Deciphering Molecular and Solvent Effects on Aqueous and Organic Solubility through Interpretable Machine Learning...

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Solubility prediction is crucial for drug development but challenging due to complex structure-property relationships.
• Accurate solubility data across diverse media (aqueous and organic) is essential for compound screening and optimization.

Purpose of the Study:

• To develop an interpretable machine learning (ML) framework for predicting drug-like compound solubility.
• To achieve accurate solubility predictions in both aqueous and organic solvents.
• To gain mechanistic insights into the factors governing solubility.

Main Methods:

• Utilized large-scale curated datasets (AqSolDB, AqSolDBc, BigSolDB, BigSolDB 2.0).
• Trained and validated CatBoost ML models using repeated 5-fold cross-validation with appropriate data splitting strategies.
• Performed feature selection, hyperparameter optimization, and SHAP analysis for interpretability.

Main Results:

• Optimized ML models showed significant performance improvements over baseline.
• Hyperparameter optimization was key to enhancing model accuracy.
• SHAP analysis revealed key drivers: polarity/H-bonding for aqueous, solvent/temperature/topology for organic solubility.

Conclusions:

• The developed ML framework provides robust and transferable solubility predictions.
• The approach is applicable to real-world solubility prediction tasks in drug discovery.
• Understanding structure-property relationships enhances predictive model reliability.