YYJMAY/temporal-ood

Temporal OOD Dataset for TCR-pMHC Binding Prediction

Dataset Description

The Temporal OOD (Out-of-Distribution) Dataset evaluates TCR-pMHC binding prediction models under temporal shift. This dataset contains SARS-CoV-2 T cell receptor sequences collected during the COVID-19 pandemic, providing a natural test of model generalization to time-lagged data.

Key Features

• Temporal Shift Testing: Data collected after training set construction
• COVID-19 Focus: SARS-CoV-2 T cell repertoire from pandemic
• Complete PMT Data: All samples include CDR3, peptide, and HLA
• Multi-Laboratory: Compiled from multiple research laboratories
• Experimentally Validated: All TCR-pMHC interactions experimentally annotated

Dataset Details

Construction Method

This dataset follows the out-of-date testing protocol introduced in FusionPMT, using the external SARS-CoV-2 repertoire from VDJdb's recent update. The sequences were:

1. Collected during COVID-19 pandemic by multiple laboratories
2. Experimentally annotated for peptide and HLA specificities
3. Filtered using standard quality control rules
4. Deduplicated to remove redundant entries
5. Validated for completeness

Statistics

• Total Samples: 13979
• Positive Samples: 1272 (9.1%)
• Negative Samples: 12707 (90.9%)
• Unique TCR Sequences: N/A
• Unique Peptide Epitopes: 239
• Unique HLA Alleles: 48

Data Format

CSV file with the following columns:

Column	Type	Description	Required
CDR3	string	TCR CDR3beta amino acid sequence	Yes
peptide	string	Peptide amino acid sequence	Yes
HLA	string	HLA allele (e.g., A*02:01)	Yes
label	int	Binding label (1=binder, 0=non-binder)	Yes
HLA_sequence	string	HLA pseudo-sequence	Optional

Peptide Length Distribution

• 7aa: 54 samples (0.4%)\n- 8aa: 1016 samples (7.3%)\n- 9aa: 9287 samples (66.4%)\n- 10aa: 2662 samples (19.0%)\n- 11aa: 729 samples (5.2%)\n- 12aa: 137 samples (1.0%)\n- 13aa: 42 samples (0.3%)\n- 20aa: 52 samples (0.4%)

Label Distribution

• Binders (label=1): 1272 samples
• Non-binders (label=0): 12707 samples
• Imbalance Ratio: ~1:10.0 (positive:negative)

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("YYJMAY/temporal-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/temporal-ood",
    filename="temporal_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/temporal-ood',
    'files': ['temporal_ood.csv'],
    'test': 'temporal_ood.csv'
}

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

Scientific Context

Temporal Shift Challenge

This dataset addresses a critical challenge in machine learning for immunology: temporal generalization. Models trained on historical data must generalize to new sequences collected at later time points. The COVID-19 pandemic provides a unique natural experiment for this evaluation.

Why Temporal OOD Matters

1. Real-world Deployment: Clinical applications require models that work on future data
2. Emerging Pathogens: New disease outbreaks generate novel epitopes
3. Distribution Drift: Immune repertoires evolve over time
4. Model Robustness: Tests whether models learn fundamental biology vs. dataset artifacts

Biological Significance

• SARS-CoV-2 Epitopes: Includes key viral peptides recognized by T cells
• Pandemic Diversity: Represents diverse patient populations and disease stages
• Laboratory Consensus: Multi-laboratory validation increases reliability
• Clinical Relevance: Direct connection to COVID-19 immune response research

Task Compatibility

• PMT Task: Yes (all samples have CDR3)
• PM Task: Yes (peptide-HLA pairs available)

All 13979 entries are suitable for TCR-peptide-MHC (PMT) binding prediction.

Quality Control

Filtering Rules Applied

1. Removed entries with missing CDR3, peptide, or HLA
2. Removed duplicate (CDR3, peptide, HLA, label) combinations
3. Validated label values (0 or 1 only)
4. Checked for empty strings in critical columns
5. Verified HLA sequence availability

Data Integrity

• No Missing Values: All required columns complete
• No Duplicates: 659 duplicates removed during preprocessing
• Valid Labels: All labels are binary (0 or 1)
• Standardized Format: Consistent with other SPRINT datasets

Comparison with Training Data

This dataset intentionally differs from training data in temporal dimension:

Aspect	Training Data	Temporal OOD
Collection Period	Pre-pandemic	During COVID-19 pandemic
Epitope Source	Various pathogens	SARS-CoV-2 dominant
Data Vintage	Historical	Recent/contemporary
Distribution	Established	Time-shifted

Benchmark Results

This dataset is used to evaluate multiple TCR-pMHC binding prediction methods in the SPRINT benchmark suite. Expected performance characteristics:

• Difficulty: Moderate to challenging due to temporal shift
• Baseline: Random classifier ~9.1% (positive class frequency)
• Evaluation Metrics: AUC, AUPR, F1, Precision, Recall

Citation

If you use this dataset, please cite:

@dataset{temporal_ood_2024,
  title={Temporal OOD Dataset for TCR-pMHC Binding Prediction},
  author={SPRINT Framework Contributors},
  year={2024},
  note={SARS-CoV-2 T cell repertoire data from VDJdb},
  url={https://huggingface.co/datasets/YYJMAY/temporal-ood}
}

And the original VDJdb paper:

@article{goncharov2022vdjdb,
  title={VDJdb in 2019: database extension, new analysis infrastructure and a T-cell receptor motif compendium},
  author={Goncharov, Mikhail and others},
  journal={Nucleic Acids Research},
  year={2022}
}

Related Datasets

• Allelic OOD: Tests generalization to rare HLA alleles
• Modality OOD: Tests cross-modality generalization (BA ↔ EL)
• FusionPMT Training: Original training data

Limitations

1. COVID-19 Bias: Heavy emphasis on SARS-CoV-2 epitopes
2. Temporal Specificity: Limited to pandemic time period
3. Imbalanced Labels: Negative samples dominate (~87%)
4. HLA Coverage: 48 alleles, may not cover all population diversity

License

MIT License - Free for academic and commercial use

Contact

For questions, issues, or contributions:

• Dataset repository: YYJMAY/temporal-ood
• SPRINT framework: https://github.com/Computational-Machine-Intelligence/SPRINT

Acknowledgments

• VDJdb team for SARS-CoV-2 repertoire data
• Multiple laboratories contributing T cell sequences during COVID-19 pandemic
• FusionPMT authors for temporal OOD protocol design