YYJMAY/pmt-interaction

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

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

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:

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:

@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