CycPeptMP

About

• Python implementation of CycPeptMP.
• CycPeptMP is an accurate and efficient model for predicting the membrane permeability of cyclic peptides.
• We designed features for cyclic peptides at the atom, monomer, and peptide levels to concurrently capture both the local sequence variations and global conformational changes in cyclic peptides. We also applied data augmentation techniques at three scales to enhance model training efficiency.

Requirements

• Python: 3.9.6
• Numpy: 1.25.0
• Pandas: 1.4.4
• Pytorch: 2.0.0 (CUDA: 11.7)
• RDKit: 2022.09.5
• Mordred: 1.2.0
• MOE: 2019.01 (commercial software)

Dataset

• Original cyclic peptide structure (SMILES) and experimentally determined membrane permeability (LogPexp) used in this study (data/CycPeptMPDB_Peptide_All.csv) were all sourced from CycPeptMPDB.
  • Li J., Yanagisawa K., Sugita M., Fujie T., Ohue M., and Akiyama Y. CycPeptMPDB: A Comprehensive Database of Membrane Permeability of Cyclic Peptides, Journal of Chemical Information and Modeling, 63(7): 2240–2250, 2023.
• Correspondence table of peptides and their constituent monomers is summarized in data/monomer_table.csv.
• Data used in this experiment, with duplicates removed (all: 7,451->7,337, PAMPA: 6,941->6,889), is summarized in desc/peptide_used.csv.
• Dataset split index is stored in data/eval_index/.

  *_ID.npy shows the CycPeptMPDB peptide ID, and *_index.npy shows the index in sorted desc/peptide_used.csv.

Input files

• Since the input file size is large, the input files for the train set and validation set have not been uploaded.
• Preparing to upload the complete input files to Zenodo (April 2025).

Code

• Testset.ipynb

  Prediction for the test set (and other assay data) shown in the paper by CycPeptMP and other baselines. Please download complete input files model/input/Trans/60/ from Google Drive.

• Newdata.ipynb

  Prediction for new data.

• Train.ipynb

  Re-searching hyperparameters/training models. Please refer to the process of Newdata.ipynb for the generation of input files.

Pre-trained weights

• Weights of CycPeptMP (60 times augmentation) for three validation runs (Fusion-60_cv*.cpt).
• Weights of fusion model with no augmentation (Fusion-1_cv*.cpt) and 20 times augmentation (Fusion-20_cv*.cpt) for three validation runs in ablation studies.

Reference

• Li J., Yanagisawa K., and Akiyama Y. CycPeptMP: Enhancing Membrane Permeability Prediction of Cyclic Peptides with Multi-Level Molecular Features and Data Augmentation, Briefings in Bioinformatics, 2024, 25(5), bbae417.

Contact

• Jianan Li: [email protected]