In silico prediction of aqueous solubility: a multimodel protocol based on chemical similarity

Aqueous solubility is one of the most important ADMET properties to assess and to optimize during the drug discovery process. At present, accurate prediction of solubility remains very challenging and there is an important need of independent benchmarking of the existing in silico models such as to suggest solutions for their improvement. In this study, we developed a new protocol for improved solubility prediction by combining several existing models available in commercial or free software packages. We first performed an evaluation of ten in silico models for aqueous solubility prediction on several data sets in order to assess the reliability of the methods, and we proposed a new diverse data set of 150 molecules as relevant test set, SolDiv150. We developed a random forest protocol to evaluate the performance of different fingerprints for aqueous solubility prediction based on molecular structure similarity. Our protocol, called a "multimodel protocol", allows selecting the most accurate model for a compound of interest among the employed models or software packages, achieving r(2) of 0.84 when applied to SolDiv150. We also found that all models assessed here performed better on druglike molecules than on real drugs, thus additional improvement is needed in this direction. Overall, our approach enlarges the applicability domain as demonstrated by the more accurate results for solubility prediction obtained using our protocol in comparison to using individual models.