AlphaFold2 structures guide prospective ligand discovery

Introduction

Structure-based docking is widely used in early-stage ligand discovery, typically relying on experimental protein structures from X-ray crystallography or cryo-electron microscopy (cryo-EM). Where experimental structures are unavailable, homology models have been used but performance can suffer when sequence identity to templates is low. Deep-learning methods such as AlphaFold2 (AF2) have produced highly accurate models at proteome scale, raising the possibility that AF2 structures could be useful for drug discovery. However, retrospective studies have reported mixed results for AF2 in ligand recognition: although AF2 models are globally accurate, small differences in binding-site side chains can impair retrospective enrichment of known ligands versus decoys. A. and colleagues set out to test prospectively whether unrefined AF2 models can support large-library docking and lead discovery on par with experimental structures. They focused on two pharmaceutically relevant membrane proteins—the σ2 receptor and the serotonin 2A (5-HT2A) G protein–coupled receptor—docked ultralarge libraries against both AF2 models and experimental structures, synthesised hundreds of top-ranked molecules from each campaign, and compared hit rates, affinities and functional profiles. The study also included retrospective docking analyses and cryo-EM determination of one AF2-derived ligand bound to 5-HT2A to evaluate how AF2-predicted conformations relate to experimentally observed ligand–receptor complexes.