A SAR and QSAR study on cyclin dependent kinase 4 inhibitors using machine learning methods

Abstract

Cyclin dependent kinase 4 (CDK4) is a promising target for cancer treatment, and developing new effective CDK4 inhibitors is of great significance in anticancer therapy. In this study, we conducted a structure activity relationship (SAR) study on 3018 CDK4 inhibitors. We applied four machine learning methods, which were Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Machine (SVM) and Deep Neural Network (DNN), to develop 18 classification models based on 3018 inhibitors (dataset 1), 18 classification models based on dataset 1 and decoys, and 24 quantitative structure–activity relationship (QSAR) models based on 1427 inhibitors (dataset 2). We obtained some optimal models. Based on dataset 1, Model A2, built by SVM and MACCS fingerprints, has a prediction accuracy (Q) of 92.68% and a Matthews correlation coefficient (MCC) of 0.874 for the test set. Based on dataset 1 and decoys, Model C2, built by SVM and MACCS fingerprints, has a Q of 98.5% and a MCC of 0.937 for the test set. Based on dataset 2, Model F7, built by SVM and MOE descriptors, has a coefficient of determination (R²) of 0.824 and a root mean squared error (RMSE) of 0.534 for the test set. For classification models, it was found that the more samples used for modelling, the more robust the models, and the better the performance of the models. Moreover, we clustered 3018 inhibitors into 12 subsets, and analysed their scaffolds and fragment features. It was found that 2-aminopyrimidine, pyridine, piperazine and cyclopentane were common scaffolds and fragments in highly active inhibitors. This study can provide guidance for the discovery and optimization of CDK4 inhibitor lead compounds.