(-0.27, 0.36)
0.77
Elevated QT × study time
-0.002
(-0.014, 0.009)
0.69
Model 5
Study time, y
-0.07
(-0.08, -0.06)
<0.0001
Elevated QT
0.04
(-0.26, 0.34)
0.78
Elevated QT × study time
-0.002
(-0.014, 0.010)
0.72
Model 1: Inverse probability of exposure weighting (IPEW) and inverse probability of attrition weighting (IPAW), with weights truncated at 1 st and 99 th percentiles, with multiple imputation.
Model 2: IPEW and IPAW, with weights not truncated, with multiple imputation.
Model 3: IPEW (no IPAW), with weights not truncated, with multiple imputation.
Model 4: Unweighted, with multiple imputation.
Model 5: Unweighted, without multiple imputation.
All models additionally adjusted for generation, alcohol use, physical activity level, education, occupation, and chest depth at Visit 1; age, height, and hypertension at Visit 2; and the presence of any APO-E4 alleles at visit 4.
The reference person was, at exam 2, 55 years old, 164 cm tall, and without a hypertension diagnosis. At exam 1 he was Nisei, with a primary education or less, had a chest depth of 19 cm, did not have a Clerical, sales, professional or managerial job, did not drink, and had a Physical Activity Index of 33. He also had no APOE-4 alleles.
Discussion
Using marginal structural models to reduce bias from confounding and participant attrition, we found that midlife QT interval was not associated with late-life CASI score approximately 25 years later, nor with cognitive decline in CASI over time in a large sample of Japanese American men from the HAAS. We confirmed this result in sensitivity analyses.
Certain aspects of left ventricular (LV) dimensions and ejection fraction have previously been associated with cognitive impairment and decline.[ 17 , 39 ] Increased LV dimensions in 211 men who were 68 years old at baseline were associated with higher risk of cognitive decline 14 years later.[ 17 ] Similarly, in 1,114 participants of the Framingham Heart Study Offspring Cohort, LV ejection fraction, an indicator of cardiac dysfunction, was associated with worse performance on neuropsychological tests related to visuospatial memory, object recognition, and executive function.[ 39 ] In addition, patients with severe LV dysfunction showed improved performance on executive and visuospatial function tests three to six months after cardiac resynchronization therapy.[ 40 , 41 ] Alterations in LV structure and function may be related to cognitive impairment through either brain hypoperfusion. An additional mechanism could be the presence of ventricular arrhythmias and the generation of thromboembolism, which could lead to cerebral infarcts and transient hypoperfusion.[ 42 ]
However, the association between ventricular arrhythmias and cognitive performance has been far less investigated. Silent myocardial ischemia and repeated ventricular premature beats were more frequent in patients with MCI and Alzheimer’s disease than in participants with normal cognition.[ 19 ] In a same sample of 33 patients with Alzheimer’s disease, 39 with MCI, and 29 controls, QT dispersion was associated with worse performance in the MMSE.[ 18 ] In a large sample of patients with normal LV ejection fraction, Coppola et al . found lower QT interval in participants with normal cognition (n = 224) compared to patients with MCI (n = 77) and Alzheimer’s disease (n = 77). These findings came from small cross-sectional studies with patients with cognitive complaints.
Research on ventricular repolarization has been even more sparse. The only study so far that had investigated the association between QT interval and cognitive performance using community-dwelling older adults was cross-sectional and included mostly very old participants (mean = 81, range of 76–85 years).[ 20 ] In that study, as in ours, repolarization measurements were not related to cognitive performance.[ 20 ] Although additional confirmation is required, our study suggests that subtle changes in the natural rhythm of the heart in midlife are unlikely to affect cognition decades later.
Our study should also be examined in light of its limitations. We did not have information on drugs that can affect the QT interval [ 43 ]. However, the low prevalence of use of drugs that may cause QT prolongation in previous studies (2–3%) suggest that our inability to consider drug use will not be a large source of bias.[ 44 ] Although we included several factors that could increase the chance of QT interval prolongation, information was missing on other clinical conditions that could influence QT interval (e.g. bundle branch block, hypokalemia and hypocalcemia, endocrine disorders). In addition, the HAAS is a cohort of Japanese American men, and our findings may not hold for other ethnicities and women. Techniques for measuring QT interval have improved since 1965–1971, so it is possible that modern measurements would be more predictive. We also cannot account for any incident prolonged QT interval between Exam 3 and the cognitive measurements. Moreover, we could not examine the association of QT interval with specific cognitive domains since we did not have a complete neuropsychological examination. Although QT interval was not associated with cognitive decline evaluated by CASI, it could be associated with vascular dementia or with cognitive decline in specific domains, such as executive function.
The strengths of this study include a large sample of participants with ECG data and cognitive evaluation over time. In addition, the causal modeling approach used in this study is a strength. MSM with IPEW and IPAW enables robust estimation of the association in contexts where we are concerned about bias due to confounding and attrition. Additionally, we examined for the first time the association of midlife incident prolonged QT interval with later cognitive performance. Since neuropathological lesions associated with dementia may start even two decades before the clinical symptoms of dementia, midlife risk factors are likely most relevant to later cognition.[ 45 ] Finally, we present here results from a community-dwelling cohort of adults; thus our results should be generalizable to similar community-dwelling populations.
In conclusion, in a prospective study of midlife ventricular function and cognition in Japanese-American men, we did not find an association of prolonged QT interval in midlife with cognitive performance or decline after 25 years of follow-up. However, future longitudinal studies with different ethnicities and women are important to confirm our findings.
Supporting information
S1 Table
Variables included in the denominator of the inverse probability weight models.
(DOCX)
S2 Table
Visit-specific estimated inverse probability of attrition and inverse probability of exposure weights.
(DOCX)
S3 Table
Mean of final non-truncated weights applied in estimation of the MSM at each visit with cognitive data.
(DOCX)
S4 Table
Weighted demographics demonstrating inverse probability weights recover distribution of characteristics at baseline (Exam 2).
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
In mammalian cells, the main antiviral defense system involves the activation of a signaling cascade relying on production of type I interferon (IFN I). This pathway depends on the recognition of extrinsic signals or pathogen associated molecular patterns (PAMPs) by dedicated host receptors. Double-stranded (ds) RNA, which can originate from viral replication or convergent transcription, is a very potent PAMP and can be sensed in the cell by various proteins among which a specific class of DExD/H-box helicases called RIG-I-like receptors (RLRs) [ 1 ]. RLRs comprise RIG-I, MDA5 and LGP2 and transduce viral infection signals to induce expression of IFN I cytokines that act in autocrine and paracrine fashions. These cytokines then trigger the expression of hundreds of interferon-stimulated genes (ISGs) to stop the virus in its tracks [ 2 ]. Among those ISGs, dsRNA-activated protein kinase R (PKR) plays an important role in antiviral defense by blocking cellular and viral translation upon direct binding to long dsRNA [ 3 ]. PKR is a serine-threonine kinase that dimerizes and auto-phosphorylates upon activation. It then phosphorylates numerous cellular targets among which the translation initiation factor eIF2α, which results in the inhibition of cap-dependent translation [ 4 ]. Accordingly, translation of many RNA viruses, including alphaviruses, is inhibited by PKR [ 5 – 7 ]. PKR is also involved in other cellular pathways including apoptosis, autophagy and cell cycle [ 3 , 8 ].
RNAi is another evolutionary conserved pathway triggered by long dsRNA sensing [ 9 ]. One key component in this pathway is the type III ribonuclease DICER, which is also essential for micro (mi)RNA biogenesis [ 10 , 11 ]. These small regulatory RNAs are sequentially produced by the two ribonucleases DROSHA and DICER, before being loaded into an Argonaute (AGO) effector protein in order to regulate their target mRNAs [ 12 ]. Whatever its substrate, be it long dsRNA or miRNA precursor, DICER relies on interacting with co-factors to be fully functional. In mammalian cells, the TAR-RNA binding protein (TRBP), a dsRNA binding protein (dsRBP), was shown to play a role in the selection of DICER substrates, its stabilization, strand selection and incorporation into AGO2 [ 13 ]. The interaction with TRBP is well characterized and depends on the helicase domain of DICER and the third dsRNA binding domain (dsRBD) of TRBP [ 14 ]. Another dsRBP, the protein activator of interferon-induced protein kinase R (PACT), was also described as an important cofactor of DICER. Although its function is not fully understood, PACT seems to also participate in miRNA loading and strand selection [ 15 , 16 ] via protein-protein interaction between the DICER helicase domain and the third dsRBD of PACT [ 17 ].
It is now common knowledge that RNAi is the main antiviral defense system in several phyla such as plants, arthropods and nematodes (reviewed in [ 18 ]). However, its exact contribution in the mammalian antiviral response remains unclear [ 19 – 21 ]. Recent studies indicate that a functional antiviral RNAi does exist in mammals in specific cases. An antiviral RNAi response was first detected in undifferentiated mouse embryonic stem cells [ 22 ] lacking the IFN response, suggesting that these two pathways could be incompatible. Indeed, in mammalian somatic cells deficient for MAVS or IFNAR, two components of the interferon response, an accumulation of DICER-dependent siRNAs derived from exogenous long dsRNA was detected [ 23 ]. In addition, the RLR LGP2 was found interacting with both DICER and TRBP, blocking respectively siRNA production and miRNA maturation [ 24 – 26 ]. Moreover, AGO4 was recently shown to be involved in antiviral RNAi against Influenza A virus (IAV), Vesicular stomatitis virus (VSV) and Encephalomyocarditis virus (EMCV) [ 27 ]. Finally, viral suppressors of RNAi (VSRs) have been shown to prevent DICER from playing an antiviral role in mammalian cells [ 28 , 29 ]. Nonetheless, several studies reported no detection of viral siRNAs in mammalian somatic cells infected with several viruses [ 30 – 32 ]. In somatic cells, only a helicase-truncated form of human DICER could produce siRNAs from IAV genome [ 33 ], but it also turned out that these siRNAs cannot confer an antiviral state [ 34 ].
Based on these conflicting observations, we decided to study the involvement of DICER during infection of human cells with the Sindbis virus (SINV). SINV is a member of the Togaviridae family in the alphavirus genus, which is transmitted by mosquitoes to mammals and can induce arthritogenic as well as encephalitic diseases [ 35 ]. It is widely used as a laboratory alphaviruses model as it infects several cell types and replicates to high titers. SINV has a positive stranded RNA genome of about 12 kb, which codes for two polyproteins that give rise to non-structural and structural proteins, including the capsid. Moreover, upon viral replication, a long dsRNA intermediate, which can be sensed by the host antiviral machinery, accumulates. Of note, SINV dsRNA can be cleaved into siRNAs in insects as well as in human cells expressing the Drosophila DICER-2 protein [ 36 ]. Nonetheless, although human DICER has the potential to interact with the viral RNA duplex, we did not find evidence that SINV dsRNA could be processed into siRNAs in somatic mammalian cells [ 30 , 36 ]. We thus hypothesized that specific proteins could interfere with DICER during SINV infection by direct interaction and limit its accessibility and/or activity. To address this hypothesis, we generated HEK293T cells expressing a tagged version of human DICER that could be immunoprecipitated in mock or SINV-infected cells in order to perform a proteomic analysis of its interactome. Among the proteins co-immunoprecipitated with DICER and that were specifically enriched upon infection, we identified dsRBPs such as ADAR1, DHX9, PACT and PKR. We further validated the direct interaction between DICER and PKR upon SINV infection. We also demonstrated that the interactions of the endogenous DICER with PKR, PACT and DHX9 could also be detected in SINV-infected, but not mock-infected, HCT116 cells. We dissected the protein domains necessary for this interaction and we found that DICER helicase domain plays a fundamental role as a recruitment platform for PKR but also for other co-factors. Finally, we also show that expression of a helicase-truncated version of DICER has a negative effect on SINV infection. Importantly, this antiviral phenotype is independent of RNAi, but requires the presence of PKR. Our results indicate that DICER interactome is highly dynamic and directly link components of RNAi and IFN pathways in modulating the cellular response to viral infection.
Results
Establishment of a HEK293T cell line expressing FLAG-HA tagged DICER
In order to be able to study the interactome of the human DICER protein during viral infection, we transduced Dicer knock-out HEK293T cells (NoDice 2.20) [ 37 ] with either a lentiviral construct expressing a FLAG-HA-tagged wild type DICER protein (FHA:DICER WT #4) or a construct without insert as a negative control (FHA:ctrl #1). After monoclonal selection of stably transduced cells, we first characterized one clone of both FHA:DICER WT and of the FHA:ctrl cell lines. We first confirmed that the expression of the tagged version of DICER restored the miRNA biogenesis defect observed in the NoDice cells ( S1A Fig ). We then monitored the phenotype of these cells during SINV infection by using as a readout of viral infection the modified version of SINV able to express GFP from a duplicated sub-genomic promoter (SINV-GFP) [ 38 ]. At 24 hours post-infection (hpi) and a multiplicity of infection (MOI) of 0.02, the GFP fluorescence observed in FHA:DICER WT #4 cells and HEK293T cells was similar. However, the NoDice FHA:ctrl #1 cells displayed a decrease in GFP signal ( Fig 1A ). Western blot analysis of GFP expression confirmed the observations by epifluorescence microscopy, i . e . a significantly lower accumulation of GFP in the absence of the DICER protein ( Fig 1B ). We therefore wished to confirm the effect of DICER loss on SINV-GFP infection in another NoDice cell line, i . e . the NoDice clone 4.25 [ 39 ], and in another clone of the NoDice 2.20 FHA:ctrl cells (NoDice FHA:ctrl #2). We observed a similar decrease of SINV-GFP infection in NoDice 2.20 cells and two independent NoDice FHA:ctrl clones compared to HEK293T cells as shown by GFP microscopy ( S1B Fig ), by titration of the virus ( S1C Fig ) and by western blot analysis ( S1D Fig ). However, the independent NoDice 4.25 Dicer knock-out clone appeared mostly unaffected compared to HEK293T cells in term of GFP accumulation and viral titer ( S1B, S1C and S1D Fig ). This suggests that, despite the observed slight effect on SINV-GFP in NoDice 2.20 cells ( Fig 1 ), DICER proviral effect is not reproductible in an independent clone and therefore could not be generalized.
10.1371/journal.ppat.1009549.g001
Fig 1
Analysis of SINV infection in HEK293T cells and characterization of FHA:DICER WT cell lines.
A. GFP fluorescent microscopy pictures of HEK293T, NoDice FHA:ctrl #1 and FHA:DICER cell lines infected (polyclonal and two clones, #4 and #17) with SINV-GFP at an MOI of 0.02 for 24 h. The left panel corresponds to GFP signal from infected cells and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with a 5x magnification. hpi: hours post-infection. B. Western blot analysis of DICER (DICER and HA) and GFP expression in SINV-GFP-infected HEK293T, NoDice FHA:ctrl #1 and FHA:DICER cell lines shown in A. Gamma-Tubulin was used as loading control. C. Mean (+/- SEM) of SINV-GFP viral titers in the same cell lines as in A infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. ns: non-significant, ordinary one-way ANOVA test with Bonferroni correction. D. Western blot analysis of DICER (DICER and HA) and AGO2 expression in HEK293T, NoDice FHA:ctrl #1 and FHA:DICER cell lines. Gamma-Tubulin was used as loading control.
In order to evaluate whether different expression levels of DICER in a NoDice background could rescue the SINV infection phenotype observed in HEK293T cells, we also infected both the FHA:DICER WT polyclonal and an independent FHA:DICER WT clone (FHA:DICER WT #17) with SINV-GFP ( Fig 1A, 1C and 1D ). We confirmed that the GFP fluorescence observed by microscopy ( Fig 1A ), as well as the viral titers and the GFP protein accumulation ( Fig 1C and 1D ) in all tested FHA:DICER lines were comparable to the ones observed in HEK293T cells. Moreover, there was no striking difference in AGO2 expression between the FHA:DICER lines ( Fig 1D ).
Altogether, these results indicate that the FHA-tagged DICER protein can functionally complement the lack of DICER in terms of miRNA biogenesis ( S1A Fig ) and can therefore be used for proteomics studies. Moreover, because we could not observe significant differences in terms of SINV infection ( Fig 1 ) between the different FHA:DICER clones tested, we decided to select one line, namely FHA:DICER WT #4, for further analysis.
Analysis of DICER interactome during SINV infection by mass spectrometry
Our molecular tool being validated, we then focused on determining the interactome of FHA:DICER during SINV infection. We wanted to look at DICER interactome at an early infection time point to isolate cellular factors that could potentially modulate either DICER accessibility or its effect on viral dsRNA. As SINV replicates quickly upon cellular entry, we chose to set up the infection conditions to a duration of 6 hours at an MOI of 2.
We performed an anti-HA immunoprecipitation experiment (HA IP) coupled to label-free LC-MS/MS analysis in FHA:DICER WT #4 cells either mock-infected or infected for 6 h at an MOI of 2 with SINV-GFP. In parallel, we performed an anti-MYC immunoprecipitation as a negative control (CTL IP). The experiments were performed in technical triplicate in order to have statistically reproducible data for the differential analysis, which was performed using spectral counts. Prior to the detailed analysis of the results, we verified that there was no confounding factor in the experimentation by performing a Principal Component Analysis (PCA). This allowed us to see that the replicates were very homogenous and that the different samples were well separated based on the conditions.
To check the specificity of the HA immunoprecipitation, we first compared the proteins identified in the HA IP with the ones identified in the CTL IP in mock-infected cells. Differential expression analysis allowed us to calculate a fold change and an adjusted p-value for each protein identified and to generate a volcano plot representing the differences between HA and CTL IP samples. Applying a fold change threshold of 2 (abs(LogFC)>1)), an adjusted p-value threshold of 0.05 and a cutoff of at least 5 spectral counts in the most abundant condition, we identified 258 proteins differentially immunoprecipitated between the two conditions out of 1318 proteins ( Fig 2A and S1 Table ). Among these, 123 proteins were specifically enriched in the HA IP. The most enriched protein was DICER, followed by its known co-factors TRBP and PACT (also known as PRKRA) [ 13 , 17 ]. We were also able to retrieve AGO2, indicating that the RISC loading complex was immunoprecipitated and that proteins retrieved in our HA IP are specific to DICER immunoprecipitation.
10.1371/journal.ppat.1009549.g002
Fig 2
LC-MS/MS analysis of DICER interactome during SINV infection.
A. Volcano plot for differentially expressed proteins (DEPs) between HA IP and CTL IP in FHA:DICER mock-infected cells. Each protein is marked as a dot; proteins that are significantly up-regulated in HA IP are shown in red, up-regulated proteins in CTL IP are shown in blue, and non-significant proteins are in black. The horizontal line denotes a p-value of 0.05 and the vertical lines the Log2 fold change cutoff (-1 and 1). DICER and its cofactors (TRBP, PACT, AGO2) are highlighted in yellow. B. Volcano plot for DEPs between SINV-GFP (MOI of 2, 6 hpi) and mock fractions of HA IP in FHA:DICER cells. Same colour code and thresholds as in A have been applied. Proteins that are discussed in the text are highlighted in yellow and SINV proteins in purple. C. Gene Ontology (GO) term enrichment of proteins up-regulated in SINV-GFP fraction of HA IP using Enrichr software [ 89 , 90 ]. The graph displays the GO term hierarchy within the "biological process" branch sorted by p-value ranking computed from the Fisher exact test. The length of each bar represents the significance of that specific term. In addition, the brighter the colour is, the more significant that term is. Viral proteins have been excluded for this analysis. D. STRING interaction network of the top 100 proteins enriched in SINV-infected vs. mock-infected cells. Proteins involved in RNA metabolic processes or the regulation thereof are indicated in red and blue respectively, proteins with a known dsRNA binding function are indicated in green. DICER is indicated by a red circle. E. Summary of the differential expression analysis of SINV-GFP vs mock fractions from HA IP in FHA:DICER cells. The analysis has been performed using a generalized linear model of a negative-binomial distribution and p-values were corrected for multiple testing using the Benjamini-Hochberg method.
We next performed the differential expression analysis of proteins retrieved in the HA IP in SINV-GFP compared to mock-infected cells. Among 1342 proteins, 296 were differentially retrieved between conditions ( Fig 2B and S2 Table ). Of these, 184 proteins, including viral ones, were at least 2-fold enriched in SINV-GFP-infected cells. GO-term analysis showed a significant enrichment in RNA binding proteins including double-stranded RNA binding proteins and RNA helicases ( Fig 2C ). We then generated a functional protein association network using STRING on the top 100 proteins enriched in SINV-infected compared to mock-infected cells ( Fig 2D ). The resulting STRING network confirmed that a limited number of these proteins are known to be interacting with DICER, but that they are all engaged in other complexes ( e . g . DHX9, DDX18) that could partly explain the presence of some candidates in the mass spectrometry data. In addition, a large number of these proteins are involved in RNA metabolic processes ( Fig 2D , in red), or in their regulation ( Fig 2D , in blue), while a whole cluster is composed of dsRNA binding proteins ( Fig 2D , in green). Among the RNA binding proteins retrieved, the top and most specific DICER interactor is the interferon-induced, double-stranded (ds) RNA-activated protein kinase PKR (also known as E2AK2), which is enriched more than 250 times in virus-infected cells ( Fig 2B and 2E ). We were also able to identify the dsRNA-specific adenosine deaminase protein ADAR-1 (also known as DSRAD), as well as PACT, which were enriched 5.9 and 4.2 times respectively in SINV-GFP-infected cells compared to mock-infected cells ( Fig 2B and 2E ). Among the isolated RNA helicases, we identified the ATP-dependent RNA helicase A protein DHX9, which is implicated in Alu element-derived dsRNA regulation and in RISC loading [ 40 , 41 ]. In order to verify if the observed interactions were specific to SINV we performed the same experiments with another virus of the Togaviridae family, the Semliki forest virus (SFV). In this analysis, we were able to retrieve ADAR-1, DHX9, PACT and PKR, specifically enriched in SFV-infected samples ( S2 Fig and S3 and S4 Tables). These results show that these interactions can be retrieved in Togaviridae -infected cells.
Taken together, our data indicate that several proteins interacting with DICER in virus-infected cells are involved in dsRNA sensing and/or interferon-induced antiviral response.
DICER and PKR interact in vivo in the cytoplasm during SINV infection
To validate the LC-MS/MS analysis, we performed a co-immunoprecipitation (co-IP) followed by western blot analysis in FHA:DICER WT #4 cells infected with SINV-GFP at an MOI of 2 for 6 h. Whereas TRBP interacted equally well with FHA:DICER in mock and SINV-GFP-infected cells, ADAR-1, PKR, DHX9 and PACT were only retrieved in the HA IP in SINV-GFP-infected cells ( Fig 3A ). We verified that these interactions could also be observed at a later time post-infection by performing the HA IP in FHA:DICER WT #4 cells infected with SINV-GFP for 24 h at an MOI of 0.02. This indicates that the specific interactions between DICER and ADAR-1, DHX9, PACT or PKR occur at an early stage of the SINV infection and remain stable in time in virus-infected cells ( S3A Fig ).
10.1371/journal.ppat.1009549.g003
Fig 3
Confirmation of LC-MS/MS analysis by co-IP and BiFC.
A. Western blot analysis of HA co-IP in mock or SINV-GFP-infected (MOI of 2, 6 hpi) FHA:DICER WT #4 cells. Proteins associated to FHA:DICER were revealed by using antibodies targeting endogenous ADAR-1, PKR, TRBP, DHX9 or PACT proteins. In parallel, an HA antibody was used to verify the IP efficiency and GFP antibody was used to verify the infection. Ponceau was used as loading control. B. Western blot analysis of HA co-IP in mock or SINV-GFP-infected (MOI of 2, 6 hpi) FHA:DICER WT #4 cells. The lysate was treated or not with RNase A/T1. Proteins associated to FHA:DICER were revealed by using antibodies targeting endogenous DHX9, p-PKR, PKR, TRBP, or PACT proteins. In parallel, an HA antibody was used to verify the IP efficiency and GFP antibody was used to verify the infection. Ponceau was used as loading control. C. Western blot analysis to validate the interaction of PKR with DICER (upper panel) and PACT (lower panel) in mock or SINV-GFP-infected HEK293T cells (MOI of 2, 6 hpi). Immunoprecipitated proteins obtained from PKR pulldowns were compared to rabbit IgG pulldowns to verify the specificity of the assay. D. Interactions between DICER and TRBP, PACT or PKR were visualized by BiFC. Plasmids expressing N-ter Venus:DICER and TRBP:, PACT: or PKR:Venus C-ter were co-transfected in NoDiceΔPKR cells for 24 h and cells were either infected with SINV at an MOI of 2 for 6 h or not. The different combinations are indicated on the left side. Reconstitution of Venus (BiFC) signal was observed under epifluorescence microscope. For each condition, the left panel corresponds to Venus signal and the right panel to the corresponding brightfield pictures. Scale bar: 100 μm. hpi: hours post-infection. E. BiFC experiment on fixed NoDiceΔPKR cells treated as in D. After fixation, cells were stained with DAPI and observed under confocal microscope. Only a merge picture of BiFC and DAPI signals of SINV-infected cells is shown here. A higher magnification of picture showing cytoplasmic localization of the interaction represented by a red square is shown in the bottom left corner. Scale bars: 20 μm and 10 μm.
In order to verify whether these interactions were mediated by RNA, we performed an anti-HA co-IP experiment on an RNase A/T1 treated total extract from FHA:DICER WT #4 cells infected with SINV-GFP at an MOI of 2 for 6 h. Since the RNase treatment was performed at relatively low salt concentration (140 mM NaCl), RNase A should cleave dsRNA [ 42 , 43 ] and we should therefore assess both ss and dsRNA-dependency in these conditions. We confirmed the efficiency of the RNase treatment by ethidium bromide staining visualisation of total RNA on an agarose gel ( S3B Fig ). TRBP equally interacted with FHA:DICER, with or without RNase treatment, in mock and SINV-GFP-infected cells ( Fig 3B ). Instead, the virus-induced interactions between DICER and PKR or PACT upon SINV-GFP infection were almost totally lost in the RNase-treated samples. Upon virus infection, PKR is phosphorylated to be activated and exert its antiviral function [ 4 ]. Using an antibody targeting the phosphorylated form of PKR (p-PKR), we looked for p-PKR before and after RNase treatment. The virus-enriched interactions between DICER and p-PKR or DHX9 were completely lost upon RNase treatment. These results therefore indicate that RNA molecules (either single- or double-stranded) facilitate DICER interaction with DHX9, PACT and PKR and its active form, although the complex may also partially interact in an RNA-independent manner.
Because of the involvement of PKR in antiviral response [ 44 ] and the fact that it shares common co-factors with DICER, namely TRBP and PACT [ 45 , 46 ], we decided to focus our analysis on the DICER-PKR interaction. To confirm the biological relevance of this interaction, we first performed a reverse co-IP to immunoprecipitate the endogenous PKR protein in HEK293T cells infected or not with SINV-GFP. While PACT interacted with PKR both in mock and in SINV-GFP-infected cells as expected ( Fig 3C ), DICER co-immunoprecipitated with the endogenous PKR only in virus-infected cells thereby confirming the specificity of the interaction between the two proteins ( Fig 3C ).
To further determine whether DICER and PKR could directly interact in vivo , we set up a bi-molecular fluorescent complementation assay (BiFC) experiment [ 47 ]. To this end, we fused the N- or C-terminal half of the Venus protein ( N-ter Venus or C-ter Venus) to DICER and to PKR but also to TRBP and PACT. Since we showed above that an N-terminally tagged DICER was functional, we fused the Venus fragments at the N-terminal end of DICER. For the other three proteins, we fused the Venus fragments at the N- or C-terminus and selected the best combination. To avoid interaction with the endogenous DICER and PKR proteins, we conducted all BiFC experiments in NoDiceΔPKR HEK293T cells [ 33 ]. In order to control the BiFC experiments, we chose to exploit the well characterized DICER-TRBP interaction, which is known to occur via the DICER DEAD-box helicase domain [ 14 ]. We therefore used the wild-type DICER protein as a positive control and a truncated version of DICER protein lacking part of this helicase domain and called DICER N1 [ 33 ] as a negative control ( S3C Fig ). We first confirmed the expression of the tagged proteins by western blot analysis ( S3D Fig ) and then, we tested the interactions between DICER and TRBP or PACT or PKR. We co-transfected the Venus constructs for 24 h and then infected cells with SINV or not for 6 h at a MOI of 2. A comparable fluorescent signal was observed both in mock- and SINV-infected cells when N-ter Venus:DICER was co-transfected with either PACT or TRBP fusion construct ( Fig 3D ). Although we initially expected an increase of the Venus fluorescence in SINV-infected cells, overall we observed a similar signal for the DICER-PKR interaction both in mock- and SINV-infected cells, probably due to the fact that both proteins are transiently overexpressed in this experiment. The same holds true for the DICER-PACT interaction that can also be seen both in mock- and SINV-infected cells.
As a control and to rule out any aspecific interactions between the different proteins tested, we also monitored the DICER-N1-TRBP interaction by BiFC. As expected, no fluorescent signal was observed in cells co-transfected with N-ter Venus:DICER N1 and TRBP:Venus C-ter ( S3E Fig ), confirming that DICER helicase domain is required for its interaction with TRBP [ 14 ] and validating the specificity of the BiFC approach.
To further confirm that the absence of PKR did not influence the interactions of TRBP or PACT with DICER, we also performed a BiFC analysis in HEK293T cells. After verifying that in this context as well, fusion proteins were expressed as expected ( S3F Fig ), we observed that the results were similar as in NoDiceΔPKR cells ( S3G Fig ).
To gain more insight into the subcellular localization of these interactions during SINV infection, we performed the BiFC experiments, fixed the cells and observed them under a confocal microscope. We observed a cytoplasmic fluorescent signal for DICER-TRBP and DICER-PACT interactions ( Fig 3E upper and middle panels), which is in agreement with their canonical localization for the maturation of miRNAs [ 10 , 48 ]. Similarly, co-transfection of DICER and PKR led to a strong Venus signal homogeneously distributed in the cytoplasm ( Fig 3E lower panel).
Collectively, these results formally confirm that DICER interacts with several RNA helicases and dsRNA-binding proteins in virus-infected cells, among which PKR, and that for the latter this interaction occurs in the cytoplasm.
DICER interactome changes upon SINV infection are not cell-type specific
To further validate our DICER interactome results and generalize them to another biological system, we performed co-IP experiments on the endogenous DICER in a different cell type. To this end, the FLAG-HA-GFP tag was knocked into (KI) the Dicer locus in human colon carcinoma cells (HCT116) by CRISPR-Cas9-mediated homologous recombination ( S4A, S4B and S4C Fig ). A guide RNA (gRNA) targeting the region corresponding to Dicer ATG and a DNA template for homologous recombination bearing the FLAG-HA-GFP sequence surrounded by the upstream and downstream arms of Dicer were used to generate the resulting cell line referred to as HCT116 KI-DICER cells. The expected insertion of the tag in one of the two Dicer alleles was assessed by PCR amplification and Sanger sequencing ( S4A, S4B and S4C Fig ). In agreement, we could detect two bands for DICER protein by western blot in the HCT116 KI-DICER cells, which confirmed that this cell line is heterozygous ( Fig 4A ).
10.1371/journal.ppat.1009549.g004
Fig 4
Confirmation of DICER interactome upon SINV infection in HCT116 KI-DICER cells.
A. Western blot analysis of DICER, AGO2, PKR and TRBP expression in HEK293T, HCT116 and HCT116 KI-DICER cell lines. Gamma-Tubulin and ponceau were used as loading controls. B . GFP fluorescent microscopy pictures of HEK293T, HCT116 and HCT116 KI-DICER cell lines infected with SINV-GFP at an MOI of 0.02, 0.1 and 1 for 24 h. The left panel corresponds to GFP signal from infected cells and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with a 5x magnification. C . miR-16 expression analyzed by northern blot in the same cell lines as in B. Expression of snRNA U6 was used as loading control. D. Western blot analysis of HA co-IP in mock or SINV-GFP-infected (MOI of 0.1, 24 hpi) HCT116 KI-DICER cells. Proteins associated to FHA-GFP:DICER were revealed by using antibodies targeting endogenous DHX9, p-PKR, PKR, PACT or TRBP proteins. The TRBP immunoblot was performed by loading the same samples on a separate membrane. In parallel, an HA antibody was used to verify the IP efficiency and GFP antibody was used to verify the infection. Ponceau was used as a loading control.
We additionally verified the expression of specific DICER-interacting proteins, such as AGO2, PKR or TRBP, in HCT116 KI-DICER cells compared to the parental HCT116 cells and to HEK293T cells ( Fig 4A ). We also measured the production of mature miRNAs, such as miR-16, by northern blot analysis and confirmed that miRNA expression is maintained in HCT116 KI-DICER cells ( Fig 4B ). Of note, the GFP inserted at the Dicer locus could not be detected by epifluorescence microscopy in the HCT116 KI-DICER cells, which probably reflects the low abundance of the DICER protein.
We then determined whether SINV-GFP infection was comparable in HCT116 cells and HEK293T cells. We infected HCT116, HCT116 KI-DICER and HEK293T cells with SINV-GFP at three different MOI (0.02, 0.1 and 1) and measured GFP fluorescence by microscopy at 24 hpi ( Fig 4C ). Both HCT116 and HCT116 KI-DICER cells expressed GFP upon infection with SINV-GFP, although with a lower intensity than HEK293T cells. We also verified by western blot analysis the accumulation of GFP and the phosphorylation of both PKR and eIF2α upon SINV-GFP infection of HCT116 KI-DICER and HEK293T cells ( S4D Fig ) and chose as optimal SINV-GFP condition of infection in HCT116 KI-DICER cells the MOI of 0.1 for 24 h.
To validate the DICER interactions observed in HEK293T FHA:DICER cells, we then performed anti-HA co-IP experiments followed by western blot analysis in HCT116 KI-DICER cells infected or not with SINV-GFP. We successfully retrieved TRBP interacting with DICER in both mock and infected cells, whereas DHX9, PKR (phosphorylated or not) and PACT were only retrieved in the HA IP in infected cells ( Fig 4D ). These results not only confirm that the endogenous DICER specifically interacts with DHX9, PACT and PKR upon SINV infection, but also that these interactions are not restricted to one specific cell type.
The helicase domain of DICER is required for its interaction with PKR
Even though DICER and PKR are likely brought together by RNA, specific protein domains might be involved in stabilizing the complex. Therefore, we next determined the domain of DICER required for its interaction with PKR. Since its helicase domain was previously shown to be involved in the interaction with TRBP and PACT [ 14 , 17 ], we speculated that it could also be implicated in binding PKR. To test this hypothesis, we cloned several versions of DICER proteins wholly or partly deleted of the helicase domain ( Fig 5A DICER N1 and N3). In addition, we also cloned the helicase domain alone ( Fig 5A DICER Hel.) and a DICER variant deleted of its C-terminal dsRNA binding domain ( Fig 5A DICER ΔdsRBD) since this domain could also be involved in protein-protein interaction [ 49 , 50 ]. We then transfected the different versions of DICER WT and the deletion mutant constructs in NoDice cells. In mock and SINV-GFP infected cells, whole cell extracts were subjected to anti-HA and anti-MYC (CTL) IP. TRBP was retrieved in both conditions with DICER WT, Hel. and ΔdsRBD ( Fig 5B and 5C ). In mock cells, PACT and PKR were only found weakly interacting with DICER WT ( Fig 5B ). In SINV-infected cells, we observed that similar to TRBP and to a lesser extent PACT, N1 and N3 mutations strongly reduced the binding of DICER with PKR ( Fig 5C lanes 2–3 and 7–8). Importantly, we also noted that the helicase domain alone could bind PKR, TRBP and PACT ( Fig 5C lanes 4 and 9). Moreover, the deletion of the dsRNA binding domain of DICER did not affect its interaction with TRBP, PACT and PKR ( Fig 5C lanes 5 and 10). We also looked for p-PKR in our co-IP ( Fig 5C panel p-PKR). We noticed that only WT DICER and its helicase domain were able to interact with p-PKR ( Fig 5C lanes 1&6 and 4&9). The fact that DICER ΔdsRBD did not interact with p-PKR ( Fig 5C lanes 5&10) is striking but could indicate that the phosphorylation of PKR may induce conformational changes preventing its interaction with some domains of DICER. These results reveal that, like for TRBP and PACT, the helicase domain of DICER is required for DICER-PKR/p-PKR interaction during SINV infection.
10.1371/journal.ppat.1009549.g005
Fig 5
Identification of DICER domains involved in DICER-PKR interaction.
A. Schematic representation of Human DICER proteins used in this study. The different conserved domains are shown in colored boxes. DUF283: Domain of Unknown Function; PAZ: PIWI ARGONAUTE ZWILLE domain; dsRBD: dsRNA-binding domain. hDICER WT is the full-length protein. hDICER N1 is deleted of the first N-terminal 495 amino acids. hDICER N3 is wholly deleted of the helicase domain. hDICER Hel. is the whole DICER’s helicase domain. hDICER ΔdsRBD is deleted of the C-terminal dsRBD. B. Western blot analysis of HA co-IP in mock NoDice 2.20 cells transfected with different versions of FHA:DICER proteins. Efficiency of immunoprecipitation was assessed using anti-HA and anti-DICER antibodies and co-IPs of TRBP, PKR and PACT were examined using appropriate antibodies. Expression of GFP in INPUT fraction was visualized as control of SINV-GFP infection. Ponceau staining of membranes is used as loading control. C. Western blot analysis of HA co-IP in NoDice 2.20 cells transfected with different versions of FHA:DICER proteins and infected with SINV-GFP (MOI of 2, 6 hpi). Efficiency of immunoprecipitation was assessed using an anti-Flag antibody and co-IPs of PKR, TRBP, p-PKR and PACT were examined using appropriate antibodies. Expression of GFP in INPUT fraction was visualized as control of SINV-GFP infection. Ponceau staining of membranes is used as loading control. The DICER Hel. band is indicated by a red asterisk. D. Plasmids expressing the different versions of DICER proteins fused to the N-terminal part of Venus and PKR:Venus C-ter plasmid were co-transfected in NoDiceΔPKR cells. Cells were treated as in Fig 3D . The different combinations are noted on the left side. The fluorescent signal was observed using an epifluorescence microscope. For each condition, the left panel corresponds to Venus signal and the right panel to the corresponding brightfield pictures. Scale bar: 100 μm. hpi: hours post-infection.
In order to confirm these co-IP experiments, we next decided to perform BiFC experiments using the same conditions as previously. In both mock and SINV-infected cells, only the combinations of DICER WT-PKR and DICER ΔdsRBD-PKR showed a strong Venus signal, while neither DICER N1 nor N3 constructs revealed an interaction with PKR ( Fig 5D ). In contrast, the DICER Hel. construct did not seem to interact with PKR in mock-infected cells but appeared to do so in SINV-infected cells as a faint Venus signal could be observed. These results therefore confirmed the co-IP observations for the DICER-PKR interaction. In addition, we also performed a BiFC experiment using the different DICER constructs with TRBP or PACT. Altogether, the BiFC results mostly fitted with the co-IP experiments for the DICER-TRBP ( S5A Fig ) and DICER-PACT ( S5B Fig ) interactions. TRBP indeed did not seem to interact with the DICER N1 and only slightly with the DICER N3. However, PACT interaction was lost with DICER N1, but not with DICER N3 in mock- and SINV-infected cells ( S5B Fig third panel). This result may be explained by the fact that DICER interacts with PACT via the helicase and DUF domains, whereas only the DICER helicase domain is required for its interaction with TRBP [ 14 , 17 ]. In agreement, the Venus signal observed between the DICER Hel. and PACT seemed weaker than the one we observed with TRBP ( S5A and S5B Fig fourth panels).
Taken together these results indicate that DICER interacts with both PKR and its phosphorylated form during SINV infection, and that this interaction requires the helicase domain of DICER.
Functional importance of DICER helicase domain during SINV infection
We then sought to study the functional role of DICER-PKR interaction during viral infection. For this purpose, we decided to use DICER helicase deletion mutants to study SINV infection. To do so, we first generated NoDice HEK293T cells stably expressing FHA-tagged DICER N1 (FHA:DICER N1) by lentiviral transduction. As for the FHA:DICER WT cell line, we first selected a clone expressing the tagged DICER N1 at a level similar to the endogenous DICER protein in HEK293T cells ( Fig 6A ). DICER N1 protein has been shown to still be able to produce miRNAs [ 33 ]. We thus verified by northern blot analysis that DICER N1 is indeed able to process miRNAs similarly to WT DICER in HEK293T and FHA:DICER cells, thereby validating the functionality of the tagged protein ( Fig 6B ). We next infected HEK293T, FHA:DICER WT #4 and FHA:DICER N1 #6 cells with SINV-GFP and measured virus accumulation by assessing GFP expression by microscopy analysis. Interestingly, the GFP protein level was drastically reduced in FHA:DICER N1 #6 cells compared to FHA:DICER WT #4 and HEK293T cells ( Fig 6C ). Encouraged by this observation, we decided to infect with SINV-GFP additional DICER deletion mutants, namely N3 and Hel. We generated stable cell lines for these various mutants by lentiviral transduction in the NoDice 2.20 background and infected those cells with SINV-GFP at an MOI of 0.02 for 24 h. We verified by western blot analysis that all selected DICER mutant clones, namely N1 #6, N3 #2.13 and Hel. #2.6, expressed the tagged protein at the expected size and at levels mostly similar to the FHA:DICER WT #4 cell line ( Fig 6D , first two panels). We also verified the DICER mutants contribution to the endogenous miRNA biogenesis by performing a northern blot analysis on miR-16 accumulation ( S6A Fig ).
10.1371/journal.ppat.1009549.g006
Fig 6
Analysis of the importance of Dicer helicase domain on SINV-GFP infection in FHA:DICER mutant stable cell lines.
A. Expression of DICER (DICER, HA), TRBP and AGO2 was analysed by western blot in HEK293T, NoDice FHA:ctrl #1, FHA:DICER WT #4 and FHA:DICER N1 #6 cell lines. Gamma-Tubulin was used as loading control. B. Northern blot analysis of miR-16 expression in the same samples as in A. Expression of snRNA U6 was used as loading control. C. Representative GFP fluorescent microscopy images of HEK293T, FHA:DICER WT #4 and FHA:DICER N1 #6 cell lines infected with SINV-GFP at an MOI of 0.02 for 24 h. The left panel corresponds to GFP signal and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with a 5x magnification. hpi: hours post-infection. D. Western blot analysis of DICER (DICER and HA), AGO2, PKR, and GFP expression in SINV-GFP-infected cells in the same condition as in C. Gamma-Tubulin was used as loading control. The asterisk correspond to aspecific bands E. Mean (+/- SEM) of SINV-GFP viral titers fold change over HEK293T cells in HEK293T, NoDice 2.20, FHA:DICER WT #4 and FHA:DICER mutants cell lines infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. * p < 0.05, ns: non-significant, ordinary one-way ANOVA test with Bonferroni correction.
We additionally verified the impact of these DICER mutants on SINV-GFP infection by measuring the GFP intensity of fluorescence by microscopy ( S6B Fig ). Our results indicate that GFP accumulation is similar in HEK293T, NoDice 2.20, FHA:DICER WT, Hel. and ctrl cells. However, almost no fluorescence was detected in FHA:DICER N1 #6 and N3 #2.13 cells compared to HEK293T cells ( S6B Fig ). The reduction of virus-encoded GFP accumulation and viral production were confirmed by western blot ( Fig 6D ) and by plaque assay, respectively (Figs 6E and S6C ).
Altogether, these results therefore indicate that expressing a helicase truncated version of DICER, which is unable to interact with PKR, appears to confer an antiviral phenotype against SINV infection.
The antiviral phenotype of the helicase-truncated DICER mutants is independent of AGO2
We finally carried out a functional analysis of the helicase-domain-truncated DICER N1 and N3 mutants to investigate the mechanism of the antiviral phenotype. First, to investigate a potential implication of the RNAi pathway, we performed a knock-down of the AGO2 protein prior to the infection of NoDice cells expressing either WT, N1 or N3 FHA:DICER. AGO2 is the main effector protein in RNA silencing pathways [ 51 ] and has been previously shown to be a crucial antiviral RNAi factor against Influenza A virus in mouse embryonic fibroblasts (MEFs) [ 52 ]. We transfected either control siRNAs, or siRNAs targeting AGO2 for 48 h in NoDice cells stably expressing either an empty vector (FHA:ctrl #2) or WT, N1 or N3 FHA:DICER constructs. Cells were then infected with SINV-GFP at an MOI of 0.02 for 24 h, and virus accumulation was first assessed by looking at GFP expression by microscopy analysis ( Fig 7A ). In all cell lines, no major difference in GFP fluorescence could be observed when comparing cells transfected with the control siRNA or AGO2-specific siRNAs. We verified the knock-down efficiency by western blot analysis and confirmed the microscopy observation by measuring GFP protein accumulation ( Fig 7B ). Finally, we measured virus accumulation by plaque assay, and we observed that the antiviral phenotype was clearly visible in FHA:DICER N1 #6 and FHA:DICER N3 #2.13 cell lines but was not complemented upon AGO2 knock-down ( Fig 7C ).
10.1371/journal.ppat.1009549.g007
Fig 7
The antiviral effect of helicase-deleted DICER mutants is independent of AGO2.
A. GFP fluorescent microscopy pictures of NoDice FHA:ctrl #2, NoDice FHA:DICER WT #4 and NoDice FHA:DICER mutant cell lines treated with two doses of siAGO2 at 20 nM for 48 hours before a 24-hour-SINV-GFP infection at an MOI of 0.02. The left panel corresponds to GFP signal from infected cells and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with 5x magnification. hpi: hours post-infection. B. Western blot analysis of DICER, AGO2 and GFP expression in SINV-GFP-infected NoDice FHA:ctrl #2, NoDice FHA:DICER WT #4 and NoDice FHA:DICER mutant cell lines shown in A. Cells were treated with two doses of siAGO2 at 20 nM for 48 hours before a 24-hour-SINV-GFP infection at an MOI of 0.02. Gamma-Tubulin was used as loading control. C. Mean (+/-SEM) of SINV-GFP viral titers in the same cell lines as in A. infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. ns: non-significant, two-tailed unpaired parametric t-test.
Altogether, these results indicate that the antiviral phenotype against SINV observed in cells expressing helicase-truncated mutant DICER proteins does not depend on the presence of AGO2, thereby ruling out an involvement of RNAi.
