Machine Learning of Partial Charges Derived from High-Quality Quantum-Mechanical Calculations

Area of Science:

• Computational Chemistry
• Materials Science

Background:

• Parametrizing small organic molecules for molecular dynamics simulations is complex due to vast chemical space.
• Current methods often involve individual parametrization using partial charges from quantum-chemical calculations, which can be computationally intensive and sensitive to calculation methods.
• The accuracy of partial charges is crucial for reliable simulation outcomes.

Purpose of the Study:

• To develop a machine learning (ML) based approach for predicting partial charges of organic molecules.
• To provide a faster and more robust method for generating molecular parameters for classical molecular dynamics (MD) simulations.
• To ensure the predicted partial charges are conformation-independent.

Main Methods:

• Developed a machine learning model trained on a diverse dataset of druglike molecules.
• Predicted partial charges from density functional theory (DFT) electron densities.
• Validated the ML-predicted charges through benchmark calculations, including free energy of hydration and liquid properties (density, heat of vaporization).

Main Results:

• The ML approach significantly accelerates the prediction of partial charges compared to traditional quantum-chemical calculations.
• Predicted partial charges are independent of molecular conformation, unlike electrostatic potential (ESP) fitting methods.
• Benchmark calculations demonstrated good quality and compatibility of ML-derived charges with standard biomolecular force fields.

Conclusions:

• Machine learning offers an efficient and accurate method for generating partial charges for molecular dynamics simulations.
• This approach broadens the applicability of MD simulations for a wider range of organic molecules.
• The conformation-independent nature of ML-predicted charges enhances their reliability and utility in computational chemistry.