transhumanist-already-exists/aida-asian-pbmc-cell-age-related-cell-sentence-diverse-60k

AIDA Asian PBMC Cell Age-Related Cell Sentence Diverse 60K Dataset

Overview

This dataset contains 60,000 diverse single-cell RNA-seq samples (50k train + 10k test) selected from 1.2 million cells using weighted diversity sampling. Each sample includes multiple gene sentence representations optimized for longevity and aging research.

Dataset Creation

Source Data

• Original dataset: longevity-genie/cell2sentence4longevity-data
• Source folder: 9deda9ad-6a71-401e-b909-5263919d85f9/train
• Files processed: 159 parquet files
• Total source rows: 1,202,886 cells
• Final dataset: 60,000 diverse samples (5% of source)

Diversity Sampling Algorithm

We applied weighted diversity sampling across 7 demographic and biological dimensions to ensure balanced representation:

Diversity Dimensions

1. Sex: male, female
2. Disease: health conditions
3. Tissue: tissue type (blood)
4. Cell type: 32 different cell types (T cells, B cells, monocytes, etc.)
5. Age: binned into decades (0-20, 21-30, 31-40, 41-50, 51-60, 61-70, 71-80, 81-90, 91+)
6. Self-reported ethnicity: 8 categories
7. Ethnicity original: original ethnicity annotations

Sampling Procedure

# 1. Create diversity key from all dimensions
diversity_key = sex | disease | tissue | cell_type | age_bin | ethnicity

# 2. Calculate frequency of each combination
frequency = count(diversity_key)

# 3. Assign inverse frequency weights (rare combinations get higher weight)
weight = max_frequency / frequency

# 4. Normalize to probabilities
probability = weight / sum(weights)

# 5. Sample without replacement using weighted probabilities
samples = weighted_random_sample(n=60000, p=probability)

This approach ensures:

• Rare combinations are proportionally over-sampled
• Common combinations are proportionally under-sampled
• Maximum diversity in the final dataset
• Reproducibility (random seed = 42)

Dataset Statistics

Overall

• Train samples: 50,000
• Test samples: 10,000
• Features: 62 columns
• Unique diversity combinations (train): 8,365 (16.7% uniqueness ratio)
• Unique diversity combinations (test): 4,933 (49.3% uniqueness ratio)

Diversity Scores (Train Split)

Normalized entropy scores (0=concentrated, 1=uniform distribution):

Dimension	Unique Values	Top 1%	Top 5%	Top 10%	Diversity Score
Sex	2	52.7%	100%	100%	0.998
Cell Type	32	4.4%	21.3%	41.1%	0.973
Age	55	5.0%	20.4%	35.4%	0.956
Self-reported Ethnicity	8	18.7%	77.8%	100%	0.942
Disease	1	100%	100%	100%	0.000
Tissue	1	100%	100%	100%	0.000
Ethnicity Original	1	100%	100%	100%	0.000

Age Distribution (Train)

• Range: 19-77 years
• Mean: 44.4 years
• Median: 43.0 years
• Std Dev: 14.3 years

Age Range	Count	Percentage
0-20	1,252	2.5%
21-30	8,434	16.9%
31-40	11,650	23.3%
41-50	12,098	24.2%
51-60	6,037	12.1%
61-70	9,638	19.3%
71-80	891	1.8%

Gene Sentence Columns

Each sample includes 5 different gene sentence representations:

1. gene_sentence_all: All expressed genes
2. gene_sentence_2000: Top 2000 expressed genes
3. gene_sentence_clock_product_2k_top_genes: Top 2k genes from clock product
4. gene_sentence_human_genage: Human aging-related genes (GenAge database)
5. gene_sentence_opengenes: Genes from OpenGenes database

These representations are optimized for different downstream tasks in longevity and aging research.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("transhumanist-already-exists/aida-asian-pbmc-cell-age-related-cell-sentence-diverse-60k")

# Access train and test splits
train_data = dataset['train']
test_data = dataset['test']

# Example: Get a sample
sample = train_data[0]
print(f"Age: {sample['age']}")
print(f"Cell type: {sample['cell_type']}")
print(f"Sex: {sample['sex']}")
print(f"Gene sentence (top 2000): {sample['gene_sentence_2000']}")

Intended Use

This dataset is designed for:

• Longevity research: Analyzing age-related gene expression patterns
• Cell type classification: Multi-class classification across 32 cell types
• Demographic modeling: Understanding biological variation across age, sex, and ethnicity
• Language model training: Pre-training or fine-tuning models on biological sequences
• Transfer learning: As a diverse foundation for specialized single-cell tasks

Limitations

• Tissue limitation: Only blood/PBMC samples (tissue diversity score = 0)
• Disease limitation: Limited disease annotations (disease diversity score = 0)
• Ethnicity: Primarily Asian population (from AIDA dataset)
• Age range: 19-77 years (limited pediatric and very elderly samples)
• Sample size: 60k samples may be small for some deep learning applications

Citation

If you use this dataset, please cite:

@dataset{aida_pbmc_diverse_60k_2025,
  title={AIDA Asian PBMC Cell Age-Related Cell Sentence Diverse 60K Dataset},
  author={Transhumanist-Already-Exists},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/transhumanist-already-exists/aida-asian-pbmc-cell-age-related-cell-sentence-diverse-60k}}
}

And the original source dataset:

@dataset{cell2sentence4longevity_2024,
  title={Cell2Sentence for Longevity Data},
  author={Longevity Genie},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/longevity-genie/cell2sentence4longevity-data}}
}

License

This dataset inherits the license from the source dataset. Please refer to the original dataset for licensing information.

Contact

For questions or issues, please open an issue on the dataset repository.

---

Dataset Statistics Generated: 2025-11-13
Sampling Method: Weighted diversity sampling with inverse frequency weighting
Random Seed: 42
Processing Threads: 256