Accurately predicting solvation free energy in aqueous and organic solvents beyond 298 K by combining deep learning and the 1D reference interaction site model

Abstract

We report a method to predict the absolute solvation free energy (SFE) of small organic and druglike molecules in water, carbon tetrachloride and chloroform solvents beyond 298 K by combining the 1 Dimensional Reference Interaction Site Model (1D-RISM) and deep learning. RISM is a statistical mechanics based method for modelling molecular solutions that is computationally inexpensive but is too inaccurate for routine SFE calculations in its common form. By replacing the 1D-RISM SFE functional with a 1D convolutional neural network (CNN) trained on RISM correlation functions, we show that predictions approaching chemical accuracy can be obtained for aqueous and non-aqueous solvents at a wide-range of temperatures. This method builds upon the previously reported RISM-MOL-INF procedure which applied RISM to accurately characterise solvation and desolvation processes through solute–solvent correlation functions [Palmer et al., Mol. Pharm., 2015, 12, 3420–3432]. Unlike RISM-MOL-INF however, the newly developed pyRISM-CNN model applied here is capable of rapidly modelling these processes in several different solvents and at a wide-range of temperatures. The pyRISM-CNN functional reduces the predictive error by up to 40-fold as compared to the standard 1D-RISM theory. Prediction errors below 1 kcal mol⁻¹ are obtained for organic solutes in carbon tetrachloride or chloroform solvent systems at 298 K and water solvent systems at 273–373 K. pyRISM-CNN has been implemented in our in-house 1D-RISM solver (pyRISM), which is made freely available as open-source software.