predict protein-protein interaction sites - rostlab · © burkhard rost (tum munich) /110...

© Burkhard Rost (TUM Munich) /1101

title: Prediction of PPI2 Predict protein-protein interaction sitesshort title: pp2_ppi2

lecture: Protein Prediction 2 - Protein function TUM Winter 2013/2014


Announcements

Videos: SciVe / www.rostlab.orgTHANKS : Tim Karl + Jona ReebSpecial lectures:

• Oct 29: Tobias Hamp• Nov 21: Tanya Goldberg• Nov 28: Arthur Dong• Dec 03+05: Marco De Vivo/Marco Punta• Dec 17+19: Andrea Schafferhans

No lecture:• Oct 10 Thu (Reformation)• Nov 12 Tue (Student assembly)• Dec 12 Thu (TUM Dies Academicus)

LAST lecture: Jan 30Examen: Feb 4 or Feb 6 (likely this room)

• Makeup: Apr 9 - morning

CONTACT: Marlena Drabik [email protected]

2

http://www.rostlab.org

http://www.rostlab.org

mailto:[email protected]

mailto:[email protected]


IV.Predict protein interactions

3


ISIS

4


IV.1a protein interactionsProtein-protein

interactions (PPI):terminology

5


Different interfaces = different physics?

PD Kwong, R Wyatt, J Robinson, RW Sweet, J Sodroski & WA Hendrickson (1998) Nature 393, 648-659.PD Kwong, R Wyatt, S Majeed, J Robinson, RW Sweet, J Sodroski & WA Hendrickson (2000) Structure 8, 1329-1339.

gp120

CD4

antibody-1antibody-2

HIV gp120 / CD4 / FAB

6


Protein association

7

A activates B activates C activates D activates ....

ABCDare

associated


Physical interaction NOT association


gp120

CD4



YES

YES

NO


IV.1b protein interactionsPPI - homology-based inference

9


Sarah Gilman

10


Physical interaction NOT association


gp120

CD4



YES

YES

NO


A B

A’ ? B’

similarity > X similarity > X

Can we transfer binding through homology?

12

Obviously, otherwise no value in model organisms ...

©Sven Mika & Burkhard Rost (Columbia New York)S Mika & B Rost 2006 PLoS Genetics, Vol 2, e29


Inter and Intra-species the same?

13

similarity > X Worm

A B

A’ B’?

A’’

similarity > X

B’’ Human?

©Sven Mika & Burkhard Rost (Columbia New York)


Much better intra-species

14S Mika & B Rost 2006 PLoS Genetics, Vol 2, e29

more similarless similar

Worm

Human

A B

?A’’ B’’

?A’ B’





Worm

Human

A B

?A’’ B’’

?A’ B’





Fruitfly (Drosophila Melanogaster)


Why?

17


“Paralogs” conserve interactions

“orthologs” don’t?

18


Model organisms pose problems for protein-protein interactions



IV.2 protein interactionsPPI de novo?

20


1999: Want to predict protein-protein partners

Simple method for prediction:


Road to predicting protein-protein partners

Implement simple method to do this failed entirely: too many false positives

Reduce false positives:

predict surface residues (PROFacc, 1999) note: 1/2 of residues -> 1/4 of false positives!


Prediction of solvent accessibility

Inside

solvent

coreprotein

• 50% of residues somehow accessible to solvent

• 10% not at all






predict residues in external interfaces (ISIS, 2004)


Predict protein-protein binding partners

Reducing false positives:predict surface residues (PROFacc, 1999)


predict residues saturated internally (PROFcon, 2004)

localization (e.g. only all nuclear, LOCtree, 2004)


Reduction by localization

27

Extra-cellular

Cytoplasm

Organelles

Mitochondria Nuclear

TMtransmembrane

Extra-cellular

Cytoplasm

Organelles

Mitochondria

Nuclear

TMtransmembrane

Y Ofran et al. & Rost 2006 Bioinformatics 22:e402-7

e.g. nuclear:=15% nuclear + 15% cytoplasm=> • come in with nuclear protein: maximally 30% of proteins to test


Predict subcellular localization: LOCtree 2: 18 classes!

