[Paper Review] Accurate Machine Learned Quantum-Mechanical Force Fields for Biomolecular Simulations

Oliver T. Unke, Martin Stöhr+9 more|arXiv (Cornell University)|May 17, 2022

Protein Structure and Dynamics17 citations

TL;DR

This paper introduces GEMS, a general framework for constructing accurate machine-learned force fields (MLFFs) by combining bottom-up and top-down molecular fragment training to capture long-range and many-body interactions. The method enables nanosecond-scale ab initio quality molecular dynamics simulations of large biomolecules—such as 25k-atom crambin in water—at orders-of-magnitude faster speeds than traditional quantum mechanics, revealing greater protein flexibility than previously assumed.

ABSTRACT

Molecular dynamics (MD) simulations allow atomistic insights into chemical and biological processes. Accurate MD simulations require computationally demanding quantum-mechanical calculations, being practically limited to short timescales and few atoms. For larger systems, efficient, but much less reliable empirical force fields are used. Recently, machine learned force fields (MLFFs) emerged as an alternative means to execute MD simulations, offering similar accuracy as ab initio methods at orders-of-magnitude speedup. Until now, MLFFs mainly capture short-range interactions in small molecules or periodic materials, due to the increased complexity of constructing models and obtaining reliable reference data for large molecules, where long-ranged many-body effects become important. This work proposes a general approach to constructing accurate MLFFs for large-scale molecular simulations (GEMS) by training on "bottom-up" and "top-down" molecular fragments of varying size, from which the relevant physicochemical interactions can be learned. GEMS is applied to study the dynamics of alanine-based peptides and the 46-residue protein crambin in aqueous solution, allowing nanosecond-scale MD simulations of >25k atoms at essentially ab initio quality. Our findings suggest that structural motifs in peptides and proteins are more flexible than previously thought, indicating that simulations at ab initio accuracy might be necessary to understand dynamic biomolecular processes such as protein (mis)folding, drug-protein binding, or allosteric regulation.

Motivation & Objective

• To overcome the limitations of existing machine-learned force fields in modeling large, heterogeneous biomolecules with long-range and many-body interactions.
• To develop a generalizable framework that integrates both bottom-up and top-down fragment training to improve transferability and accuracy in complex systems.
• To enable long-timescale, high-accuracy molecular dynamics simulations of biomolecules—such as peptides and proteins—using MLFFs trained on quantum-mechanical reference data.
• To assess whether ab initio-level accuracy is necessary to capture dynamic biomolecular processes like protein folding and allostery.

Proposed method

• The GEMS framework trains a machine learning model on ab initio reference data from both small molecular fragments (bottom-up) and larger, representative structural motifs (top-down) to learn multi-scale interactions.
• The model uses a deep learning architecture—specifically, a message-passing neural network (e.g., SchNet) trained on DFT-computed energies and forces for diverse molecular configurations.
• Fragment-based training ensures that short-range, covalent, and non-covalent interactions are accurately captured across different chemical environments.
• The method combines fragment-derived representations with global system context to reconstruct the potential energy surface of large biomolecules with ab initio accuracy.
• Simulations are performed using the SchNetPack MD toolbox with a 0.5 fs timestep and no bond constraints, enabling fully unconstrained dynamics.
• Temperature and pressure control are enforced via Nosé-Hoover chain thermostats and Parrinello-Rahman barostats in NVT and NPT ensembles, respectively.

Experimental results

Research questions

• RQ1 Can machine-learned force fields trained on multi-scale fragments achieve ab initio-level accuracy for large biomolecular systems?
• RQ2 How do the dynamics of peptides and proteins simulated with GEMS compare to those obtained with classical force fields like AmberFF?
• RQ3 To what extent do long-range and many-body interactions influence protein conformational flexibility and folding pathways?
• RQ4 Is ab initio-level accuracy necessary to correctly describe dynamic biomolecular processes such as protein folding and ligand binding?

Key findings

• GEMS enables nanosecond-scale molecular dynamics simulations of systems with over 25,000 atoms in aqueous solution at ab initio accuracy, previously unattainable with standard DFT methods.
• Simulations of AceAla15Nme show that the folding pathway involves transient hydrogen-bonded intermediates, with the helical conformation stabilized by backbone H-bonding patterns.
• Crambin simulations with GEMS reveal significantly greater structural fluctuations compared to AmberFF, indicating that classical force fields may underestimate protein flexibility.
• The method achieves high accuracy in predicting energies and forces, with GEMS outperforming AmberFF across a range of conformational states of AceAla15Nme.
• The use of both bottom-up and top-down fragments allows the model to generalize across diverse chemical environments and capture long-range effects critical in biomolecular dynamics.
• The framework is scalable and applicable to complex systems such as the ACE2/SARS-CoV-2 RBD complex, enabling simulations of variant-specific mutations with high fidelity.