Quantitative Prediction of Human Immunodeficiency Virus Drug Resistance

Abstract

Drug resistance of pathogens, including viruses, is one of the reasons for decreased efficacy of therapy. Considering the impact of HIV type 1 (HIV-1) on the development of progressive immune dysfunction and the rapid development of drug resistance, the analysis of HIV-1 resistance is of high significance. Currently, a substantial amount of data has been accumulated on HIV-1 drug resistance that can be used to build both qualitative and quantitative models of HIV-1 drug resistance. Quantitative models of drug resistance can enrich the information about the efficacy of a particular drug in the scheme of antiretroviral therapy. In our study, we investigated the possibility of developing models for quantitative prediction of HIV-1 resistance to eight protease inhibitors based on the analysis of amino acid sequences of HIV-1 protease for 900 virus variants. We developed random forest regression (RFR), support vector regression (SVR), and self-consistent regression (SCR) models using binary vectors containing values from 0 or 1, depending on the presence of a specific peptide fragment in each amino acid sequence as independent variables, while fold ratio, reflecting the level of resistance, was the predicted variable. The SVR and SCR models showed the highest predictive performances. The models built demonstrate reasonable performances for eight out of nine (R2 varied from 0.828 to 0.909) protease inhibitors, while R2 for predicting tipranavir fold ratio was lower (R2 was 0.642). We believe that the developed approach can be applied to evaluate drug resistance of molecular targets of other viruses where appropriate experimental data are available.

Keywords

• drug resistance
• viral infections
• HIV
• machine learning
• self-consistent regression