transhumanist-already-exists/aida-asian-pbmc-cell-sentence-top2000

AIDA Asian PBMC Cell Sentences (Top 2000 Genes)

Dataset Description

This dataset contains 1,265,624 single cells from peripheral blood mononuclear cells (PBMCs) of 619 healthy donors across 5 Asian countries, transformed into "cell sentences" - space-separated gene symbols ordered by expression level.

Each cell is represented as a sequence of the top 2,000 most highly expressed genes, enabling language model-style analysis of single-cell transcriptomics data.

Source

This dataset is derived from the Asian Immune Diversity Atlas (AIDA) project:

• Original Dataset: AIDA 5-country PBMC dataset
• Source Portal: CZ CELLxGENE Discover
• Dataset ID: 9deda9ad-6a71-401e-b909-5263919d85f9
• Download URL: https://datasets.cellxgene.cziscience.com/9deda9ad-6a71-401e-b909-5263919d85f9.h5ad

AIDA Project

The Asian Immune Diversity Atlas (AIDA) is a multi-national single-cell reference atlas of circulating immune cells from healthy donors across 5 Asian countries (India, Japan, South Korea, Singapore, Thailand), comprising over 1.2 million cells from 619 donors.

Dataset Statistics

• Total Cells: 1,265,624
• Countries: 5 (India, Japan, South Korea, Singapore, Thailand)
• Donors: 619 healthy donors
• Age Range: 19-77 years (54 unique ages)
• Tissue: Blood (PBMC)
• Cell Types: Multiple immune cell types
• Technology: 10x Genomics 5' v2
• Reference Genome: GRCh38
• Genes per Cell Sentence: 2,000 (top expressed)
• Total Size: ~12 GB

Country Distribution

Country	Cells	Percentage
🇸🇬 Singapore (SG)	394,523	31.2%
🇰🇷 South Korea (KR)	386,792	30.6%
🇯🇵 Japan (JP)	302,255	23.9%
🇹🇭 Thailand (TH)	135,978	10.7%
🇮🇳 India (IN)	46,076	3.6%

Dataset Splits

The dataset is pre-split into train and test sets with stratification by age:

• Train: 1,202,342 cells (95.0%)
• Test: 63,282 cells (5.0%)

Each age group has exactly 5% of cells in the test set, ensuring proportional representation across all 54 age groups (19-77 years).

Transformation Pipeline

The original h5ad file was processed through the following steps:

1. Gene Mapping: Converted Ensembl IDs to HGNC gene symbols using official HGNC mappings
2. Cell Sentence Generation: For each cell:
   • Extracted expression values for all genes
   • Sorted genes by expression level (descending)
   • Selected top 2,000 genes
   • Converted to space-separated string of gene symbols
3. Age Extraction: Parsed donor age from development_stage field
4. Format Conversion: Saved as parquet format for efficient loading

See TRANSFORMATION_PIPELINE.md for detailed documentation.

Dataset Schema

The dataset contains 50 columns:

Key Columns

• cell_sentence (string): Space-separated gene symbols ordered by expression (top 2,000 genes)
  • Example: "MALAT1 EEF1A1 RPL13 RPL41 RPS27 RPL10 RPS12 RPL34 RPS3A RPLP1..."
• age (int): Donor age in years (19-77)
• cell_type (category): Cell type annotation (T cell, B cell, NK cell, etc.)
• sex (category): Donor sex (male, female)
• donor_id (category): Unique donor identifier
• nCount_RNA (float): Total UMI counts per cell
• nFeature_RNA (int): Number of genes detected per cell
• pMito (float): Percentage of mitochondrial reads

Plus 42 additional metadata columns including donor demographics, sample processing details, and cell annotations.

Loading the Dataset

Using Hugging Face Datasets

from datasets import load_dataset

# Load the dataset (includes train/test splits)
dataset = load_dataset("transhumanist-already-exists/aida-asian-pbmc-cell-sentence-top2000")

# Access the data
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['cell_sentence', 'age', 'cell_type', 'sex', ...],
#         num_rows: 1202342
#     }),
#     test: Dataset({
#         features: ['cell_sentence', 'age', 'cell_type', 'sex', ...],
#         num_rows: 63282
#     })
# })

# View a sample from train set
sample = dataset['train'][0]
print(f"Cell type: {sample['cell_type']}")
print(f"Age: {sample['age']}")
print(f"Country: {sample['Country']}")
print(f"Cell sentence (first 100 chars): {sample['cell_sentence'][:100]}")

Use Cases

This dataset is suitable for:

• Cell Type Classification: Train language models to predict cell types from gene expression
• Cell Representation Learning: Learn embeddings of cells using transformer models
• Gene Pattern Analysis: Study co-expression patterns across different cell types
• Cross-population Studies: Compare with other AIDA subsets (Japan, Korea, Singapore, Thailand)
• Zero-shot Cell Type Prediction: Use pre-trained language models for cell annotation

Citation

If you use this dataset, please cite:

Original AIDA Dataset

Asian Immune Diversity Atlas (AIDA)
CELLxGENE Collection: ced320a1-29f3-47c1-a735-513c7084d508
https://cellxgene.cziscience.com/collections/ced320a1-29f3-47c1-a735-513c7084d508

Related Publications

• AIDA Consortium. "Asian diversity in human immune cells." Cell (2025)
• More information: https://www.a-star.edu.sg/gis/home/press-releases/press-releases-2025/scientists-assemble-world-s-first-immune-cell-atlas-from-diverse-asian-populations

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

• License URL: https://creativecommons.org/licenses/by/4.0/
• SPDX: CC-BY-4.0

License Terms

You are free to:

• ✅ Share: Copy and redistribute the material in any medium or format
• ✅ Adapt: Remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• ⚠️ Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made

See the full license text for details.

Modifications

This dataset has been modified from the original AIDA h5ad file:

1. Added cell_sentence column: Top 2,000 expressed genes as space-separated gene symbols
2. Added age column: Extracted from development_stage field
3. Converted Ensembl IDs to HGNC gene symbols
4. Converted format from h5ad to parquet

The original expression matrix is not included. For the full expression data, please download the original h5ad file from CELLxGENE.

Related Resources

• AIDA CELLxGENE Collection: https://cellxgene.cziscience.com/collections/ced320a1-29f3-47c1-a735-513c7084d508
• Human Cell Atlas: https://www.humancellatlas.org/
• HGNC Gene Nomenclature: https://www.genenames.org/

Contact

For questions about this dataset transformation, please open an issue in the GitHub repository.

For questions about the original AIDA data, please refer to the AIDA project documentation.