Data-driven prediction of lipophilic drug solubility in supercritical CO2 using an adaptive ensemble learning architecture.

INTRODUCTION: Predicting how lipophilic pharmaceuticals dissolve in supercritical carbon dioxide (SC-CO2) remains a central challenge for process development, largely because experimental measurements are time-consuming and traditional modeling approaches often fail to generalize across chemically diverse compounds.

METHODS: This study examines the SC-CO2 solubility of four representative drugs Sirolimus, Tacrolimus, Rifampin, and Teriflunomide and introduces an adaptive computational framework designed to improve predictive power across varied molecular structures. The approach combines multiple boosting-based regressors within an ensemble scheme that incorporates a DST-inspired aggregation step and a weighted integration mechanism. Model hyperparameters are optimized using a bio- inspired Electric Eel Foraging strategy. Reliability of the final model was evaluated through repeated cross-validation, comparative statistical tests, and uncertainty estimation.

RESULTS AND DISCUSSION: The resulting ensemble achieved high predictive accuracy (R 2 = 0.99, RMSE = 0.089) with consistent performance across the investigated compounds. By integrating molecular descriptors, thermodynamic features, and process conditions within a unified ensemble framework, the approach provides a practical computational tool for supporting solubility estimation in SC-CO2 systems.