Aqueous solubility prediction based on weighted atom type counts and solvent accessible surface areas.

In this work, four reliable aqueous solubility models, ASM-ATC (aqueous solubility model based on atom type counts), ASM-ATC-LOGP (aqueous solubility model based on atom type counts and ClogP as an additional descriptor), ASM-SAS (aqueous solubility model based on solvent accessible surface areas), and ASM-SAS-LOGP (aqueous solubility model based on solvent accessible surface areas and ClogP as an additional descriptor), have been developed for a diverse data set of 3664 compounds. All four models were extensively validated by various cross-validation tests, and encouraging predictability was achieved. ASM-ATC-LOGP, the best model, achieves leave-one-out correlation coefficient square (q2) and root-mean-square error (RMSE) of 0.832 and 0.840 logarithm unit, respectively. In a 10,000 times 85/15 cross-validation test, this model achieves the mean of q2 and RMSE being 0.832 and 0.841 logarithm unit, respectively. We believe that those robust models can serve as an important rule in druglikeness analysis and an efficient filter in prioritizing compound libraries prior to high throughput screenings (HTS).