CycPeptPPB

The official implementation of the CycPeptPPB.
CycPeptPPB is a predictor of Plasma Protein Binding rate for cyclic peptide with high performance focusing on residue-level features and circularity.

Requirements

• Python: 3.7.9
• RDKit: 2020.03.1
• Chainer: 7.1.0

Code

• EXAMPLE.ipynb

  Jupyter notebook with an example of prediction (trained weight is required).

• cut_ring.py

  Divide the main chain of the cyclic peptide into substructures.
  The target bonds of division are amide bonds and disulfide bonds.

• generate_input.py

  Generate prediction model input feature map.
  In the paper, we used descriptors computed from MOE software, but since MOE is a commercial software, CycPeptPPB implementation on this site used descriptors computed by RDKit.
  MOE descriptors used for the model in the paper: logP(o/w), PEOE_VSA-1, logS.
  RDKit descriptors used in the CycPeptPPB implementation on this site: MolLogP, PEOE_VSA6, EState_VSA3.

• generate_model.py

  Generate prediction model.
  You need to add the trained weights file of the model such as "model_weight/model.npz".

• draw_saliency_2Dmol.py

  Draw a 2D molecular heatmap of Saliency Score.
  This function is only feasible when CyclicConv is not used (Baseline model & CycPeptPPB model 2).

• get_output.py

  Make a prediction.
  You can change the variables use_augmentation(=True) and use_CyclicConv(=False) to specify the model to use.

  • use_augmentation=False, use_CyclicConv=False → Baseline model (1DCNN)
  • use_augmentation=False, use_CyclicConv=True → CycPeptPPB model 1 (CyclicConv)
  • use_augmentation=True, use_CyclicConv=False → CycPeptPPB model 2 (Augmentated 1DCNN)
  • use_augmentation=True, use_CyclicConv=True → CycPeptPPB model 3 (Augmentated CyclicConv)

Pretrained weights

• Pretrained weights are not available.

Prediction performance

• Prediction accuracy of external test data (DrugBank dataset):
• MOE descriptors version we used in the paper:
  • Baseline model (1DCNN): MAE=6.55, R=0.89.
  • CycPeptPPB model 1 (CyclicConv): MAE=15.60, R=0.66.
  • CycPeptPPB model 2 (Augmentated 1DCNN): MAE=4.79, R=0.92.
  • CycPeptPPB model 3 (Augmentated CyclicConv): MAE=8.97, R=0.87.

Reference

If you find CycPeptPPB useful, please consider citing this publication;

• Li J, Yanagisawa K, Yoshikawa Y, Ohue M, Akiyama Y. Plasma protein binding prediction focusing on residue-level features and circularity of cyclic peptides by deep learning. Bioinformatics, 38(4): 1110-1117, 2022. doi:10.1093/bioinformatics/btab726

Contact

• Jianan Li: [email protected]