Decoding Clade-Specific bNAb-HIV-1 Env Interactions Using Machine Learning

Human immunodeficiency virus type 1 (HIV-1) exhibits extensive genetic diversity, posing a major challenge to AIDS vaccine development. Broadly neutralizing antibodies (bNAbs) isolated from HIV-1-infected individuals offer valuable templates for designing broadly protective immunogens. Here, structural and energetic descriptors extracted from experimentally resolved bNAb/HIV-1 Envelope (Env) protein-protein complexes were used to develop machine-learning (ML) models predicting binding free energies ({triangleup}G) and dissociation constants (KD) across multiple HIV-1 clades. Among the algorithms evaluated, extreme gradient boosting (XGBoost) achieved the best predictive performance, with training and test R2 values of 0.92 and 0.658, respectively. Applying the optimized model to HIV-1 datasets revealed clade-dependent binding patterns, with clade B showing the greatest variability and clade G exhibiting signatures associated with reduced neutralization potency. Predicted {triangleup}G of different bNAb classes followed the trend gp120-gp41 interface {approx} fusion domain > CD4 binding site (CD4bs) > V1V2 loop {approx} V3-glycan, consistent with experimentally reported affinity profiles of benchmark bNAbs. Clade C exhibited intermediate {triangleup}G with PGT121, VRC01, and 10-1074 maintaining high-affinity interactions, whereas PG9 and 2G12 show reduced binding affinity. These findings offer quantitative insight into clade-specific bNAb binding energetics and demonstrate the potential of machine-learning approaches to guide HIV-1 immunogen design and vaccine development.