Developing a Naïve Bayesian Classification Model with PI3Kγ structural features for virtual screening against PI3Kγ: Combining molecular docking and pharmacophore based on multiple PI3Kγ conformations

Abstract

Phosphatidylinositol 3-kinase gamma (PI3Kγ) plays a critical role in immune signaling, thus identifying PI3Kγ as a potential therapeutic target. However, developing selective PI3Kγ inhibitors is hampered by the highly conserved structure of the ATP-binding pocket. Focused effort would be needed to improve upon the γ-subtype selectivity of the inhibitors; therefore, in the present study, a naïve Bayesian classification (NBC) model with PI3Kγ structural features that integrates molecular docking and pharmacophore based on multiple PI3Kγ conformations was developed for virtual screening against PI3Kγ to find novel selective PI3Kγ inhibitors. First, the active PI3Kγ inhibitors/decoy dataset was used to prove whether molecular docking or pharmacophore, integrating multiple PI3Kγ conformations always has higher prediction accuracy than that of any single conformation. Second, both internal cross-validation and external prediction revealed that the NBC model combining molecular docking and pharmacophore could significantly improve the enrichment of active PI3Kγ inhibitors. Then, an analog dataset based on JN-PK1 (a reference compound) was constructed and submitted to virtual screening using the optimal NBC model. Finally, a novel inhibitor with higher PI3Kγ inhibitory activity than JN-PK1 was identified through a series of biological assays, showing both good accuracy and significant reliability of the NBC model with the PI3Kγ structural features. We hope that the developed virtual screening strategy will provide valuable guidance for the discovery of novel selective PI3Kγ inhibitors.

Keywords: Machine learning; Molecular docking; Naïve Bayesian classification; PI3Kγ; Pharmacophore; Selective inhibitor; Virtual screening.