Combination of\nCoevolutionary Information and Supervised\nLearning Enables Generation of Cyclic Peptide Inhibitors with Enhanced\nPotency from a Small Data Set

Computational generation\nof cyclic peptide inhibitors using machine\nlearning models requires large size training data sets often difficult\nto generate experimentally. Here we demonstrated that sequential combination\nof Random Forest Regression with the pseudolikelihood maximization\nDirect Coupling Analysis method and Monte Carlo simulation can effectively\nenhance the design pipeline of cyclic peptide inhibitors of a tumor-associated\nprotease even for small experimental data sets. Further <i>in\nvitro</i> studies showed that such <i>in silico</i>-evolved cyclic peptides are more potent than the best peptide inhibitors\npreviously developed to this target. Crystal structure of the cyclic\npeptides in complex with the protease resembled those of protein complexes,\nwith large interaction surfaces, constrained peptide backbones, and\nmultiple inter- and intramolecular interactions, leading to good binding\naffinity and selectivity.