electricsheepafrica/ssa-breast-data-quality-benchmarks

SSA Breast Data Quality Benchmarks (Sequencing & Sample QC, Synthetic)

Dataset summary

This module provides a synthetic data quality benchmark dataset for breast cancer sequencing studies, focusing on:

• Batch effects and inter-laboratory variation (sequencing center, platform, batch, run year).
• RNA quality metrics (RNA integrity number, yield, read depth, %Q30) and QC flags.
• DNA quality metrics (yield, coverage, %≥20x, duplication, contamination) and QC flags.

It is designed to support method development and teaching around technical variation, QC pipelines, and batch-effect correction, especially in multi-ancestry contexts including Sub-Saharan Africa.

All records are fully synthetic, parameterized from published multi-center NGS performance studies, RIN literature, and FFPE vs fresh-frozen comparisons.

Cohort design

Sample size and populations

• Total N: 8,000 synthetic sequencing libraries.

• Populations:

  • SSA_West: 1,600
  • SSA_East: 1,600
  • SSA_Central: 1,200
  • SSA_Southern: 1,200
  • AAW (African American women): 1,200
  • EUR: 800
  • EAS: 400

• Sex:

  • Female ≈95%, Male ≈5% (to allow mixed-sex QC modeling).

This dataset is not intended to represent incidence or prevalence; it is a QC sandbox for realistic multi-ancestry technical variation.

Sequencing centers and platforms

Variables:

• sequencing_center:

  • Lab_A_EUR (European reference center)
  • Lab_B_USA (US reference center)
  • Lab_C_EAS (East Asian reference center)
  • Lab_SSA_1, Lab_SSA_2 (SSA regional centers)

• sequencing_platform:

  • HiSeq4000
  • NovaSeq6000
  • NextSeq550

Center assignment depends on population (e.g., SSA samples enriched in Lab_SSA_1/2; EUR samples enriched in Lab_A_EUR; EAS in Lab_C_EAS). Platforms differ by center (e.g., more NovaSeq in reference centers, more HiSeq in SSA labs).

Additional technical variables:

• library_batch – integer batch identifier (1–12).
• run_year – uniform across 2012–2024 to emulate protocol drift / upgrades.

RNA quality metrics

For each sample, RNA metrics are present with probability has_rna_prob ≈ 0.85 (some libraries are DNA-only).

Variables:

• has_rna – boolean.
• rna_source – FFPE, Fresh_frozen, Blood_PAXgene (distribution by center).
• rna_rin – RNA Integrity Number (1–10 scale):
  • Base mean around ~7.2 with SD ~1.0.
  • Source effects: FFPE ≈ −2, Fresh_frozen ≈ +0.4, Blood_PAXgene ≈ +0.2.
  • Center effects: e.g., Lab_A_EUR slightly higher RIN than Lab_SSA_1/2.
• rna_yield_ng – expected input RNA mass (log-normal distribution with source effects).
• rna_read_count_millions – approximate RNA-seq depth:
  • Base ~40M reads, with platform effects (e.g., more reads on NovaSeq than NextSeq).
• rna_pct_q30 – % bases at Q30 or higher:
  • Base ~90%, with platform and center shifts reflecting inter-lab variation.

QC flags based on thresholds (RIN ≥6.5, ≥20M reads, ≥80% Q30):

• rna_qc_rin_pass
• rna_qc_read_count_pass
• rna_qc_q30_pass
• rna_qc_pass – overall RNA QC pass (all criteria met).

DNA quality metrics

For each sample, DNA metrics are present with probability has_dna_prob ≈ 0.90.

Variables:

• has_dna – boolean.
• dna_source – FFPE, Fresh_frozen, Blood (distribution by center).
• dna_yield_ng – DNA input mass (log-normal with source effects; FFPE yields lower on average).
• dna_mean_coverage – mean on-target coverage:
  • Base ~80×, lower for FFPE and SSA centers, higher for fresh-frozen and reference centers.
• dna_pct_bases_20x – % of target bases with ≥20× coverage (base ~92%).
• dna_duplication_rate – fraction of duplicated reads:
  • Base ~0.15, elevated in FFPE and some SSA centers; slightly reduced in reference labs.
• dna_contamination_pct – estimated sample contamination (%), generally low (~1%).

QC flags based on thresholds (e.g., coverage ≥30×, ≥80% bases at ≥20×, duplication ≤0.25, contamination ≤5%):

• dna_qc_coverage_pass
• dna_qc_pct20x_pass
• dna_qc_duplication_pass
• dna_qc_contamination_pass
• dna_qc_pass – overall DNA QC pass (all criteria met).

