Challenges in Predicting Aqueous Solubility of Organic Molecules Using the COSMO-RS Model

The conductor-like screening model for real solvents (COSMO-RS) is an emerging tool for predicting the thermodynamic properties of small molecules, such as activity coefficients, vapor pressure, and solubility in mixed solvents. We have evaluated the validity and performance of the COSMO-RS model in predicting aqueous solubility using a high-fidelity experimental data set containing 1852 small organic molecules. Our analysis indicates that overreliance on distributions of the screening charge densities on the outer surface of the molecule, known as the sigma-profile, leads to frequent overestimation or underestimation, rendering the predicted aqueous solubility of organic molecules unreliable. Despite our efforts to refine the inherent parameters of the COSMO-RS model using a comprehensive experimental data set, the prediction accuracy remained limited (R-2 similar to 0.5). We analyzed the correlation between prediction accuracy and common molecular descriptors, such as dipole moment, molar mass, symmetry of charges, and types of charge centers. A notable challenge of COSMO-RS prediction arises from the electronegative atoms, which likely generate uneven charge distributions in the sigma-profile, leading to overtly strong hydrogen bond corrections and thereby causing overestimation of aqueous solubilities.