QSPR Modeling of Partition Coefficients and Henry’s Law Constants for 75 Chloronaphthalene Congeners by Means of Six Chemometric Approaches—A Comparative Study

Abstract

n -octanol/water and n-octanol/air partition coefficients were calculated for 75 chloronaphthalenes (CNs) by means of quantitative structure-property relationship (QSPR) strategy to fill significant lacks in the empirical data. The QSPR models based on quantum-chemical descriptors computed on the level of density functional theory using B3LYP functional and 6-311++G** basis set. For each property, six models were identified using chemometric approaches such as: multiple regression method, principal component regression, partial least square regression, partial least square regression with initial elimination of the uninformative variables, partial least square regression with variable selection by a genetic algorithm (GA-PLS), and neural networks with variable selection by a genetic algorithm (GA-NN). They were calibrated and validated using the experimentally measured values of logKOW available for 16 congeners and the values of logKOA existing for 43 congeners. The models were compared regarding to their complexity and prediction ability. For best predictive model logKOW values of 75 CNs varied from 3.93 to 6.68, while that of logKOA, from 5.93 to 11.64. Root mean square errors of prediction for the best (GA-NN) models were 0.065 and 0.091, respectively. Further, values of logKAW and KH of CNs were calculated based on predicted logKOW and logKOA data. Depending on the CN congener logKAW varied from −1.68 to −5.21 and that of KH from 0.02 to 51.24. The errors of partitioning data computed in this study were of the same order of magnitude as reported for experimentally derived partitioning data, which confirmed applicability of the proposed modeling scheme for successful determination of logKOW and KOA. Accordingly, a new procedure of the computational partitioning data generation based on partial least square regression with variable selection by a genetic algorithm (GA-PLS) and neural networks with variable selection by a genetic algorithm (GA-NN) was optimized and proposed for future use.