Towards AshGC, a graph neural network charge model for Open Force Field

Biomolecular force fields typically model electrostatic interactions with point partial charges assigned using expensive ab initio or semi-empirical approaches, which scale poorly to large molecules such as proteins. While protein force fields typically encode pre-computed charges for standard residues, this approach cannot account for non- standard components like covalently bound ligands or post-translational modifications and assigning charges to these systems remains a non-trivial task. Open Force Field has been developing a graph neural network-based method for efficient assignment of self- consistent charges to both small and large molecules for Open Force Field force fields. This charge model, called AshGC, has been fit to produce charges of semi-empirical quality at low cost. Here we describe the model, its architecture, featurization, training process, and release plan. AshGC is currently available as a release candidate under the name "openff-gnn-am1bcc-0.1.0-rc.3.pt" and can be used with Open Force Field infrastructure.