Harnessing protein folding neural networks for peptide–protein docking

3.0 科研～小助 2025-09-01 5 4 2.79MB 12 页 1知币

侵权投诉

ARTICLE

Harnessing protein folding neural networks for

peptide–protein docking

Tomer Tsaban 1,2, Julia K. Varga1,2, Orly Avraham1,2, Ziv Ben-Aharon 1, Alisa Khramushin1&

Ora Schueler-Furman 1✉

Highly accurate protein structure predictions by deep neural networks such as AlphaFold2

and RoseTTAFold have tremendous impact on structural biology and beyond. Here, we show

that, although these deep learning approaches have originally been developed for the in silico

folding of protein monomers, AlphaFold2 also enables quick and accurate modeling of

peptide–protein interactions. Our simple implementation of AlphaFold2 generates

peptide–protein complex models without requiring multiple sequence alignment information

for the peptide partner, and can handle binding-induced conformational changes of the

receptor. We explore what AlphaFold2 has memorized and learned, and describe speciﬁc

examples that highlight differences compared to state-of-the-art peptide docking protocol

PIPER-FlexPepDock. These results show that AlphaFold2 holds great promise for providing

structural insight into a wide range of peptide–protein complexes, serving as a starting point

for the detailed characterization and manipulation of these interactions.

https://doi.org/10.1038/s41467-021-27838-9 OPEN

1Department of Microbiology and Molecular Genetics, Institute for Biomedical Research Israel-Canada, Faculty of Medicine, The Hebrew University of

Jerusalem, Jerusalem, Israel.

These authors contributed equally: Tomer Tsaban, Julia K. Varga, Orly Avraham. ✉email: ora.furman-schueler@mail.huji.ac.il

NATURE COMMUNICATIONS | (2022) 13:176 | https://doi.org/10.1038/s41467-021-27838-9 | www.nature.com/naturecommunications 1

1234567890():,;

Peptide–protein interactions are highly abundant in living

cells and are important for many biological processes1.Itis

estimated that up to 40% of interactions in cells are medi-

ated by peptide–protein interactions, or peptide-like interaction:2

short segments, isolated or embedded within unstructured regions

that mediate binding to a partner3. In addition, peptides are often

used for biotechnological applications, drug delivery, imaging, as

therapeutic agents, and other applications4,5, by binding proteins

and mediating or blocking interactions.

Determining the 3-dimensional structure of these peptide–-

protein complexes is an important step for their further study.

They can provide the basis to identify hotspot residues that are

crucial for binding6–8, and by mutating these hotspots, the

functional importance of a given interaction can be uncovered9.

They could help to better understand disease-causing mutations

and also serve as a starting point for the design of strong and

stable peptidomimetics10,11.

However, peptide-mediated interactions pose signiﬁcant chal-

lenges, both for their experimental as well as their computational

characterization: These interactions are in many cases weak,

transient, and considerably inﬂuenced by their context, resulting

in often noisy experiments. Widely used structure determination

methods (e.g., X-ray crystallography) are not applicable to many

of these interactions. Computational modeling, and particularly

blind peptide–protein docking12, is hindered by the lack of

known structure for the peptide side, in contrast to classical

domain-domain docking, where the structure of the free indivi-

dual domains is usually deﬁned. In order to succeed in the study

and design of peptide–protein interactions, we must gain a better

understanding of the peptide conformational preferences.

One way to approach this challenge is based on the observa-

tion that a peptide bound conformation is often present in solved

monomer structures13. Based on this ﬁnding, we developed the

high-resolution blind peptide docking protocol, PIPER-

FlexPepDock (PFPD)13. First, a representative ensemble of

fragments is extracted from monomer structures using the

Rosetta Fragment Picker14, which takes into account both

sequence and (predicted) secondary structure similarity. Then

this ensemble is rigid-body docked onto the receptor with the

PIPER protocol15, followed by short local reﬁnement by Rosetta

FlexPepDock16, which simultaneously optimizes internal peptide

and rigid-body degrees of freedom. Numerous other peptide

docking approaches have since been developed12,17,many

focusing on efﬁcient low-resolution docking18,19, others lever-

aging information about protein interfaces to ﬁnd matches for

similar interface patches20–22.

