Artificial intelligence has transformed protein structure prediction, with AlphaFold2 (AF2) generating models with near-experimental accuracy. However, as AF2 was trained to generate a single structural model, this method does not capture the conformational flexibility of proteins, limiting its utility for structure-based drug design. Here, we explore strategies to generate diverse ensembles of binding-site models suitable for structure-based virtual screening against G protein-coupled receptors (GPCRs), a major class of drug targets. Our AFsample2T approach uses multiple sequence alignment column masking in the receptor binding site to reduce coevolutionary signals, leading to greater structural heterogeneity in the generated models. We demonstrate that ensembles of AFsample2T models capture multiple relevant binding-site conformations and reproduce experimentally observed conformational variability. Evaluation of AF2-based models in structure-based virtual screening using docking of actives and decoys shows that considering ensembles of diverse binding-site models substantially improves ligand enrichment. Our results provide guidelines for using AF2-based models in structure-based ligand discovery for GPCRs, and the AFsample2T approach can readily be applied to other protein classes.
Mitjavila-Domènech et al. (Mon,) studied this question.