Assessment of AlphaFold structures and optimization methods for virtual screening

Abstract

Recent advancements in artificial intelligence such as AlphaFold, have enabled more accurate prediction of protein three-dimensional structure from amino acid sequences. This has attracted significant attention, especially for the application of AlphaFold in drug discovery. However, how to take full advantage of AlphaFold to assist with virtual screening remains elusive. We evaluate the AlphaFold structures of 51 selected targets from the DUD-E database in virtual screening. Our analyses show that the virtual screening performance of about 35% of the AlphaFold structures is equivalent to that of DUD-E structures, and about 25% of the AlphaFold structures yield better results than the DUD-E structures. Remarkably, AlphaFold structures produce slightly better results than the Apo structures. Moreover, we develop a new consensus scoring method based on Z-score standardization and exponential function, which shows improved screening performance compared to traditional scoring methods. By implementing a multi-stage virtual screening process and the new consensus scoring method, we are able to improve the speed of virtual screening by about nine times without compromising the enrichment factor. Overall, our results provide insights into the potential use of AlphaFold in drug discovery and highlight the value of consensus scoring and multi-stage virtual screening.