The Development Of Machine Learning-based Virtual Screening Of PI3Kγ Inhibitors

PI3Kγ（Phosphatidylinositol 3-kinase gamma）belongs to the IB subclass of the PI3K family.A large number of studies have shown that PI3Kγis highly expressed in the hematopoietic system,especially in leukocytes,which makes PI3Kγa potential drug target for the treatment of hematological tumors,immune,inflammation,and other related diseases.Therefore,the development of selective PI3Kγinhibitors has received increasing attention.Currently,PI3Kγinhibitors are mainly designed based on the ATP pocket,but the structures of the ATP pocket between different PI3K subtypes are highly conserved,which greatly hinders the development of selective PI3Kγinhibitors.Nowadays,some unique structural features of PI3Kγhave been gradually revealed,which brings some new ways of discovering novel PI3Kγinhibitors.Therefore,the technology which could integrate the structural features of PI3Kγwould help the development of PI3Kγinhibitors,among them,computer-aided drug design（CADD）has been attracted much interest because of its high efficiency and economic characteristics.Thus,in this present study,a multi-PI3Kγconformational virtual screening method based on machine learning was developed to find novel PI3Kγinhibitors at the molecular level.The main work and results of this thesis are as follows:（1）Appropriate protein structures and molecular docking programs are critical to ensure the accuracy of molecular docking-virtual screening.Therefore,some mainstream docking software were chosen to conduct molecular docking on the multiple PI3Kγproteins,and then,the software with the best“sampling power”and“screening power”was retained.Finally,naive Bayesian classification（NBC）models were built by combining the molecular docking results,and the predictive ability of the NBC models was estimated by the validation database.The final results show that CDOCKER and Glide are both appropriate docking programs for the PI3Kγsystem,and virtual screening integrating multiple PI3Kγconformations always has higher prediction accuracy than that of any single conformation.（2）Val882 has been demonstrated as a key structural feature of PI3Kγprotein,which always forms hydrogen bond interaction with PI3Kγinhibitors.Therefore,based on the research work of（1）,Val882 and its hydrogen bond interaction were introduced to the virtual screening model through molecular docking and pharmacophore,and a PI3Kγstructural feature automated-selected virtual screening strategy was finally constructed based on the NBC model.And then,an analogs database based on JN-PK1,which is a PI3Kγinhibitor identified through our previous study,was constructed and then submitted to the virtual screening by using the NBC model.After a series of biological experimental evaluations,we found a compound JN-S2,which has a higherγ-selective inhibitory effect(IC₅₀=2.45μM)than the reference JN-PK1(IC₅₀=7.49μM).The results overall showed that the NBC model with PI3Kγstructural features,which integrates molecular docking and pharmacophore based on multiple PI3Kγconformations,contains a reliable and efficient prediction ability to discover novel PI3Kγinhibitors.（3）The unique structural characteristics of PI3Kγinhibitors would help to further improve the accuracy of virtual screening against PI3Kγ.Here,the key“γ-selective”structural characteristics of the reported highly selective PI3Kγinhibitors were first obtained through the calculation of molecular fingerprints and molecular characteristics.And then,several virtual screening models were constructed through incorporating these structural features.Finally,the Drugbank library was used as a validation dataset to perform virtual screening,top-scored 20 small molecules were chosen for comparative analysis.The results suggest that the model integrating the structural features of PI3Kγand the inhibitors would significantly improve the virtual screening enrichment of active PI3Kγinhibitors(EF_1%=48.17).