A Multi‐complex Virtual Screening Approach Exploring Machine Learning to Identify Potential PI3Kγ Inhibitors

Phosphoinositide 3‐kinase gamma (PI3Kγ) plays a crucial role in regulating cellular physiological activities. However, the development of selective PI3Kγ inhibitors has proven particularly challenging due to the high sequence homology among residues near the ATP‐binding pocket across class I PI3K isoforms. Previous studies have demonstrated that an ensemble machine learning strategy can improve the success rate of virtual screening against multiple PI3Kγ proteins. In this study, we aimed to investigate the effects of different protein structures and varying protein quantities on the outcomes of virtual screening by constructing three distinct groups of multiple proteins. Three distinct virtual screening strategies utilizing Naïve Bayesian Classification (NBC), based on multicomplex molecular docking and pharmacophore modeling, were developed employing three different protein sets. Internal and external validation demonstrated that the NBC model, using multiple crystal structures of human PI3Kγ protein complexes bound to selective PI3Kγ inhibitors, exhibits superior predictive capability. Furthermore, the optimal model was employed to conduct virtual screening on the ChEMBL database, leading to the identification of several compounds with significant potential as PI3Kγ inhibitors. We anticipate that these findings will provide valuable insights and a robust computational framework for the design and optimization of novel PI3Kγ inhibitors.