The antiviral phenotype due to the DICER helicase-domain deletion requires PKR
In order to determine the functional role of the PKR-DICER interaction in the antiviral response to SINV, we generated NoDiceΔPKR cells stably expressing either the full length FHA:DICER WT or the helicase deletion mutants FHA:DICER N1 or N3, or the empty vector as a control (FHA:ctrl) by lentiviral transduction. After monoclonal selection of each cell line, we infected them with SINV-GFP at an MOI of 0.02 and assessed virus accumulation by looking at GFP fluorescence by microscopy analysis ( Fig 8A ). As expected, an increase in GFP fluorescence was observed in NoDiceΔPKR FHA:ctrl cells compared to HEK293T cells at 24 hpi. In contrast we could not observe any difference in GFP fluorescence between NoDiceΔPKR FHA:ctrl cells and those expressing FHA:DICER WT, FHA:DICER N1 or N3 proteins. To verify whether any significant difference in terms of virus accumulation could be observed in NoDiceΔPKR cells expressing WT or helicase truncated DICER proteins, we measured GFP protein levels by western blot analysis ( Fig 8B ) and virus production by plaque assay ( Fig 8C ). As opposed to the observations done in NoDice cells expressing PKR ( Fig 6 ), both GFP accumulation and viral titers remained unchanged between NoDiceΔPKR FHA:ctrl cells and those expressing FHA:DICER WT, N1 or N3 constructs. Taken together, these results demonstrate that the antiviral phenotype of helicase-truncated DICER mutants depends on the presence of PKR. Therefore, our data suggest that the helicase domain of DICER sequesters PKR and when this interaction is lost, the antiviral effect of PKR is exacerbated, thereby explaining the phenotype observed in cells expressing helicase-truncated DICER mutants.
10.1371/journal.ppat.1009549.g008
Fig 8
The antiviral effect of helicase-deleted DICER mutants requires PKR.
A. GFP fluorescent microscopy pictures of HEK293T, NoDiceΔPKR:ctrl and FHA:DICER mutant cell lines infected with SINV-GFP at an MOI of 0.02 for 24 h. The left panel corresponds to GFP signal from infected cells and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with 5x magnification. hpi: hours post-infection. B. Western blot analysis of DICER (DICER and HA), AGO2, PKR and GFP expression in SINV-GFP-infected HEK293T, NoDiceΔPKR FHA:ctrl and FHA:DICER mutant cell lines shown in A. Gamma-Tubulin was used as loading control. C. Mean (+/-SEM) of SINV-GFP viral titers in the same cell lines as in A. infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. ns: non-significant, ordinary one-way ANOVA test with Bonferroni correction.
Discussion
The role of DICER in antiviral defense in human cells remains a topic of intense discussion [ 21 , 22 , 53 , 54 ]. In particular there have been contradictory reports regarding its capacity to produce siRNAs from viral RNAs [ 31 , 37 , 55 , 56 ]. These observations could be due to the fact that several mammalian viruses potentially encode VSR proteins, thereby masking the effect of RNAi [ 22 , 28 , 29 , 52 , 57 ]. Another putative but non-exclusive explanation could be that there is a mutual regulation of type I IFN and RNAi pathways [ 58 , 59 ]. Thus, it has already been shown that PACT can regulate MDA5 and RIG-I during virus infection and therefore the induction of type I IFN response [ 60 , 61 ]. To date, it is not clear whether the activity of the DICER protein as well could be regulated by potential interactors, or inversely whether it could itself modulate the activity of proteins involved in the IFN pathway. To answer this question, we determined the changes in the interactome of human DICER upon SINV and SFV infections. This analysis allowed us to reveal that a lot of proteins associating with DICER during viral infection are dsRNA-binding proteins and RNA helicases. A number of these proteins are known to be involved in antiviral defense pathways, thereby indicating the possible formation of one or several complexes between DICER and these proteins, which are very likely brought together by the accumulation of dsRNA during virus infection.
Among these proteins, we chose to focus on the well-known ISG PKR, which is involved in many cellular pathways such as apoptosis, cellular differentiation, development and antiviral defense [ 4 , 8 , 62 , 63 ]. PKR is one of the main actors of the Integrative Stress Response (ISR) in human cells, and its activation or inhibition needs to be tightly regulated in order to have a properly balanced response to stress. Our results indicate that DICER interacts via its helicase domain with PKR in the cytoplasm during SINV infection. The helicase domain of DICER, which is also required for its interaction with TRBP and PACT, belongs to the helicase superfamily 2, which is also found in RLRs such as RIG-I, MDA5 or LGP2 [ 64 , 65 ]. These proteins act as sensors of viral infection and through the activation of proteins such as MAVS, mediate the induction of type I IFN pathway [ 65 ]. We hypothesize that even though the human DICER helicase has evolved mainly to act in miRNA/siRNA pathways, it still retained the capacity to act as an RLR. However, as opposed to RIG-I and MDA5, our data suggest that DICER would act more as an inhibitor rather than inducer of the immune response. Therefore, we propose that this domain serves as a platform for the recruitment of different proteins to diversify the functions of DICER.
One such regulatory effect appears to be on the antiviral activity of PKR, as cells expressing a truncated form of DICER unable to interact with PKR become resistant to SINV infection. This is in agreement with previous observations that ectopic expression of the Drosophila DICER2 protein in human cells perturbs IFN signaling pathways and antagonizes PKR-mediated antiviral immunity [ 36 ]. Although the precise molecular mechanism involved will require further work to be fully deciphered, it seems that the two proteins are likely brought together via their interaction with RNA, most probably of viral origin. Indeed, we showed that the co-IP interaction was partially RNase sensitive. However, we confirmed that the interaction is not artificially created during the co-immunoprecipitation procedure, since we could show that DICER and PKR interact in BiFC assay, a technique that favors the detection of direct interactions [ 47 ]. Most of the time, the inhibition of PKR activity relies on its inhibition to bind to dsRNA or to auto-phosphorylate. For example, the human tRNA-dihydrouridine synthase 2 (hDus2) binds the first dsRBD of PKR and prevents its activation [ 66 ]. TRBP binds dsRNAs but also PKR directly hindering its dimerization. In normal condition, TRBP is also associated with PACT thus preventing PKR activation by PACT [ 67 – 70 ]. Since we showed that DICER can bind the activated phospho-PKR, we hypothesize that this interaction does not result in the inhibition of PKR autophosphorylation. In fact, in condition of infection with a high virus dose, we showed that phospho-PKR levels are similar in cells expressing DICER WT or helicase deletion mutants N1 and N3, but the activated PKR does not associate with these truncated versions of DICER. Therefore, one possibility could be that DICER interaction with PKR prevents the latter from acting upon some of its targets, which remain to be identified, to fine-tune the antiviral response.
As of now, we cannot formally rule out that the effect of DICER on PKR is mediated by other proteins. TRBP and PACT have been shown to regulate PKR activity, the former normally acting as a repressor and the latter as an activator [ 46 , 68 , 70 ]. Interestingly, in lymphocytic Jurkat cells infected by HIV-1, PACT can also act as a repressor of PKR [ 71 ]. It is thus tempting to speculate that these two proteins participate in the formation of the DICER-PKR complex. However, our results show that this may not necessarily be the case. Indeed, in the BiFC experiment, the DICER N3 mutant still interacted with PACT but not with PKR indicating that PACT binding is not sufficient to confer the association with PKR.
Besides PKR, other proteins were specifically enriched upon viral infection in the DICER IP. These are also interesting candidates to explain the putative regulatory role of DICER. Among these proteins, DHX9 and ADAR-1 are especially intriguing. DHX9, also known as RNA helicase A (RHA), associates with RISC, helping the RISC loading [ 41 ]. Moreover, DHX9 is directly involved in removing toxic dsRNAs from the cell to prevent their processing by DICER [ 40 ]. It has also been implicated in HIV-1 replication and knockdown of DXH9 leads to the production of less infectious HIV-1 virions [ 72 – 74 ]. Finally, DXH9 interacts with and is phosphorylated by PKR in MEFs. This phosphorylation precludes the association of DHX9 with RNA, thus inhibiting its proviral effect [ 75 ]. In light of these observations and ours, we can speculate that the inhibitory effect of DICER on PKR activity could also be linked to DHX9 phosphorylation. ADAR-1 is one of the well-known RNA-editing factors [ 76 ]. ADAR-1 is linked to both miRNA biogenesis [ 77 – 79 ] and virus infection. Indeed, ADAR-1 has an antiviral effect against Influenza virus, but most of the time, its depletion leads to a decrease of the viral titer, as was reported for VSV or HIV-1 [ 80 , 81 ]. It has been shown that ADAR-1 and PKR interact directly during HIV-1 infection. This interaction triggers the inhibition of PKR activation, and thus a reduction of eIF2α phosphorylation leading to an increase of virus replication [ 5 , 82 ]. Interestingly, over-expression of ADAR-1 enhances drastically the replication of the alphaviruses Chikungunya virus (CHIKV), and Venezuelan equine encephalitis virus (VEEV) most likely by interfering with the IFN induction [ 83 ].
One hypothesis to explain the virus resistance phenotype of the DICER N1 and N3 cell lines could be an increased processivity of these truncated proteins on long dsRNA substrates [ 33 ], which would render DICER RNAi proficient. However, our results are not in favor of this hypothesis, since we show that knocking-down AGO2 does not allow to make cells expressing DICER N1 or N3 more sensitive to SINV infection. AGO2 being the only slicer-proficient Argonaute protein expressed at physiological levels in HEK293T cells, we can confidently conclude that the observed phenotype is RNAi-independent.
Finally, we demonstrated that the phenotype of helicase-truncated DICER isoforms depends on PKR expression, because it was completely lost in PKR knockout cell lines. We therefore propose that, at least during infection with SINV, DICER prevents PKR to be fully active by interacting with and potentially sequestrating it. Deciphering the exact molecular mechanism at play will require additional studies in order to get the full picture. Nevertheless, by assessing the interactome of DICER during SINV infection, we have unveiled a new, PKR-dependent, role for the helicase domain of DICER in regulating the cellular response to viral infection.
Material and methods
Plasmids, cloning and mutagenesis
Plasmids used for BiFC experiments were a gift from Dr. Oliver Vugrek from the Ruđer Bošković Institute and described in [ 47 ]. The cDNAs of TRBP, PACT and PKR were respectively amplified from (pcDNA-TRBP Addgene #15666) [ 16 ], (pcDNA-PACT Addgene #15667) [ 16 ], (pSB819-PKR-hum Addgene #20030) [ 84 ], and cloned into the four pBiFC vectors by Gateway recombination. DICER N1, N3, Hel. and ΔdsRBD were generated by PCR mutagenesis from pDONR-DICER described in [ 36 ] and cloned into the four pBiFC and pDEST-FHA vectors by Gateway recombination. plenti6 FHA-V5 vector was modified from plenti6-V5 gateway vector (Thermo Fisher Scientific V49610) by Gibson cloning. DICER WT, N1, N3 and Hel. from pDONR plasmids were cloned into plenti6 FHA-V5 by Gateway recombination. All primers used are listed in S5 Table .
Cell lines
HEK293T, HEK293T/NoDice (2.20 and 4.25), and HEK293T/NoDiceΔPKR cell lines were a gift from Pr. Bryan Cullen and described in [ 33 , 39 ]. HCT116 cell line was a gift from Dr. Christian Gaiddon.
Generation of Flag-HA-GFP-DICER knock-in cell line by CRISPR/Cas9
To generate the knock-in cell line, the sequence of Flag-HA-GFP was amplified by PCR from the Flag-HA-GFP plasmid [ 85 ]. DNA sequences corresponding to 1 Kb upstream (left homology arm) and downstream (right homology arm) the starting codon (ATG) of DICER gene were amplified from HCT116 cell genomic DNA using primer pairs listed in S5 Table . The three PCR products were gel-purified and cloned into a linearized pUC19 by In-fusion cloning (Clontech) to obtain the template for homologous recombination (LarmDICER-FlagHAGFP-RarmDICER).
Design of the guide RNA targeting the region between Dicer 5’-UTR and its first coding exon for CRISPR/Cas9 mediated knock-in was carried out using the CRISPOR Design Tool [ 86 ]. Annealed oligonucleotides corresponding to the gRNA ( S5 Table ) were cloned into the vector pX459 (Addgene #48139) which also encodes S . pyogenes Cas9 with 2A-Puro.
The sequence of the donor plasmid was additionally mutagenized to disrupt the PAM sequence of the right homology arm to avoid its cleavage by the gRNA.
To obtain the knock-in (KI) cell line, 5 x 10 5 HCT116 cells were seeded in a 6 well plate with Dulbecco’s modified Eagle medium (DMEM, Gibco, Life Technologies) supplemented with 10% fetal bovine serum (FBS, Clontech) in a humidified atmosphere of 5% CO 2 at 37°C and transfected after 24 hours with the pX459-gRNADicerNterm-Cas9-2A-Puro plasmid and the Leftarm-FlagHAGFP-RightarmDICER donor plasmids at the ratio of 1 to 1 (6 micrograms plasmids in total) using Lipofectamine 2000 according to the manufacturer’s instructions. 24 hours later, puromycin (1 mg/mL) was added to the cells to increase the KI efficiency and genomic DNA was isolated from individual colonies few days later.
The presence of the Flag-HA-GFP tag in frame with hDICER coding sequence was confirmed by sequencing PCR amplicon from KI cell gDNA. Expression of Flag-HA-GFP N-terminal tagged Dicer protein in the KI cells was confirmed by western blot.
Cell culture and transfection
Cells were maintained in Dulbecco’s modified Eagle medium (DMEM, Gibco, Life Technologies) supplemented with 10% fetal bovine serum (FBS, Clontech) in a humidified atmosphere of 5% CO 2 at 37°C. Transfection was performed using Lipofectamine 2000 (Invitrogen, Thermo Fisher Scientific) according to the manufacturer’s instructions.
Lentivirus production and generation of stable cell lines
The lentiviral supernatant from single transfer vector was produced by transfecting HEK293T cells (ATCC CRL-3216) with 20 μg of the transfer vector, 15 μg of pMDLg/p RRE and 10 μg of pRSV-Rev packaging plasmids (Addgene #12251 and Addgene #12253) and the pVSV envelope plasmid (Addgene #8454) using Lipofectamine 2000 reagent (Invitrogen, Thermo Fisher Scientific) according to the manufacturer’s protocol. Standard DMEM medium (Gibco, Life Technologies) supplemented with 10% Fetal bovine serum (FBS, Gibco, Life Technologies) and 100 U/mL of penicillin-Streptomycin (Gibco, Life Technologies) were used for growing HEK293T cells and for lentivirus production. One 10 cm plate of HEK293T cells at 70–80% confluency was used for the transfection. The medium was replaced 8 hours post-transfection. After 48 hours the medium containing viral particles was collected and filtered through a 0.45 μm PES filter. The supernatant was directly used for transfection or stored at -80°C. A 6 well plate of HEK293T/NoDice or HEK293T/NoDiceΔPKR cells at 80% confluency was transduced using 600 μL of lentiviral supernatant either expressing FHA:DICER, N1, N3, Hel. or empty vector, supplemented with 4 ug/mL polybrene (Sigma) for 6 hours. The transduction media was then changed with fresh DMEM for 24 hours and the resistant cell clones were selected for about 6 weeks with blasticidin (15 μg/mL for NoDice or 10 μg/mL for NoDiceΔPKR) and subsequently maintained under blasticidin selection.
Viral stocks, virus infection
Viral stocks of SINV or SINV-GFP were produced as described in [ 36 ]. Cells were infected with SINV or SINV-GFP at an MOI of 0.02, 0.1, 1 or 2 and samples were collected at different time points as indicated in the figure legends.
Analysis of viral titer by plaque assay
Vero R cells were seeded in 96-well plates format and were infected with 10-fold serial dilutions infection supernatants for 1 hour. Afterwards, the inoculum was removed, and cells were cultured in 2.5% carboxymethyl cellulose for 72 hours at 37°C in a humidified atmosphere of 5% CO 2 . Plaques were counted manually under the microscope and viral titer was calculated according to the formula: PFU/mL = #plaques/ (Dilution*Volume of inoculum) . All data and statistics pertaining to plaque assay analysis can be found in S6 Table .
Western blot analysis
Proteins were extracted from cells and homogenized in 350 μL of lysis buffer (50 mM Tris-HCl pH 7.5, 150 mM NaCl, 5 mM EDTA, 1% Triton X-100, 0.5% SDS and Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich). Proteins were quantified by the Bradford method and 20 to 30 μg of total protein extract were loaded on 4–20% Mini-PROTEAN TGX Precast Gels (Bio-Rad). After transfer onto nitrocellulose membrane, equal loading was verified by Ponceau staining. For PVDF membrane, equal loading was verified by Coomassie staining after transfer and blotting. Membranes were blocked in 5% milk and probed with the following antibodies: anti-hDicer (1:500, F10 Santa Cruz, sc-136979) and anti-hDicer (1:1000, A301-937A, Bethyl), anti-TRBP (1:500, D-5 Santa Cruz, sc-514124), anti-PKR (1:2500, Abcam ab32506), anti-PACT (1:500, Abcam, ab75749), anti-HA (1:10000, Sigma, H9658), anti-DHX9 (1:500, Abcam, ab26271), anti-p-eIF2 (1:1000, Ser-52 Santa Cruz, sc-601670), anti-hADAR-1 (1:500 Santa Cruz, sc-271854) anti-p-PKR (1:1000 Abcam ab81303) anti-GFP (1:10000, Roche, 11814460001) and anti-Tubulin (1:10000, Sigma, T6557). Detection was performed using Chemiluminescent Substrate (Pierce, Thermo Fisher Scientific) and visualized on a Fusion FX imaging system (Vilber).
RNA extraction and northern blot analysis
Total RNA was extracted using Tri-Reagent Solution (Fisher Scientific; MRC, Inc) according to the manufacturer’s instructions. Northern blotting was performed on 10 μg of total RNA. RNA was resolved on a 12% urea-acrylamide gel, transferred onto Hybond-NX membrane (GE Healthcare). RNAs were then chemically cross-linked to the membrane during 90 min at 65°C using 1-ethyl-3-[3-dimethylaminopropyl]carbodiimide hydrochloride (EDC) (Sigma Aldrich). Membranes were prehybridized for 30 min in PerfectHyb plus (Sigma Aldrich) at 50°C. Probes consisting of oligodeoxyribonucleotides (see S5 Table ) were 5′-end labeled using T4 polynucleotide kinase (Thermo Fisher Scientific) with 25 μCi of [γ-32P]dATP. The labeled probe was hybridized to the blot overnight at 50°C. The blot was then washed twice at 50°C for 20 min (5× SSC/0.1% SDS), followed by an additional wash (1× SSC/0.1% SDS) for 5 min. Northern blots were exposed to phosphorimager plates and scanned using a Bioimager FLA-7000 (Fuji).
Immunoprecipitation
Immunoprecipitation experiments were carried out either on tagged proteins or on endogenous proteins.
Tagged proteins
Cells were harvested, washed twice with ice-cold 1× PBS (Gibco, Life Technologies), and resuspended in 550 μL of lysis buffer (50 mM Tris-HCl pH 7.5, 140 mM NaCl, 1.5 mM MgCl 2 , 0.1% NP-40), supplemented with Complete-EDTA-free Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich). Cells were lysed by 30 min incubation on ice and debris were removed by 15 min centrifugation at 2000 g and 4°C. An aliquot of the cleared lysates (50 μL) was kept aside as protein Input. Samples were divided into equal parts (250 μL each) and incubated with 15 μL of magnetic microparticles coated with monoclonal HA or MYC antibodies (MACS purification system, Miltenyi Biotech) at 4°C for 1 hour under rotation (10 rpm). Samples were passed through μ Columns (MACS purification system, Miltenyi Biotech). The μ Columns were then washed 3 times with 200 μL of lysis buffer and 1 time with 100 μL of washing buffer (20 mM Tris-HCl pH 7.5). To elute the immunoprecipitated proteins, 95°C pre-warmed 2x Western blot loading buffer (10% glycerol, 4% SDS, 62.5 mM Tris-HCl pH 6.8, 5% (v/v) 2-β-mercaptoethanol, Bromophenol Blue) was passed through the μ Columns. Proteins were analyzed by western blotting or by mass spectrometry.
Endogenous proteins
mock or SINV-GFP-infected HEK293T cells (MOI of 2) were lysed 6 hours post-infection using immunoprecipitation buffer (50 mM Tris-HCl [pH 7.5], 150 mM NaCl, 5 mM EDTA, 0.05% SDS, 1% triton) supplemented with Complete-EDTA-free Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich). Lysates were treated for 20 min at 37°C with 1 μL of DNase I (Thermo Fisher Scientific) using its buffer (10 mM MgCl 2 , 5 mM CaCl 2 and 1 μL of ribolock). Lysates were cleared by centrifugation at 16000 g, 10 min at 4°C. Supernatants were precleared 1 h at room temperature with magnetic beads blocked with BSA (Thermo Fisher Scientific) to avoid aspecific binding. Lysates were incubated overnight on wheel at 4°C with immunoprecipitation buffer containing magnetic Protein A DynaBeads (Invitrogen, Thermo Fisher Scientific) conjugated with human PKR antibody (Abcam) or negative control rabbit IgG (Cell signaling, Ozyme). Beads were washed 3 times with immunoprecipitation buffer, 3 times with wash buffer (50 mM Tris-HCl [pH 7.5], 200 mM NaCl, 5 mM EDTA, 0.05% SDS, 1% triton, supplemented with Complete-EDTA-free Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich) and twice with cold PBS 1X (Gibco, Life Technologies). Beads were eluted with 2x western blot loading buffer and incubated for 10 min at 95°C under agitation. Proteins were analyzed by western blotting.
RNase treatment followed by co-IP
On tagged proteins : Cells were harvested, washed twice with ice-cold 1× PBS (Gibco, Life Technologies), and resuspended in 550 μL of lysis buffer (50 mM Tris-HCl pH 7.5, 140 mM NaCl, 1.5 mM MgCl 2 , 0.1% NP-40), supplemented with Complete-EDTA-free Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich). Cells were lysed by 30 min incubation on ice and debris were removed by 15 min centrifugation at 2000 g and 4°C. Lysate was treated or not with RNase A/T1 mix (Thermo Fisher Scientific) and place at 37°C 30 min. An aliquot of the cleared lysates (25 μL) was kept aside as protein Input and another aliquot (25 μL) was kept to assess RNase treatment efficiency. Co-IP was led as previously described.
Total RNA was extracted using Tri-Reagent Solution (Fisher Scientific; MRC, Inc) according to the manufacturer’s instructions. RNA integrity upon treatment was verified on an 1% agarose gel containing ethidium bromide 10 mg/mL (Invitrogen, Thermo Fisher Scientific) and revealed under UV on Gel DocEZ system (Bio-Rad).
siRNA transfection
20 nM of human AGO2 or non-targeting control siRNA (Horizon discovery) were transfected in 130000 NoDice FHA:ctrl #2, NoDice FHA:DICER WT #4, N1 #6 or N3 #2.13 cells using Lipofectamine 2000 transfection reagent (Invitrogen, Thermo Fisher Scientific) according to the manufacturer’s instructions. After 24 hours, the cells were again transfected with 20 nM of the same siRNA and incubated overnight. Cells were infected or not with SINV-GFP at an MOI of 0.02 for 24 h. Proteins and supernatants were collected and analyzed by western blotting and plaque assay, respectively.
BiFC assay
Experiments were carried out in two different ways. For non-fixed cells, NoDiceΔPKR or HEK293T cells were seeded at the density of 1.2 x 10 5 cells per well in a 24-well plate. After 16 hours, cells were transfected with equimolar quantities of each plasmid forming BiFC couples. After 24 hours, cells were infected with SINV at an MOI of 2 and pictures were taken 6 hours post-infection using ZOE fluorescent cell imager (Bio-Rad). Proteins were collected with lysis buffer (50 mM Tris-HCl pH 7.5, SDS 0.05%, Triton 1%, 5 mM EDTA, 150 mM NaCl) supplemented with Complete-EDTA-free Protease Inhibitor Cocktail (complete Mini; Sigma Aldrich), and subjected to western blot analysis. For fixed cells, NoDiceΔPKR cells were seeded at the density of 8.10 4 cells per well in 8-well Millicell EZ Slides (Merck Millipore), transfected and infected as described previously. At 6 hours post-infection, cells were fixed with 4% formaldehyde and 0.2% glutaraldehyde for 10 min. Cells were then washed with 1× PBS (Gibco, Life Technologies) and stained with 10 μg/μL DAPI (Invitrogen, Thermo Fisher Scientific) in 1× PBS solution (Invitrogen, Thermo Fisher Scientific) for 5 min. Fixed cells were mounted on a glass slide with Fluoromount-G mounting media (Southern Biotech). Images were acquired using confocal LSM780 (Zeiss) inverted microscope with an argon laser (514x nm) and with ×40 immersion oil objective. All pictures obtained from BiFC experiments were treated using FigureJ software (NIH).
Mass spectrometry analysis
Protein extracts were prepared for mass spectrometry as described in a previous study [ 87 ]. Each sample was precipitated with 0.1 M ammonium acetate in 100% methanol, and proteins were resuspended in 50 mM ammonium bicarbonate. After a reduction-alkylation step (dithiothreitol 5 mM–iodoacetamide 10 mM), proteins were digested overnight with sequencing-grade porcine trypsin (1:25, w/w, Promega, Fitchburg, MA, USA). The resulting vacuum-dried peptides were resuspended in water containing 0.1% (v/v) formic acid (solvent A). One sixth of the peptide mixtures were analyzed by nanoLC-MS/MS an Easy-nanoLC-1000 system coupled to a Q-Exactive Plus mass spectrometer (Thermo-Fisher Scientific, USA) operating in positive mode. Five microliters of each sample were loaded on a C-18 precolumn (75 μm ID × 20 mm nanoViper, 3 μm Acclaim PepMap; Thermo) coupled with the analytical C18 analytical column (75 μm ID × 25 cm nanoViper, 3 μm Acclaim PepMap; Thermo). Peptides were eluted with a 160 min gradient of 0.1% formic acid in acetonitrile at 300 nL/min. The Q-Exactive Plus was operated in data-dependent acquisition mode (DDA) with Xcalibur software (Thermo-Fisher Scientific). Survey MS scans were acquired at a resolution of 70K at 200 m/z (mass range 350–1250), with a maximum injection time of 20 ms and an automatic gain control (AGC) set to 3e6. Up to 10 of the most intense multiply charged ions (≥2) were selected for fragmentation with a maximum injection time of 100 ms, an AGC set at 1e5 and a resolution of 17.5K. A dynamic exclusion time of 20 s was applied during the peak selection process.
Database search and mass-spectrometry data post-processing
Data were searched against a database containing Human and Viruses UniProtKB sequences with a decoy strategy (GFP, Human and Sindbis Virus SwissProt sequences as well as Semliki Forest Virus SwissProt and TrEMBL sequences (releases from January 2017, 40439 sequences)). Peptides were identified with Mascot algorithm (version 2.3, Matrix Science, London, UK) with the following search parameters: carbamidomethylation of cysteine was set as fixed modification; N-terminal protein acetylation, phosphorylation of serine / threonine / tyrosine and oxidation of methionine were set as variable modifications; tryptic specificity with up to three missed cleavages was used. The mass tolerances in MS and MS/MS were set to 10 ppm and 0.02 Da respectively, and the instrument configuration was specified as “ESI-Trap”. The resulting .dat Mascot files were then imported into Proline v1.4 package ( http://proline.profiproteomics.fr ) for post-processing. Proteins were validated with Mascot pretty rank equal to 1, 1% FDR on both peptide spectrum matches (PSM) and protein sets (based on score). The total number of MS/MS fragmentation spectra (Spectral count or SpC) was used for subsequent protein quantification in the different samples. All data have been deposited to the ProteomeXchange Consortium [ 88 ].
Exploratory and differential expression analysis of LC-MS/MS data
Mass spectrometry data obtained for each sample were stored in a local MongoDB database and subsequently analyzed through a Shiny Application built upon the R/Bioconductor packages msmsEDA (Gregori J, Sanchez A, Villanueva J (2014). msmsEDA: Exploratory Data Analysis of LC-MS/MS data by spectral counts. R/Bioconductor package version 1.22.0) and msmsTests (Gregori J, Sanchez A, Villanueva J (2013). msmsTests: LC-MS/MS Differential Expression Tests. R/Bioconductor package version 1.22.0). Exploratory data analyses of LC-MS/MS data were thus conducted, and differential expression tests were performed using a negative binomial regression model. The p-values were adjusted with FDR control by the Benjamini-Hochberg method and the following criteria were used to define differentially expressed proteins: an adjusted p-value < 0.05, a minimum of 5 SpC in the most abundant condition, and a minimum fold change of 2 (abs(LogFC) > 1). GO term analysis was performed using the EnrichR web-based tool ( http://amp.pharm.mssm.edu/Enrichr ). The direct interaction network for proteins enriched in SINV-infected cells was generated using the STRING database ( https://string-db.org ).
Supporting information
S1 Fig
Analysis of SINV-GFP infection in FHA:DICER cell lines at different MOI and time points.
A. miR-16 expression analyzed by northern blot in HEK293T, NoDice FHA:ctrl #1 and FHA:DICER WT #4 cell lines. Expression of snRNA U6 was used as loading control. B. Representative GFP pictures of HEK293T, NoDice 2.20, NoDice 4.25, NoDice FHA:ctrl #1 and NoDice FHA:ctrl #2 cells infected with SINV-GFP at an MOI of 0.02 for 24 h. The left panel corresponds to GFP signal and the right panel to a merge of GFP signal and the corresponding brightfield. Pictures were taken with a 5x magnification. hpi: hours post-infection. C. Mean (+/- SEM) of SINV-GFP viral titers in cells infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. * p < 0.05, ns: non-significant, ordinary one-way ANOVA test with Bonferroni correction. D . Western blot analysis of DICER, AGO2 and GFP expression in SINV-GFP-infected cells shown in B. Gamma-Tubulin was used as loading control.
(TIF)
S2 Fig
LC-MS/MS analysis of DICER interactome during SFV infection.
A. Volcano plot for differentially expressed proteins (DEPs) between HA IP and CTL IP in FHA:DICER mock-infected cells. Each protein is marked as a dot; proteins that are significantly up-regulated in HA IP are shown in red, up-regulated proteins in CTL IP are shown in blue, and non-significant proteins are in black. The horizontal line denotes a p-value of 0.05 and the vertical lines the Log2 fold change cutoff (-1 and 1). DICER and its cofactors (TRBP, PACT, AGO2) are highlighted in yellow. B. Left panel: Volcano plot for DEPs between SFV (MOI of 2, 6 hpi) and mock fractions of HA IP in FHA:DICER cells. Same colour code and thresholds as in A were applied. Proteins that are discussed in the text are highlighted in yellow and SFV proteins in purple. C. Summary of the differential expression analysis of SFV vs mock fractions from HA IP in FHA:DICER cells. The analysis has been performed using a generalized linear model of a negative-binomial distribution and p-values were corrected for multiple testing using the Benjamini-Hochberg method.
(TIF)
S3 Fig
Confirmation of LC-MS/MS analysis by co-IP and BiFC controls.
A. FHA:DICER WT #4 cells were infected with SINV-GFP at an MOI of 0.02 for 24 h and a HA co-IP was performed. Eluted proteins were resolved by western blot and IP efficiency was assessed using an HA antibody. In parallel, co-IPed proteins were visualized using appropriate antibodies. GFP antibody was used to verify the infection and Ponceau staining serves as loading control. B. 1% agarose gel analysis of RNA extracted from INPUT of the co-IP in Fig 3B . Ribosomal RNA integrity was compared to a control HEK293T cell line. RNAs were revealed using ethidium bromide under UV. C. Schematic representation of Human DICER proteins used for BiFC positive and negative controls. The different conserved domains are shown in colored boxes. DUF283: Domain of Unknown Function; PAZ: PIWI ARGONAUTE ZWILLE domain; dsRBD: dsRNA-binding domain. hDICER WT is the full-length protein. hDICER N1 is deleted of the first N-terminal 495 amino acids. D. Expression of BiFC plasmids was assessed by western blot. DICER proteins (WT and N1) and PKR were visualized using antibodies targeting endogenous proteins, whereas TRBP and PACT were detected using GFP antibody. Antibody targeting the SINV coat protein (CP) was used as infection control. Ponceau staining was used as loading control. E. Positive and negative BiFC controls on fixed NoDiceΔPKR cells. After co-transfection, cells were infected with SINV at an MOI of 2 for 6 h and fixed. After fixation, cells were stained with DAPI and observed under confocal microscope. Merge pictures of BiFC and DAPI signals of SINV-infected cells are shown. A higher magnification of images showing the interaction represented by a red square is shown in the bottom left corner. Scale bars: 20 μm and 10 μm. F. Expression of BiFC plasmids was assessed by western blot. DICER, PKR, TRBP and PACT were detected using GFP antibody. Antibody targeting the SINV coat protein (CP) was used as infection control. Gamma-Tubulin was used as loading control. The asterisk corresponds to an aspecific band. G. Interactions between DICER and TRBP, PACT or PKR were visualized by BiFC. Plasmids expressing N-ter Venus:DICER and TRBP:, PACT: or PKR:Venus C-ter were co-transfected in HEK293T cells for 24 h and cells were either infected with SINV at an MOI of 2 for 6 h or not. The different combinations are indicated on the left side. Reconstitution of Venus (BiFC) signal was observed under epifluorescence microscope. For each condition, the left panel corresponds to Venus signal and the right panel to the corresponding brightfield pictures. Scale bar: 100 μm.
(TIF)
S4 Fig
Confirmation of DICER interactome upon SINV infection in HCT116 KI-DICER cells.
A. Schematic representation of DICER WT and Flag-HA(FHA)-GFP knocked-in (KI) alleles. FHA sequence is in purple, GFP in green, DICER 5’UTR in orange and DICER coding region in yellow. The gRNA used to generate the KI was designed to target the first coding exon of DICER gene. B. PCR on genomic DNA extracted from WT and KI cells. C. An oligo outside the homologous recombination region and an oligo within the GFP tag were used to verify the presence of a 1040 bp amplicon in HCT116 KI-DICER clone. Sequencing results corresponding to this region are shown. D. Western blot analysis of DICER, p-PKR, PKR and p-eIF2α expression in mock or SINVGFP-infected HEK293T and HCT116 KI-DICER cell lines at an MOI of 2 for 6 h or 16 h and 0.02 for 24 h. GFP antibody was used to verify the infection. Ponceau and gamma-Tubulin were used as loading controls.
(TIF)
S5 Fig
Interaction analysis between the different versions of DICER and TRBP or PACT using BiFC assay.
NoDiceΔPKR cells were co-transfected for 24 h with plasmids expressing the different versions of DICER proteins fused to the N-terminal part of Venus and either TRBP:Venus C-ter ( A ) or PACT:Venus C-ter ( B ). Cells were then infected with SINV at an MOI of 2 for 6 h and Venus signal was observed under epifluorescence microscope. The left panel corresponds to Venus signal and the right panel to the corresponding brightfield picture. Pictures were taken with a 5x magnification. hpi: hours post-infection. Scale bar: 100 μm.
(TIF)
S6 Fig
Analysis of the importance of Dicer helicase domain on SINV-GFP infection in FHA:DICER mutant stable cell lines.
A . Northern blot analysis of miR-16 expression in HEK293T, NoDice 2.20, NoDice FHA:ctrl #2, FHA:DICER WT polyclonal, FHA:DICER N1 #6, FHA:DICER Hel. #2.6, and FHA:DICER N3 #2.13. Expression of snRNA U6 was used as loading control. B. Representative GFP fluorescent microscopy images of HEK293T, NoDice 2.20, FHA:DICER mutants cell lines infected with SINV-GFP at an MOI of 0.02 for 24 h. The left panel corresponds to GFP signal and the right panel to a merge picture of GFP signal and brightfield. Pictures were taken with a 5x magnification. hpi: hours post-infection. C . Mean (+/- SEM) of SINV-GFP viral titers over FHA:DICER WT #4 cells in FHA:DICER N1 #6, FHA:DICER N3 #2.13, NoDice FHA:ctrl #2 and NoDice 2.20 cell lines infected at an MOI of 0.02 for 24 h (n = 3) from plaque assay quantification. *** p < 0.001, ns: non-significant, ordinary one-way ANOVA test with Bonferroni correction.
(TIF)
S1 Table
Top 100 proteins that are differentially immunoprecipitated in mock-infected FHA:DICER cells by the HA and Myc (CTL) antibodies.
Related to Fig 2 .
(XLSX)
S2 Table
Top 100 proteins that are differentially immunoprecipitated with the HA antibody in SINV-infected vs mock-infected FHA:DICER cells.
Related to Fig 2 .
(XLSX)
S3 Table
Top 100 proteins that are differentially immunoprecipitated in mock-infected FHA:DICER cells by the HA and Myc (CTL) antibodies, in the SFV infection experiment.
Related to S2 Fig .
(XLSX)
S4 Table
Top 100 proteins that are differentially immunoprecipitated with the HA antibody in SFV-infected vs mock-infected FHA:DICER cells.
Related to S2 Fig .
(XLSX)
S5 Table
List of primers used in this study.
(XLSX)
S6 Table
Data and statistical tests details used in plaque assays shown in Figs 1 , 6 , 7 , 8 , S1 and S6 .
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Foot-and-mouth disease (FMD) is a highly infectious condition that affects domestic and wild cloven-hoofed animal species, including cattle, sheep, goats, and pigs [ 1 ]. This disease has devastating economic consequences, resulting from the restrictions on international trade of animals and animal products from infected countries, as well as the high costs associated with the implementation of outbreak controlling mechanisms [ 2 , 3 ].
FMD transmission factors include contact with infected fomites and personnel, high animal density, high contact rates between domestic animals and wildlife, and lack of compliance with biosecurity measures [ 4 – 7 ], among others. Nevertheless, direct contact between infected and susceptible animals constitutes one of the principal mechanisms for FMD virus dissemination [ 6 ]. Importantly, this disease may spread among animals for long periods without showing any observable signs, imposing an additional challenge for devising prevention and control strategies [ 6 ].
Livestock movement is a common practice for most agricultural systems and links to trade and the need to access resources for animal sustainment [ 8 , 9 ]. However, this practice increases the probability of direct contact between FMD-infected and susceptible animals despite its importance. Therefore, during the last years, several studies aimed to characterize how livestock movement features may serve as an indicator for the FMD dissemination risk [ 10 ].
The sanitary authorities at the country level must systematically gather information about livestock movements, inline with the recommendation of the World Organization for Animal Health in the International Animal Health Code [ 11 ]. More specifically, they must register information related to the official livestock trades [ 12 , 13 ]. These records represent connections between livestock holdings, markets, and abattoirs, among others [ 14 ], which allow the construction of livestock transportation networks. The structure of these networks can critically influence the dynamics of transmission of many infectious diseases, including FMD [ 14 ].
In these transportation networks, the nodes represent places where animals may have contact with infectious agents, for instance, holdings or markets located in specific areas. The livestock movements between these correspond to edges [ 14 , 15 ]. Different works characterized these transportation networks for previous FMD outbreaks, e.g., the well-documented 2001 United Kingdom FMD outbreak [ 16 ]. These studies identified that multiple nodal network properties related to disease spreading. For example, nodes with significant levels of FMD susceptibility reported high values for different centrality measurements, including in and out-degree, ingoing and outgoing infection chain, and betweenness, among others [ 14 ]. Other works on simulation also showed that some of these nodal centrality properties relate to infectious disease propagation [ 15 , 17 ]. In principle, these local descriptions may allow the identification and prioritization of critical areas related to future FMD spread over time. However, despite its utility for describing some regional FMD epidemiological aspects, such a descriptive approach may be restricted to devising surveillance strategies because epidemiological vigilance requires the prioritization of the critical areas related to FMD [ 18 ]. This restriction emerges from the conflicting nature of the different centrality measurements [ 17 ], which likely result in different prioritizations of risk areas, limiting the possibility of constructing a unified rank for surveillance design. In addition, depending on the disease under study, network properties considered may differ. For instance, network properties related to fast spreading diseases, such as FMD, may vary from properties used to describe risk in bacterial origin diseases such as bovine brucellosis [ 19 ]. Therefore, computing a single high-risk area prioritization from multiple nodal movement network descriptions which reflect particular disease spreading properties constitutes a challenge not only for the disease characterization but also for devising cost-effective epidemic control alternatives [ 20 , 21 ].
The main objective of this work was to construct an FMD risk regional transmission ranking based on livestock movement data and quantitatively evaluate how this ranking may benefit the prediction of future FMD disease spread. In contrast with previous works aimed to describe nodal network features likely related to virus transmission [ 17 , 22 ], this work aims to use these nodal descriptions to construct a single ranking of FMD high-risk areas. The proposed ranking relies on super-spreaders, nodes that maximize their impact on other nodes during an outbreak [ 23 ], and which can be described by integrating multiple centrality nodal measures [ 24 ]. We hypothesized that a super-spreaders-based ranking (SSBR) may provide valuable information to address FMD in different stages, including early disease spreading and posterior disease proliferation [ 23 ]. To explore this hypothesis, we studied the prediction capacity of the SSBR of an actual FMD outbreak detected in the region of Cesar (Colombia). Our results show that the SSBR based on livestock movement data from 2016 helps prioritize areas subsequently affected by an FMD outbreak reported in 2018. The main contributions of our work are, first, the construction of a ranking of FMD high-risk areas based on cattle movement data and the notion of super-spreaders, and second, the quantitative evaluation of the prediction capacity of this ranking over an actual FMD outbreak. These findings may have implications for designing cost-effective planning mechanisms for disease breakout control and FMD epidemiology understanding.
Materials and methods
Fig 1 shows the methodology proposed for computing the ranking of areas with a high risk of FMD affectation. First, the mobilizations of animals on a daily scale were determined for the construction of cattle mobilization networks for non-overlapping monthly periods. Then, a set of nodal centrality properties were computed on these networks to describe different aspects of the FMD disease risk spreading. Next, these properties were combined to identify the super-spreader nodes and their corresponding SSBR. Finally, the predictive capacity of this ranking to detect an actual FMD outbreak was quantitatively evaluated.
10.1371/journal.pone.0284180.g001
Fig 1
Workflow of the proposed methodology to compute rankings of FMD high-risk areas.
First, cattle movement records allowed the construction of livestock mobilization networks. Then, centrality measurements, including betweenness and degree, allowed to build multiple rankings of risk. Next, a unique SSBR resulted from combining these rankings. Finally, based on the SSBR, predictions of possible future outbreaks were computed and compared against an actual FMD outbreak.
Livestock mobilization network
The information system for mobilization guides (SIGMA—Sistema de información para Guías de Movilización Animal— https://sigma.ica.gov.co/ ) provided livestock movement records used in this study. The Instituto Colombiano Agropecuario (ICA), the principal sanitary authority in Colombia administers this system. SIGMA contains the so-called sanitarian guides of internal livestock mobilization . These guides constitute the official authorization allowing cattle movements inside the country. Each guide includes the date, source, destination, species type, and the batch size of the livestock movements (number of transported animals). Due to use of organizational data, informed consent was not applicable.
This study focused on historical data of cattle movements from 1 January 2016 to 31 December 2016 in the Cesar (Colombia) region. This study concentrated on this region because later, in September 2018, it reported an FMD outbreak involving some of its municipalities [ 25 , 26 ]. Therefore, this study aimed to predict these areas using the proposed ranking approach. The study conformed to guidelines of the Comite de Bioética at Universidad Nacional de Colombia, permitting exemption from full ethical review.
SIGMA mobilization data allowed the construction of livestock mobility networks. More specifically, the nodes corresponded to municipalities that served as the origin or destination of animal movements. A movement with at least one animal defined a link between nodes. The total number of animal mobilizations in non-overlapping monthly consecutive periods originating from the same municipality and ending at a different municipality characterized each link. This time partition size, i.e., about four weeks, assumes that it is unlikely that an FMD epidemic could persist without being identified beyond this time [ 14 – 16 ].
For the animal movement characterization, undirected, directed, and weighted directed networks of livestock movements were constructed [ 14 ]. Undirected networks described livestock movements without considering direction. Directed networks distinguished between the origin and destination of the livestock movement. Finally, weighted-directed networks accounted for the number of animals transported. Markets and holdings were considered different nodes in the graph. Data for different holdings were aggregated into a single holding per municipality. Livestock data movements for markets were also combined into a single marked node per municipality when they were present. Because of the low transmission risk in abattoirs, movements to these were not considered for the transportation graph construction.
Super-spreading nodes identification
The underlying livestock movement network may serve as a proxy to study the animal contact phenomena. In particular, it may provide valuable information about the possible dynamics of the FMD disease spread [ 14 ]. The present analysis focused on identifying and ranking the so-called super-spreaders nodes [ 27 ]. These nodes maximize their epidemiological impact on other nodes [ 24 ]. Therefore, these nodes may link to a higher risk of FMD spreading and are worthy of being characterized.
Network centrality properties
Recent evidence suggests that the super-spreading phenomenon is closely associated with high centrality values in the nodes of the transportation network [ 24 ]. The work herein presented focused on particular nodal centrality measurements that capture livestock transportation features linked to these super-spreaders and that may increase FMD spreading chance.
Despite the large number of alternatives for nodal centrality measurements on these networks [ 28 ], only a few are good predictors of the super-spreader phenomena [ 17 ]. Therefore, this work concentrates on centrality described by degree and betweenness features, which proved to be good predictors of super-spreading [ 17 ]. This apriori selection of a few network features may also help minimize overfitting risk [ 29 ].
The degree of centrality corresponds to the number of connections each node has to other nodes [ 30 ]. In the livestock transportation network, the degree may describe the number of potential direct contacts per holding. For FMD, animals in nodes with a high degree, i.e., with many connections, are likely to become infected early in an epidemic outbreak [ 15 ]. The betweenness centrality estimates the extent to which a node lies on paths between other nodes by considering its importance for the shortest paths through the network [ 30 ]. In the FMD case, the nodes with high betweenness centrality are likely to accelerate the spread of infection through the network during livestock transportation [ 15 ]. Therefore, targeting nodes with higher degree and betweenness may help early identification and rapid disease control [ 24 ].
The risk of contagious and spreading FMD disease can be affected by the direction of the livestock movement. For instance, livestock movements toward abattoirs are less likely to spread diseases than livestock movements to markets [ 6 ]. Similarly, when the number of transported animals is large, the risk of contact with infected animals increases [ 6 ]. These two factors were accounted for by computing the degree and betweenness measurements in the three networks: undirected, directed, and weighted [ 31 ]. Depending on the kind of network, the definition of the centrality measurements may vary. For instance, in directed networks, two types of degrees may be considered, namely, in-degree and out-degree.