Another way to approach the global peptide docking challenge

is to view the binding of a peptide to its partner as the ﬁnal step of

protein folding, complementing the receptor surface with a

missing piece23. Indeed, functional proteins can be reconstituted

experimentally from short fragments of the original sequence,

indicating that covalent linkage is not necessarily a prerequisite for

monomer folding24,25. We and others have successfully modeled

peptide–protein interactions using this principle, by ﬁnding frag-

ments in monomer structures and on protein-protein interfaces

that could complement structural patches derived from the surface

of a given receptor20–22,26. These concepts lay the groundwork for

novel approaches in peptide–protein docking, where the vast

information inherently stored in folded monomer structures is

efﬁciently integrated in the search space for peptide docking.

The advances in the ﬁeld of protein structure prediction in

recent years open up exciting opportunities to fully leverage such

information. The development and application of deep learning

(DL) neural network (NN) architectures to predict monomeric

protein structures provided us with highly accurate computa-

tional models as particularly showcased by the last CASP14

experiment27. AlphaFold2 (AF2) developed by Google Deepmind

was able to generate models of exceptional accuracy, approaching

the resolution of crystallography experiments28. Signiﬁcantly

improved modeling was also reported for RoseTTAFold, devel-

oped by RosettaCommons, that followed ideas from AF2 and also

implemented fully continuous crosstalk between 1D, 2D and 3D

information29. Most importantly, AF2, as well as RoseTTAFold,

are now freely available to the scientiﬁc community30,31, opening

up powerful avenues for protocol development and applications

to many biological systems that were not amenable to structural

characterization in the past. These are truly exciting times!

Can such NNs also model peptide–protein interactions, and not

only monomers? If peptide–protein interfaces are indeed abundant

in monomer structures, and if indeed peptide–protein interactions

can be captured as protein folding as stated above, RoseTTAFold

and AF2 should, in principle, also allow for the modeling of

peptide–protein complex structures. Moreover, they could alleviate

the lack of data impairing the ability to fully employ DL for

peptide–protein interactions. We note that both RoseTTAFold and

AF2 NNs were trained on single chain protein structural data, and

both use Multiple Sequence Alignments (MSA) as a critical step in

structure prediction. Prediction of protein-protein complexes was

shown to be possible given an informative MSA27,29,32, and it has

also been explored whether it is indeed necessary to provide paired

sequences for successful extraction of interface information33,34.As

both methods heavily rely on good quality MSA, the main chal-

lenge would be to accurately predict the peptide conformation.

Mainly due to their short length, creating an effective MSA for

these regions is challenging.

Here we present a global peptide–protein docking approach

that incorporates the biological concept of peptide–protein

interactions mimicking protein folding and harnesses NNs

trained to predict monomeric protein structures. We show that

by connecting the peptide to the receptor (e.g., by a poly-glycine

linker), monomer folding NNs generate accurate peptide–protein

complex structures (a similar idea was proposed in parallel by

others35). This is possible thanks to the ability of AF2 to (1)

accurately identify unstructured regions36 and model these as

extended linkers, and (2) predict peptide-receptor complexes

without a multiple sequence alignment for the peptide partner,

as we demonstrate in this study. Best performance is obtained

by combining our linker-based strategy with modeling of

peptide–protein complexes by presenting two separate chains to

AF2. The latter has been implemented for the modeling of homo-

and hetero-multimers in several recent studies on AF236,37.

We perform a short calibration on a small representative,

previously well-studied set of protein-peptide interactions, con-

sisting of peptides with and without known binding motifs13.We

then provide a detailed comparison to the currently top-

performing global peptide docking protocol PFPD13. We then

assess the protocol on an extensive, non-redundant set of curated

peptide–protein complexes consisting of 96 interactions, each

involving a distinct fold. Finally, we explore speciﬁc types of

interactions of special interest, including examples in which

peptide binding induces a large conformational change in the

receptor upon binding. The latter are very challenging to model

using docking, but easily amenable to AF2 which models the

complex as a whole. Beyond presenting an approach to dock

peptides, this study provides another view on what AF2 may have

learned beyond memorization.

Results

Adapting NN-based structure prediction to peptide docking.

By adding the peptide sequence via a poly-glycine linker to the

C-terminus of the receptor monomer sequence, we mimicked

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27838-9

2NATURE COMMUNICATIONS | (2022) 13:176 | https://doi.org/10.1038/s41467-021-27838-9 | www.nature.com/naturecommunications

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

1 知币 4人已下载

立即下载

Harnessing protein folding neural networks for peptide–protein docking.pdf

共12页,预览4页

还剩页未读，继续阅读

Harnessing protein folding neural networks for peptide–protein docking

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

推荐作者

热门标签

举报选择: