A Novel Sequence-Structure Approach for Accurate Prediction of Resistance to HIV-1 Protease Inhibitors

Phenotypic tests are useful for measuring the degree of resistance of HIV-1 protease mutants to commercially available inhibitors. However, these tests are expensive and time-consuming, and phenotyping has been performed on only a fraction of the nearly 400 distinct patient-derived protease mutants. We employed a computational mutagenesis methodology, incorporating both sequence and structure information, to generate a feature vector representation for each of the isolated protease mutants. Training sets were prepared for seven protease inhibitors, each consisting of protease mutants with known phenotypes. Four machine-learning algorithms were implemented, and random forest performed best at distinguishing between sensitive/resistant mutants based on area under the ROC curve (0.81 -0.92). Trained models were used to make predictions about recently assayed protease mutants and displayed 83% agreement with the experimental data. The results suggest that susceptibility of unassayed protease mutants to each inhibitor can be reliably predicted.