Discovery Of CDK2 Inhibitors Based On Machine Learning And Molecular Dynamics Simulations

Aberrations in the cell cycle are among the hallmarks of cellular pathology.Among these,dysregulation of proliferation occurs when cells undergo mitosis mechanically,leading to abnormalities in the cell cycle.Therefore,cancer treatment strategies targeting the cell cycle have been highly attractive.Cyclin-dependent kinases(CDKs)are key regulatory enzymes that control cell cycle transitions and division,and thus,CDKs have long been considered potential targets for cancer therapy.Recent evidence suggests that inhibiting CDK2 activity may be feasible for treating certain tumor diseases such as neuroblastoma and colon cancer.CDK2 drives cells into the S(DNA synthesis)and M(mitosis)phases of the cell cycle.While the role of CDK2 in tumorigenesis is controversial,the development of CDK2 inhibitors has become a major focus of cancer research due to increasing reports confirming its impact on cancer.Currently,with no drugs targeting CDK2,so the development of CDK2 inhibitors is necessary.Therefore,the aim of this study was to explore potential inhibitors of CDK2 by developing a classification model for CDK2 inhibitors using a large activity database and machine learning algorithms.Based on indicator analysis(AUC,ACC,SE,SP,etc.),extreme gradient boosting tree model(XGBoost)based on circular fingerprint(ECFP6),random forest(RF)model,and attentive fingerprint(Attentive FP)based on graph neural network algorithm showed good predictive performance in the CDK2 dataset.In order to explain the reliability of the model prediction results,the Shapley Additive Explanations(SHAP)algorithm was used to interpret the model.Through the interpretation of the model,some active fingerprint fragments,such as fragments at positions 1476,1292,and 1582,were identified in the CDK2 dataset,which can provide guidance for the optimization and modification of lead compounds.After comparing the predictive performance and generalization ability of different models,we selected the XGBoost model to screen an unknown active database,which showed good stability in predicting the activity and inactivity of compounds.We screened 1152 novel compounds from a 50,000 dataset in the Enamine database using the machine learning model and sorted their affinity in CDK2 by molecular docking and scoring function.We found that all the screened compounds had docking scores in the CDK2 ligand below-7.0 kcal/mol.To avoid the single structure of the compounds selected by docking score screening,we used fingerprint clustering to divide the compounds into four categories and selected one compound with a higher docking score from each category.Then,the four compounds were subjected to drug-likeness analysis and molecular dynamics simulation to verify their feasibility in drug development,their interaction in the receptor,and binding free energy.The four potential CDK2 inhibitors(Z1766368563,Z363564868,Z1891240670,and Z2701273053)selected through screening have been analyzed for their drug-likeness and have shown good drug-like properties.In addition,molecular dynamics simulations have demonstrated that these compounds have high binding free energies,indicating that they can stably bind to CDK2 and serve as lead compounds for further optimization and modification.