Novel deep learning energy-based model to predict protein dynamics

Abstract

We have developed a deep learning (DL) model employing artificial neural network that can learn and predict the potential energy surface (PES) of a protein ensemble from a molecular dynamics (MD) trajectory alone. Our unique model, which uses a state-of-the-art DL method called transformer, identifies important atoms only from Cartesian coordinates of alpha carbons, without including explicitly bond, angle, and dihedral information. Instead, it uses a DL based classical potential energy (DL energy) to predict PES of a protein ensemble.