---
dataset_info:
  features:
    - name: sample_id
      dtype: string
    - name: tissue
      dtype: string
    - name: subject_id
      dtype: string
    - name: sex
      dtype: int64
    - name: age
      dtype: string
    - name: death_time
      dtype: string
    - name: estimated_age
      dtype: float64
  license: cc-by-4.0
  language:
    - en
  size_categories:
    - 1K<n<10K
  task_categories:
    - tabular-regression
    - tabular-classification
configs:
  - config_name: default
    data_files:
      - split: train
        path: "data/gtex_bulkformer_estimated_age_tpm.parquet"
---

# aging-expressions: Age-Stratified Gene Expression Dataset

## Dataset Description

This dataset provides **age-stratified gene expression data** derived from **ARCHS4** and **GTEx** databases, specifically curated for **fine-tuning BulkFormer** and other bulk RNA-seq deep learning models. The dataset contains TPM-normalized expression values for protein-coding genes, enriched with comprehensive demographic metadata including precise age information.

**Key Features:**
- 🎯 **Optimized for BulkFormer**: Uses the same ~20,000 protein-coding gene subset as BulkFormer training data
- 📊 **TPM Normalization**: Expression values normalized using BulkFormer gene lengths for consistency
- 👥 **Age-Filtered**: Only samples with numeric age in years (excludes gestational age, non-numeric values)
- 🧬 **Ensembl IDs**: Genes identified by stable Ensembl IDs for cross-dataset compatibility
- 🔬 **Rich Metadata**: Age, sex, tissue, disease, and treatment information extracted from sample characteristics

This dataset enables:
- Fine-tuning BulkFormer for age prediction tasks
- Training age-related gene expression models
- Comparative analysis across tissues and demographics
- Transfer learning for bulk RNA-seq applications

### Dataset Summary

**Sources:**
- **GTEx Analysis v10** (Release v10, RNASeQCv2.4.2): 9,662 samples across 54 tissue types
- **ARCHS4 v2.2+**: 111,000+ age-filtered samples across 62+ tissue types (human)

**Gene Selection:**
- **Total Genes**: ~20,000 protein-coding genes (BulkFormer subset)
- **Gene IDs**: Ensembl gene identifiers for consistency
- **Normalization**: TPM using BulkFormer gene length information
- **Source**: `bulkformer_gene_info.csv` containing curated protein-coding genes

**Sample Characteristics:**
- **Age Information**: Numeric age in years (filtered from ARCHS4 characteristics)
- **GTEx Age Range**: 20-79 years (6 age brackets: 20-29, 30-39, 40-49, 50-59, 60-69, 70-79)
- **ARCHS4 Age Range**: 1-122 years (actual numeric ages from sample metadata)
- **Tissues**: Blood, brain, liver, lung, skin, muscle, and 50+ other tissue types
- **Format**: Parquet files (optimized for fast loading with Polars/Pandas)

### Key Features

1. **Optimized for BulkFormer Fine-Tuning**:
   - Same ~20,000 protein-coding gene subset used in BulkFormer pre-training
   - TPM normalization using identical gene lengths as BulkFormer
   - Ensembl IDs matching BulkFormer's expected input format
   - Ready for fine-tuning age prediction, tissue classification, or disease detection models
   - Compatible with other bulk RNA-seq deep learning architectures

2. **Rich Demographic Metadata**:
   - **GTEx samples**: Sample/Subject IDs, tissue type, sex (1=male, 2=female), age brackets, estimated age, death circumstances
   - **ARCHS4 samples**: Sample IDs, numeric age in years, sex (male/female), tissue type, disease status, treatment information
   - All metadata automatically extracted and validated from sample characteristics
   - Age filtering ensures only samples with reliable numeric age values

3. **TPM-Normalized Expression**:
   - Raw counts converted to TPM (Transcripts Per Million) for comparability
   - Normalization formula: TPM = (counts / gene_length) * 1e6 / sum(counts / gene_length)
   - Uses BulkFormer's curated gene lengths for consistency
   - Ready for log1p transformation: `log(TPM + 1)` for deep learning

4. **Protein-Coding Gene Focus**:
   - Filtered to ~20,000 high-confidence protein-coding genes
   - Excludes pseudogenes, lncRNAs, and other non-coding RNAs
   - Gene set matches BulkFormer's training data for transfer learning
   - Ensembl IDs provide stable identifiers across datasets

