Integration of Multicomplex‐Based Pharmacophore Modeling and Molecular Docking in Machine Learning‐Based Virtual Screening: Toward the Discovery of Novel PI3K Inhibitors

Abstract The phosphatidylinositol‐3 kinase (PI3K) pathway is a crucial intracellular signaling pathway within living cells. The hyperactivation of PI3K signaling cascades is a common occurrence in human cancers, rendering PI3K a promising therapeutic target. Although several PI3K inhibitors are already available on the market, the adverse side effects of current therapies continue to highlight the necessity for the development of novel PI3K inhibitors. In this study, a virtual screening strategy employing naïve Bayesian classification (NBC) models, based on multicomplex‐based molecular docking and pharmacophore modeling, is developed. First, the docking accuracy and scoring reliability of four docking software are assessed, and Glide demonstrated higher predictability for PI3K inhibitors. Second, pharmacophore models are generated based on the current reported PI3K‐inhibitor interactions, and five pharmacophore hypotheses displayed significant capability in discriminating active PI3K molecules from inactive ones. Subsequently, three NBC models are constructed based on molecular docking and/or pharmacophore models, and the validation results showed that the NBC model, combining multicomplex‐based molecular docking and pharmacophore, significantly improved the hit rate of virtual screening against PI3K. Finally, the optimal NBC model is employed for virtual screening against the ChEMBL database, leading to the identification of multiple molecules with high potential as active PI3K inhibitors.