Explainable Supervised Machine Learning Model To Predict Solvation Gibbs Energy.

Area of Science:

• Computational Chemistry
• Machine Learning Applications

Background:

• Accurate prediction of solvation free energy (ΔGsol) remains a challenge for computational models.
• Existing machine learning (ML) methods offer speed but lack explanatory power for broad chemical predictions.

Purpose of the Study:

• To develop a novel supervised ML model for predicting ΔGsol.
• To achieve a favorable speed-accuracy trade-off with enhanced explanatory insights.

Main Methods:

• Utilized two ensemble regressor algorithms for ML model development.
• Employed open-source chemical features encoding electronic, structural, and surface area descriptors.
• Integrated molecular properties and chemical interaction features for analysis.

Main Results:

• Achieved high accuracy in ΔGsol prediction, outperforming benchmark Neural Network methods.
• Identified increasing polar surface area and decreasing polarizability as key solute descriptors.
• Demonstrated a maximum absolute error of 0.22 ± 0.02 kcal mol⁻¹.
• Validated model performance on external databases and through solvent hold-out tests.

Conclusions:

• The developed ML model offers a fast and accurate approach for predicting ΔGsol.
• The model provides valuable explanatory insights into descriptor importance.
• This method holds potential for predicting other thermodynamic properties in computational chemistry.