STRINGCross-species integration of known and

predicted protein-protein interactions

Lars Juhl JensenEMBL Heidelberg

STRING provides a protein network based on integration of diverse types of evidence

Genomic neighborhood

Species co-occurrence

Gene fusions

Database imports

Exp. interaction data

Microarray expression data

Literature co-mentioning

Inferring functional modules fromgene presence/absence patterns

Restingprotuberances

Protractedprotuberance

Cellulose

© Trends Microbiol, 1999

CellCell wall

Anchoring proteins

Cellulosomes

Cellulose

The “Cellulosome”

Genomic context methods

© Nature Biotechnology, 2004

Formalizing the phylogenetic profile method

Align all proteins against allAlign all proteins against all

Calculate best-hit profileCalculate best-hit profile

Join similar species by PCAJoin similar species by PCA

Calculate PC profile distancesCalculate PC profile distances

Calibrate against KEGG mapsCalibrate against KEGG maps

Predicting functional and physical interactions from gene fusion/fission events

Find in A genes that matcha the same gene in B

Find in A genes that matcha the same gene in B

Exclude overlappingalignments

Exclude overlappingalignments

Calibrate againstKEGG maps

Calibrate againstKEGG maps

Calculate all-against-allpairwise alignments

Calculate all-against-allpairwise alignments

Inferring functional associations from evolutionarily conserved operons

Identify runs of adjacent geneswith the same direction

Identify runs of adjacent geneswith the same direction

Score each gene pair based onintergenic distances

Score each gene pair based onintergenic distances

Calibrate against KEGG mapsCalibrate against KEGG maps

Infer associationsin other species

Infer associationsin other species

Score calibration against a common reference

• Many diverse types of evidence– The quality of each is judged by

very different raw scores

– Quality differences exist among data sets of the same type

• Solved by calibrating all scores against a common reference– Scores are directly comparable

– Probabilistic scores allow evidence to be combined

• Requirements for the reference– Must represent a compromise of

the all types of evidence

– Broad species coverage

Integrating physical interaction screens

Complexpull-down

experiments

Complexpull-down

experiments

Yeast two-hybriddata sets are

inherently binary

Yeast two-hybriddata sets are

inherently binary

Calculate scorefrom number of

(co-)occurrences

Calculate scorefrom number of

(co-)occurrences

Calculate scorefrom non-shared

partners

Calculate scorefrom non-shared

partners

Calibrate against KEGG mapsCalibrate against KEGG maps

Infer associations in other speciesInfer associations in other species

Combine evidence from experimentsCombine evidence from experiments

Mining microarray expression databases

Re-normalize arraysby modern methodto remove biases

Re-normalize arraysby modern methodto remove biases

Buildexpression

matrix

Buildexpression

matrix

Combinesimilar arrays

by PCA

Combinesimilar arrays

by PCA

Calculate pairwiselinear correlation

coefficients

Calculate pairwiselinear correlation

coefficients

Calibrateagainst

KEGG maps

Calibrateagainst

KEGG maps

Inferassociations inother species

Inferassociations inother species

?

Source species

Target species

Evidence transfer based on “fuzzy orthology”

• Orthology transfer is tricky– Correct assignment of orthology

is difficult for distant species

– Functional equivalence cannot be guaranteed for in-paralogs

• These problems are addressed by our “fuzzy orthology” scheme– Confidence scores for functional

equivalence are calculated from all-against-all alignment

– Evidence is distributed across possible pairs according to confidence scores in the case of many-to-many relationships

Multiple evidence types from several species

Getting more specific – generally speaking

• Benchmarking against one common reference allows integration of heterogeneous data

• The different types of data do not all tell us about the same kind of functional associations

• It should be possible to assign likely interaction types from supporting evidence types

• The aim: to construct an accurate, qualitative models of biological systems or processes

• The models should be accurate even at the level of individual interactions

• This allows specific, testable hypotheses to be made based on high-throughput experimental data

Yeast culture Microarrays Gene expression Expression profile

600 periodically expressed genes (with associated peak times) that encode “dynamic

proteins”

The parts listNew analysis

Getting the parts list

Cho & Spellman et al.

Constructing a reliable protein network

• The stickiness of an interaction was scored based on its local network topology

• We benchmarked these scores for each individual data set against a common reference

• Impossible interactions were eliminated based on subcellular localization data

• By restricting the network to a particular system the error rate is further reduced

Cell cycle microarray

data

Physical PPI interactions with

confidence scores

Expand the set of proteins to include non-periodic proteins that are strongly connected to

periodic proteins

Raw DataNode selection

List of periodically expressed proteins

with peak time

Interactions

Require compatible compartments and high confidence

Extract cell cycle network

Extracting a cell cycle interaction network

The temporal interaction network

Interacting proteins are expressed close in time

Two thirds of the dynamic proteins lack interactions but likely participate in transient interactions

Static proteins comprise a third of the interactions at all times of the cell cycle

Their time of action can be predicted from interactions with dynamic proteins

Static proteins play a major role

Cdc28p and its interaction partners

Just-in-time synthesis vs. just-in-time assembly

Most dynamic proteins are expressed just before they are needed to carry out their function

Most complexes also contain static proteins

Just-in-time assembly of complexes appear to be the general principle

The time of assembly is controlled synthesizing the last subunits just-in-time

Assembly of the pre-replication complex

Network as a discovery tools

The network enables us to place 30+ uncharacterized proteins in a temporal interaction context

Quite detailed hypotheses can be made concerning the their function

The network also contains entire novel modules and complexes

Transcription is linked to phosphorylation

A genome-wide screen identified 332 Cdc28p targets, which include– 6% of all yeast proteins

– 8% of the static proteins

– 27% of the dynamic ones

A similar correlation was observed with predicted PEST regions

This suggests a hitherto undescribed link between transcriptional and post-translational control

Conclusions

• Genomic context methods are able to infer the function of many prokaryotic proteins from genome sequences alone

• Integration of large-scale experimental data allows similar predictions to be made for eukaryotic proteins

• Benchmarking is a prerequisite for data integration

• It is possible to construct highly reliable models through careful integration of high-throughput experimental data

• Try STRING at http://string.embl.de

Acknowledgments

• The STRING team– Christian von Mering– Berend Snel– Martijn Huynen– Daniel Jaeggi– Steffen Schmidt– Sean Hooper– Julien Lagarde– Mathilde Foglierini– Peer Bork

• New context methods– Jan Korbel– Christian von Mering– Peer Bork

• Cell cycle analysis– Ulrik de Lichtenberg– Thomas Skøt Jensen– Anders Fausbøll– Søren Brunak

Thank you!