Background:  HIV drug resistance is conferred by >100 mutations at ~50 sites in reverse transcriptase (RT) and protease, which are the molecular targets of therapy. These mutations emerge in complex patterns and often have synergistic and antagonistic interactions. We applied 3 supervised learning methods to RT and protease genetic sequence data to predict reductions in HIV drug susceptibility.

Methods:  Susceptibility to 9 RT inhibitors were available for ~300 sequenced RT isolates and to 5 protease inhibitors for ~500 sequenced protease isolates. Susceptibility results indicated the fold-increase in the drug concentration required to inhibit a wild type virus isolate by 50%. Because of the high-dimensionality of the protease and RT sequence data, 4 alternate genotypic feature sets were explored:  6 protease and 6 RT mutations strongly associated with drug resistance based on in vitro experiments; 19 protease and 16 RT mutations accepted by most experts to be associated with drug resistance; 45 protease and 42 RT positions recently found to be statistically associated with treatment failure; and all 99 protease and all 240 RT positions. Decision trees (C4.5), neural networks, and support vector machines were trained to classify virus isolates as susceptible, low-level resistant, or high-level resistant to each of the 14 drugs.

Results:  A minimum of 19 protease and 16 RT mutations were required to obtain optimal predictions for each of the 14 drugs using each of the three learning methods. The median prediction accuracy into the 3 categories (susceptible, low-level resistant, and high-level resistant) was 69.5% (range 63% to 84%) for decision trees and 73% (range 64% to 88%) for neural networks, and 72% for support vector machines (64% to 92%). Discriminating between susceptible and low or high-level resistance was more accurate than discriminating between low-level and high-level resistance.

Conclusions:  This study demonstrates the applicability of three learning methods for using HIV genotypes to predict drug susceptibility phenotype. The study identifies subsets of mutations that capture the variability required for phenotypic prediction. Analyses are underway to compare the importance of individual mutations for classification by each of the methods.

Keywords: Genotype-Phenotype; Drug Resistance; Machine Learning