Machine learning-based quantitative structure–retention relationship models for predicting the retention indices of volatile organic pollutants

In this research, a dataset including 206 volatile organic compounds was used to develop quantitative structure–retention relationship models for predicting the retention indices of volatile organic compounds on DB-5 stationary phase. A total of 141 molecules were put in train set to build models and 65 molecules were put in test set to validate models, externally. By using stepwise-multiple linear regression, two descriptors including X1sol (solvation connectivity index chi-1) and AAC (mean information index on atomic composition) were selected to create linear and nonlinear quantitative structure–retention relationship models. Multiple linear regression, epsilon-support vector regression and deep learning-based artificial neural network were used as modeling techniques. All models were validated by calculating several statistical parameters for both train and test sets that show created models have high predictive power. R2 values for the test set of multiple linear regression, epsilon-support vector regression and deep learning-based artificial neural network models were 0.90, 0.94 and 0.94, respectively. Results show the Van der Waals interactions of molecules with methyl groups in DB-5 stationary phase and the electrostatic interactions of atoms with partial negative charge in molecules with the hydrogen atoms of phenyl groups in DB-5 stationary phase are responsible for the separation of volatile organic compounds in DB-5 stationary phase. Finally, these created models were used to predict the retention indices of 694 volatile organic compounds that had no retention index data on DB-5 stationary phase.