A New Model for Solubility Prediction to Guide Solvent Selection for Process Development

Solvent selection to obtain adequate yield and impurity rejection is a critical activity in the development of crystallization processes for pharmaceutical molecules. This activity requires the determination of solubilities for a solute molecule (moreover a polymorph of a solute molecule) in a variety of different solvents and solvent combinations. This is referred to as solvent screening. There are roughly fifty different solvents commonly used in synthetic chemical manufacturing. Although experimentally determining the solubility of a single solute in each of these solvents is an approachable (be it impractical) problem, experimental screening becomes unapproachable for mixtures of two or more solvents. Conventionally, intuition has driven the reduction of the complete set of binary blends to a more tractable set of mixed solvent systems on a case-by-case basis. As such the majority of solvent combinations remaining unexplored for a given molecule. A predictive model was developed to better facilitate the exploration of single solvent systems and blends.

This model was developed starting from a mechanistic perspective for crystal growth and dissolution, which is most applicable at temperatures far from the melt. The model can be implemented based on either the regression of a limited set of solubility data and the subsequent prediction of solute characteristics (magnitude of polar, dispersive and hydrogen bonding interactions, size) or using molecular modelling to approximate these characteristics directly. The regression approach has proven to be efficient (requiring minutes) and accurate for single solvent systems. This model has been extended to mixtures of two solvents, and coupled with crystallization process design calculations to predict yield, and provide a rough process outline to the end user. The intention being that once a “good” system has been found (determined using the model) it can be verified experimentally. This model has been applied to several systems (~20 solutes and 200 solute/solvent combinations) including the modelling of a solubility data set for paracetamol and caffeine collected within the Enabling Technologies Consortium.