Comparative analysis and integration of HIV drug resistance mutation testing and prediction tools

Abstract

The pressure of antiretroviral therapy (ART) induces the selection of drug resistant variants in patients with human immunodeficiency virus 1 (HIV-1). Interpreting patient HIV-1 sequencing data allows personalization of treatment, avoiding clinical complications associated with drug resistance. Rules-based methods are the standard for drug resistance interpretation, although machine learning-based methods have also been developed as an alternative. Moreover, minority variants are relevant for the assessment of drug resistance, pointing to next-generation sequencing (NGS) techniques as the most appropriate technologies for this application given its low frequency detection limit. Only a limited number of pipelines are available for drug resistance interpretation from NGS data. Several improvements could thus be proposed with respect to the current pipelines. In this project, we compared the performance of seven different drug resistance interpretation methods on 22 ART drug datasets. We showed that the combination of rules-based HIVDB method, linear regression, and random forest in an ensemble approach achieved the best performance among all methods. Then, we proved the robustness of this ensemble method to incomplete sequence coverage for protease inhibitor (PI), nucleoside reverse transcriptase inhibitor (NRTI) and non-nucleoside reverse transcriptase inhibitor (NNRTI) drugs. Finally, we integrated these results into a standalone software tool to be incorporated into V-pipe, a software capable of analyzing genetically variable viral NGS samples. The incorporation into V-pipe paves the way for building an NGS-based pipeline for clinical and research purposes that outperforms the current state-of-the-art.