Abstract
Accurate prediction of solubility limits of organic compounds in mixed solvents is critical for the design and optimization of chemical and pharmaceutical processes. Recent advances in machine learning have enabled fast and reliable prediction of physicochemical properties of molecules, including solubility. In this work, we present a Bayesian Optimization framework to identify optimal solvent combinations, compositions, and temperatures that maximize the solubility of active pharmaceutical ingredients. The optimization strategy leverages a multicomponent directed-edge message passing neural network trained on solvent mixtures to predict solubility in ternary systems consisting of a solute and two solvents. To enable efficient Bayesian optimization, we represented the solvents in a continuous space and compare three different strategies: integer enumeration, numerical descriptors, and deep embeddings. The proposed approach was tested on a dataset comprising 14 299 points in solvent mixtures. The results indicate that integer enumeration and embedding-based representations achieve the best performance in identifying solvent mixtures that maximize solubility for each compound at a fixed number of model evaluations. While demonstrated on ternary systems, the framework is readily extensible to mixtures with a larger number of components. With this work we aim to identify which chemical representation is most suitable for similar optimization problems, providing a solid foundation for further and broader exploration, for example in the optimization of crystallization processes.