In summary, and accounting for the kind of network particularities, eight centrality measurements were computed for characterizing super-spreading behavior: 1) degree on undirected, 2) in-degree on directed, 3) out-degree on directed, 4) degree on weighted, 5) in-degree on weighted, 6) out-degree on weighted, 7) betweenness on undirected, and 8) betweenness on weighted. In principle, any of these measures computed on the corresponding transportation network, may provide a ranking of FMD risk for the nodes. However, previous work suggests that no single centrality measurements perform as the best predictor of disease spreading [ 24 ], mainly because different measures have different objectives. However, in most cases, epidemiological control requires a single rank for planning [ 24 ].
Super-spreaders identification
Borda’s count aggregation method provided a single super-spreader ranking for the nodes [ 24 , 32 ]. This method employs as input a set of lists of ranks R = { r 1 , r 2 , ⋯, r n }. Each ranked list r i has K items possibly in a different order, with K being the number of nodes in the transportation network. For the super-spreaders identification problem, the ranked lists resulted from computing the ascending order of the nodes for each centrality feature. Then the method defines a new rank based on an aggregated rank value defined for each element j as:
B ( j ) = ∑ i ∈ R K - r i j
where r i j is the position of the j -th node in the rank r i . These B ( j ) are then reordered to provide Borda’s count rank.
As previously discussed, degree and betweenness characterization may have different epidemiological purposes, namely, modeling the potential first contagion and the possibility of disease propagation. In this work, these two objectives were modeled by using a hierarchical Borda’s aggregation. In particular, the degree (i.e., features 1) to 6) in Section) and betweenness-related ranks (i.e., features 7) and 8) in Section) were combined using Borda’s method, resulting in two different ranks. These two ranks were then combined again using Borda’s aggregation to construct the final aggregated SSBR.
The 2018 FMD outbreak in Cesar (Colombia)
The super-spreaders’ approach may provide a single ranking suitable for prioritizing areas with higher FMD risk, more specifically, for quantifying FMD risk in areas where the disease is not present yet [ 33 ]. However, good risk indicators should help predict areas where this risk may materialize [ 33 ], i.e., the risk should be higher for regions later affected by FMD disease. Therefore, in contrast to previous studies that provided only risk descriptions, this study investigated the predictive capacity of the proposed ranking to predict regions for which FMD risk later materialized. For this, we used the mobilization data from 2016 to compute risk rankings and then studied its prediction capacity on an actual FMD outbreak detected in 2018 at the department of Cesar, Colombia.
This outbreak was detected in October 2018 after two cows (from a holding of 216) tested positives for FMD type O in the municipality of San Diego [ 25 ]. Posteriorly, FMD infections were also detected in the municipality of Valledupar. The sanitary authority confirmed this infection, deployed epidemiological control protocols, and defined a quarantine control area [ 25 , 26 ]. The area included the municipalities of Valledupar, La Paz (Robles), San Diego, and Agustín Codazzi, all located in Cesar (Colombia) [ 25 , 26 ]. Officially, the outbreak originated with the illegal introduction of animals from the Bolivarian Republic of Venezuela. However, it cannot be discarded that there was a history of disease circulation in the area before the emergence of the outbreak [ 34 ]. Map at Fig 2 shows the municipalities affected by the 2018 outbreak at Cesar. In particular, primary (red) and secondary (organge) foci. We evaluated if the 2016 SSBR predicted the two FMD-affected municipalities in 2018.
10.1371/journal.pone.0284180.g002
Fig 2
Municipalities affected by FMD in the Cesar (Colombia) department in 2018.
Primary and secondary foci are marked as red and orange points, respectively. Quarantine (green) and outbreak (yellow) areas are also shown.
Data were extracted and organized with Python 3.4.7. Calculations of measurements were performed using the Python module NetworkX 2.4. Super-spreader characterization, rankings, and municipality classification were computed using Python 3.4.7.
Results
This work studied the construction of an network risk ranking based on the animal transportation network and its capacity to characterize areas affected by an actual FMD outbreak. First, we described the general features of the animal transportation data herein studied. Then, we reported the monthly SSBR computed using the 2016 transportation data. Following, we report these rankings’ predictive capacity in characterizing municipalities where the risk later materialized, i.e., the municipalities affected by FMD in 2018. We also studied the stability of the SSBR across different months. Finally, we studied the distributions for the centrality measurements considered.
Livestock transportation network
A total of 21,254 livestock movements were reported between municipalities in the region of Cesar (Colombia) in 2016. Of these movements 9,854 (to markets and holdings) among 29 nodes in 25 municipalities were considered for the construction of the transportation network. A total of 175,068 animal movements were studied. Movements by truck were 97% and by walking, 3%.
Network risk rankings
Fig 3 shows the monthly SSBR (top 15) computed using the proposed approach using the 2016 livestock transportation data. The figure shows in yellow the municipalities later affected by the 2018 FMD outbreak. As observed, the proposed method ranks as high risk posteriorly affected by FMD or considered at high risk by the sanitary authority [ 25 , 26 ]. More specifically, the proposed rankings tagged as high risk the municipalities posteriorly affected by the disease. S1 Table reports the municipalities ranked by the SSBR.
10.1371/journal.pone.0284180.g003
Fig 3
Network risk rankings computed for 2016 months.
Each ranking corresponds to different municipalities displayed ordered depending on their level of risk. Yellow flow lines indicate the municipalities later affected by an outbreak in the same region in 2018.
Valledupar, a municipality with an elevated risk in 2018, led the ranking most months [ 26 ]. San Diego, the municipality where the disease first emerged [ 25 ], showed also high the hierarchy throughout the year.
SSBR predictive capacity to characterize materialized risk
Supervised classification allowed a quantitative evaluation of the rankings’ predictive capacity to characterize areas where the risk later materialized [ 35 ]. Specifically, we aimed to describe how well the SSBR obtained with 2016 livestock mobilization data predicted the municipalities that materialized the risk by FMD in 2018 [ 25 , 26 ]. The orders induced by the 2016 centrality-based risk rankings provided the cut-offs for studying the true-positive rate or hit rate, i.e., the percentage of municipalities with materialized risk in 2018 correctly targeted by the 2016 proposed ranking from the total of municipalities in risk, and the false-positive rate or false alarm rate, i.e., the percentage of municipalities with materialized risk in 2018 incorrectly targeted by the 2016 rankings from the set of municipalities with no risk. The predictive ability of these classifiers was measured using the area under the receiver operative curve (ROC) [ 35 ]. For constructing this ROC curve, the number of municipalities prioritized by the proposed ranking was considered as the threshold. Additionally, to provide a quantitative measure of not highlighting a region affected by FMD. The negative predictive values (NPV) for monthly classifiers were computed. In particular, considering three classifiers constructed with the top-three areas provided by the SSBR, the mean and standard deviation of the NPVs across months were calculated. NPV was computed as True Negative/(False Negative + True Negative) [ 36 ].
Fig 4 shows the mean of the ROCs for the months under study. As observed, the classifiers based on SSBR correctly targeted municipalities with high risk in 2018, with low false alarm rates. The mean of the area under the ROC for all the months was 0.91 ± 0.07 ( mean ± 1 std . dev .). The maximum area under the ROC was 0.98 in September, and the lowest performance resulted in July with 0.80.
10.1371/journal.pone.0284180.g004
Fig 4
Mean and standard deviation of the ROC curves for different months.
In this case, the task aimed to classify municipalities affected by the disease in 2018. Therefore, monthly ROC curves were constructed by considering varying cut-offs in the risk rankings proposed, i.e., changing the number of municipalities prioritized by the proposed ranking.
The mean and standard deviation of the NPVs for the top-1 classifiers, i.e., risk classifiers that considered only the first region ranked by the SSBR were 0.66 ± 0.47. The mean and standard deviation of the NPVs for classifiers based on the top-2 ranking, i.e., risk classifiers relying on the first two regions ranked by the SSBR were 0.41 ± 0.18. At the same time, the mean and standard deviation of the NPVs for classifiers based on the top-3 ranking were 0.36 ± 0.21. Indicating that the proposed SSBR may miss some regions affected by FMD. For instance, see the month of July in Fig 3 .
Finally, Table 1 reports the mean and standard deviation of the area under the ROC for the SSBR method and some commonly used alternative approaches for constructing risk rankings.
10.1371/journal.pone.0284180.t001
Table 1
The area under the ROC for five different methods for ranking regions likely affected by FMD.
The proposed method SSBR combines different degree and betweenness features. The Borda degree and betweenness combine degree and betweenness features, respectively. Two rankings rely on the density of animals and the number of farms.
Ranking method
Mean
Standard deviation
SSBR
0.91
0.07
Borda ′ sdegree
0.9
0.06
Borda ′ sbetweenness
0.89
0.06
Densityofanimals
0.89
0
Numberoffarms
0.8
0
In particular, four different rankings were considered: two for degree and betweenness features, obtained by combining the corresponding features using Borda’s count aggregation, and two for the density of animals and the number of farms. As observed, the proposed approach improved disease characterization performance compared to other methods related to individual features and underlying productive variables.
Rankings stability across months
Fig 3 shows that SSBRs varied across months, indicating that the risk may also change along the year, as livestock transportation dynamics may also vary during the year. To investigate these variations, we studied the stability of monthly rankings. For this, we computed Spearman’s rank-order correlation coefficient ρ between pairs of months [ 37 ], which assesses how well the relationship between two monthly FMD risks is described using the monotonic functions induced by the SSBR, i.e., the similarity of FMD risks obtained in the different months.
Fig 5 shows the pairwise monthly stability measured by ρ . As observed, the FMD risk SSBR from January to July were highly stable, with stability values above 0.82. The stability obtained for August compared with the rest of the months decreased but remained high. The SSBR from September to December were highly stable, with values above 0.82.
10.1371/journal.pone.0284180.g005
Fig 5
Spearman’s rank-order correlation coefficient of the FMD SSBR computed for pairs of months in 2016.
High values on this measure represent high stable ranks.
Distribution of centrality measurements
Fig 6 shows the distributions for the eight centrality measurements considered for the 2016 months. As expected, different centrality features exhibit different distribution shapes, reflecting specific livestock transportation network dynamic attributes.
10.1371/journal.pone.0284180.g006
Fig 6
Distributions of different network centrality measurements.
These distributions were computed per month for features related to degree and betweenness.
For instance, degree distributions (Panel Degree at Fig 6 ) are more symmetric than betweenness distributions (Panel Betweenness at Fig 6 ), which have positive skewness and kurtosis values, likely related to the existence of nodes serving as intermediate in the network. Similarly, centrality measurement distributions change over months, evidencing a time-varying dynamic.
Discussion
This work studied the construction of rankings of FMD risk from livestock transportation data and their predictive capacity to anticipate future disease outbreaks. A dataset consisting of more than 20,000 animal mobilizations from a region subsequently affected by FMD was analyzed using super-spreader rankings to describe the risk of contact. In contrast to previous livestock transportation network-based descriptions which mainly focused on features linked to FMD transmission, this study describes for the first time the construction of a single SSBR suitable for decision making, which can predict future outbreaks with high accuracy.
FMD is one of the most economically devastating livestock diseases because of the resulting production losses and the severe restrictions on the animal trade [ 2 ]. Therefore, developing strategies to anticipate and mitigate the disease is challenging for the productive sector and the sanitary authorities [ 38 ]. Ideally, these strategies should focus on high-risk areas in scenarios of limited resources [ 21 , 39 , 40 ]. Therefore, the prioritization of these areas is a significant requirement. Nevertheless, because of the complexity of the disease transmission mechanisms, the risks of FMD outbreaks are hard to establish [ 6 ]. Our results show that a single SSBR based on livestock transportation may provide high-performance values in determining future areas with FMD outbreaks. The simplicity of this ranking, essentially an ordered list of municipalities (see Fig 3 ), and its high predictive capacity for future outbreaks (see Fig 4 ) makes it suitable to be used in decision-making tasks.
Previous works on livestock transportation networks showed that high centrality values in nodes link to an increased risk of FMD transmission [ 14 , 22 ]. However, although centrality is widely considered a valuable attribute for describing the risk, there is no consensus on how to define it [ 41 ]. Also, previous works show scenarios in which, depending on the disease nature (e.g. highly contagious) various centrality measures may result in different risk prioritizations [ 42 , 43 ]. Therefore, using a single centrality to quantify the risk of nodes in the livestock transportation network threatens loss of insights obtained from other centrality measures. This work overcomes this limitation by combining multiple centrality measures, resulting in a single ranking, see Fig 1 . Other authors proposed a similar approach for combining different nodal centrality measures methods, for instance, in the pig transportation network, to construct a risk index [ 40 ]. Our results complement these works by providing for the first time evidence of the predictive power of these indices in an actual outbreak. As Fig 4 shows, the use of the FMD risk index may successfully predict the areas with high risk in an actual outbreak. Importantly, our results suggest that this prediction capacity remains high for most months, implying that the livestock transportation-based index is quite informative about the materialization of FMD risk.
The proposed risk index may improve decisions regarding biosecurity and biocontainment in several ways. Some of the most common strategies to contain FMD outbreaks include culling, increasing biosecurity measurements, and vaccination, among others, all of which can benefit from this index [ 38 ]. For instance, in the case of FMD introduction, the proposed risk index may contribute to defining the culling area as part of the slaughter control policy [ 38 ]. In the case of an FMD outbreak, the need to act fast implies that culling should be performed on clinical and epidemiological grounds without the benefit of laboratory testing to confirm diagnoses [ 38 ]. Our results suggest that the FMD SSBR information may complement this evidence by being a good predictor of high FMD risk areas. Importantly, previous works indicate that removing these nodes reduces potential disease transmission [ 40 ]. Similarly, this FMD risk index may also inform prevention strategies, such as improved biosecurity and vaccination, by helping to focus and prioritize resources [ 21 ].
Fig 1 shows that the FMD risk index varied across months. However, the stability measure in Fig 5 suggests that the risk ranking is stable. Previous works have reported also changes in dynamic temporal networks [ 22 , 40 ], but with a different approach. The common pattern is likely related to existing transport infrastructure and commercial network trades and the variations in productive and commercial dynamics [ 44 ]. For instance, the nodes identified in the analysis matched well with the road infrastructure, and some of them, including Valledupar, show the most intense cattle commercial activities. These variations can also be exploited, for instance, to guide the design of an epidemiological surveillance system focused on nodes with intense trade between them or to plan health programs by allowing selection of relevant nodes across time [ 45 ].
The proposed ranking corresponds to an ascending-ordered list of municipalities with risks related to FMD obtained per month (see Fig 3 ). Nevertheless, practical uses of these lists in epidemiology-related tasks, such as surveillance and disease control, require focusing resources on particular municipalities. Naturally, this selection may concentrate on the top areas of the ranking as they provide a higher predictive capacity of FMD risk, see Fig 4 . However, it is worth recalling that the number of top municipalities highly depends on the particular application (for example, vigilance or control), type of resource, and resource capacities [ 46 ]. Therefore, the proposed ranking represents an additional input for improving planning in epidemiological tasks. Finally, strategies for determining the threshold (number of municipalities) from the ROC can be considered when there are no resource limitations [ 35 ].
This work has some limitations. First, it is worthy to remark that FMD risk indices are just one epidemiological tool that should be used in conjunction with other epidemiological information. Other relevant vulnerability factors, such as livestock production systems, market infrastructure, and the biophysical environment, should inform FMD control policies [ 47 ]. Second, illegal trades may increase transmission risk, especially within international borders [ 48 ]. Our work is based only on official exchanges registered by the sanitary authority, providing a limited view of the animal transportation risk. In addition, our analysis also ignores the existence of alternative transmission pathways, for instance, induced by the movement of people or fomites [ 49 ]. Future work should explore the characterization of the risk linked to these illegal trades and these alternative transmission pathways. Finally, the unit of analysis, i.e., the municipality, is quite large. Future work may include a detailed analysis of movements among farms, abattoirs and markets.
Conclusion
We studied the construction of network risk rankings to describe livestock mobility. For this, we proposed the construction of a monthly livestock mobility network characterized by different centrality measurements that capture features related to disease transmission. These measurements were combined into a single ranking to describe super-spreaders on the transportation network. Our results show for the first time that the rank computed on historical data provides a high predictive capacity for future FMD outbreaks. The proposed rankings vary across months, likely linking to variations in the commercial animal dynamic. However, the stability of rankings across most of the year is high. Further work should integrate these ranks with other epidemiological information sources to devise cost-effective control policies for FMD.
Supporting information
S1 Table
Rankings of municipalities computed for 2016 months.
(XLS)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Enterohemorrhagic Escherichia coli (EHEC) cause severe diseases in humans worldwide. Shiga toxins are regarded as their main virulence factor. However, EHEC possess various further virulence factors that mediate adherence or interfere with host defense [1] , [2] . One of these additional virulence factors is the plasmid-encoded extracellular serine protease EspP which belongs to the serine protease autotransporter of Enterobacteriaceae (SPATE) family [3] . Five subtypes of EspP have been described (EspPα-EspPε [4] , [5] , from which the translocation-competent and proteolytically active subtype EspPα (Uniprot Accession Number: Q7BSW5) is associated with highly virulent strains and isolates from patients with severe disease [4] , [6] . EspPα exhibits serine protease activity. In addition to porcine pepsin A and EHEC-Hemolysin [3] , [7] , EspPα cleaves the human plasma proteins apolipoprotein A-I, the complement factors C3 and C5, and coagulation factor V [3] , [8] , [9] . EspPα-mediated cleavage of complement factors has been demonstrated to significantly reduce complement activation [9] . In addition, the degradation of factor V has been suggested to interfere with blood coagulation possibly leading to prolonged bleeding during EHEC infection [3] .
The E . coli secreted protease, island-encoded (EspI) is a further member of the SPATE family and is secreted by Shiga toxin-producing E . coli (STEC) [8] . Notably, EspI has been found in less pathogenic E . coli serotypes [8] , [10] , [11] . The physiological function of EspI is yet unknown and to date only two substrates have been identified, namely porcine pepsin A and human apolipoprotein A-I [8] .
Serine protease inhibitors (serpins) are structurally closely related proteins which modulate different important protease cascades by irreversible inactivation of serine proteases. They are involved in inflammatory host defense, complement activation, and blood coagulation [12] , [13] . Serpins share an exposed reactive center loop (RCL) that serves as a pseudosubstrate for the target protease. Cleavage of the reactive serpin bond initiates a conformational rearrangement of the serpin structure that leads to distortion and inactivation of the target protease by formation of an irreversible covalent serpin-protease complex [14] . α1-protease Inhibitor (α1-PI, Uniprot Accession Number: P01009) is the archetypal member of the serpin family and the most abundant serpin in human plasma. Its main physiological target is neutrophil elastase [15] . α1-antichymotrypsin, (α1-AC, Uniprot Accession Number : P01011) which is closely related to α1-PI, [16] , [17] mainly inhibits cathepsin G and mast cell chymases [15] , [18] . α2-antiplasmin (α2-AP, Uniprot Accession Number: P08697) is the main physiological inhibitor of plasmin and thus influences fibrinolysis following blood coagulation [19] , [20] . Antithrombin III (ATIII, Uniprot Accession Number: P01008) inhibits thrombin, FIXa, and FXa - proteases of the blood coagulation pathway - which is considerably faster in the presence of its cofactor heparin [21] – [24] . Angiotensinogen (AGT, Uniprot Accession Number: P01019) is a non-inhibitory serpin that does not target proteases [25] . Via proteolytic processing by renin, AGT releases the vasopressor peptide angiotensin I which is further converted to angiotensin II [26] , [27] . An overview of serpin functions and nomenclature is given in Table 1 .
10.1371/journal.pone.0111363.t001 Table 1
Serpins used in this study.
Serpin
Systematicname
Main Targetproteases
Function
Reference
α1-Protease Inhibitor
SERPINA1
neutrophilelastase
Protection of tissue duringinflammation, deficiencyresults in emphysema
[15] , [56] – [58]
α1-Antichymotrypsin
SERPINA3
CathepsinG, mast cell chymases
Deficiency may result inemphysema, possiblecontribution to Alzheimer
[15] , [18] , [59] – [61]
Angiotensinogen
SERPINA8
-
Non-inhibitory, reninsubstrate, release ofangiotensin I
[25] , [62]
α2-Antiplasmin
SERPINF2
plasmin
Regulation of fibrinolysis
[19] , [45]
Antithrombin III
SERPINC1
thrombin,FIXa, FXa
Most important inhibitorof the coagulation pathway
[21] – [24]
Given are the systematic serpin name, target proteases, and general function.
Serpins are therefore highly relevant concerning their regulatory function as pseudosubstrates that inactivate serine proteases by formation of serpin-enzyme-complexes. In addition, cleavage of serpins without formation of an inhibitory complex has been described in literature for different metalloproteases. The human matrix metalloproteinase-3, e.g., cleaves α1-AC, α2-AP, and plasminogen activator inhibitor-1 [28] , [29] while human matrix metalloproteinase-9 cleaves α1-PI [30] . The bacterial 56-kDa proteinase from Serratia marcescens also cleaves α1-PI, α2-AP, ATIII, and C1 esterase inhibitor (C1-INH) [31] , [32] . C1-INH is also specifically cleaved by StcE, a metalloprotease found in highly pathogenic EHEC [33] . Surprisingly, interference of StcE with C1-INH also results in enhanced inhibition of complement-mediated lysis irrespective of cleavage of this serpin [34] , [35] . Interference with serpin function in the human host during bacterial infection is therefore a further pathogenicity mechanism.
Notably, we describe here the specific cleavage of various serpins from human plasma by the bacterial serine protease EspPα and compare this activity with the related SPATE EspI. Presented data further support the hypothesis that EspPα mediates virulence by interaction with key regulatory proteins of host defense and blood coagulation. In addition, we developed a photometrical assay for the analysis of serpin activity and applied matrix assisted laser desorption ionization-time of flight-mass spectrometry (MALDI-TOF-MS) and electrospray ionisation-fourier transform mass spectrometry (ESI-FTMS) for the direct elucidation of proteolytic cleavage sites.
Materials and Methods
Pseudonymized residual sample material from voluntary blood donations from the Transfusion medicine of the University Clinics Münster was used. Blood donors approved prior to donation that residual sample material can be used for scientific studies. The Ethics Committee of the Medical Faculty of the University of Münster was informed and approved the study design.
Proteins
EspPα was purified from clone HB101 (WB4–5k) containing espP from E . coli O157:H7 strain EDL933 [3] . The inactive EspP mutant S263A served as a negative control [36] and EspI was purified in the same way from clone DH5α/pZH4 containing espI from E . coli O91:H − strain 4797/97 [4] , [8] . Protein precipitation from culture supernatants was performed as described previously [4] . Briefly, protein pellets were dissolved in 20 mM Tris buffer containing 50 mM NaCl (pH 6.5). Proteins were purified using HiPrep 16/10 DEAE FF, HiTrap Benzamidine FF (HS), and HiPrep 16/60 Sephacryl S-200 HR columns (GE Healthcare) according to the manufactureŕs instructions. Protein preparations were diluted to 1 µg/µL with phosphate buffered saline (PBS, 100 mM NaCl, 4.5 mM KCl, 7.0 mM Na 2 HPO 4 , 3.0 mM KH 2 PO 4 , pH 7.4).
Purified serpins were purchased from Merck Millipore and dissolved according to the manufactureŕs instructions in the following buffers: α1-PI, 30 mM Na 3 PO 4 , 300 mM NaCl, pH 6.5, α1-AC, 20 mM Tris, 250 mM NaCl, 4.5 mM KCl, 7.0 mM Na 2 HPO 4 , 3.0 mM KH 2 PO 4 , pH 7,4, α2-AP, 20 mM Bis-Tris, 200 mM NaCl, pH 6.4, ATIII, 100 mM NaCl, 4.5 mM KCl, 7.0 mM Na 2 HPO 4 , 3.0 mM KH 2 PO 4 , pH 7.4, AGT, 50 mM Na 3 PO 4 , 150 mM NaCl, pH 7.0.
Plasma fractionation
Plasma samples (fresh frozen plasma, FFP) were stabilized with 17–23% (v/v) citrate-phosphate-dextrose (CPD) and were derived from whole blood donations using standard separation procedures for blood banks.
Plasma was diluted with 20 mM Na 3 PO 4 buffer (pH 7.0) and depleted using HiTrap Protein A FF and HiTrap Blue HP (GE Healthcare) according to the manufactureŕs instructions. The depleted plasma was further fractionated using HiPrep 16/10 DEAE FF via gradient elution ranging from 100% buffer A (20 mM Tris, 50 mM NaCl, pH 8.0) to 70% buffer B (20 mM Tris, 500 mM NaCl, pH 8.0). The protein fraction eluting from 15–40% buffer B was used for further experiments.
Cleavage of Substrates
To determine cleavage of substrates by EspPα or EspI, fractionated plasma (25 µg) or serpins (5 µg or 10 µg) were incubated (15 h, 37°C) with 1.5 µg of purified protease in 30 µL PBS buffer. ATIII was incubated in the same way after addition of 25 µg/mL (4.8 units/mL) unfractionated heparin (Merck). Proteins were either separated via sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), digested in-gel and analyzed using matrix assisted laser desorption ionization-time of flight-mass spectrometry (MALDI-TOF-MS) or subjected directly to MS analysis.
SDS-PAGE
After denaturation, proteins were separated on a 7.5% SDS-PAGE gel using a glycine (19.2 mM) containing buffer [37] or on a 13.3% SDS-PAGE gel using a tricine (100 mM) containing buffer [38] and stained with Coomassie Blue.
In-gel-digestion
In-gel-digestion was performed as described before [39] . Briefly, gel pieces were cut out, proteins were reduced using dithiothreitol (10 mM), alkylated with iodoacetamide (55 mM) and digested (15 h, 37°C) with trypsin (13 ng/µL, Promega). Peptides were extracted and desalted using ZipTip C 18 Pipette Tips (Merck Millipore) according to the manufactureŕs instructions. Peptides were eluted with 40% acetonitrile (MeCN)/1% formic acid (FA) and 70% MeCN/1% FA (5 µL each) and eluates were combined.
Mass spectrometric analysis
In-gel-digests or incubation mixtures (0.5 µL) were mixed with 0.5 µL α-cyano-4-hydroxycinnamic acid (Sigma-Aldrich, 10 µg/µL in 50% MeCN/1% trifluoroacetic acid) and 0.5 µL of the mixture were spotted on a MALDI target (MTP 384 target plate ground steel, Bruker). Samples were analyzed using a Bruker autoflex speed in positive mode.
To determine the accurate masses of the largest α2-AP fragment, the incubation mixture was desalted using ZipTip C 18 Pipette Tips as described before and measured using a Thermo LTQ Orbitrap XL in positive static nanospray mode (sheath gas flow rate 15 arb.u., aux gas flow rate 10 arb.u., sweep gas flow rate 5 arb.u.).
Determination of EspPα and α1-PI activity
Potential functional consequences of the interaction between α1-PI and EspPα were analyzed by measuring the activities of both proteins after coincubation. To investigate effects of α1-PI on EspPα protease activity, both proteins were incubated together (15 h, 37°C) at equimolar concentrations. Preincubated (15 h, 37°C) EspPα or α1-PI were used as controls. The remaining EspPα protease activity was then determined by the incubation (15 h, 37°C) of an aliquot containing 1 µg EspPα (either preincubated alone or with α1-PI) with 2 mM of the chromogenic peptide substrate Suc-Ala-Ala-Pro-Leu-pNA (Bachem) in 100 µL PBS (pH 7.4) and 5% dimethyl sulfoxide (DMSO). Active EspPα releases para-nitroaniline (pNA) from the peptide which is detected at 405 nm using a FLUOstar Optima plate reader (BMG Labtech). PBS was used as a buffer control.
The effect of EspPα-mediated cleavage on α1-PI serpin activity was determined by coincubation (15 h, 37°C) of α1-PI and EspPα in a molar ratio of 4∶1. Again, incubations (15 h, 37°C) of EspPα or α1-PI alone were used as controls. To assess remaining serpin activity of α1-PI, the coincubation mixture and controls were incubated (5 h, 37°C) with trypsin (Promega) at a molar ratio of α1-PI and trypsin of 4∶1. Active α1-PI inhibits trypsin protease activity. The remaining serpin activity was therefore assessed indirectly by determination of reduced trypsin activity using aliquots of coincubation mixtures and controls containing 0.25 µg trypsin and incubation (2 h, 37°C) with 2 mM of the chromogenic peptide Bz-Arg-pNA (Bachem) in 100 µL PBS (pH 7.4) containing 5% DMSO. Active trypsin releases pNA and absorbance was measured at 405 nm using a FLUOstar Optima plate reader. PBS was used as a buffer control.
Results and Discussion
Purification of EspPα and S263A
EspPα and the inactive EspPα mutant S263A were purified from culture supernatants using ammonium sulfate precipitation and liquid chromatography. Purity was verified via SDS-PAGE ( Fig. 1 , lane 5 and 6). EspPα shows a band at ∼104 kDa representing the intact EspPα and a band at ∼80 kDa which was identified by MALDI-TOF-MS as autoproteolysis product. S263A samples showed a pronounced protein band at ∼104 kDa and a weaker band at ∼85 kDa which was identified as a truncated form of S263A. The autoproteolyis product of EspPα remains active even after long term incubation ( Figure S1 ). Proteolytic activity of purified EspPα and the inactive S263A were assessed using a chromogenic oligopeptide substrate. As expected, all EspPα samples were proteolytically active while S263A showed no proteolytic activity ( Fig. S1 ).
10.1371/journal.pone.0111363.g001 Figure 1
Identification of substrates in plasma.
Fractionated plasma (25 µg) was incubated (15 h, 37°C) with EspPα or S263A (1.5 µg) and separated via SDS-PAGE using a glycine buffer. M, molecular weight marker, *, EspPα autodegradation product, α, α1-PI, α*, α1-PI degradation product.
Identification of EspPα substrates in plasma
To identify physiological relevant substrates of EspPα, fractionated plasma was incubated either with EspPα or the EspPα negative control S263A ( Fig. 1 , lane 3 and 4). Incubation with EspPα resulted in loss of a pronounced 50 kDa band in plasma and the occurrence of a degradation product with a molecular weight of ∼45 kDa in SDS-PAGE. The according protein band was digested in-gel and subjected to MALDI-TOF-MS analysis and unambiguously identified as α1-PI (Aldente score 235.7, sequence coverage 69% to α1-PI (UniProtKB: P01009)).
EspPα cleaves various serpins
To determine if further serpins are cleaved by EspPα, different serpins were incubated with EspPα or S263A and cleavage was monitored by SDS-PAGE. EspPα degrades α1-PI, α1-AC, and the non-inhibitory serpin AGT into a large (>40 kDa) and a small (<10 kDa) fragment ( Fig. 2 a–f), while incubation of α2-AP leads to several degradation products ( Fig. 2 g, h). None of the incubations led to pronounced formation of an inhibitory serpin-enzyme complex. Interestingly, the anticoagulatory serpin ATIII was not degraded by EspPα ( Fig. 2i, j ).
10.1371/journal.pone.0111363.g002 Figure 2
Cleavage of various serpins by EspPα.
Serpins (5 µg) were incubated (15 h, 37°C) with EspPα or S263A (1.5 µg). Degradation products were separated via SDS-PAGE using a glycine buffer (a, c, e, g, i) or tricine buffer system (b, d, f, h, j). Proteolytic serpin fragments formed by EspPα are indicated by an arrow. a, b α1-PI is degraded to a large and small fragment (∼45 kDa and ∼4 kDa, respectively), c, d cleavage of α1-AC in two fragments, e, f the AGT band with the highest molecular weight is cleaved in two fragments, g, h large and small fragments (∼55–57 kDa and ∼4–7 kDa) formed by α2-AP cleavage. i, j ATIII is not cleaved by EspPα. Incubation of α1-PI with EspPα leads to a weak formation of an inhibitory enzyme-serpin complex as marked by **. M, molecular weight marker, *, autodegradation product of EspPα.
Activity of α1-PI and EspPα after incubation
We next determined the functional consequences of the coincubation of serpin and SPATE protease by use of the bona fide serpin α1-PI and EspPα. The remaining EspPα-activity following incubation with α1-PI was assessed in a photometrical assay using the chromogenic EspPα substrate Suc-Ala-Ala-Pro-Leu-pNA. Incubation with α1-PI had no influence on the proteolytic activity of EspPα ( Fig. 3a ), demonstrating that α1-PI does not target EspPα.
10.1371/journal.pone.0111363.g003 Figure 3
Activity of EspPα and α1-PI after coincubation.
a, Determination of EspPα activity. EspPα and α1-PI were preincubated (15 h, 37°C) at equimolar concentrations and remaining activity of EspPα was analyzed by incubation of an aliquot of the mixture with the chromogenic substrate Suc-Ala-Ala-Pro-Leu-pNA. Activity was measured via released para -nitroaniline and normalized to EspPα. n = 9 for EspPα and EspPα+α1-PI or n = 6 for α1-PI, respectively. b, α1-PI activity (measured as inhibitory potential on trypsin) after incubation with EspPα. α1-PI and EspPα or S263A were preincubated at a molar ratio of serpin∶enzyme = 4∶1. Remaining inhibitory activity of α1-PI on trypsin was analyzed by incubation at a molar ratio of α1-PI∶trypsin = 4∶1. Trypsin activity was measured via release of para -nitroaniline from the chromogenic substrate Bz-Arg-pNA. c, SDS-PAGE analysis of conincubations. α1-PI, EspPα, S263A, and trypsin were incubated as in b) and mixtures were separated via SDS-PAGE (12% SDS-PAGE gel, glycine buffer). M, molecular weight marker, *, EspPα autodegradation product, **, inhibitory complex of α1-PI and trypsin, +, trypsin was directly subjected to SDS-PAGE without incubation.
The remaining inhibitory potential of α1-PI following incubation with EspPα was analyzed using trypsin as a serpin target. Although neutrophil elastase is the physiological target for α1-PI, trypsin also forms an irreversible inhibitory complex with the serpin and can therefore be used as an indicator for α1-PI activity [40] . Active α1-PI inhibits the proteolytic activity of trypsin and consequently loss of α1-PI serpin activity results in high proteolytic activity in the assay. Trypsin activity was determined by photometrical detection of the cleavage of the trypsin substrate Bz-Arg-pNA.
Incubation of trypsin with α1-PI or α1-PI preincubated with S263A resulted in nearly complete loss of trypsin activity ( Fig. 3b ), demonstrating that the employed α1-PI shows high serpin activity and that the inactive EspPα mutant S263A does not affect α1-PI. In contrast, α1-PI preincubated with EspPα did not reduce trypsin activity in the following assay ( Fig. 3b ). This demonstrates that EspPα-mediated α1-PI cleavage leads to loss of the inhibitory serpin activity. Corresponding results were obtained using SDS-PAGE ( Fig. 3c ). Incubation of α1-PI with trypsin leads to the formation of a serpin-enzyme-complex ( Fig. 3c , lane 10). After incubation with EspPα, α1-PI is not able to form this complex with trypsin. Instead, the large α1-PI fragment is further degraded by trypsin ( Fig. 3c , lane 6). EspPα as well as S263A were completely degraded when incubated with trypsin, demonstrating that neither EspPα nor S263A directly interfere with trypsin activity ( Fig. 3c , lanes 2 and 4). In addition, α1-PI does not interact with S263A (no serpin enzyme complex) ( Fig. 3c , lane 7) but is cleaved by EspPα ( Fig. 3c , lane 5). The addition of trypsin to the mixture of α1-PI and S263A led to incomplete degradation and occurrence of several degradation bands in SDS-PAGE. This is due to the fact that degradation of S263A by trypsin and the inhibition of trypsin by α1-PI occur in parallel resulting in only incomplete S263A degradation ( Fig. 3c , lane 8).
EspPα cleaves inside the reactive center loop
The loss of activity of α1-PI but not EspPα is based on cleavage of α1-PI without formation of an inhibitory serpin-enzyme-complex. To further understand how EspPα-mediated cleavage affects the inhibitory function, we determined the cleavage sites in α1-PI and the other serpins included in this study. To this end, large and small fragments of cleaved serpins were separated using SDS-PAGE, in-gel-digested and subjected to MALDI-TOF-MS analysis. Figure 4 shows the peptide mapping of EspPα cleavage products of α1-PI. The large α1-PI fragment consists of the N -terminal part of the serpin ( Fig. 4a and b ), while the C -terminal part from residue 383 to 418 forms the small fragment ( Fig. 4 a, and c). EspPα cleavage occurs at the active site of the serpin between 382 Met and 383 Ser as demonstrated by the occurrence of the non-tryptic peptide 1′(SIPPEVK) and the complete sequence coverage for the small fragment ( Fig. 4c ). Sequence coverage of degradation products of the other serpins are given in Figure S2 .
10.1371/journal.pone.0111363.g004 Figure 4
Peptide mapping of EspPα cleavage products of α1-PI.
α1-PI fragments were subjected to tryptic in-gel-digest and generated peptides were analyzed via MALDI-TOF-MS. a, Sequence coverage of α1-PI fragments. Peptides of the large fragment are given in bold and numbered 1–25. Peptides of the small fragment are given in italics and numbered 1′-6′. Note the newly formed N -terminus of the small fragment (SIPPEVK, underlined). b, MALDI-TOF-MS spectrum of the large fragment of α1-PI. Inset: SDS-PAGE gel, glycine buffer. Fragment used for peptide mapping is marked by arrow. c, MALDI-TOF-MS spectrum of the small fragment of α1-PI. Inset: SDS-PAGE gel, tricine buffer. Fragments used for peptide mapping are marked by arrow. α1-PI peptides are numbered according to a, T, trypsin autoproteolysis products, E, EspPα autoproteolysis products.
Direct MALDI-TOF-MS analysis of small fragments
Not all cleavage sites can be identified via in-gel-digest. Tryptic peptides might be too small when cleavage occurs close to lysine or arginine residues or when several cleavage sites are in close proximity to each other. As all small fragments formed by EspPα-cleavage show a molecular weight below 10 kDa, we applied direct MALDI-TOF-MS analysis to determine the exact mass of the small serpin fragments to elucidate and confirm cleavage sites ( Fig. 5 ). For the small α1-PI fragment we observed a signal for the proton adduct of the α1-PI sequence 383 Ser- 418 Lys (m/z 4133.333) confirming the cleavage site determined via in-gel-digest. In addition, signals representing the Na + adduct and the oxidized Na + adduct of the according α1-PI fragment sequence were observed ( Fig. 5a ). α1-AC shows a similar spectrum with a pronounced signal at m/z 4623.419 demonstrating cleavage C-terminal of 383 Leu at the reactive bond ( Fig. 6b ), which is in good accordance with data from in-gel-digest ( Figure S2 ). For AGT, we already observed three bands in SDS-PAGE (intact AGT and two non-proteolytic fragments) when incubated without protease ( Fig. 2e ). Accordingly, signals of two small AGT fragments were observed in MALDI-TOF-MS ( Fig. 5c , right lane). Incubation with EspPα led to degradation of intact AGT and occurrence of the corresponding small fragment in MALDI-TOF-MS ( Fig. 2e and Fig. 5c , left lane). For α2-AP, proteolytic cleavage into several fragments is observed in SDS-PAGE (see Fig. 2g and 2h and Fig. 5d ) after incubation with EspPα. Four distinct signals are seen in the MS spectrum indicating 4 cleavage sites. As the resolution for the signal at m/z 5308.3 is too low to determine the monoisotopic mass, we measured this sample in addition via nanospray-ESI-FTMS. Table 2 summarizes EspPα cleavage sites and their positions within the respective serpin. Measurement of α2-AP after incubation with EspPα via nanospray ESI-FTMS is described in Table 3 .
10.1371/journal.pone.0111363.g005 Figure 5
Direct analysis of the small cleavage product of serpins via MALDI-TOF-MS.
Serpins were incubated with EspPα and directly analyzed via MALDI-TOF-MS. a, MALDI-TOF-MS spectrum of α1-PI fragment. Inset: Detailed view of the signal representing the small α1-PI fragment. b, MALDI-TOF-MS spectrum of α1-AC. c, MALDI-TOF-MS spectrum of AGT. Left lane: Spectrum after incubation with EspPα, right lane: Spectrum after incubation of AGT without EspPα. *, signals represent non-proteolytic fragments also found after incubation of AGT without EspPα. d, MALDI-TOF-MS spectrum of α2-AP. Inset: Detailed view of the m/z window 2160–2260 representing signals (M H), (M Na), (M Na+O) of the cleavage site in the N -terminal extension of α2-AP are exemplarily shown. (M H), proton adduct of small serpin fragment, (M Na), Na adduct of small serpin fragment, (M Na+O), Na adduct oxidized at one methionine residue.
10.1371/journal.pone.0111363.g006 Figure 6
Cleavage of serpins by EspI.
Serpins (5 µg) were incubated (15 h, 37°C) with EspI (1.5 µg). Degradation products were separated via SDS-PAGE using a glycine buffer (a, c, e, g, i) or a tricine buffer (b, d, f, h, j). a, b α1-PI is cleaved into two fragments (∼45 kDa and ∼4 kDa), c, d α1-AC is cleaved into two fragments, e, f AGT is not cleaved by EspI, g, h α2-AP is not cleaved by EspI, i, j ATIII is cleaved only with very low efficiency. Note the formation of inhibitory serpin-enzyme-complexes after incubation with α1-PI and α1-AC. M, molecular weight marker, *, autodegradation product of EspI, **, inhibitory serpin-EspI-complex. Serpin fragments are indicated by an arrow.
10.1371/journal.pone.0111363.t002 Table 2
Serpin cleavage sites determined by MALDI-TOF-MS.
Serpin
m/z determined
Theoretical mass
Deviation (ppm)
Sequence
Position
α1-PI
4133.333
4133.234
+24
380 IPM-SIP 385
Reactive bond
α1-AC
4623.419
4623.495
−16
381 TLL-SAL 386
Reactive bond
AGT
4299.351
4299.293
+14
444 QQL-NKP 449
Reactive center loop
α2-AP
2181.123
2181.097
+12
45 SPL-TLL 50
N -terminal extension
α2-AP
3489.789
3489.788
<1
458 QSL-KGF 463
C -terminal extension
α2-AP
3602.870
3602.872
−1
457 LQS-LKG 462
C -terminal extension
α2-AP
5308.3 (average)
5307.9 (average)
+75
442 REL-KEQ 447
C -terminal extension
Given are masses determined by MALDI-TOF-MS directly after incubation of serpin with EspPα, theoretical masses, mass deviation, according sequence, and position inside the serpin sequence. Numeration is according to the serpin precursor. 10.1371/journal.pone.0111363.t003 Table 3
α2-AP cleavage site determined by ESI-FTMS.
m/z determined
m/z theoretical
Charge state (z)
Deviation (ppm)
Sequence
Position
884.9507
884.9512
6
−1
442 REL-KEQ 447
C -terminal extension
758.6725
758.6736
7
−1
442 REL-KEQ 447
C -terminal extension
663.9624
663.9653
8
−4
442 REL-KEQ 447
C -terminal extension
Given are masses of the large α2-AP fragment as determined by nanospray ESI-FTMS.
α1-PI and α1-AC are cleaved at their reactive bonds (position of reactive sites are described in [41] , [42] ), leading to loss of serpin function. In both molecules the reactive bonds are exposed in the RCL and serve as pseudosubstrates for the targeted proteases. In case of EspPα, the serpins are not able to form a stable inhibitor-enzyme-complex and therefore release the intact EspPα after cleavage. Although AGT as non-inhibitory serpin does not contain a reactive bond, it is structurally closely related to the other serpins and is also cleaved in the RCL, indicating that a reactive bond is not necessary for EspPα-mediated serpin degradation. This is further underlined for α2-AP, which is cleaved at four positions outside the RCL (for RCL position see [43] ). Cleavage sites are located at the N - and C -terminal extensions 25 aa downstream the N -terminus and 46, 31, and 30 aa upstream the C -terminus (see Table 2 ). Intriguingly, both the N - and C -terminal extensions are vital for the functional relevant binding of α2-AP to other proteins [19] , [44] , [45] .
Cleavage of serpins by EspI
Purified EspI samples showed a protein band at ∼110 kDa (intact EspI) as well as two EspI autoproteolysis products at ∼50 and 45 kDa, respectively. Similar to EspPα, autoproteolysis products remain active. Serpins were incubated with purified EspI in the same way as described for EspPα. Incubation of α1-PI and α1-AC with EspI led to degradation of these serpins. Notably, EspI also forms a pronounced inhibitory complex with both protease inhibitors resulting in only incomplete serpin degradation ( Fig. 6 a-d). In contrast to EspPα, EspI does not cleave α2-AP and AGT ( Fig. 6e–h ). Cleavage of ATIII occurred only with very low efficiency ( Fig. 6i ) and might not be relevant under physiological conditions.
To determine the cleavage sites of α1-PI and α1-AC, we subjected incubation mixtures of serpins and EspI to direct MALDI-TOF-MS analysis. Serpin cleavage occurred at the reactive bond leading to signals at m/z 4155.400 (α1-PI, 20 ppm deviation according to calculated m/z) and 4623.509 (α1-AC, 19 ppm deviation according to calculated m/z), respectively (data not shown).
Conclusions
EspPα is an EHEC virulence factor that belongs to the SPATE family. As suggested for SPATEs in general, EspPα most likely mediates its virulence via cleavage and inactivation of host proteins. Here, we present a method for the rapid determination of EspPα-mediated cleavage sites in various human plasma serpins via MALDI-TOF-MS as well as a photometrical assay to analyze serpin functionality after proteolytic cleavage. Concerning the functional consequences, degradation of α2-AP might lead to bleeding disorders. This serpin is the primary physiological inhibitor of plasmin and deficiency has been shown to result in uncontrolled fibrinolysis and severe hemorrhagic complication [44] , [45] . α2-AP harbors a 42 aa N -terminal and a 55 aa C -terminal extension [19] , [46] . While the N -terminal extension is cross-linked to fibrin, the very C -terminal 491 Lys residue mediates binding to plasmin [47] . EspPα cleaves between 47 Leu and 48 Thr releasing part of the N -terminal extension and at three different sites inside the C -terminal extension leading to release of a polypeptide containing 491 Lys. Together, this most likely leads to loss of function of α2-AP. The role of α1-PI in thrombosis is not well understood. However, α1-PI is able to inhibit activated protein C. In pediatric ischemic stroke patients elevated levels of α1-PI have been found and were discussed to contribute to this thrombotic disease in children [48] , [49] . ATIII is the main anticoagulatory serpin. Although it is able to interfere with virtually all proteolytic coagulation factors, its main targets are thrombin, FIXa, and FXa. Intriguingly, it is the only serpin in this study that is not cleaved by EspPα. Despite the structural similarity of serpins, EspPα specifically cleaves only selected serpins. More specific, procoagulatory serpins such α2-AP and α1-PI are efficiently degraded while the anticoagulatory ATIII is not affected at all. Together with data demonstrating that EspPα cleaves coagulation factor V [3] , this underlines the hypothesis that interference with blood coagulation (and possibly also inflammatory host responses) [50] might be one of the major functions of EspPα which might contribute to formation of hemorrhages observed during EHEC infection.
