Abstract

Harnessing protein folding neural networks for peptide–protein docking: what have we learned?

Peptide-mediated interactions play crucial roles in cellular regulation. Challenged by the flexibility of the peptide on the one hand and the often transient and weak character of the interaction on the other, they pose special challenges, both for modeling and experimental efforts. Recent advances in Deep Learning, such as by Deepmind Alphafold2, are revolutionizing computational structural biology, bringing to reach high accuracy models of full proteomes. 

Evidence has accumulated that the binding of peptides to their receptors could be seen as monomer complementation, i.e., a final step of monomer folding. Based on this concept, we have developed in the recent years two novel, top-performing peptide docking protocols as assessed on a comprehensive benchmark and validated set. The first, PatchMAN (1), applies a fast search to match patches on the receptor surface for structural  motifs in solved structures. These can then be used to extract complementing fragments that serve as starting point for peptide refinement. The second approach uses a slight modification of Alphafold2 to model the peptide either as separate unit, or connected by a poly-glycine linker to the c-terminus of the receptor (2).  Importantly, we succeed in modeling these interactions at high accuracy, even though no information on the multiple sequence alignment of the peptide partner is available.

In my presentation, I will shortly introduce these approaches and their performance, and then discuss the underlying reasons for success, and failure. This can teach us about the basic principles of this interesting and important type of interactions between proteins. In turn it also teaches us lessons on what Alphafold2 may have learned beyond memorization.

(1) Alisa Khramushin, Ziv Ben-Aharon, Tomer Tsaban, Julia K Varga, Orly Avraham, Ora Schueler-Furman (2022). Matching protein surface structural patches for high-resolution blind peptide docking. PNAS 2022. 119: e2121153119.
(2) Tomer Tsaban, Julia Varga, Orly Avraham Ziv Ben-Aharon, Alisa Khramushin, Ora Schueler-Furman (2022) Harnessing protein folding neural networks for peptide–protein docking. Nat Commun 2022, 13:2021.08.01.454656.