How good are AlphaFold models for docking-based virtual screening?

A crucial component in structure-based drug discovery is the availability of high-quality three-dimensional structures of the protein target. In molecular docking, whenever experimental structures were not available, in silico structural homology modeling has been the method of choice. However, using computationally predicted structures adds a further degree of uncertainty to the docking process. Recently, AlphaFold (AF), an artificial intelligence-based modeling tool, has shown impressive results in terms of model accuracy within the field of ab initio protein structure prediction. This outstanding success prompted us to evaluate how accurate AF models are from the perspective of docking-based drug discovery. We compared the performance of AF models in high-throughput docking (HTD) to their corresponding experimental PDB structures using a benchmark set of 16 targets spanning different protein families and binding site properties. Four docking programs and two consensus techniques were used to evaluate the HTD performance. The AF models showed consistently worse performance than their corresponding PDB structures, with zero enrichment factor values in several cases. While AlphaFold shows a remarkable ability to predict protein architecture and binding site anatomy, we conclude that this is not enough to guarantee that AF models can be reliably used for HTD purposes. Moreover, we show that very small variations at the side chain level of essential ligand-binding residues have a large impact on the outcome of HTD, what suggests that post-modeling refinement strategies might be key to increase the chance of success of AF models in prospective HTD campaigns.