Enabling Biomolecular Simulations with Neural Network Potentials in GROMACS

Lukas MüllenderBerk Hess, Erik Lindahl · 2026

Read Enabling Biomolecular Simulations with Neural Network Potentials in GROMACS by Lukas MüllenderBerk Hess, Erik Lindahl (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-21441.txt.

Introduction

Lukas Müllender, Berk Hess, and Erik Lindahl connect neural network potentials to GROMACS for hybrid machine-learning/molecular-mechanics simulation. Their interface lets PyTorch-trained potentials supply energies and forces for chosen subsets or whole systems, open to descriptor-based or message-passing designs. It plugs into GROMACS sampling and free-energy workflows. Demos cover peptide torsional landscapes, solvation free energies, and protein-ligand systems, with water-box benchmarks across designs.

Catalog record

CREATORLukas MüllenderBerk Hess, Erik Lindahl
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46290
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.21441

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enabling Biomolecular Simulations with Neural Network Potentials in GROMACS Lukas MüllenderBerk Hess, Erik Lindahl Abstract ABSTRACT. Neural network potentials (NNPs) are rapidly changing the landscape of state-of-the-art molecular dynamics (MD) simulations. To make full use of this development, the community needs flexible, easy-to-use interfaces firmly integrated with existing methodologies. To address this, we here present an interface for hybrid machine learning/molecular mechanics (ML/MM) simulations implemented in the widely used MD code GROMACS. The interface enables NNPs trained in the PyTorch framework to contribute energies and forces during MD simulations, either for selected subsets or entire molecular systems. By defining a flexible set of model inputs and outputs, the interface is agnostic to specific NNP architectures and can accommodate a wide range of descriptor-based and message-passing models. In particular, the design integrates NNP inference seamlessly into the extensive GROMACS molecular simulation ecosystem, providing users with the capability to straightforwardly combine NNPs with existing advanced sampling and free energy workflows. We demonstrate the capabilities of the interface using several representative applications, including enhanced sampling of peptide torsional free energy landscapes, absolute solvation free energy calculations, and protein–ligand simulations. We also run performance benchmarks on water boxes for several different NNP architectures. Our interface is available in recent GROMACS releases, and we believe it will provide a practical foundation for incorporating machine learning potentials into production MD simulations of biomolecular systems. 1 Introduction Since their development in the 1970s [ 64 ] , hybrid quantum mechanical/molecular mechanical (QM/MM) simulations have established themselves as a powerful computational tool, combining the enthalpic accuracy of electronic structure methods with the powerful sampling capabilities of molecular dynamics (MD) simulations [ 35 ] . They provide an ideal trade-off between accuracy and efficiency especially in biomolecular systems, which are highly heterogeneous and often necessitate large solvent boxes best treated classically [ 50 ] . However, QM/MM simulations remain computationally intensive as a result of the bottleneck of ab initio QM calculations, even for QM regions of moderate size, making the treatment of many biological and chemical phenomena costly and time-consuming. In recent years, the development of a class of machine learning (ML) models called machine learning interactomic potentials (MLIPs), or more specifically neural network potentials (NNPs), is promising to address this issue, by reproducing energies and forces with QM accuracy at a significantly reduced computational cost [ 59 , 6 ] . Since the seminal work of Behler and Parrinello [ 5 ] , many different architectures have demonstrated excellent predictive accuracy on a wide range of available benchmarks, including ANI [ 53 , 15 ] , AIMNet [ 72 , 2 ] , MACE [ 3 ] , PhysNet [ 60 ] , NequIP [ 4 ] and others. Generally, applications of MD simulations with NNPs fall into two main classes, each with their own challenges: Full ML simulations, where the entire system is modeled by an NNP similar in spirit to ab initio MD, and hybrid ML/MM simulations akin to QM/MM approaches. The first has found use in biomolecular applications [ 61 , 31 , 28 ] and particularly in materials science, where they are now the de-facto standard due to the lack of broadly applicable empirical force fields [ 39 ] . However, these simulations face challenges related to high compute and memory requirements, which requires scaling to multiple GPUs [ 41 , 19 ] , as well as the accurate incorporation of long-range interactions, which is an area of ongoing research [ 58 , 65 , 30 , 51 ] . In the context of biomolecular simulations, where widely applicable force fields and water models are available, employing a more expensive model for the description of e.g. solvent-solvent interactions which are ultimately of lesser interest might not be the most efficient use of computational resources. For this reason, it remains desirable to perform hybrid ML/MM simulations, where only a central region of interest is modeled with the NNP, while the rest of the system is treated with classical force fields. Such approaches have already shown promise for example in its application to conformational and binding free energies of protein–ligand complexes [ 33 , 47 ] , alchemical free energy calculations [ 47 , 48 , 29 ] and vibrational spectroscopy [ 22 , 49 , 68 ] . These works give rise to the nontrivial question of how to treat long-range atomic interactions across the ML–MM boundary. Many works have implemented a scheme analogous to mechanical embedding (ME) [ 33 , 47 , 16 , 20 ] , which is straightforward to implement with NNPs but known to perform insufficiently in many QM/MM applications [ 35 , 50 ] . There, the state-of-the-art is represented by electrostatic embedding, which has been adapted to NNPs by the EMLE method [ 70 , 69 ] and others [ 22 , 49 , 45 ] . When it comes to the practical application of this breadth of different competing approaches in high-performance MD simulation engines, there is a clear need for flexible software interfaces that can accommodate the architectural diversity, while keeping compatibility with existing workflows. While direct implementations and targeted optimizations exist for specific architectures [ 31 , 20 ] , a more viable alternative is to rely on popular ML frameworks like PyTorch [ 42 ] or JAX [ 12 ] . This strategy has been taken by several implementations of NNP interfaces in popular MD codes like ASE [ 26 ] , OpenMM [ 16 ] , LAMMPS [ 54 ] and Amber [ 43 ] . Furthermore, packages that offer training capabilities for specific NNP architectures often also implement interfaces to MD engines of their choice, like MACE [ 3 ] , DeePMD-kit [ 67 ] and chemtrain-deploy [ 19 ] . In this paper, we present a general and flexible interface for performing ML/MM simulations in the widely used molecular simulation toolkit GROMACS [ 1 ] . Our interface, which we call nnpot , allows for NNPs trained within the popular PyTorch framework to contribute to the force calculation for an arbitrary subset of atoms in the system, or its entirety. It is actively maintained as part of the main GROMACS repository since its 2025 release, with the 2026 version adding significant updates to model compatibility and applicability. To be used with the interface, an arbitrary NNP only has to conform to a user-defined set of expected inputs and outputs, allowing for a high degree of flexibility. While this approach allows for straightforward customization, it might provide a somewhat high barrier of entry for users less familiar with machine learning frameworks, which is why we also provide a GitHub repository with an extensive how-to guide and several examples. We note that while our examples focus on popular NNP architectures like ANI and MACE, our interface is designed to be general and to work with any NNP that conforms to the API set by the interface. After briefly reviewing the main concepts of NNPs and ML/MM simulations, as well as an overview of the implementation and main features of our interface, we showcase its capabilities on a range of different applications: To check the compatibility with enhanced sampling techniques available in GROMACS, we

Read the complete text