Accurate Electrostatics for Biomolecular Systems through Machine Learning

Accurate models for electrostatic interactions are fundamental for force field calculations in drug and material design. Given good mod- els of the entire charge distributions, e.g. from the quantum-chemical electron density, the electrostatic interaction between molecules can be calculated using Coulomb’s law. Here we show that the popular method of fitting charges to the electrostatic potential in just a few layers around molecules is flawed due to lack of information, as ex- plained by the Poisson equation. Instead, we employ machine learn- ing using the Alexandria Chemistry Toolkit (ACT) to generate charge models that reproduce electrostatic and induction energies from sym- metry adapted perturbation theory calculations for charged amino-acid side chain analogs with inorganic ions and water. The ACT enables ra- tional design of physics-based models through force field science, and it is demonstrated how the energy components predicted by novel mod- els, either with or without explicit polarization, can be made more accurate than existing models.