Data-Driven Critical Evaluation of the General Solubility Equation

Solubility is a widely measured physicochemical property and a highly sought-after descriptor in the evaluation of molecules for drug design and development, medicinal chemistry, and agrochemical sciences. Consequently, a wide range of in silico models based on deductive and inductive reasoning have been developed. However, the General Solubility Equation (GSE) stands out as a simple, thermodynamically derived equation to predict intrinsic solubility (log S0, in mol/L) of molecules from their melting point and n-octanol/water partition coefficient (log PN). Despite its widespread use, the applicability of this equation has not been rigorously evaluated to establish when it can or cannot produce accurate predictions. Therefore, a large-scale evaluation of the GSE was conducted using an extensive database of fully experimental melting point and log PN data for 2740, mostly drug-like compounds. A systematic analysis based on cheminformatics descriptors was conducted to identify the primary factors that allow the GSE to achieve predictions within the experimental uncertainty range (±1 log S unit). This information has been used to create GSESolver: a supervised machine learning classification model that, based on the selected descriptors, can predict whether the GSE will perform accurately solely from its SMILES string. GSESolver tool can be used by anyone in the scientific community to quickly assess molecule solubility and it was developed under the FAIR principles (Findable, Accessible, Interoperable, and Reusable) in a Google Colab notebook to ensure ease of use. Finally, the utility of the GSESolver tool was exploited in medicinal chemistry applications for the in silico prediction of the Maximum Absorbable Dose (MAD) of potential drugs, and in a rational agrochemical design application where solubility is a key factor.