YYJMAY/allelic-ood

Allelic OOD Dataset

Dataset Description

The Allelic OOD (Out-of-Distribution) dataset is designed to evaluate TCR-pMHC binding prediction models on rare and novel HLA alleles that are infrequent or completely absent in training data. This tests the model's ability to generalize to new HLA variants (allelic generalization).

Key Features

• Allelic OOD Testing: Contains samples with HLA alleles that appear at ≤1% frequency in both PM and PMT training sets (rare alleles), plus alleles completely absent from training (zero-shot alleles)
• Two OOD Types:
  • Rare Alleles: HLA alleles with ≤1% frequency in training (intersection of PM and PMT trains)
  • Zero-Shot Alleles: HLA alleles completely absent from both PM and PMT training sets
• Dual Task Support:
  • PMT Task: Samples with CDR3 sequences (~10% of data) for TCR-Peptide-MHC prediction
  • PM Task: All samples (100%) for Peptide-MHC prediction
• Peptide Length Control: All peptides filtered to 8-14 amino acids to isolate allelic OOD from length OOD
• Standardized HLA Format: All HLA alleles normalized to A*02:01 format
• Quality Controlled: Deduplicated and validated data

Dataset Structure

Data Format

CSV file with the following columns:

Column	Type	Description	Required
CDR3	string	TCR CDR3beta amino acid sequence	Optional (for PMT task)
peptide	string	Peptide amino acid sequence (8-14 aa)	Yes
HLA	string	HLA allele (e.g., A02:01, B07:02)	Yes
label	int	Binding label (1=binder, 0=non-binder)	Yes
ood_type	string	OOD category: 'rare' or 'zero'	Yes
HLA_sequence	string	HLA pseudo-sequence	Optional

Dataset Statistics

• Total Samples: 62205
• Positive Samples: 28167 (45.3%)
• Negative Samples: 34038 (54.7%)
• Unique HLA Alleles: 32
  • Rare alleles (≤1% in training): 21
  • Zero-shot alleles (absent in training): 11
• Samples with CDR3 (PMT): 6094 (9.8%)
• Samples without CDR3 (PM only): 56111 (90.2%)
• Peptide Length Range: 8-14 amino acids

HLA Allele Distribution

The dataset focuses on rare and zero-shot HLA alleles:

• B*18:01: 40311 samples
• A*32:01: 1965 samples
• B*14:02: 1871 samples
• B*38:01: 1850 samples
• C*16:01: 1734 samples
• B*27:01: 1626 samples
• B*35:08: 1465 samples
• A*02:07: 1311 samples
• C*14:02: 1300 samples
• C*01:02: 1237 samples

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("YYJMAY/allelic-ood")
df = dataset['train'].to_pandas()

Load with Pandas

import pandas as pd
from huggingface_hub import hf_hub_download

file_path = hf_hub_download(
    repo_id="YYJMAY/allelic-ood",
    filename="allelic_ood.csv",
    repo_type="dataset"
)
df = pd.read_csv(file_path)

Use with SPRINT Framework

from sprint.core.dataset_manager import DatasetManager

manager = DatasetManager()
dataset_config = {
    'hf_repo': 'YYJMAY/allelic-ood',
    'files': ['allelic_ood.csv'],
    'test': 'allelic_ood.csv'
}

files = manager.get_dataset('allelic_ood', dataset_config)
test_file = files['test']

Construction Method

This dataset was constructed by:

1. Training Set Analysis: Analyzing HLA frequency distribution in both PMT and PM training sets
2. Rare HLA Identification:
   • Rare: HLA alleles appearing at ≤1% frequency in both PM and PMT training sets
   • Zero-shot: HLA alleles completely absent from both training sets
3. Data Collection: Extracting samples with rare/zero-shot HLAs from multiple OOD test sources
4. Peptide Length Filtering: Restricting to 8-14 amino acids to isolate allelic shift from length shift
5. Format Standardization: Normalizing HLA naming conventions (HLA-A02:01 → A02:01)
6. Quality Control: Deduplication and validation

Citation

If you use this dataset, please cite:

@dataset{allelic_ood_2024,
  title={Allelic OOD Dataset for TCR-pMHC Binding Prediction},
  author={SPRINT Framework Contributors},
  year={2024},
  url={https://huggingface.co/datasets/YYJMAY/allelic-ood}
}

License

MIT License

Contact

For questions or issues, please open an issue on the dataset repository.