Deep learning of Protein Sequence Design of Protein-protein Interactions

Abstract

MotivationAs more data of experimentally determined protein structures is becoming available, data-driven models to describe protein sequence-structure relationship become more feasible. Within this space, the amino acid sequence design of protein-protein interactions has still been a rather challenging sub-problem with very low success rates - yet it is central for the most biological processes.ResultsWe developed an attention-based deep learning model inspired by algorithms used for image-caption assignments for sequence design of peptides or protein fragments. These interaction fragments are derived from and represent core parts of protein-protein interfaces. Our trained model allows the one-sided design of a given protein fragment which can be applicable for the redesign of protein-interfaces or the de novo design of new interactions fragments. Here we demonstrate its potential by recapitulating naturally occurring protein-protein interactions including antibody-antigen complexes. The designed interfaces capture essential native interactions with high prediction accuracy and have native-like binding affinities. It further does not need precise backbone location, making it an attractive tool for working with de novo design of protein-protein interactions.AvailabilityThe source code of the method is available at https://github.com/strauchlab/iNNterfaceDesignSupplementary informationSupplementary data are available at Bioinformatics online.

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

“…The procedure of increasing of the length of the fragment can be repeated a few times on either end. Amino acid sequence (AAS) design for the new backbones is then performed by the deep NN-based PepSeP1 or PepSeP6 methods 21,22 . These two methods were specifically developed for iNNterfaceDesign.…”

“…Amino acid sequence (AAS) design for the new backbones is then performed by the deep NN-based PepSeP1 or PepSeP6 methods 21,22 . These two methods were specifically developed for iNNterfaceDesign.…”

“…Our design workflow consists of 3 methods within the iNNterfaceDesign pipeline. The first method, PepBB, predicts initial 6-residue backbone fragments of peptide ligands; second PepBBE elongates the initial backbones and then the third model, PepSep1 or PepSep6 22 produces amino acid sequences for the designed motifs. We discussed the performance of the core part PepBB, which is the most crucial starting component in detail.…”

Deep learning of protein sequence design of protein–protein interactions

“…Perturbed backbones were either reverted to their native amino acid sequence (c3, annotated as native* within the main text to highlight the backbone perturbation) or designed with PepSeP1 (c4). HP (highly perturbed) backbones (c5) are generated using the iNNterfaceDesign method ( Syrlybaeva and Strauch, 2022 )…”

“…Besides simply testing the methods on all-glycine peptide ligands extracted from native interactions, we also tested them on artificially highly perturbed (HP) peptide variations of all-glycine peptide ligands as part of our case studies. Backbones were generated by the iNNterfaceDesign method ( Syrlybaeva and Strauch, 2022b , Supplementary Fig. S3 ).…”