5. **Quality Assurance**:
   - GTEx: Analysis v10 with rigorous QC from GTEx consortium
   - ARCHS4: Uniformly processed using STAR alignment and Kallisto quantification
   - Age validation: Excludes gestational age, age in weeks/months/days
   - Memory-optimized processing with 50% reduction in RAM usage

## Dataset Structure

### Data Files

| File Name | Rows | Columns | Size | Description |
|-----------|------|---------|------|-------------|
| `gtex_bulkformer_estimated_age_tpm.parquet` | 9,662 | 18,255 | ~350 MB | GTEx samples with TPM-normalized gene expression |

### Data Fields

#### Metadata Columns (7 columns)

| Column | Type | Description | Example Values |
|--------|------|-------------|----------------|
| `sample_id` | string | GTEx sample identifier | "GTEX-QMFR-1926-SM-32PL9" |
| `tissue` | string | Tissue type | "Blood", "Brain", "Muscle" |
| `subject_id` | string | GTEx subject/donor identifier | "GTEX-QMFR" |
| `sex` | int64 | Biological sex (1=male, 2=female) | 1, 2 |
| `age` | string | Age bracket from GTEx | "20-29", "50-59", "70-79" |
| `death_time` | string | Hardy scale death circumstances | "Ventilator Case", "Fast death" |
| `estimated_age` | float64 | Midpoint of age bracket | 24.5, 54.5, 74.5 |

#### Gene Expression Columns (18,248 columns)

- **Column Names**: Gene symbols (e.g., `TP53`, `BRCA1`, `APOE`)
- **Data Type**: float64
- **Values**: TPM-normalized gene expression (0 to ~1,000,000)
- **Transformation**: Ready for log1p transformation: `log(TPM + 1)`

### Data Splits

This dataset does not have predefined train/test splits. Users should create their own splits based on their research needs:

- **Age-stratified splits**: Ensure balanced age distribution across splits
- **Tissue-stratified splits**: For tissue-specific analyses
- **Subject-level splits**: Avoid data leakage by splitting at subject level (not sample level)

### Example Record

```python
{
  "sample_id": "GTEX-QCQG-0006-SM-5SI8M",
  "tissue": "Blood",
  "subject_id": "GTEX-QCQG",
  "sex": 2,
  "age": "50-59",
  "death_time": "Ventilator Case",
  "estimated_age": 54.5,
  "TP53": 125.34,
  "BRCA1": 12.56,
  "APOE": 234.78,
  # ... 18,245 more genes
}
```

## Data Processing Pipeline

This dataset is created from **ARCHS4** and **GTEx** raw data using the `aging-expressions` package. The processing pipeline extracts age information, normalizes expression to TPM using BulkFormer gene lengths, and filters to protein-coding genes for optimal compatibility with deep learning models.

### Source Data

**ARCHS4 Data** (for tissue-specific parquet files):
1. **ARCHS4 Human Gene Expression**: `human_gene_v2.latest.h5` (~50GB)
   - 67,000+ genes × 500,000+ samples
   - Raw counts from GEO/SRA uniformly processed with Kallisto
   - Source: https://maayanlab.cloud/archs4/

2. **ARCHS4 Sample Metadata**: Embedded in HDF5 file
   - Sample characteristics from GEO
   - Contains age, sex, tissue, disease, and treatment information
   - Extracted using metadata enrichment algorithms

**GTEx Data** (for consolidated parquet file):
1. **GTEx Expression Data**: `GTEx_Analysis_v10_RNASeQCv2.4.2_gene_tpm.gct.gz`
   - Raw TPM values from GTEx portal
   - 25,150 genes × 9,662 samples

2. **GTEx Phenotype Data**: `GTEx_Analysis_v10_Annotations_SubjectPhenotypesDS.txt`
   - Subject-level demographic information
   - Age brackets, sex, death circumstances

3. **GTEx Sample Attributes**: `GTEx_Analysis_v10_Annotations_SampleAttributesDS.txt`
   - Sample-level tissue annotations

**BulkFormer Gene Information**:
- **BulkFormer Gene Info**: `bulkformer_gene_info.csv`
  - ~20,000 protein-coding genes with Ensembl IDs
  - Gene lengths for TPM normalization
  - Source: BulkFormer model repository

### Processing Steps

The complete pipeline: **Download** → **Prepare** → **Upload**

```bash
# 1. Download raw data
uv run download archs4 human
uv run download gtex all
uv run download bulkformer data

# 2. Prepare TPM-normalized parquet files
uv run prepare archs4  # Creates tissue-specific files
uv run prepare gtex    # Creates consolidated file

# 3. Upload to HuggingFace (optional)
uv run upload
```

