Harnessing protein folding neural networks for peptide-protein docking

Abstract

Highly accurate protein structure predictions by the recently published deep neural networks such as AlphaFold2 and RoseTTAFold are truly impressive achievements, and will have a tremendous impact far beyond structural biology. If peptide-protein binding can be seen as a final complementing step in the folding of a protein monomer, we reasoned that these approaches might be applicable to the modeling of such interactions. We present a simple implementation of AlphaFold2 to model the structure of peptide-protein interactions, enabled by linking the peptide sequence to the protein c-terminus via a poly glycine linker. We show on a large non-redundant set of 162 peptide-protein complexes that peptide-protein interactions can indeed be modeled accurately. Importantly, prediction is fast and works without multiple sequence alignment information for the peptide partner. We compare performance on a smaller, representative set to the state-of-the-art peptide docking protocol PIPER-FlexPepDock, and describe in detail specific examples that highlight advantages of the two approaches, pointing to possible further improvements and insights in the modeling of peptide-protein interactions. Peptide-mediated interactions play important regulatory roles in functional cells. Thus the present advance holds much promise for significant impact, by bringing into reach a wide range of peptide-protein complexes, and providing important starting points for detailed study and manipulation of many specific interactions.

Improved Docking of Protein Models by a Combination of Alphafold2 and ClusPro

“…A similar protocol was applied to protein-peptide docking by Ko and Lee [8] and by Tsaban et al [9]. Both groups linked the peptide sequence to the protein sequence via a polyglycine linker, and used AlphaFold2 without any modifications.…”

“…In some cases, the method obviously failed with the poly-glycine linker throwing the peptide segment into space. In other cases the models correctly identified the binding pocket, but showed errors of peptide rotation or translation [9]. Tsaban et al also compared the performance of AlphaFold2 to that of the physics-based peptide docking protocol PIPER-FlexPepDock [10] and observed an almost orthogonal behavior in terms of successes and failures.…”

A Perspective on the Prospective Use of AI in Protein Structure Prediction

“…94 A simple implementation already showed interesting predictive capacity, including in cases where the peptide induces a large conformational change of the protein and docking therefore most likely fails, and without the need for a peptide MSA. 95 AlphaFold-Multimer performs better than AF2 at protein− peptide complex prediction, and sampling a larger part of the conformational space by enforcing dropout at inference time in AlphaFold-Multimer further increased the quality of protein− peptide complex models. 96,97 Using a protein−peptide complex benchmark that is not redundant with the AF2 training set, AlphaFold-Multimer achieves only 40% success rate in identifying the correct site and structure of the interface when the full-length partners are used as input; combining input fragments of size 100 or 200 amino acids and different strategies for building the MSAs, this success rate can rise up to 90%.…”