Combination of least absolute shrinkage and selection operator with Bayesian Regularization artificial neural network (LASSO-BR-ANN) for QSAR studies using functional group and molecular docking mixed descriptors

Abstract

A combination of least absolute shrinkage and selection operator (LASSO) with Bayesian Regularization feed-forward artificial neural network (LASSO-BR-ANN) was used as a new approach in the quantitative structure-activity relationship (QSAR) studies. A mixture of the docking derived descriptors with the simple functional group (structural) features was also introduced as a new ensemble of descriptors for accurate QSAR modeling. The performance of introduced approaches was tested with QSAR modeling of the biological activities (pEC50) of 73 azine derivatives as new non-nucleoside reverse transcriptase inhibitors (NNRTIs) for treatment of HIV disease. Molecular docking descriptors (MDDs) were generated from ligand-receptor interactions and functional group features derived using Dragon 5.5 software. The dataset was divided into three sets of training, validation, and test data. LASSO, as a penalized regression method, was applied to the training data set for the selection of the most relevant descriptors among the mixture of the structural and MDDs. LASSO selected descriptors were used as inputs in the construction of the Bayesian Regularization artificial neural network (BR-ANN) model. The results showed that the addition of functional group properties to the MDDs improves the accuracy of the model. Under the optimum conditions, LASSO-BR-ANN was successfully applied for the prediction of PEC50 values for compounds in the external test set with mean square error (MSE) and coefficient of determination (R2) values of 0.07 and 0.88, respectively. Some of the prediction statistical parameters of the model were calculated and all of them were in their acceptable ranges, which confirm the validity of the proposed QSAR model.