Predicting the Impact of Drug Resistance Mutations on Inhibitor Potency with Molecular Dynamics and Machine Learning.

Many enzymes are vital drug targets in diseases such as cancer and pathogenic infections; however, mutations can drastically disrupt inhibitor binding to confer resistance. Resistance mutations primarily occur at the inhibitor binding site, but accompanying distal mutations can exacerbate resistance to render even highly potent inhibitors obsolete. While structure-based analysis is often critical to drug design, the enzyme-inhibitor complex structures provide little insight as to why the mutated enzyme variant is resistant. In this study we are using a very potent inhibitor, darunavir, bound to a series of 28 variants of HIV-1 protease with affinities from picomolar to micromolar, to develop an accurate method for predicting resistance. Our optimized strategy involves rerefining cocrystal structures to ensure accurate inhibitor geometries and combining parallel molecular dynamics simulations with feature selection to develop a robust machine learning model to predict loss of potency and resistance. The best performing models included only physics-based features of intramolecular interactions, with four specific features largely distal from the active site sufficient to predict binding affinity within 1 kcal/mol of the experimental value, far better than models based on either the structures or sequences alone. Thus, we demonstrate a strategy to robustly predict the loss of binding potency due to unseen drug-resistance mutations.