SP4: In silico methodsPartner 16A EMBL (Russell, Bork)Partner 1 (CRG Serrano)Partner 5 (NKI Perrakis)Partner 10 (HU Margalit)Partner 12 (CCNet)Partner 17 IRB (Aloy)Partner 3A (Paris-Sud, Janin)

SP4 In silico methods

• WP4.1: Target identification & annotationPartners: EMBL-Bork/Russell, HU, CCNet, IRB

• WP4.2: Complex modelingPartners: EMBL-Russell, IRB, Gif, CRG

• WP4.3: Interface to the scientific community & scientific data managementPartners: NKI, EMBL-Russell, CCNet

WP4.1: Target identification & annotation

Partners EMBL-HD, HU, CCN, IRB

Activities in:• Interaction prediction (HU, EMBL-Bork)• Complex prediction & ranking, the ‘list of 20’

(IRB)• Complex visualisation (CCNet/EMBL)• Data gathering (CCNet)

(e.g. protein-chemical interactions)• Gel processing (EMBL)

Complex database, web interface & wiki

Matthew Betts (EMBL)

The list of 20

Aloy group (IRB)

Experimental tests on the 20

Aloy, Seraphin, van Tilburgh & Dziembowski groups

WP4.2: Complex modelingPartners EMBL-HD, IRB, Gif, CRG

Activities in:• Complex modelling

– Automated procedures (EMBL-Russell/IRB)– Interaction prediction via structure (EMBL-Russell/IRB)

• New methods for modelling– FoldX (CRG/EMBL)– High-throughput Docking (IRB)

• Analyses, individual models (Everybody)

Building a complex from pieces

Aloy et al, Curr. Opin. Struct. Biol, 2005.

For 636 complexes in yeast 3505 : proteins modelable 419 : complexes single subunit models 224 : 2+ subunit models 122 : 3+ subunit models

Damien Devos (EMBL)

Example: MCM complex model agrees with existing EM data

eIF2 /eIF2B complex

Preliminary reconstructions (Bettina Boettcher)

GCD7GCD2

GCN3

SUI2

SUI4

SUI3

GCN3

GCD6

GCD1

SUI4

GCD7

SUI4SUI3

GCD1

Sub-complexes

Intact complex MS (Carol Robinson)

Modelling (Damien Devos)

The 3D-Repertoire Model Gallery

New interactions and principles arising from 3DR SP4

Defining new interfaces:424 candidate interfaces to date

Complex 1

Complex 2 Complex 3

New complex

Complex 1,2 & 3

Complex 1Complex 2

Complex 1 & 2

New complex

SuperimpositionSuperimposition

A common shape denotes a similar fold

Example: Transcription factor SPX dimer

Domain 1, 1z3eA.c.47.1.12-1-trans3 (chain A) Transcriptional regulator SPX (B.subtilis) Domain 2, 1z3eA.c.47.1.12-1-trans4 (chain B) Transcriptional regulator SPX (B.subtilis) Domain 3, 1z3eB.a.60.3.1-1-trans1p (chain C) RNA polymerase alpha (B.subtilis) Domain 4, 1z3eB.a.60.3.1-1-trans2p (chain D) RNA polymerase alpha (B.subtilis) Domain 5, 1lb2E.a.60.3.1-1-trans1 (chain E) RNA polymerase alpha (E.coli) Domain 6, 1lb2B.a.60.3.1-1-trans2 (chain F) RNA polymerase alpha (E.coli)

E.coli dimer in one protein, forms nice interface in B.subtilis – good evidence from other sources (Myco TAP)

Complex 1

Complex 2

Complex 3

New interface

Enabling/disabling loops can predict mulimerization state

Homodimer E. Coli Guanylate kinase

Monomer S. Cerevisiae

Guanylate kinase

E. Coli Guanylate kinase

V. Cholerae Guanylate kinase

Yeast Guanylate kinase

Mouse Guanylate kinase

Pig Guanylate kinase

Bovine Guanylate kinase

Homodimers

enabling loop

Monomers

When modelling fails – docking?

• The Aloy group (IRB) is currently running many tens of thousands of docking experiments using Mare nostrum, the largest supercomputer in Europe

• Aim is to identify promising docking candidates to help model key interactions of interest

Modelling versus docking• We can model an interaction structure if there is a

previously determined structure containing parts homologous to the two interacting proteins

• We can predict an interaction structure by docking if we have structures or models for parts of the interacting proteins

homology

homology

Large-scale Docking36 million possible yeast protein interactions

Large-scale Docking

Proteins Interactions all vs all 100

CPUs 1000 CPUs

100 CPUs

1000 CPUs

pdb 90% 8899 9645 39596100 24747,5 2474,7 197980,5 19798,0 pdb 30% 5224 5836 13645088 8528,2 852,8 68225,4 6822,5

Interaction types 2782 3737 3869762 2418,6 241,8 19348,8 1934,8 yeast xray orfs 79 97 3120,5 1,9 0,2 15,6 1,6

Unrefined Refined

Using FoldX to assess docked or modelled interactions

Good Interaction, but many clashes, Model is not so good but could be rescuedBy backbone moves/further docking

WP4.3: Interface to the scientific community and scientific data management

Partners: EMBL-HD, NKI, [CCN]

Activities in:• Web site maintenance

– New data (copy number, structural annotation)– Various optimisation– YeastWiz

• Complex target DB– Resting period for software development– Needs data. Listen to Tassos.