Having a closer look at EspPα cleavage sites, it is notable that more than 70% (5 of 7) of cleavage sites identified in this study occur after Leu. This is in good accordance to already reported EspPα cleavage sites [3] , [9] , [7] , [51] , indicating that substrate cleavage is most favorable C -terminal to Leu. In α2-AP, cleavage also occurs after 459 Ser. This residue, however, is positioned next to 460 Leu after which EspPα cleaves, too. The second non-Leu cleavage site is C-terminal to 382 Met in α1-PI. The 382 Met- 383 Ser bond, however, is the reactive bond exposed in the RCL and required to react with target proteases. Similarly, α1-AC is cleaved at the reactive bond that consists of a Leu-Ser motif which is also located in the exposed RCL. Cleavage of the non-inhibitory AGT shows that a reactive bond is not strictly required for substrate recognition by EspPα but cleavage also occurs inside the corresponding reactive center loop. In contrast, α2-AP is not cleaved in the RCL but inside the N - and C -terminal extensions which are vital for α2-AP functionality. Though the crystal structure of α2-AP has only been solved for a N -terminally truncated murine form, it seems that the C -terminal extension consists of a flexible loop because it could not be modeled into electron density maps [52] . Perhaps, this structural flexibility seen in the reactive center loops and in the C -terminal extension of α2-AP is required for substrate recognition by EspPα. Figure 7 shows crystal structures of the serpins that are cleaved by EspPα [52] – [55] .
10.1371/journal.pone.0111363.g007 Figure 7
Crystal structures of serpins cleaved by EspPα.
Serpins are shown as cartoons. RCL is indicated in black, approximate cleavage sites are encircled. Non-resolved parts of the crystal structures are indicated by dots (c, RCL of AGT, d, RCL of α2-AP and the N - and C -terminal extension of α2-AP). a, human α1-PI, b, cleaved human α1-AC, the RCL is indicated by dots, c, human angiotensinogen, d, murine truncated α2-AP Δ43 , the N -terminal extension of native α2-AP is indicated by dots.
EspI shows significant differences in substrate specificity compared to EspPα. α1-PI and α1-AC are also cleaved at their reactive bonds which should lead to loss of function of these serpins. However, serpin cleavage and release of the protease is not complete for EspI, most probably due to the pronounced formation of an inhibitory serpin-enzyme-complex of EspI with α1-PI and α1-AC. In contrast, EspPα completely degrades both serpins and forms only small amounts of the inhibitory complex only with α1-PI which does not significantly reduce EspPα activity. In addition, AGT and α2-AP, which are degraded by EspPα at positions other than the reactive bond, are not degraded by EspI. Concerning the functional differences of both SPATE proteases, EspPα is able to cleave serpins specifically within accessible loop structures and is notably not inhibited by the analyzed serpins, while EspI is only able to interact with the reactive bond of α1-PI and α1-AC. The latter interactions show equilibria between EspI inhibition and serpin degradation. Taking into account the high amounts of serpins such as α1-PI in plasma, EspI activity might be strongly reduced in this milieu in vivo, while serpin degradation and inactivation might be a relevant function of EspPα also during infection.
In summary, we established a rapid method to determine cleavage sites of small proteolytic fragments via MALDI-TOF-MS. Functional implications have been investigated in a newly developed photometrical assay using chromogenic peptide substrates. EspPα degrades and thereby inactivates different plasma serpins which, in case of α2-AP, might lead to bleeding disorders or in case of α1-PI and α1-AC might interfere with the acute phase reaction during inflammatory host response. Cleavage occurs in flexible regions most favorable C -terminal to Leu. Comparison of EspPα and EspI indicate different functions of this SPATE also in vivo.
Supporting Information
Figure S1
Activity of EspPα and S263A. a, Determination of EspPα and S263A activity directly after purification. EspPα or S263A was incubated (15 h, 37°C) with the chromogenic substrate Suc-Ala-Ala-Pro-Leu-pNA. Activity was measured via released para -nitroaniline and normalized to EspPα. PBS was used as control. n = 2, b, Determination of EspPα activity after preincubation. Purified EspPα was preincubated for 15 h at 37°C resulting in the formation of autoproteolysis products (see Fig. 3c , lane1). To assess remaining proteolytic activity of autoproteolysis products the preincubated sample was incubated with the chromogenic substrate Suc-Ala-Ala-Pro-Leu-pNA (15 h, 37°C). Again, activity was measured via released para -nitroaniline and normalized to EspPα. PBS was used as control. n = 2.
(TIFF)
Figure S2
Peptide mapping of EspPα cleavage products of the serpins. Serpin fragments were subjected to in-gel-digest and analyzed via MALDI-TOF-MS. Peptides of the large fragment are given in bold. Peptides of the small fragments are given in italics, a, sequence coverage of α1-AC fragments, b, sequence coverage of AGT, c, sequence coverage of α2-AP. Note that in the small fragments of AGT and α2-AP no serpin peptides were found.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Inclusion bodies (IBs) are dense, electron-refractile particles of aggregated proteins found in the cytoplasmic space of bacterial cells [1] . Hydrophobic heterologous proteins expressed at high levels in bacterial cells are likely to accumulate in IBs [2] , [3] . IBs vary in diameter from 0.5–1.3 µm and are more dense (∼1.3 mg/mL) than many other cellular components, which make the particles easy to separate from disrupted cells by high-speed centrifugation for protein refolding [4] , [5] .
In general, the proteins in IBs are functionally inactive. However, recent studies have shown that they are not necessarily inactive, and some exhibit substantial levels of activity in E. coli [2] , [3] , [6] – [8] . For example, certain enzymes fused to a viral capsid protein or an ionic self-assembling peptide generated active IBs that had high levels of catalytic activity [6] – [8] . Accordingly, we found that a family II cellulose binding domain (CBD) from Cellulomonas fimi induced the formation of active IBs when fused with β-glycoside hydrolyzing enzymes. The enzymatic activity of these IBs was 30%–40% of that of the soluble enzymes [9] . In addition, a family IIIa CBD has also been used to form active IBs with high D-amino acid oxidase activity [10] . The family II CBD in IBs also exhibited significant binding affinity towards insoluble celluloses [9] .
In this study, the family II CBD from C. fimi was used to generate IBs displaying functional leucine zipper proteins (LZs) as bait for localizing soluble cytosolic proteins in E. coli ( Fig. 1A ). LZs are universal, two-stranded, α-helical heterodimers that are found in diverse DNA binding proteins and dimerization domains [11] , [12] . Therefore, the heterodimer formation between LZs was expected to recruit soluble, functionally active proteins to IBs ( Fig. 1B ). As a soluble model protein, monomeric red fluorescent protein 1 (mRFP1) [13] was used to allow for rapid and quantitative analysis in living cells. Imaging and flow cytometric analyses showed that protein localization increased according to the binding affinity between the LZ proteins, consistent with the observations of a report that showed that dimerization of coil proteins caused the co-purification of soluble enzymes in IB fractions [14] . Our affinity-based localization of cytosolic proteins to active IBs is expected to be useful for many biotechnology applications: for example, the in vivo matrix can be used to localize enzymes for sequential reactions to the same locations in cells, thereby adjusting the local concentration of the enzymes and reducing intermediate loss through diffusion and side reactions [15] – [17] . In addition, as the localization of interacting proteins to IBs can be easily identified, this study provided a new platform for investigating protein-protein interactions in living cells, using fluorescence microscopy or flow cytometry [18] .
10.1371/journal.pone.0097093.g001 Figure 1
Controlled localization of functionally active proteins to inclusion bodies (IBs) using leucine zippers (LZs).
A. Interactions between anti-parallel leucine zippers (LZ). Dashed lines indicate the charge-charge interactions of glutamic acid (E) and lysine (K). B. Representation of controlled localization in bacterial cell. The red fluorescent protein, which is dispersed throughout the cytosol, is localized to IBs by a specific molecular interaction.
Materials and Methods
Materials
The family II CBD was cloned from the exoglucanase (cex) of Cellulomonas fimi KCTC 9143. The EGFP gene was obtained from the commercial plasmid pEGFP-N1 (Clontech, Mountain View, CA, USA). The pRFP plasmid, which contains the gene for monomeric red fluorescent protein 1 ( mrfp ) [13] , was a kind gift from Dr. R. Tsien (UCSD, USA). Genes encoding two anti-parallel LZs, used for the bait and prey, were cloned from pET11a-Z-NGFP and pMRBAD-Z-CGFP [12] , respectively, which were provided by Dr. L. Regan (Yale University, USA). E. coli DH5α (Takara Bio, Ohtsu, Japan) and BL21(DE3) (Novagen, Gibbstown, NJ, USA) cells were used as the cloning host and the expression host, respectively. All restriction enzymes were purchased from Roche Applied Science (Indianapolis, IN, USA), and T4 DNA Ligase was purchased from Fermentas (Glen Burnie, MD, USA).
DNA manipulation
All primers were synthesized by Bioneer Co. (Daejeon, Korea) ( Table S1 ). The EGFP gene was amplified from pEGFP-N1, and then cloned into the Nde I and Xho I sites of pET21a (Invitrogen, Carlsbad, CA, USA) to yield pEGFP. The EGFP-CBD gene was prepared using overlap PCR and was inserted into the Nde I and Hin dIII sites of pET21a to yield plasmid pEGFP-CBD. The bait and prey LZs were fused to the EGFP-CBD and mRFP genes, respectively, by overlap extension PCR ( Fig. S1 ). The resulting bait-EGFP-CBD and prey-mRFP genes were then inserted into pET21a to yield pCN20-CBD. pC20-CBD, a bait-less variant of pCN20-CBD was constructed by using EGFP-CBD instead of bait-EGFP-CBD . Four variants of the pCN20-CBD plasmid (pCN8-CBD, pCN31-CBD, pCN50-CBD, and pCN1000-CBD) were constructed by introducing known mutations into the prey moiety as shown in Table 1 [12] , using a QuikChange mutagenesis kit (Stratagene, La Jolla, CA, USA).
10.1371/journal.pone.0097093.t001 Table 1
Amino acid sequences and affinity information of mutant CZs.
No.
CZ peptide
Mutations
K D (µM)
CN8-CBD
EQL K KKLQALEKKLAQLEWKNQALEK E LAQ
4/27
8
CN20-CBD
EQLEKKLQALEKKLAQLEWKNQALEKKLAQ
None
20
CN31-CBD
EQLEKKLQALEKKLAQLEWKNQAL K KKLAQ
25
31
CN50-CBD
EQLEKKLQAL K KKLAQLEWKNQALEKKLAQ
11
50
CN1000-CBD
EQLEKKLQALEK E LAQLEWKNQAL K K E LAQ
13/25/27
1000
Mutation sites are underlined. The CZ peptide sequences and KD's were adopted from the results of Magliery TJ et al. [12] .
Protein expression and western blotting analysis
E. coli BL21(DE3) cells were cultivated at 37°C in LB medium containing ampicillin (50 µg/mL). Protein expression was induced with 0.5 mM IPTG when the cultures reached an OD 600 of 0.5, and the cells were incubated for an additional 6 h. The cells were harvested by centrifugation at 16,300× g for 10 min and then disrupted by sonication on ice.
The protein expression was analyzed by SDS-PAGE and western blotting. Aliquots of cell lysates were electrophoresed on 12% SDS-polyacrylamide gels and electro-transferred to polyvinylidene fluoride membranes (Millipore, Billerica, MA, USA). The membranes were hybridized with an anti-GFP mouse antibody (Sigma-Aldrich, St. Louis, MO, USA) and an anti-groEL antibody as the internal standard (Abcam, Cambridge, MA, USA), followed by an HRP-conjugated anti-mouse IgG goat antibody (Bio-Rad, Hercules, CA, USA) prepared in TBST buffer (20 mM Tris-HCl, 100 mM NaCl, and 0.1% Tween-20, pH 7.5) containing 5% skimmed milk. The hybridized bands were identified by colorimetric detection using an Opti-4CN substrate kit (Bio-Rad).
Imaging and fluorescence analyses
Cells were observed with an Axio Observer microscope (Carl Zeiss, Oberkochen, Germany) at ×1,000 magnification under differential interference contrast (DIC) imaging conditions. Fluorescence imaging was also performed using the same microscope fitted with a GFP filter (excitation BP 470/20, beam splitter FT 493, emission BP 505–530) and a rhodamine filter (excitation BP 546/12, beam splitter FT 580, emission LP 590) for EGFP and mRFP, respectively. Image acquisition and region-of-interest analyses were performed using MetaMorph software (Molecular Devices, Sunnyvale, CA, USA). At least 5 cells per image were selected and subjected to region-of-interest analyses. All ROI data were presented as means ± standard error of the mean.
Flow cytometry
Flow cytometric analyses were performed using a FACSCalibur flow cytometer (BD Biosciences, Franklin Lakes, NJ, USA). The gate was set based on side scatter channel (SSC) and forward scatter channel (FSC) parameters, and the EGFP and mRFP signals were detected using FL1 (530/30 nm) and FL2 (585/42 nm) photomultiplier tubes (PMTs), respectively. The overlap of the EGFP and mRFP signals was minimized using a compensation option. A total of 10 4 cells were counted for each sample and the data were collected using BD CellQuest Pro software (version 4.0.2; BD Biosciences). Cell sorting was performed using a FACSAria Cell Sorter (BD Biosciences) at KRIBB, Jeonbuk Branch (Jeongeup, Korea).
Electron microscopy and Zeta-potential analysis
For SEM imaging, purified CBD-IBs were fixed in a mixture of 2.5% paraformaldehyde and 2.5% glutaraldehyde in a 100-mM sodium phosphate buffer (pH 7.2) for 2 h, post-fixed with 1% osmium tetroxide in the same buffer for 1 h, dehydrated in graded ethanol, substituted with isoamyl acetate, and then critical point dried in CO 2 . The samples were then coated with gold in a SC502 sputter coater (Quorum Technologies Ltd, East Sussex, UK) and observed under a Quanta 250 FEG scanning electron microscope (FEI, Hillsboro, OR, USA) at KRIBB (Daejeon, Korea).
The size and zeta-potential of the EGFP-IBs were measured using a Malvern Zetasizer Nano ZS (Malvern Instruments, Malvern, UK) at the National Nanofab Center (Daejeon, Korea). The protein solution was diluted with 10 mM Tris-HCl (pH 8.0), and 0.75 mL of the diluted solution was added to disposable zetasizer cuvettes for the measurements. The experiments were performed in triplicate and the data were processed using Zetasizer Nano software (version 6.01; Malvern Instruments).
Results
Generation of functional IBs
The CBDs include three to four aromatic residues that are exposed to bulk liquid on the surface of the protein ( http://www.pdb.org ; PDB ID: 1exg) [19] , which may cause rapid aggregation of the protein. As previously mentioned, C-terminal fusions of the family II CBD from C. fimi formed active IBs retaining 30%–40% of the original activity while maintaining the ability to bind insoluble celluloses [9] . In the current study, E. coli cells expressing a fusion of the CBD with EGFP exhibited one or two fluorescent IBs in microscopic images ( Fig. 2A ). When cells expressing either EGFP or EGFP-CBD were compared by flow cytometry, the fluorescence intensity of the EGFP-CBD cells was estimated to be 10%–20% of that in cells expressing soluble EGFP ( Fig. S2 ), although the expression of both proteins (as detected by western blotting) was similar. When the EGFP-CBD cells were sonicated in Tris buffer (50 mM Tris-HCl, pH 8.0 and 200 mM NaCl) to break the IBs into smaller pieces, the fluorescence intensity increased up to 2 folds in proportion to the sonication time ( Fig. 2B ). Therefore, the IBs are estimated to contain higher amounts of properly folded/native-like protein than that observed in flow cytometry. The low detection of fluorescence in IBs is discussed further in the Discussion section.
10.1371/journal.pone.0097093.g002 Figure 2
Microscopic observation of controlled localization to CBD IBs.
A. Imaging expressed EGFP and EGFP-CBD in E. coli . The left panel represents the western blot images after treatment with anti-GFP and anti-GroEL antibodies. The EGFP band is indicated by the black arrows. Scale bar = 5 µm. B. Increased fluorescence following sonication of EGFP-CBD IBs. E. coli cells expressing EGFP-CBD were treated by sonication in a Tris buffer (50 mM Tris-HCl, pH 8.0, 200 mM NaCl) and the fluorescence intensity analyzed using a Cary Eclipse fluorometer. The inset represents a SEM image of the EGFP-CBD IBs. C. Microscopic images of E. coli cells with no interaction (top) and interaction (bottom) between LZs. Scale bar = 5 µm.
Localization of soluble proteins to IBs
The possibility of active IBs as a matrix to recruit soluble cytosolic proteins was tested by displaying a bait LZ that can bind to prey LZs in cytosol ( Figs. 1A and 1B ). LZ is a super-secondary structure that generates adhesion forces between α-helices. A single LZ consists of multiple leucine residues at approximately 7-residue intervals, which forms an amphipathic alpha helix with a hydrophobic region on one side. This hydrophobic region provides an area for dimerization, allowing the motifs to combine. Therefore, fusion proteins tagged with prey LZs may form a two-stranded α-helical coiled-coil heterodimer with the bait LZ in active IBs ( Fig. 1A ). A monomeric red fluorescent protein 1 (mRFP1) was used as a model prey protein to take advantage of its easy detection in living cells. The bait-EGFP-CBD and prey-mRFP1 genes were cloned into pET21a in a polycistronic manner to balance the relative expression of the bait and prey. When these bait and prey proteins were co-expressed in E. coli cells, the red fluorescence was clearly localized to the IBs (lower row in Fig. 2C ), whereas the red fluorescence remained dispersed in cells without the bait LZ (upper row in Fig. 2C ), showing that localization was dependent on the bait LZ.
Next, the effect of LZ binding affinity was investigated using different combinations of LZs (shown in Table 1 ). The leucine residue is essential for duplex formation, whereas ionic interactions between oppositely charged residues affected binding affinity. We examined five different bait and prey pairs that were designed by Magliery et al. [12] with K D values of 8, 20, 31, 50, and 1,000 µM. As anticipated, more red fluorescence was observed to localize to the IBs when bait-prey pairs with smaller K D values were used for the co-expression experiments ( Fig. 3 ). When region-of-interest (ROI) analysis was applied to the cellular images ( Fig. 4A ), red fluorescence in cytosol decreased as the prey-mRFP1 protein localized to the IBs. Consequently, the mean yield of localization to IBs, ROI 2 vs . ROI 1 , was calculated from at least five single cell images and a high yield of 0.65 was estimated for CN8 (K D = 8 µM), which was nearly the same as the mean yield for EGFP-CBD ( Fig. 4B ). The yield for CN1000 (K D = 1,000 µM) was approximately 0.30. Therefore, the higher the affinity of the bait for the prey, the more prey-mRFP localized to the IBs. In all the experiments, the expression levels of the bait-EGFP-CBD and prey-mRFPs were similar (as shown by SDS-PAGE analyses) ( Fig. S3 ).
10.1371/journal.pone.0097093.g003 Figure 3
Effects of binding affinity between LZs.
Microscopic images of E. coli cells containing LZ pairs with varying affinities (K D = 8, 20, 31, 50, and 1,000 µM). Scale bar = 5 µm.
10.1371/journal.pone.0097093.g004 Figure 4
Region of interest (ROI) analysis of microscopic images.
A. Cellular fluorescence decreased in proportion to the binding affinity between LZs in IBs. B . Comparison of localization yield to IBs. The fluorescence in IBs was normalized to the total cellular fluorescence, ROI 2 /ROI 1 , where ROI 1 is the cellular area and ROI 2 is the IB area of the cell. More than five cells per image were examined for the ROI analysis. Error bars show the standard deviations from 5 independent measurements of the cells.
The localization of red fluorescence to IBs was also investigated by flow cytometry. When the cytometric results were drawn on FL1 vs. FL2 dot plots, the mRFP intensity (FL2) decreased as the binding affinity increased ( Fig. 5A ), whereas the EGFP intensity (FL1) increased. For example, the mean intensity of mRFP for the CN8-CBD cell populations was about 40% of that for cells with no bait in the CBD IBs ( Fig. 5B ). This result was consistent with the microscopic observations in Fig. 4A .
10.1371/journal.pone.0097093.g005 Figure 5
Flow cytometric analyses of controlled localization to IBs.
A. FL1 vs. FL2 dot plot of cells containing LZ pairs with varying binding affinities (K D = 8, 20, 31, 50, and 1,000 µM). EGFP* indicates the E. coli cells that expressed soluble EGFP and soluble mRFP (the controls). The dashed line indicates the sorting gate. B. Comparison of the fluorescence intensity of cells expressing different leucine zipper pairs. C . Western blot analysis of CN20-CBD cells sorted by the FACSAria. Lane 1, CN20-CBD cells; lane 2, C20-CBD cells; lane 3, a mixture of CN20-CBD and C20-CBD cells before sorting; lane 4, a mixture of CN20-CBD and C20-CBD cells after sorting. Error bars show the standard deviations from 5 independent measurements of the cells.
Finally, we attempted to purify cells with IB-localized red fluorescence using a single cell sorter, the FACSAria. For this experiment, equal amounts of cells with (pCN20-CBD) and without bait (pC20-CBD) were mixed, and the specific cells within a predetermined gate (dashed areas in Fig. 5A ) were recovered. The collected cells were then analyzed by western blotting using an anti-GFP antibody ( Fig. 5C ). Lanes 1 and 2 show the control bands for bait-EGFP-CBD and EGFP-CBD, respectively. Before sorting, both proteins were observed in the cells (lane 3), whereas after sorting, the band corresponding to bait-EGFP-CBD was enriched in the recovered cells (lane 4), indicating selective sorting of cells with red fluorescent IBs due to protein-protein interactions between the bait and prey LZs.
High physical stability of fluorescent IBs
The functional IB particles were extracted from the CN20-CBD cells and the C20-CBD cells by sonication and washed twice with a solution containing 0.5% Triton X-100 detergent in a Tris buffer (50 mM Tris-HCl, pH 8.0 and 200 mM NaCl). Microscopic observation showed that the CN20-CBD IBs contained both green and red fluorescent IB particles, while the bait-less C20-CBD IBs contained only green fluorescent IBs because the prey-mRFP was washed out ( Fig. 6 ). Therefore, the interactions in the active IBs were highly specific and were maintained during sonication and washing.
10.1371/journal.pone.0097093.g006 Figure 6
Comparison of fluorescent IBs purified from CN20-CBD and C20-CBD cells.
Scale bar = 5 µm.
The physical stability of the active IBs was investigated using a zeta potential analyzer (Zetasizer Nano). The particle size was approximately 0.45–0.5 µm in diameter and the zeta potentials were estimated at approximately −56.8 mV. Zeta potentials larger than ±40 indicate that colloidal particles are stable in solution, while particles with a zeta potential smaller than ±30 tend to coagulate or flocculate easily [20] . Therefore, the IB particles in this study remained physically stable under both in vivo and in vitro conditions.
Discussion
Synthetic biology, an emerging field, involves the design and construction of new genetic devices for use in research and industry [15] , [21] . One successful device applied to metabolite production is the synthetic protein scaffold [16] , [17] . When a heterologous or synthetic pathway is introduced, the host cell can suffer from flux imbalance, intermediate loss, and chemical toxicity [22] . Therefore, constructing synthetic scaffolds may improve the metabolite conversion rate by increasing the local enzyme concentration and reducing intermediate loss caused by diffusion or side reactions. In this regard, CBD IBs could be useful as a synthetic matrix in E. coli cells. The target proteins can be recruited to the synthetic IB matrix via bait and prey interactions between LZs ( Fig. 1 ), which are a well-known domain consisting of only 30 amino acids. LZs such as E, K coil proteins have been used previously to immobilize active enzymes in polyhydroxybutyrate synthase IBs [14] . In this study, the affinities between LZs were controlled by changing the amino acid sequences. In addition, mRFP1 was used as a soluble target protein because it is easy to detect without cell disruption. Imaging and flow cytometric analyses showed that prey localization was dependent on the binding affinity between the bait and prey LZ proteins; the prey protein exhibited only marginal localization to the IBs when the K D of the LZs was 1,000 µM ( Fig. 3 ); as the K D decreased, localization increased sharply and it reached a maximum level when the K D was 8 or 20 µM. Eventually, we established a quantitative method to evaluate the localization of cytosolic proteins to IBs in situ by using LZs with different affinities ( Fig. 3 ), which provides useful implications for the generation of synthetic matrices with designed compositions.
Localization of EGFP to IBs resulted in a large decrease in the fluorescence signal compared to the signal for soluble EGFP ( Fig. S2 ), which is approximately half of the activity retention observed for catalytic enzymes in a previous study. The fluorescence intensity increased 2-fold when the particles were broken into smaller pieces by sonication ( Fig. 2B ). Based on literature reviews and our results in Fig. S4 , the reason for the decreased fluorescence in the IBs is thought to be related with the scattering of the excitation light by the highly refractile surfaces of the IB particles [23] , [24] and/or a shortened fluorescence lifetime in the densely packed environment [25] . In general, IBs are more dense (∼1.3 mg/mL) than any other cellular component.
Investigations of protein-protein interactions (PPIs) are crucial in modern biological science research [26] , and there is growing interest in the development of high throughput technologies [18] , [27] . In the method developed here, proteins with different affinity of LZs localized to IBs were quantitatively analyzed in living cells using flow cytometry ( Fig. 5 ), while the E, K coil proteins in IB fractions was detected by electrophoretic methods after cell disruption in previous study [14] . Therefore, the current method can be applied usefully for high throughput screening of PPI inhibitors, comparisons of interacting protein partners, and engineering binding affinities in bacterial cells.
Conclusions
Fluorescent proteins localized in IBs exhibited high intrinsic activity; however, their activity was somewhat suppressed when localized to IBs formed by fusion with the CBD from C. fimi exoglucanase. The signal intensity on microscopic images or in high throughput flow cytometry was dependent on the binding affinities of the interacting pairs. This controlled localization to IBs in living cells can be useful for the collective localization of cytosolic proteins in E. coli for sequential reactions. In addition, easy detection of protein localization to the IBs may provide a new platform for the rapid analyses of PPIs in bacterial cells.
Supporting Information
Figure S1
Construction of CN20-CBD (A) and C20-CBD (B).
(TIF)
Figure S2
Flow cytometric analyses of cells expressing EGFP and EGFP-CBD proteins. The dark green and light green signals indicate cells expressing EGFP and EGFP-CBD, respectively.
(TIF)
Figure S3
SDS-PAGE analysis of different leucine zipper proteins in E. coli cells (CN8-CBD, CN20-CBD, CN31-CBD, CN50-CBD, and CN1000-CBD). The upper and lower arrows indicate the size of the NZ-EGFP-CBD and CZ-mRFP proteins, respectively.
(TIF)
Figure S4
Side and forward scattering analyses of E. coli cells expressing EGFP (A) and EGFP-CBD (B).
(TIF)
Table S1
Primers used in this study.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Most cells under unaltered conditions of growth are able to maintain their size within a strict range, and a current view sustains that the cell cycle and cell growth machineries should be interconnected by specific molecular mechanisms ensuring cell size homeostasis [ 1 – 5 ]. Budding yeast cells control their size mainly at Start [ 6 , 7 ], when a G1 cyclin, Cln3, acts as the most upstream activator [ 8 ]. Cyclin Cln3 forms a complex with Cdc28, the cell cycle Cdk in budding yeast, which phosphorylates the transcriptional inhibitor Whi5 and induces a transcriptional wave in circa 200 genes to trigger cell cycle entry [ 9 ]. Cln3 modulates cell volume at Start in a precise, dose-dependent manner [ 10 – 12 ], which suggests that mechanisms regulating its levels or activity likely play important roles in cell size determination. In this regard, Cln3 is present at low and nearly constant amounts throughout G1 [ 8 , but see 13 , 14 ], and its nuclear levels are restrained by retention at the ER [ 15 , 16 ] and ubiquitin-mediated degradation by the proteasome [ 17 , 18 ].
It has long been known that cell size increases linearly with ploidy in fungi [ 19 – 21 ], plants [ 22 ], and animals [ 23 , 24 ], a function that is maintained across the enormous DNA content variation among eukaryotes [ 25 ] and has been used to infer ploidy in the fossil record [ 26 ]. Although ploidy has direct implications in cell growth and development, the underlying mechanisms that set cell size as a function of ploidy remain elusive [ 27 ]. Here, we describe a pathway linking the centromere (CEN) to the Start network in budding yeast. Briefly, we have found that an excess number of CENs increases degradation of Cln3 in the nucleus by a mechanism that involves physical and functional interactions between Cdc4, the specific F-box protein that targets Cln3 to SCF for ubiquitination, and Mad3, a centromeric signaling protein.
Results and discussion
In control experiments in which the size of yeast cells was carefully measured, we had previously observed that the presence of an empty yeast centromeric plasmid (YCp) produced a slightly larger volume at budding. Interestingly, this effect was exacerbated by increasing the number of empty centromeric vectors with different auxotrophic markers, suggesting that G1 length could be modulated by a genetic determinant present in these extrachromosomal DNA molecules. After ruling out possible effects due to plasmid-borne auxotrophic markers ( S1A Fig ), we analyzed newborn daughter cells during cell cycle entry in time-lapse experiments and found that, while initial volume was very similar, YCp caused a strong delay in G1 and a larger cell size at budding ( S1B and S1C Fig ). To assess the effects of YCp copy number at the single-cell level, we inserted a TEF1p -driven transcription unit expressing green fluorescent protein (GFP) in YCp vectors and mCherry in chromosome 5 and used different approaches to increase the number of centromeric sequences in the cell, some of them in a conditional manner ( Fig 1A ). We first analyzed cells in the simplest scenario, i.e., containing three GFP-expressing YCp vectors. Budding volume of control cells displayed a large variability [ 21 , 28 ] but steadily increased with the GFP/mCherry ratios ( Fig 1B , see S1 Data for a detailed statistical analysis). Intriguingly, the observed trend was compatible with the doubling in budding volume displayed by diploid cells. A yeast episomal plasmid (YEp), which is present at much higher copy numbers, did not significantly alter budding volume ( Fig 1C ), thus pointing to the autonomous-replicating sequence (ARS) or the CEN as the YCp-specific genetic determinants modulating cell size at budding. To discern between these possibilities, we used a yeast CEN placed immediately downstream from the inducible GAL1 promoter as a conditional CEN that, by growing cells under conditions that activate (galactose) or repress (glucose) transcription from the GAL1 promoter, can be switched off or on, respectively [ 29 ]. We introduced this conditional CEN ( Fig 1A ) into three different YCp vectors and observed that, under permissive conditions, cell volume at budding increased with a much steeper slope compared to unmodified YCp ( Fig 1D ). To rule out possible topological effects due to the circular conformation of YCp vectors, we used a linear yeast artificial chromosome (YAC) containing a conditional CEN to obtain a wide range of copy numbers per cell. As shown in Fig 1E , budding volume correlated with YAC copy number in a similar manner to that obtained with YCp vectors. Moreover, as this effect was also observed with a circular YAC derivative ( S2 Fig ), we were able to rule out possible additional effects of telomeric sequences. Finally, introducing conditional CENs into chromosomes 4 and 7 caused a significant increase in the budding volume of newborn daughter cells obtained by differential gradient centrifugation when allowed to enter the cell cycle under permissive conditions ( Fig 1F ). As previously described [ 30 , 31 ], high copies of centromeric vectors caused a short mitotic delay ( S3A Fig ) that depended on the spindle-assembly checkpoint (SAC) [ 32 , 33 ]. However, this delay was much shorter than that observed during cell-cycle entry ( S1C Fig ), suggesting that elevated CEN copies have a greater impact in G1. Accordingly, additional CEN sequences caused a small but significant increase in the proportion of cells in G1 phase in asynchronous cultures ( S3B Fig ). Overall, these data indicate that CEN number modulates G1 length in daughter cells and regulates their size at budding.
10.1371/journal.pbio.2005388.g001
Fig 1
CEN number effects on cell size.
( A ) Scheme showing the different approaches used to assess and manipulate CEN number. ( B ) Yeast cells endogenously expressing mCherry were transformed with one (small purple dots) or three (small blue dots) GFP-expressing YCp vectors, and cell size at budding was determined as a function of vector copy number (GFP/mCherry ratio). Individual budding volumes were binned, and mean values (large orange circles, N = 50) and a regression line are plotted. The mean budding size for wild-type diploid cells (which have 16 additional CENs compared to haploid cells) is also plotted (black diamond). ( C ) Cells carrying YEp (green circles) or YCp (orange circles) vectors were analyzed as in (B) to determine cell size at budding as a function of copy number. ( D ) Cells carrying YCp–CEN GALp (red circles) or YCp (orange circles) vectors were analyzed as in (B) to determine cell size at budding as a function of copy number under permissive conditions for the additional conditional CEN GALp CEN. ( E ) Cells carrying a YAC–CEN GALp artificial chromosome were grown at restrictive conditions for the conditional CEN GALp CEN to obtain a wide range of copies per cell, returned to permissive conditions, and analyzed as in (B) to determine cell size at budding as a function of copy number. Individual budding volumes (small blue dots) were binned, and mean values (large orange circles, N = 50) and a regression line are plotted. The mean budding size for wild-type diploid cells is also plotted (black diamond). ( F ) Newborn daughter cells carrying additional conditional CEN GALp CENs in chromosomes 4 and 7 were grown under permissive conditions until they entered the cell cycle. Individual budding volumes ( N = 100) and median values are plotted. Correlation analysis and pairwise comparisons were performed with nonparametric tests as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere; GFP, green fluorescent protein; YCp, yeast centromeric plasmid; YEp, yeast episomal plasmid.
Budding yeast cells mainly determine their size at Start [ 4 ]. Thus, we reasoned that signals originating from the CEN could target specific components of the Start network. YCp vectors clearly increased budding volume in cells deficient in Whi5 ( Fig 2A and 2B ), thus ruling out this transcriptional repressor of the G1/S regulon [ 34 , 35 ]. By contrast, cells lacking Cln3, the most upstream G1 cyclin [ 8 , 10 , 36 ] acting at Start, did not increase their size further, indicating that Cln3 is essential in the mechanisms that allow centromeric signals to modulate cell size. Overexpression of wild-type Cln3, which causes a strong nuclear accumulation of this G1 cyclin [ 15 , 37 ], also suppressed the YCp-mediated effects on budding size. However, a Cln3–1 hyperstable mutant that also reaches high levels but lacks the C-terminal nuclear-localization signal (NLS) that is essential for nuclear import of Cln3 [ 38 ] was as sensitive as wild type to the presence of YCp ( Fig 2A and 2B ). Supporting the notion that centromeric-dependent effects take place in the nucleus, a different hyperstable Cln3 ΔPEST mutant protein that retains the C-terminal NLS and strongly accumulates in the nucleus [ 15 , 37 ] fully suppressed YCp-mediated effects in cell size. Together, our data point to the idea that centromeric-dependent signals target, directly or indirectly, the yeast G1 cyclin in the nucleus.
10.1371/journal.pbio.2005388.g002
Fig 2
Exceeding CENs modulate cell size in a Cln3-dependent manner.
( A ) Cells with the indicated genotypes carrying three YCp vectors (3YCp) or none (ctrl) were analyzed to determine cell size at budding. Individual data ( N > 400), and median values are plotted. ( B ) Cells with the indicated genotypes carrying three YCp vectors were analyzed as in Fig 1B to determine cell size at budding as a function of copy number. Individual budding volumes (small dots) were binned, and mean values (large circles, N = 50) and a regression line are plotted. Correlation analysis and pairwise comparisons were performed with nonparametric tests as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
A high-throughput two-hybrid analysis in budding yeast [ 39 ] had revealed an interaction between the Cln3 cyclin and Mad3, a component of the kinetochore-signaling network involved in the SAC [ 40 , 41 ]. Thus, we tested whether centromeric signaling proteins could have a role in modulating budding size as a function of YCp copy number ( Fig 3A ). The budding size of cells lacking either Mad3 or Bub3 was absolutely refractory to increasing copies of YCp while, contrarily, kinase Bub1 did not have any effect. These results suggest that Mad3/Bub3 inhibit Cln3 function in a Bub1-independent manner, thus defining a mechanism different to that executing the SAC.
10.1371/journal.pbio.2005388.g003
Fig 3
Exceeding CENs require centromeric Mad3/Bub3 signaling proteins to modulate cell size.
( A ) Cells with the indicated genotypes carrying three YCp vectors were analyzed as in Fig 1B to determine cell size at budding as a function of copy number. Individual budding volumes (small dots) were binned, and mean values (large circles, N = 50) and a regression line are plotted. ( B ) Newborn daughter cells with the indicated genotypes were analyzed to determine cell size at budding. Mad3 overexpression ( oMAD3 ) was attained by inducing a GAL1p – MAD3 construct with 1 mM estradiol in newborn cells expressing the Gal4–hER–VP16 transactivator. Individual data ( N > 300) and median values are plotted. Correlation analysis and pairwise comparisons were performed with nonparametric tests as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
Newborn daughter haploid and diploid cells lacking Mad3 displayed a smaller volume at budding compared to wild type ( Fig 3B ). However, size reduction was only moderate compared to the difference between haploid and diploid wild type, and mad3 cells reduced their size normally from diploid to haploid status. These data suggest that either Mad3 is not required per se in the sensing mechanism or cells must have additional or backup mechanisms to adjust cell size to ploidy (see below). Although Mad3 could not be overexpressed to much higher levels compared to the endogenous copy ( S4A and S4B Fig ), budding size displayed a clear increase under these mild overexpression conditions ( Fig 3B ). Considered together, these data reinforce the notion of an inhibitory role for Mad3 in cell cycle entry and cell size determination at budding.
Mad proteins use different but complementary mechanisms to modulate degradation of Cdc20 targets by the anaphase-promoting complex (APC/Cdc20), including mitotic cyclins [ 32 , 33 ], which suggests that the Mad3-dependent effects of YCp vectors on budding volume could be mediated by degradation of Cln3. Supporting this idea, Skp1 is a highly expressed centromeric protein that is also present in SCF, the E3 ubiquitin ligase required to degrade Cln3 [ 18 ]. We found that the presence of YCp vectors strongly increased the degradation rate of Cln3 in promoter shut-off experiments, and more importantly, this effect required Mad3 ( Fig 4A, 4B , S5A and S5B Fig ). To support these findings further, we used a partially hyperstable and hypoactive mutant (Cln3–11A) fused to mCitrine that has no gross effects on cell cycle progression [ 42 ] but allows detection of this cyclin in G1 cells by fluorescence microscopy to monitor Cln3 degradation specifically in the nucleus. Although cells expressing mCitrine–Cln3–11A displayed an increased volume at budding when compared to wild-type cells, the presence of YCp vectors caused a similar relative increment in their budding size ( S6A Fig ), which validated its use. Notably, by measuring mCitrine–Cln3–11A levels in G1 cells after cycloheximide addition, we found that the presence of YCp vectors also increased the degradation rate of this G1 cyclin in the nucleus in a Mad3-dependent manner ( Fig 4C and S6B Fig ). Accordingly, mCitrine–Cln3–11A steady-state levels were strongly decreased by YCp in the nucleus of G1 cells within the same volume range ( Fig 4D ). Since Cln3 is rate-limiting for triggering Start and setting the critical size at budding, these results would explain why the presence of YCp vectors causes a larger cell size. Next we analyzed the interaction between Mad3 and Cln3 by affinity purification and found that they yielded relative coprecipitation efficiencies similar to Cln3 and Cdc4ΔFbox ( Fig 4E ), the adaptor protein that recruits Cln3 to SCF in the nucleus [ 18 ]. Interestingly, we were able to detect an interaction between Cdc4 and Mad3 ( Fig 4F ), which suggests that Mad3 is present with Cdc4 in SCF complexes. Mad3 contains a GLEBS domain that is known to interact with Bub3 and, as a likely consequence, with Skp1 [ 43 ], and we found that Mad3 lacking the GLEBS domain does not efficiently interact with either Cdc4 or Cln3 ( Fig 4E and 4F ). Finally, modulation of budding size as a function of YCp copy number was strongly dampened by deleterious SCF mutations or deletion of the Mad3 GLEBS domain ( Fig 4G and S7 Fig ), supporting the essential role of a SCF–Cdc4/Mad3 complex in boosting Cln3 degradation to modulate cell size at budding as a function of CEN copy number.
10.1371/journal.pbio.2005388.g004
Fig 4
Degradation of cyclin Cln3 by exceeding CENs: Mad3 physical and functional interactions with SCF.
( A ) Analysis of Cln3 stability by promoter shut-off experiments in the presence (orange circles) or absence (gray circles) of two YCp–CEN GALp vectors in wild-type cells grown under permissive conditions. After tetracycline addition, cells were collected at the indicated times, and obtained Cln3–6FLAG levels are plotted relative to an unspecific cross-reacting band (asterisk) used as loading control. ( B ) Analysis of Cln3 stability in Mad3-deficient cells as in (A). ( C ) Analysis of mCitrine–Cln3–11A stability by time-lapse microscopy in the presence (orange circles) or absence (gray circles) of three YCp vectors. Nuclear levels of mCitrine–Cln3–11A in cells were determined at the indicated times after cycloheximide addition, and mean values ( N = 100) are plotted. ( D ) Analysis of mCitrine–Cln3–11A accumulation in the nucleus in the presence (orange circles) or absence (gray circles) of three YCp vectors. Nuclear levels of mCitrine–Cln3–11A were determined in G1 daughter cells with 50–60 μm 3 of volume. Individual data ( N = 90) and median values are plotted. ( E ) Cell extracts (input) and GST PDs of cdc4ts grr1 cells expressing Cln3–13myc and GST fusions to Cdc4ΔFbox, Mad3, or Mad3ΔGLEBS were analyzed by immunoblotting with either αmyc (top panels) or αGST (bottom panel) antibodies. ( F ) Cell extracts (input) and GST PDs of cells expressing Mad3–3HA or Mad3 ΔGLEBS–3HA and either GST or GST–Cdc4 were analyzed by immunoblotting with either αHA (top panels) or αGST (bottom panel) antibodies. ( G ) Cells with the indicated genotypes carrying three YCp vectors were analyzed as in Fig 1B to determine cell size at budding as a function of copy number. Individual budding volumes (small dots) were binned, and mean values (large circles, N = 50) and a regression line are plotted. Correlation analysis and pairwise comparisons were performed with nonparametric tests as described in Materials and methods. Underlying data can be found in S1 Data . GST, glutathione S-transferase; PD, pulldown; YCp, yeast centromeric plasmid.
In summary, we have uncovered a pathway that links centromeric signaling proteins to G1 cyclin stability and, hence, cell size determination in budding yeast ( Fig 5 ). SCF–Cdc4 is estimated to be at low levels in the nucleus of yeast cells, and we envisage that Mad3, which is present at much higher levels, could act as a co-adaptor to increase the affinity of Cdc4 for Cln3. Strikingly, Cdc4 and Cdc20 display a high degree of similarity (34.2%) and contain WD40 segments that are used to interact with client proteins. However, the interaction of Mad3 would have different outcomes: (1) prevent Cdc20 from binding its targets in metaphase and (2) acting as an adaptor bridging Cln3 to Cdc4 in G1.
10.1371/journal.pbio.2005388.g005
Fig 5
CEN signaling proteins cooperate with SCF–Cdc4 to enhance degradation of the yeast G1 cyclin and modulate cell size in budding yeast.
When present in excess, centromeres accelerate Cln3 degradation in the nucleus with the essential participation of Mad3, a centromeric signaling protein that interacts with Cdc4 and requires SCF function to modulate cell size as a function of centromere number. CEN, centromere.
Mad3 is present at rather constant levels throughout the cell cycle [ 41 ], and Spc105, a scaffold protein involved in Mad3 activation at kinetochores by SAC [ 32 ], is already present in CENs in G1 [ 44 ]. Thus, by mechanisms different from those operating the SAC, Mad3 could be specifically activated in G1 at the kinetochore and sustain degradation of Cln3 at levels proportional to the number of CENs during G1 progression. Alternatively, we would like to speculate that the pathway uncovered here could belong to a Mad3-dependent checkpoint triggered by the excess of a kinetochore component that, being synthesized as a function of cell mass, would act as ploidy-mass reporter. Since Mad3-deficient or overexpressing cells do not display strong alterations in cell size, Mad3 would have a role as an effector of the checkpoint, not as sensor. Furthermore, the uncovered mechanism could be used to ensure that CENs congregate at the spindle-pole body (SPB) [ 45 ] before cell-cycle entry in budding yeast. While structural determinants of centromeric DNA are strikingly different in yeast and mammalian cells, kinetochore structural and signaling proteins are very well conserved. For this reason, we envisage that the mechanism operating in budding yeast could also exist across the evolutionary scale.
Previously proposed mechanisms to adjust cell size to ploidy [ 3 ] have not received sufficient experimental support. Although Whi5 is expressed at levels that depend on ploidy [ 42 ], diploid cells lacking one WHI5 copy are larger than haploid wild-type cells. On the other hand, introduction of additional Cln3-targeted promoters delays cell-cycle entry and increases cell size at budding [ 46 ]. However, it remains unclear whether titration of Cln3 by genome duplication is sufficient to produce a diploid cell size. We propose that, most likely with the contribution of these mechanisms, CEN-dependent degradation of Cln3 may play a pivotal role in scaling size with ploidy, a universal property of cells.
Materials and methods
Growth conditions and strain constructions
Cells were grown in SC medium with 2% glucose at 30 °C unless stated otherwise. Late G1-arrested cells were obtained by treating exponential cultures at OD 600 = 0.5 with 5 μg/ml α factor for 105 min at 30 °C. Conditional CEN GALp CENs were inhibited by addition of 2% galactose to culture medium. Cycloheximide was added at 25 μg/ml to inhibit protein synthesis. Cln3–3HA half-life was analyzed in tet -promoter shut-off experiments by adding tetracycline to 1 μg/ml [ 47 ]. MAD3 overexpression was attained by inducing a GAL1p-MAD3 construct with 1 mM estradiol in cells expressing the Gal4–hER–VP16 transactivator [ 48 ]. Yeast parental strains and methods used for chromosomal gene transplacement and PCR-based directed mutagenesis have been described [ 37 ]. Centromeric plasmids and yeast artificial chromosomes were obtained by multiple-fragment recombination [ 49 ] in yeast cells. Conditional CEN GALp CENs ( GAL10p–CEN4 ) were inserted in chromosomes 4 and 7 by CRISPR/Cas9-driven recombination [ 50 ]. The ΔGLEBS mutant of Mad3 lacked the C-terminal 155 amino acids. Cln3–1 [ 11 ] and Cln3 ΔPEST [ 15 ] are both hyperstable mutant proteins, but only Cln3 ΔPEST retains the C-terminal NLS [ 38 ]. The Cln3–11A mutant protein is a hypoactive and hyperstable cyclin that contains 11 amino acid substitutions (R108A, T420A, S449A, T455A, S462A, S464A, S468A, T478A, S514A, T517A, T520A) [ 42 ]. The Cdc4ΔFbox protein has been already described [ 18 ].
