Purpose. A quantitative structure-property relationship (QSPR) was developed to predict drug solubility in binary mixtures of polyethylene glycol (PEG) 400 and water. The ability of the QSPR model to predict solubility was assessed and compared to the classic log-linear cosolvency model.
Methods. The solubility of 122 drugs, ranging in log P from -2.4 to 7.5, was determined in 0%, 25%, 50%, and 75% PEG (v/v in water) by the shake-flask method. Solubility data from 84 drugs were fit by linear regression using the following molecular descriptors: molecular weight, volume, radius of gyration, density, number of rotatable bonds, hydrogen-bond donors, and hydrogen-bond acceptors. The multiple linear regression model was optimized by a genetic algorithm guided selection method. The remaining 38 compounds were used to test the predictability of the model.
Results. QSPR-based models developed at each volume fraction with the training set compounds showed a reasonable correlation coefficient (r) of ∼0.9 and a root mean square (rms) error of <0.5 log unit. The model predicted solubility values of ∼78% of the testing set compounds within 1 log unit. The log-linear model was as effective as the QSPR-based model in predicting the testing set solubilities; however, many drugs, as expected, showed significant deviation from log-linearity.
Conclusions. The QSPR model requires only the chemical structure of the drug and has utility for guiding vehicle identification for early preclinical in vivo studies, especially when compound availability is limited and experimental data such as aqueous solubility and melting point are unknown. When experimental data are available, the log-linear model was verified to be a useful predictive tool.