The development of deep learning tools such as AlphaFold2 (AF2), which enable structural predictions of proteins with high confidence, provides a potential alternative or substitute for experimentally derived structures. AF2 was also envisioned to help identify new ligands through virtual screening. However, retrospective docking analysis comparing AF2-generated versus experimental structures revealed mixed results in predicting ligand-binding sites. Lyu et al. performed a prospective analysis to test the performance of AF2 with the σ2 receptor and 5-HT2A receptor as candidates. The AF2 models for these receptors were predicted before the experimental structures were reported. The two datasets showed good concordance, potentially removing biases that may have occurred in the retrospective analysis. The team performed large-scale docking screening using chemically diverse libraries ranging from 490 million to 1.6 billion molecules against the AF2-generated receptor model and the experimental structure. For both receptors, the hit rate, affinities and Ki values between AF2 and experimental screening campaigns were similar, despite the lack of similarities in chemical structures between the screens. Cryo-electron microscopy (cryo-EM) analysis of an AF2-specific hit for 5-HT2A confirmed the predicted docking interaction, showing that AF2 models could be used to identify new ligands. In addition, the 5-HT2A screen revealed that the AF2 model produced compounds with higher potency and selectivity relative to the cryo-EM structure. Although the generality of these findings to other proteins remains unclear, the findings from Lyu et al. support the potential utility of docking with AF2-derived structures for ligand discovery.
Original reference: Science 384, eadn6354 (2024)
This is a preview of subscription content, access via your institution
