QSAR study of heparanase inhibitors activity using artificial neural networks and Levenberg–Marquardt algorithm

A linear and non-linear quantitative structure–activity relationship (QSAR) study is presented for modeling and predicting heparanase inhibitors' activity. A data set that consisted of 92 derivatives of 2,3-dihydro-1,3-dioxo-1H-isoindole-5-carboxylic acid, furanyl-1,3-thiazol-2-yl and benzoxazol-5-yl acetic acids is used in this study. Among a large number of descriptors, four parameters classified as physico-chemical, topological and electronic indices are chosen using stepwise multiple regression technique. The artificial neural networks (ANNs) model shows superiority over the multiple linear regressions (MLR) by accounting 87.9% of the variances of antiviral potency of the heparanase inhibitors. This paper focuses on investigating the role of weight update functions in developing ANNs. Levenberg–Marquardt (L–M) algorithm shows a better performance compared with basic back propagation (BBP) and conjugate gradient (CG) algorithms. The accuracy of 4-3-1 L–M ANN model was illustrated using leave-one-out (LOO), leave-multiple-out (LMO) cross-validations and Y-randomization. The mean effect of descriptors and sensitivity analysis show that log P is the most important parameter affecting the inhibitory behavior of the molecules.

A linear and non-linear quantitative structure–activity relationship (QSAR) study is presented for modeling and predicting heparanase inhibitors' activity. A data set that consisted of 92 derivatives of 2,3-dihydro-1,3-dioxo-1H-isoindole-5-carboxylic acid, furanyl-1,3-thiazol-2-yl and benzoxazol-5-yl acetic acids is used in this study. Among a large number of descriptors, four parameters classified as physico-chemical, topological and electronic indices are chosen using stepwise multiple regression technique. The artificial neural networks (ANNs) model shows superiority over the multiple linear regressions (MLR) by accounting 87.9% of the variances of antiviral potency of the heparanase inhibitors. This paper focuses on investigating the role of weight update functions in developing ANNs. Levenberg–Marquardt (L–M) algorithm shows a better performance compared with basic back propagation (BBP) and conjugate gradient (CG) algorithms. The accuracy of 4-3-1 L–M ANN model was illustrated using leave-one-out (LOO), leave-multiple-out (LMO) cross-validations and Y-randomization. The mean effect of descriptors and sensitivity analysis show that log P is the most important parameter affecting the inhibitory behavior of the molecules.Figure optionsDownload as PowerPoint slide