OpenMM 8: Molecular Dynamics Simulation with Machine Learning Potentials.

Area of Science:

• Computational chemistry and biophysics
• Molecular dynamics simulations
• Machine learning applications in science

Background:

• Machine learning (ML) is increasingly vital in molecular simulation.
• Traditional simulation methods face limitations in accuracy and computational cost.
• The OpenMM toolkit is a widely used platform for molecular dynamics.

Purpose of the Study:

• Introduce new machine learning (ML) capabilities in the OpenMM toolkit.
• Enable the integration of arbitrary PyTorch models for force and energy calculations.
• Provide a user-friendly interface for applying ML potentials in simulations.

Main Methods:

• Integration of arbitrary PyTorch models within OpenMM for force field calculations.
• Development of a higher-level interface for utilizing general-purpose, pretrained ML potentials.
• Implementation of optimized CUDA kernels and custom PyTorch operations for enhanced simulation speed.

Main Results:

• Demonstrated successful simulations of cyclin-dependent kinase 8 (CDK8) and a green fluorescent protein chromophore.
• Achieved significant speed improvements in molecular dynamics simulations.
• Showcased the practical application of ML potentials for improved simulation accuracy.

Conclusions:

• The latest OpenMM version facilitates the practical use of machine learning in molecular simulations.
• ML potentials offer a way to enhance simulation accuracy without substantial computational overhead.
• These advancements make ML-driven molecular dynamics more accessible and efficient.