Estimation of aqueous solubility of organic compounds with QSPR approach

Purpose. To derive a QSPR model for estimation of aqueous solubility of organic compounds. Methods. Solubility data for 930 diverse compounds was investigated with principal component regression analysis. This set of compounds consists of pharmaceuticals, pollutants, nutrients, herbicides, and pesticides. The diversity of this collection was analyzed using MACCS fingerprint and BCUT chemistry space. Results. The training set of the solubility data is as diverse as the Available Chemicals Directory, and more diverse than the MDL Drug Data Report. Forty-six molecular descriptors were screened using a genetic algorithm. A QSPR model with a squared correlation coefficient (r(2)) of 0.92, a root mean square error of 0.53 log molar solubility (log S-w), an average absolute estimation error of 0.36 log S-w, and a cross-validated q(2) of 0.91 was derived. The QSPR model was validated with a test set of 249 compounds not included in the training set. The absolute estimation error for the test set of compounds was 0.39 log S-w. Conclusions. A highly predictive QSPR model for estimating aqueous solubility was derived and validated. This model can be used to estimate aqueous solubility for virtual screening and combinatorial library design.