Extreme Gradient Boosting Combined with Conformal Predictors for Informative Solubility Estimation.

Area of Science:

• Computational Chemistry
• cheminformatics
• Machine Learning in Chemistry

Background:

• Accurate prediction of chemical compound solubility is crucial for drug discovery and chemical process design.
• Existing methods often focus on prediction accuracy alone, neglecting the reliability and applicability domain of predictions.

Purpose of the Study:

• To develop and validate a comprehensive four-step computational methodology for predicting experimental solubility.
• To establish an applicability domain for solubility predictions across large chemical databases.
• To quantify prediction uncertainty and accuracy classes for chemical compounds.

Main Methods:

• Utilized the extreme gradient boosting (XGB) algorithm for predicting solubility and identifying key molecular descriptors.
• Performed applicability domain (AD) testing on large datasets (Drugbank, PubChem, COCONUT) using curated and uncurated AquaSolDB data.
• Applied conformal prediction to generate narrow prediction intervals and validate them against experimental solubility values.

Main Results:

• Achieved prediction accuracy (RMSE) of 0.59-0.76 Log(S) for water and 0.62-0.79 Log(S) for organic solvents.
• Demonstrated that over 95% of approximately 500,000 compounds fall within the established applicability domain.
• Successfully estimated individual error margins and accuracy classes for solubility predictions.

Conclusions:

• The developed four-step approach provides a robust framework for reliable solubility prediction and uncertainty quantification.
• This method extends beyond typical solubility prediction studies by incorporating applicability domain assessment and prediction interval generation.
• The methodology enables accurate solubility prediction and error estimation for vast chemical libraries without requiring experimental solubility data.