A comparative QSPR study on aqueous solubility of polycyclic aromatic hydrocarbons by GA-SVM, GA-RBFNN and GA-PLS

Abstract

A novel method to develop quantitative structure-property relationship (QSPR) models of organic contaminants was proposed based on genetic algorithm (GA) and support vector machine (SVM). GA was used to perform the variable selection and SVM was used to construct QSPR models. In this study, GA-SVM was applied to develop the QSPR model for aqueous solubility (S w , mol$L -1 ) of polycyclic aromatic hydrocarbons (PAHs). The R 2 (0.98) of the model developed by GA-SVM indicated a good predictive precision for lg S w values of PAHs. According to leave-one-out (LOO) cross validation, the results of GA-SVM were compared with those of genetic algorithm-radial based function neural network (GA-RBFNN) and genetic algorithm-partial leastsquares (GA-PLS) regression. The comparisons showed that the cross validation correlation coefficient ðQ 2 LOO ¼ 0:92Þ and root mean square error of LOO cross validation (RMSE LOO = 0.49) of GA-SVM were the highest and lowest, respectively, which illustrated that GA-SVM was more suitable to develop QSPR model for the lg S w values of PAHs than GA-RBFNN and GA-PLS.