Optimal Selection of Support Vector Regression Parameters and Molecular Descriptors for Retention Indices Prediction

The quantitative structure-retention relationship (QSRR) was used for the prediction of retention indices of compounds in gas chromatography. 252 compounds containing boiling points (BP) was extracted from Molecular Operating Environment (MOE) database. After calculation of molecular descriptors of all compounds, genetic algorithm (GA) was used to select an optimal subset of the molecular descriptors. We investigated the predictive performance of four methods: GA on MLR (GA-MLR), the subset selected by GA-MLR was used to train SVR (GA-MLR-SVR), GA on SVR (GA-SVR) and GA on SVR with optimizing parameters (GA-SVR-Para). Twenty in-silicon experiments were conducted on each method. The experimental results show that the GA-SVR and GA-SVR-Para have better predictive performance with small variations. Among these four QSRR models, GA-SVR-Para achieved the best performance with a R2 > 0.98.