Time-lapse microscopy
Yeast cells were analyzed by time-lapse microscopy in 35-mm glass-bottom culture dishes (GWST-3522, WillCo) essentially as described [ 28 ] using a fully-motorized Leica AF7000 microscope. Time-lapse images were analyzed with the aid of BudJ, an ImageJ (Wayne Rasband, NIH) plugin that can be obtained from www.ibmb.csic.es\home\maldea to obtain cell dimensions and fluorescence levels in cellular and nuclear compartments [ 28 ]. Briefly, cell boundaries are detected as pixels markedly darker compared to both the surrounding background and the cell interior. Once outliers have been removed, an ellipse is fitted to the obtained boundary pixel array, and major and minor axes are used to calculate the cell volume assuming a prolate as shape. The same cell is followed through consecutive time-lapse images by using the center of the ellipse as seed point to obtain radial profiles in the following image.
Affinity purification and immunoblotting
GST-tagged proteins were affinity purified with glutathione beads (GE Healthcare) from cell extracts as described [ 37 ]. Immunoblot analysis [ 51 ] was performed with antibodies against HA (12CA5, Roche), FLAG (M2, Sigma), myc (9E10, Sigma), and GST (polyclonal, Millipore).
Miscellaneous
Small daughter cells were isolated from Ficoll gradients as described [ 52 ]. DNA content distributions were obtained by Fluorescence Activated Cell Sorting [ 51 ].
Statistical tests
Pairwise comparisons were performed with non-parametric tests. Specifically, median cell volumes at budding were compared with a Mann–Whitney U test. On the other hand, correlation of cell volume at budding with GFP/mCherry ratios was analyzed with a Spearman rank test. For pairwise analysis, data were subject to bootstrap resampling ( N = 100), and the resulting median slopes were compared by a Mann–Whitney U test. For both median and regression analysis, the resulting p values are shown in the corresponding figure panels.
Supporting information
S1 Fig
CEN number effects in G1 phase.
( A ) Newborn daughter wild-type cells with three YCp vectors (3YCp) or none (ctrl) or from a prototrophic URA3 LEU2 TRP1 derivative (3AUX) were analyzed to determine cell size at budding. Individual data ( N > 300) and median values (vertical lines) are plotted. ( B ) Newborn daughter cells with three YCp vectors (3YCp) or none (ctrl) were analyzed by time-lapse microscopy to determine initial and budding volumes. Individual data ( N > 90) and median values (vertical lines) are plotted. ( C ) G1 lengths corresponding to cells analyzed in panel B. Individual data ( N > 90) and median values (horizontal lines) are plotted. Pairwise comparisons were performed with non-parametric tests as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
(TIF)
S2 Fig
CEN number effects by a conditional-centromeric circular chromosome on cell size.
Cells carrying a YAC–CEN GALp artificial circular chromosome with no telomeric sequences were grown at restrictive conditions for the conditional CEN GALp CEN to obtain a wide range of copies per cell, returned to permissive conditions and analyzed as in Fig 1B to determine cell size at budding as a function of copy number. Individual budding volumes (small gray dots) were binned, and mean values (large orange circles, N = 50) and a regression line are plotted. The mean budding size for wild-type diploid cells is also plotted (black diamond). Nonparametric correlation analysis was performed as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere.
(TIF)
S3 Fig
CEN number effects in G2/M phases.
( A ) Wild-type or Mad3-deficient cells with three YCp vectors (3YCp) or none (ctrl) were arrested in late G1 with α factor and released into fresh medium to determine the percentage of binucleate cells at the indicated times. ( B ) DNA content distributions of wild-type cells carrying the indicated vectors or none (ctrl) under permissive conditions for CEN GALp CENs. Bars at the top correspond to the respective percentage of G1 cells in each sample. Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
(TIF)
S4 Fig
Overexpression of MAD3 under the GAL1 promoter.
( A ) Immunoblot analysis of GAL1p -driven Mad3–6FLAG levels at different times after GAL1p induction with 1 mM estradiol. Extracts from cells expressing Mad3–6FLAG at endogenous levels and untagged cells were also loaded as reference. A Coomassie Blue–stained major band is shown as loading control. ( B ) Quantification of Mad3–6FLAG levels shown in panel (A). Underlying data can be found in S1 Data .
(TIF)
S5 Fig
Degradation of cyclin Cln3 by exceeding CENs.
( A ) Analysis of Cln3 stability by promoter shut-off experiments in the presence (orange circles) or absence (gray circles) of two YCp–CEN GALp vectors in wild-type cells grown under permissive conditions. After tetracycline addition, cells were collected at the indicated times, and obtained Cln3–6FLAG levels are plotted relative to an unspecific cross-reacting band (asterisk) used as loading control. ( B ) Analysis of Cln3 stability in Mad3-deficient cells as in (A). Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
(TIF)
S6 Fig
YC effects on mCitrine–Cln3–11A and stability in Mad3-deficient cells.
( A ) Cells expressing mCitrine–Cln3–11A carrying three YCp vectors (3YCp) or none (ctrl) were analyzed to determine cell size at budding. Individual data ( N > 400) and median values (vertical lines) are plotted. Pairwise comparisons were performed with a nonparametric method as described in Materials and methods. ( B ) Analysis of mCitrine–Cln3–11A stability in Mad3-deficient cells. Nuclear levels of mCitrine–Cln3–11A were determined by time-lapse microscopy in mad3 cells and in the presence (orange circles) or absence (gray circles) of three YCp vectors after cycloheximide addition as in Fig 4C . Mean values obtained from individual cells ( N = 100) are plotted. Underlying data can be found in S1 Data . YCp, yeast centromeric plasmid.
(TIF)
S7 Fig
Cell size effects by exceeding CENs in SCF-deficient cells.
Cells with the indicated genotypes carrying three YCp vectors were analyzed as in Fig 1B at the restrictive temperature for cdc53ts and cdc34ts alleles to determine cell size at budding as a function of copy number. Individual budding volumes (small dots) were binned, and mean values (large circles, N = 50) and a regression line are plotted. Correlation pairwise comparisons were performed with a nonparametric test as described in Materials and methods. Underlying data can be found in S1 Data . CEN, centromere; YCp, yeast centromeric plasmid.
(TIF)
S1 Data
Source data for all plots in manuscript.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Multiple Sclerosis (MS) is an inflammatory demyelinating disease of the brain and spinal cord with a presumed autoimmune etiology. The pathology of MS features focal areas of inflammatory infiltration and demyelination in which oligodendrocytes (OLs) – the myelinating cells of the central nervous system (CNS) – are depleted [1] . During the early stages of MS, some repair of this damage is still possible, likely due to partial remyelination [2] initiated by surviving or newly formed OLs (generated from progenitor cells) [3] . In the primary and secondary progressive forms of MS, a high degree of cortical demyelination has been observed [4] , [5] , but active cortical lesions show only mild lymphocytic infiltrates [6] . All approved treatments for MS are immunoregulatory, which are able to reduce the inflammatory component of the disease. Unfortunately, these treatments are not effective against progressive clinical disability. However, new treatments that target OLs and myelin sheaths have come under serious consideration [7] , [8] . Approaches that directly protect myelin-producing OLs and enhance remyelination may improve long-term outcomes and reduce the rate of axonal damage in MS patients.
Galanin (GAL) is a 29-amino-acid neuroendocrine peptide that is widely distributed throughout the rat, mouse and human CNS, where it functions alongside more “classical” neurotransmitters [9] , [10] . It has been demonstrated that GAL acts as a survival and growth-promoting factor for different types of neurons in the peripheral nervous system (PNS) and the CNS [11] . GAL has also been implicated in the control of neurogenesis (i.e., the proliferation, differentiation and/or migration of neural stem cells) in both normal and injured brains [12] , [13] , [14] , [15] . High levels of GAL and GAL receptors (GalR1 and GalR2) are expressed in the subventricular zone (SVZ), the rostral migratory stream (RMS), the subgranular zone of the dentate gyrus (SGZ) and in oligodendrocyte precursor cells (OPCs) of the corpus callosum [12] , [13] , [16] , [17] , [18] . Recently, it has been demonstrated that GAL is markedly upregulated in MS lesions including shadow plaques in post-mortem brain tissue from chronic MS sufferers exclusively in microglia [19] , although not all microglia were galanin positive. In the same study, GAL was also uperegulated in the CNS of mice with acute inflammation in the EAE model, although here it was exclusively in oligodendrocytes.
One of the best characterized demyelinating mouse models is that of C57BL/6J mice with cuprizone (CPZ), a copper chelator, added to their diet [20] . CPZ, when given in small doses, acts as a neurotoxin that specifically induces the death of OLs and causes inflammation in the resulting demyelinated areas; these effects are reversible, and the removal of CPZ from the mouse feed permits remyelination. The lesions that occur in the brains of CPZ-treated mice are similar to type III and type IV lesions observed in MS [21] . In this mouse model, focal inflammation and demyelination occurs in the brain, however, T lymphocytes do not appear to play a role in these processes [20] .
We have previously reported the creation of a GAL-Tg mouse in which GAL is over-expressed in the anterior pituitary gland causing circulating GAL levels to be chronically elevated [22] , [23] , and to a much lesser extent prolactin (PRL) and growth hormone (GH). In the present study, using the CPZ-mediated demyelination model of MS, we demonstrate that GAL has pronounced neuroprotective effects with respect to demyelination and remyelination and that it directly prevents OL death. Moreover, we also found changes in the expression of the GAL receptor GalR1 during the demyelination and remyelination processes. These results suggest that GAL and/or GAL receptors could represent next-generation therapeutic targets for the treatment of MS.
Results
Effects of the over-expression of GAL on body weight
Initially (at 8–9 weeks of age) Gal-Tg mice were heavier than WT mice, with average body weights of 29.9 ± 4.3 g and 24.9 ± 2.4 g, respectively (mean ± SD).By six weeks later, Gal-Tg and WT mice had gained 27.6±18.3% and 21.3±13.5%, respectively, of their body weight compared to their initial weight( Fig. 1A ). In comparison, after treatment with CPZ diets, both Gal-Tg and WT mice lost a significant amount of body weight in the first week. Tg mice lost approximately 5% of their weight after one week, while WT mice lost approximately 10% of their weight after two weeks. However, Tg mice started to gain weight after one week on the diet, while WT mice continued losing weight until the third week ( Fig. 1B ). Following withdrawal of the CPZ challenge, mice of both genotypes gained weight immediately (but at different rates) ( Fig. 1B ). There was a statistically significant difference in the rate of body-weight increase between Gal-Tg and WT mice. Two-way mixed-measures ANOVA revealed a significant interaction (F(6,138) = 9.83, p<0.0001) between normal body-weight growth rate and genotype, indicating that the rate of body-weight growth was dependent on genotype. Another two-way mixed-measures ANOVA analysis revealed that, during CPZ challenge, there was a significant influence of genotype on changes in body weight.
10.1371/journal.pone.0033901.g001
Figure 1
The effect of GAL over-expression on body weight.
Body weight was measured at the age of 8–9 weeks old as the starting point 0 w. The raw body weight growth in normal control groups (A) and the CPZ-treated groups (B) is shown. Both the cuprozone groups and the cuprizone-challenge plus 3-week recovery groups were measured at the same time. Data are expressed as the mean ± SEM. (n = 12–17 per group). * p<0.05, ** p<0.01, *** p<0.001.
CPZ-induced demyelination in MBP staining is attenuated in GAL-Tg mice
Consistent with previous studies [20] , [24] , our findings revealed that CPZ challenge caused significant loss of myelin in WT mice, particularly in the CC and in the junction area between the CC and the cerebral cortex ( Fig. 2B , Fig. 2b ). However, this myelin breakdown was significantly reduced in Gal-Tg mice ( Fig. 2E , Fig. 2d ). In Gal-Tg mice, myelin degradation in the external capsule of the CC (as demonstrated by the doubled arrow in the whole brain pictures of Fig. 2b , Fig. 2d ) were obviously reduced compared with WT mice. In WT mice, after six weeks of CPZ-induced challenge, only a few myelin fibers remained, which were restricted to layer four of the cortex. On the other hand, both WT and Gal-Tg mice from the recovery groups showed dense MBP staining ( Fig. 2C, 2F ). A two-way ANOVA analysis revealed that genotype significantly affected MPB staining in the control, demyelination and remyelination groups (F(2,16) = 9.56, p = 0.0019). Post-hoc comparison tests found that Gal-Tg mice had a significantly greater level of MBP staining (based on the optical density values) after six weeks of the CPZ challenge. One-factor ANOVA tests on the WT mouse data revealed that the CPZ-induced challenge significantly affected the myelin staining results (F(2,6) = 28.86, p<0.001). We also used Luxol Fast Blue (LFB) staining on normal and CPZ-challenged animals to verify the IHC results, which showed that the integrity of the myelin fibers was similar to that observed in the IHC staining ( Fig. 2 , images I–IV). Western blotting analysis of MBP protein level changes were consistent with the changes in MBP staining (data not shown).
10.1371/journal.pone.0033901.g002
Figure 2
Over-expression of GAL blocks CPZ-induced myelin breakdown.
(A)–(F) are photographs of the cerebral cortex and the CC, while (a)–(d) are the photographs of whole brain. Both Tg and WT mice were given 0.3% CPZ for six weeks (6wCPZ), and then the two groups of mice were allowed to recover for three weeks on a normal food diet (6wCPZ+3wR). Mice brains were processed for IHC staining using an antibody against MBP. Consistent with previous studies, we found extreme demyelination in WT mice after six weeks of CPZ challenge: compare (B) and (b) with the control groups (A) and (a). However, MBP staining of Tg mice brains, (E) and (d), indicated that myelin breakdown was not significantly different compared with controls, (D) and (c). After three weeks on a normal diet, the WT mice recovered well (as expected) (C). To verify the IHC staining results, we also used luxol fast blue staining on WT CLT (I), WT 6wCPZ (II), Tg CLT (III) and Tg 6wCPZ (IV) samples (6 µm paraffin sections). The bar graphs represent the measurements of optical density of MBP IHC staining in the cerebral cortex and the CC areas. Arrows show the aca area that was used for color-intensity standardization. Data are expressed as mean ± s.e.m values. (n = 3–5 per group). ** p<0.01, *** p<0.001.
Increased circulating GAL alleviates demyelination-related pathogenesis
The loss of mature OLs was attenuated in GAL-Tg mice. We first examined the mature OLs that survived the CPZ-induced challenge. IHC staining demonstrated that, in control animals, there were widely distributed GST-π positive cells in the cortex (data not shown) and linearly distributed cells in the CC ( Fig. 3A, 3D ). As expected, after six weeks of CPZ-induced challenge in WT mice, only a few weakly positive cells remained in the CC, and a small number of positive cells remained in the cortical areas (restricted to cortex layer four). However, in Gal-Tg mice, there were still many positive cells after six weeks of CPZ-induced challenge ( Fig. 3E, 3H ). Furthermore, many positive cells could be observed in WT mice from the recovery group, while no obvious differences were observed in Gal-Tg mice between the demyelination and recovery groups ( Fig. 3E, 3F ). A two-way ANOVA test on the cell numbers revealed that genotype significantly influenced the number of GST-π-positive cells under the control, demyelination and recovery conditions (F(2,12) = 4.40, p = 0.0370). Post-hoc comparison tests found a significant difference between WT and Tg mice from the 6wCPZ group (p<0.01). A one-way ANOVA test on the WT groups showed that the mean number of positive cells was significantly different across the groups, and Tukey's test revealed significant differences between the WT CLT versus the WT6wCPZ groups (p<0.001) and the WT 6wCPZ versus the WT 6wCPZ+3wR groups (p<0.001). However, a one-way ANOVA test revealed no significant difference among the Tg groups, although the results showed a trend indicating impairment after the CPZ-induced challenge.
10.1371/journal.pone.0033901.g003
Figure 3
Increased levels of galanin attenuated CPZ-induced oligodendrocyte loss.
Mature oligodendrocytes were detected with IHC using a GST-π antibody. Photographs (A–F) were taken from the knee region of the CC, and (G) and (H) are examples of the full-size pictures taken of WT and Tg brains from the 6wCPZ group. The three CC images in the upper panels (A–C) show the WT mice from the CTL, 6wCPZ and 6wCPZ+3wR groups, and the three images in the middle panels (D–F) show the Tg mice from the same groups. In (A–F), high-magnification micrographs were also taken of the CC area using an oil-immersion lens, as shown in the inserts. (I) A bar chart displaying the numbers of GST-π positive cells, which were counted manually. **p<0.01, ***p<0.001, n = 3 in each group. The longer scale bar represents a length of 200 µm and the shorte scale bar represents 500 µm.
Proliferation and accumulation of OPCs were low in the Gal-Tg mice after CPZ-induced challenge
The distribution of OPCs was revealed by IHC for PDGFR-α, one of the two commonly used cell markers for OPC identity. Under normal conditions in the control groups, there were few PDGFR-α-positive cells in the brains of WT and Tg mice ( Fig. 4A, 4D ). After CPZ-challenge, the number of PDGFR-α-positive cells began to increase in the CC of WT mice ( Fig. 4B ). Furthermore, this positive staining was significantly higher in the brains of WT mice from the recovery group ( Fig. 4C ) (p<0.05). In contrast with WT mice, the numbers of PDGFR-α-positive cells in Tg mice were similar among the different groups ( Fig. 4D–F ).
10.1371/journal.pone.0033901.g004
Figure 4
Increased number of PDGFR-α positive cells in WT mice that underwent the CPZ-induced demyelination challenge.
WT and Tg mice were given CPZ challenge for six weeks (6wCPZ, B and E) while the control mice (CLT, A and D) received normal CPZ-free rodent chow. After six weeks of challenge, two groups of animals (6wCPZ+3wR, C and F) were allowed to recover for three weeks on a normal CPZ-free diet. The bar graph shows the results of the optical-density measurements of the PDGFR-α-positive cells; a significant difference was observed between the 6wCPZ+3wR WT and Tg groups. The scale bar represents 200 µm in the low-magnification. Data are expressed as mean ± SEM values (n = 3 per group). *p<0.05, ** p<0.01.
In GAL-Tg mice the invasion of CPZ-induced reactive astrogliosis was restricted
Without any experimental challenge, most of the GFAP positive cells were limited to the CC area in both WT and Tg mice brains As expected, after the CPZ-induced challenge, increased numbers of GFAP-positive astrocytes could be detected in all regions of the brain in both WT and Tg mice. More precisely, the distribution of the reactive astrocytes in WT mice was diffused through all layers of the cerebral cortex, but in Tg mice, the invasion of reactive astrocytes was limited within layer four of the cerebral cortex, i.e., no labeling was seen in cerebral cortex layers five and six. In the recovery groups, the reactive astrogliosis (in terms of the density of staining) in WT mice was reduced, although the distribution was still dispersed in the cortex, but the diffusion in Tg mice were still restricted in the limited regions.
The GalR1 receptor is more highly expressed during the demyelination and remyelination (recovery) phases
The expression levels of GAL receptors were normalized to the levels observed in WT mice from the control group. The expression levels of GalR1 were 2.7- and 6.4-fold higher in WT and Tg mice, respectively, in the demyelination groups. The expression levels of GalR1 were increased even further (2.9- and 5.8-fold higher in WT and Tg mice, respectively) in the recovery groups ( Fig. 5A ). In contrast with GalR1 expression, the expression of GalR2 was increased 1.4- and 2.1-fold in WT and Tg mice, respectively, in the demyelination groups; expression of GalR2 increased 2.2- and 3.3-fold in WT and Tg mice, respectively, in the recovery groups ( Fig. 5B ).
10.1371/journal.pone.0033901.g005
Figure 5
The differential expression of GalR1 and GalR2 in the CC area.
RNA samples were extracted from CC areas as demonstrated in Figure 6 . The gene expression levels among the groups were normalized to the WT CLT levels (set equal to 1). (A) The expression of GalR1 among the groups. (B) The expression of GalR2 among the groups. Data are expressed as the mean ± SEM values. (n = 3–6 per group). * p<0.05, ** p<0.01.
Discussion
GAL has diverse biological functions in the nervous system, and the potential for its involvement in myelin development had been demonstrated by Shen et al. [15] . The present study reveals a potential direct effect for GAL in myelin protection against cuprizone induced demyelination.
Unlike other GAL transgenic models, the circulating levels of GAL in our Gal-Tg mice were 10-fold higher than those observed in WT mice [22] . This dramatic increase in circulating GAL provides a unique tool to study its neuromodulatory role. The Blood-Brain Barrier (BBB) was once assumed to be impermeable to peptides. However, it is now widely accepted that peptides are capable of crossing the BBB by both non-saturable and saturable transport mechanism to the extent that they can affect events in the CNS. There exist several potential means whereby circulating GAL could enter the brain. First, the circumventricular organs could provide a broad passageway [25] . Additionally, active transport mechanisms have been described elsewhere, which also protect the substances from rapid degradation [26] . Furthermore, GAL has been shown elsewhere to cross the BBB, as intravenous administration of GAL has quick acting anti-depressant activity and affects sleep EEG recordings [27] .
The present study indicates that GAL-receptor signaling plays a key role in modulating CPZ-induced demyelination. Using our GAL over-expressing transgenic mouse model, we identified a novel attenuation of OLs against CPZ-induced demyelination, and we found that this influence was exerted independently of progenitor cells. Alleviation of myelin breakdown in the GAL-Tg mice was observed to be significant.
Myelin basic protein (MBP) is a structural protein of myelin sheath. Furthermore, it is believed to be a vitally important functional protein for myelin assembly and maintenance [28] . MBP IHC staining, which was used in the present study, is a common method for quantitatively measuring myelin sheath levels [29] . Thus, after a six-week exposure to CPZ challenge, the significant loss of MBP staining in WT mice is indicative of severe demyelination, and the attenuation of MBP-staining loss in Tg mice suggests an alleviation of demyelination. The DAB-dye-based IHC staining results were also verified using Luxol Fast Blue staining, another commonly used histological myelin staining method. The loss of MBP protein in WT mice was also demonstrated by western blotting with the same antibody used for the IHC staining (data not shown).
Because the experimental demyelination induced by CPZ is caused by the apoptosis of OLs [20] , we next examined the number of mature OLs in the different groups. GST-π is a myelin- and OL-associated enzyme in the brain that is used as a biomarker for mature OLs [30] , [31] . For the first time, we show that (with respect to CPZ-induced demyelination) the loss of GST-π positive cells observed in WT mice was greatly alleviated in our GAL- Tg mice. Interestingly, the distribution of protected OLs in the neocortex of the GAL-Tg mice corresponded to the GAL-stained cells in the same area, indicating a direct protective effect of GAL on OL survival. In this study, we also found that the oligodendrocyte precursor cells (OPCs) were highly proliferative and accumulated in response to CPZ-induced oligodendrocyte degradation in the WT mice, in contrast to a minimal response in the Tg mice. It has been suggested that the infiltrating proliferative OPCs are recruited to reverse the loss of mature OLs by proliferation and differentiation and that OL death (in the lesions) is characterized by a repopulation of OPCs prior to new OL formation [32] . These findings support the hypothesis that the lower levels of OPC proliferation and accumulation seen in Tg mice are caused by an improvement in OL survival.
These results provide important insights into the mechanisms by which GAL-receptor signaling influences the phenotypes of central nervous system (CNS) demyelination. Our data strongly indicate that GAL has the capacity to influence the outcome of primary insults that directly target OLs, as opposed to cases where immune activation is the primary pathogenic event. This is an important distinction, as EAE (the principal animal model of central demyelination) is driven by immune cell (principally T cell) activation. Evidence has already been provided to suggest that the influence of GAL in EAE is associated with OLs [19] , which is of particular importance given the fact that OL death can be an early event in MS (even preceding immune cell infiltration) [33] . Thus, our study provides evidence for the potential use of GAL as a therapeutic target of oligodendro-cytopathy in MS and compliments the study of Dr. Wynick's group.
One limitation of the study is that in our Gal-Tg mouse besides chronic increase of serum GAL there is an increase, to a much lesser extent, of PRL and GH that they might act as confounders to the observed phenotype. While though both hormones have been described to have an effect on myelination [34] , [35] , [36] , [37] , this effect is mainly through the proliferation of OL progenitors (OPS). In our study the most pronounced effect in the Gal-Tg is not the proliferation of OPs, since PDGFR-α-positive cells did not change, but the attenuation of OLs death. We believe that this effect is mainly due to increased GAL levels, although further studies are needed to confirm that. Furthermore, the increased lipid levels of the obese phenotype of our transgenic mice might have also contributed to an accelerated myelination. Studies in progress are defining this issue as well.
In the context of the CPZ model (where the BBB is thought to be intact), the apparent efficacy of increased levels of circulating GAL [38] provides an important insight into the pharmacodynamics of this protein. Our data suggest that (in the context of MS) GAL might be able to protect OLs in regions anatomically distant from the discrete active plaques where the BBB is actually disrupted; this could help eliminate progressive neurodegeneration, independent of the degree of acute inflammatory activity.
It is well accepted that GAL exerts its biological effects through the three G-protein coupled receptors, GalR1, GalR2 and GalR3. These three receptors activate distinct G-proteins that participate in different GalR-signaling pathways, which results in diverse biological functions being activated upon the binding of GAL or other ligands [39] . In this study, a gene expression assay revealed that the expression of GalR1 and GalR2 was differentially altered in the demyelination and remyelination phases, suggesting these two receptors have distinct roles during these two phases. The change in GalR1 expression during the initial challenge phase may be indicative of a protective effect in response to demyelination, while the trend of GalR2 inrease in the later recovery phase, although not significant compared to wild type, may imply the existence of a neurotrophic effect during recovery. Interestingly, the differential involvement of GalR1 and GalR2 in response to a pathological challenge has also be observed in seizure development – the involvement of GalR1 was found to be important during the initiation phase, while GalR2 was found to be more important during seizure maintenance [40] . However, the gene expression assay in this study does not provide direct evidence concerning the mechanisms of GalR1 and GalR2 action in the demyelination and remyelination processes. Further studies are needed to investigate the roles of these two receptors and their downstream effects.
In summary, we have provided evidence that elevated circulating GAL can greatly alleviate CPZ-induced demyelination. Furthermore, we show that the attenuation of demyelination-related impairments (from the effect of elevated GAL) was directly related to a reduction in the death of mature OLs. Gene expression assays also revealed differential responses of GalR1 and GalR2 to demyelination. These findings strongly support the hypothesis that elevated circulating GAL can alleviate CPZ-induced demyelination via GalR1 receptor. Demyelination is the final pathological downstream event in MS, and it is at the core of the diverse symptoms and disabilities caused by this disease. Over the decades, many contributions have been made to the immunoregulatory treatment of MS. However, new strategies for myelin protection are considered to be the best hope for next-generation therapies. Overall, our findings strongly suggest potential therapeutic effects for GAL, with respect to myelin protection, in the treatment MS.
Materials and Methods
Experimental animals
All mice, including wild-type (WT; C57BL/6) and homozygous transgenic mice (GAL-Tg) maintained on a C57BL/6, were housed in the University of Manitoba animal facility in a temperature controlled environment (20°C under a 12 h light/dark cycle). Food and drinking water were available ad libitum. Twenty five male WT mice were obtained from Charles River (Montreal, QC, Canada); GAL-Tg mice were generated as previously reported [22] , [23] using a 320 bp fragment of the rat growth hormone (GH) promoter fused to the full-length rat preprogalanin cDNA clone. Before the initiation of the experiment, homozygous Gal-T mice were backcrossed to wild type WT; C57BL/6 from Charles River and brought again to homozygosity. Thirty six GAL-Tg male mice were used in this study, to avoid the effects of fluctuating steroid hormones in female mice on the expression of GAL [23] , [41] . 3–6 mice were used per group and the whole experiment was repeated twice. All procedures were in accordance with the Animal Protocol Review Board of the University of Manitoba which has approved this study under the protocol #10-013/1 (principal investigator Dr.Vrontakis). Cuprizone-induced demyelination
To induce demyelination, both WT and GAL-Tg mice at 8–9 weeks of age were fed with a diet containing 0.3% (w/w) CPZ (C9012-25G, Sigma-Aldrich) for six weeks. No lethality was observed due to treatment. CPZ powder was well blended into milled LabDiet rodent chow (Prolab® RMH 3000 5P00, 22.0% crude protein, 5.0% crude fat, 5.0% crude fiber, 6.0% ash, and 2.5% added minerals) using a food processor (43-1976-0, KitchenAid), then the blended diet powder was re-pelleted by extrusion through a 30 ml plastic tube. The CPZ diet was freshly prepared and given to the mice twice a week while monitoring for body weight. To allow for remyelination, animals that had received the CPZ diet were put back on a normal diet for an additional three wks. The experimental mice were divided into the following three groups: control (CLT), demyelination (6wCPZ) and remyelination (6wCPZ+3wR). Body weight and general behaviors (grooming, activity, etc.) were monitored twice a week.
Immunohistochemistry
For the Immunohistochemistry (IHC) analysis, half of animals in each group were anesthetized by isoflurane inhalation, perfused intracardially with 0.1 M Phosphate buffered saline (PBS; pH 7.4) containing 50 U/ml heparin sodium, followed by 4% paraformaldehyde prepared in 0.1 M PBS. The brains were then dissected out and post-fixed in the same fixation solution at 4°C overnight, followed by cryoprotective treatment in 0.1 M PBS containing 30% sucrose at 4°C (until the brains no longer floated). The dehydrated brains were snap-frozen on dry ice and stored at −80°C until sectioning. Serial coronal sections (25 µm thick) of the frozen brains were made using the Leica SM2400 sliding microtome equipped with a freezing plate. The free-floating sections were then kept in 24-well plates containing 0.1 M PBS at 4°C. Sections corresponding to levels 185–195 of the High Resolution Mouse Brain Atlas (Sidman et al.; http://www.hms.harvard.edu/research/brain/atlas.html ) were used in this study. Staining was performed using the avidin-biotin-peroxidase complex technique. Procedures were carried out on a slow shaker at room temperature, unless otherwise indicated. Briefly, free-floating sections were washed in 50 mM Tris Buffered Saline (TBS, pH 7.4) three times (5 min each) followed by pre-treatment with 3% hydrogen peroxide for 30 min. Sections were then washed in TBS and incubated for 1 h in a blocking solution consisting of 5% normal goat serum and 0.3% Triton-X100 in TBS. Subsequently, the sections were incubated in the same blocking solution overnight at 4°C with primary antibodies ( Table 1 ). After washing in TBS containing 0.1% Tween-20 (TBST), sections were incubated with biotinylated secondary antibodies (all secondary antibodies were purchased from Vector Laboratories, Ontario, CA) for 1 h, followed by three washes in TBST. Sections were further treated with the peroxidase-coupled avidin-biotin complex (ABC Peroxidase Staining Kit, Cat # 32020, Pierce) for 1 h. Finally, immune-precipitated sections were visualized using the DAB Peroxidase Substrate Kit (3,3′-diaminobenzidine; Cat # sk-4100, Vector Lab). After mounting on the clean pre-coated slides (Superfrost Plus Microscope Slides, Cat # 12-550-15, Fisher), sections were then air-dried on a 37°C slide warmer overnight. The following day, sections were post-treated with a series of increasing concentration alcohols from 70% to 100%, followed by two treatments with xylene. Control slides were processed in parallel with experimental slides using the same agents, except without primary antibodies. No positive staining was observed in the control slides.
10.1371/journal.pone.0033901.t001 Table 1
Primary antibodies used for immunohistochemistrical staining.
Target
Dilution ratio
Catalog #
Supplier
Anti-MBP (C-16)
Myelin
1∶1000
Sc-13914
Santa Cruz
Anti-GST-π
Mature OL
1∶1000
610719
BD Biosciences
Anti-GFAP
Astrocyte
1∶1000
MAB360
Millipore
Anti-CD140a
PDGFR-α/OPC
1∶500
558774
BD Biosciences
Abbreviations: MBP – myelin basic protein; GST-π – Glutathione S-Transferase π form; GFAP – glial fibrillary acidic protein; PDGFR-α – alpha-type platelet-derived growth factor receptor; OPC – oligodendrocyte precursor cell; OL – oligodendrocyte.
RNA preparation
All instruments used in the brain dissections were RNase-decontaminated by wrapping with RNaseZap® (Cat # 9780, Ambion). Mice were anesthetized by isoflurane inhalation and were quickly decapitated. The fresh brains were quickly removed and placed into a stainless-steel mouse-brain holder (Adult Mouse Brain Slicer Matrix, Cat # BSMAS005-1, ZIVIC Instruments). On the holder, a 2.5 mm thick coronal slab was made. The first coronal cut was 5 mm away from the edge of olfactory bulb, and the second cut was 2.5 mm (5 section slice intervals of the brain holder) away from the first cut. The selected brain structures (the corpus callosum and part of the cortex) were then dissected from the brain slab, using a stereomicroscope on ice ( Fig. 6 ). Isolated tissues were quickly transferred into 2.0 ml RNase-free microtubes containing 0.6 ml of ice-cold TRIzol (Cat # 15596026, Invitrogen). Total RNA was extracted from each homogenized tissue sample (processed using Homogenizer Power Gen 125, Generator Flat 5×95, Fisher Scientific) using the phenol/chloroform extraction protocol provided by Invitrogen (TRIzol Reagent). Subsequently, the total RNA was purified using the RNeasy Mini Kit (Cat # 74004, QIAGEN) according to the manufacturer's protocol; DNA-decontamination treatment was included. The purified RNA samples were dissolved in RNase-free water and were stored at -80°C until processed.
10.1371/journal.pone.0033901.g006
Figure 6
RNA sample sources.
The illustration on the left represents the mouse brain. The top dashed line represents the first coronal cut, 5 mm away from the edge of olfactory bulb. The bottom dashed line represents the second coronal cut, 2.5 mm away from the first cut. The illustration on the right represents the brain section isolated on the left. The tissue within the dashed rectangle, containing mainly corpus callosum and part of the cortex, was used for RNA extraction. Abbreviations: CC – corpus callosum; LV – lateral ventricle; aca – anterior commissure.
Real-time reverse transcription polymerase chain reaction
Total RNA was quantified using a spectrophotometer (NanoDrop, Cat # ND-1000, Thermo, Fisher Scientific), and 1 µg of total RNA was used to generate single-stranded cDNA. The concentrated total RNA samples were diluted to 0.1 µg/µl in at 10 µl volume. The diluted RNA samples were denatured at 70°C for 10 min and then added to the reverse transcription (RT) master-mix ( Table 2 ). For RT, the reaction conditions were as follows: 25°C for 10 min, 42°C for 50 min, and 72°C for 15 min with a 0°C hold at the end.
10.1371/journal.pone.0033901.t002 Table 2
Primer pairs provided by SABiosciences™.
Symbol
Description
Band size
Ref pos
Exon
Cat#.
Gal
Galanin
139
607
6
PPM25148E
GalR1
Galanin receptor 1
151
1368
3
PPM04847A
GalR2
Galanin receptor 2
155
1261
2
PPM05170A
GAPDH
Glyceraldehyde-3-phosphate dehydrogenase
140
309
3
PPM02946E
The previously generated cDNA templates were diluted with 50 µl RNase-free water. For qPCR, the primer pairs were designed and generated by SABiosciences™, QIAGEN; the housekeeping gene GAPDH (Glyceraldehyde-3-phosphate dehydrogenase) was used as the internal quantitative control ( Table 2 ). Each PCR consisted of 12.5 µl RT 2 SYBR Green/ROX qPCR MasterMix (Cat # PA-012-24, SABiosciences, QIAGEN), 6.5 µl RNase-free water and 1 µl Primer Mix. PCRs for each gene of interest were run in triplicate using the StepOne™ Real-Time PCR System (Applied Biosystems™). The PCR cycling program was as follows: 95°C for 10 min and 40 cycles of 95°C for 15 s and 60°C for 1 min. The melt curve program was as follows: 95°C for 15 s, 60°C for 1 min, 65°C to 95°C at 2°C/min and 95°C for 12 s.
For validation of the primers, after each PCR cycling program a default melting program was run to make sure that disassociation curves for each pair of primers contained a single peak and the agorose gels of the amplified product revealed single band corresponding to the predictable amplicon length. To determine amplification efficiency a calibration curve was performed prior to the initiation of the experiments with excellent results.
Data collection and processing
The body weight of the experimental animals was monitored twice a week. Body weight increases of the individual mice were calculated in terms of percentage of body weight increase compared to the initial body weight.
IHC staining slides were digitally captured using a ZEISS AxioImager A1 light microscope equipped with an AxioCam ICC3 digital camera, as well as with an OLYMPUS BH-2 light microscope equipped with an OLYMPUS Q-color5 digital camera. The software programs used for these two microscopes were AxioVision Rel. 4.8 version (Carl Zeiss) and ImagePro Plus 5.0 version (MediaCybernetics), respectively. The MBP-immunostaining revealed many fine fibers, which ran parallel to the cortex and were heavily condensed in the CC. We measured the optical density of the MBP-immunostaining according to NIH ImageJ protocols for quantitative comparison [42] . Because the intensity of the DAB color varied slightly between samples (due to natural variations between similarly treated samples), the optical density measurements were standardized for color intensity using the optical density of the aca (as demonstrated by the single-arrow in Fig. 2 ), an area where MBP staining was consistent, regardless of CPZ challenge. All of these data collection methods were also applied to PDGFR-α staining because proliferation of the positive cells in the experimental groups increased so dramatically that we could not accurately count them. Analysis of GST-π staining was based on actual counts by an experimental-blind personnel on whole fields of view.
Statistical analysis
The statistic calculations and graphs were made using GraphPad Prism® version 5. Body weight data were analyzed using two-factor mix-measures analysis of variance (two-factor mix-measures ANOVA) assuming that genotype was the factor for independent grouping and that time elapsed was the other factor for repeated measures. All IHC data were analyzed using two-way independent ANOVA assuming that genotype was the factor for independent grouping and using groupings for different levels of experimental manipulation as another independent factor. All two-way ANOVA tests were accompanied by Bonferroni's multiple comparison tests. When considering variances caused by either factor (e.g., body weight changes within WT groups) one-factor ANOVA tests accompanied by Tukey's multiple comparison tests were applied. A p-value less than 0.05 (difference/effect) was considered to be significant.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Endothelins (ET-1, ET-2 and ET-3) are vasoactive peptides found in many tissues [1] . The first member of this family, ET-1 is present in brain endothelial cells [2] , neurons [3] , and astrocytes [4] , and its secretion increases in several pathologies, such as cerebral ischemia [5] , Alzheimer disease [6] , HIV infection [7] , [8] , reactive gliosis [4] , [9] , [10] and astrocytic tumours [11] . In astrocytes, ET-1 behaves as a growth factor, exerting important biological effects such as changes in protein content and morphology [12] , the induction of proliferation [13] , [14] , [15] , [16] and the increase in the rate of glucose uptake [16] , [17] , [18] .
The regulation of glucose uptake in astrocytes is an important aspect of brain function since glucose taken up by astrocytes is used not only by astrocytes but also to supply the neurons with metabolic substrates required to sustain synaptic transmission [19] , [20] . Astrocytes are connected through gap junctions [21] , composed mainly of connexin43 (Cx43) [22] . This intercellular communication provides the basis for several important astrocytic functions [23] . For instance, gap junction channels allow the passage from cell to cell of glucose and other metabolites, contributing to the distribution of metabolic substrates from blood to different regions [17] , [20] , [23] . Various physiological and pathological signals promote changes in gap junctional communication and Cx43 expression (for a review, see [23] ). One of these signals is ET-1, which rapidly inhibits gap junctional communication between astrocytes and after long-term exposure (24 h) causes the down-regulation of Cx43 [24] , [25] . In fact, we have shown that the down-regulation of Cx43 exerted by ET-1 is involved in the increase in the rate of glucose uptake observed in astrocytes [26] . This effect includes the up-regulation of the glucose transporter GLUT-1 and the induction of GLUT-3, an isoform not normally expressed in astrocytes [18] , [26] . Intracellular glucose is quickly phosphorylated by hexokinase (Hx) to glucose-6-phosphate, which is a charged molecule that cannot pass back through the plasma membrane and becomes trapped within the cell. Both type I (Hx-1) and type II (Hx-2) hexokinase are up-regulated by ET-1 in astrocytes [18] , [26] .
Hypoxia-Inducible Factor (HIF)-1α/β heterodimer is a master transcription factor for several genes involved in glucose uptake, angiogenesis, glycolysis, pH balance and metastasis. Among the genes activated by HIF-1 are GLUT-1, GLUT-3, Hx-1 and Hx-2 (revised in [27] ). While HIF-1β is stable and constitutively expressed, HIF-1α is highly regulated, as well as susceptible to oxygen-dependent degradation due to the sequential action of oxygen-dependent prolyl hydroxylases and the VHL ubiquitin ligase. Therefore, under hypoxic conditions HIF-1α is stabilized, dimerizes with HIF-1ß and activates target genes. It should be mentioned that although HIF is mainly activated under hypoxia, several factors activate HIF-1α under normoxic conditions. Intriguingly, endothelins are among the factors with the ability to activate HIF-1α under normoxic conditions in melanoma and ovarian carcinoma cells [28] , [29] , [30] . Furthermore, oncogenes, such as c-Src prevent hydroxylation-dependent ubiquitinylation of HIF-1α, thus stabilizing it under normoxic conditions [31] , [32] , [33] .
The intracellular carboxyl tail of Cx43 interacts with a large number of signalling and scaffolding proteins [34] , [35] , thereby regulating cell functions such as cell adhesion, migration and proliferation [36] , [37] , [38] , [39] . One of these interacting proteins is the non-receptor tyrosine kinase c-Src [40] , [41] , [42] . Interestingly, we have recently shown that the interaction between Cx43 and c-Src promotes the inactivation of the oncogenic activity of c-Src [43] . Although it is well described that ET-1 increases the rate of glucose uptake by a Cx43-dependent mechanism that includes the up-regulation of GLUT-1, GLUT-3, Hx-1 and Hx-2 [18] , [26] , so far there have been no studies examining the mechanism by which Cx43 regulated these genes. Since GLUT-1, GLUT-3, Hx-1 and Hx-2 are targets of the transcription factor HIF-1α, and this factor can be activated by c-Src, which interacts with Cx43, in this work we aimed to investigate the participation of HIF-1α and c-Src on ET-1 modulation of the rate of glucose uptake in astrocytes.
Results
ET-1 up-regulates HIF-1α by decreasing Cx43 expression in astrocytes
In previous work we showed that ET-1 increased the rate of glucose uptake in astrocytes. Thus, ET-1 up-regulated GLUT-1 and Hx-1 and induced the expression of isoforms not normally expressed in astrocytes, such as GLUT-3 and Hx-2 [18] , [26] . Since these proteins are target genes of HIF-1α, in this work we investigated whether the treatment with ET-1 modified HIF-1α levels in astrocytes, as it has been shown in other cell types [28] , [29] , [30] . Our results show that the treatment with 0.1 µM ET-1 strongly increased (by about 150%) the levels of HIF-1α ( Figures 1A and B ). This effect was evident after two hours of treatment and persisted for at least 48 h.
10.1371/journal.pone.0032448.g001
Figure 1
Effect of ET-1 on the expression of HIF-1α in astrocytes.
Astrocytes were incubated in the absence (control) or presence of 0.1 µM ET-1 for the indicated times. Then, the expression of HIF-1α and Cx43 was analysed by Western blot. A ) Representative Western blot of HIF-1α, Cx43 and GAPDH showing that ET-1 up-regulated HIF-1α and down-regulated Cx43. B ) HIF-1α quantification. C ) Cx43 quantification. The results are expressed as percentages of the level found in the controls at time 0. ***p<0.001 versus the corresponding controls.
The inhibition of gap junctional communication promoted by ET-1 in astrocytes is well documented [12] , [16] , [17] , [44] , [45] , [46] , [47] . Thus, ET-1 triggers very fast changes (within 3–10 minutes) in the gap junctional communication and in the phosphorylation status of Cx43 [48] . Changes in gap junctional communication and in the phosphorylation status of Cx43 are associated with changes in Cx43 interaction with other proteins and with Cx43 endocytosis [41] , [42] , [48] . Thus, a prolonged (24 hours) exposure to ET-1 reduces Cx43 expression in astrocytes [24] , [25] . Interestingly, the down-regulation of Cx43 promoted by ET-1 coincided with the up-regulation of HIF-1α ( Figure 1C ). Since our previous work indicated that the effect of ET-1 on glucose uptake was due to the reduction in Cx43 expression, we investigated the effect of decreasing Cx43 expression on HIF-1α levels. Thus, by using specific siRNA against Cx43 [26] , [43] we found that 48 h after the transfection with Cx43-siRNA the level of Cx43 was reduced by about 50% ( Figures 2A and 2C ) and the expression of HIF-1α was increased by about 80% ( Figure 2B ), when compared to astrocytes transfected with a non-targeting siRNA (NT-siRNA). Consequently, our results show that silencing Cx43 up-regulated HIF-1α. In addition, Figure 3 shows that the difference in HIF-1α levels between ET-1 and the control was not statistically significant in cells transfected with Cx43-siRNA, indicating that ET-1 was not able to further increase significantly the levels of HIF-1α in Cx43-silenced astrocytes.
10.1371/journal.pone.0032448.g002
Figure 2
Effect of silencing Cx43 on the expression of HIF-1α in astrocytes.
Astrocytes were transfected with NT-siRNA or with Cx43-siRNA. At the indicated times the expression of HIF-1α and Cx43 was analysed by Western blot. A ) Representative Western blot of HIF-1α, Cx43 and GAPDH showing that the decrease in Cx43 expression was concomitant with HIF-1α up-regulation. B ) HIF-1α quantification. C ) Cx43 quantification. The results are expressed as percentages of the level found in the NT-siRNA condition at time 0. ***p<0.001 versus the corresponding NT-siRNA values.
10.1371/journal.pone.0032448.g003
Figure 3
Effect of ET-1 on the expression of HIF-1α in Cx43-silenced astrocytes.
