Molecular docking and QSAR analysis of naphthyridone derivatives as ATAD2 bromodomain inhibitors: application of CoMFA, LS-SVM, and RBF neural network

摘要

In this research, molecular docking and QSAR models based on comparative molecular field analysis-like, least squares-support vector machine, and radial basis function neural network were used to investigate the interactions of naphthyridone derivatives with the binding site of ATAD2 and predict their activities. First molecular docking was used to investigate binding interactions between molecules with the greatest, the lowest and with medium activities and the binding site of ATAD2, then comparative molecular field analysis was used to model and predict their activities. Squared correlation coefficient (R (2)) for training and test sets of comparative molecular field analysis-like model and its leave-one-out cross validation (Q (2)) were 0.87, 0.83, and 0.78, respectively. The contributions of steric and electrostatic fields in the building of model were 49.64 and 50.36 %, respectively. Comparative molecular field analysis contour maps were extracted and interpreted to help the design of new molecules with greater activity. Principal component analysis was performed on comparative molecular field analysis descriptors and extracted scores were used as input variable to develop more reliable least squares-support vector machine and radial basis function neural network models. R (2) values for training and test sets of least squares-support vector machine were 0.82 and 0.84, respectively, and Q (2) parameter for its training set was 0.82. These results indicate least squares-support vector machine has slightly higher predictive power compared to the comparative molecular field analysis model. R (2) values for training and test sets of radial basis function neural network model were 0.89 and 0.90, respectively, and its squared correlation coefficient for leave-one-out cross validation was 0.87 that shows radial basis function neural network model has the greatest predictive power. All models have been validated with several statistical parameters and their applicability domains show all models were reliable.