[Paper Review] Machine-learned molecular mechanics force field for the simulation of protein-ligand systems and beyond

Kenichiro Takaba, Iván Pulido+12 more|arXiv (Cornell University)|Jul 13, 2023

Protein Structure and DynamicsBiochemistry, Genetics and Molecular Biology114 references13 citations

TL;DR

Espaloma-0.3 is a graph-neural-network–based, end-to-end differentiable Class I MM force field that learns from large-scale quantum chemical data to accurately reproduce energies and forces, and to self-consistently parametrize proteins and ligands.

ABSTRACT

The development of reliable and extensible molecular mechanics (MM) force fields -- fast, empirical models characterizing the potential energy surface of molecular systems -- is indispensable for biomolecular simulation and computer-aided drug design. Here, we introduce a generalized and extensible machine-learned MM force field, exttt{espaloma-0.3}, and an end-to-end differentiable framework using graph neural networks to overcome the limitations of traditional rule-based methods. Trained in a single GPU-day to fit a large and diverse quantum chemical dataset of over 1.1M energy and force calculations, exttt{espaloma-0.3} reproduces quantum chemical energetic properties of chemical domains highly relevant to drug discovery, including small molecules, peptides, and nucleic acids. Moreover, this force field maintains the quantum chemical energy-minimized geometries of small molecules and preserves the condensed phase properties of peptides, self-consistently parametrizing proteins and ligands to produce stable simulations leading to highly accurate predictions of binding free energies. This methodology demonstrates significant promise as a path forward for systematically building more accurate force fields that are easily extensible to new chemical domains of interest.

Motivation & Objective

• Motivate the need for more accurate and extensible molecular mechanics force fields for biomolecular simulations and drug discovery.
• Introduce a scalable, differentiable framework (espaloma-0.3) that replaces discrete atom typing with continuous representations learned from quantum chemistry data.
• Demonstrate that espaloma-0.3 can extend to new chemical domains (small molecules, peptides, nucleic acids) and preserve quantum chemical minima and condensed-phase properties.
• Show that the approach yields superior accuracy in reproducing quantum chemical energies and forces compared with traditional force fields.

Proposed method

• Use graph neural networks to generate continuous atomic embeddings from chemical graphs, replacing discrete atom typing.
• Employ symmetry-preserving pooling (Janossy pooling) to produce continuous bond, angle, and torsion representations.
• Train neural networks to predict MM parameters (bonds, angles, torsions, charges) from invariant embeddings in an end-to-end differentiable setup.
• Incorporate quantum chemical forces into training and regularization strategies to stabilize learning and improve transferability.
• Retain OpenFF 2.0 parameters for non-valence terms while predicting valence terms and charges to enable rapid MM energy computation in standard MD packages.
• Generate AM1-BCC-like partial charges via a charge equilibration scheme and target AM1-BCC ELF10 charges for training.

Figure 1 : Espaloma is an end-to-end differentiable molecular mechanics parameter assignment scheme for arbitrary organic molecules. Espaloma ( e xtensible s urrogate p otenti al o ptimized by m ess a ge-passing ) is a modular approach for directly computing molecular mechanics force field parameter

Experimental results

Research questions

• RQ1 Can espaloma-0.3 provide a self-consistent, differentiable framework to parametrize proteins and ligands within a single force field?
• RQ2 Does training on a large, diverse QC dataset enable accurate reproduction of quantum chemical energies and forces across small molecules, peptides, and nucleic acids?
• RQ3 Does the approach preserve quantum chemical energy minima and condensed-phase properties to support reliable biomolecular simulations?
• RQ4 What is the performance of espaloma-0.3 compared with established force fields (GAFF, OpenFF, Amber) on relevant benchmarks?
• RQ5 How well can the model extend to new chemical domains without a rise in complexity or loss of accuracy?

Key findings

• Espaloma-0.3 outperforms traditional force fields in reproducing quantum chemical energies and forces across diverse datasets.
• The model preserves quantum chemical energy minima and yields MM-optimized geometries that are close to QM references on industry benchmarks.
• Espaloma-0.3 maintains condensed-phase properties for peptides and self-consistently parametrizes proteins and ligands for stable simulations.
• The approach enables rapid retraining and extension to new chemical domains without performance penalties, outperforming baselines even when considering ionization and resonance-related concerns.
• Training on a large QC dataset (1.1M energy/force calculations, 17,000 species) can be completed in a single GPU day, highlighting scalability.
• Compared to ff14SB and RNA.OL3 baselines, espaloma-0.3 shows superior energy and force RMSEs across multiple biomolecular categories.

Figure 2 : espaloma-0.3 preserves the location of quantum chemical energy minima. An industry standard benchmark of gas-phase QM-optimized geometries (the OpenFF Industry Benchmark Season 1 v1.1 [ 20 ] from QCArchive), comprising 9728 unique molecules and 73 301 conformers, was used to compare the s