YYJMAY/pmt-interaction

---
license: mit
task_categories:
- text-classification
tags:
- tcr
- tcr-pmhc
- peptide
- mhc
- immunology
- binding-prediction
- pmt
size_categories:
- 100K<n<1M
---

# PMT Benchmark Dataset

## Dataset Description

The PMT (Peptide-MHC-TCR) benchmark dataset for training and evaluating TCR-pMHC binding prediction models. This dataset contains TCR CDR3 sequences, peptide antigens, HLA alleles, and binary binding labels.

### Dataset Summary

This is the official PMT training and in-distribution (ID) test set from the SPRINT framework. The data has been cleaned, deduplicated, and standardized for reproducibility.

- **Training Set**: 474,881 samples
- **ID Test Set**: 4,564 samples
- **Task**: Binary classification (TCR-pMHC binding prediction)
- **Modality**: TCR CDR3 + Peptide + MHC (PMT task)

## Dataset Structure

### Data Files

- `train.csv`: Training data (474,881 samples)
- `id_test.csv`: In-distribution test data (4,564 samples)

### Data Format

CSV files with the following columns:

| Column | Type | Description |
|--------|------|-------------|
| CDR3 | string | TCR CDR3beta amino acid sequence |
| peptide | string | Peptide antigen sequence (8-15 aa) |
| HLA | string | HLA allele (standardized format: A*02:01) |
| label | int | Binding label (1=binder, 0=non-binder) |
| HLA_sequence | string | HLA pseudo-sequence (optional) |

### Dataset Statistics

#### Training Set

- **Total Samples**: 474,881
- **Positive Samples**: 33,129 (7.0%)
- **Negative Samples**: 441,752 (93.0%)
- **Unique HLAs**: 78
- **Unique Peptides**: 638
- **Unique TCRs**: 32,853

#### ID Test Set

- **Total Samples**: 4,564
- **Positive Samples**: 321 (7.0%)
- **Negative Samples**: 4,243 (93.0%)
- **Unique HLAs**: 12
- **Unique Peptides**: 190
- **Unique TCRs**: 1,283

## Usage

### Load with Hugging Face Datasets

```python
from datasets import load_dataset

# Load training data
dataset = load_dataset("YYJMAY/pmt-interaction", split="train")
train_df = dataset.to_pandas()

# Load test data
dataset = load_dataset("YYJMAY/pmt-interaction", split="test")
test_df = dataset.to_pandas()
```

### Load with Pandas

```python
import pandas as pd
from huggingface_hub import hf_hub_download

# Download training file
train_path = hf_hub_download(
repo_id="YYJMAY/pmt-interaction",
filename="train.csv",
repo_type="dataset"
)
train_df = pd.read_csv(train_path)

# Download test file
test_path = hf_hub_download(
repo_id="YYJMAY/pmt-interaction",
filename="id_test.csv",
repo_type="dataset"
)
test_df = pd.read_csv(test_path)
```

### Use with SPRINT Framework

The SPRINT framework automatically downloads and uses this dataset:

```bash
python scripts/run_benchmark.py --method METHOD --dataset pmt --mode train
python scripts/run_benchmark.py --method METHOD --dataset pmt --mode eval
```

## Data Quality

### Preprocessing

- **Deduplication**: All duplicate entries removed based on (CDR3, peptide, HLA, label)
- **HLA Standardization**: All HLA alleles normalized to standard format (e.g., A*02:01)
- **Missing Values**: No missing values in required columns
- **Label Validation**: All labels are binary (0 or 1)

### Peptide Length Distribution

Training set peptide lengths: 8-15 amino acids
Test set peptide lengths: 8-15 amino acids

## Construction

This dataset was curated and cleaned as part of the SPRINT benchmarking framework:

1. Collected from multiple public TCR-pMHC datasets
2. Standardized HLA allele naming conventions
3. Removed duplicates and incomplete entries
4. Split into training and in-distribution test sets
5. Validated for data quality and consistency

## Tasks

This dataset is designed for:

- **PMT (Peptide-MHC-TCR) Task**: Predict TCR-pMHC binding using all three components
- **Binary Classification**: Classify as binder (1) or non-binder (0)
- **Model Benchmarking**: Evaluate model performance on standardized data

## Limitations

- Only includes class I MHC (HLA-A, HLA-B, HLA-C)
- Limited to TCR CDR3beta sequences
- Binary labels (no binding affinity values)
- Peptide length range: 8-15 amino acids

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{pmt_benchmark_2024,
title={PMT Benchmark Dataset for TCR-pMHC Binding Prediction},
author={SPRINT Framework Contributors},
year={2024},
url={https://huggingface.co/datasets/YYJMAY/pmt-interaction}
}
```

## License

MIT License

## Contact

For questions or issues, please open an issue in the SPRINT repository.

## Related Datasets

- Allelic OOD: YYJMAY/allelic-ood
- Temporal OOD: YYJMAY/temporal-ood
- Modality OOD: YYJMAY/modality-ood