transhumanist-already-exists/aida-asian-pbmc-cell-age-related-cell-sentence-balanced-120k

Cell2Sentence for Longevity - Balanced Dataset (120k)

This dataset is a carefully curated and balanced subset of single-cell transcriptomics data, designed for training foundation models to understand cellular aging and longevity patterns.

Data Sources

Primary Data Source

• CZI CellxGene Collection: Tabula Sapiens - A multiple-organ, single-cell transcriptomic atlas of humans
• Original collection provides comprehensive single-cell RNA-seq data across multiple human organs and donors

Gene Annotations

The dataset includes specialized gene sentence embeddings from two authoritative longevity databases:

1. gene_sentence_opengenes: Top age-related genes ranked by expression from Open Genes Database

   • Curated collection of genes associated with aging and longevity
   • Expression-based rankings for age-related biological processes

2. gene_sentence_human_genage: Human aging genes from GenAge Database

   • The Human Ageing Genomic Resources (HAGR) collection
   • Genes potentially associated with human aging

Dataset Statistics

• Total samples: 119,792
• Train samples: 95,846 (500 donors)
• Test samples: 23,946 (125 donors)
• Unique cell types: 32
• Samples per donor: 192 (median), 191.7 (mean)

Train/Test Split

• Train: 80% of donors (500 donors)
• Test: 20% of donors (125 donors)
• No donor overlap: Complete separation between train and test sets

Data Preparation Pipeline

1. Diversity-Preserving Donor Split

We created a stratified 80/20 train/test split at the donor level to ensure proportional representation of demographic and clinical characteristics:

Stratification factors:

• Age (binned into groups)
• Disease status
• Sex
• Self-reported ethnicity
• Smoking status

This approach ensures that the test set is representative of the overall donor population distribution.

2. Balanced Sampling Strategy

To create a balanced dataset with equal donor representation:

Target: 120,000 samples total (192 samples per donor × 625 donors)

Sampling algorithm:

1. Each donor contributes exactly 192 samples (or all available if fewer)
2. Samples are distributed proportionally across cell types for each donor
3. Two-pass redistribution algorithm:
   • First pass: Allocate target samples to each cell type (192 ÷ number of cell types)
   • Second pass: If any cell type has insufficient samples, redistribute the shortfall to other cell types from the same donor

Results:

• 623 out of 625 donors achieve exactly 192 samples
• 2 donors have fewer samples due to limited data availability in source:
  • JP_RIK_H007: 71 samples (only 71 cells available in original data)
  • SG_HEL_H001: 159 samples (only 159 cells available in original data)

3. Technical Implementation

Parallel processing:

• All operations use 64-worker parallel processing for efficiency
• HuggingFace datasets library with num_proc=64 for fast data loading
• Efficient shuffling using HF datasets .shuffle() method

Reproducibility:

• Random seed: 42 for train set
• Random seed: 1042 for test set
• All sampling operations are deterministic

Cell Type Distribution

The dataset includes 32 unique cell types spanning multiple organ systems:

• Immune cells (T cells, B cells, NK cells, monocytes, dendritic cells)
• Blood cells (erythrocytes, platelets)
• Tissue-specific cells (hepatocytes, epithelial cells, endothelial cells)
• And more...

Donor Diversity

The dataset represents diverse demographic and clinical characteristics:

• Age range: Multiple age groups from young adults to elderly
• Geographic diversity: Multiple countries and ethnicities
• Health status: Both healthy donors and various disease conditions
• Smoking status: Balanced representation of smokers and non-smokers

Smoking Status Distribution

• Train: 71.6% non-smokers, 23.0% smokers, 5.4% unknown
• Test: 67.2% non-smokers, 29.6% smokers, 3.2% unknown
• Difference: ±6.6% (acceptable variation for stratified sampling)

Data Format

Each sample contains:

• Gene expression data: Transcriptomic profiles
• Cell metadata: Cell type, donor ID, tissue information
• Donor metadata: Age, sex, disease, ethnicity, smoking status
• Gene annotations: Longevity-related gene sentence embeddings

Use Cases

This dataset is designed for:

• Training foundation models for cellular aging
• Understanding cell-type-specific aging patterns
• Predicting longevity biomarkers
• Cross-donor generalization studies
• Age-related disease research

Citation

If you use this dataset, please cite:

1. Original data source:

   • The Tabula Sapiens Consortium. (2022). The Tabula Sapiens: A multiple-organ, single-cell transcriptomic atlas of humans. Science.
   • CellxGene Collection: https://cellxgene.cziscience.com/collections/ced320a1-29f3-47c1-a735-513c7084d508

2. Gene annotations:

   • Open Genes Database: https://open-genes.com/
   • GenAge Database: Tacutu, R., et al. (2018). Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Research.

Files

• train.parquet (16GB): Training dataset with 95,846 samples
• test.parquet (3.9GB): Test dataset with 23,946 samples
• stats.json: Detailed dataset statistics

Preprocessing Scripts

The complete data preparation pipeline is available in our repository, including:

• create_diversity_split.py: Stratified donor split generation
• create_balanced_dataset_hf.py: Balanced sampling with redistribution
• check_smoking_distribution.py: Demographic validation
• check_donor_overlap.py: Train/test separation verification

License

Please refer to the original CZI CellxGene data license and terms of use for the Tabula Sapiens collection.

Acknowledgments

• Chan Zuckerberg Initiative for the CellxGene Data Portal
• The Tabula Sapiens Consortium for the original dataset
• Open Genes and GenAge teams for longevity gene annotations