During the past few years the field of deep learning (DL) has been rapidly improving due mostly to advances in hardware (increasing GPU power) and also due to new theoretical developments in design and architecture of deep neural networks. This progress has made it possible to apply DL tools to tackle scientific problems previously out of reach. The field of molecular dynamics is an example of a domain where DL has the potential to introduce important developments, and fast atomistic neural network potentials have already been shown to reproduce the accuracy of density functional theory energy and force calculations at a fraction of the computational cost. The Roitberg group has developed one such family of potentials, the ANI family (implemented in C++/CUDA code and PyTorch). The potential energy surface predicted by these potentials is smooth even for dihedrals and nonbonded interactions, and agrees well with the underlying level of theory used for training. Here I present a recently implemented intuitive and lightweight interface that joins these neural networks with the GPU accelerated AMBER molecular dynamics package, and show preliminary results of simulations of biomolecules in explicit solvent using refined ANI models through said interface.

Introducing AMBER + TorchANI, a deep learning model for molecular dynamics of biomolecules

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.