EspalomaCharge: Machine Learning-Enabled Ultrafast Partial Charge Assignment.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Artificial Intelligence

Background:

• Atomic partial charges are essential for molecular dynamics (MD) simulations, influencing electrostatic interactions and molecular behavior.
• Current methods for assigning partial charges, like AM1-BCC, are computationally expensive, limiting their application to large systems.
• The accuracy of partial charge assignment directly impacts the reliability of MD simulation results.

Purpose of the Study:

• To develop a significantly faster yet accurate method for predicting atomic partial charges.
• To enable the creation of consistent charge models for large molecules and biomolecular systems.
• To facilitate the development of next-generation self-consistent biomolecular force fields.

Main Methods:

• A hybrid approach combining graph neural networks (GNNs) with a charge equilibration (QEq) method.
• GNNs predict molecule-specific atomic electronegativity and hardness parameters.
• Analytical QEq determines optimal partial charges while conserving total molecular charge.

Main Results:

• The hybrid GNN-QEq model achieves speeds orders of magnitude faster than traditional methods.
• The accuracy of the predicted partial charges is comparable to existing AM1-BCC implementations.
• The method exhibits linear scaling with the number of atoms, making it suitable for large systems.

Conclusions:

• The developed hybrid approach offers a computationally efficient and accurate solution for atomic partial charge assignment.
• This method enables the use of consistent charge models for small molecules and biopolymers.
• The implementation in EspalomaCharge provides a practical tool for researchers, integrating with existing workflows.