**Detailed Processing Steps:**

1. **Age Filtering** (ARCHS4):
   - Extract age from sample characteristics (`characteristics_ch1` field)
   - Filter to samples with numeric age in years (exclude gestational age, age in months/weeks)
   - Validate age range (1-122 years, oldest verified human)
   - Result: 111,000+ samples with reliable age information

2. **Metadata Extraction**:
   - **ARCHS4**: Extract age, sex, tissue, disease, treatment from GEO characteristics
   - **GTEx**: Join phenotype and sample attributes, compute estimated age from brackets
   - Normalize tissue names (e.g., "whole blood" → "blood", "pbmc" variations → "pbmc")
   - Encode categorical variables consistently

3. **Gene Filtering**:
   - Filter to genes present in BulkFormer gene list (~20,000 protein-coding genes)
   - Map gene symbols to Ensembl IDs using BulkFormer gene info
   - Remove genes without valid Ensembl IDs or gene length information
   - **ARCHS4**: 67,000 → ~20,000 genes retained
   - **GTEx**: 25,150 → ~18,000 genes retained

4. **TPM Normalization**:
   - **ARCHS4**: Convert raw counts to TPM using formula: `TPM = (counts / gene_length_kb) * 1e6 / sum(counts / gene_length_kb)`
   - **GTEx**: Data already in TPM format, validate consistency
   - Use BulkFormer gene lengths for normalization (identical to BulkFormer training)
   - Memory-optimized implementation (50% RAM reduction via inline calculations)

5. **Ensembl ID Mapping**:
   - Replace gene symbols with Ensembl IDs as column names
   - Ensures compatibility across datasets and model versions
   - Genes without Ensembl mappings are dropped with warnings

6. **Quality Control**:
   - Validate sample IDs match between expression and metadata
   - Check for missing values in critical columns
   - Ensure data type consistency (float64 for expression, proper types for metadata)
   - Tissue-level validation (minimum sample counts, unique tissue names)

7. **Output Generation**:
   - Save to Parquet format using streaming mode (memory efficient)
   - **ARCHS4**: One file per tissue (e.g., `blood_bulkformer_age_tpm.parquet`)
   - **GTEx**: Single consolidated file (`gtex_bulkformer_estimated_age_tpm.parquet`)
   - Lazy evaluation ensures minimal memory footprint during writes

### Age Bracket Mapping

| GTEx Age Range | Estimated Age | Count (approx) |
|----------------|---------------|----------------|
| 20-29 | 24.5 | ~500 |
| 30-39 | 34.5 | ~1,000 |
| 40-49 | 44.5 | ~1,800 |
| 50-59 | 54.5 | ~2,800 |
| 60-69 | 64.5 | ~2,500 |
| 70-79 | 74.5 | ~1,000 |

## Usage

### Loading the Dataset

#### Using Polars (Recommended)

```python
import polars as pl

# Load the full dataset
df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

print(f"Shape: {df.shape}")
print(f"Columns: {df.columns[:10]}")
```

#### Using Pandas

```python
import pandas as pd

# Load the full dataset
df = pd.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

print(f"Shape: {df.shape}")
print(df.head())
```

#### Using HuggingFace Datasets

```python
from datasets import load_dataset

dataset = load_dataset("longevity-genie/aging-expressions", split="train")
df = dataset.to_pandas()
```

### Example Analyses

#### 1. Age-Stratified Gene Expression

```python
import polars as pl

df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

# Analyze TP53 expression across age groups
age_analysis = df.group_by("age").agg([
    pl.count().alias("n_samples"),
    pl.col("TP53").mean().alias("tp53_mean"),
    pl.col("TP53").std().alias("tp53_std"),
    pl.col("BRCA1").mean().alias("brca1_mean"),
])

print(age_analysis.sort("age"))
```

#### 2. Tissue-Specific Expression

```python
import polars as pl

df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

# Get blood samples only
blood_samples = df.filter(pl.col("tissue") == "Blood")

# Analyze aging genes in blood
aging_genes = ["CDKN2A", "TP53", "TERT", "SIRT1", "FOXO3"]
blood_aging = blood_samples.select(["estimated_age", "sex"] + aging_genes)

print(blood_aging.describe())
```