28

TatyanaGoldberg

T Goldberg, T Hamp & B Rost (2012) submitted

k-mer profile kernel SVM

SVM$

SVM$SVM$SVM$

SVM$SVM$SVM$SVM$ SVM$

SVM$

SVM$SVM$

SVM$

SVM$

SVM

SVM$

SVM$

SVM$ SVM$

SVM$

SVM

SVM$

SVM$

SVM$

SVM$

SVM$SVM$SVM

SVM$

SVM$

SVM$

SVM$SVM$SVM$

SVM$SVM$SVM$SVM$

SVM$

SVM$ER$

EXT$ GOLGI$ NUC$CYT$VAC$

PERO$

MITO$

CHLO$PLAST$

PLAS$ GOLGI$ VAC$

ER$ NUC$

PERO$

MITO$CHLO$

SVM$SVM$

Non$membrane$

Eukaryo1c$Protein$Sequence$

Transmembrane$

Intra:cellular$ Secretory$pathway$ Intra:cellular$

TobiasHamp

B Alberts et al. 1994 The Cell Garland







predict residues in protein-substrate interfaces (active)








predict protein domains/improve alignments








predict protein domains/improve alignments

Put it all together and predict binding partners

TobiasHamp


IV.3 protein interactions sites

PPI - data collection

32


Different interfaces, different physics!


gp120

CD4



33


Different interfaces, different physics!


gp120

CD4



At least 6 types of interfaces differ in sequence!

Internal (inter-domain and intra-domain)External homomers (permanent/transient)External heteromers (permanent/transient)

Y Ofran & B Rost (2003) J Mol Biol 325, 377-87

33


How to answer the question?

34

??

???


PDB: numbers

Molecules of experimentally determined structure (3D co-ordinates)www.pdb.org• check out: Molecule of the MonthStat 2010/04:• ~65,000 structures• 60K proteins• 2K DNA/RNA• 3K complexes• 56K X-ray• 8K NMR• 0.3K Electron microscopy

35

http://www.pdb.org

http://www.pdb.org


extract interactionshow?

36


How 1: take all or subset?

37


Remove redundancy

38

PDB

CD-Hit / UniqueProt, e.g. 70% PIDE


Histogram of protein family sizes

39

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34

#



39

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34

#

?



40

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34

#


Zipf law / Power law

41

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34

Wikipedia:George Kingsley Zipf (Jan 2, 1902-Sep 25, 1950; Harvard)

#


Size of protein families (assumption)

42

Wikipedia:George Kingsley Zipf (Jan 2, 1902-Sep 25, 1950; Harvard)

PDB “nature”


Size of protein families: assumed relation

43

PDB1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35

“nature”


Size of protein families: assumed relation

43

PDB1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35

“nature”

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35


Problems with redundancy reduction

What approximates “nature”?

How far to move “is” toward “should be”?

44


Develop method

1. PDB->Unique

45


How 2:distance

46


Different interfaces = different physics


gp120

CD4



47




gp120

CD4



48

gp120

CD4


<6.5 Å?


Develop method

1. PDB->Unique2. parse heavy atoms <6.5 Ångstrøm (0.65 nm)

49

gp120

CD4





gp120

CD4



50


PDB chains

Each atom in the PDB belongs to a residueEach residue belongs to a chainChains may have “breaks”

51




gp120

CD4



52


Map PDB to annotations

53

PDB SWISS-PROT

BLAST 10-10, >90% PIDE over >90% of length

chain A maps to SPa, chain B maps to SPbif (SPa=SPb) assume A and B from same proteinelse A and B from two different proteins


Develop method

1. PDB->Unique2. parse heavy atoms <6.5 Ångstrøm (0.65 nm)3. map chains to SWISS-PROT, distinguish transient protein-protein interactions from others

54


Remove redundancy for transient PP

55

PDB-PP



Develop method

1. parse heavy atoms <6.5 Ångstrøm (0.65 nm)2. map chains to SWISS-PROT, distinguish transient protein-protein interactions from others3. PDB sub(PP)->Unique

56


Develop method

1. parse heavy atoms <6.5 Ångstrøm (0.65 nm)2. map chains to SWISS-PROT, distinguish transient protein-protein interactions from others3. PDB sub(PP)->Unique

NOW we have a data set and can apply machine learning

57


PPI interfaces use local segments

58Y Ofran & B Rost (2003) FEBS Lett 544, 236-9


1st question:external and

internal interactions the same biophysics?

