Group Contribution and Machine Learning Approaches to Predict Abraham Solute Parameters, Solvation Free Energy, and...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Physical Chemistry

Background:

• Accurate prediction of solute properties and solvation thermodynamics is crucial for various chemical applications.
• Existing methods for predicting Abraham solute parameters and solvation energies have limitations in accuracy and scope.
• Development of robust and efficient computational tools is needed to advance molecular modeling and property prediction.

Purpose of the Study:

• To develop and evaluate novel computational models for predicting Abraham solute parameters and solvation free energy and enthalpy.
• To compare the performance of a group contribution method (SoluteGC) and machine learning models (SoluteML, DirectML).
• To provide publicly accessible databases and prediction models for solute properties and solvation thermodynamics.

Main Methods:

• A group contribution method (SoluteGC) using atom-centered functional groups with strain corrections.
• Machine learning models (SoluteML and DirectML) employing directed message passing neural networks.
• Training and evaluation on extensive datasets comprising 8366 solute parameters, 20,253 solvation free energies, and 6322 solvation enthalpies.

Main Results:

• The DirectML model demonstrated superior performance in predicting solvation free energy and enthalpy, achieving accuracy comparable to advanced quantum chemistry methods.
• SoluteGC and SoluteML also provided useful predictions, with combined model performance exceeding individual model accuracy.
• Identification of uncertain predictions by comparing the three models was feasible.

Conclusions:

• The developed DirectML model offers a highly accurate and efficient approach for predicting solvation thermodynamics.
• The suite of models (SoluteGC, SoluteML, DirectML) provides versatile tools for diverse computational chemistry needs.
• Publicly accessible databases (SoluteDB, dGsolvDBx, dHsolvDB) and prediction models are released to facilitate research.