Prediction of HIV-1 drug susceptibility phenotype from the viral genotype using linear regression modeling.

ABSTRACT

Linear regression modeling on a database of HIV-1 genotypes and phenotypes was applied to predict the HIV-1 resistance phenotype from the viral genotype. In this approach, the phenotypic measurement is estimated as the weighted sum of the effects of individual mutations. Higher order interaction terms (mutation pairs) were included to account for synergistic and antagonistic effects between mutations. The most significant mutations and interactions identified by the linear regression models for 17 approved antiretroviral drugs are reported. Although linear regression modeling is a statistical data-driven technique focused on obtaining the best possible prediction, many of these mutations are also known resistance-associated mutations, indicating that the statistical models largely reflect well characterized biological phenomena. The performance of the models in predicting in vitro susceptibility phenotype and virologic response in treated patients is described. In addition to a high concordance with in vitro measured fold change, which was the primary aim of model design, the models per drug show good predictivity of therapy response for regimens including that drug, even in the absence of other clinically relevant factors such as background regimen.