TorchMD: A Deep Learning Framework for Molecular Simulations.

Stefan Doerr,Adrià Pérez,Frank Noe,Cecilia Clementi,Andreas Krämer,Gianni De Fabritiis,Maciej Majewski,Toni Giorgino

doi:10.1021/acs.jctc.0c01343

Abstract

Molecular dynamics simulations provide a mechanistic description of molecules by relying on empirical potentials. The quality and transferability of such potentials can be improved leveraging data-driven models derived with machine learning approaches. Here, we present TorchMD, a framework for molecular simulations with mixed classical and machine learning potentials. All force computations including bond, angle, dihedral, Lennard-Jones, and Coulomb interactions are expressed as PyTorch arrays and operations. Moreover, TorchMD enables learning and simulating neural network potentials. We validate it using standard Amber all-atom simulations, learning an ab initio potential, performing an end-to-end training, and finally learning and simulating a coarse-grained model for protein folding. We believe that TorchMD provides a useful tool set to support molecular simulations of machine learning potentials. Code and data are freely available at github.com/torchmd.

Highlights

Classical molecular dynamics (MD) is a compute-intensive technique that enables quantitative studies of molecular processes
We demonstrated TorchMD, a PyTorch-based molecular dynamics engine for biomolecular simulations with machine learning capabilities
We have shown several possible applications ranging from Amber all-atom simulations to endto-end learning of parameters and a coarse-grained neural network potential for protein folding

Summary

Introduction

Classical molecular dynamics (MD) is a compute-intensive technique that enables quantitative studies of molecular processes. The SchNet architecture,[4,5] for instance, learns a set of features using continuous filter convolutions on a graph neural network and predicts the forces and energy of the system. A key feature of using SchNet is that the model is inherently transferable across molecular systems This has been extended to learn a potential of mean force which involves averaging of a potential over some coarse-grained degrees of freedom,[6−12] which pose challenges in their parametrization.[13,14] molecular modeling on a more granular scale has been tackled by socalled coarse-graining (CG) approaches before,[15−20] but it is interesting in combination with DNNs

Methods

Results

Conclusion