59



60



gp120

CD4


Stru

ctur

e: H

endr

icks

on la

b



60



gp120

CD4


Stru

ctur

e: H

endr

icks

on la

b

At least 6 types of interfaces differ in sequence!

Internal (inter-domain and intra-domain)External homomers (permanent/transient)External heteromers (permanent/transient)

Y Ofran & B Rost (2003) J Mol Biol 325, 377-87


Interface types: composition

61Y Ofran & B Rost (2003) J Mol Biol 325, 377-87

gp120

CD4



Interface types differ in composition


gp120

CD4



Interface types differ in composition


gp120

CD4


They obviously differ!But, are these differences meaningful?


How to answer the meaningful question?

64

??

???


Are these differences statistically significant?

Y Ofran & B Rost 2005 submitted


Are these differences statistically significant?Chi-square test:

known problem: small data setshere millions of points





all differences < 10-300-> SIGNIFICANT





all differences < 10-300-> SIGNIFICANT

… unfortunately also:proteins [a-b] vs [c-d]1 vs 2 authors random subsets ...



Find-self test (statistical significance)



Find-self test on six types of interfaces


gp120

CD4



IV.4 protein interactionsPPI predict binding

sites

68


Machine learning

69


how to choose the input features?

70


ask your friend(ideally in the group)

71


Strength of prediction reflects reliability?

72



0.6

0.4

0.9

0.1

weakstrong

72


More complex system to predict structure

Sequence PSI-BLAST FilterPROFsecPROFacc

1999

73


What it takes to realize this as a server

2004

PDB SWISS-PROT TrEMBL HSSP


Much more complex system for function

75


Much more complex system for function

75

2004


Few features

76

Profilepredicted 1D structure• secondary structure• solvent accessibility• membrane regions• disorderpredicted aspects of function


Publish not to perish -

tie it up

77


Are we there yet?

78


Let neural networks figure it out ...

79


Let neural networks figure it out ...

79

Train

Test


Cross-validation: how?

80

PDB-PP



Random split not enough

81


avoid overlap training/cross-

training vs. testing

82


Now, are we there yet?

83


PP interfaces predicted from sequence

Y Ofran & B Rost 2003 FEBS Lett 544:236-9Y Ofran & B Rost 2007 Bioinformatics e13-16

84


0.6

0.4

0.9

0.1

weakstrong

85





86



87

Accuracy:>94% for 1 in 10>70% for 2 in 10

Y Ofran & B Rost (2003) FEBS Lett 544, 236-9


Successful prediction: skp1-skp2

88

Uniquitin ligase skp1-skp2 complexSchulman, B.A. et al. (2000) Nature 408, 381-6

Green: 2 correctly predicted residues

(pocket binding TRP109 of SKP-2 F-box protein)

Y Ofran & B Rost (2003) FEBS Lett 544, 236-9

Accuracy:>94% for 1 in 10>70% for 2 in 10


Prediction system

Level 1:Neural networksinput: alignment profile/predicted secondary structure + accessibility (PROF)/predicted sequence complexity/overall features (protein length, amino acid composition, asf.)output: 2 units: is or is not P=P

Level 2:Neural networks using input from previous levelLevel 3:simple clustering

89


PPI interfaces use local segments

90Y Ofran & B Rost (2003) FEBS Lett 544, 236-9




91




92


PPI hot spots?

93


Interaction HOT SPOTS

94

residues that are essential for protein-protein interactionsoperational: • 1. residue in the interface• 2. mutation of the residue knocks out interaction



Very strong =

hot spots?


95


Prediction of hot spots for CD4

• alanine scan for V1 domain of CD4 (bound to gp120)(A Ashkenazi et al. & DJ Capon (1990) PNAS 87, 7150)red: observed

Y Ofran & B Rost 2007 PLoS CB 3:e11996


Prediction of hot spots for CD4

• alanine scan for V1 domain of CD4 (bound to gp120)(A Ashkenazi et al. & DJ Capon (1990) PNAS 87, 7150)red: observed

• structure:PD Kwong et al. & WA Hendrickson (2000) Structure 8, 1329-1339.

purple: predicted(Y Ofran & B Rost (2006) ISIS submitted)



enough to publish?

