Accurately predicting solubility curves via a thermodynamic cycle, machine learning, and solvent ensembles

Determining solubilities of organic molecules is critical in various fields such as pharmaceuticals, agrochemicals, and environmental science. Knowing how a solute will dissolve in different solvents and at different temperatures is essential for drug formulation, synthesis, purification, and crystallization. Hard-to-estimate solubility limits currently hinder the design of new processes, making innovation more expensive. We propose a fast and general method for predicting the solubilities of neutral organic molecules in a wide range of solvents and temperatures. Our method uses a thermodynamic fusion cycle to combine machine learning predictions of the activity coefficient, fusion enthalpy, and melting point temperature. This method was tested on a combined dataset with more than 100,000 experimental solubility values, showing better or comparable performance to competing methods on many solubility benchmarks even at elevated temperatures. We also introduce reference ensembling to leverage all available experimental solubilities for a given solute in estimating its solubility in a different solvent. Reference ensembling is also shown to enhance the robustness of models trained directly on solubility data.