Boosting the predictive performance with aqueous solubility dataset curation.

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Drug Discovery

Background:

• Intrinsic solubility is crucial for drug bioavailability.
• Current AI solubility prediction faces data scarcity and quality issues.
• Lack of unified measurements hinders AI and physics-based model integration.

Purpose of the Study:

• To curate diverse aqueous solubility datasets for improved AI model training.
• To evaluate the performance of enhanced deep learning models on curated data.
• To compare AI-driven solubility prediction with physics-based approaches.

Main Methods:

• Collected and curated seven aqueous solubility datasets.
• Developed and applied a dataset curation workflow.
• Evaluated two expanded deep learning methods and compared with a state-of-the-art physics-based approach using RMSE, Pearson, and Spearman correlations.

Main Results:

• Improved RMSE scores were observed across all curated thermodynamic datasets.
• Expanded Chemprop achieved high Pearson (0.930) and Spearman (0.947) correlation coefficients, comparable to physics-based methods.
• Performance metrics showed steady improvement with increasing data points.

Conclusions:

• A curated dataset and enhanced AI models significantly improve solubility prediction accuracy.
• AI models offer computational advantages for rapid molecule evaluation in drug discovery.
• This work facilitates better decision-making during hit identification and lead optimization stages.