97


Hot spots reliably predicted from sequence!

worst:~60%right

hottest of hot = no error!



What makes it work?

Evolutionary information:• Optimally choosing profile

• Explicitly using conserved residues

(Predicted) 1D Structure important: good prediction + used correctly• Surface residues

• Secondary structure

Mark low-complexity and sticky

Filtering “isolated predictions”


Hot spots prediction requires full information

100

0 10 20 30 40 50 60 70 80 90

89

85

36

35

12

Sequence+Evolution+Exp. structure

Sequence+Evolution+Pred. structure

Evolution only

Sequence only

Hydrophobic Moment



Functionally important residues - interactions sites

101

© Marco Punta & Yanay Ofran & Burkhard Rost (Columbia New York)

..LNDRA. ..LNDRA...---P-.



Find non-homologous competitive binder

102


IV.5 protein interactionsPPI - hubs

103


Ta-Tsen Soong

Eyal MozesSara Gilman


Connect micro- and macro-level

105

micro level:residuesRIGHT: more hotspots

macro level:networks

UP: more partners


Date- and Party-hubs

106

Hubs: promiscuous proteins

Date/Party hubsNotation introduced by Marc VidalJD Han et al. & M Vidal 2004 Nature 430:88-93

• Date hubs interactions at different times/same location?

• Party hubs interactions at same time/different location


More hotspots -> more party-hub like!

107

Non-hubsParty hubsDate hubs

micro: more hotspots

macro: more partners

Y Ofran, A Schlessinger & B Rost submitted



107

0

0.125

0.250

0.375

0.500

926

42







107

0

0.125

0.250

0.375

0.500

926

42


00.10.20.30.40.5

9 26 42





More unstructured -> more date-hub like!

108






More unstructured -> more date-hub like!

108

0

0.14

0.28

0.42

0.56

0.70

NORSnet






Examples for Date & Party hubs

109Y Ofran, A Schlessinger & B Rost submitted

FUS3 MAP kinase - date hub (PDB 2b9f) right complex with MSG5binding motif (light blue) ISIS unstructured

ABC10-beta subunit of RNA polymerase - party hub (PDB 1r9sJ

right: RNA Polymerase II elongation complex (ABC10-

beta in red)


Lecture plan (PP2 function)01: 2013/10/15: no lecture02: 2013/10/17: welcome: who we are03: 2013/10/22: Intro - function 1: concepts04: 2013/10/24: Intro - function 2: homology05: 2013/10/29: Tobias Hamp: Homology-based prediction of function06: 2013/10/31: no lecture Reformation07: 2013/11/05: Tobias Hamp: Homology-based prediction of function 208: 2013/11/07: Intro - function 3: motifs09: 2013/11/12: no lecture: SVV (student reps)10: 2013/11/14: Localization 111: 2013/11/19: Localization 212: 2013/11/21: Localization 3 - Tatyana Goldberg13: 2013/11/26: Arthur Dong: Protein-protein interaction networks14: 2013/11/28: Localization 4 - Tatyana Goldberg15: 2013/12/03: Marco De Vivo (ISS Genoa) - Drug Design16: 2013/12/05: Marco Punta (Pfam, EBI) - Pfam17: 2013/12/10: Exercise/Presentations18: 2013/12/12: no lecture: Dies Academicus / Protein-protein interaction 119: 2013/12/17: Andrea Schafferhans: 3D function prediction20: 2013/12/19: Andrea Schafferhans: Docking21-24: no lectures - winter break (2013/12/23 - 2014/01/06)25: 2014/01/07: Protein-protein interaction 126: 2014/01/09: Protein-protein interaction 227: 2014/01/14: Protein-protein interaction 3 / Protein-DNA/RNA interaction28: 2014/01/16: Protein-DNA/RNA interaction 229: 2014/01/21: SNP effect 1 30: 2014/01/23: SNP effect 231: 2014/01/28: SNP effect 332: 2014/01/30: wrap-up33: 2014/02/04: examen34: 2014/02/06: no lecture

110

predict protein-protein interaction sites - rostlab · © burkhard rost (tum munich) /110...

Documents