Simulating protein-ligand binding with neural network potentials

Abstract

Computational methods have been developed to predict the structures and energetics of protein-ligands complexes. However these methods are limited by the accuracy and transferability of the molecular mechanical (MM) models used to calculate the potential energy. Neural network potentials (NNPs) eliminate the need for parameterization and avoid many of the limiting assumptions of MM models. We evaluated the accuracy of ANI-type NNP models for predicting the potential energy surface of biaryl torsions. The ANI-2X and ANI-1ccX NNPs were found to be more accurate and reliable than popular molecular mechanical models. We then developed a new method where the NNP is used to describe the intramolecular terms of a ligand while a conventional MM model is used to describe the environment. This method was found to be effective for predicting the binding pose of ligands bound to proteins and could be used to calculate the conformational component of the binding energy. We also show that these methods can be used to re�ne low-resolution cryo-EM structures of protein-ligand complexes.