A GraphTheoretic Method for Mining Functional Modules in Large Sparse Protein Interaction Network

1
A Graph-Theoretic Method for Mining Functional
Modules in Large Sparse Protein Interaction
Network

• Zhang et al.
• Presented by Youngik Yang

2
Motivation

• Network modules do not occur by chance,
  identification of modules is likely to capture
  the biologically meaningful interaction

3
Existing Methods

• Hierarchical clustering methods
• Prediction Methods
• Combinde methods (enumeration of complete
  sub-graphs, superparamagnetic clustering and
  Monte Carlo simulation)

4
Problem on existing methods

• Partition algorithms each protein belongs to
  only one specifiic module not suitable for
  finding overlapping modules.
• PPI networks are very sparse, while most methods
  only identify strongly connected subgraphs as
  modules, so only a few modules were detected

5
Real Networks Figures from Palla et al, Nature,
2005
6
(No Transcript)
7
Approach CPM LGT

• Clique Percolation Method CPM based on clique
  can reveal overlapping module structure of
  complex networks.
• Shortcoming a 3-clique structure.
• the spoke-like module can not be detected
• only a few modules can be detected in large
  sparse PPI network (fly, yeast, worm, etc).
• Line Graph Transformation (LGT) is introduced to
  overcome the shortcoming

8
Data

• Data collected from various sources such as
  MIPS, PreBIND, BIND, GRID, and released papers
  from Nucleic Acids Research and Science
• Preprocessing - remove self-interactions and
  duplicated interactions

9
Procedure

• Step1. Compute line graph L(G)
• Step 2. Apply CPM on the L(G)
• Step 3. Resulting modules in L(G) are transformed
  back to modules in G
• Step 4. Merge two heavily overlapped modules

10
Method Illustration
11
Clique Percolation Method (CPM)

• k-clique community
• a union of all k-cliques
• a series of adjacent k-cliques (where adjacency
  means sharing k-1 nodes)
• k-clique community can be considered as a usual
  module because of its dense internal linkage and
  sparse external linkage with other part of the
  whole network.

12
(No Transcript)
13
Line Graph Transformation (LGT)

• Problem on CPM
• Too restrictive to detect proper modules in
  sparse networks. E.g. spoke-like modules can not
  be detected.
• Transform nodes into edges.
• Retain information of the original network
• more highly structured than the original network.
  So it is much more convenient than directly using
  clique percolation clustering.

14
Reverse Transformation (RLG)

• Edges in G which correspond to the nodes of a
  module in L(G) will form a subnet of the original
  network G
• Add the lost edges within the nodes of the subnet
  to form modules in the original PPI network.

15
Post-processing

• Merging is executed for two modules which have a
  large overlap

16
(No Transcript)
17
Validation of protein complexes

• P-value the probability that a given set of
  proteins is enriched by a given functional group
  merely by chance
• , where module M contains k proteins in a
  function category F, and the PPI network contains
  N proteins

18
Cont.

• By minimizing the probability Pol of a random
  overlap between a computational group and an
  experimental group, we can determine the
  best-matching experimental complex for a module.
• , where C, M are the sizes of an experimental
  complex and a computed module respectively

19
Results
20
Proteins in the same module have the same
localization
21
Functional annotation of network modules
22
Matching with experimentally determined complexes
Cellular complexes (550.1.136)
Coat complexes II (260.30.20)
Membrane complex (290.10)
23
(No Transcript)
24
Statistical properties of overlapping modules