A hub-attachment based method to detect functional modules from confidence-scored protein interactions and expression profiles Authors: Chia-Hao Chin 1,4,

A hub-attachment based method to detect functional modules from

confidence-scored protein interactions and expression profiles

Authors: Chia-Hao Chin1,4, Shu-Hwa Chen1, Chin-Wen Ho4, Ming-Tat Ko1,5, Chung-Yen Lin1,2,3,5

1. Institute of Information Science, Academia Sinica, Taiwan2. Division of Biostatistics and Bioinformatics, National Health Research Institutes, Taiwan3. Institute of Fishery Science, College of Life Science, National Taiwan University, Taiwan 4. Department of Computer Science and Information Engineering, National Central University, Taiwan5. Research Center of Information Technology Innovation, Academia Sinica, Taiwan

Outline

• Goal

• Method

• Experiment results

Detecting functional modules

Identify functional modules by parsing Protein-Protein Interaction (PPI) networks into densely connected regions

+

A more reliable PPI

C1 C2 C3 C4

V1 0.2 0.4 0.3 0.6

V2 0.4 0.4 0.8 0.8

V3 0.3 0.4 0.7 0.9

V1

V2

V3

V1 V2 V3

V1 - 0.5 0.7

V2 0.5 - 0.9

V3 0.7 0.9 -

Pearson correlation threshold = 0.6

Gene expression data

A PPI network

The overview of HUNTER

An ExampleModule seeds generation

Modules amalgamation

Module seed growth

module seeds

grown modules

final modules

Module seed generation

• Four cases for this stage

input graph

contain expression data

Unweighted Weighted

No Case 1 Case 2

Yes Case 3 Case 4

Module seed generation(1/4)

• Case 1 : – Input data is an unweighted graph.

• Find a maximum connected component of the subgraph induced by v's neighbors.

v

The Union of the vertex set of a maximum connected component and vertex v is a module seed .

Union vertices of this sugraph and vertex v.

This is a maximum connected component of the subgraph induced by v's neighbors.

This is the subgraph induced by v's neighbors. It is composed of three connected components.

A q-connected module

• A vertex set U V is q-connected if the probability is at least q for all W U with at least one edge that connects W with U \ S. [Ulitsky et. al. 2009]

a

b

c

0.8

0.6

0.7

p( {a}, {b, c} ) = 1 - (1-0.8)*(1-0.6) = 0.92

p( {a, b}, {c} ) = 1 - (1-0.8)*(1-0.7) = 0.94

p( {a, c}, {b} ) = 1 - (1-0.6)*(1-0.7) = 0.88

If q = 0.9, then this graph is not q-connected.

Module seed generation(2/4)

• Case 2 : – Input data is a weighted graph.

• Find a maximum q-connected component of the subgraph induced by v's neighbors.

v

This subgraph is q-connected, and the vertex set of it is a module seed.0.1

0.8

0.8

0.7

0.60.7

1.0

0.6

0.80.7

0.8

0.8

If a threshold q = 0.9, then this induced subgraph is not q-connected.

If a threshold q = 0.9, then this induced subgraph is q-connected.

If a threshold q = 0.9, then this induced subgraph is not q-connected.

Is this subgraph q-connected?

Is this subgraph q-connected?

Find a maximum q-connected component of the subgraph induced by v's neighbors.

Module seed generation(3/4)

• Case 3 : – Input data is composed of an unweighted graph and gene expression dat

a.• Find a maximum connected component of the subgraph induced by

v's neighbors, where the Pearson correlation of any pair of vertices is greater than a threshold.

v

In this subgraph, the Pearson correlation of each pair of vertices is greater than a threshold, and the vertex set of it is a module seed

A blue dashed line means its Pearson correlation is less than a threshold t = 0.6

A green dashed line means itsPearson correlation is larger thana threshold t = 0.6

Check each subgraph by usinggene expression data.

Module seed generation(4/4)

• Case 4 : – Input data is composed of a weighted graph and gene expression data.

• Find a maximum connected component of the subgraph induced by v's neighbors, where the Pearson correlation of any pair of vertices is greater than a threshold.

v The vertex set of this subgraph is a module seed.

A blue dashed line means its Pearson correlation is less than a threshold t = 0.6

A green dashed line means its Pearson correlation is larger than a threshold t = 0.6

This induced subgraph is not q-connected.

