Deciphering Complex Mechanisms of Resistance and Loss of Potency through Coupled Molecular Dynamics and Machine Learning

Drug resistance threatens many critical therapeutics through mutations in the drug target. The molecular mechanisms by which combinations of mutations, especially those remote from the active site, alter drug binding to confer resistance are poorly understood and thus difficult to counteract. A machine learning strategy was developed that coupled parallel molecular dynamics simulations with experimental potency to identify specific conserved mechanisms underlying resistance. Physical features were extracted from the simulations, analyzed, and integrated into one consistent and interpretable elastic network model. To rigorously test this strategy, HIV-1 protease variants with diverse mutations were used, with potencies ranging from picomolar to micromolar to the drug darunavir. Feature reduction resulted in a model with four specific features that predicts for both the training and test sets inhibitor binding free energy within 1 kcal/mol of the experimental value over this entire range of potency. These predictive features are physically interpretable, as they vary specifically with affinity and diagonally transverse across the protease homodimer. This physics-based strategy of parallel molecular dynamics and machine learning captures mechanisms by which complex combinations of mutations confer resistance and identify critical features that serve as bellwethers of affinity, which will be critical in future drug design.