Learning patterns of HIV-1 resistance to broadly neutralizing antibodies with reduced subtype bias using multi-task learning

Abstract

The ability to predict HIV-1 resistance to broadly neutralizing antibodies (bnAbs) will increase bnAb therapeutic benefits. Machine learning is a powerful approach for such prediction. One challenge is that some HIV-1 subtypes in currently available training datasets are underrepresented, which likely affects models' generalizability across subtypes. A second challenge is that combinations of bnAbs are required to avoid the inevitable resistance to a single bnAb, and computationally determining optimal combinations of bnAbs is an unsolved problem. Recently, machine learning models trained using resistance outcomes for multiple antibodies at once, a strategy called multi-task learning (MTL), have been shown to improve predictions. We develop a new model and show that, beyond the boost in performance, MTL also helps address the previous two challenges. Specifically, we demonstrate empirically that MTL can mitigate bias from underrepresented subtypes, and that MTL allows the model to learn patterns of co-resistance to combinations of antibodies, thus providing tools to predict antibodies' epitopes and to potentially select optimal bnAb combinations. Our analyses, publicly available at https://github.com/iaime/LBUM, can be adapted to other infectious diseases that are treated with antibody therapy.