Retention prediction modeling of ginsenosides on a polyvinyl alcohol-bonded stationary phase at subambient temperatures using multiple linear regression and artificial neural network.

The development of retention prediction models for the seven ginsenosides (Rf, Rg1, Rd, Re, Rc, Rb2 and Rb1) on a polyvinyl alcohol (PVA)-bonded stationary phase at subambient temperatures is presented. The models were derived using multiple linear regression (MLR) and artificial neural network (ANN) using the logarithm of the retention factor (log k) as the dependent variable. Using stepwise MLR, the retention of the analytes under all temperature conditions was satisfactorily described by a three-predictor model; the predictors being the percentage of acetonitrile (%MeCN) in the mobile phase, the number of hydrogen bond donors (HBD) and the ovality (Ov) of the compounds. These predictors account for the contribution of the solute-related variables (HBD and Ov) and the influence of the mobile phase composition (%MeCN) on the retention behavior of the ginsenosides. The MLR models produced adequate fits, as proven by the high calibration R2 values of the predicted versus the observed log k (> 0.95) and good predictive properties, as indicated by the high cross-validated q2 (> 0.93) and high R2 (> 0.95) values obtained from the test set. ANN modeling was also conducted using the predictors that were derived from MLR as inputs and log k as the output. A comparison of the models derived from both MLR and ANN revealed that the trained ANNs showed better predictive abilities than the MLR models in all temperature conditions as demonstrated by their higher R2 values for both training and test sets and lower average percentage deviation of the predicted log k from the observed log k of the test compounds. The ANN models also showed excellent performance when applied to the prediction of the seven ginsenosides in different sample matrices.