Retention prediction of adrenoreceptor agonists and antagonists on unmodified silica phase in hydrophilic interaction chromatography

The development of retention prediction models for adrenoreceptor agonists and antagonists chromatographed on an unmodified silica stationary phase under the hydrophilic interaction chromatographic (HILIC) mode at three pH conditions (3.0, 4.0 and 5.0) is described. The models were derived using multiple linear regression (MLR) and an artificial neural network (ANN) using the logarithm of the retention factor (log k) as the dependent variable. In addition to the effects of the solute-related variables (molecular descriptors), the percentage of acetonitrile (%ACN) was also used as a predictor to gauge the influence of the mobile phase on the retention behavior of the analytes. Using stepwise MLR, the retention behavior of the studied compounds at pH 3.0 were satisfactorily described by a four-predictor model; the predictors being the %ACN, the logarithm of the partition coefficient (log D), the number of hydrogen bond acceptors (HBA), and the magnitude of the dipole moment (DipolMag). In addition to these four predictors, the total absolute atomic charge (TAAC) was found to be a significant predictor of retention at pH 4.0 and 5.0. Among the five descriptors, %ACN had the strongest effect on the retention, as indicated by its higher standardized coefficient than those obtained for the other four predictors. The inclusion of these four predictors which are related to the molecular properties of the compounds (log D, HBA, DipolMag, and TAAC) suggested that hydrophilic interactions, hydrogen bonding and ionic interactions are possible mechanisms by which analytes are retained on the studied system. The reliability and predictive ability of the derived MLR equations were tested using cross-validation and a test set which was not used when fitting the model. The models derived from MLR produced adequate fits, as proven by the high R ² values obtained for all calibration and training sets (0.9497 and above), and their good predictive power, as indicated by the high cross-validated q ² (0.9465 and above) and high R ² (0.9305 and above) values obtained for the test sets. ANN prediction models were also derived using the predictors derived from MLR as inputs and log k as output. A comparison of the models derived from both ANN and MLR revealed that the trained ANNs showed better predictive abilities than the MLR models, as indicated by their higher R ² values and their lower root mean square error of predictions (RMSEP) for both training and test sets under all pH c