Learning coevolutionary models for protein design

Abstract

The patterns of residue coevolution encoded in protein sequence alignments have been successfully exploited for protein structure prediction. Here, we explore an inverse problem: given a protein structure, can we infer the coevolutionary patterns which allow sequences to fold to that structure? To do so, we build on previous work on the coevolutionary design of proteins folding to a given structure, combining it with tools of deep learning. We evaluate different machine learning approaches and neural network architectures for the task of learning sequence information from protein structure.