When Does Machine Learning Add Value over Theory? Predicting API Solubility in Binary Mixtures with COSMO-RS and...

Area of Science:

• Computational chemistry
• Pharmaceutical sciences
• Machine learning

Background:

• Predicting active pharmaceutical ingredient (API) solubility in aqueous-organic mixtures is crucial for drug formulation.
• Physics-based models like COSMO-RS are valuable but struggle with complex, non-ideal systems.
• The added value of machine learning (ML) over established theory needs clear delineation.

Purpose of the Study:

• To evaluate when ML-based approaches offer significant improvements over COSMO-RS for API solubility prediction.
• To compare the performance of COSMO-RS with a hybrid ML/COSMO-RS workflow (DOOIT2).
• To provide practical guidance for formulation scientists on selecting appropriate predictive models.

Main Methods:

• Comparison of COSMO-RS and DOOIT2 (hybrid ML/COSMO-RS) on two datasets: diverse APIs and acid-centered solutes.
• Utilized API-out Structured Group K-Fold validation for robust generalization assessment.
• Included newly measured solubilities in aqueous 4-formylmorpholine mixtures.

Main Results:

• For the homogeneous acid series, COSMO-RS showed strong performance (RMSD=0.321, R²=0.925), with no significant improvement from DOOIT2.
• For the diverse API dataset, DOOIT2 reduced RMSD from 0.686 to 0.527 and increased R² from 0.829 to 0.849.
• Prediction uncertainty was linked to the low-solubility region, not solely molecular weight.

Conclusions:

• ML-based corrections like DOOIT2 provide tangible benefits for solubility prediction in complex, diverse API systems.
• For simpler, homogeneous systems, established physics-based models offer sufficient accuracy.
• Findings clarify the practical utility of ML in pharmaceutical formulation, guiding model selection.