TorchMD: A Deep Learning Framework for Molecular Simulations.

Area of Science:

• Computational chemistry
• Molecular modeling
• Machine learning in science

Background:

• Molecular dynamics (MD) simulations offer mechanistic insights into molecular behavior using empirical potentials.
• The accuracy and generalizability of these potentials can be enhanced through data-driven machine learning (ML) models.

Purpose of the Study:

• Introduce TorchMD, a novel framework for molecular simulations.
• Facilitate the integration of classical and ML potentials within a unified simulation environment.
• Enable the development and application of neural network potentials (NNPs) in molecular simulations.

Main Methods:

• Developed TorchMD using PyTorch, expressing all force computations (bond, angle, dihedral, Lennard-Jones, Coulomb) as PyTorch arrays and operations.
• Integrated capabilities for learning and simulating NNPs.
• Validated the framework through standard Amber all-atom simulations, ab initio potential learning, end-to-end training, and coarse-grained protein folding simulations.

Main Results:

• Demonstrated the flexibility of TorchMD in handling diverse simulation types, from all-atom to coarse-grained models.
• Successfully learned and simulated systems using NNPs, showing potential for improved accuracy.
• Validated the framework's performance against established simulation methods.

Conclusions:

• TorchMD provides a versatile and powerful toolset for molecular simulations incorporating ML potentials.
• The framework supports the advancement of NNP development and application in computational chemistry.
• TorchMD is publicly available, encouraging wider adoption and contribution within the research community.