DASH: Dynamic Attention-Based Substructure Hierarchy for Partial Charge Assignment.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Molecular modeling

Background:

• Accurate partial charges are crucial for molecular simulations.
• Existing methods can be computationally expensive or lack generalizability.
• Graph neural networks (GNNs) show promise for predicting molecular properties.

Purpose of the Study:

• To develop a computationally efficient and accurate method for assigning atomic partial charges.
• To create a software-independent approach that integrates with existing parametrization pipelines.
• To leverage GNNs for improved molecular charge assignment.

Main Methods:

• A graph neural network (GNN) was trained to predict atomic partial charges using quantum-mechanical (QM) calculations.
• A dynamic attention-based substructure hierarchy (DASH) was constructed from GNN attention values.
• The DASH approach was integrated into the Open force field (OpenFF) parametrization pipeline.

Main Results:

• The DASH approach achieves the same accuracy as the parent GNN model.
• The method provides a significant speed-up in partial charge assignment.
• DASH is software-independent and easily integrated into existing workflows.

Conclusions:

• The DASH method offers a robust, efficient, and accurate solution for atomic partial charge assignment.
• Open-source availability of the implementation, tree, and training data facilitates adoption.
• This work advances the field of molecular parametrization and computational chemistry.