Global and local computational models for aqueous solubility prediction of drug-like molecules

The aim of this study was to develop in silico protocols for the prediction of aqueous drug solubility. For this purpose, high quality solubility data of 85 drug-like compounds covering the total drug-like space as identified with the ChemGPS methodology were used. Two-dimensional molecular descriptors describing electron distribution, lipophilicity, flexibility, and size were calculated by Molconn-Z and Selma. Global minimum energy conformers were obtained by Monte Carlo simulations in MacroModel and three-dimensional descriptors of molecular surface area properties were calculated by Marea. PLS models were obtained by use of training and test sets. Both a global drug solubility model (R(2) = 0.80, RMSE(te) = 0.83) and subset specific models (after dividing the 85 compounds into acids, bases, ampholytes, and nonproteolytes) were generated. Furthermore, the final models were successful in predicting the solubility values of external test sets taken from the literature. The results showed that homologous series and subsets can be predicted with high accuracy from easily comprehensible models, whereas consensus modeling might be needed to predict the aqueous drug solubility of datasets with large structural diversity.