# GEOMeta: Large-Scale Human Bulk RNA-seq Dataset with Curated Metadata

## Overview

GEOMeta is a large-scale human bulk RNA-seq dataset derived from the [ARCHS4](https://maayanlab.cloud/archs4/) resource, containing transcript abundance profiles for ~474,000 samples across training, test, and held-out test splits. Each sample is annotated with standardized metadata including sex, organ system, disease category, age group, and experimental setting.

## Dataset Construction

Transcript abundance profiles were obtained from the ARCHS4 human reference HDF5 file (`human_gene_v2.5.h5`). Genes were read in batches of 1,000 per chunk to reduce memory usage, with two known problematic gene index ranges (indices 56,000–58,000) excluded. Gene identifiers (Ensembl ID, gene symbol) and sample annotations (GSM accession, sample name, series ID) were extracted from the same HDF5 file and assembled into an [AnnData](https://anndata.readthedocs.io/) object.

Curated metadata tables were then merged by GSM accession to add standardized attributes. Duplicate GSM entries were removed (keeping first occurrence).

The script used to generate the AnnData files is provided at `data/generate_h5ad.py`.

## Files

| File | Description | Samples |
|------|-------------|---------|
| `data/human_gene_v2.5_train_nhmerged.h5ad` | Training split | ~436,000 |
| `data/human_gene_v2.5_test_nhmerged.h5ad` | Test split | ~24,500 |
| `data/human_gene_v2.5_test_2024_nhmerged.h5ad` | Held-out test split (2024 samples) | ~13,600 |
| `data/GEO_HUMAN_500K_training.csv` | Metadata for training split | ~436,000 |
| `data/GEO_HUMAN_500K_test.csv` | Metadata for test split | ~24,500 |
| `data/GEO_HUMAN_500K_test_2024.csv` | Metadata for held-out test split | ~13,600 |
| `data/generate_h5ad.py` | Script to reproduce AnnData files from source HDF5 | — |

## AnnData Structure

### `.obs` (sample-level metadata)

| Field | Description |
|-------|-------------|
| `geo_accession` | GSM accession ID |
| `sample` | Sample name |
| `series_id` | GSE series ID |
| `gender` | Sex (standardized) |
| `organ_system` | Organ/tissue system |
| `disease` | Disease label (`Normal` for healthy samples) |
| `age` | Age group |
| `Experimental_Setting` | Experimental context (e.g., `In Vitro`, `In Vivo`) |
| `Broad_Disease_Category` | Broad disease category (grouped from `disease`) |

### `.var` (gene-level metadata)

| Field | Description |
|-------|-------------|
| `ensembl_id` / `ensembl.gene` | Ensembl gene ID (index) |
| `symbol` | Gene symbol |

### `.X`

Raw transcript abundance matrix (samples × genes), sparse format. Gene coverage spans 65,186 genes (67,186 total minus the 2,000 excluded problematic indices).

## Metadata CSV Columns

| Column | Description |
|--------|-------------|
| `GSE_ID` | GEO Series accession |
| `GSM_ID` | GEO Sample accession |
| `year` | Year of submission |
| `contact_country` | Submitting country |
| `instrument_model` | Sequencing platform |
| `RNA_Library` | Library type (e.g., `mRNA-based`) |
| `Exp_Setting` | Experimental setting |
| `GSE_Pert` | Whether the series involves a perturbation |
| `Disease` | Disease label |
| `Organ` | Organ/tissue |
| `Gender` | Sex |
| `Age_Group` | Age group |

## Usage

```python
import anndata as ad

adata_train = ad.read_h5ad("data/human_gene_v2.5_train_nhmerged.h5ad")
adata_test  = ad.read_h5ad("data/human_gene_v2.5_test_nhmerged.h5ad")
adata_test_2024 = ad.read_h5ad("data/human_gene_v2.5_test_2024_nhmerged.h5ad")

print(adata_train)
# AnnData object with n_obs × n_vars = ~436000 × 65186
```

## Source

- ARCHS4: https://maayanlab.cloud/archs4/
- Raw HDF5 file: `human_gene_v2.5.h5` (available from ARCHS4)