0.8 0.1

0.8

0.8

0.7

0.60.7

1.0

0.6

0.80.7 0.8

We check whether this subgraph is q-connected.

We check each subgraph by using gene expression data.

This subgraph is q-connected.

Module growth

• After creating a module seed, we join the neighbors of the module seed if most of their adjacent nodes also belong to the module seed.

v

w

A module seed

v

w

A grown module

Module amalgamation

• we merge any two modules if they have too many common proteins

grown module 1 grown module 2 A final module

Functional Group Verification Using Gene Ontology

Gene Ontology• Three separate ontologies:

• Biological Process• Molecular Function• Cellular Component

• Organized as a DAG describing gene products (proteins and functional RNA)

• GO Annotation•A GO term is associated with a gene or gene product to form a GO annotation.

http://www.yeastgenome.org/help/GO.html

p-value

• Given a gene ontology and term t, the p-value is the probability of observing x or more proteins in the cluster c.– N: the number of proteins annotated to a term of the GO ontology.– M: the number of proteins annotated to the GO term t.– n : the number of proteins of the cluster c.– x : the number of proteins of the cluster c which are annotated to the

GO term t.

-value

n

i x

M N M

i n ip

N

n

NMn

x

F-measure

• For each method, we measured– Sensitivity: the fraction of annotations that are enriched in

at least one module at p-value < 10-4 [Ulitsky et.al. 2009].

– Specificity: the fraction of modules enriched with at least one annotation at p-value < 10-4 [Ulitsky et. al. 2009].

We compare our method with three newly developed methods

• CEZANNA [Ulitsky et. al. 2009]

• CMC [Liu et. al. 2009]

• Core [Leung et. al. 2009]

Check experiment results by GO

Check experiment results by golden standard databases

• p-value: Given a golden standard database and complex g, the p-value is

the probability of observing x or more proteins in the cluster c.– N: the number of proteins in a golden standard database.– M: the number of proteins in a complex g of the golden standard database.– n : the number of proteins of the cluster c.– x : the number of proteins of the cluster c which also belong to the

complex g.

-value

n

i x

M N M

i n ip

N

n

NMn

x

Check experiment results by golden standard databases

RNA Polymerase I

RNA Polymerase III

RNA Polymerase II

Common module for RNA polymerase I, II, III

Common module for RNA polymerase I, III

Common regulatory unit for RNA polymerase I, II

TFIIF for RNA polymerase II

A cluster of our prediction on yeast PPI

Threshold

• q-connected

– We set q as 0.95 corresponds to an "error probability" of 0.05.

• correlation threshold t – Initiation

• A complete graph

• given a cutoff threshold

– Remove those edges whose Pearson correlation are less or equal than the threshold.

0.7

0.9

0.6

0.80.6

0.6

cutoff threshold = 0.6

Clustering coefficient

ki: degree of node i

Ei: edges between neighbors of node i’s

The density of the network surrounding node i, characterized as the number of triangles through i.

i

The center node has 8 (grey) neighborsThere are 4 edges between the neighbors

C = 2*4 /(8*(8-1)) = 8/56 = 1/7K is the number of nodes whose degree are larger than 1.

A threshold for Pearson correlation • The authors conjectured that the removed links are lik

ely to be noise as long as the difference between the observed clustering coefficient and its randomized counterpart increases monotonically [Elo et. al. 2007].

A threshold r0 = 0 r1 = 0.01 r100 = 1

threshold

C( ri ) – C0( ri )

the first local maximumC*

References

• Elo LL, Jarvenpaa H, Oresic M, Lahesmaa R, Aittokallio T: Systematic construction of gene coexpression networks with applications to human T helper cell differentiation process. Bioinformatics 2007, 23(16):2096-2103.

• Liu G, Wong L, Chua HN: Complex discovery from weighted PPI networks. Bioinformatics 2009, 25(15):1891-1897.

• Leung HC, Xiang Q, Yiu SM, Chin FY: Predicting protein complexes from PPI data: a core-attachment approach. J Comput Biol 2009, 16(2):133-144.

• Ulitsky I, Shamir R: Identifying functional modules using expression profiles and confidence-scored protein interactions. Bioinformatics 2009, 25(9):1158-1164.

Thank you for your attention!