FragWise

Overview

FragWise predicts the MS2 fragmentation setting that is most likely to improve downstream quantification for a given precursor. It provides a potential framework for integration with real-time search platforms, where peptide sequences identified during data acquisition could be used to guide precursor-specific fragmentation in real time.

The model predicts probabilities for 10 conditions:

• CID 22, 25, 28, 31, 34
• HCD 22, 25, 28, 31, 34

🐱 The setting with the highest probability is reported as the predicted best setting for each precursor.

Quick start

FragWise requires Python 3.12.

conda create -n fragwise python=3.12
conda activate fragwise
pip install -r requirements.txt

The trained 5-fold models are included in model/.

Run:

python scripts/predict.py \
    --input example/example_input.csv \
    --output example/example_output.csv

Input

CSV with one row per peptide precursor:

Column	Description
sequence	Peptide sequence. Oxidized Met: M*, M(ox), or M[+16].
charge	Precursor charge, e.g. 2, 3, 4.
peptide_id	Optional.

See example/example_input.csv.

Output

• predicted_best_setting — predicted best fragmentation setting
• predicted_best_prob — predicted probability
• P_<setting> — probabilities for all 10 settings
• clean_sequence — sequence used for feature calculation

Possible settings:

CID_22  CID_25  CID_28  CID_31  CID_34
HCD_22  HCD_25  HCD_28  HCD_31  HCD_34

See example/example_output.csv.

Important

TMTpro only. FragWise was trained on TMTpro-labeled precursors. Non-TMT and label-free data are not supported. Peptides with unrecognized modification notation are skipped with a warning. For reproducibility, use the versions in requirements.txt. In particular:

xgboost==3.2.0

Older XGBoost versions may produce different predictions.

Model

• Algorithm: XGBoost, 5-fold ensemble.
• Target: Predict the best MS2 fragmentation setting for each peptide precursor.
• Conditions: CID/HCD at NCE 22, 25, 28, 31, or 34.
• Features: 9 features from sequence and charge.
• Prediction: Probabilities from the 5 models are averaged, and the setting with the highest probability is reported.

Files

fragwise/
├── README.md              
├── requirements.txt       
├── LICENSE
├── model/
│   └── fragwise_fold0-4.json              5-fold XGBoost ensemble
├── scripts/
│   └── predict.py         reads CSV -> computes features -> predicts -> writes CSV
├── example/
│   ├── example_input.csv   10 example peptides                      
│   └── example_output.csv  
└── docs/
    └── workflow.md       

License

MIT — see LICENSE.