Astrocytes were transfected with NT-siRNA or with Cx43-siRNA as indicated. After 48 h, cells were incubated in the absence (control, C) or presence of 0.1 µM ET-1 for 24 h. HIF-1α levels were analysed by Western blot. Note that the differences between ET-1 and Control in Cx43-silenced astrocytes were lower than those found between ET-1 and Control in NT-siRNA or in non-transfected astrocytes. The results are expressed as percentages of the level found in control non-transfected cells. ***p<0.001 versus the corresponding controls; # p<0.05 versus the corresponding non-transfected cells. n.s. not significant.
HIF-1α mediates the increase in glucose uptake promoted by ET-1
Since our results indicate that the axis ET-1/Cx43 up-regulates HIF-1α, which is a transcription factor involved in the regulation of glucose uptake, our next goal was to investigate the participation of HIF-1α in the increase in the rate of glucose uptake promoted by ET-1 in astrocytes. To do so, the expression of HIF-1α was down-regulated by 3 different and specific siRNAs against HIF-1α (HIF-1α-siRNA) and the results compared with those obtained with NT-siRNA. Our results show that 48 h after the transfection with HIF-1α-siRNA the expression of HIF-1α was reduced by about 50% when compared with NT-siRNA ( Figure 4A ) and this effect persisted for at least 72 h ( Figure S1 ). Next, the effect of ET-1 was tested in astrocytes transfected with NT-siRNA or with HIF-1α-siRNA for 48 h. Our results show that the up-regulation of HIF-1α promoted by ET-1 in astrocytes transfected with NT-siRNA was reduced in HIF-1α-silenced astrocytes ( Figure 4A ). In a similar way, silencing HIF-1α strongly reduced the up-regulation of GLUT-1, GLUT-3, Hx-1 and Hx-2 promoted by ET-1 in astrocytes transfected with NT-siRNA ( Figure 4A ). Importantly, this effect was also observed when the rate of glucose uptake was analyzed. Thus, the rate of glucose uptake was reduced by HIF-1α-siRNA both in the control and in the ET-1 treated astrocytes ( Figure 4B ), suggesting that HIF-1α participates in the effect of ET-1 on glucose uptake.
10.1371/journal.pone.0032448.g004
Figure 4
Effect of ET-1 on glucose uptake in HIF-1α-silenced astrocytes.
Astrocytes were transfected with NT-siRNA or with HIF-1α-siRNA. After 48 h, astrocytes were incubated in the absence (control, C) or presence of 0.1 µM ET-1 for 24 h. A ) HIF-1α, Hx-1, GLUT-3, Hx-2, GLUT-1 and GAPDH Western blots and quantification. The results are the means ± SEM of at least three independent experiments and they are expressed as percentages of the level found in the control NT-siRNA. B ) Glucose uptake expressed as pmol of 2-deoxyglucose (2-DG) taken up per hour and per milligram of protein. The results show that the down-regulation of HIF-1α levels promoted by HIF-1α-siRNA decreased the rate of glucose uptake and the expression of GLUT-1, GLUT-3, Hx-1 and Hx-2, both in the control and in the ET-1 treated astrocytes. ***p<0.001, **p<0.01 and *p<0.05 versus the corresponding controls (C); ### p<0.001, ## p<0.01 and # p<0.05 versus the corresponding NT-siRNA.
c-Src mediates the increase in glucose uptake promoted by ET-1
Our recent work shows a direct relationship between Cx43 and c-Src activity [43] . Since c-Src has been shown to up-regulate HIF-1α [31] , [32] , [33] , we investigated whether c-Src mediates the effect of ET-1 and Cx43 on glucose uptake. To measure c-Src activity we analysed the levels of c-Src phosphorylated at Tyr-416 (Y416 c-Src), the active form of this tyrosine kinase [49] , [50] . Our results show that ET-1 transiently increased c-Src activity. Thus, ET-1 rapidly increased Y416 c-Src (within 6 minutes) and this effect was maintained for two hours, decreasing thereafter ( Figure 5A ). Meanwhile, the total amount of c-Src remained constant. Interestingly, silencing Cx43 also increased c-Src activity without changing the amount of total c-Src ( Figure 5B ).
10.1371/journal.pone.0032448.g005
Figure 5
Effect of ET-1 and Cx43 on c-Src activity in astrocytes.
A ) Astrocytes were incubated in the absence (control, C) or presence of 0.1 µM ET-1 for the indicated times, proteins were extracted and Y416 c-Src and total c-Src were analysed by Western blot. The results are expressed as percentages of the level found in the control and they show that ET-1 rapidly and transiently increased Y416 c-Src without affecting significantly total c-Src. ***p<0.001; *p<0.05 versus control. B ) Astrocytes were transfected with NT-siRNA or with Cx43-siRNA. After 48 h, proteins were extracted and Y416 c-Src, total c-Src and Cx43 were analysed by Western blot. The results are expressed as percentages of the level found in the control and they show that silencing Cx43 increased Y416 c-Src without affecting significantly total c-Src. **p<0.01 versus NT-siRNA.
In order to investigate whether the activation of c-Src was responsible for HIF-1α up-regulation, the activity of c-Src was inhibited by PP2 and the inactive analogue PP3 was used as a control. Thus, while in the presence of PP3, ET-1 immediately increased c-Src activity, in the presence of PP2 the effect of ET-1 on c-Src was abrogated ( Fig. 6 ). Next, the effect of ET-1 and Cx43 on HIF-1α was tested when c-Src was inhibited by PP2. Our results show that ET-1 was not able to up-regulate HIF-1α ( Figure 7A ) when c-Src was inhibited by PP2. Interestingly, the effect of ET-1 on the rate of glucose uptake was abrogated when c-Src was inhibited by PP2 ( Figure 7B ). Finally, our results show that silencing Cx43 did not up-regulate HIF-1α when the activity of c-Src was inhibited by PP2 ( Figure 7C ).
10.1371/journal.pone.0032448.g006
Figure 6
Inhibition of c-Src activity in astrocytes by PP2.
Astrocytes were preincubated with 100 ng/µL PP2 (c-Src inhibitor) or 100 ng/µL PP3 (inactive analogue) for 1 h and these agents were maintained for the rest of the experiment. Then, cells were incubated in the absence or presence of 0.1 µM ET-1 for the indicated times and proteins were extracted for the analysis of Y416 c-Src and total c-Src by Western blot. The results are expressed as percentages of the level found in the control (PP3, time 0) and they show that PP2 inhibited the activation of c-Src (Y416 c-Src) promoted by ET-1. ***p<0.001; *p<0.05 versus the corresponding time 0.
10.1371/journal.pone.0032448.g007
Figure 7
Effect of ET-1 and Cx43 on HIF-1α expression and glucose uptake when c-Src is inhibited.
Astrocytes were preincubated with 100 ng/µL PP2 (c-Src inhibitor) or 100 ng/µL PP3 (inactive analogue) for 1 h. Then, cells were incubated in the absence (control) or presence of 0.1 µM ET-1 for 24 h. A ) HIF-1α Western blot and quantification. The results are expressed as percentages of the level found in the controls treated with PP3 and they show that the inhibitor of c-Src PP2 prevented the up-regulation of HIF-1α promoted by ET-1. ***p<0.001 versus the absence of ET-1. B ) Glucose uptake expressed as pmol of 2-deoxyglucose taken up per hour and per milligram of protein. The results show that the inhibitor of c-Src PP2 prevented the increase in the rate of glucose uptake promoted by ET-1. ***p<0.001 versus the absence of ET-1. C ) Astrocytes were preincubated with 100 ng/µL PP2 or 100 ng/µL PP3 for 1 h. Then, cells were transfected with NT-siRNA or with Cx43-siRNA and after 48 h HIF-1α was analysed by Western blot. The results are expressed as percentages of the level found in the PP3 NT-siRNA and they show that the inhibitor of c-Src PP2 prevented the up-regulation of HIF-1α promoted by silencing Cx43. ***p<0.001 versus the corresponding NT-siRNA.
Discussion
Endothelins are involved in several important pathologies of the CNS such as ischemia, gliomas, reactive gliosis and Alzheimer's disease. One of their targets in the CNS is astrocytes, in which endothelins exert a broad range of biological effects. In this sense, the mitogenic effect of ET-1, its ability to inhibit gap junctional communication, to down-regulate Cx43, the protein forming gap junction channels, and to increase the rate of glucose uptake is well known [14] , [15] , [16] , [44] . In previous work we showed that ET-1 increased the rate of glucose uptake in astrocytes by a mechanism that includes the up-regulation of the glucose transporters GLUT-1 and GLUT-3 and the enzymes required to phosphorylate glucose, Hx-1 and Hx-2 [18] . In the present study we sought deeper into the signalling pathway responsible for this process. Our results show that ET-1 promoted an increase in the levels of HIF-1α in astrocytes under normoxic conditions, as it has been previously shown in melanoma and ovarian carcinoma cells [28] , [29] , [30] . When HIF-1α is silenced in astrocytes, the effects of ET-1 on GLUT-1, GLUT-3, Hx-1, Hx-2 and on the rate of glucose uptake were strongly reduced. Since HIF-1α is a transcription factor for GLUT-1, GLUT-3, Hx-1 and Hx-2 [27] , it could be proposed that the effect of ET-1 on the rate of glucose uptake in astrocytes is mediated by HIF-1α.
We have previously shown that the effect of ET-1 on glucose uptake is mediated by the reduction in the expression of Cx43 [26] . Agreeing with this, in this study we show that silencing Cx43 promoted an increase in the levels of HIF-1α and reduced the effect of ET-1 on HIF-1α, indicating that Cx43 participates in the effect of ET-1 on HIF-1α. Cx43 is the main protein forming gap junction channels in astrocytes, which implies that the protein is critical for important functions of astrocytes in the brain [23] , [51] . In fact, Cx43 expression can be reduced in response not only to ET-1 but also to various physiological and pathological stimuli (for a review, see [23] ). This study contributes to the identification of the signalling pathway evoked after Cx43 down-regulation that results in increased glucose uptake in astrocytes. In this sense, although the relationship between Cx43 and several signalling pathways is well known [34] , so far this is the first evidence showing a relationship between Cx43 and HIF-1α.
HIF-1α is mainly regulated by hypoxia but it can also be up-regulated by several mechanisms, including the activity of c-Src [31] , [32] , [33] , a non-receptor tyrosine kinase involved in the regulation of cell proliferation. Interestingly, the intracellular carboxyl tail of Cx43 interacts with c-Src [40] , [41] , [42] , and this interaction modulates reciprocally their activities. Thus, phosphorylation of Cx43 by c-Src reduces gap junctional communication [40] , [52] , [53] while the increase in Cx43 levels inhibits c-Src activity, at least in rat glioma cells [43] . Agreeing with this, our results show that both the treatment with ET-1 and with Cx43-siRNA increased c-Src activity in astrocytes. In addition, when c-Src activity was inhibited neither ET-1 nor silencing Cx43 were able to up-regulate HIF-1α. Consequently, the effect of ET-1 on glucose uptake was abrogated when c-Src is inhibited. It is well documented that the inhibition of gap junctional communication in astrocytes promoted by ET-1 and other gap junction uncouplers increases proliferation and glucose uptake [16] , [18] , [24] , [26] , however the molecular mechanism linking these cellular events is unknown. In this study we provide evidence that c-Src is activated by ET-1 and by silencing Cx43. c-Src is a well known regulator of cell proliferation [54] and this study and the work carried out by other laboratories [31] , [32] , [33] show that c-Src can activate HIF-1α and consequently glucose uptake. Therefore, it could be proposed that c-Src could be the mediator that links the increase in cell proliferation and glucose uptake found in astrocytes after inhibiting gap junctional communication or reducing Cx43 expression.
Furthermore, our results confirm that the level of Cx43 expression is important to regulate c-Src activity. Thus, the up-regulation of Cx43 in glioma cells reduces the high c-Src activity found in these cells [43] , while in this study we show that silencing Cx43 activates c-Src in astrocytes. It should be mentioned that this effect could be due to the absence of Cx43 function (gap junctional communication) or to the absence of Cx43 interaction with other proteins, such as c-Src. Changes in Cx43 expression are accompanied by changes in cell proliferation, thus restoration of Cx43 in glioma cells decreases the rate of proliferation [43] , [51] , [55] , [56] , while silencing Cx43 increases the rate of astrocyte proliferation [26] . Consequently, it could be proposed that the interaction between Cx43 and c-Src could be an important step in the regulation of cellular events, such as cell proliferation. In this context, the interaction between Cx43 and c-Src has been proposed to be responsible for the regulation of P2Y1 purinergic receptors, which are involved in glial calcium signal transmission and in the migration of neural progenitor cells [57] . Together, these data suggest a relevant role of the interaction between Cx43 and c-Src in the CNS.
As mentioned above, endothelins are involved in several pathologies of the CNS such as gliomas [11] . It should be mentioned that reduced expression of Cx43 [58] , [59] , [60] , [61] , high c-Src activity [62] and activation of the HIF-1 pathway [63] are all common features of gliomas. In addition, glioma cells, like many other cancer cells adapt their metabolism to the tumour environment, a process known as “Warburg Effect” that begins by an increase in the rate of glucose uptake and a metabolic shift to aerobic glycolysis, which is associated with a survival advantage as well as the generation of substrates necessary in rapidly proliferating cells [64] . For instance, enzymes such as Hx-2 are key mediators of aerobic glycolisis and promote tumour growth in gliomas [65] . In this study, we suggest that these events could be linked in a common pathway. Thus, our results suggest that ET-1 by down-regulating Cx43 activates c-Src, which in turn up-regulates HIF-1α leading to the transcription of the machinery required to increase the rate of glucose uptake in astrocytes ( Figure 8 ). These metabolic changes are probably designed to sustain the higher rate of cell proliferation observed under these circumstances [66] . ET-1 participates in the progression of different tumours [67] , however, whether the presence of ET-1 in gliomas [11] activates the pathway reported in this study ( Figure 8 ) to promote the growth and progression of these tumours remains to be elucidated and warrant further exploration.
10.1371/journal.pone.0032448.g008
Figure 8
Proposed mechanism.
Since Cx43 inhibits c-Src activity [43] , it could be proposed that ET-1 by reducing Cx43 activates c-Src, which in turn increases HIF-1α. HIF-1α dimerizes with HIF-1ß and translocates to the nucleus. The transcriptional activity of HIF induces the synthesis of the machinery required to augment the rate of glucose uptake in astrocytes, i.e., the glucose transporters, GLUT-1 and GLUT-3 and the enzymes required for glucose phosphorylation, type I and type II hexokinase (Hx-1 and Hx-2).
Materials and Methods
Materials
Dulbecco's modified Eagle medium (DMEM), penicillin, streptomycin, poly-L-lysine, ET-1, and protease inhibitors were obtained from Sigma-Aldrich Chemical Co. (Madrid, Spain). Fetal calf serum (FCS), DNase, bovine serum albumin and trypsin were from Boehringer Mannheim (Barcelona, Spain). Optimen and Lipofectamine 2000 were purchased from Invitrogen (Barcelona, Spain). Monoclonal antibody against glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and small interfering RNAs (siRNAs) were obtained from Genelink (New York, USA). 2-Deoxy-D[1- 14 C]glucose was from Amersham Pharmacia Biotech (Barcelona, Spain). The c-Src inhibitor, PP2 and the control PP3 were obtained from Calbiochem (Nottingham, United Kingdom). Mouse monoclonal antibody against Cx43 (610062) was from Transduction Laboratories, Inc. (BD Bioscience Pharmigen, San Diego, CA, U.S.A.). Mouse monoclonal antibody against Hx-1 (MAB1532) and Hx-2 (MAB1629) and rabbit polyclonal antibody against GLUT-1 (AB1340) and GLUT-3 (AB1344) were from Chemicon International Inc (Madrid, Spain). Rabbit polyclonal antibody against total c-Src (2108) and rabbit polyclonal antibody against Y416 c-Src (2101) were from Cell Signaling Technology (Boston, EEUU). Rabbit polyclonal antibody against HIF-1α (NB100-479) was from Novus Biologicals (Colorado, EEUU). Polyvinylidene fluoride membranes (PVDF) were from Millipore Corporation (Bedford, U.S.A.). The Bio-Rad protein assay and polyacrylamide were from Bio-Rad (Madrid, Spain). X-ray films were from Fujifilm (Madrid, Spain). Other chemicals were purchased from Sigma-Aldrich Chemical Co. (Madrid, Spain) or Merck (Barcelona, Spain).
Animals
Albino Wistar rats, fed ad libitum on a stock laboratory diet (49.8% carbohydrates, 23.5% protein, 3.7% fat, 5.5% wt/vol minerals and added vitamins and amino acids), were used for the experiments. Rats were maintained on a 12-h light–dark cycle. Postnatal day 1 newborn rats were used to prepare astrocyte cultures. The animals were obtained from the animal facility of the University of Salamanca and their use for this study was approved by the bioethics committee of this institution.
Cell cultures
Astrocytes in primary culture were prepared from the forebrains of 1- to 2-day-old Wistar rats as previously described [68] . Briefly, animals were decapitated and their brains immediately excised. After removing the meninges and blood vessels, the forebrains were placed in Earle's balanced solution containing 20 mg/mL DNase and 0.3% (w/v) bovine serum albumin. The tissue was minced, washed, centrifuged and incubated in 0.025% (w/v) trypsin and 60 mg/mL DNase for 15 min at 37°C in a shaking water bath. Trypsinization was completed by adding DMEM containing 10% (v/v) FCS. The tissue was then dissociated by trituration, passing it eight times through a siliconized Pasteur pipette, and the supernatant cell suspension was recovered. This procedure was repeated and the resulting cell suspension was centrifuged. The culture medium was DMEM (with 5 mM glucose) supplemented with 10% (v/v) FCS, penicillin (50 U/mL) and streptomycin (50 mg/mL). The cell pellet was resuspended in a known volume of culture medium and plated onto poly-L-lysine-coated Petri dishes or flasks at a density of 1.5×10 5 cells/cm 2 . Cells were incubated at 37°C in an atmosphere of 95% air/5% CO 2 with 90–95% humidity. After 3 days, cytosine arabinoside (10 mM) was added to the culture medium for 2 days to prevent the growth of microglia and cells from the O-2 lineage. The culture medium was renewed with a fresh one twice a week. Under our experimental conditions, 90–95% of the cells were astrocytes as determined by immunostaining against glial fibrillary acidic protein [69] , [70] , [71] . Experiments were carried out after 21 days in culture.
Cell treatments
The compounds tested were 0.1 µM ET-1, 100 ng/mL PP3 or 100 ng/mL PP2. PP2 and PP3 were preincubated for 1 hour and were present in the solutions used throughout the experiments. Incubations were performed for the indicated times in culture medium at 37°C in an atmosphere of 95% air/5% CO 2 with 90–95% humidity.
Transfection of siRNA
Conditions were similar to those reported previously [26] , [43] . Cells were transfected with the double-strand siRNA (50 nM) complexed with 2.5 µl/ml Lipofectamine 2000 in culture medium without antibiotics. The sequences of siRNAs were as follow: Cx43-siRNA: sense 5′-gcugguuacuggugacagatt-3′ and antisense 5′-ucugucaccaguaaccagctt-3′ ; HIF-1α siRNA: 5′-cugauaacgugaacaaauatt-3′ and antisense 5′-uauuuguucacguuaucagtt-3′ and a validated NT-siRNA provided by Ambion used as a negative control. Other siRNA sequences for HIF-1α were tested (sense 5′-cauugaagaugaaaugaaatt-3′, antisense 5′-uuucauuucaucuucaaugtt-3′ and sense 5′-cuguugaucuuauaaugautt-3′ , antisense 5′-aucauuauaagaucaacagtt-3′ ) and the same phenotype was observed ( Figure S1 ). The cells were maintained in the presence of the oligonucleotides in culture medium without antibiotics and after 6 h, the medium was replaced with DMEM plus 10% FCS with antibiotics. The extent of siRNA-mediated down-regulation of protein expression was evaluated in Western blots. Cell treatments were performed 48 hours after siRNA transfections.
2-Deoxyglucose uptake
Conditions were similar to those reported previously [17] , [18] , [26] . After the indicated treatments, cells were incubated with DMEM containing 2-deoxy-D[1- 14 C]glucose (750 dpm/pmol) for 30 min. Then, cells were washed with ice-cold PBS and were lysed by adding 500 µL of 10 mM NaOH containing 0.1% Triton X-100. An aliquot was assayed for [ 14 C] by liquid scintillation counting (efficiency 95%) and another aliquot was used to measure protein concentration.
Western blotting
Twenty-four hours after the treatments, proteins were extracted from the cells using 2% sodium dodecyl sulphate (SDS) in 5 mM Tris–HCl, pH 6.8, containing 2 mM EGTA, 2 mM EDTA, 2 mM phenylmethylsulphonyl fluoride, 0.5 µg/mL antipain, 0.5 µg/mL pepstatin, 0.5 µg/mL amastatin, 0.5 µg/mL leupeptin, 0.5 µg/mL bestatin, 0.5 µg/mL of trypsin inhibitor, sodium fluoride (NaF) 1 mM and sodium orthovanadate (Na 3 VO 4 ) 100 µM. The protein extract (80 µg) was applied to a 10% SDS–Polyacrilamide gel under reducing conditions and then transferred to a PVDF membrane. The membranes were cut into several strips to be immunoblotted with distinct antibodies, thus allowing for comparative analysis of the amount of each protein in the same sample. Membranes were then blocked with 10% fat-free dried milk in TTBS and then exposed to primary antibody against Cx43 (1∶100), Hx-1 (1∶500), Hx-2 (1∶500), GLUT-1 (1∶500), GLUT-3 (1∶500), c-Src (1∶500), Y416 c-Src (1∶500) or HIF-1α (1∶500), for at least 4 h. Mouse antibody against GAPDH (1∶5000) was used as a loading control. Peroxidase-conjugated anti-rabbit IgG or peroxidase conjugated anti-mouse IgG were used and developed with a chemiluminiscent substrate. Membranes were exposed to X-ray films. When necessary, blots were stripped and re-probed with other antibodies. X-ray films were scanned and densitometry analysis of the bands was performed using image-analyzer software (SCION IMAGE, based on NIH Image, Wayne Rasband, National Institutes of Health, Bethesda, MD, USA). The amounts of GAPDH recovered in each sample served as loading control and the values for each protein were normalized to their corresponding GAPDH level. The results are expressed as percentages of the values found in the controls.
Protein Determinations
Protein concentrations were determined by the Bradford method [72] , using bovine serum albumin (BSA) as the standard.
Statistical analyses
The results are the means ± S.E.M. of at least three independent experiments. Statistical analyses were carried out with Student's t -test or One-Way ANOVA followed by Tukey test when comparing more than two variables. Values were considered significant when p<0.05.
Supporting Information
Figure S1
Silencing HIF-1α in astrocytes by siRNA. Astrocytes were transfected with NT-siRNA or with 3 different sequences of siRNA specific for HIF-1α (sequence 1: sense 5′-cauugaagaugaaaugaaatt-3′ , antisense 5′-uuucauuucaucuucaaugtt-3′ ; sequence 2: sense 5′-cugauaacgugaacaaauatt-3′ and antisense 5′-uauuuguucacguuaucagtt-3′ ; sequence 3: sense 5′-cuguugaucuuauaaugautt-3′ and antisense 5′-aucauuauaagaucaacagtt-3′ ). At the indicated times the proteins were extracted and HIF-1α levels were analysed by Western blot. The results are expressed as percentages relative to the level found in cells transfected with NT-siRNA. ***p<0.001 versus NT-siRNA. Sequence 2 was selected for the following experiments because it showed the higher down-regulation after 48 h and the reduction was maintained for 72 h.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Genkwanin is one of the major non-glycosylated flavonoids in some herbs which have anti-inflammatory activities, such as Genkwa Flos ( Daphne Genkwa Sieb. et Zucc.) [1] , rosemary ( Rosmarinus officinalis L.) [2] and the leaves of Cistus laurifolius L. [3] . Previous pharmacological studies have found that genkwanin has a variety of pharmacological effects including anti-bacterial [4] , [5] , antiplasmodial [6] , radical scavenging [7] , chemopreventive [8] and inhibiting 17β-Hydroxysteroid dehydrogenase type 1 [9] activities. Although Pelzer et al . [10] has reported that genkwanin could inhibit the development of cotton-pellet-induced granuloma in rat, the potential molecular mechanisms of the anti-inflammatory activity of genkwanin remain obscure.
Inflammation is a central feature of many pathophysiological conditions in response to tissue injury and host defenses against invading microbes [11] . Key events in the inflammatory process include the expression of inflammatory cytokines, chemokines, and other mediators [12] . Macrophages play a central role in host defense against pathogen microbes by recognizing bacterial components and resulting in the activation of an arsenal of anti-microbiol effectors and initiation of the inflammatory cascade [13] , [14] . LPS, a major component of the outer membranes in Gram-negative bacteria, can be recognized by a TLR4 receptor complex [15] . Stimulation of TLR4 by LPS triggers the recruitment of adaptor protein MyD88, which in turn transmits a series of signaling cascades that lead to the activation of mitogen-activated protein kinase (MAPK) [16] - [19] . The MAPK is a group of highly conserved serine/threonine protein kinases, including p38, ERK1/2 and JNK. Once activated, MAPK phosphorylate downstream protein kinases and transcription factors, leading to the production of proinflammatory cytokines, such as iNOS, TNF-α, IL-1β and IL-6, etc.
The objective of the present paper is to clarify the anti-inflammatory mechanisms of genkwanin in LPS-activated RAW264.7 macrophages. Our results indicate that genkwanin suppresses the production of inflammatory mediator in LPS-activated RAW264.7 macrophages mainly through mediating microRNA-101 (miR-101)/MAPK phosphatase 1 (MKP-1)/MAPK pathway.
Materials and Methods
Materials
The murine macrophage RAW264.7 cell line was purchased from American Type Culture Collection (ATCC, Rockville, MD, USA). Genkwanin (≥98%) was purchased from Rochen Co. (Shanghai, China) and dissolved in DMSO at the concentration of 10 mg/mL. We confirmed that genkwanin from other source, the National Institutes for Food and Drug Control (Beijing, China), exhibited equivalent effects in crucial experiments. Mouse TNF-α and IL-6 ELISA kits were obtained from Biolegend Co. (San Diego, California, USA). Mouse IL-1β ELISA kit was obtained from Excell Technology Co. (Shanghai, China). Antibodies for iNOS, MKP-1 and β-actin were obtained from Santa Cruz Biotechnology, Inc. (Santa Cruz, California, USA). Antibodies for MAPK and Akt were from Cell Signaling Technology (Danvers, Colorado, USA). The plasmids for NFκB-TA-luc, AP1-TA-luc and their controls GL6-TA were from Beyotime Institute of Biotechnology (haimen, Jiangsu, China). The luciferase assay system and lipofectamine™ 2000 reagent were purchased from Promega Co. (Madison, Tennessee, USA) and Invitrogen (New York, California, USA), respectively. Horseradish peroxidase-conjugated anti-rabbit or mouse IgG secondary antibodies were obtained from Jackson ImmunoResearch Laboratories, Inc. (Lancaster, Philadelphia, USA). miR-101 qPCR kit with U6 snRNA (control), dsRNA mimic and ssRNA inhibitor for mmu-miR-101a were obtained from Genepharma (Shanghai, China). All other reagents were of analytical grade.
Cell culture and treatment
RAW264.7 macrophages were grown in DMEM supplemented with 10% heat-inactivated FBS and 1.0% penicillin-streptomycin solution in a humidified incubator with 5.0% CO 2 at 37°C. When performing genkwanin treatment, genkwanin was first added to the culture medium and then mixed thoroughly (the final concentrations of DMSO were ≤0.15%). The cell culture medium was replaced with the medium already mixed with genkwanin.
Measurement of nitrite
Cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 24 h. The nitrite concentration in the medium was measured as an indicator of NO production according to the Griess reaction. 100 µL of each supernatant was mixed with the same volume of Griess reagent (1% sulfanilamide in 5% phosphoric acid and 0.1% N-1-naphthylethylenediamine dihydrochloride in water). The absorbance was measured at a wavelength of 540 nm after incubation for 10 min. The nitrite concentration was calculated with reference to a standard curve of sodium nitrite generated from known concentrations. L-NAME was used as a positive control.
Measurement of iNOS enzyme activity
The activity of iNOS was assayed as previously described [20] with slight modifications. The cells were plated in a 25 cm 2 culture flask and incubated with LPS (Sigma, Escherichia coli 055: B5; 10 ng/mL) for 12 h. After being washed twice by PBS, the cells were harvested and plated in a 48-well plate, and incubated in the presence or absence of genkwanin at different concentrations for a further 12 h. The iNOS activity was assayed by measuring the nitrite level in the supernatant by Griess method. L-NAME was used as a positive control.
ELISA for TNF-α, IL-1β and IL-6
For the measurements of TNF-α, IL-6 and IL-1β, RAW264.7 macrophages were pretreated with genkwanin for 2 h and then stimulated with LPS (10 ng/mL) for 24 h. TNF-α, IL-6 and IL-1β in the cell supernatants were assayed using ELISA kits according to the manufacturer's instructions. The concentrations were calculated from the standard curves.
Quantitative real-time PCR (RT-qPCR)
The cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 4 h. The RNA extraction and RT- q PCR assays for the mRNA levels of iNOS, TNF-α, IL-1β and IL-6 were performed as we previously described [21] . RT- q PCR assay for the miR-101 level was performed according to the manufacturer's protocol. The cycling conditions were as follows: 95°C for 3 min, and then 40 cycles of 95°C for 12 s, 62°C for 40 s. The levels of iNOS, TNF-α, IL-1β and IL-6 were normalized to β-actin. The level of miR-101 was normalized to U6 snRNA.
Transfection of plasmids, dsRNA mimic and ssRNA inhibitor for mmu-miR-101a
The sense strand and the antisense strand of dsRNA mimic for mmu-miR-101a were 5′-UAC AGU ACU GUG AUA ACU GAA-3′ and 5′-CAG UUA UCA CAG UAC UGU AUU-3′, respectively. The strand of ssRNA inhibitor against mmu-miR-101a was 5′-UUC AGU UAU CAC AGU ACU GUA-3′. Plasmids for pNFκB-TA-luc, pAP1-TA-luc and their controls pGL6-TA, and RNAs were transfected into RAW264.7 macrophages using lipofectamine™ 2000 as described in the manufacturer's protocol.
Luciferase reporter assay
RAW264.7 macrophages were transfected with the luciferase reporter pAP-1-TA-luc (A) and pNF-κB-TA-luc (B). 24 h after transfection, the cells were pretreated with different concentrations of genkwanin for 2 h and then exposed to LPS (10 ng/mL). After 6 h, the cells were lysed and luciferase activity was measured using the Luciferase Assay System.
Western blot analysis
RAW264.7 macrophages were pretreated with genkwanin at indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 1 h (for p-p38, p-JNK, p-ERK1/2, MKP-1 and p-Akt assays) or 24 h (for iNOS assay). The western blot assay was performed as we previously described [21] . Proteins in nucleus or cytoplasm were extracted and separated by SDS-PAGE and transferred to polyvinylidene difluoride membranes. The membranes were blocked at room temperature for 4 h with 5.0% nonfat dry milk, and then incubated with each primary antibody at 4°C overnight. After washing, the membranes were incubated with HRP-conjugated secondary antibodies for 2 h at room temperature. The blots were visualized using enhanced chemiluminescence, and data were analysed using the Gel Doc EQ System (Bio-Rad).
Statistical analysis
Data represent the mean ± SD of at least three independent experiments, each experiment was performed in triplicate. One-way ANOVA was used to determine the statistical significance between different groups. A student t -test was used when only two groups were compared. Differences were considered to be significant at p <0.05.
Results
Genkwanin inhibits LPS-induced NO production and suppresses iNOS at the transcriptional and translational levels
Cell viability analysis showed that genkwanin did not affect the cell viability up to a concentration of 50 µM ( Figure S1 ). NO, a small diffusible molecular generated by iNOS in activated macrophages, is closely related to many inflammatory diseases [22] – [25] . Thus, to investigate the effect of genkwanin on inflammation, we first measured supernatant NO production in LPS-stimulated RAW264.7 macrophages. As shown in Figure 1A , genkwanin inhibited the LPS-induced production of NO in a concentration-dependent manner. As we known, iNOS only expresses in the present of external stimulus [26] . To assay the effect of genkwanin on iNOS enzyme activity, we pretreated the cells with LPS for 12 h and then removed LPS. In a 48-well plate, the re-plated cells could not produce new iNOS in the absent of LPS. Under this condition, all change of NO production was attributed to the change of iNOS enzyme activity rather than iNOS mass. As shown in Figure 1B , genkwanin could not significantly affect the activity of iNOS.
10.1371/journal.pone.0096741.g001 Figure 1
Effects of genkwanin on NO production and iNOS in LPS-activated RAW264.7 macrophages.
(A) Effects of genkwanin on LPS-induced NO production. Cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 24 h. After treatment, nitrite levels in the medium were measured by Griess reaction. L-NAME was used as a positive control. (B) Effects of genkwanin on iNOS enzyme activity. Cells were pretreated with LPS (10 ng/mL) for 12 h and then exposed to genkwanin at the indicated concentrations for a further 12 h without LPS. Nitrite levels in the medium were measured. L-NAME was used as a positive control. (C) Effects of genkwanin on iNOS mRNA expression. Cells were pretreated with the indicated concentrations of genkwanin for 2 h and then exposed to LPS (10 ng/mL) for 4 h. mRNA of iNOS was determined by RT- q PCR analysis. (D) Effects of genkwanin on iNOS protein levels. Cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 24 h. After treatment, cellular proteins were prepared and the iNOS protein levels were determined by Western blot analysis. Bars represent mean ±SD of three independent experiments. ## p <0.01 vs. normal control group; ** p <0.01 vs. LPS alone.
Thus, we next investigated the inhibitory effects of genkwanin on iNOS mRNA and protein levels. As shown in Figure 1C–D , LPS stimulation of RAW264.7 macrophages resulted in a dramatic increase in iNOS at the transcriptional ( Figure 1C ) and translational ( Figure 1D ) levels. Treatment with genkwanin concentration-dependently inhibited the LPS-induced increase in iNOS mRNA expression and protein levels.
Genkwanin suppresses LPS-induced TNF-α, IL-1β and IL-6 at the transcriptional and translational levels
The effect of genkwanin on the production of proinflammatory cytokines was examined. As shown in Figure 2A , genkwanin suppressed the productions of TNF-α, IL-1β and IL-6 in LPS-stimulated RAW264.7 macrophages in a concentration-dependent manner. We next analysed the effects of genkwanin on the mRNA quantities of TNF-α, IL-1β and IL-6 by RT- q PCR. As shown in Figure 2B , genkwanin consistently down-regulated the LPS-induced transcription of TNF-α, IL-1β and IL-6 mRNA in a concentration-dependent manner.
10.1371/journal.pone.0096741.g002 Figure 2
Effects of genkwanin on TNF-α,IL-1β and IL-6 in LPS-activated RAW264.7 macrophages at the transcriptional and translational levels.
(A) The cells were pretreated with the indicated concentrations of genkwanin for 2 h and then exposed to LPS (10 ng/mL) for 24 h. The levels of TNF-α, IL-1β and IL-6 in the supernatant were determined by ELISA. (B) The cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 4 h. The mRNA expressions of TNF-α, IL-6 and IL-1β were determined by RT- q PCR analysis. ## p <0.01 vs. normal control group; ** p <0.01 vs. LPS alone. Bars represent mean ±SD of three independent experiments.
Genkwanin suppresses the LPS-induced phosphorylation of p38 and JNK via the up-regulation of MKP-1 expression
Since the induction of proinflammatory mediators by LPS is known to be predominantly regulated by NF-κB and AP-1 [27] – [30] , we investigated if genkwanin exerts anti-inflammatory activities by affecting these two pathways. As shown in Figure 3 , genkwanin significantly suppressed the AP-1 signaling pathway ( Figure 3A ) but had little effect on the NF-κB signaling pathway ( Figure 3B ). Thus, we next explored the MAPK, which mainly act upstream of AP-1, to determine the target of genkwanin. MAPK signal transduction pathways are classified into three components. Hence, the effects of genkwanin on LPS-induced phosphorylation of p38, ERK1/2 and JNK were investigated. The results indicate that genkwanin suppresses the phosphorylation of p38 and JNK in a concentration-dependent manner, but little affects ERK1/2 phosphorylation ( Figure 4A ).
10.1371/journal.pone.0096741.g003 Figure 3
Effects of genkwanin on LPS-induced AP-1 and NF-κB activities.
RAW264.7 macrophages were transfected with the luciferase reporter pAP-1-TA-luc (A) and pNF-κB-TA-luc (B). 24 h after transfection, the cells were pretreated with genkwanin for 2 h and then exposed to LPS (10 ng/mL) at the indicated concentrations. After 6 h, the luciferase activity was determined. ## p <0.01 vs. normal control group; ** p <0.01 vs. LPS alone. Bars represent mean ±SD of three independent experiments.
10.1371/journal.pone.0096741.g004 Figure 4
Effects of genkwanin on LPS-induced MAPK and MKP-1.
(A) Effects of genkwanin on the LPS-induced phosphorylation of p38, JNK and ERK1/2. RAW264.7 macrophages were pretreated with genkwanin at indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 1 h. Total proteins were extracted for the Western blot analysis. (B) Effects of genkwanin on MKP-1 mRNA in LPS-activated RAW264.7 macrophages. The cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 1 h. The mRNA expression of MKP-1 was determined by RT- q PCR analysis. (C) Effects of genkwanin on MKP-1 protein level in LPS-activated RAW264.7 macrophages. Cells were pretreated with genkwanin at indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 1 h. Cell lysates were analysed by Western blot. ## p <0.01 vs. normal control group; ** p <0.01 vs. LPS alone. Bars represent mean ±SD of three independent experiments.
Originally identified as an immediate early gene, MKP-1 was then found to be a dual specificity phosphatase acting as a negative regulator of ERK1/2, JNK and p38 MAPK activities, with predominant effects on the latter two [31] – [36] . Thus, we next explored the effect of genkwanin on MKP-1. As shown in Figure 4B–C , LPS stimulation induced the expression of MKP-1 at the transcriptional and translational levels. Pretreatment with genkwanin markedly up-regulated the expression of MKP-1 without affecting MKP-1 mRNA.
Genkwanin exerts anti-inflammatory effects mainly through decreasing miR-101 production
Our above results have demonstrated that genkwanin up-regulates MKP-1 at the posttranscriptional level. It was previously found that MKP-1 as a target of miR-101 which can repress MKP-1 protein expression [37] . Thus, we next evaluated the effect of genkwanin on miR-101 expression by RT- q PCR. As shown in Figure 5A , LPS induced the expression of miR-101, but pretreatment with genkwanin significantly decreased miR-101. We also analysed the effect of miR-101 on the protein level of MKP-1 in LPS-activated RAW264.7 macrophages which had been transfected with dsRNA mimic or ssRNA inhibitor for mmu-miR-101a (miR-101 mimic or miR-101 inhibitor). As shown in Figure 5B , transfection with miR-101 mimic significantly inhibited the production of MKP-1 protein, while miR-101 inhibitor markedly increased MKP-1 protein. These results are consistent with the previous reports that miR-101 is a negative regulator of MKP-1 expression [37] .
10.1371/journal.pone.0096741.g005 Figure 5
Genkwanin exerts anti-inflammatory effects mainly through decreasing miR-101 production.
(A) Effect of genkwanin on miR-101 expression in LPS-activated RAW264.7 macrophages. The cells were pretreated with genkwanin at the indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 90 min. miR-101 expression was determined by RT- q PCR analysis. (B) Effect of miR-101 on MKP-1 protein level. RAW264.7 macrophages were transfected with miR-101 mimic or inhibitor or their negative controls, and then stimulated with LPS (10 ng/mL) for 1 h. Cell lysates were analysed by Western blot. (C–D) Effects of genkwanin on supernatant TNF-α in LPS-stimulated RAW264.7 macrophages which have been transfected with ssRNA inhibitor (C) or dsRNA mimic (D) for mmu- miR-101a. RAW264.7 macrophages were transfected with miR-101 inhibitor or mimic or their negative controls, and then stimulated with LPS (10 ng/mL) for 24 h in the presence or absence of genkwanin (50 µM). Supernatant TNF-α was measured by ELISA. ## p <0.01 vs. resting cells transfected with miR-101 inhibitor or mimic; * p <0.05 vs. LPS-treated cells transfected with miR-101 inhibitor or mimic; ΔΔ p <0.01 vs. resting cells transfected with negative controls; @@ p <0.01 vs. LPS-treated cells transfected with negative controls. (E) Effect of genkwanin on the LPS-induced p-Akt. Cells were pretreated with genkwanin at indicated concentrations for 2 h and then exposed to LPS (10 ng/mL) for 1 h. Cell lysated were assayed by Western blot analysis using Akt and p-Akt antibodies. ## p <0.01 vs. normal control group. Bars represent mean ±SD of three independent experiments.
To understand how genkwanin suppressed inflammation via miR-101, we transfected miR-101 inhibitor into RAW264.7 macrophages. In the resulting miR-101-deficient cells, the TNF-α production in response to LPS was significantly decreased as compared with the cells transfected with ssRNA negative control (NC) ( Figure 5C ). Genkwanin potently decreased LPS-induced supernatant TNF-α with an inhibition rate (IR) value of ∼38% {IR% = [(TNF-α LPS - TNF-α LPS+Genkwanin )/TNF-α LPS ] ×100} in ssRNA NC cells. However, in miR-101-deficient cells, genkwanin only slightly decreased LPS-induced supernatant TNF-α with an IR value of ∼10%, indicating that miR-101 played a predominant role in the anti-inflammatory activity of genkwanin.
Next, we transfected miR-101 mimic into RAW264.7 macrophages. In the resulting miR-101-abundant cells, the TNF-α production in response to LPS was significantly increased as compared with the cells transfected with dsRNA NC ( Figure 5D ). Genkwanin significantly decreased LPS-induced TNF-α with an IR value of ∼36% in dsRNA NC cells, while in miR-101-abundant cells, the IR value dropped to ∼7%, indicating that genkwanin was not effective against exogenous miR-101. Similar effects of miR-101 inhibitor and mimic on supernatant NO, IL-1β and IL-6 were also observed ( Figure S2 ). Next, we evaluated the effects of genkwanin on the phosphorylation level of Akt in LPS-activated RAW264.7 macrophages. As shown in Figure 5E , LPS could induce the phosphorylation of Akt, but genkwanin did not appreciably affect the level of phospho-Akt (p-Akt).
Discussion
Genkwanin (4′,5-dihydroxy-7-methoxyflavone), as one of the major bioactive components in Genkwa Flos, is used as a representative index for the quality control of this herb and are included in the State Pharmacopoeia Commission of the People's Republic of China [1] . Many of its structurally similar analogues, such as apigenin [38] , acacetin [39] , chrysin [40] , baicalein [41] , wogonin [42] , luteolin [43] and velutin [44] ( Figure 6 ), were reported to suppress proinflammatory mediators in LPS-stimulated macrophages. Analysis of the structure-activity relationships of flavones showed that the anti-inflammatory effects could be enhanced by the methoxylation of the 5- or 7-hydroxyl groups on the A-ring or non-methoxylation of the 3′-hydroxyl groups on the B-ring [45] . Coincidentally, the chemical structure of genkwanin includes these dispositions, such as the 5-OCH 3 and 3′-OH. Indeed, our results show that genkwanin (12.5 µM - 50 µM) potently decreases LPS-induced proinflammatory mediators, such as iNOS, TNF-α, IL-1β and IL-6 at the transcriptional and translational levels in RAW264.7 macrophages without cytotoxicity ( Figures 1 and 2 ), indicating genkwanin's excellent anti-inflammatory potency.
10.1371/journal.pone.0096741.g006 Figure 6
Chemical structures of genkwanin, apigenin, acacetin, wogonin, chrysin, baicalein, luteolin and velutin.
Of course, due to the minor structural differences, the anti-inflammatory mechanisms of the analogues are diverse. For example, apigenin [46] , baicalein [47] and luteolin [43] exert their anti-inflammatory effects mainly by inactivating NF-κB, while wogonin [48] can block JNK phosphorylation. Specially, acacetin [39] and velutin [44] can inactivate both NF-κB and MAPK. In this study, after carrying out Western blot assays by using phosphorylation antibodies, we found that genkwanin decreased the LPS-induced phospho-p38 (p-p38) and phospho-JNK (p-JNK) levels, but had no effect on phospho-ERK1/2 (p-ERK1/2) ( Figure 4A ).
Reversible activation MAPK requires phosphorylation on threonine and tyrosine residues of the activation domain of p38, JNK and ERK1/2. They are negatively regulated by a family of dual-specificity (threonine/tyrosine) phosphatases known as the MAPK phosphatases (MKPs) [49] . MKP-1, a stress-responsive MKP, localizes to the nucleus through its N terminus [50] and preferentially dephosphorylates activated p38 and JNK relative to ERK1/2. In LPS-stimulated mouse macrophages, MKP-1 shows a transient expression pattern with rapid induction, followed by a quick return to basal levels [51] . It is a critical negative regulator of macrophage signaling in response to inflammatory stimuli and is responsible for switching off the production of proinflammatory cytokines [51] – [53] . Therefore, the differential regulations of genkwanin on MAPK phosphorylation strongly suggest that the upstream regulator may be MKP-1. As expected, our results show that genkwanin increases the MKP-1 expression at the posttranscriptional level ( Figure 4B–C ).
MicroRNAs (miRNAs), the short (∼22 nucleotides) non-coding RNAs, play a central role in the regulation of gene expression at the posttranscriptional level via an RNA interference mechanism [54] . Recently, it was found that miR-101, a tumor-related miRNA, repress MKP-1 expression by binding to the 3′ untranslated region of MKP-1 in a direct and sequence-specific manner [37] . In our study, we also found the negative regulatory effect of miR-101 on MKP-1 protein ( Figure 5B ). Moreover, in response to LPS, the supernatant TNF-α, NO, IL-1β and IL-6 levels of miR-101 deficient cells is decreased ( Figures 5C and S 2), while these levels of miR-101-abundant cells is increased ( Figures 5D and S 2). Based on these slight effects of genkwanin on LPS-induced TNF-α, NO, IL-1β and IL-6 in exogenous miR-101-abundant cells ( Figure 5D ), we infer that genkwanin up-regulates MKP-1 protein may be attributed to the decrease of miR-101 production. In LPS-stimulated miR-101-deficient cells, genkwanin still somewhat suppresses supernatant TNF-α, NO, IL-1β and IL-6 ( Figures 5C and S 2), which suggests another mechanism, not depending on miR-101, remains possible. As predicted, genkwanin not only can decrease the phosphorylation level of JNK ( Figure 4A ), but also can directly inhibit the activity of p-JNK ( Figure S3 ).
