Predicting the biological activities of triazole derivatives as SGLT2 inhibitors using multilayer perceptron neural network, support vector machine, and projection pursuit regression models

• Three nonlinear methods were used to study the pIC50 of triazole derivatives.
• Several validation techniques were used to demonstrate reliability.
• Relative significance of the descriptors to each model was studied.
• Interpretation of descriptors enhances insights on the pIC50 of triazole derivatives.

Quantitative structure–activity relationship (QSAR) studies were performed in this work to predict the pIC50 of non-glycoside sodium-dependent glucose cotransporter-2 (SGLT2) inhibitors (46 triazole derivatives). Four descriptors were selected from the pool of DRAGON descriptors using the enhanced replacement method. Three nonlinear regression methods—multilayer perceptron neural network (MLP NN), support vector machine (SVM), and projection pursuit regression (PPR)—were then used to build the QSAR models. The performance of the obtained models was assessed through the cross-validation and external validation of the test set. PPR produced a better model than MLP NN and SVM with coefficients of determination of 0.962 and 0.871 and root mean square errors of 0.162 and 0.471 for the training and test sets, respectively. This study developed simple and efficient approaches to predict the pIC50 of triazole derivatives and provided some insights into their structural features for the development of more potential SGLT2 inhibitors.