General structure-activity relationship patterns for the ligands of cyclin-dependent kinases as tools for virtual screening of PubChem database

Cyclin-dependent kinases (CDKs) are a multifunctional family of enzymes, which can modify different protein substrates involved in the cell cycle progression (1). The identification of selective CDK inhibitors is very important for developing anticancer drugs. The main aim of this study is to find general structureselectivity relationship patterns for CDKs molecules and applying such rules for virtual screening purposes. We collected 12704 active and inactive ligands of CDK1, CDK2, CDK4, CDK5, and CDK9 from the Binding DB to achieve this goal. For every molecule, a broad range of molecular descriptors was computed, encompassing topological, constitutional, and both 2D and 3D descriptors by using DRAGON 5.5 software (5) and we used the variable importance in projection (VIP) approach for selecting discriminative molecular features (6). Certain key factors, including the hydrophobicity, the number of secondary amides (aliphatic), and C-043, have been identified as significant parameters in characterizing the inhibitory effects of CDK inhibitors. The C-043 is a common molecular descriptor between three active/inactive classification groups for CDK1, CDK4, and CDK9. This descriptor represents the number of X--CR..X fragments in a molecule and it is placed in the block of atom-centered descriptors. We used two machine learning methods for the classification of ligands using their activities including counter propagation artificial neural networks (CPANN) and supervised Kohonen networks (SKN). We downloaded two million random molecules from the PubChem database to evaluate multivariate classifiers for ligand-based virtual screening (7). The average values of the enrichment factor (EF10%) for the CPANN and SKN were 5.22% and 7.41%, respectively. In addition, the average values of the area under the receiver operating characteristic (ROC) curves were greater than 0.78 and 0.87 for the CPANN and SKN models, respectively. The study has found that it is possible to define and distinguish specific subspaces within chemical space that are relevant to CDK ligands. These subspaces have their own diagnostic boundaries encompassing all molecules and chemical compounds closely associated with CDK ligands. This knowledge can help guide the development of new CDK ligands with enhanced efficacy and selectivity.