README.md

---
language:
- en
tags:
- biology
- immunology
- tcr
- peptide
- t-cell-receptor
- binding-prediction
task_categories:
- text-classification
size_categories:
- 10K<n<100K
---

# PT Interaction Dataset

## Dataset Description

The **PT (Peptide-TCR) interaction dataset** is designed for training and evaluating T-Cell Receptor (TCR) binding prediction models with full TCR sequence information. This dataset contains paired peptide sequences and complete TCR alpha/beta chain sequences (including all 6 CDR regions: A1-A3, B1-B3), along with binary binding labels.

### Key Features

- **Full TCR Information**: Contains all 6 CDR regions (A1, A2, A3, B1, B2, B3) for both alpha and beta chains
- **Binary Labels**: Binding labels (0=non-binder, 1=binder)
- **HLA Allele Information**: MHC allele context for each peptide-TCR pair
- **Peptide Length Range**: 8-12 amino acids
- **CDR3β Length Range**: 5-23 amino acids
- **Training Set**: 43,378 samples (13.62% positive, 86.38% negative)
- **Test Set**: 2,956 samples (13.97% positive, 86.03% negative)

### Dataset Statistics

| Split | Samples | Positives | Negatives | Unique TCRs | Unique HLAs |
|-------|---------|-----------|-----------|-------------|-------------|
| Train | 43,378 | 5,906 (13.62%) | 37,472 (86.38%) | 10,414 | 10 |
| ID Test | 2,956 | 413 (13.97%) | 2,543 (86.03%) | 2,511 | 10 |

### Data Format

Each row contains the following columns:

- `peptide`: Amino acid sequence of the peptide (8-12 aa)
- `A1`, `A2`, `A3`: CDR1α, CDR2α, CDR3α sequences
- `B1`, `B2`, `B3`: CDR1β, CDR2β, CDR3β sequences
- `binder`: Binary binding label (0=non-binder, 1=binder)
- `allele`: HLA allele (e.g., A*02:01, B*07:02)

### Example Data

```python
{
    "peptide": "KLGGALQAK",
    "A1": "SSVPPY",
    "A2": "YTSAATLV",
    "A3": "AVKWSSNYKLT",
    "B1": "SQVTM",
    "B2": "ANQGSEA",
    "B3": "SVGSGDHGEQF",
    "binder": 0,
    "allele": "A*03:01"
}
```

## Dataset Construction

### Data Sources

The PT dataset is curated from multiple publicly available TCR-peptide binding databases and experimental studies, including:
- VDJdb: A curated database of T-cell receptor sequences
- McPAS-TCR: Manually curated catalog of pathology-associated TCR sequences
- IEDB: Immune Epitope Database
- Published experimental validation studies

### Quality Control

1. **TCR Leakage Prevention**: Train and test splits are carefully constructed to ensure no TCR overlap based on CDR3β sequences
2. **Duplicate Removal**: All duplicate (peptide, B3, binder) combinations are removed
3. **Length Filtering**: Only peptides of length 8-12 amino acids are included
4. **HLA Standardization**: All HLA alleles follow the format "A*02:01" (without "HLA-" prefix)
5. **Data Validation**: All sequences are validated for amino acid composition

### Split Strategy

- **ID Test**: Random split preserving the same peptide/HLA/TCR distribution as training
- **No TCR Leakage**: Train and test sets are strictly disjoint based on CDR3β sequences

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("YYJMAY/pt-interaction")

# Access splits
train_data = dataset['train']
test_data = dataset['test']

# Convert to pandas DataFrame
import pandas as pd
train_df = pd.DataFrame(train_data)
test_df = pd.DataFrame(test_data)
```

### Training Example

```python
from datasets import load_dataset
import pandas as pd

# Load training data
dataset = load_dataset("YYJMAY/pt-interaction", split="train")
df = pd.DataFrame(dataset)

# Prepare features
X = df[['peptide', 'A1', 'A2', 'A3', 'B1', 'B2', 'B3', 'allele']]
y = df['binder']

# Train your model
# model.fit(X, y)
```

### Evaluation Example

```python
from datasets import load_dataset
import pandas as pd
from sklearn.metrics import roc_auc_score, accuracy_score

# Load test data
dataset = load_dataset("YYJMAY/pt-interaction", data_files="id_test.csv")
df = pd.DataFrame(dataset['train'])  # HF loads single files as 'train' split

# Make predictions
X_test = df[['peptide', 'A1', 'A2', 'A3', 'B1', 'B2', 'B3', 'allele']]
y_test = df['binder']

# predictions = model.predict(X_test)
# print(f"AUC: {roc_auc_score(y_test, predictions):.4f}")
# print(f"Accuracy: {accuracy_score(y_test, predictions > 0.5):.4f}")
```

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{pt_interaction_dataset,
  title={PT Interaction Dataset: Peptide-TCR Binding Prediction},
  author={SPRINT Benchmark Contributors},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/YYJMAY/pt-interaction}}
}
```

## Related Datasets

- **PM Dataset**: Peptide-MHC binding (no TCR information)
- **PMT Dataset**: Peptide-MHC-TCR with CDR3β only
- **Allelic OOD**: Out-of-distribution test for rare HLA alleles
- **Temporal OOD**: Out-of-distribution test for COVID-19 era data
- **Modality OOD**: Cross-modality generalization (BA vs EL)

## License

This dataset is released under the MIT License. The original data sources may have their own licenses.

## Contact

For questions or issues, please open an issue on the [SPRINT GitHub repository](https://github.com/Computational-Machine-Intelligence/SPRINT).

## Dataset Card Authors

SPRINT Benchmark Team

## Dataset Version

- **Version**: 1.0
- **Last Updated**: 2025-01-19