Molecular mechanics (MM)-based simulations rely on force field (FF) parameters to describe the intra- and intermolecular interactions between atoms in a biomolecular system. The biomolecular FFs, based on empirical potentials, have been optimized to describe proteins and common biomolecules reliably, but the parametrization of drug-like molecules remains challenging. For a more accurate description of small molecules, a solution is to rely on quantum mechanics (QM)/MM simulations, where the ligand is treated at the QM level, and the rest of the system using the classical MM. However, the computational cost of the QM/MM simulations limits the simulations length and system size. A newly developed method uses neural network potentials (NNP), trained by machine learning (ML) on data sets of description of ligands at the QM level of precision. The hybrid ML/MM simulations offer a precision of the intramolecular description of the ligand close to the QM precision but at a much lower computational cost. The new module of CHARMM-GUI Hybrid ML/MM builder creates input files for the simulations of protein-ligand complexes using a NNP to describe the ligand and a classical FF for the rest of the system. The module handles protein-ligand complex systems in solution and in the membrane and offers a user-friendly interface to create the input files. The goal of Hybrid ML/MM builder is to help the user setting up hybrid ML/MM simulations with various NNPs and software.
Szczepaniak et al. (Sun,) studied this question.