End-to-end differentiable construction of molecular mechanics force fields

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Drug Discovery

Background:

• Traditional molecular mechanics (MM) potentials rely on inflexible, human-curated chemical perception rules (atom types).
• This limits the optimization of parameters and extensibility for new molecules.
• Existing methods struggle to efficiently integrate quantum chemical data for force field development.

Purpose of the Study:

• To develop an alternative approach for MM force field parameterization using graph neural networks (GNNs).
• To create a modular, end-to-end differentiable system for generating and extending force fields.
• To improve the accuracy and efficiency of parameter prediction and partial charge assignment.

Main Methods:

• Utilized graph neural networks to learn chemical environments and generate continuous atom embeddings.
• Employed invariance-preserving layers to predict valence and nonbonded parameters from atom embeddings.
• Developed an end-to-end differentiable pipeline for force field construction and optimization.

Main Results:

• The GNN approach successfully reproduced legacy atom types and accurately extended existing MM force fields.
• New force fields were constructed self-consistently for biomolecules and small molecules directly from quantum mechanics.
• The 'espaloma' package demonstrated superior accuracy in alchemical free energy calculations and faster partial charge prediction.

Conclusions:

• Graph neural networks offer a flexible and accurate alternative to traditional atom typing schemes for MM force fields.
• The developed method enables rapid, high-fidelity force field generation and parameterization.
• The 'espaloma' package provides an open-source tool for advancing computational chemistry and drug discovery.