WP4.3: Web site

Matthew Betts (EMBL)

Yeast Wiz (CCNet)

Interactions of known structure

www.3drepertoire.org/yeastwizWindows XPLinuxManual (rather beta)

Accounts enabled Monday for everybodySuggestions for new data promising

Matthew Betts (EMBL), Tomasz Ignasiak (CCNet)

Mediator complex in yeastwiz (7 proteins from yesterday – two clicks)

3D

3D

Current data contributions

CCN-DB

EMBL HU

IRB

STRING (interactions)SMART (orthologues)3D Interaction predictionsOrthologyModelsEtc.

Protein-protein (>10 sources, manual)Protein-chemical (Manual, TM)

Dom-dom profilesContext interactionsEnabling loops

Docking solutionsList of 20

3DR

WP4.3: Interface to the scientific community and scientific data management

WIKI

TargetDB

Need target db stuff

The exosome

Damien Devos (w Carol Robinson)

Analysis of a “gold set” of 61 models of known interactions by FoldX

Easy case : Good Interaction Energy, few clashesGood Model

Analysis of a “gold set” of 61 models of known interactions by FoldX

Bad Interaction, loads of clashes, interpenetrating mainchainsBad Model

Analysis of a “gold set” of 61 models of known interactions by

FoldX

Bad Interaction, many clashes, but Model could be rescued by some backbone moves/ further docking

Analysis of a “gold set” of 61 models of known interactions by

FoldX

Good Interaction, many clashes, interpenetrating mainchains, gaps in the structureBad Model

Analysis of a “gold set” of 61 models of known interactions by FoldX

Easy case : Good Interaction Energy, few clashes : Good Model

Bad Interaction, many clashes - interpenetrating mainchains, gaps in the structure : Bad Model- mainchains too close on a large region but this can be solved bybackbone moves/further docking (could improve the model?)

Good Interaction, many clashes: - interpenetrating mainchains, gaps in the structure : Bad Model- mainchains too close on a large region but this can be solved bybackbone moves/further docking

The magnitude of the local clashes correlate with the possibility to rescue or not a model (mild clashes on a lot of residues), but still there are exceptions.

Could we really skip a step of visualization?

From protein-protein interactions to domain-domain interactions and back

Hanah MargalitThe Hebrew University of Jerusalem

domain pairs

protein-protein interactions

Modularity in protein-protein interactions

fine tuners Yes YesNo No

positive datasetreliable protein-protein interactions

reliable pairs of proteins that do not interact

negative dataset

What are the fine tuners of domain-domain recognition?

Homodimers and monomers provide an ideal dataset

Domains that mediate homodimerization Domains that mediate homodimerization are found also in monomersare found also in monomers

homodimers: co-localized co-expressed interact

monomers: co-localized co-expressed do not interact

Database of 50 homodimers/monomerswith the same domain for which structural data is available

Phosphorylations PP

monomersInterface residuesubstitutions

Different fine-tuners determine theself-interaction potential of domains

homodimers

Enabling loops mediate homodimerization

Homodimer E. Coli Guanylate kinase

Monomer S. Cerevisiae

Guanylate kinase

E. Coli Guanylate kinase

V. Cholerae Guanylate kinase

Yeast Guanylate kinase

Mouse Guanylate kinase

Pig Guanylate kinase

Bovine Guanylate kinase

Homodimers

enabling loop

Monomers

Disabling loops prevent homodimerization

Monomer: Bovine inositol polyphosphate 1-phosphatase

Homodimer: Bovine inositol monophosphatase

DL

Monomers

Homodimers

Loop profiles

A multiple-sequence alignment with locations of potential loops

Presence AND absence are informative

monomerprotein 2

homodimerprotein 1

enablingloop

disablingloop

The ‘core set’

64 / 73 are consistent (88%, p-value ≤ 3.2•10-6)

Test set

experimental oligomeric state

loop profile

monomer

dimer

72/80 are consistent (90%, p-value ≤ 5•10-6)

80 proteins with documented oligomeric state based on experimental data

monomer

95

dimer

363

Large-scale prediction of domain-domain interaction

pfkB carbohydrate kinase domain proteins

EL

DL DL DL

108 homodimers

31 monomers

monomer

homodimer

core

Homodimer

Monomertest

>1000predictable

Boundary loops

There are enabling/disabling loops that are located outside domain boundaries

Metallo-beta-lactamase domain

Dominance of disabling over enabling loops

RNase Z (B. Subtilis)ccrA (B. Fragilis)

Summary

1. Enabling/disabling loops are newly discovered fine-tuners of domain-domain interaction

2. Their presence/absence is highly preserved in evolution, implying that prevention of unwanted interactions is an evolutionary constraint

3. Prediction of self-interaction potential of domains according to loop profiles is highly

accurate (~90%)

Extension of the analysis

Homodimers of multi-domain proteins

Heterodimers of proteins with self-interacting domains

Heterodimers of proteins with different domains