Molecular Dynamics and Generative Neural Networks to Explore Protein Conformational Space

Determining the different conformational states of a protein and the transition paths between them is key to fully understanding the relationship between biomolecular structure and function. This can be accomplished by sampling protein conformational space with molecular simulation methodologies. Despite advances in computing hardware and sampling techniques, simulations always yield a discretised representation of this space, with transition states undersampled proportionally to their associated energy barrier. We present neural networks that learn a continuous conformational space representation from example structures. We define a physics-based loss function that, applied to network generated structures, discourages physically implausible conformations. We show that this network, trained with simulations of distinct protein states, can correctly predict a biologically relevant transition path, without any example on the path provided. We also show we can transfer features learnt from one protein to others, which results in superior performances, and requires a surprisingly small number of training examples. Our methods are compiled into the software package molearn that aims to lower the barrier to entry for training and analysing these networks while simultaneously providing a platform for rapid development within our framework. In particular we present a version of our loss function using OpenMM to efficiently calculate energies and forces. Using molearn we profile our networks showing that they are able handle different atom selections and multimers. Additionally, we demonstrate the ability of this loss function to effectively ‘deblur’ Sinkhorn trained networks. Overall, this work demonstrates how generative models can be effectively exploited to characterise proteins conformational spaces, thereby providing precious insight into their biological function.