中文 | English

PEP2CCS

How to Integrate Non-Sequence Features for Peptide Collision Cross Section Prediction with Deep Learning?, IEEE BIBM 2024. DOI: 10.1109/BIBM62325.2024.10822507.
Zhimeng Tian, Zizheng Nie, Yong Zhang, Daming Zhu, Xuefeng Cui · xfcui@email.sdu.edu.cn

Installation

Conda and Python 3.9 are required. Dependencies are pinned in environment.yml, including PyTorch 2.2.2.

bash run_setup.sh
conda activate PEP2CCS

Prediction

bash run_prediction.sh data/examples/test_data.csv
python scripts/predict.py --data /path/to/peptides.csv --output results/mine.csv --device auto

Results go to results/predict.csv. When CCS is present, the terminal reports MSE and the median absolute relative deviation, (CCS / Pred CCS) * 100 - 100. Input needs Sequence, Charge, and m/z. Charges are 2, 3, or 4; sequences use the standard amino acids and fit in 64 positions. See the data schema.

Released weights are attached to release v1.0.0. The first run downloads the archive it needs and extracts it under checkpoints/.

Reproducing the experiments

python scripts/run_experiment.py exp1   # exp2 … exp5

Metrics are written to results/<experiment>/metrics.csv. Paper numbers use the full example table. --limit 128 --device cpu only checks that a run starts.

Experiment	Comparison
Exp1	Primary model
Exp2	Primary model, a second Transformer checkpoint, and a 3-layer bidirectional LSTM
Exp3	Injecting m/z or length into the charge token
Exp4	Amino-acid composition, an alphabetical vocabulary, and a second set of physical coefficients
Exp5	Seven pooling operators

Details are in the experiment notes, data/README.en.md, and checkpoints/README.en.md.

Training

scripts/train.py follows configs/model.yaml: Adam, learning rate ×0.1 at epoch 200, a 1% validation split with seed 1115, MSE on log CCS, and a stop after 15 epochs without improvement.

python scripts/train.py --train-csv /path/to/train.csv --output checkpoints/rerun.pt

Layout

configs/        hyperparameters, physical coefficients, experiment configs
data/examples/  example test table
checkpoints/    extracted weights (downloaded from the release)
src/pep2ccs/    data, model, training, evaluation
scripts/        predict, train, run experiments
docs/           data schema and experiment notes

PYTHONPATH=src python -m unittest discover -s tests -v

Citation

@inproceedings{tian2024pep2ccs,
  author    = {Tian, Zhimeng and Nie, Zizheng and Zhang, Yong and Zhu, Daming and Cui, Xuefeng},
  title     = {How to Integrate Non-Sequence Features for Peptide Collision Cross Section Prediction with Deep Learning?},
  booktitle = {IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  year      = {2024},
  doi       = {10.1109/BIBM62325.2024.10822507},
  url       = {https://ieeexplore.ieee.org/abstract/document/10822507}
}

MIT License