Technical variation & batch effects

The dataset encodes several layers of technical variation:

• Center-level shifts:

  • Higher mean RIN and Q30 in Lab_A_EUR and Lab_B_USA than in Lab_SSA_1/2.
  • Higher mean coverage and lower duplication in reference centers; more variable metrics in SSA labs.

• Platform effects:

  • NovaSeq oriented towards higher depth and slightly higher Q30.
  • NextSeq with somewhat lower depth/Q30, capturing platform-typical performance.

• Source effects:

  • FFPE samples have lower RIN, higher duplication, and slightly lower effective coverage.
  • Fresh-frozen samples have higher coverage and lower duplication.

These patterns follow the literature on batch effects, ABRF NGS platform benchmarks, and FFPE vs fresh-frozen DNA/RNA.

Files and schema

Data table

Files:

• data_quality.parquet
• data_quality.csv

Each row is one library/sample with:

• Demographics & identifiers:

  • sample_id
  • population, region, is_SSA, is_reference_panel
  • sex

• Technical factors:

  • sequencing_center
  • sequencing_platform
  • library_batch
  • run_year

• RNA metrics:

  • has_rna
  • rna_source
  • rna_rin
  • rna_yield_ng
  • rna_read_count_millions
  • rna_pct_q30
  • rna_qc_rin_pass
  • rna_qc_read_count_pass
  • rna_qc_q30_pass
  • rna_qc_pass

• DNA metrics:

  • has_dna
  • dna_source
  • dna_yield_ng
  • dna_mean_coverage
  • dna_pct_bases_20x
  • dna_duplication_rate
  • dna_contamination_pct
  • dna_qc_coverage_pass
  • dna_qc_pct20x_pass
  • dna_qc_duplication_pass
  • dna_qc_contamination_pass
  • dna_qc_pass

For samples without RNA or DNA (has_rna=False or has_dna=False), the corresponding metrics are NaN and QC flags are False.

Generation

The dataset is generated using:

• data_quality_benchmarks/scripts/generate_data_quality.py

with configuration in:

• data_quality_benchmarks/configs/data_quality_config.yaml

and literature inventory in:

• data_quality_benchmarks/docs/LITERATURE_INVENTORY.csv

Key generation logic:

1. Base sample table – multi-ancestry populations, sex, center, platform, batch, run year.
2. RNA QC assignment – conditional on has_rna and driven by source, center, and platform:
   • RIN values and yields are drawn from normal/log-normal distributions with additive effects.
   • Depth and Q30 reflect platform and center performance.
   • RIN/read count/Q30 thresholds define pass/fail flags.
3. DNA QC assignment – conditional on has_dna:
   • Yield, coverage, %≥20x, duplication, and contamination vary by source and center.
   • Coverage, %≥20x, duplication, and contamination thresholds define DNA QC pass/fail.

Validation

Validation is performed with:

• data_quality_benchmarks/scripts/validate_data_quality.py

and summarized in:

• data_quality_benchmarks/output/validation_report.md

Checks include:

• C01–C02 – Sample size and population counts vs configuration.
• C03 – Sequencing center distributions by population.
• C04 – RNA RIN distribution mean vs expected.
• C05 – RNA QC pass rate within a target range (not too permissive, not too strict).
• C06 – DNA coverage and duplication patterns for FFPE vs Fresh_frozen (Fresh_frozen higher coverage, FFPE higher duplication).
• C07 – DNA QC pass rate within a target range.
• C08 – Center effect magnitude: RIN and coverage differences between best and worst centers meet minimum thresholds.
• C09 – Missingness in key variables (demographics, technical factors, QC metrics).

Intended use

This dataset is intended for:

• Developing and testing QC pipelines for RNA-seq and DNA-seq.
• Benchmarking batch effect detection and correction methods.
• Teaching about:
  • Inter-laboratory variation and its impact on downstream analyses.
  • The role of RIN, coverage, duplication, and contamination in sequencing QC.
  • How technical covariates (center, platform, batch) interact with biological covariates (population).

It is not intended for:

• Estimating actual failure rates of real-world labs.
• Drawing conclusions about specific centers or platforms.
• Clinical QC decision-making for patient samples.

Ethical considerations

• All data are synthetically generated from literature-based parameter sets; no real sample identifiers or QC runs are used.
• Center and platform labels are generic; they must not be interpreted as referring to specific institutions.
• The dataset aims to enable more robust analyses under technical variability, not to evaluate real labs.

License

• License: CC BY-NC 4.0.
• Free for non-commercial research, method development, and education with attribution.

Citation

If you use this dataset, please cite:

Electric Sheep Africa. "SSA Breast Data Quality Benchmarks (Sequencing & Sample QC, Synthetic)." Hugging Face Datasets.

and, where applicable, key QC/batch-effect references such as Leek et al. (2010), Chen et al. (2011), ABRF NGS performance studies, and RIN/FFPE quality literature.