PI3K/Akt is known to regulate proinflammatory cytokine expression, but its exact role (positive versus negative) is controversial. Some studies have demonstrated that the PI3K/Akt pathway negatively regulates TLR-induced MAPK activation and proinflammatory cytokine production [55] – [58] . Other reports, however, have displayed that the PI3K/Akt implicated as a positive regulator of TLR-induced inflammatory response [59] – [61] . Zhu and his colleagues [37] proposed that PI3K/Akt negatively regulated the expression of MKP-1 through the induction of miR-101. However, our results indicated that genkwanin could not significantly affect the level of p-Akt ( Figure 5E ), suggesting that the phosphorylation of Akt may be not responsible for the effect of genkwanin on miR-101 production.
Taken together, our results demonstrate that the anti-inflammatory effect of genkwanin may be mainly attributed to the down-regulation of the LPS-induced miR-101, thus increasing the protein expression of MKP-1, which dephosphorylates p38 and JNK in RAW264.7 macrophages ( Figure 7 ). To our knowledge, genkwanin is the first compound derived from plant source shown to exert its anti-inflammatory activities mainly through the decrease of miR-101 production.
10.1371/journal.pone.0096741.g007 Figure 7
Proposed mechanism by which genkwanin exerts anti-inflammatory activity in LPS-activated RAW264.7 macrophages.
The gray color indicates the targets of genkwanin.
Supporting Information
Figure S1
Effect of genkwanin on cell viability. RAW264.7 macrophages were incubated with genkwanin for 24 h and the cell viability were evaluated by MTT assay. Data represent the mean ± SD of three independent experiments.
(TIF)
Figure S2
Effects of genkwanin on supernatant NO, IL-1β and IL-6 in LPS-stimulated RAW264.7 macrophages which have been transfected with ssRNA inhibitor or dsRNA mimic for mmu-miR-101a. RAW264.7 macrophages were transfected with miR-101 inhibitor or mimic or their negative controls, and then stimulated with LPS (10 ng/mL) for 24 h in the presence or absence of genkwanin (50 µM). Supernatant NO (A–B), IL-1β (C–D) and IL-6 (E–F) were measured. ## p <0.01 vs. resting cells transfected with miR-101 inhibitor or mimic; * p <0.05 vs. LPS-treated cells transfected with miR-101 inhibitor or mimic; ΔΔ p <0.01 vs. resting cells transfected with negative controls; @@ p <0.01 vs. LPS-treated cells transfected with negative controls.
(TIF)
Figure S3
Effect of genkwanin on p-JNK activity. RAW264.7 macrophages were treated with or without LPS (10 ng/mL) for 1 h. The intracellular p-JNK was extracted and purified by immunoprecipitation. The obtained p-JNK was treated with genkwanin for 15 min at room temperature and then incubated with c-Jun protein and ATP substrate. The effect of genkwanin on p-JNK activity was assayed by Western blot analysis and represented as the blots of p-c-Jun. All of the extraction and purification of p-JNK and the kinase activity assay were performed according to the manufacturer's instructions of KinaseSTAR JNK Activity Assay Kit (BioVision, Inc., San Francisco, California, USA). ## p <0.01 vs. normal control group; ** p <0.01 vs. LPS alone. Bars represent mean ±SD of three independent experiments.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
With the increasing global competition and growing complexity of products, the division of labour is becoming increasingly specialized. As a result, the core firm requires joint development involving customers, suppliers and research institutes to overcome these limitations. Through cross-organizational collaborative product design, it can realize the maximization of resource integration and knowledge sharing as well as the improvement of design efficiency. However, in the process of collaborative product design (CPD), the diversity of design agent and interdependence and mutual restriction between tasks make the collaborative product design process quite complicated. Therefore, design task and resource should be reasonably allocated to shorten the development cycle and reduce cost.
There is a great amount of research work on the task and resource allocation of a collaborative design project. Some of these research studies focused on task identification, task relationship analysis and task scheduling based on Petri Nets[ 1 ] and Design Structure Matrix (DSM)[ 2 – 3 ]. Other research studies focused on the establishment of a task and resource dynamic scheduling optimization model and a model solution based on heuristic algorithm[ 4 ] and intelligent algorithm, such as the Genetic Algorithm[ 5 ], Ant Colony Optimization[ 6 ], Particle Swarm Optimization[ 7 ], Artificial Bee Colony[ 8 ]. Pang et al. [ 9 ] established a design task net and constructed a task assignment model from tasks to team members based on the principle of equilibrium-moderation. Li et al. [ 10 ] proposed a two-stage multi-agent resource allocation method, including the arbitration of manager agent and design agent selection according to task priority function. Regard collaborative production tasks as a directed weighted complex network, Yu et al. [ 11 ] proposed an evolution model for simulating collaborative production task state to perturbations. In order to deal with the collaboration between task decomposition and task scheduling, Liu et al. [ 12 ] put forward a new method for task granularity quantitative analysis, which is used to guide the coarse-grained task decomposition and recombine the subtasks with low cohesion coefficient. Currently, in the study of capacity and matching degree for CPD, Frillman et al. [ 13 ] proposed a competency model for engineers functioning in a PLM environment that emphasized individuals' competencies. Wu et al. [ 14 ] proposed a resource capability measuring method and resource capability deployment mechanism by mapping resource task capability item (RTCI) to resource physical capability item (RPCI). Combined with cost and productivity considerations, Tanuchporn et al. [ 15 ] proposed a multi-objective ergonomic workforce scheduling model to minimize the number of utilized workers and the total worker-task changeover, maximize the total worker-task fit score. Based on agent simulation, Zhang and Li [ 16 ] simulated the human working behaviours in a collaborative product development process, where the design agent selected her/his partner according to the ability and character matching degree. Furthermore, Li and Zhang [ 17 ] analysed the static single category resource scheduling problem and the multi-category resource static scheduling problem. Based on ontology and service capabilities, He and Hu [ 18 ] proposed matching rules and algorithms of manufacturing tasks and services. However, these research studies did not consider matching between tasks and the collaboration team.
For a collaborative product design project, the project is decomposed into tasks first, and then, the tasks are allocated to the collaborative team. Next, the tasks are decomposed into sub-tasks or more detailed tasks; these sub-tasks or detailed tasks are then assigned to individuals. The previous research studies have focused on the matching between a task and an individual based on task priority or designer preference. The question arises, taking the design team as a whole, from the perspective of system engineering, what is the method to realize reasonable task-team matching? Furthermore, partner selection or task assignment requires measurement of the collaborative team comprehensive capacity. This concept refers to not only individual competency but also members’ cooperation. In addition, for task allocation, it is necessary to evaluate the capability of collaborative team while considering the cost.
In the sections that follow, the capacity model of a collaborative team is presented first. Next, the 2-tuple linguistic evaluation method is adopted to evaluate the capacity of the collaborative team. Subsequently, the matching degree (MD) is defined. Afterwards, a scheduling model considering matching degree is established, and the improved genetic-annealing algorithm is designed to solve the scheduling model. An example is solved successfully to illustrate the feasibility and validity of the proposed method and model. Finally, conclusions are presented.
Team capacity evaluation based on the 2-tuple linguistic method
Capacity model of collaborative team
Collaborative product design, as a multi-agent and knowledge-intensive activity, emphasizes collaborative work between design teams. Moreover, creative customers and suppliers are involved. These innovative design agents have different background knowledge, experience, skill level and interests, i.e., each team has its own special abilities and resources. Therefore, collaborative product design requires not only reasonable design task decomposition but also reasonable matching between innovation team and task, such matching has important influence on the efficiency and cost of product design.
Capacity reflects the skill or ability sets necessary for the relevant tasks. The capacity model requires a description of the capacity elements for a task. When finding an appropriate team to conduct a design task, team capacity should be considered. For collaborative work, information sharing, goal congruence, decision synchronization, resource sharing, collaborative communication, and joint knowledge creation are significant and interconnecting elements[ 19 – 20 ]. Moreover, they are the prerequisite elements. Thus, the capacity model of a collaborative team is constructed as shown in Fig 1 .
10.1371/journal.pone.0200753.g001
Fig 1
Capacity model of collaborative product design.
In the model, the basic resources of a collaborative product design team are information resources, hardware and software resource, brand resource and social net resource. The information resource includes available technical information and industry information. Important customers, government, and partners in other industries constitute the team’s social net resource. The comprehensive capacity consists of team learning capability, communication capability, team executive capability, technical capability, service consciousness, and management capability. Learning capability and communication capability are more important than the others at this level. The core capacities are team innovation capability and collaboration capability. Team collaboration requires good communication and executive ability as well as excellent team management. Learning capability and technical capability are important prerequisites and serve as the foundation for innovation. Finally, high efficiency and high quality are the ultimate goals.
Team capacity evaluation based on the 2-tuple linguistic method
For capacity evaluation, the common methods are based on fuzzy mathematics theory, such as AHP and triangular fuzzy numbers. However, in these methods, fuzzy operation based on the extension principle increases the fuzziness of the results and causes information loss or distortion. In addition, evaluation experts often adopt natural language to express their preference, e.g., they use ‘‘high”, ‘‘average” and ‘‘low” to evaluate the team capacity, or they use ‘‘very high”, ‘‘high”, ‘‘average”, ‘‘low” and ‘‘very low” to express their evaluation results. In other words, different experts can express their evaluation information at different levels of granularity. The 2-tuple linguistic method can effectively aggregate natural language evaluation information of different levels of granularity to avoid information loss and make the result more precise[ 21 – 22 ]. Thus, the 2-tuple linguistic method is adopted to evaluate the competencies of the collaborative team.
The 2-tuple linguistic method represents the linguistic evaluation information by means of a two-tuple ( s i , α i ), where s i is a linguistic label from predefined linguistic term set S = { s 0 , s 1 , …, s g }; α i is the value of symbolic translation, α i ∈[-0.5,0.5); and g +1 is the granularity of the set S . For example, a set S = { s 1 , s 2 , s 3 , s 4 , s 5 } represents the evaluation information set. The meanings of linguistic terms s 1 , s 2 , s 3 , s 4 , and s 5 are ‘‘very high”, ‘‘high”, ‘‘average”, ‘‘low” and ‘‘very low”, respectively.
Definition 1
A real number β ∈[0, g ] is a number value representing the aggregation result of the linguistic symbols. The function Δ used to obtain the 2-tuple linguistic information equivalent to β is defined as:
Δ : [ 0 , g ] → S × [ − 0.5 , 0.5 ) , Δ ( β ) = { s k , k = r o u n d ( β ) α k = β − k , α k ∈ [ − 0.5 , 0.5 )
(1)
where round () is the rounding operator, S k has the closest index label to β , α k is the value of the symbolic translation.
In contrast, the 2-tuple linguistic variable can be converted into the crisp value β by the inverse function Δ -1 :
Δ − 1 : S × [ − 0.5 , 0.5 ) → [ 0 , g ] , Δ − 1 ( s k , α k ) = k + α k = β
(2)
Definition 2
Let S = {( s 1 , α 1 ), ( s 2 , α 2 ), …, ( s m , α m )} be a 2-tuple linguistic variable set at a given granularity, the arithmetic average operator of the set is computed as follows:
( s ¯ , α ¯ ) = Δ [ 1 m ∑ j = 1 m Δ − 1 ( s j , α j ) ] , s ¯ ∈ S , α ¯ ∈ [ − 0.5 , 0.5 )
(3)
Definition 3
Let S = {( s 1 , α 1 ), ( s 2 , α 2 ), …, ( s t , α t )} be a set of 2-tuples and C = {( c 1 , ß 1 ), ( c 2 , ß 2 ),…, ( c t , ß t )} be the linguistic weighting vector of 2-tuple ( s k , α k )( k = 1,2,…, t ). The extended 2-tuple weighted geometric (ET-WG) operator is defined as follows[ 23 – 24 ]:
( s ˜ , α ˜ ) = E T _ W G C ( ( s 1 , α 1 ) , ( s 2 , α 2 ) , … , ( s t , α t ) ) = Δ ( ∏ k = 1 t ( Δ − 1 ( s k , α k ) ) Δ − 1 ( c k , β k ) ∑ k = 1 t Δ − 1 ( c k , β k ) )
(4)
Definition 4
Let ( s ˜ 1 , α ˜ 1 ) , ( s ˜ 2 , α ˜ 2 ) , …, ( s ˜ u , α ˜ u ) be the two-tuple linguistic information with different granularities that will be aggregated. u is the number of groups. The improved EOWA operator is defined as:
( s * , α * ) = I E O W A ( ( s ˜ 1 , α ˜ 1 ) , ( s ˜ 2 , α ˜ 2 ) , … , ( s u ˜ , α ˜ u ) ) = Δ ( λ ′ i ( Δ − 1 ( s ˜ i , α ˜ i ) ) )
(5)
where ( s ˜ i , α ˜ i ) is the evaluation information with the i th maximum granularity, and λ ′ i is the i th maximum number in array λ . λ = ( λ 1 , λ 2 , …, λ u ) is the weight of EOWA operator that is quantified by the fuzzy operator E(r) :
λ i = E ( i / u ) − E ( ( i − 1 ) / u ) , i = 1 , 2 , … , u
E ( r ) = { 0 r < a ( r − a ) / ( b − a ) a ≤ r ≤ b 1 r > b
(6)
where a , b , and r ∈[0, 1] correspond to the fuzzy linguistic quantitative principle of “half”, “most” and “as much as possible”, respectively, with the parameters ( a , b ) taking on values of (0, 0.5), (0.3, 0.8), and (0.5, 1), respectively.
The specific evaluation steps are as follows:
Step 1 . The experts with the same granularity are divided into a group. The weight evaluation result of expert k ( k = 1, 2,…, t ) for capacity is denoted as ( c k y , β k y ). The evaluation result of team j for task i in capacity given by expert k is denoted as ( c k i j y , β k i j y ). According to Eq (4) , the integrated information of group with the same granularity, denoted as ( s ˜ i j y , α ˜ i j y ) , is obtained.
Step 2 . Obtain the weight vector λ ’ = ( λ 1 ′ , λ 2 ′ , … , λ u ′ ) according to the improved EOWA operator, and then, aggregate the integrated information ( s ˜ i j y , α ˜ i j y ) according to Eq (5) to obtain the comprehensive evaluation information of team j for task i in capacity y , denoted as ( s i j y , α i j y ). Next, the weight vector is converted into a crisp value g i j y .
Scheduling model for CPD
Matching degree between task and team
The matching degree refers to measure of fitness between elements. For example, when matching a project task with the collaborative team, if the matching degree is too low, then the collaborative team’s capabilities and resources are not adequate to allow them to complete the task. A higher matching degree ensures that the team can accomplish the tasks high-efficiency and high-quality, but it also means higher cost. To address this trade-off, this paper constructs a task-team matching degree model of collaborative product design project.
The task-team matching degree model is constructed in two ways: one is based on the personnel capability matching degree of collaborative team (the comprehensive capacity and core capacities in the capacity model), and the other is based on the available resources matching degree (the basic resources in the capacity model).
The matching degree between task i and team j at the dimension of personnel capabilities, denoted as TC ij , is defined as follows:
T C i j = ∑ p = 1 8 α i p ( 1 ± | ( g i j p ) 2 - ( e i p ) 2 | ( e i p ) 2 )
(7)
where p denotes the p th personnel capability, α i p is the weight of the p th personnel capability for task i , g i j p is the evaluation value of the p th personnel capability of team j for task i , and e i p is the required value of the p th personnel capability for task i . In Eq (7) , if g i j p > e i p , then take “+”; otherwise, take “-”.
Some available resources can be quantified, such as hardware and software. Thus, the matching degree calculation model between project task i and collaborative team j at the dimension of available resource, denoted as TR ij , is defined as follows:
T R i j = ∑ r = 1 4 β i r * g i j r e i r
(8)
where r denotes the r th resource, β i r is the weight of the r th resource for task i , g i j r is the available amount of the r th resource of team j for task i , and e i r is the required amount of the r th resource for task i .
Furthermore, the matching degree (MD ij ) between task i and team j is defined as:
M D i j = ω i 1 * T C i j + ω i 2 * T R i j = ω i 1 ( ∑ p = 1 8 α i p ( 1 ± | ( g i j p ) 2 - ( e i p ) 2 | ( e i p ) 2 ) ) + ω i 2 ( ∑ r = 1 4 β i r * g i j r e i r )
(9)
where w i 1 and w i 2 are the weights of the personnel capability and the available resource for task i , respectively.
Scheduling model
In a collaborative innovation project, through rational resource selection and configuration according to the project tasks’ requirement, optimal duration and cost are achieved.
Parameters:
PT: the project duration;
C : the project cost;
T : the set of project tasks, T = { T 1 , T 2 ,…, T m };
G : the set of collaborative teams, G = { G 1 , G 2 ,…, G n }, where n is the number of collaborative teams;
S = { s t 1 , s t 2 ,…, s ti …, s tm , s tm +1 }, where s ti denotes the start time of task i , and task m +1 is a virtual task;
MD ij : the matching degree between task i and team j ;
t Ni : the standard expected execution time of task i ;
Δ t i : the maximum shortened amplitude of execution time for task i ;
t ij : the expected time of collaborative team j to execute task i .
For collaborative product design, the shortened duration often leads to increased costs. Chen et al. [ 25 ] proposed a linear relationship between the activity time reduction and the cost increases to transfer the time-cost trade-off problem into a linear programming problem. Thus, the optimization objective is as follows:
min f ( x ) = a 1 * P T + a 2 * C = a 1 * S t m + 1 + a 2 * C
(10)
Constraints:
x i j = { 1 , team j complete task i 0 else
(11)
∑ j = 1 n x i j = 1
(12)
∑ i ∈ A t x i j = 1
(13)
e r m i n i ≤ e r i ≤ e r m a x i
(14)
S t q = max min ( S t i + t i j ) , T i ∈ B ( q )
(15)
t i j = { t N i MD ij , MD ij ≤ 1.0 Max { t N i MD ij , ( t N i − Δ t i ) } , MD ij > 1.0
(16)
In the objective function f ( x ), a 1 and a 2 are the weights of project duration and cost, respectively. Constraint ( 12 ) expresses the resource constraint. Constraint ( 13 ) ensures that task i is only performed by one collaborative team. Constraint ( 14 ) ensures that one collaborative team can only perform one task at a period, A t denotes the collection of tasks that are conducted at time t . Constraint ( 15 ) is the time constraint, and B (q) is the precedence activities setoff task q . Eq (16) is the time taken for collaborative team j to finish task i while considering the matching degree.
The improved GA
The issue proposed in this paper is a combinatorial optimization problem. However, it is different from traditional combinatorial optimization problems because the encoding cannot be repeated. A collaborative team can execute several tasks as long as the tasks do not overlap in one period. To solve the problem, the genetic algorithm is improved, where genetic operators are used to represent the individual of feasible solution in the encoding process. Single-coding in the solution space not only eliminates the decoding process between gene space and solution space but also can enhance the accuracy and reduce the complexity of computation process.
The steps of improved genetic algorithm are as follows:
Coding
Adopting decimal single coding, each gene locus represents the task code, and the number on the gene locus represents the corresponding matching collaborative team, as shown in Fig 2 .
Fitness function
The fitness function of GA is known as the evaluation function; it is used to determine the quality of individual. In this paper, the objective function is set as fitness function F( x ).
F ( x ) = f ( x )
Selecting the initial population
Randomly generate a certain number of individuals. Next, remove the repeated individuals and the individuals who do not meet the constraints, choose the best individual into the initial population and select a -1 individual from the remaining individuals randomly. These individuals compose an initial population with number of a . The probability ( p i ) that can be selected is set as follows:
p i = F i ∑ F i
(17)
Crossover operator
Multi-point crossover is adopted. In the process of evolution, if the current individual fitness is lower than the average fitness, then the individual evolution is invalid. To improve the search speed, it is necessary to improve individual crossover probability. Therefore, the adaptive crossover probability strategy is adopted. The crossover probability( p c ) is defined as
p c = { p c 1 − ( p c 1 − p c 2 ) ( F i − F a v ) ( F max − F a v ) F i ≥ F a v p c 1 F i < F a v
(18)
where F av and F max are the average fitness and the largest fitness, respectively.
Mutation operator
Execute mutation operation for each individual, the gene changes at a certain probability and varies from 1 to n ( n is the total number of collaborative team). In the process of mutation, single point mutation is used the first half of the individual, and multi-point mutation is adopted in the second part.
Selection operator
The previous generation population, the population after crossover and the population after mutation constitute the selection set. Remove the individuals of the population that do not meet the constraints. Next, the best individuals of the preceding generation population, crossover population and mutation population are retained. For the remaining individuals, two individuals are selected randomly, and one of them is chosen using the simulated annealing operator with probability exp (-Δc/θ) to bring into the next generation, and the other is taken back.
Repeat the above procedure until the number of the next generation reaches a , and then go to the next round.
Termination condition, output the optimal
10.1371/journal.pone.0200753.g002
Fig 2
Coding.
When meet one of the conditions, the iteration is stopped:
Fitness of the best individual and the group are no longer rising;
The number of iterations reaches the preset number.
The procedure of improved genetic algorithm is shown in Fig 3 .
10.1371/journal.pone.0200753.g003
Fig 3
The procedure of the improved genetic algorithm.
Case study
First, we conducted an experiment on our scheduling optimization algorithm of mobile phone collaborative product design. The relationship of design task is shown in Fig 4 . A total of 15 tasks were included in the project, and 20 collaborative teams were available.
10.1371/journal.pone.0200753.g004
Fig 4
Task relationship.
Standard time and the maximum shorten time of the tasks are shown in Table 1 .
10.1371/journal.pone.0200753.t001
Table 1 Standard execution time and the maximum shorten time of the design tasks.
T 1
T 2
T 3
T 4
T 5
T 6
T 7
T 8
T 9
T 10
T 11
T 12
T 13
T 14
T 15 (Days)
t Ni
4
5
5
6
30
30
25
7
5
15
5
1
1
1
4
Δ t i
1
2
3
2
2
3
2
1
2
3
2
0.5
0.2
0.5
2
The matching degrees between the collaborative teams and the tasks are shown in Tables 2 and 3 .
10.1371/journal.pone.0200753.t002
Table 2 Matching degree between collaborative teams (G 1 -G 10 ) and tasks(T 1 - T 15 ).
G 1
G 2
G 3
G 4
G 5
G 6
G 7
G 8
G 9
G 10
T 1
1.859
0.514
1.358
1.608
1.149
1.446
0.468
1.465
1.022
1.259
T 2
1.604
0.911
0.725
1.422
1.719
1.209
0.665
0.570
0.541
1.131
T 3
1.054
1.698
0.902
0.774
0.780
1.563
0.522
0.819
1.178
0.758
T 4
1.595
0.918
0.490
1.173
1.905
1.128
1.469
1.667
1.202
0.677
T 5
1.469
1.224
0.612
1.688
1.891
1.605
1.168
1.230
0.730
0.790
T 6
0.862
0.630
1.155
0.680
1.472
1.150
0.920
1.589
1.688
0.829
T 7
0.957
1.701
1.953
0.714
1.477
1.743
1.722
1.870
1.298
0.915
T 8
1.371
0.289
0.709
1.401
1.476
0.798
1.322
0.798
0.609
0.513
T 9
1.139
1.245
0.863
1.858
1.892
0.946
0.708
0.848
0.951
0.540
T 10
1.822
1.995
0.660
0.860
1.039
0.918
0.694
0.980
1.501
1.213
T 11
1.589
1.035
1.780
0.860
1.393
1.608
1.062
1.495
1.060
0.895
T 12
0.358
0.913
0.660
0.977
0.796
0.654
1.912
0.641
1.910
0.833
T 13
1.845
0.743
1.063
1.892
1.254
1.076
1.083
1.301
1.437
1.492
T 14
1.049
1.282
1.588
1.251
0.759
1.644
0.556
0.411
1.160
0.919
T 15
0.483
1.864
1.982
1.816
1.561
1.409
0.960
0.747
1.176
0.922
10.1371/journal.pone.0200753.t003
Table 3 Matching degree between collaborative teams (G 11 -G 20 ) and tasks(T 1 - T 15 ).
G 11
G 12
G 13
G 14
G 15
G 16
G 17
G 18
G 19
G 20
T 1
1.427
2.267
1.906
1.487
1.039
1.300
1.973
1.814
0.502
2.296
T 2
2.244
1.852
1.338
2.211
1.360
1.882
1.902
0.879
0.852
1.745
T 3
1.743
1.511
1.229
1.784
0.608
1.020
1.623
1.853
1.492
1.370
T 4
1.638
1.467
0.590
1.364
0.860
2.319
1.666
1.882
0.933
1.378
T 5
1.594
1.311
0.813
2.023
0.569
1.442
2.475
1.980
1.362
1.490
T 6
1.291
1.004
0.930
1.351
0.555
2.383
1.208
0.748
0.644
1.674
T 7
1.112
1.824
1.710
1.151
0.962
2.191
1.374
0.536
1.259
1.864
T 8
2.366
1.244
0.786
1.924
0.906
1.454
1.231
1.169
1.150
1.673
T 9
2.486
1.448
1.475
1.023
1.540
1.340
1.066
1.543
1.074
1.419
T 10
1.499
1.385
1.805
1.150
1.168
1.919
2.175
1.746
1.387
1.364
T 11
1.076
2.077
1.149
1.065
1.102
2.326
1.166
1.861
0.583
1.203
T 12
1.561
1.803
1.531
1.336
1.377
1.456
2.295
1.004
0.974
1.868
T 13
2.346
1.815
1.227
2.234
0.933
1.945
2.286
1.117
0.534
1.125
T 14
1.568
1.638
1.658
1.367
0.644
2.193
1.052
1.502
1.528
1.285
T 15
2.017
2.001
0.582
1.534
0.972
1.219
2.128
1.346
0.775
1.129
The task costs are listed in Tables 4 and 5 .
10.1371/journal.pone.0200753.t004
Table 4 The cost that the collaborative teams (G 1 - G 10 ) require to complete the task.
G 1
G 2
G 3
G 4
G 5
G 6
G 7
G 8
G 9
G 10 (10 4 )
T 1
6
8
7
8
7
10
7
7
8
8
T 2
8
6
7
6
9
9
6
5
8
7
T 3
6
7
7
5
7
6
7
5
6
6
T 4
8
9
10
9
11
11
10
8
10
9
T 5
18
19
17
17
20
18
16
23
21
17
T 6
23
26
27
27
24
25
22
20
25
24
T 7
15
18
16
16
18
17
18
19
16
18
T 8
12
10
10
13
13
11
14
15
14
12
T 9
6
7
8
7
7
5
7
9
9
7
T 10
18
17
16
21
19
18
20
19
18
17
T 11
5
4
5
4
6
7
5
5
7
6
T 12
2
3
4
5
2
3
5
6
4
3
T 13
7
4
5
7
4
4
5
7
6
5
T 14
3
5
3
4
5
6
5
6
5
4
T 15
10
7
10
9
8
9
8
12
9
9
10.1371/journal.pone.0200753.t005
Table 5 The cost that the collaborative teams (G 11 - G 20 ) require to complete the task.
G 11
G 12
G 13
G 14
G 15
G 16
G 17
G 18
G 19
G 20 (10 4 )
T 1
10
7
10
8
9
10
10
8
6
10
T 2
10
8
6
9
6
9
10
6
5
8
T 3
9
5
6
5
7
7
9
7
5
8
T 4
11
10
11
12
9
11
11
9
7
12
T 5
30
18
21
19
23
24
35
20
16
24
T 6
35
27
22
25
22
26
32
23
20
26
T 7
16
10
16
15
18
16
16
17
19
17
T 8
15
15
13
14
11
15
13
13
11
13
T 9
10
8
8
9
8
6
10
8
5
7
T 10
26
19
15
20
18
19
27
15
14
17
T 11
6
6
5
7
5
7
6
5
4
5
T 12
5
3
4
6
5
5
5
3
2
7
T 13
5
7
7
4
7
7
6
7
5
6
T 14
8
3
5
5
3
4
7
3
3
7
T 15
15
11
11
9
10
12
17
10
8
9
The parameter configurations of the improved GA were as follows: the initial population size was 20, P c 1 was 0.85, P c 2 was 0.65, the mutation probability was 0.9, the maximum number of iteration was 800, a 1 was 0.6, and a 2 was 0.4. Based on the data above, the procedures of the improved genetic algorithm were written in Matlab and run on a PC with an Intel Core 2.4 GHz CPU, 4GB RAM, the optimal programme is shown in Table 6 .
10.1371/journal.pone.0200753.t006
Table 6 Tasks—Team matching programme.
Task Number
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Collaborative team
1
19
14
19
7
8
12
3
19
19
2
5
5
12
2
Under this matching programme, the objective optimal value is 74.10, the duration is 45.7days and the cost is 1,180,000 RMB. The solution obtained by GA is {1 19 4 1 7 8 12 2 19 13 2 1 5 12 2}. The fitness curve of the improved GA and that of the traditional GA are shown in Fig 5 . The optimal was achieved at the 458th and the 622nd iteration by the improved and the traditional genetic algorithm, respectively. The result of the comparison revealed the advantage of the improved algorithm in finding the optimal and convergence speed, as shown in Table 7 .
10.1371/journal.pone.0200753.g005
Fig 5
Fitness curves of the improved GA and the GA.
10.1371/journal.pone.0200753.t007
Table 7 Comparison of theimproved GA and the GA.
Algorithm
Fitness
Run time(s)
Iteration
GA
75.45
32.6
622
Improved GA
74.65
20.4
458
The project task allocation and schedule plan is shown in Fig 6 .
10.1371/journal.pone.0200753.g006
Fig 6
Project task allocation and timing chart.
Conclusions
In this paper, a competence evaluation and a scheduling model of collaborative product design were studied based on matching degree. In the competence model, the collaborative team capacity is composed of core competency, basic competency and basic resource. Variable competencies or resources have different effects on the matching degree. The 2-tuple linguistic method was used to avoid information loss and make the evaluation result more precise. The scheduling model considering matching degree was established considering matching degree, project duration and cost. In the improved algorithm, single-coding strategy, multi-point mutation and crossover are adopted.
Although the case study demonstrated that the proposed approach is a useful tool to obtain the reasonable programme, there are still limitations in the approach, such as the subjectivity of evaluation and the precision of resource quantization. Furthermore, during the process of collaborative product design, there may be resource conflicts and partner selection conflicts. In the future, more work on the encouragement and collaboration mechanism for collaborative design should be performed.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Notice of Republication
This article was republished on April 18 2014, to correct the spelling error in the title. Please download this article again to view the correct version. The originally published, uncorrected article and the republished, corrected article are provided here for reference.
Supporting Information
File S1
Originally published, uncorrected article.
(PDF)
File S2
Republished corrected article.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
The Model for End Stage Liver Disease (MELD) has been used as the method of liver graft allocation since 2002. Liver grafts are distributed geographically based on MELD score in Organ Procurement and Transplant Network (OPTN) sharing areas consisting of local, regional and national tiers. It has been shown that patients with a MELD score ≥ 35 have wait list mortality rates that were similar to Status 1 candidates, who represent a cohort of candidates with acute liver failure most likely to die within 7 days without liver transplant [ 1 ]. To facilitate liver graft allocation to those patients with high wait list mortality, regional sharing for patients with a MELD ≥ 35 was implemented in June of 2013 with the goal of increasing life saving liver transplant for the sickest patients and decreasing death on the waiting list (“Share 35”).
Initial publications looking at the overall effects of “Share 35” demonstrated no reduction in wait list mortality in the 2 years following its implementation [ 2 , 3 ]. When a subanalysis was performed looking at patients with MELD ≥ 35, a reduction of 90 day waitlist mortality (66% versus 59%) was observed. While this reduction was heralded as an encouraging result, some authors have stressed that the effects of any policy change must be examined for all patients awaiting LT [ 3 ].
LT has been established as an effective treatment for patients with hepatocellular carcinoma (HCC) within specific size criteria (Milan criteria) since the seminal report by Mazzaferro and colleagues [ 4 ]. Patients with HCC within Milan criteria receive an “exception” MELD score, in an attempt reflect candidates expected wait list mortality due to tumor progression. Multiple previous studies have suggested that non-HCC candidates have significantly higher wait list drop-off rates than HCC patients due to mortality and clinical deterioration [ 5 – 7 ]. The primary goal of the present study was to assess the impact of the Share 35 policy on patients undergoing LT for HCC at the 2 years mark post-implementation.
Materials and Methods
After approval from the Mayo Clinic Institutional Review Board, data were obtained and extracted from the United Network of Organ Sharing (UNOS) Standard Analysis and Research file. The study population included all patients on the waitlist for LT in the United States from June 18, 2011 to June 18, 2015. Prior to Share 35 in the United States liver allografts were allocated to patients with the highest MELD score in sequential sharing areas consist of first local, then regional, and finally national tiers. The Share 35 policy was implemented in June 2013 to achieve broader sharing whereby the sickest waitlist candidates (patients with a MELD score ≥ 35) are first prioritized in a tiered manner regionally before any local candidates with MELD scores less than 35 are offered the livers. For the majority of analyses, data were provided for 2 eras; the 2 years pre-implementation of “Share 35” (Era 1) (6/18/2011 to 6/17/2013) and the 2 years post-implementation (Era 2) (6/18/2013 to 6/18/2015). Share 35 was implemented in June 2013 and therefore the dates were chosen so that we had 2 equal time periods with a minimum of 1 year of follow-up.
Donor and recipient factors were examined, including all the components of the liver donor risk index (DRI), donor sex, donor body mass index (BMI), recipient age, recipient BMI, recipient sex, recipient etiology of liver disease, biologic Model for End-Stage Liver Disease (MELD) score at transplant, match MELD score at transplant (match MELD is the score with which the organ was allocated; it also includes exception scores for cancers or other indications), presence of hepatocellular cancer (HCC) as a secondary diagnosis, re-transplantation, mechanical ventilation at the time of transplant and medical condition at the time of transplant [ 8 ]. Extended criteria grafts (ECD) were defined as a DRI > 1.7 [ 9 , 10 ].
Graft survival was calculated from the time of transplant until death, graft loss, or date of last follow-up. The occurrence and the date of death were obtained from data reported to the Scientific Registry of Transplant Recipients (SRTR) by transplant centers and were completed by data from the US Social Security Administration and from the Organ Procurement Transplant Network (OPTN).
Wait-list outcomes were analyzed with previously defined methods [ 2 , 11 ]. Briefly, removals for death as well as for “too sick” were treated as deaths. Patients that had tumor progression were included in the delisting for “too sick” definition. Patients’ wait list status was therefore classified into 1 of 3 categories: death, transplanted or still waiting/other. For the wait-list analysis a wash-out period was used. The pre-Share 35 era cohort listing dates were shortened to avoid overlap with the post-Share 35 cohort by 180 days. The Post-Share 35 cohort was also shortened in a similar manner, so that both eras were equal time intervals.
All statistical analyses were performed using STATA 12 (Stata Corp., College Station, TX). Differences between groups were analyzed using the unpaired t test for continuous variables and by the χ 2 test or continuity correction method for categorical variables. Wilcoxon rank-sum was used for variables that did not display a normal distribution. Survival curves for patient or graft survival were generated using the Kaplan-Meier method and compared by the log-rank test. All statistical tests were two-sided and differences were considered significant when p < 0.05.
Results
In Era 1 a total of 12,636 LT were performed of which 2916 (23.0%) were performed for HCC compared to 13533 LT of which 3029 (22.4%) were for HCC in Era 2. There was no difference in the proportion of LTs performed for HCC in the 2 eras (p = 0.18). No difference in the median match MELD score was seen for HCC patients between Era 1 and Era 2 (25 [range 6–40] vs. 25 [range 6–40]; p = 0.12). For HCC patients the median wait-time in era 1 was 185 days compared to 195 days in era 2. No significant difference in wait list time for HCC patients was seen between Era 1 and Era 2 in any of the eleven UNOS regions. Similarly, no difference in wait list time was seen for non-HCC patients in any of the eleven UNOS regions except for Region 1 where wait time increased (206 vs. 277 days; p = 0.02). When a sub-analysis was performed for all patients with a MELD ≥ 35, no difference in wait list time was seen for any of the eleven Regions. There was a slight statistical increase in the proportion of patients newly listed with HCC exception MELD between the 2 eras (14.3% vs.15.0%; p = 0.04); however this statistical difference is of minimal clinical relevance.
Data was analyzed to determine the number of cases in each Region in which organs were regionally shared for HCC exception MELD scores ≥ 35 (“Share 35”). In Era 1, prior to “share 35” policy implementation, there were no organs shared for this reason. In Era 2, no organs were regionally shared for HCC exception MELDs ≥ 35 in Regions 3, 4, 6, 7, 10, 11. In Regions 1, 2, 8, and 9 less than 2% or less of liver grafts were regionally shared for HCC exception MELD scores ≥ 35 (Region 1: n = 3 [2%]; Region 2: n = 2 [0.6%]; Region 8: n = 2 [0.9%]; Region 9: n = 2 [1.2%]). In Region 5, 33 [8.3%] liver grafts were regionally shared for exception MELD scores ≥ 35.
Competing risk analysis for patients’ wait-listed for HCC in the two eras can be seen in Fig 1 . Candidates in Era 2 were more likely to die while waiting (7.2% vs. 5.3%; p = 0.005) within 15 months, while there was no difference in the overall likelihood of getting transplanted (75% vs. 74%; p = 0.42). Change in the death rate for waitlisted HCC patients varied significantly between regions. Regions 4 and 5 had the largest increase (5% to 11%; p = 0.006 and 7% to 14%; p = 0.001, respectively) while Region 3 saw a reduction in waitlist mortality (5% to 2%; p = 0.06). Post-transplant graft survival by 12 months for patients with HCC did not change between the 2 eras (p = 0.51), nor did post-transplant patient survival (p = 0.21) ( Figs 2 and 3 ).
10.1371/journal.pone.0170673.g001
Fig 1
Wait list outcomes in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
Candidates in Era 2 were more likely to die while waiting (7.2% vs. 5.3%; p = 0.005) within 15 months, while there was no difference in the overall likelihood of getting transplanted (75% vs. 74%; p = 0.42).
10.1371/journal.pone.0170673.g002
Fig 2
Graft survival in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
10.1371/journal.pone.0170673.g003
Fig 3
Patient survival in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
Patients with a MELD score ≥ 35 were less likely to die while waiting in era 2 (28%) compared to era 1 (31%) (p<0.001) and were more likely to have been transplanted (42% vs 30%; p<0.001). For non-HCC patients with a MELD <35 there was no difference in mortality rate between era 1 and era 2 (10% vs 11%; P = 0.11), while there was a lower likelihood of getting transplanted (52% vs. 46%; p<0.001).
Mean donor DRI for patients transplanted for HCC increased from 1.41 to 1.44 (p < 0.001), while DRI decreased for those patients transplanted with a MELD ≥ 35 between Era 1 and Era 2 (1.44 vs. 1.40; p = 0.007). Patients transplanted for HCC received a higher proportion of ECD grafts (21.1% vs. 25.0%; p < 0.001), grafts from DCD donors (5.9% vs. 8.5%; p < 0.001) and grafts from PHS increased risk donors (13.3% vs. 21.8%;p < 0.001) between Era 1 and Era 2 ( Table 1 ). Recipient characteristics in patients undergoing LT for HCC in Era 1 and Era 2 can be seen in Table 2 . Patients transplanted with a MELD ≥ 35 had a decreased proportion of ECD organs (20.2% to 16.8%) between Era 1 and Era 2. Non-HCC patients with a MELD<35 had a slight non-significant trend of increased proportion of ECD grafts (22.6% vs 24.0%; p = 0.06) between the Eras. The distribution of graft DRI for HCC patients undergoing LT in the 2 eras can be seen in Fig 4 .
10.1371/journal.pone.0170673.g004
Fig 4
Distribution of donor risk index in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
10.1371/journal.pone.0170673.t001
Table 1 Donor characteristics in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
Pre-Share 35
Post-Share 35
p-value
N = 2916
N = 3029
DRI
1.41 ± 0.34
1.44 ± 0.36
<0.001
ECD (DRI > 1.7)
615 (21.1%)
757 (25.0%)
<0.001
Donor age ≥ 70 years
111 (3.8%)
145 (4.8%)
0.06
DCD
172 (5.9%)
257 (8.5%)
<0.001
PHS Increased Risk
388 (13.3%)
660 (21.8%)
<0.001
Share Type
Local
2420 (83%)
2393 (79%)
<0.001
Regional
437 (15%)
545 (18%)
0.002
National
58 (2%)
121 (4%)
<0.001
Regional Share 35
NA
42 (1.4%)
NA
DRI: Donor Risk Index; ECD: Extended Criteria Donor; DCD: Donation after Cardiac Death Donor; PHS: Public Health Service.
10.1371/journal.pone.0170673.t002
Table 2 Recipient characteristics in patients undergoing liver transplantation for hepatocellular carcinoma in Era 1 (Pre-Share 35) and Era 2 (Post-Share 35).
Pre-Share 35
Post-Share 35
Recipient Characteristics
N = 2916
N = 3029
p value
Age at transplant (years)
58.8 ± 7.6
60.0± 7.3
<0.001
Body mass index
28.6± 5.3
28.6± 5.2
>0.99
Gender (male)
2227 (76%)
2333 (77%)
0.55
Diagnosis
Hepatitis C virus serology
1875 (64%)
1877 (62%)
0.06
EtOH
52 (1.6%)
56 (1.9%)
0.85
NASH
46 (1.6%)
67 (2.2%)
0.07
Calculated MELD score †
12 (6–48)
11 (6–50)
0.09 ††
Match MELD score †
25 (6–40)
25 (6–40)
0.12 ††
Race/ethnicity
White
1977 (68%%)
2063 (68%%)
0.80
Black
282 (9.7%)
317 (10%)
0.31
other
657 (23%)
649 (21%)
0.30
†median (range)
††Wilcoxon rank-sum used.
Discussion
Broader regional sharing though the “Share 35” policy change was implemented with the goal of increasing life saving LT for the sickest patients on the wait list. Initial publications examining the effects of “Share 35” have shown several positive results, including reduction in 90-day mortality for patients with MELD scores ≥ 35 [ 2 , 12 ]. While these initial results are encouraging, it is important to fully explore the effects of “Share 35” for all patient groups.
The present study specifically examined the effects of “Share 35” on patients who underwent LT for HCC by comparing the 2 years pre-implementation of “Share 35” with the 2 years post-implementation. The proportion of LT performed for patients with HCC did not change following implementation of “Share 35” nor did the waiting time. Despite these findings, a higher rate of death/delisting for “too sick” for patients on the wait list for HCC in the post “Share 35” era (5.3% vs. 7.2% at 15 months) was observed. Given the relatively stable number of available liver grafts, this finding highlights the reality that higher transplant rates for one cohort of patients inevitably results in less organ availability for another. The present study did demonstrate a slight increase in the proportion of patients newly listed with HCC MELD exception score between the 2 periods (14.3% vs.15.0%). This increase in listed HCC patients, associated with no increase in the proportion of LT performed for patients with HCC, likely accounts for the higher death rate for HCC patients on the wait list. It must, however, be acknowledged that this rate of death is still significantly lower than the death rate for patients with high biologic MELD scores and perhaps is an unavoidable consequence of attempts to reduce the disparity in waitlist mortality for HCC and non-HCC patients awaiting LT [ 13 ]. It will also be important to follow results of wait listed patients with HCC in light of the recently implementation changes to HCC MELD exception points nationally. Modelling that was performed prior to the policy change that resulted in a 6 month delay in receiving HCC exception points, was designed to increase dropout (death) for patients with HCC on the waiting list. These models demonstrated an increase in wait list death for patients with HCC and a decrease in wait list death for non-HCC patients [ 14 ]. Proponents of the new HCC exception point policies would argue that HCC patients who died/dropped off the wait list likely had more advanced tumor biology and would have had an inferior result if they had been transplanted [ 15 ]. While this may be the case with some patients it should also be noted that modelling demonstrated that with longer delays, the biological MELD for HCC patients at transplant also increased [ 14 ]. It is conceivable that some patients with HCC that died while waiting may also have died as a result of their underlying liver disease.
The newly implemented modifications to the MELD exception points to patients with HCC also created a cap to the exception score of 34, to avoid the allocation of regionally “Share 35” organs to patients with exception MELD scores ≥ 35. The present study demonstrated that almost no patients received regionally shared liver grafts for MELD exception scores ≥ 35 in any of the eleven UNOS regions except Region 5. In Region 5, 8.3% of LT for HCC received liver grafts as a result of “Share 35”. With capping of MELD exception at 34 for patients with HCC, for regions with high allocation MELD scores, it will become even more imperative to attempt to increase the usage of higher DRI organs such as DCD grafts for patients with HCC.
Another important finding of the present study was that between the 2 eras studied, there was a change in the liver graft quality that HCC patients received. An increase the usage of ECD and DCD livers was observed for patients with HCC undergoing LT while a decrease was observed for patients with a MELD score ≥ 35. This finding suggests a perhaps expected shift of the higher quality liver grafts allocated to patients with the highest MELD scores. In order to maintain stable wait times for patients with HCC, in the post-implementation of “Share 35” era, transplant programs have adjusted their behavior by using more ECD and DCD organs for these patients. Increased death rates for patients on the wait list for HCC may also have played a role in the acceptance of more ECD and DCD grafts for this cohort. A previous study examining the effects of “Share 35” at 1 year post implementation demonstrated liver grafts with higher DRI were more likely to be shared within a region in pre-Share 35 era, while in post-Share35 era, liver grafts with lower DRI were more likely to be shared [ 12 ].
Previous studies examining the effects of “Share 35” have demonstrated no overall change in the DRI or CIT of liver grafts used for LT [ 2 ]. We also did not find any overall increase in DRI when all transplants were examined. This suggests “Share 35” did not increase the total number of higher DRI organs being used, but instead simply resulted in a modification in which recipients these organs are being used for. It should be noted, that despite no substantial change in DRI, initial publications on “Share 35” did show a small decrease in overall liver graft discard rates [ 2 ]. This may suggest that by optimizing donor-recipient selection we can perhaps more effectively utilize higher DRI liver grafts while still maintaining good post-transplant outcomes. Patients with HCC who have relatively preserved synthetic function may represent ideal candidates for DCD or other higher DRI organs (ECD and DCD) as they can more safely “weather the storm” of mild delayed graft function [ 9 ]. Indeed, previous authors have demonstrated excellent results for patients with HCC receiving DCD grafts [ 16 ]. Despite the use of a higher proportion of ECD and DCD organs for patients with HCC, the present study did not show any change in graft or patient survival by 12 months for these patients when comparing the 2 eras. This may suggest that as we continue to search for new ways to increase the available liver donor pool, patients with HCC may represent a suitable group to use higher DRI organs. The increased wait list death rate for this population may also change dogmatic thinking that HCC patients have time to wait for a better organ.
