README.md

---
license: cc-by-nc-4.0
size_categories:
- 100K<n<1M
---

# Pep2Prob: Peptide-Specific Fragment Ion Probability Prediction Dataset

Pep2Prob is a comprehensive dataset designed to predict peptide-specific fragment ion probability in tandem mass spectrometry (MS/MS) based proteomics studies. This dataset addresses the limitations of conventional global statistical approaches by enabling the development of models that can predict fragmentation probabilities based on peptide sequence context.

## 📊 Dataset Overview and Highlights

- The dataset provides **Fragment ion probability statistics for 610,117 unique peptide precursors** derived from over 183 million high-resolution HCD spectra
- Diverse representation of precursors with varying lengths (6-40 amino acids) and charge states (1-8)
- High-quality annotations derived from validated peptide assignments with 0.1% false discovery rate
- A novel train-test split scheme is adapted to minimize structural similarity between entries in the training and the testing set.

## 🔗 Code and Documentation

- **GitHub Repository**: [https://github.com/Bandeira-Lab/pep2prob-benchmark](https://github.com/Bandeira-Lab/pep2prob-benchmark)
<!-- - **Paper**: [Pep2Prob Benchmark: Predicting Fragment Ion Probability for MS²-based Proteomics](https://openreview.net/forum?id=3mOtYJWr90) -->
- **Documentation**: Complete usage instructions and examples available in the GitHub repository

## ⚠️ Important Data Access Notice

**HuggingFace Statistics Issue**: The repository statistics incorrectly show ~3.15 million entries due to a platform counting error. HuggingFace sums across our five train-test splits (610,117 × 5 ≈ 3.15M) instead of recognizing these as the same unique precursors distributed into separate folds.

**Recommended Data Access Methods**:

1. **Using the downloader in our GitHub** (recommended):
```bash
# Download from GitHub
git clone https://github.com/Bandeira-Lab/pep2prob-benchmark.git
cd pep2prob-benchmark
pip install -r requirements.txt             (if missing any requried packages)
python data/download_data.py
```

2. **Manual download** for specific splits using `wget` with the URL.

## 📁 Dataset Structure

```
Pep2Prob/
├── data/pep2prob_dataset.parquet           # Info for all the precursors
├── meta_data/            
│   ├── X_columns.parquet                   # Input feature metadata
│   └── Y_columns.parquet                   # Target variable metadata
├── train_test_split_set_1/                 # Train-test split 1
│   ├── train_indices.parquet
│   └── test_indices.parquet
├── train_test_split_set_2/                 # Train-test split 2
│   ├── train_indices.parquet
│   └── test_indices.parquet
├── train_test_split_set_3/                 # Train-test split 3
│   ├── train_indices.parquet
│   └── test_indices.parquet
├── train_test_split_set_4/                 # Train-test split 4
│   ├── train_indices.parquet
│   └── test_indices.parquet
└── train_test_split_set_5/                 # Train-test split 5
    ├── train_indices.parquet
    └── test_indices.parquet
```

### File Descriptions

- **`pep2prob_dataset.csv`**: Main dataset containing fragment ion probability statistics for 610,117 unique peptide precursors, derived from over 183 million high-resolution HCD MS/MS spectra
- **`X_columns.csv`**: Metadata describing input features (peptide sequence information, amino acid properties, fragment ion types, physicochemical features)
- **`Y_columns.csv`**: Metadata describing target variables (probability values for different fragment ion types)
- **`train_test_split_set_X/`**: Five pre-defined train-test splits ensuring no sequence similarity between training and testing sets, preventing data leakage

## 🎯 Train-test split Methodology

Our dataset uses a sophisticated sequence-similarity-aware splitting strategy:

1. **Graph-based clustering**: Precursors are grouped based on sequence similarity (identical sequences, shared prefixes/suffixes of length 6)
2. **Component-based splitting**: Connected components are distributed across five balanced folds
3. **No data leakage**: Ensures similar peptides appear in only one split
4. **Robust evaluation**: Enables fair assessment of model generalization to novel peptide sequences

## 🔬 Applications

This dataset enables:

- **Model Development**: Training peptide-specific fragment ion probability prediction models
- **Benchmarking**: Standardized evaluation of machine learning approaches with varying complexity
- **Proteomics Enhancement**: Improving peptide identification algorithms and tools for library search, database search, mass spectrum prediction, de novo sequencing...
- **Fragmentation Research**: Exploring relationships between peptide sequence context and fragmentation behavior
- **Quality Control**: Identifying problematic spectra through unexpected fragmentation patterns

## 📖 Citation

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

```bibtex
@misc{xu2025pep2prob,
      title={Pep2Prob Benchmark: Predicting Fragment Ion Probability for MS$^2$-based Proteomics}, 
      author={Hao Xu and Zhichao Wang and Shengqi Sang and Pisit Wajanasara and Nuno Bandeira},
      year={2025},
      eprint={2508.21076},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2508.21076}, 
}
```

## 📄 License

This dataset is released under the CC-BY-NC-4.0 license. See LICENSE file for details.

## 🤝 Contributing

We welcome contributions! Please see our [GitHub repository](https://github.com/Bandeira-Lab/pep2prob-benchmark) for:
- Bug reports and feature requests
- Usage examples and tutorials
- Benchmark improvements and new baseline models
- Documentation enhancements

<!-- ## ⭐ Acknowledgments -->

<!-- This dataset was constructed from publicly available mass spectrometry data in the MassIVE repository, with curation based on the MassIVE Knowledge Base. We thank the proteomics community for sharing high-quality data that enables this research. -->