#### 3. Sex-Stratified Analysis

```python
import polars as pl

df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

# Compare gene expression between sexes
sex_comparison = df.group_by("sex").agg([
    pl.count().alias("n_samples"),
    pl.col("XIST").mean().alias("xist_mean"),  # X-inactivation gene
    pl.col("DDX3Y").mean().alias("ddx3y_mean"),  # Y chromosome gene
])

print(sex_comparison)
```

#### 4. Prepare for Machine Learning

```python
import polars as pl
import numpy as np

df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

# Separate metadata from expression
metadata_cols = ["sample_id", "tissue", "subject_id", "sex", "age", "death_time", "estimated_age"]
gene_cols = [c for c in df.columns if c not in metadata_cols]

# Log-transform expression data
df_log = df.with_columns([
    (pl.col(gene).log1p().alias(gene)) for gene in gene_cols
])

# Create train/test split (stratified by age)
from sklearn.model_selection import train_test_split

train_df, test_df = train_test_split(
    df_log, 
    test_size=0.2, 
    stratify=df_log["age"], 
    random_state=42
)

print(f"Train: {len(train_df)}, Test: {len(test_df)}")
```

#### 5. Integration with BulkFormer

```python
import polars as pl
from bulkformer import BulkFormer, extract_features

df = pl.read_parquet("hf://datasets/longevity-genie/aging-expressions/data/gtex_bulkformer_estimated_age_tpm.parquet")

# Prepare expression matrix for BulkFormer
metadata_cols = ["sample_id", "tissue", "subject_id", "sex", "age", "death_time", "estimated_age"]
gene_cols = [c for c in df.columns if c not in metadata_cols]

# Extract expression matrix
expr_df = df.select(gene_cols)

# Use BulkFormer for feature extraction or downstream tasks
# (See BulkFormer documentation for details)
```

## Dataset Statistics

### Sample Distribution

- **Total Samples**: 9,662
- **Total Subjects**: ~800 (each subject contributes multiple tissue samples)
- **Sex Distribution**: 
  - Male (sex=1): ~60%
  - Female (sex=2): ~40%

### Tissue Coverage

54 GTEx tissues including:
- **Blood**: Whole blood samples
- **Brain**: Multiple brain regions (cortex, cerebellum, etc.)
- **Heart**: Atrial and ventricular tissue
- **Liver**: Hepatic tissue
- **Muscle**: Skeletal muscle
- **Adipose**: Subcutaneous and visceral fat
- **Skin**: Sun-exposed and not sun-exposed
- **And 47 more tissues**

### Gene Coverage

- **Total Genes**: 18,248 protein-coding genes
- **Gene ID Type**: Ensembl gene IDs (via gene symbols)
- **Expression Range**: 0 to ~1,000,000 TPM
- **Median Genes Detected per Sample**: ~15,000 (TPM > 0)

## Data Source & Citation

### GTEx Project

This dataset is derived from **ARCHS4** and the **GTEx (Genotype-Tissue Expression) Project**.

### ARCHS4

**ARCHS4** (All RNA-seq and ChIP-seq Sample and Signature Search) provides uniformly processed gene expression data from GEO and SRA.

**ARCHS4 Citation**:
```bibtex
@article{lachmann2018massive,
  title={Massive mining of publicly available RNA-seq data from human and mouse},
  author={Lachmann, Alexander and Torre, Denis and Keenan, Alexandra B and Jagodnik, Kathleen M and Lee, Hoyjin J and Wang, Lily and Silverstein, Moshe C and Ma'ayan, Avi},
  journal={Nature communications},
  volume={9},
  number={1},
  pages={1366},
  year={2018},
  publisher={Nature Publishing Group}
}
```

**ARCHS4 Portal**: https://maayanlab.cloud/archs4/

**Data Version**: ARCHS4 v2.2+ (continuously updated)

### GTEx

The **GTEx Project** is supported by the Common Fund of the Office of the Director of the National Institutes of Health, and by NCI, NHGRI, NHLBI, NIDA, NIMH, and NINDS.

