Abstract
Broadly neutralizing antibodies (bNAbs) against HIV-1 are under evaluation for both prevention and therapy. HIV-1 sequence diversity observed in most HIV-infected individuals and archived variations in critical bNAb epitopes present a major challenge for the clinical application of bNAbs, as preexistent resistant viral strains can emerge, resulting in bNAb failure to control HIV. In order to identify viral resistance in patients prior to antibody therapy and to guide the selection of effective bNAb combination regimens, we developed what we believe to be a novel Bayesian machine-learning model that uses HIV-1 envelope protein sequences and foremost approximated glycan occupancy information as variables to quantitatively predict the half-maximal inhibitory concentrations (IC50) of 126 neutralizing antibodies against a variety of cross clade viruses. We then applied this model to peripheral blood mononuclear cell-derived proviral Env sequences from 25 HIV-1-infected individuals mapping the landscape of neutralization resistance within each individual's reservoir and determined the predicted ideal bNAb combination to achieve 100% neutralization at IC50 values <1 μg/ml. Furthermore, predicted cellular viral reservoir neutralization signatures of individuals before an analytical antiretroviral treatment interruption were consistent with the measured neutralization susceptibilities of the respective plasma rebound viruses, validating our model as a potentially novel tool to facilitate the advancement of bNAbs into the clinic.