Unlocking the application potential of AlphaFold3-like approaches in virtual screening

Abstract

AlphaFold3 (AF3) has revolutionized the paradigm for protein-ligand complex structure prediction, yet its potential for structure-based virtual screening (VS) remains largely underexplored. Herein, we present a systematic assessment of AF3-like approaches for VS applications, using AF3, Protenix and Boltz-2 as representative models. Initial benchmarks on the well-established DEKOIS2.0 datasets demonstrate AF3’s exceptional screening capability, driven solely by its intrinsic confidence metrics for compound ranking. While third-party scoring schemes do not improve efficacy, both AF3 and Protenix prove robust as pose generators. Further analysis reveals performance declines in three more challenging cases: progressive exclusion of chemically similar active ligands from test sets, evaluation on a novel GPCR dataset with limited structural representation in model training, and assessment on a subset of LIT-PCBA dataset whose inactive compounds were experimentally verified. Despite these limitations, these models consistently surpass conventional docking tools in accuracy in most cases. Pose analysis further indicates that most predictions adopt physically plausible conformations, albeit with minor structural artifacts. This study highlights the promise and current constraints of AF3-like methods in VS, offering practical insights for their deployment in modern drug discovery.