Predicting absolute aqueous solubility by applying a machine learning model for an artificially liquid-state as proxy...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of aqueous solubility is crucial for drug development.
• Traditional quantitative structure-activity relationship (QSAR) models often struggle with absolute solubility predictions.
• Quantum mechanics (QM)-derived descriptors offer a more detailed representation of molecular properties.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting the absolute solubility of drug-like compounds.
• To compare the performance of QM-derived descriptors against traditional cheminformatics descriptors.
• To explore cost-effective alternatives for solubility prediction.

Main Methods:

• Utilized machine learning algorithms (e.g., random forest, gradient boosting) combined with QM-derived COSMO-RS descriptors and Morgan fingerprints.
• Employed two prediction strategies: a hypothetical pathway involving an artificial liquid state and direct solubility prediction.
• Trained models on Bayer in-house compound data and validated using external solubility challenge datasets.

Main Results:

• Machine learning models with QM-derived descriptors significantly improved absolute solubility prediction accuracy compared to existing QSAR models.
• QM-derived descriptors demonstrated superiority over standard cheminformatics descriptors.
• Fragment-based COSMOquick calculations provided a low-cost alternative with only a minor decrease in prediction quality.

Conclusions:

• Machine learning integrated with QM-derived descriptors represents a powerful approach for accurate absolute solubility prediction.
• QM-derived descriptors are more effective than cheminformatics descriptors for this task.
• Cost-effective QM-based methods can be utilized without substantial loss of predictive performance.