Binding Affinity of HIV-1 Protease Inhibitors: Insights from Machine Learning Models of Crystallographic Structures

HIV-1 protease inhibitors are major players among the antiviral therapies used against acquired immunodeficiency syndrome (AIDS). As such, the crystallographic coordinates of hundreds of HIV-1 protease complexes with miscellaneous ligands have been resolved and uploaded to the Protein Data Bank. Herein, machine learning (ML) models were constructed based on the crystallographic coordinates of 291 HIV-1 protease inhibitor complexes and >2500 molecular descriptors and subsequently used to predict the binding affinity of 274 additional HIV-1 protease–inhibitor complexes for which inhibition constants were not measured. The resulting accuracy scores of the ML models were >0.85. Three ML models, each based on 8–9 features, were analyzed in detail. These models were based on KBest with Random Forest, Recursive Feature Elimination with Random Forest, and Sequential Feature Selection with Support Vector Machine. The strength of our ML models lies in their ability to capture the chemical essence of key factors that influence binding. These include HIV-1 protease–inhibitor properties related to charge distribution, ability to form hydrogen bonds, and three-dimensional topology. Additionally, our models consider important properties of HIV-1 protease, including the symmetry of its active site and mutations in its flaps. The findings of the study highlight the contribution of a comprehensive analysis of accumulated experimental data to enhance the structural understanding of this important molecular system.