ai-department-lpnu/gtex-single-cell-rnaseq

GTEx Single-Cell RNA-seq Dataset

This repository provides tools to create a Hugging Face dataset from GTEx single-nucleus RNA-seq data, transforming the hierarchical H5AD format into a flat, ML-ready structure.

Overview

Data Source

The data comes from GTEx's snRNA-seq atlas:

• Source: GTEx Portal
• Publication: Eraslan et al., Science 2022 - "Single-nucleus cross-tissue molecular reference maps toward understanding disease gene function"
• Content: 209,126 cells from 16 individuals across 10 tissue types
• Original Format: H5AD (AnnData)
• File Size: 1.9 GB (compressed)

Dataset Statistics

Metric	Value
Total Cells	209,126
Genes per Cell	17,695
Donors	16 individuals
Tissue Types	10 tissues
Total Expression Values	~3.7 billion

Gene Name Mapping

Since expression_data is a flat list of 17,695 values, gene names are provided separately in gene_names.txt (one gene name per line, where line number = gene index).

Example - Finding expression for a specific gene:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

REPO_ID = "ai-department-lpnu/gtex-single-cell-rnaseq"

dataset = load_dataset(REPO_ID, split="train")

gene_names_path = hf_hub_download(
    repo_id=REPO_ID,
    repo_type="dataset",
    filename="data/gene_names.txt",
)
with open(gene_names_path) as f:
    gene_names = [line.strip() for line in f]

gene_to_index = {g: i for i, g in enumerate(gene_names)}

cell = dataset[0]

gene_idx = gene_to_index["SAMD11"]
samd11_expression = cell["expression_data"][gene_idx]

print(f"SAMD11 expression: {samd11_expression}")

Dataset Schema

The dataset includes expression data plus all 48 metadata fields from the original GTEx data:

Core Fields

expression_data

• Type: List of floats
• Length: 17,695 values per cell
• Format: Flat 1D list (no nested structure)
• Content: Normalized gene expression values
• Example: [1.23, 0.0, 2.45, ...]

Metadata Fields (48 fields total)

The dataset preserves all metadata from the original H5AD file, including:

Key Biological Metadata

• tissue: Tissue type (e.g., "Lung", "Heart", "Brain")
• Age_bin: Donor age range (e.g., "60-69")
• Sex: Donor biological sex
• Broad cell type: High-level cell type annotation
• Granular cell type: Detailed cell type annotation
• individual: Donor identifier

Sample Information

• Sample ID: Unique sample identifier
• Participant ID: GTEx participant identifier
• Tissue Site Detail: Detailed tissue location
• prep: Sample preparation batch

Quality Metrics

• nGenes: Number of genes detected
• nUMIs: Number of UMIs (transcripts) detected
• PercentMito: Mitochondrial gene percentage
• PercentRibo: Ribosomal gene percentage
• RIN score: RNA integrity number
• Autolysis Score: Tissue degradation metric
• Sample Ischemic Time (mins): Time before preservation

Technical Annotations

• scrublet: Doublet detection flag
• scrublet_score: Doublet score
• batch: Sequencing batch
• barcode: Cell barcode
• leiden: Leiden clustering assignment
• leiden_tissue: Tissue-specific clustering

Alignment Statistics

• introns, junctions, exons: Read mapping stats
• sense, antisense, intergenic: Read categories
• exon_ratio, intron_ratio, junction_ratio: Mapping ratios

And more... (see full list of 48 fields in the dataset)

Data Flow

Pipeline Overview

The dataset creation follows a simple 4-step pipeline:

GTEx Portal → Download → Load & Transform → HF Dataset

Step	Input	Output	What Happens
1. Download	GTEx URL	H5AD file (1.9 GB)	Stream download from Google Storage
2. Load	H5AD file	AnnData object	Read hierarchical structure into memory
3. Transform	AnnData object	Flat dictionaries	Extract & flatten 209K cells to simple structure
4. Save	Flat dictionaries	HF Dataset	Write to Arrow format on disk

Key Transformation: Hierarchical HDF5 (209K cells × 17K genes + 48 metadata fields) → Flat structure (49 columns: expression_data + all 48 metadata fields)

Data Transformation Steps

Step 1: H5AD File Structure

GTEx_8_tissues_snRNAseq_atlas_071421.public_obs.h5ad
│
├── X (Expression Matrix)
│   ├── Shape: (209126, 17695)
│   ├── Type: Sparse CSR or Dense Array
│   └── Data: Normalized gene expression
│
├── obs (Cell Metadata) - 209,126 rows × 48 columns
│   ├── tissue          ← All fields preserved
│   ├── Age_bin         ← All fields preserved
│   ├── Sex             ← All fields preserved
│   ├── Broad cell type ← All fields preserved
│   └── ... (44 more fields)
│
└── var (Gene Metadata) - 17,695 rows
    └── Gene names and annotations

Step 2: Load and Extract

# Load H5AD
adata = anndata.read_h5ad(file)

# Convert sparse to dense
expression_matrix = adata.X.toarray()  # (209126, 17695)

# Extract metadata
tissues = adata.obs['tissue'].values   # (209126,)
ages = adata.obs['Age_bin'].values     # (209126,)

Step 3: Transform to Flat Structure

Each cell is transformed from the AnnData structure to a simple dictionary:

# Before: Cell in AnnData
expression_matrix[i]  # numpy array (17695 genes)
metadata[i]           # All 48 metadata fields from obs DataFrame

# After: Flat dictionary
{
    'expression_data': [1.23, 0.0, 2.45, ...],  # List of 17,695 floats
    'tissue': 'Lung',                            # String
    'Age_bin': '60-69',                          # String
    'Sex': 'Female',                             # String
    'Broad cell type': 'Epithelial',             # String
    ... # + 44 more metadata fields
}

Step 4: Create HF Dataset

Dataset({
    features: {
        'expression_data': Sequence(Value('float64')),
        'tissue': Value('string'),
        'Age_bin': Value('string'),
        'Sex': Value('string'),
        'Broad cell type': Value('string'),
        ... # + 44 more metadata fields
    },
    num_rows: 209126
})

Expression Data Format

Each cell's expression_data is a flat list of 17,695 float values:

expression_data = [
    1.234,   # Gene 1 expression
    0.0,     # Gene 2 expression (not expressed)
    2.456,   # Gene 3 expression
    ...      # 17,692 more values
]

Characteristics:

• Sparsity: ~70-90% zeros (typical for single-cell data)
• Range: Usually 0 to ~15 (log-normalized values)
• Type: Float (64-bit in HF dataset)
• Size: 17,695 values per cell × 8 bytes = ~141 KB per cell

Citation

If you use this dataset, please cite the original GTEx publication:

@article{eraslan2022single,
  title={Single-nucleus cross-tissue molecular reference maps toward understanding disease gene function},
  author={Eraslan, G{\"o}kcen and Drokhlyansky, Eugene and Anand, Shankara and others},
  journal={Science},
  volume={376},
  number={6594},
  pages={eabl4290},
  year={2022},
  publisher={American Association for the Advancement of Science}
}

And the GTEx Consortium:

@article{gtex2020gtex,
  title={The GTEx Consortium atlas of genetic regulatory effects across human tissues},
  author={GTEx Consortium},
  journal={Science},
  volume={369},
  number={6509},
  pages={1318--1330},
  year={2020},
  publisher={American Association for the Advancement of Science}
}

License

The GTEx data is available under the GTEx Data Use Agreement.