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

Improving membrane permeability is crucial in cyclic peptide drug discovery. Although the approach based on molecular dynamics (MD) simulation is widely used, it is computationally expensive. Alternatively, machine learning can predict membrane permeability at negligible cost, but it requires a larger dataset. There are only 7991 experimental values of membrane permeability available at the newly developed database. Another challenge in predicting membrane permeability using machine learning arises from the unique stable conformation of each cyclic peptide, which is strongly related to membrane permeability but difficult to predict from chemical structure. Therefore, we developed a machine learning protocol using 3D descriptors obtained from MD simulations in addition to 2D descriptors obtained from chemical structure of cyclic peptides, respectively, to generate a universal model with a realistic computational cost. We targeted 252 peptides across four datasets and, to calculate their 3D descriptors, predicted their conformation outside the membrane, at the water/membrane interface, and in the membrane by MD simulations based on the replica exchange with solute tempering/replica exchange umbrella sampling method using 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 a Pearsons correlation coefficient R = 0.77 and root mean square error (RMSE) = 0.62. The important descriptors included those that describe the hydrophilicity and hydrophobicity of the peptide, conformational differences between inside and outside the membrane, and the degree of freedom of the peptide. We confirm the models ability to predict the membrane permeability of peptides that differ in chemical structure from the training data by predicting the external data consisting of 24 peptides and obtained R = 0.76, RMSE = 1.14. Furthermore, we extracted one of the four datasets of the training data, re-trained the model, and performed the prediction of the permeability coefficients of the extracted dataset. The results showed the models generic nature with R = 0.61 and RMSE = 0.74, when using position-specific (PS) 3D, and 2D descriptors. In such situations descriptors based on conformations obtained from MD are essential for the prediction.