| چکیده انگلیسی مقاله |
Cyclin-dependent kinases (CDKs) are key regulators of the cell cycle that provides domains essential for enzymatic activity. To date, many kinase inhibitors have been approved for cancer treatment. Due to wide range of activities and functions for all kinases, there is a vital need for development of selective kinase inhibitors for targeting only some specific types of them for effective treatments. In the present contribution, we seek to find selectivity in structural patterns of inhibitor molecules for three important isoforms of kinases [2, 3]. In this regard, ~5086 molecules were loaded from the Binding database for CDK1, CDK2 and CDK5 enzymes. Referring to various articles and sources, the ranges of bioactivity (IC50) for different receptors have been determined and the molecules were divided into two active and inactive categories in each class [4]. A wide range of molecular descriptors including, topological, constitutional, 2D- and 3D- descriptors were calculated for each molecule. The datasets were randomly split into 70% calibration (training) and 30% test sets. The best set of molecular descriptors were selected using variable importance in projection (VIP) algorithm. Counter propagation artificial neural network (CPANN), k-nearest neighbor (KNN), partial least squares-discriminant analysis (PLS-DA), supervised Kohonen networks (SKN), and support vector machine (SVM), were applied for modeling the activity of the molecules using the VIP-selected set of molecules. The performances of the models were evaluated by standard metrics derived from the confusion matrix and the values of sensitivity, specificity, precision and non-error rate were calculated for training and test sets. The accuracy values which refer to the ratio of correctly classified compounds were calculated to measure the overall performance of classifiers. The values of prediction accuracy in the test set for SKN and SVM methods were more than 80%. All of the optimized classification models represented high statistical quality and predictive ability with accuracy greater than 78% for the test sets. The high accuracy values of the obtained classifiers for the training and test sets demonstrate that the information provided is reliable for describing and predicting the activity of CDK inhibitors. It helps for designing molecules with better therapeutic potency and reduced side effects. The results in this work suggest some important molecular features that help medicinal chemists to develop selective inhibitors for different isoforms of kinases [2, 3]. |
