Predicting the solubility of active pharmaceutical ingredients (APIs) is essential throughout drug development. However, state-of-the-art modeling approaches require system-specific data sets for parameter estimation and are resource intensive. This work introduces a new method that integrates adaptive machine learning (ML) methods with PC-SAFT modeling to vastly reduce requirements of experimental data. Instead of extensive experimental campaigns of solubility measurements, only the molecular structure and melting properties of the API are needed - information onten available in literature or easily measured. The ML framework applied in this work supplies PC-SAFT parameters for APIs. With solvent parameters already available from literature, this novel approach provided highly accurate solubility estimations for 21 APIs in pure solvents (R2 = 0.83 without using any binary data and R2 = 0.98 using a single binary data point), as well as for mixed solvents, closely matching literature data. Compared to prior models, this hybrid method is more generalizable, consistent, and efficient, streamlining the workflow and providing reliable predictions with minimal experimental effort. By making a thermodynamic-based solubility assessment available early in process development, it outperforms state-of-the-art models that demand significantly more experimental input. The results of this work indicate that the newly developed ML framework can be efficiently applied to provide PC-SAFT parameters for APIs with minimal need of or even without using any experimental solubility data, which can be used to achieve thermodynamics-based access to API solubility in a very early stage of process development. This approach does not only provide solubility data in pure solvents but also in solvent mixtures.
Habicht et al. (Mon,) studied this question.