Machine-Learned Electrostatic Potentials for Accurate Hydration Free Energy Calculations

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Physical Chemistry

Background:

• Free energy calculations are crucial in computational chemistry but limited by partial charge accuracy.
• Semiempirical methods like AM1-BCC often yield inaccurate charges for polar species, impacting hydration free energy predictions.
• Existing methods struggle with charge assignment reproducibility and accuracy for diverse molecular conformations.

Purpose of the Study:

• To establish the link between inaccurate AM1-BCC charges and poor hydration free energy calculations.
• To develop a rapid and accurate method for predicting partial charges using machine learning.
• To introduce a novel method for assigning ensemble-averaged charges and improve free energy calculations.

Main Methods:

• Utilized an XGBoost regressor trained on atomic descriptors to predict charges from high-fidelity DFT calculations (PBE0-D3(BJ)/def2-TZVP).
• Developed the Boltzmann Percentile method combining the predictive model with molecular dynamics simulations for ensemble charge assignment.
• Calculated hydration free energies on the FreeSolv dataset using both traditional and the new charge assignment methods.

Main Results:

• The XGBoost model accurately predicts DFT-level charges, improving electrostatic descriptions.
• The Boltzmann Percentile method yields charges robust to conformational variations.
• Achieved a root mean squared error of 1.69 kcal/mol for hydration free energies, significantly outperforming AM1-BCC (3.05 kcal/mol).

Conclusions:

• The proposed machine learning approach and Boltzmann Percentile method offer a realistic enhancement for free energy calculations.
• This method provides accurate and reproducible charge assignments at a computational cost comparable to semiempirical methods.
• Enables more reliable molecular dynamics simulations in condensed phases by improving charge accuracy.