**GTEx Citation**:
```bibtex
@article{gtex2020,
  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}
}
```

**GTEx Portal**: https://gtexportal.org/

**Data Version**: GTEx Analysis Release v10 (2022-06-06)

### aging-expressions Package

This dataset was processed using the `aging-expressions` Python library:

**GitHub**: https://github.com/longevity-genie/aging-expressions

**Citation**:
```bibtex
@software{aging_expressions,
  title={aging-expressions: Age-stratified gene expression analysis toolkit},
  author={Longevity Genie Team},
  year={2025},
  url={https://github.com/longevity-genie/aging-expressions}
}
```

### BulkFormer Model

Gene filtering and compatibility are based on the **BulkFormer** foundation model:

**BulkFormer Citation**:
```bibtex
@article{bulkformer2025,
  title={BulkFormer: A large-scale foundation model for human bulk transcriptomes},
  author={[Authors]},
  journal={bioRxiv},
  year={2025},
  doi={10.1101/2025.06.11.659222}
}
```

**GitHub**: https://github.com/your-org/BulkFormer

## License

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

### Conditions

- ✅ **Share**: Copy and redistribute the material in any medium or format
- ✅ **Adapt**: Remix, transform, and build upon the material for any purpose
- ✅ **Attribution**: You must give appropriate credit to both:
  - The GTEx Project (original data source)
  - The aging-expressions package (data processing)

### GTEx Data Use Agreement

Users of this dataset must also comply with the **GTEx Data Use Certification Agreement**:
- Data is for research purposes only
- Do not attempt to identify individual participants
- Acknowledge GTEx in publications

For full terms, see: https://gtexportal.org/home/dataUseAgreement

## Limitations & Considerations

### Age Limitations

- **Age Brackets**: GTEx provides age as 10-year ranges, not exact ages
- **Age Range**: Limited to 20-79 years (no samples < 20 or > 79)
- **Estimated Age**: Midpoints are used (e.g., 54.5 for "50-59"), which introduces uncertainty

### Sample Considerations

- **Post-mortem Tissue**: All GTEx samples are from deceased donors
- **Death Circumstances**: Variable (recorded in `death_time` column)
- **Tissue Quality**: Quality varies by death circumstances and preservation
- **Multiple Samples per Subject**: Same individual contributes multiple tissues (avoid subject leakage in train/test splits)

### Gene Expression Limitations

- **Gene Coverage**: 18,248 genes (not all human genes)
- **Gene Filtering**: Limited to BulkFormer gene set
- **TPM Normalization**: Assumes gene length and library size corrections are accurate
- **Batch Effects**: Potential batch effects across collection sites and dates

### Demographic Limitations

- **Sex Only**: No gender identity information
- **Limited Diversity**: GTEx v10 is primarily from donors of European ancestry
- **No Disease Status**: Donors are generally healthy (post-mortem collection)

## Ethical Considerations

### Privacy

- **De-identified Data**: All GTEx data is de-identified
- **No Protected Health Information (PHI)**: Sample IDs are not linkable to individuals
- **IRB Approved**: GTEx project has IRB approval from all participating sites

### Bias & Fairness

- **Demographic Bias**: Dataset skews toward European ancestry, male donors, and middle-aged individuals
- **Tissue Availability Bias**: Some tissues are underrepresented due to collection feasibility
- **Research Use Only**: Not suitable for clinical decision-making

### Responsible Use

- **Research Purposes**: This data is for research only, not clinical diagnostics
- **Model Validation**: Models trained on this data should be validated on independent cohorts
- **Transparency**: Report dataset characteristics when publishing results

## Updates & Maintenance

### Version History

- **v1.0** (2025-01): Initial release with GTEx v10 data
  - 9,662 samples, 18,248 genes
  - TPM-normalized expression
  - Age-stratified metadata

### Future Plans

- **ARCHS4 Integration**: Add ARCHS4 human and mouse samples
- **DEE2 Metadata**: Incorporate SRA metadata for cross-dataset analysis
- **Tissue Subsets**: Create tissue-specific subsets for faster loading
- **Age Prediction Features**: Add BulkFormer-derived age predictions

## Contact & Support

### Questions & Issues

- **GitHub Issues**: https://github.com/longevity-genie/aging-expressions/issues
- **Discussions**: https://github.com/longevity-genie/aging-expressions/discussions

### Contributing

Contributions are welcome! See the [GitHub repository](https://github.com/longevity-genie/aging-expressions) for:
- Bug reports
- Feature requests
- Documentation improvements
- Additional analyses

### Acknowledgments

This dataset was created by the **Longevity Genie** team as part of our mission to accelerate aging research through open data and reproducible analysis.

Special thanks to:
- **GTEx Consortium** for making this invaluable resource available
- **BulkFormer team** for the foundation model and gene annotations
- **Open-source community** for tools like Polars, Pandas, and HuggingFace

---

**Last Updated**: January 2025  
**Dataset Version**: 1.0  
**Curator**: Longevity Genie Team