| Abstract | We report on a novel approach, combining a knowledge-based statistical contact potential with machine learning tools, for developing accurate predictive models of HIV-1 protease (PR) functional changes due to single or multiple amino acid replacements. Motivation stems from the nature of structure-function relationships, suggesting that accurate in silico evaluation of HIV-1 PR structural changes upon mutation correlate with experimentally measured changes in activity and susceptibility to inhibitors. Delaunay tessellation, a well-established computational geometry procedure, is applied to a diverse collection of protein structures in order to derive a four-body statistical potential. This potential is then used to empirically calculate residue compatibility scores at all HIV-1 PR positions. A subsequently formulated computational mutagenesis methodology quantifies the relative environmental change from wild type at every position in an HIV-1 PR mutant, which provide the components of a feature vector representation for the mutant. For activity change, a dataset of 536 HIV-1 PR single missense mutants, experimentally categorized by R. Swanstrom et al. based on their degree of pol processing capability relative to the wild type protein, is used to train and evaluate the performance of neural network, decision tree, support vector machine, and random forest supervised classification predictive models. Multiple datasets are prepared in the case of HIV-1 PR inhibitor susceptibility, each based on the experimentally measured degree of resistance of HIV-1 PR mutants to seven commercially available protease inhibitor medications. The HIV-1 PR mutants were isolated and sequenced from patients enrolled in clinical trials and consist of single as well as multiple residue replacements, and fold levels of resistance are available from the online Stanford University HIV Drug Resistance Database. Individual regression (actual fold level values) and supervised classification (resistance categories) predictive mo dels are trained for each protease inhibitor. |
|---|
