Zahra Rostami - Department of Chemistry, Payame Noor University (PNU), P. O. Box, ۱۹۳۹۵-۳۶۹۷ Tehran), Iran
Eslam Pourbasheer - Department of Chemistry, Payame Noor University (PNU), P.O. Box ۱۹۳۹۵-۳۶۹۷, Tehran, Iran

A comparative workflow, including linear and non-linear QSAR models, was carried out to evaluate the predictive accuracy of models and predict the inhibition activity of a series of aryl-substituted isobenzofuran-1(3H)-ones. The data set consisted of 34 compounds was classified into the training and test sets, randomly. Molecular descriptors were selected using the genetic algorithm (GA) as a feature selection tool. Various linear models based on multiple linear regression (MLR), principle component regression (PCR) and partial least square (PLS) and non-linear models based on artificial neural network (ANN), adaptive network-based fuzzy inference system (ANFIS) and support vector machine (SVM) methods were developed and compared. The accuracy of the models was studied by leave-one-out cross-validation (Q_LOO^2), Y-randomization test and group of compounds as external test set. Six descriptors were selected by GA to develop predictive models. With respect to the linear models, GA-PCR method was more accurate than the reset with statistical results of 〖 R〗_train^2=0.883, R_test^2=0.897,〖 R〗_(adj,train)^2=0.829,〖 R〗_(adj,test)^2=0.849,〖 F〗_train=24.07 and F_test=34.17. In case of non-linear models, GA-SVM (R_train^2=0.992 and R_test^2=0.997) showed high predictive accuracy for the inhibitory activity. It was found that the selected descriptors have the major roles in interpretation of biological activities of the compounds.

QSAR, genetic algorithms, global optimization, SVM