Beyond the cutoff: Hybrid ML/MM electrostatics for neural network potentials

Abstract

Atomistic Neural Network Potentials (NNPs) have been developed to predict molecular properties and electronic ground-state energies of small molecules with kcal/mol accuracy. However, despite their excellent scalability and low computational cost, atomistic NNPs are intrinsically local, making them unreliable for modeling long-range interactions in condensed-phase systems relevant to biological, engineering, and pharmaceutical applications. Recently, we demonstrated how to explicitly incorporate information on long-range interactions into the ANI NNP by retraining it in the presence of electrostatic potentials arising from the molecular environment. Effectively, the introduced embedded ANI/MM model is similar in spirit to quantum mechanics/molecular mechanics. Here, we extend this line of work by developing and training the ANI/MM network to predict binding energies for two protein–ligand complexes. We show that this network predicts forces with an error of less than 1 kcal/mol/Å, opening the possibility of using it for geometry optimizations and molecular dynamics. The resulting ANI/MM NNP outperforms the accurate, ab initio–fitted classical force field Q-Force and exhibits good transferability to new solutes, provided that the training set includes relevant structural fragments of the target molecule. Together, these findings demonstrate that hybrid ML/MM neural architectures offer a promising route toward chemically accurate, scalable modeling of complex molecular systems.