DiffNets: Self-supervised deep learning to identify the mechanistic basis for biochemical differences between protein variants

Abstract

A mechanistic understanding of how mutations modulate proteins’ biochemical properties would advance our understanding of biology, provide insight for engineering proteins with particular functions, and facilitate efforts in precision medicine. However, such mechanistic insight remains elusive in many cases. For example, experimentally-derived structures of protein variants with dramatically different behaviors are often nearly identical, suggesting that one must consider the entire ensemble of structures that a protein adopts. Molecular dynamics (MD) simulations provide access to such ensembles, but identifying the relevant features of these complex entities remains difficult. Here we develop DiffNets, a deep learning framework that combines supervised autoencoders with expectation maximization to identify the structural preferences that are responsible for the biochemical differences between protein variants. As a proof of principle, we show that DiffNets identify the important structural preferences that confer increased stability to TEM β-lactamase variants without any a priori knowledge of the relevant structural features.