Development and evaluation of an effective solubility prediction model for pharmaceuticals in organic solvents using...

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Pharmaceutical Sciences

Background:

• Accurate prediction of active pharmaceutical ingredient (API) solubility is crucial for drug development and formulation.
• Existing models often struggle to capture the complex temperature-dependent solubility behavior of APIs in diverse organic solvents.

Purpose of the Study:

• To evaluate the predictive performance of a machine learning model based on the XGBoost framework for API-like molecule solubility in organic solvents.
• To assess the model's ability to capture temperature-dependent solubility across a wide temperature range.

Main Methods:

• Utilized a dataset of 30 binary mixtures with solubility data for four solutes (butyl paraben, fenofibrate, risperidone, fenoxycarb) for training, testing, and validation.
• Employed 10 descriptors including solute properties (heat of fusion, melting temperature, heat capacity, Hansen solubility parameter), solvent properties (dielectric constant, boiling temperature), and temperature.
• Incorporated a constraint on the XGBoost algorithm to accurately model temperature dependence.

Main Results:

• Achieved low root mean square errors (RMSLE) for calculated solubility: 0.05 (train), 0.09 (test), 0.13 (validation).
• The model predicted butamben solubility with an average RMSLE of 0.41.
• The XGBoost model demonstrated significantly improved solubility modeling and prediction capabilities compared to Flory-Huggins and NRTL-SAC models for all studied compounds.

Conclusions:

• The XGBoost machine learning model effectively predicts the solubility of API-like molecules in organic solvents, accurately capturing temperature dependence.
• This approach offers a superior alternative to traditional thermodynamic models for solubility prediction in pharmaceutical research and development.