Abstract:Machine Learning Force Fields (MLFFs) promise to enable general molecular simulations that can simultaneously achieve efficiency, accuracy, transferability, and scalability for diverse molecules, materials, and hybrid interfaces. A key step toward this goal has been made with the GEMS approach to biomolecular dynamics [Sci. Adv. 10, eadn4397 (2024)]. This work introduces the SO3LR method that integrates the fast and stable SO3krates neural network for semi-local interactions with universal pairwise force fields designed for short-range repulsion, long-range electrostatics, and dispersion interactions. SO3LR is trained on a diverse set of 4 million neutral and charged molecular complexes computed at the PBE0+MBD level of quantum mechanics, ensuring a comprehensive coverage of covalent and non-covalent interactions. Our approach is characterized by computational and data efficiency, scalability to 200 thousand atoms on a single GPU, and reasonable to high accuracy across the chemical space of organic (bio)molecules. SO3LR is applied to study units of four major biomolecule types, polypeptide folding, and nanosecond dynamics of larger systems such as a protein, a glycoprotein, and a lipid bilayer, all in explicit solvent. Finally, we discuss the future challenges toward truly general molecular simulations by combining MLFFs with traditional atomistic models.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to simultaneously achieve efficiency (Efficiency), accuracy (Accuracy), transferability (Transferability) and scalability (Scalability) in molecular simulations, namely the EAST requirements. Existing methods usually make significant compromises in these aspects, resulting in a limited scope of application. This paper addresses this challenge by proposing a new machine - learning force field (MLFF) - the SO3LR method. SO3LR combines the fast and stable SO3krates neural network for semi - local interactions and a general pairwise force field designed for short - range repulsion, long - range electrostatics and dispersion interactions. In this way, SO3LR aims to provide a general molecular simulation tool that can be widely applied to different molecules, materials and hybrid interfaces while maintaining computational efficiency, data efficiency and high precision. Specifically, the paper addresses the following key issues: 1. **Constructing an atomic force field model that meets the EAST requirements**: Traditional force field models are either based on approximate but fast mechanical expressions or on accurate but extremely computationally costly ab - initio electronic structure calculations. Both methods sacrifice accuracy and efficiency. SO3LR aims to overcome these limitations by combining a machine - learning model and physically - inspired pairwise potential energy terms. 2. **Describing semi - local and long - range interactions**: SO3LR describes semi - local multi - body interactions through the SO3krates model and handles short - range repulsion, long - range electrostatics and dispersion interactions through physical pairwise terms. This combination ensures the wide applicability of the model in different chemical spaces. 3. **Comprehensive coverage of the dataset**: SO3LR is trained on a dataset containing 4 million neutral and charged molecular complexes at the PBE0 + MBD quantum - mechanical calculation level, covering a comprehensive range of covalent and non - covalent interactions. 4. **Scalability for large - scale simulations**: SO3LR can be scaled to systems of 200,000 atoms on a single GPU with reasonable high precision, applicable to the chemical space of organic (biological) molecules. Through these improvements, SO3LR has demonstrated its performance and potential in the simulations of small - molecule units, the folding of polyalanine systems, the dynamics of liquid water and large - molecule systems, especially in the explicit - solvent simulations of biomolecules.

Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields

Accurate Machine Learned Quantum-Mechanical Force Fields for Biomolecular Simulations

Towards Linearly Scaling and Chemically Accurate Global Machine Learning Force Fields for Large Molecules

sGDML: Constructing Accurate and Data Efficient Molecular Force Fields Using Machine Learning

Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields

Accurate global machine learning force fields for molecules with hundreds of atoms

Efficient Machine Learning Force Field for Large-Scale Molecular Simulations of Organic Systems

Efficient interatomic descriptors for accurate machine learning force fields of extended molecules

From Peptides to Nanostructures: A Euclidean Transformer for Fast and Stable Machine Learned Force Fields

Crash Testing Machine Learning Force Fields for Molecules, Materials, and Interfaces: Molecular Dynamics in the TEA Challenge 2023

Machine learning for molecular simulation

PhyNEO: A Neural-Network-Enhanced Physics-Driven Force Field Development Workflow for Bulk Organic Molecule and Polymer Simulations.

Crash Testing Machine Learning Force Fields for Molecules, Materials, and Interfaces: Model Analysis in the TEA Challenge 2023

Accurate and scalable graph neural network force field and molecular dynamics with direct force architecture

EL-MLFFs: Ensemble Learning of Machine Leaning Force Fields

Accurate and scalable multi-element graph neural network force field and molecular dynamics with direct force architecture

Accurate interatomic force field for molecular dynamics simulation by hybridizing classical and machine learning potentials

Toward a general neural network force field for protein simulations: Refining the intramolecular interaction in protein

Molecular Force Fields with Gradient-Domain Machine Learning (GDML): Comparison and Synergies with Classical Force Fields

Improving machine learning force fields for molecular dynamics simulations with fine-grained force metrics

The emergence of machine learning force fields in drug design