Beyond Partitioning: Using Force Field Science to Evaluate Electrostatics Models.

Area of Science:

• Computational molecular science
• Physical chemistry
• Drug discovery and materials design

Background:

• Accurate electrostatic and induction interaction models are fundamental for molecular simulations.
• Existing methods for deriving atomic charges, such as partitioning electron density and fitting to electrostatic potential (ESP), have limitations.
• Force field calculations often rely on monomer-based charge models, which may not optimally predict interaction energies.

Purpose of the Study:

• To evaluate and improve methods for deriving atomic charges for force field calculations.
• To develop physics-based force fields that directly predict electrostatic and induction interaction energies.
• To leverage machine learning for enhanced force field parameterization.

Main Methods:

• Evaluation of charge derivation methods: partitioning electron density and ESP fitting.
• Comparison of different charge models, including positive point charges (PC) with distributed negative charges (Gaussian or Slater).
• Application of machine learning with the Alexandria Chemistry Toolkit to train physics-based models on symmetry-adapted perturbation theory (SAPT) interaction energies.

Main Results:

• ESP-fitted models combining PC and distributed charges improved predictions by ~30% over PC alone (RMSD 12 kJ/mol).
• A nonpolarizable model trained directly on SAPT dimer energy components achieved an RMSD of 3 kJ/mol.
• The developed approach enables direct comparison and optimization of force field models.

Conclusions:

• Directly training physics-based force fields on SAPT interaction energies using machine learning significantly enhances accuracy.
• This methodology provides a robust framework for developing accurate and predictive molecular force fields.
• The improved force fields will accelerate progress in computational molecular science for various applications.