Developing and Benchmarking Sage 2.3.0 with the AshGC Neural Network Charge Model.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Drug Discovery

Background:

• Assigning partial atomic charges is crucial for molecular simulations but computationally expensive.
• Existing methods often rely on quantum mechanics, which scales poorly and is sensitive to molecular geometry.

Purpose of the Study:

• Introduce Open Force Field (OpenFF) AshGC, a novel graph convolutional neural network charge model.
• Develop the Sage 2.3.0 small molecule force field, consistent with AshGC charges.
• Provide efficient, conformer-independent charges for molecules of all sizes.

Main Methods:

• Developed AshGC, a graph convolutional neural network for charge assignment.
• Parametrized the Sage 2.3.0 force field with Lennard-Jones and valence parameters consistent with AshGC.
• Benchmarked performance across gas-phase geometry optimization and protein-ligand binding free energies.

Main Results:

• AshGC produces conformer-independent charges of semiempirical quality at linear cost.
• Sage 2.3.0 shows comparable performance to previous releases, with improved condensed-phase properties.
• AshGC charges align with established methods like AM1-BCC, with minor deviations in specific chemical areas.

Conclusions:

• AshGC and Sage 2.3.0 offer an efficient and accurate approach to charge assignment and molecular simulation.
• The new force field is compatible with existing Amber protein force fields.
• All data and code are publicly available for reproducibility.