Quantitative predictions of gas chromatography retention indexes with support vector machines, radial basis neural networks and multiple linear regression

Support vector machines (SVM), radial basis function neural networks (RBFNN) and multiple linear regression (MLR) methods were used to investigate the correlation between GC retention indexes (RI) and physicochemical descriptors for both 174 and 132 diverse organic compounds. The correlation coefficient r2 between experimental and predicted retention index for training and test sets by SVM, RBFNN and MLR is 0.986, 0.976 and 0.971 (for 174 compounds), 0.986, 0.951 and 0.963 (for 132 compounds) respectively. The results show that non-linear SVM derives statistical models have similar prediction ability to those of RBFNN and MLR methods. This indicates that SVM can be used as an alternative modeling tool for quantitative structure–property/activity relationship (QSPR/QSAR) studies.