CCL: TorchMD - A deep learning framework for molecular simulations

 Sent to CCL by: "Franck  Chevalier" [f.chevalier:-:acellera.com]
 This week our paper on TorchMD has been published in the Journal of Chemical
 Theory and Computation
 (https://www.acellera.com/md-simulation-blog-news/). TorchMD
 is a framework for molecular simulations
 that enables users to do research faster in force-field development as well as
 integrate neural network
 potentials seamlessly into the dynamics with the simplicity and power of
 PyTorch. It is the result of the
 collaboration of Acellera with Universitat Pompeu Fabra, Freie Universitt
 Berlin,  Rice University, CNR-
 IBF- Italy, Universit degli Studi di Milano and ICREA.
 Molecular dynamics simulations provide a mechanistic description of molecules by
 relying on empirical
 potentials. The quality and transferability of such potentials can be improved
 by leveraging data-driven
 models derived with machine learning approaches. With TorchMD, we aim to provide
 the glue which
 connects classical molecular simulations with machine learning methods. All
 force computations
 including bond, angle, dihedral, Lennard-Jones and Coulomb interactions are
 expressed as PyTorch
 arrays and operations. It also enables learning and simulating with neural
 network potentials. TorchMD
 has been validated 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.
 What will come next?
 The end-to-end differentiability of parameters, which TorchMD offers, is a
 feature that projects such as
 the Open Force Field Initiative can already start exploiting. For faster,
 production-ready simulations, we
 are working on facilitating the integration of machine learning potentials in
 OpenMM and ACEMD3.
 Meanwhile, we believe that TorchMD can play an important role by facilitating
 experimentation between
 ML and MD fields, speeding up the model train-evaluate prototyping cycle, and
 promoting the adoption of
 data-based approaches in molecular simulations. All the code used to produce and
 evaluate the models
 is available for practitioners at github.com/torchmd. We invite you to use it
 and drop us a line on your
 experience.