Geno2pheno: Interpretable and accurate HIV antibody resistance prediction

Abstract

BackgroundAntiretroviral therapy (ART) is a life saving option for people living with HIV-1 (PLWH) and is effective against many viral strains. The most common ARTs involve combinations of drugs targeting viral or cellular proteins. Most of these drugs have to be taken daily. An alternative to ARTs with established inhibitors comprises broadly neutralizing antibodies (bNAbs). However, bNAbs share the problem of viral resistance with protein inhibitors. We developed a web service geno2pheno[bNAbs] that allows users to upload viral genotypes and estimates the respective resistance to many common bNAbs. The service uses trained statistical models to classify the virus into sensitive and resistant, respectively or to regress the IC50. MethodsWe used two linear models as well as two neural nets for each task and multi-task (MT) learning to train both models for IC50 prediction and classification simultaneously. During multi-task learning we penalize divergence of class and IC50 score in addition to the loss individual to each of the models. FindingsWe compared the linear models of geno2pheno[bNAbs] to other state-of-the-art methods like recurrent neural nets and self-attention, and found them to be competitive in regard to accuracy and have the benefit of fast computation and being easily interpretable in regard to features, i.e., positions on the envelope. InterpretationWe developed a web service for the prediction of antibody resistance (geno2pheno[bNAbs]) to HIV-1, which is free to use and can be extended to other viruses, like Sars-Cov2, in the future.