Datasets:

FuhaiLiAiLab
/

OmniCellTOSG_Dataset

pretty_name: OmniCellTOSG
dataset_name: omnicelltosg
dataset_summary: >
  OmniCellTOSG is a large-scale Text–Omic Signaling Graph (TOSG) dataset for
  single-cell learning.

  It integrates sharded expression matrices, graph topology (full/internal/PPI
  edges), and textual

  entity metadata (names, descriptions, sequences) with optional precomputed
  embeddings. It supports

  graph-aware pretraining and downstream tasks such as cell-type annotation,
  disease status, and gender classification.
annotations_creators:
  - no-annotation
language_creators:
  - found
language:
  - en
multilinguality:
  - monolingual
source_datasets:
  - original
  - external
size_categories:
  - '>1M'
task_categories:
  - other
task_ids:
  - multi-label-classification
  - explanation-generation
tags:
  - single-cell
  - transcriptomics
  - foundation-models
license: other
license_url:
  - https://cellxgene.cziscience.com/tos
  - https://doi.org/10.1038/s41591-024-03150-z
  - https://www.ncbi.nlm.nih.gov/geo/info/citations.html#third-party
homepage: https://github.com/FuhaiLiAiLab/OmniCellTOSG
repository: https://github.com/FuhaiLiAiLab/OmniCellTOSG
paper: https://arxiv.org/pdf/2504.02148
point_of_contact: Heming Zhang
dataset_type: multimodal-graph
configs:
  - config_name: default
    data_files: cell_metadata_with_mappings.csv
pretty_format: true

OmniCellTOSG

🧭 Overview

OmniCellTOSG is a large-scale Text–Omic Signaling Graph (TOSG) resource for single-cell foundation model pretraining and omics analysis. It combines:

Expression matrices (sharded .npy for scalable IO)
Graph topology (full, internal, and PPI edges)
Textual metadata (entity names, descriptions, sequences) with precomputed embeddings

Supported tasks include graph–language foundation model pretraining, cell-type annotation, disease status and gender classification, plus core signaling inference.

📁 Dataset Structure

OmniCellTOSG_Dataset/
├── expression_matrix/
│ ├── braincellatlas_brain_part_0.npy
│ ├── braincellatlas_brain_part_1.npy
│ ├── cellxgene_blood_part_0.npy
│ ├── cellxgene_blood_part_1.npy
│ ├── cellxgene_lung_part_0.npy
│ ├── cellxgene_small_intestine_part_0.npy
│ └── ... (additional *.npy shards)
├── cell_metadata_with_mappings.csv
├── cell_metadata_with_mappings.parquet
├── edge_index.npy
├── internal_edge_index.npy
├── ppi_edge_index.npy
├── s_bio.csv
├── s_desc.csv
├── s_name.csv
├── x_bio_emb.npy
├── x_desc_emb.npy
└── x_name_emb.csv

Notes:

Files in expression_matrix/*.npy are sharded partitions of single-cell expression matrices; merge shards (concatenate/stack) to reconstruct the full matrix for a given source/organ.

cell_metadata_with_mappings.csv contains standardized per-cell annotations (e.g., tissue, disease, sex, cell type, ontology mappings).

edge_index.npy, s_bio.csv, s_name.csv, and s_desc.csv provide the graph topology (COO [2, E]) and entity metadata (biological sequences, names, descriptions).

x_bio_emb.npy, x_desc_emb.npy, and x_name_emb.csv are precomputed entity embeddings ([#entities × D], encoder-dependent) aligned to the CSVs—use these to skip on-the-fly embedding.

⚙️ Installation

If you only need dataset loading/extraction, download the standalone loader package from the Releases page.

🚀 Quick Start

from CellTOSG_Loader import CellTOSGDataLoader

conditions = {"tissue_general": "brain", "disease_name": "Alzheimer's Disease"}

ddataset = CellTOSGDataLoader(
    root=args.dataset_root,
    conditions=conditions,
    task=args.task,                          # "disease" | "gender" | "cell_type"
    label_column=args.label_column,          # "disease" | "gender" | "cell_type"
    sample_ratio=args.sample_ratio,          # mutually exclusive with sample_size
    sample_size=args.sample_size,
    shuffle=args.shuffle,
    stratified_balancing=args.stratified_balancing,
    extract_mode=args.extract_mode,          # "inference" | "train"
    random_state=args.random_state,
    train_text=args.train_text,
    train_bio=args.train_bio,
    correction_method=args.correction_method, # None | "combat_seq"
    output_dir=args.output_dir,
)

# --- Access outputs ---
if args.extract_mode == "inference":
    X = dataset.data                         # pandas.DataFrame (expression/features)
    y = dataset.labels                       # pandas.DataFrame
    metadata = dataset.metadata              # pandas.DataFrame (row-aligned metadata)
else:
    X = dataset.data                         # dict: {"train": X_train, "test": X_test}
    y = dataset.labels                       # dict: {"train": y_train, "test": y_test}
    metadata = dataset.metadata              # dict: {"train": meta_train, "test": meta_test}

all_edge_index = dataset.edge_index                   # full graph (COO [2, E])
internal_edge_index = dataset.internal_edge_index     # optional transcript–protein edges
ppi_edge_index = dataset.ppi_edge_index               # optional PPI edges     
x_name_emb, x_desc_emb, x_bio_emb = pre_embed_text(args, dataset, pretrain_model, device) # Prepare text and seq embeddings