Limitations of the present study include its reliance on registry data and lack of granularity that can be obtained at single program level. In addition, due to the fact that data was only available for 2 years following implementation of “Share 35”, many outcomes were truncated at shorter follow-up intervals.
In conclusion, the present study demonstrates that there has been no significant change to wait time for patients listed for HCC following implementation of “Share 35”. Transplant program behavior has changed resulting use of a higher proportion of ECD and DCD liver grafts for patients with HCC, while lower DRI liver grafts are preferentially allocated to patients with higher MELD scores through regional “Share 35”. Despite these changes, no significant difference in post-transplant graft or patient survival has been observed. A higher rate of waitlist mortality was observed in patients with HCC in the post-“Share 35” era, however, wait list mortality rate still remains substantially lower than that observed for non-HCC patients.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku:
|
Introduction
Recent studies (e.g. [ 1 ]) indicate that the general public is increasingly using a combination of online platforms (e.g. news websites/apps and social media) and traditional news avenues (e.g. local and cable TV, radio, and print newspapers) to get its news. This results in a novel news ecosystem, supporting highly and actively engaged news consumers [ 2 – 4 ]. Within this news ecosystem, cyberspace is complementing traditional sources to communicate news [ 5 – 7 ], raise public awareness [ 8 , 9 ], and form and shape opinions and agendas [ 10 – 12 ] across a diverse range of issues.
This highly participatory cyber-physical news awareness ecosystem is fostering digital activism. Its manifestations range from encouraging and enabling the crowd to generate, publish, and disseminate focused information, to fostering the formation of online communities that jointly pursue activities in the cyberspace, and often the physical space too [ 13 ]. We are repeatedly witnessing this in the use of social media to support political activism in the streets of cities across the globe [ 14 – 19 ], the establishment of peer-to-peer information exchange communities [ 20 , 21 ], or even the formation of online communities to support cyberspace projects like Wikipedia [ 22 ]. While this powerful cycle linking information consumption with the raising of awareness and the resulting digital activism will continue to be essential to the future of our society, the mechanisms that drive this process still remain fuzzy.
One popular expression of such digital activism is in the form of volunteered contributions to online collaborative activities. In order to better understand the motivating factors that drive such activities, certain studies have focused on individual platforms separately. Nov [ 23 ] built on earlier work of Clary [ 24 ] that had identified the major motivational functions that drive volunteerism overall (e.g. altruistic values, social interaction, enhancing one’s understanding and career-related benefits) to show that these general motivations apply to Wikipedia contributions as well. Xu and Li [ 22 ] classified these functions under two broad categories, namely the will to contribute content, and the interest in participating in that online community. Budhathoki and Haythornthwaite [ 25 ] examined the motivational factors that drive contributions on OpenStreetMap (hereafter referred to as OSM in this paper), a freely accessible and editable map of the world [ 26 – 28 ]. That study reported important positive factors relating to personal, yet the shared need to contribute to open source projects, the co-integration of individuals into open-source and geographic knowledge communities, along with the need to be attentive to participation taking places within the OSM community. It is also important to note that the motivation for contributors within these different communities can vary [ 29 , 30 ].
The above studies have been crucial in providing a better understanding of the various factors that drive people to become digitally active in online collaborative communities. However, relatively few studies have examined digital activism in the context of the previously described broader news awareness ecosystem. Zastro [ 31 ], for example, noted that during the Haitian earthquake in 2010, volunteer mappers used a combination of different news media (i.e. news reports, social media and text messages from survivors) to collect information on the damaged status of buildings. This information was then added to the OSM platform. Westrope et al. [ 32 ] examined the possible impact of media coverage on mapping activities following Typhoon Haiyan in the Philippines in 2013. That study showed that in areas that received substantial media coverage (e.g. Tacloban City) 92% of the surveyed buildings had over-represented damage reported in the OSM edits. In contrast, in areas that received less media coverage only 56% to 76% of the buildings had over represented damage reported in the OSM edits. As the authors of that study suggest, the large over-representation in damage buildings in Tacloban City could be attributed, at least in part, to the priority given to this city by the Humanitarian OpenStreetMap Team (HOT) OSM community. In a related study, Dittus et al., [ 33 ], compared OSM activity for various Humanitarian HOT events, showing that the two most publicized media events, Typhoon Haiyan, and the Nepal earthquake in 2015, had the largest newcomer mapping recruitment rates. The authors of that study suggested that news coverage of these events may have assisted in attracting a much larger number of newcomers. Similar findings of the influence of media coverage and an increase in newcomer enrollment in the OSM platform were reported by Begin et al., [ 34 ]. While these studies shed some light on the possible relation between OSM activity and news media coverage, a more in-depth analysis of this relation is still needed.
The pursuit of a deeper understanding of the relationship between the different components of the complex news awareness ecosystem and digital activism is becoming a substantial research challenge. Wikipedia edit patterns, for example, have been compared to breaking news [ 35 ], and news media coverage of an issue to increases in searches for and edits of related articles in Wikipedia [ 36 ]. Such patterns have also been used to define entity-specific news tickers and timelines [ 37 ], with page views further used to detect popular topics related to users’ interest [ 38 ]. Google Trends [ 39 ], a tool for measuring public agenda [ 40 , 41 ], has also been found to be useful for quantifying trading behavior in financial markets [ 42 , 43 ], for disease surveillance and health care research [ 44 , 45 ], and as a tool for behavior analytics, such as predicting non-cigarette tobacco use [ 46 ]. Moreover, Ratkiewicz et al., [ 47 ] show that bursts of online search volume activity extracted from Google Trends tend to be correlated to similar bursts of editing activity on Wikipedia. Such studies highlight just how complex this news awareness ecosystem is.
In the context of this paper we use the term news awareness ecosystem to refer to the ensemble of sources, ranging from traditional newsrooms to grassroots citizen journalism, that provides the public with possible exposure to news, and from which public awareness can emerge. A key characteristic of this ecosystem is that rather than being monolithic it is multifaceted, multimodal, and evolving. This term builds on the idea of the “news ecosystem” that has become prevalent in recent years. While an agreed upon definition of this term is still lacking, several attempts towards the construction of such a definition have been made. For instance, Anderson [ 48 ] provides a brief genealogy of the term and its origins, and discusses several concepts that may support the construction of such a definition. Alternatively, Picard [ 49 ] provides a detailed discussion of the forces that reshaped news journalism into what is referred to as the new (news) ecosystem.
Taking a more non-contemporaneous view of the news cycle, Nghiem et al., [ 50 ] showed that, depending on the specific topic, news coverage can be responsive to or lag an occurring trend (as defined by Google Trends search volume data). In some cases, albeit very few, a build-up of news activity was also followed by a decrease in search volume. Such responsive behavior of news patterns has been documented for trending topics in social media as well [ 51 , 52 ]. Althoff et al., [ 53 ] compared trending data from Google (Search, News and Trends), Twitter and Wikipedia to show that combined, these different data platforms can be used to forecast trending topics, even though temporal activity patterns (comprising spikes and cumulative build-up) tend to vary across these platforms. Other work by Al Emadi et al., [ 54 ] showed the strong response of the online mapping communities in the aftermath of natural disasters, e.g. in the form of increased online volunteering activity in the MicroMapper [ 55 ] crowdsourcing platform.
Digital activism and geography have long been intertwined, giving rise to a range of online user activities, which in the context of this paper we refer to as geo-activism (in the remainder of the paper we will use the term digital- and geo-activism interchangeably). From crowdsourcing crisis information mapping, to support disaster relief and recovery efforts [ 56 , 57 ] and collectively editing geographical data in OSM [ 27 ], volunteered geographical information (VGI) has emerged as a substantial new mechanism for the general public to contribute geographical content by mapping roads, buildings, and other artifacts [ 58 – 60 ]. At the same time, it is important to recognize that the general public is also increasingly engaged in geo-activism while being exposed to the news media. As suggested in previous studies [ 61 – 63 ], many forms of online news media have the power to shape public opinion, affect other media sources, and foster engagement.
Consequently, a key premise of our work is that geo-activism should be considered in the context of the news media rather than as a separate, independent phenomenon, in order to advance our understanding of the mechanisms that drive it. This would allow us, for example, to explore how the ebb and flow of information in the public media sphere affects digital geo-activism. At a time where VGI is establishing itself as a rich supplementary—and sometimes only—source of geospatial information [ 64 , 65 ], a better understanding of the mechanisms driving geo-activism will allow us to further harness its power.
Accordingly, our objective in this paper is to study links between media coverage and geo-activism by focusing on the question of geographical saliency within the news awareness cycle. The basic argument is that news features of broad community interest but with a certain, narrow geographical footprint build up over time awareness and interest for that location, leading to bursts of volunteered geographical contributions for it. We pursue this task by comparing patterns of online volunteered geographical edits in OSM to relevant news stories, and to corresponding edit patterns in Wikipedia. Therefore, the research question at the core of this paper is whether extreme instances of public awareness deficit or surplus (resulting from sustained relevant trends) tend to trigger digital geo-activism bursts.
Towards this goal, we present in this paper a case study that focuses on several refugee camps around the world to determine whether salient features in news media affect digital activism in OSM. This line of inquiry allows us to examine the possible association and interconnectedness between macro level global awareness and news coverage and digital activism, with a hyperlocal geographical focus.
A number of factors render refugee camps particularly suitable for such a study. First, they match very well the above-mentioned motivational factors that drive digital activism as they were identified by Nov [ 23 ]. Second, compared to previously studied events like natural disasters, refugee camps differ in the sense that they are not abrupt events that occur over a period of few hours (e.g. earthquake) or days (e.g. flooding), which would render the relationship between media coverage and digital activism trivial, as event, news coverage, and response practically coincide. Instead, these refugee camps are set up and operate over a period of years [ 60 ] and public awareness to them builds over time. Third, they offer the advantage of having a geographically distinct footprint in terms of size and location relevant to their surroundings while being associated with a narrow thematic focus at the same time (unlike large cities, which may have numerous themes occurring at the same time). Finally, in the aftermath of massive displacement of populations due to civil war or conflict (e.g. the Syrian crisis that began in 2011 and other on-going ethnic tensions), refugee camps represent a topic that is becoming increasingly important in terms of media coverage and public awareness [ 66 ].
The remainder of this paper is organized as follows. In Section 2, the datasets and methods used to address our research question are presented followed by an in-depth discussion of the results in Section 3. Finally, Section 4 provides a discussion of our results and concludes with an outlook of future work.
Materials and methods
Refugee camps
The current concentration of refugees is highest in African and Middle Eastern countries compared to other world regions [ 67 ]. These regions also contain the most populous camps as well [ 68 ]. Accordingly, a set of 8 refugee camps was selected for our study, located in Africa, Middle East, and Europe, as shown in Fig 1 . In addition to reflecting the global distribution of refugees, these camps offer diversity in terms of their size, population, and date of establishment. Camp age (date of establishment) is of particular interest, as it allows us to study camps at different stages of news cycles and interest.
10.1371/journal.pone.0206825.g001
Fig 1
Study areas (centroid location of camp).
Satellite image courtesy of the DigitalGlobe Foundation.
Table 1 shows more detailed information for each camp including their established date and camp status, their size (area), population, and population density (See S1 File for references). As expected, larger camps tend to host more refugees, and the correlation between camp size and population has a Pearson r value of 0.70 with a p-value of 0.03 (with Kakuma being the one exception). However, in general, there is no clear observable association between the age of the camps and their population density or size.
10.1371/journal.pone.0206825.t001
Table 1 Refugee camps.
Camp
Country
Established Date/Status
Area (km 2 )
Population
Population Density (People per km 2 )
Dadaab
Kenya
1992 with an addition in 2011 / Active
40.7
242,998
5,970
Kakuma
Kenya
1991/Active
6.4
171,085
26,732
Nyarugusu
Tanzania
1996/Active
25.9
78,519
3,032
Calais
France
2015/Closed on October 2016
0.61
9,000
14,754
Yida
South Sudan
2011/Active
20.4
55,012
2,697
Bidibidi
Uganda
2016/Active
360.9
270,000
748
Oncupinar
Turkey
2012/Active
0.71
15,000
21,127
Zaatari
Jordan
2012/Active
6.1
79,827
13,086
Data sources
To analyze patterns of geo-activism as it relates to these refugee camps, we used the patterns of relevant information contributions in OSM. OSM is a prototypical example of a VGI platform allowing anyone to map features on the Earth’s surface. It was launched in July of 2004 and presently, as of August 2018, has almost 5 million registered users [ 69 ]. While being general in its scope, OSM has grown over time to be a particularly rich source of geographical data, and especially so in support of humanitarian response activities. Some recent examples include OSM contributions relevant to the Ebola outbreak in West Africa in 2014 [ 70 ], the Nepal earthquake in 2015 [ 71 ], and slums in Sub-Saharan Africa [ 65 ].
As OSM is a collaborative platform of information contribution, it tends to exhibit activity patterns similar to those found in other non-geographical platforms. For example, a common thread in many online collaborative communities is that the level of participation is not even: few users tend to contribute massive amounts of information, whereas the large majority tends to contribute less. Previous studies have confirmed that this pattern is also present in the OSM community [ 72 , 73 ]. Similar patterns of contributor activity have also been found with other online crowdsourcing communities such as Wikipedia [ 74 – 77 ], which are often characterized by bursts of activity [ 78 , 79 ]. When it comes to OSM, the spatial and temporal frequency of contributions can vary based on numerous factors, including, the diversity of contributors [ 27 , 80 ], issues associated with the digital divide [ 81 – 83 ], the social structure of contributor communities [ 84 – 85 ], the direct intervention of social groups such as mapping parties [ 86 , 87 ], and bulk imports into the OSM platform [ 88 ], among others.
For the purpose of our study, we used OSM data that were extracted from the planet history file ( https://planet.openstreetmap.org/planet/full-history ), which contains all OSM edits for the entire world. An OSM edit herein is defined as any create, delete or modify operation to any OSM node, way or relation feature. The OSM planet history file also contains other useful information, such as the location (latitude and longitude) of each node, object versions (from October 2007 onwards), contributor user ID, and timestamp. Refugee camp OSM data were clipped from the history file using spatial and temporal parameters. In order to delineate the footprint of each camp we use polygons that demarcated their spatial extent, and use these polygons to select the corresponding edits. Camp polygons were manually traced using high resolution satellite imagery over each camp. In terms of time window, we selected edits during the current decade (01/01/2010 to 05/31/2017). From among all our camps, only Bidibidi had edits (only 2) prior to our study window, and so in practice our study addresses the whole OSM history of these camps.
In conjunction with the collection of OSM edit activity, edit activity was collected for the Wikipedia page entries that correspond to each of the refugee camps in Table 1 . A Wikipedia edit in the context of this research refers to any change (i.e. creation, modification or deletion operation) made to the content of a Wikipedia page. In our study the edit activity of each page was collected from the history page of each camp’s Wikipedia page using the same time window parameters noted above for OSM. Only English pages were used in this study, a caveat that will we will later revisit in the Discussion section. Wikipedia is the world’s largest free online encyclopedia, allowing anyone to create new and edit existing articles. As of August 2018, more than 34 million users were registered with almost 46 million pages, and with the number of entries now approaching 6 million [ 89 ]. A key motivation to explore the edit activity of Wikipedia alongside the edit activity of OSM is that while both platforms rely on digital activism through crowdsourcing efforts, the former focuses on digital activism in a more general sense while the latter focuses specifically on geo-activism. This difference will enable our analysis to compare and contrast digital activism patterns in two substantially different crowdsourcing environments.
In order to analyze relevant news media activity to OSM edit activities we also collected relevant data from Google News and Google Trends. While Google News conveys expressions of media coverage, Google Trends is considered here as a proxy for public interest. News coverage data were captured from Google News [ 90 ], an online news aggregator for up-to-date news stories from all over the world. As of 2012, Google stated that it draws from more than 50,000 news sources with more than one billion unique users connected each week to its news content [ 91 ]. Prior research has also compared news extracted from Google with other platforms such as LexisNexis, showing that Google News provided broader worldwide coverage than its counterparts [ 92 ]. Google News articles were searched using the keywords “ name of camp ” AND “ refugee ” to help filter non-relevant stories. Articles were further filtered manually by the authors as needed, to remove irrelevant data. Between 6% and 26% (average of 13%) of articles were removed during the data cleaning process. Filtering was initially done by one of the authors who collected the data, and later on two other authors assisted to further remove articles that were deemed non-relevant. The remaining relevant news articles were then binned into daily counts.
To capture information on public interest, we used Google Trends data. Google Trends provides information on how often a particular keyword was searched using Google’s web search engine. This data has also been used in previous studies as a proxy measure for public interest in a variety of topics [ 40 , 50 , 93 , 94 ]. In order to access Google Trends data for each camp, we used the corresponding online portal and the camp’s name as the keyword, to capture searched volumes for each camp. This data provided by Google, is available as a normalized series of values between 0 and 100, based on the popularity of users’ search interest within the specified temporal range. A summary of the various steps used to extract and preprocess the different data sources (OSM, Wikipedia, Google News, and Google Trends) are presented in Fig 2 . Further, in S2 File , we show the total number of OSM edits, Wikipedia edits and Google News articles at the daily level for the period 01/01/2010 to 05/31/2017. With respect to Google Trends, data at the daily level is limited to a temporal search window within 269 days. Following this, the data is provided at weekly and then the monthly aggregated level depending on the temporal search window. As such, the Google Trends data is not included in S2 File . An initial visual comparison of the data at the daily level did not reveal any apparent patterns across the different data sources.
10.1371/journal.pone.0206825.g002
Fig 2
Data extraction and preprocessing.
In the data extraction step, data on OSM, Wikipedia, Google News, and Google Trends is collected from their various online platforms. The multiscale analysis step involves a progressive refinement approach: Data is first examined at the monthly level and the strongest extremum points on the awareness curve are identified. These points are then assigned a score that is comprised of its magnitude and a weight that is proportional to the duration of consistent trend along the awareness curve prior to the detected extremum point. A search time window is then defined on the awareness peak with the highest score and used for working with data at the weekly analysis level. A similar approach is used to identify a search window for working with data at the next (finer) temporal granular analysis level.
Overview of multiscale and multiplatform comparison approach
Given the objective of this research, our approach is based on using four basic measures to explore the possible associations between news media coverage and crowdsourced activity: Google Trends index, Google News volume, OSM edits activity, and Wikipedia edits activity. While the first two measures serve as a proxy of the general public’s interest and availability of news around specific themes, the latter two measures serve as a proxy of the level of digital activism around specific thematic features over space and time. A key premise of our approach is that comparing and contrasting these measures can provide additional insight into the possible relationship between the ebb and flow of news and activity patterns in digital activism in platforms such as OSM and Wikipedia.
While gauging the levels of digital activism is relatively straightforward (e.g. tracking the number of edits in Wikipedia), measuring the ebb and flow of news in the media ecosystem is more challenging as there is no such single readily available measure. To overcome this issue, we propose to consider the difference between the normalized Google Trends index and the normalized Google News metric as a measure of the overall public awareness with respect to a specific theme (or a set of themes). Here, we refer to the term awareness simply as the “knowledge that something exists, or understanding of a situation or subject at the present time based on information or experience” [ 95 ]. The reasoning behind using this difference measure is that while Google News represents the availability of news (i.e. information) regarding a theme, Google trends represents the public’s active pursuit of information about the theme. As news stories emerge, evolve, re-emerge, and eventually subside in the media, the public’s information seeking activities may increase, decrease, or remain unchanged over time [ 50 ]. Consequently, a surplus (or deficit) of public awareness can be built-up over time through the availability (or unavailability) of news stories that consistently increase (or reduce) the public’s online information seeking activities. A prolonged period of awareness surplus growth would be one where the normalized metrics of Google Trends grow faster than the corresponding metrics of Google News, implying that public interest on the topic grows faster than news coverage. Conversely, a prolonged period of awareness deficit would be one where news coverage outpaces public search interest. In this context of this paper, a min-max approach was used to rescale all four data metrics to values between 0 and 1.
The ability to track trends in public awareness overall enables us to explore the potential associations between such trends—and in particular trend changes—in awareness and digital activism. Specifically, our interest lies in the question of how are trend changes in public awareness related to user activity in OSM and Wikipedia ? The emerging argument from this question is that media coverage drives participation, by informing the general public of evolving/developing situations, thus setting an agenda and leading to digital activism. Since such a mechanism can occur at different time scales, our approach is based on consecutively examining the different measures at 3 levels of time granularity, namely monthly, weekly and daily, in order to identify possible associations at the finest temporal granularity considered here. Starting from the monthly level, extremum points in public awareness activity are detected, and the strongest extremum point is identified. The time stamp of this point is then used together with a search time window around it to define a new search interval at the next (weekly) granularity level. This process is then repeated for identifying a search time window at the daily time granularity, in which possible relations between public awareness trend changes and OSM and Wikipedia activity are explored. A more detailed description of the extremum points detection and selection process is provided below.
Our analysis process begins by examining the public awareness curve at the monthly level and then proceeds using a progressive refinement mechanism towards an examination of data at the daily level. In order to capture broad trends in public awareness level, we model the public awareness data using multivariate adaptive regression splines (MARS). MARS is particularly suitable for this type of data since it makes no assumptions about the underlying distribution of the data and provides an intuitive approach for understanding the intrinsic complicated data mapping in high-dimensional data patterns [ 96 ]. MARS is represented as a combination of basis functions expressed as [ 97 ]:
y = β 0 + ∑ m = 1 M β m h m ( X )
(1)
where y is the dependent variable (i.e. our public awareness measure), X is the independent variable (time), β 0 is the intercept parameter, and β m is the coefficient applied to each basis function h m (X) , which are summed over M non-constant terms used for defining the number of basis functions. In this case, M is determined in a data-driven manner based on the optimum number of basis functions required to fit the data. While MARS supports polynomial basis functions, in our implementation linear basis functions are used for simplicity.
Using the derived MARS model, extremum points in the public awareness curve are detected by estimating the second derivative f ″( x ) as a central finite difference approximation:
f ″ ( x ) ≈ f ( x i + Δ x ) - 2 f ( x i ) + f ( x i - Δ x ) Δ x 2
(2)
where x i represents the i th value of x in the data series defined by the predicted values in Eq 1 , Δx is the step size in temporal units (i.e. monthly, weekly or daily dependent on the scale of analysis) around point x i , where i = 0,1,2,…, n . At each temporal granularity level Δx is set to 1 time unit, e.g. at the monthly level a Δx value of 1 month is used.
As the detection of extremum points along the public awareness curve may result in multiple local maxima or minima points, a pruning process is applied in order to capture the most prominent extremum points. This process begins by identifying all extremum points and then for each point suppressing other weaker extremum points (in terms of their absolute magnitude) within a local window centered around each top extremum point. In our analysis this was set to ±1.5 months, ±1 week, and 0 for the monthly, weekly and daily granularity, respectively. In the case of Google Trends, this data was first captured at weekly granularity to identify local windows around each extremum point at the monthly analysis level. Following this, at the next level of analysis (i.e. weekly), the Google Trends data was captured at daily granularity. No local windows were used at the daily analysis level.
Following the pruning process, each of the remaining extremum points (in terms of their absolute magnitude) is assigned a score that is comprised of its magnitude and a weight that is proportional to the duration of consistent trend (positive or negative) along the awareness curve prior to the detected extremum point. This weighting scheme is guided by previous work [ 98 ], which suggested that the longer a theme is circulated within the media, the greater its potential for influencing the public agenda (which may result in digital activism). Based on these calculated scores, the awareness peak with the highest score (strongest extremum point) is selected, and is then used for defining a search time interval ΔT at the next (finer) temporal granularity level.
The selection of ΔT in our analysis is guided by prior research on agenda setting theory in communication [ 99 ], in which the question of the time that it takes for the public to respond to news stories was explored. Such prior work suggested that it may take as little as a few days [ 100 ] to as much as 2 to 6 months [ 98 , 101 ] for changes in the media agenda to become fully realized into public agenda. In other studies, e.g. [ 93 ], a period of 50–70 days was suggested. Informed by these prior studies, our analysis is based on repeating the search for the for the highest scoring awareness peak using a progressively refined ΔT . Specifically, ΔT was set to ±12 months and ±4 months around each highest scoring awareness peak in the monthly and weekly time granularity, respectively.
As our objective relates to the comparison of the awareness curve to the two crowdsourcing platforms, it is also necessary to detect extremum points in the OSM and Wikipedia edit activity data. For convenience, and due to the overall stepwise nature of this data, we consider the cumulative edit activity time series for these two platforms and approximate each edit activity curve using the Douglas-Peucker line simplification algorithm [ 102 ]. Then, using the simplified edit activity curves, we detect the top 5 extremum points ranked by their overall magnitude. In some cases, fewer than 5 significant peaks were extracted as a result of the line simplification process.
Results
Given the objectives of this paper and the proposed analysis approach, this section summarizes the results that were derived along three main themes. First, we outline the key trends in the edit activity for the eight camp sites. Following this, we present the results of an analysis of the edit activity patterns in the local context of each camp site. Finally, we examine how trends in public awareness relate to OSM and Wikipedia edit activity patterns.
Edit activity trends in OSM
The overall number of OSM edits for all 8 refugee camps for the period 01/01/2010 to 05/31/2017 is shown in Table 2 . Comparing these numbers with the data in Table 1 reveals a strong correlation between camp size (spatial extent) and number of OSM edits, with a Pearson r correlation of 0.89 and a p-value of 0.002.
10.1371/journal.pone.0206825.t002
Table 2 Total OSM edits per camp for the study period (01/01/2010–05/31/2017).
Camps
Total OSM Edits
Dadaab
31,283
Kakuma
11,147
Nyarugusu
74,416
Calais
5,965
Yida
84,053
Bidibidi
198,916
Oncupinar
708
Zaatari
69,981
In order to delve into the driving forces behind public participation in OSM we further look at the temporal variations of these contributions. Fig 3 shows the patterns of temporal activity in OSM edits for the 8 camps selected. The data in Fig 3 has been normalized to values between 0 and 1 to account for different scales of edit activity between camps.
10.1371/journal.pone.0206825.g003
Fig 3
OSM edit patterns for the 8 camps studied (01/01/2010–05/31/2017).
Fig 3 shows that edit activity in OSM tend to occur in bursts (spikes in the charts), similar to the what was observed in Wikipedia [ 47 ]. This observation is consistent with Barabasi [ 103 ], who suggested that such an activity pattern tend to follow a non-Poisson pattern, which is common in human activities in general. Following this theory, edit activity bursts could then be seen as following a natural pattern of contributors’ interest (as a measure of priority) in mapping refugee camps. Furthermore, as such OSM edit activity bursts represent epochs in time in which users’ activity was focused on specific localized camp sites, one could argue that these bursts represent instances when these sites become salient features/attention landmarks [ 100 ] in geographic space. A more detailed discussion of this saliency property is provided in the next section.
In addition, it is interesting to observe that the OSM edit activity bursts in Fig 3 do not tend to coincide in time across camps. This is expected as it is rather unlikely that users that edit different camp sites will do so at the exact same time, and suggests that geo-activism in OSM around this theme tends to be asynchronous across different locations. This asynchronous nature could also be attributed, in part, to the specific history of each camp site. Some camps, such as Bidibidi, exhibit OSM edit activity bursts towards the end of our study period, while more established camps, such as Dadaab, exhibit several bursts of activity spread throughout the study period. For camps such as Nyarugusu, a single distinct edit activity burst is apparent. Finally, there were no clearly observed co-occurrences of OSM edit activity in Fig 3 and camp establishment dates as noted in Table 1 .
Geographical saliency: Camps as local activity hotspots
In order to assess the degree to which a refugee camp becomes a local geographical salient artifact, overshadowing interest on its immediate surroundings, we evaluate the extent to which OSM editing activity within their boundaries exceeds the editorial activity in their immediate surroundings. In that sense, camps then become local activity “hotspots” of digital geo-activism. In order to estimate the OSM editing activity both within and around the camps, we define a set of 4 zones for each camp: a camp zone Z 0 and 3 surrounding aerial zones, Z 1 , Z 2 , and Z 3 . The relationship between camp and surrounding zones is shown Fig 4 .
10.1371/journal.pone.0206825.g004
Fig 4
A schematic example of a camp (aggregate of black polygons), its minimum bounding box (dashed white line B) and its surrounding zones Z 1 (light gray), Z 2 (aggregate of light and mid gray), and Z 3 (aggregate of Z 2 and darkest gray).
Formally, zone Z 0 is defined as the spatial footprint of the camp, as captured by the set of (one or more) polygons P = {p 1 , p 2 , …, p n } (n≥1) that delineate it. Based on this, zone Z 0 is defined as:
Z 0 = ∪ α ( p i ) , i = 1 , 2 , … , n
(3)
where α(p i ) is the area enclosed by polygon p i , and ∪ is the union operator. Given P, the set of minimum bounding boxes of p 1 , p 2 , …, p n are then derived, and the maximum side length (width or height) d max among all bounding boxes is found. In addition, the minimum bounding box B that encloses P is also derived. Using d max and B the zones Z 1 , Z 2 , and Z 3 (respectively) are derived by repeatedly applying a dilation operator (⊕) on the area enclosed by B (denoted by α(B)) using a square structuring element D with a dimension d max as follows:
Z 1 = α ( B ) ⊕ D + α ( B ) - Z 0 ; Z 2 = Z 1 ⊕ D - Z 0 ; Z 3 = Z 2 ⊕ D - Z 0
(4)
We use these zones to compare OSM editing activity within the camp versus its 3 surrounding zones (Z 1 , Z 2 , Z 3 ) during a ±4 month period around the strongest extremum point extracted from the weekly data, and examined at the daily granularity for each camp. The temporal window used for each refugee camp is shown in Table 3 with the results from this analysis summarized in S3 File . These results exhibit a sharp decline in editing activity along the transition from the camps outwards, to their surrounding zones. On average, for five of the eight camps the drop in number of OSM edits from Zone Z 0 to Z 1 was 93%. The remaining three camps, namely Kakuma, Calais and Bidibidi, have edits in Z 1 exceeding edits within Z 0 . In the case of Calais, this finding can be explained by the camp’s neighboring synonymous city being part of the Calais Z 1 zone. Similarly, in the case of Bidibidi, the increase in the number of OSM edits can be attributed to the fact that Bidibidi is surrounded by other refugee camps that are characterized by substantial edit activity, and are included in Bidibidi’s corresponding Zones Z 1 to Z 3 .
10.1371/journal.pone.0206825.t003
Table 3 Temporal window (±4 month period) around the strongest extremum point extracted from the weekly data at the daily granularity.
Site
Start date
Stop date
Dadaab
10/16/14
6/15/15
Kakuma
10/6/16
6/5/17
Nyarugusu
1/25/15
9/24/15
Calais
5/7/15
1/6/16
Yida
9/25/15
5/24/16
Bidibidi
9/23/16
5/22/17
Oncupinar
1/30/16
9/29/16
Zaatari
5/19/16
1/18/17
The drop in the number of OSM edits becomes even more pronounced when considering the normalized edit metrics per area for each zone. On average, for the eight camps studied here, Z 1 , Z 2 , and Z 3 zones cover an area that is 18, 54, and 108 times larger than the area of Z 0 , respectively. In order to account for these zone area variations, the number of OSM edits is normalized to be per km 2 , as shown in S3 File . With the exception of Kakuma, the data shows that on average the number of OSM edits per km 2 drops as one moves from Z 0 outwards by 96.37% (Z 1 ), 97.96% (Z 2 ), and 98.42% (Z 3 ). This pattern was also similar for Kakuma, however, the drop in the number of edits per km 2 was much lower, moving from Z 0 outwards by 2.71% (Z 1 ), 63.99% (Z 2 ), and 81.91% (Z 3 ). These results, therefore, support the notion that the camps are indeed serving as local activity hotspots, attracting OSM edits from the corresponding volunteer community beyond what would be expected by their surrounding areas.
Similar to the number of OSM edits, an analysis of the number of contributors per km 2 was carried out, as summarized in S3 File . These results exhibit a trend similar to the one found in the number of OSM edits per km 2 : the number of contributors per km 2 also drops as one moves from Z 0 outwards by 95.07% (Z 1 ), 97.59% (Z 2 ), and 98.09% (Z 3 ). Interestingly, in the case of Calais, the number of edits per km 2 is increasing moving from Z 0 outwards. Once again, this reversal in trend can be explained by the proximity of this refugee camp to a large urban area in a developed country.
In order to complete the saliency analysis, S3 File also lists the number of individual camp contributors who were active in the surrounding zones. As can be seen from these results, overall, the number of contributors who were active in the each of the camp zones (zone Z 0 ) who were also active in other zones is approximately 40% (the values range between 25%-50%), further suggesting the role of camps as local OSM edit activity hotspots.
Public awareness versus OSM and Wikipedia edit activity
In order to assess the possible relationship between news media coverage and digital activism in OSM, we compare OSM edit activity to three additional data sources, namely Wikipedia edits, Google News and Google Trends for each camp using the progressive refinement approach described earlier. As stated previously, in the context of this comparison, Google News is used as an indicator of media coverage, conveying how frequently a refugee camp appeared in news. In contrast, Google Trends is used as an indicator of broad public interest in this camp. Finally, Wikipedia edit activity represents an example of a non-geographic form of digital activism.
Fig 5 shows the time series of all four data sources (OSM edits, Wikipedia edits, Google News items, and Google Trends indicator) for each of the eight camps during a ±4 months period around the strongest extremum point of each camp (at the weekly temporal granularity– Table 3 ). Based on these time series data, the public awareness curve was calculated (as the cumulative difference of Google Trends and Google News activity) for each camp and the strongest extremum points were detected. Then, using the progressive refinement approach presented earlier, we examined the relationship between the public awareness curve and the OSM and Wikipedia edit activity in a time window of ±4 months around the extremum point. Fig 6 depicts the public awareness curve (magenta line) along with the cumulative OSM and Wikipedia edit activity (black and red lines, respectively). In these graphs we also show splines (dashed blue lines) fitted to the public awareness curves to better visualize the overall trends in these curves. In the context of our approach, because we analyze extreme cases of public awareness, any significant deviation between the two input variables (i.e. Google Trends and Google News) at the daily analysis level is viewed as a possible predictor of activism activity.
10.1371/journal.pone.0206825.g005
Fig 5
OSM, Wikipedia, Google News, and Google Trends time series during a ±4 months period around the strongest extremum point of each camp.
The figures show that whereas OSM and Wikipedia entries tend to come in bursts, Google News and Trends display a more sustained type of activity.
10.1371/journal.pone.0206825.g006
Fig 6
The public awareness curve versus the cumulative OSM and Wikipedia edit activity during a ±4 months period around the strongest extremum point of each camp.
For camps such as Nyarugusu, OSM and Wikipedia bursts overlap with public awareness surplus. In other camps, such as Bidibidi, OSM edit activity bursts coincide with public awareness deficit.
As the graphs in Fig 6 show, digital activism bursts tend to be associated with periods of consistent build-up of surplus or deficit in public awareness. This tendency is particularly evident in the Dadaab, Nyarugusu and Yida camps, where OSM edit activity bursts tend to overlap with public awareness surplus, and in the Bidibidi camp, where OSM edit activity bursts coincide with public awareness deficit. Such patterns of activity are also evident with respect to Wikipedia edits for some camps: for example, Dadaab, Nyarugusu and Bidibidi. However, in other camps, namely Kakuma, Calais and Zaatari, Wikipedia edits exhibit a rather prolonged sustained effort of edit activity compared to the activity bursts in OSM (note that for Yida and Oncupinar no Wikipedia edits were made during the ±4 months period around the strongest extremum point).
To further examine the association between public awareness trends and OSM and Wikipedia edit activity, we derived the time gap between the most significant extremum point in OSM and Wikipedia extremum points to the closest extremum point in the public awareness curve of each camp at the three temporal granularity levels. The time gaps that were found at the monthly, weekly, and daily time granularities as a result of the progressive refinement process along with the range of time gap values across the eight camp sites are provided in Table 4 . As can be seen from this table, at the finest (daily) temporal granularity, the average time gap is approximately between 11 and 12 days.
10.1371/journal.pone.0206825.t004
Table 4 Summary of the time gap values derived at the monthly, weekly, and daily levels for OSM and Wikipedia.
Platform
Time granularity units
Average time gap (and range)
OSM
Monthly
104.6 (0, 334)
Weekly
43.0 (0, 126)
Daily
10.8 (1, 32)
Wikipedia
Monthly
108.8 (0, 304)
Weekly
44.3 (0, 133)
Daily
12.0 (1, 27)
Discussion
Today’s age of the participatory news consumer [ 104 ] has been steadily blurring the lines between digital content consumption and production [ 105 , 106 ]. An emerging manifestation of this change is the bridging of the gap between the omnipresence of news in one’s daily life and one’s resulting expression of activism [ 107 , 108 ]. Such expressions of activism have previously been studied in the context of the shaping of public policy (e.g. [ 109 , 110 ]), political campaigns (e.g. [ 62 ]), environment (e.g. [ 111 ]) and climate change (e.g. [ 112 ]) issues. However, little is still known about the impact that news media coverage has on digital activism, especially as it relates to online crowdsourcing platforms such as OSM and Wikipedia.
Our objective in this paper was to advance our understanding of the complex interrelationships that link media coverage and digital activism by focusing in particular on the geographical dimension of news media and the manifestation of digital activism—in particular edit activity—in both geographic and non-geographic crowdsourcing platforms. In order to pursue this goal, we used refugee camps as a test case. As noted earlier, refugee camp sites are particularly suitable for this objective due to their conceptual alignment to the motivational factors that drive volunteerism, and due to their distinct geographical locations, that render their OSM and Wikipedia contributions exclusively related to the camps themselves. Our analysis focused on two interrelated themes, namely the saliency of refugee camps as geographically distinct subjects of digital activism, and the possible co-occurrences between public awareness trends and digital activism. In both cases edit activity of contributors in OSM and Wikipedia was considered to be a manifestation of digital activism.
Considering the issue of saliency, we examined the OSM edit activity in each camp site and a set of three surrounding zones, both in terms of the number of edits and in terms of unique contributors. It was found that, in general, both the total number of edits and the number of edits per km 2 drops substantially around camp sites compared to the edit activity within them. Additionally, a similar decay was found with respect to the number of unique contributors that were engaged in edit activities in camp sites versus the zones surrounding these sites. These results suggest that refugee camp sites tend to serve as geographically salient features that attract purposeful digital activism. Moreover, the decay in the number of unique contributors around the periphery of the camp sites suggest that the camp sites become salient objects of awareness to which OSM contributors pay specific attention. These findings give rise to the idea that the geographic saliency and awareness saliency are interdependent in the context of digital activism.
Focusing on the notion of awareness saliency and digital activism, we then explored the relationship between public awareness and evidence of digital activism related to refugee camps. Using a public awareness measure that was derived from Google News and Google Trends, we compared trend changes in public awareness to patterns of edit activity in the OSM and Wikipedia crowdsourcing platforms. Our findings indicate that in these platforms digital activism bursts tend to take place during periods of build-up of public awareness surplus or deficit, with an average time gap of approximately 11 to 12 days from extremum points in OSM and Wikipedia activity curves to the closest extremum point in the public awareness curve. It is important to note that the average time gap values were consistent across the two crowdsourcing platforms for all tree time granularities that were examined (namely monthly, weekly, and daily). However, our analysis shows that these two platforms do not always share similar activity patterns. Specifically, the results suggest that OSM edit activity within refugee camps tends to be concentrated in distinct bursts, while Wikipedia edit activity is often characterized by a gradual sustained edit activity effort. This difference suggests that while the user communities in both platforms are potentially exposed to the same public awareness trends, the response of each community may not be the same.
While our analysis did not address directly the issue of motivation, the results of our public awareness analysis highlight the multifaceted nature of motivation in the context of crowdsourcing. Specifically, our results indicated that edit activity bursts can occur during periods of sustained surplus or deficit in public awareness. These seemingly contradicting findings can be explained by two complementary theories in mass communication related to activism, namely agenda setting theory [ 99 ] and corrective action theory [ 113 ]. A manifestation of the former is the finding that periods of consistent public awareness surplus lead to increased saliency of the corresponding refugee camps as attention artifacts, which in turn lead to edit activity bursts (as is the case with Dadaab, Nyarugusu, and Yida). A manifestation of the latter is the finding that periods of consistent public awareness deficit (as is the case with the rest of the camps), which increases the saliency of the camp sites as attention artifacts due to the perceived lack of coverage of the topic in the news media, and leads to edit activity.
Combined, these results suggest the potential of a novel stimulus-awareness-activism (SA 2 ) framework in today’s participatory digital age. This framework, as presented in Fig 7 , is built on three primary constructs: (1) stimulus, (2) awareness, and (3) activism. In this framework, stimulus is provided by news media coverage of a specific topic (expressed through Google News metrics). We argue that over time, such coverage leads to awareness , whereby the public seeks additional information on the topic (expressed through Google Trends metrics). When awareness grows faster than news coverage a build-up of awareness surplus occurs: a topic resonates with the public, and in a sense goes viral. When awareness growth lags in comparison to news coverage awareness deficits occurs: a topic fails to capture the public’s interest and slowly fades away. Build-ups of awareness surplus or deficits lead to activism. Such activism can be manifested either offline (e.g. participating in crowdfunding efforts or volunteering for a non-government organization) or online (e.g. participating in OSM or Wikipedia activities). One could reasonably expect that this process is cyclical in nature, as activism is likely to lead to increased news coverage, providing renewed stimulus and creating a feedback loop in this stimulus-awareness-activism (SA 2 ) framework. In terms of the individuals involved in such activities, it’s important to note that while a large population of individuals may be exposed to the stimulus, only a portion of this population may develop awareness, and an even smaller portion will engage in activism. Additionally, it is important to point out that new individuals may become exposed and develop awareness as the topic gains saliency in the news awareness ecosystem.
10.1371/journal.pone.0206825.g007
Fig 7
Stimulus-awareness-activism (SA 2 ) framework.
In this paper, we studied the framework that links news coverage to awareness and activism by focusing in particular on the question of geographical saliency. Certain news stories tend to have a specific geographic dimension associated with them, and refugee camps are an excellent example of this, both explicitly and implicitly. Explicitly because these camps are physical constructs and have a specific location that they occupy (and in some cases are even named after that location), while implicitly because they are stops along geographical pathways that take the refugees from an origin location (e.g. the homeland that they had to abandon) to a destination location (their intended final destination). Accordingly, geographical saliency is rather prominent when considering these types of stories. However, geography is also prominent in most other news stories [ 114 ], just like for example geographical content is prominent in the vast percentage of Wikipedia entries [ 115 ].
As with any study, this study has some noteworthy limitations. Here, we highlight several such limitations that could be further investigated and refined. The first relates to the number of refugee camps studied. In our paper, although we selected refugee camps from around the world, only 8 camps were used. However, a larger cross-section of camps would be useful in exploring the relationship in news media coverage and digital geo-activism in greater breadth. Such research can also benefit from a much longer-term study of these variables, which can further be used to better understand the movement of these variables overtime and their possible association with other exogenous factors (e.g. crisis events). Second, while our analysis was done in the context of refugee camps, further analysis is required in order to explore whether similar patterns can be observed in other contexts (e.g. natural and manmade disasters, and disease outbreaks). Third, as noted above, the examination of motivational factors is not a straightforward process and requires further investigation (e.g. large-scale surveys of the motivational factors of OSM contributors).
Another, fourth, issue relates to the focus of this study only on the news awareness ecosystem in the English language, primarily due to the overwhelming pervasiveness of English in online content compared to other languages [ 116 ]. As the behavior of the news awareness ecosystem in other languages may be different, further analysis of the possible relationship between public awareness and digital activism across different languages is needed. A fifth issue we highlight concerns the use of the volume of Google News items in the public awareness measure used in this work. By considering only the volume and not the content or the impact of each news item our approach takes a simplistic view in which all news items are regarded as equal. However, in practice it is possible that some news items (or news outlets) may become more influential than others, which may result in a different pattern of digital activism. Examining this issue requires a separate line of inquiry that involves the development of appropriate measures for estimating the influence of news items as well as content analysis.
A sixth related issue is that our study only considers extreme bursts of activity in public awareness that tend to trigger digital geo-activism. However, the input variables used in determining such public awareness are expected to be in a perpetual state of fluctuation, with their own unique circadian patterns and influenced by various factors such as seasonality and crisis situations, among others. A more in-depth study analyzing these specific patterns would therefore be of interest. Another related limitation of our study is that for some camps, namely Yida and Oncupinar, there were no Wikipedia edits for these camps for the specific search window used when examined at the daily granular level. In the case of Oncupinar, recent reports have shown that Wikipedia editing access in Turkey was blocked by government authorities in April 2017 [ 117 ]. This period, however, is beyond our study period. Nevertheless, it is difficult to assess the impact that such actions may have on inactive periods of edit activity in online platforms such as Wikipedia since for example, technology measures exist that may nullify their effect (e.g. [ 118 – 119 ]). Recent studies, for example, have shown that digital censorship may also have the opposite effect, that is, there is an increase drive to access more information, and thus innovative ways to overcome such restrictions emerge (e.g. [ 120 ]).
Further work could also address the degree to which geographical saliency drives digital activism in news stories that relate to geographical areas that are not as monothematic as refugee camps, but rather are often featured in news stories for a wide variety of issues. For example, a megacity may find itself in the news following a major disaster, yet at the same time, it may also be featured for the numerous other activities/issues that are associated with it. Studying such multi-thematic geographical areas will allows us to further refine the SA 2 framework. Such refinements will allow us to devise more effective communication campaigns that will harness the power of the crowd in an organized manner to build responses to societal needs, such as mapping uncharted parts of the Sub-Saharan Africa to better study the birth and spread of exotic diseases at the human-environment interface. Even more importantly, such studies will offer us a better understanding of how our societies function across the cyber-physical news awareness ecosystem that is becoming the prevailing paradigm when interacting with the news.
Supporting information
S1 File
Population references for refugee camps.
(DOCX)
S2 File
OSM edits, Wikipedia edits and Google News articles at the daily level for the period 01/01/2010 to 05/31/2017.
(DOCX)
S3 File
OSM edits within camps and surrounding areas (±4 month period) around the strongest extremum point extracted from the weekly data at the daily granularity.
(DOCX)
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.