Protocol for Membrane Permeability Prediction of Cyclic Peptides Using Descriptors Obtained from Extended Ensemble Molecular Dynamics Simulations and Chemical Structures.

Improving the membrane permeability is crucial in cyclic peptide drug discovery. Molecular dynamics (MD) simulations are widely used to analyze membrane permeability but are computationally expensive. Machine learning offers a low-cost alternative, but its performance is limited by the size of available experimental data sets and the difficulty of incorporating peptide-specific three-dimensional (3D) structural information. Although the 3D conformations of cyclic peptides are closely associated with membrane permeability, they are highly sensitive to subtle changes in the chemical structure. Therefore, we developed a machine learning protocol that combines 3D descriptors derived from conformations obtained by MD simulations with 2D descriptors derived from cyclic peptide chemical structures, aiming to improve generalizability while reducing the simulation cost relative to the direct MD-based permeability prediction. We targeted 252 peptides across four data sets and calculated 3D descriptors from peptide conformations sampled outside the membrane, at the water/membrane interface, and in the membrane using replica exchange with solute tempering/replica exchange umbrella sampling simulations with 16 replicas. For machine learning, six different algorithms were used, ranging from simple methods, such as ridge regression, to more sophisticated methods, such as XGBoost. The best prediction performance was obtained using XGBoost, with Pearson's correlation coefficient