Prediction of aqueous solubility from SCRATCH

Abstract

Several methods have been proposed for the prediction of aqueous solubility. This study proposes the SCRATCH model for the aqueous solubility estimation of a compound directly from its structure. The algorithm utilizes predicted melting points and predicted aqueous activity coefficients for the solubility estimation, reflecting the truly predictive nature of the model. It uses two additive, constitutive molecular descriptors (enthalpy of melting and aqueous activity coefficient) and two non-additive molecular descriptors (symmetry and flexibility). The melting point prediction is trained on over 2200 compounds whereas the aqueous activity coefficient is trained on about 1640 compounds, making the model very rigorous and robust. The model is validated using a 10-fold cross- validation.A comparison with the General Solubility Equation suggests that the SCRATCH predicted aqueous solubilities have a slightly more average absolute error. This could result due to the fact that SCRATCH uses two predicted parameters whereas the GSE utilizes only one predicted property. Although the GSE is simpler to use, the drawback of requiring an experimental melting point is overcome in SCRATCH which can predict the aqueous solubility of a compound just based on its structure and no experimental values.