Parameters (`CellTOSGDataLoader`)

root (str, required) — Filesystem path to the dataset root (e.g., ../OmniCellTOSG/CellTOSG_dataset_v2).
conditions (dict, required) — Metadata filters used to subset rows
(e.g., {"tissue_general": "brain", "disease": "Alzheimer's disease"}).
task (str, required) — Downstream task type: "disease" | "gender" | "cell_type".
label_column (str, required) — Target label column (e.g., "disease", "gender", "cell_type").
extract_mode (str, required) — Extraction mode:
- "inference": extract a single dataset for inference/analysis (no train/test split)
- "train": extract a training-ready dataset and generate splits (e.g., train/test)
sample_ratio (float, optional) — Fraction of rows to sample (0–1). Mutually exclusive with sample_size.
sample_size (int, optional) — Absolute number of rows to sample. Mutually exclusive with sample_ratio.
shuffle (bool, default: False) — Shuffle rows during sampling/composition.
stratified_balancing (bool, default: False) — Enable stratified sampling / class balancing based on label_column.
random_state (int, default: 2025) — Random seed for reproducibility (sampling, shuffling, splitting).
train_text (bool, default: False) — Controls text feature output:
- False: return precomputed text embeddings (if available)
- True: return raw text fields for custom embedding
train_bio (bool, default: False) — Controls biological sequence feature output:
- False: return precomputed sequence embeddings (if available)
- True: return raw sequences for custom embedding
correction_method (str or None, default: None) — Correction method:
- None: no correction
- "combat_seq": apply ComBat-Seq
output_dir (str, optional) — Directory for loader outputs (extracted expression matrix, label，splits).

Returns:

extract_mode="inference":

dataset.data: pandas.DataFrame

dataset.labels: pandas.DataFrame

dataset.metadata: pandas.DataFrame

extract_mode="train":

dataset.data: dict ({"train": X_train, "test": X_test})

dataset.labels: dict ({"train": y_train, "test": y_test})

dataset.metadata: dict ({"train": meta_train, "test": meta_test})

edge_index, internal_edge_index, ppi_edge_index: graph topological information

Either raw text/sequence fields (s_name, s_desc, s_bio) or their precomputed embeddings (x_name_emb, x_desc_emb, x_bio_emb), returned according to the train_text/train_bio flags.

🧪 Pretraining

python pretrain.py

🏋️ Training Examples (CLI)

Disease status (AD, brain)

# Alzheimer's Disease (Take AD as an example)
python train.py \
  --downstream_task disease \
  --label_column disease \
  --tissue_general brain \
  --disease_name "Alzheimer's Disease" \
  --sample_ratio 0.1 \
  --train_base_layer gat \
  --train_lr 0.0005 \
  --train_batch_size 3 \
  --random_state 42 \
  --dataset_output_dir ./Data/train_ad_disease_0.1_42

Gender (AD, brain)

# Alzheimer's Disease (Take AD as an example)
python train.py \
  --downstream_task gender \
  --label_column gender \
  --tissue_general brain \
  --disease_name "Alzheimer's Disease" \
  --sample_ratio 0.1 \
  --train_base_layer gat \
  --train_lr 0.0005 \
  --train_batch_size 2 \
  --random_state 42 \
  --dataset_output_dir ./Data/train_ad_gender_0.1_42

Cell type annotation (LUAD, lung)

# Lung (LUAD) (Take LUAD as an example)
python train.py \
  --downstream_task cell_type \
  --label_column cell_type \
  --tissue_general "lung" \
  --disease_name "Lung Adenocarcinoma" \
  --sample_ratio 0.2 \
  --train_base_layer gat \
  --train_lr 0.0001 \
  --train_batch_size 3 \
  --random_state 42 \
  --dataset_output_dir ./Data/train_luad_celltype_0.2_42

Signaling inference

python analysis.py

⚖️ Licensing & Attribution

This dataset aggregates data from CellxGENE, the Brain Cell Atlas, GEO and HCA. Use of these resources is governed by their respective terms and citation policies:

CellxGENE Terms of Service — Follow the platform’s ToS for data access, reuse, and sharing.
🔗 https://cellxgene.cziscience.com/tos
Brain Cell Atlas (citation required)
Cite:
Xinyue Chen#, Yin Huang#, Liangfeng Huang#, Ziliang Huang#, Zhao-Zhe Hao#, Lahong Xu, Nana Xu, Zhi Li, Yonggao Mou, Mingli Ye, Renke You, Xuegong Zhang, Sheng Liu*, Zhichao Miao*. A brain cell atlas integrating single-cell transcriptomes across human brain regions. Nat Med (2024). https://doi.org/10.1038/s41591-024-03150-z
GEO Citation Policy — Follow NCBI GEO guidelines for citing datasets and third-party analyses.
🔗 https://www.ncbi.nlm.nih.gov/geo/info/citations.html#third-party
HCA Data Use Agreement
🔗 https://data.humancellatlas.org/about/data-use-agreement

Note: You are responsible for ensuring compliance with the licenses/terms and for providing appropriate attribution to each source in any publications or derived works.

📚 Citation

If you use OmniCellTOSG, please cite:

@misc{omnicelltosg2025,
  title  = {OmniCellTOSG: A Text–Omic Signaling Graph Dataset for Single-Cell Learning},
  author = {Zhang, Heming and Li, Fuhai and collaborators},
  year   = {2025},
  note   = {Dataset on Hugging Face},
  url    = {https://huggingface.co/FuhaiLiAiLab}
}