Machine-learning-guided library screening and drug resistance profiling of HIV-1 protease inhibitors.
TL;DR
Coupling ML-based large-scale prioritization with structure-based and dynamical validation can prospectively enrich for resistance-robust HIV-1 protease inhibitors and provide a generalizable framework for anticipating drug resistance in rapidly evolving viral targets.
Abstract
HIV-1 protease inhibitors are central to antiretroviral therapy, yet their long-term efficacy is undermined by the rapid emergence of drug-resistant variants. Although virtual screening of small-molecule libraries and resistance prediction have each been used to guide inhibitor design, there is still no integrated framework that jointly optimizes intrinsic potency and robustness to protease mutations at ultra-large chemical scale. Here, we present a multistage, machine learning (ML)-first hierarchical virtual screening pipeline that enables billion-scale prioritization of the ZINC20 chemical space while reserving computationally intensive structure-based analyses for the most promising candidates. Ensemble ML models for wild-type (WT) activity were first used to evaluate the full library (∼1.4 billion compounds) by ML inference, progressively reducing the chemical space to 19,911 candidates with predicted sub-nanomolar activity. These prioritized candidates were then assessed for cross-variant resistance using an LGBM-Chemprop drug-isolate fold-change (DIF) model, followed by molecular docking, MM/GBSA binding free energy calculations, all atom molecular dynamics (MD) simulations, and Neural Relational Inference (NRI) analysis on increasingly focused candidate subsets. This workflow prioritizes three computationally predicted mutation-tolerant candidates, of which ZINC000408994641 shows the most consistent in silico profile, with favorable or comparable predicted binding energetics and efficiency metrics relative to darunavir and tipranavir, and limited predicted loss of affinity across clinically relevant protease variants. These results demonstrate that coupling ML-based large-scale prioritization with structure-based and dynamical validation can prospectively enrich for resistance-robust HIV-1 protease inhibitors and provide a generalizable framework for anticipating drug resistance in rapidly evolving viral targets.