Combining PC-SAFT and ML to Access Unknown API Solubilities.

Area of Science:

• Computational chemistry
• Chemical engineering
• Machine learning applications in drug development

Background:

• Predicting active pharmaceutical ingredient (API) solubility is crucial but current methods demand extensive, system-specific data.
• Existing state-of-the-art models are resource-intensive, limiting their early application in drug development.

Purpose of the Study:

• To develop a novel method integrating machine learning (ML) with PC-SAFT modeling to predict API solubility.
• To significantly reduce the experimental data requirements for solubility prediction.

Main Methods:

• Integration of adaptive machine learning (ML) methods with PC-SAFT modeling.
• Utilizing only molecular structure and melting properties of the API, with pre-existing solvent parameters.
• The ML framework generates PC-SAFT parameters for APIs.

Main Results:

• Highly accurate solubility estimations for 21 APIs in pure and mixed solvents were achieved.
• Achieved R² values of 0.83 (without binary data) and 0.98 (with one binary data point).
• The hybrid method demonstrated superior generalizability, consistency, and efficiency compared to prior models.

Conclusions:

• The developed ML framework efficiently provides PC-SAFT parameters for APIs with minimal or no experimental solubility data.
• Enables thermodynamics-based API solubility assessment early in process development.
• Outperforms existing models by requiring significantly less experimental input for reliable predictions.