Protocol for Membrane Permeability Prediction of Cyclic Peptides Using Descriptors Obtained from Extended Ensemble...

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Membrane biophysics

Background:

• Cyclic peptide drug discovery requires accurate prediction of membrane permeability.
• Molecular dynamics (MD) simulations are accurate but computationally expensive for permeability analysis.
• Existing machine learning models struggle with limited data and incorporating 3D structural information.

Purpose of the Study:

• To develop a cost-effective machine learning protocol for predicting cyclic peptide membrane permeability.
• To improve model generalizability by integrating 3D conformational and 2D chemical structure descriptors.
• To reduce the computational cost associated with direct MD-based permeability predictions.

Main Methods:

• Generated 3D descriptors from cyclic peptide conformations using replica exchange simulations (16 replicas).
• Combined 3D descriptors with 2D descriptors from chemical structures.
• Trained and evaluated six machine learning algorithms, including XGBoost, on 252 peptides across four datasets.

Main Results:

• The XGBoost model achieved the best performance with R = 0.77 and RMSE = 0.62.
• Key predictive descriptors included hydrophilicity/hydrophobicity, conformational changes, and flexibility.
• Validated generalization on an external dataset (24 peptides) yielding R = 0.76 and RMSE = 1.14.

Conclusions:

• Integrating MD-derived 3D descriptors with 2D descriptors improves machine learning model generalizability for cyclic peptide membrane permeability prediction.
• This hybrid approach offers a reduced-simulation-cost alternative to direct MD permeability predictions.
• The findings demonstrate the potential for accurate external data prediction using relatively small training sets.