Necessity of Explicit Long-Range Electrostatics in NNP-Based Biomolecular Simulations

Determine whether explicit long-range electrostatics (e.g., via Particle-Mesh Ewald with partial charges) must be incorporated into biomolecular simulations driven by neural network potentials, particularly for systems with charge interactions spanning tens of angstroms such as membrane proteins, or whether implicit treatment by learned potentials with sufficient receptive field and message passing suffices.

Background

In biomolecular systems, charge interactions can extend over long distances, and classical molecular dynamics commonly employs explicit long-range electrostatics (e.g., PME) with partial charges. Partial charges, however, are not quantum observables and entail additional computational cost, while modern neural network potentials may implicitly capture charge distributions. The authors highlight that the need for explicit long-range electrostatics in NNP-based simulations remains unresolved.