One-sided design of protein-protein interaction motifs using deep learning

Abstract

Protein-protein interactions are part of most processes in life and thereby the ability to generate new ones to either control, detect or inhibit them has universal applications. However, to develop a new binding protein to bind to a specific site at atomic detail without any additional input is a challenging problem. After DeepMind entered the protein folding field, we have seen rapid advances in protein structure predictions thanks to the implementation of machine learning algorithms. Neural networks are part of machine learning and they can learn the regularities from their input data. Here, we took advantage of their capabilities by training multiple neural networks on co-crystal structures of natural protein complexes. Inspired by image caption algorithms, we developed an extensive set of NN-based models, referred to as iNNterfaceDesign. It predicts the positioning and the secondary structure for the new binding motifs and then designs the backbone atoms followed by amino acid sequence design. Our methods are capable of recapitulating native interactions, including antibody-antigen interactions, while they also capable to produce more diverse solutions to binding at the same sites. As it was trained on natural complexes, it learned their features and can therefore also highlight preferential binding sites, as found in natural protein-protein interactions. Our method is generally applicable, and we believe that this is the first deep learning model for one-sided design of protein-protein interactions.