Detecting protein complexes in a PPI network: a gene ontology based multi-objective evolutionary...

3036 Mol. BioSyst., 2012, 8, 3036–3048 This journal is c The Royal Society of Chemistry 2012

Cite this: Mol. BioSyst., 2012, 8, 3036–3048

Detecting protein complexes in a PPI network: a gene ontology based

multi-objective evolutionary approachw

Anirban Mukhopadhyay, Sumanta Ray* and Moumita De

Received 27th March 2012, Accepted 14th August 2012

DOI: 10.1039/c2mb25302j

Protein complexes play an important role in cellular mechanism. Identification of protein

complexes in protein–protein interaction (PPI) networks is the first step in understanding the

organization and dynamics of cell function. Several high-throughput experimental techniques

produce a large amount of protein interactions, which can be used to predict protein complexes in

a PPI network. We have developed an algorithm PROCOMOSS (Protein Complex Detection

using Multi-objective Evolutionary Approach based on Semantic Similarity) for partitioning the

whole PPI network into clusters, which serve as predicted protein complexes. We consider both

graphical properties of a PPI network as well as biological properties based on GO semantic

similarity measure as objective functions. Here three different semantic similarity measures are

used for grouping functionally similar proteins in the same clusters. We have applied the

PROCOMOSS algorithm on two different datasets of Saccharomyces cerevisiae to find and

predict protein complexes. A real-life application of the PROCOMOSS is also shown here by

applying it in the human PPI network consisting of differentially expressed genes affected by

gastric cancer. Gene ontology and pathway based analyses are also performed to investigate the

biological importance of the extracted gene modules.

1 Introduction

A PPI network can be described as a complex system of proteins

linked by interactions. The simplest representation takes the form

of an undirected graph consisting of nodes and edges,1 where

proteins are represented as nodes and the interaction of two

proteins is represented as adjacent nodes connected by an edge.

The protein complexes in a PPI network are assemblages of

proteins that interact with each other at a given time and place,

forming a dense region in the PPI networks.

Several techniques based on graph clustering, finding dense

regions, or clique finding have been proposed to discover

protein complexes in PPI networks.2–5 Molecular Complex

Detection (MCODE), proposed by Bader et al.,6 detects

densely connected regions in the PPI network by giving weight

to each vertex, corresponding to its local neighborhood den-

sity. Then, starting with the top weighted vertex (seed vertex),

it includes the vertices whose weight is above a given threshold

in the cluster, recursively. The Markov Cluster algorithm

(MCL) proposed in ref. 7 converges toward a partitioning of

the graph, with a set of high-flow regions (the clusters)

separated by boundaries with no flow. In ref. 8 Restricted

Neighborhood Search Clustering (RNSC), a cost-based local

search algorithm is proposed that explores the solution space

to minimize a cost function calculated according to the

number of intra-cluster and inter-cluster edges. Starting from

an initial random solution, RNSC iteratively moves a vertex

from one cluster to another if this move reduces the general cost.

Recently in ref. 9 a clustering with overlapping neighborhood

expansion (ClusterONE) has been introduced for detecting

potentially overlapping protein complexes from protein–protein

interaction data. This algorithm consists of three major steps:

first, starting from a single seed vertex, a greedy procedure adds

or removes vertices to find groups with high cohesiveness. In the

second step, merging between each pair of groups is done based

on the extent of overlap between each pair of groups for which

the overlap score is above a specified threshold. In the third step,

a postprocessing is done by discarding complex candidates that

contain less than three proteins or whose density is below a given

threshold.

In general it has been observed that the proteins constituting

a complex are functionally similar and they carry out some

common biological activity. Motivated by this observation, in

this article a multi-objective algorithm PROCOMOSS (Protein

Complex Detection using Multi-objective Evolutionary Approach

based on Semantic Similarity) is developed. PROCOMOSS

optimizes both graph based density metric and GO-semantic

similarity based metric simultaneously to find dense protein

Department of Computer Science and Engineering, University ofKalyani, Kalyani, India. E-mail: [email protected],[email protected], [email protected] Electronic supplementary information (ESI) available: The code andother related materials are available at http://kucse.in/procomoss/. SeeDOI: 10.1039/c2mb25302j

MolecularBioSystems

Dynamic Article Links

www.rsc.org/molecularbiosystems PAPER

http://dx.doi.org/10.1039/c2mb25302j

http://dx.doi.org/10.1039/c2mb25302j

This journal is c The Royal Society of Chemistry 2012 Mol. BioSyst., 2012, 8, 3036–3048 3037

complexes containing functionally similar proteins. NSGA-II,10,11

a popular multi-objective GA,12 has been utilized as an under-

lying optimization tool.

The results are collected by applying PROCOMOSS in

the protein–protein interaction (PPI) data downloaded from

two different high throughput datasets DIP and MIPS. The

performance of PROCOMOSS is compared with that of other

methods such as MCODE,6 RNSC,8 MCL,7 and clusterONE9

and Affinity propagation.13 The different semantic similarity

measures we have used in our PROCOMOSS algorithm are

also have been utilized in the Affinity propagation for grouping

proteins into modules. We also perform a Gene Ontology and

pathway based analysis of the predicted modules identified by

PROCOMOSS in the differentially expressed genes extracted

from the gastric cancer dataset downloaded from www.biolab.

si/supp/bi-cancer/projections/info/GSE2685.htm.

2 Multi-objective optimization using GA

The multi-objective optimization problem can be stated as

follows:12,14–16 find the vector �x� ¼ ½x�1; x�2; . . . ; x�n�T of the

decision variables satisfying the m inequality constraints:

gi( %x) Z 0, i = 1,2,. . .,m, p equality constraints hi( %x) = 0,

i = 1,2,. . .,p that optimizes the vector function %f( %x) = [f1( %x),

f2( %x),. . ., fk( %x)]T. The constraints define the feasible region F

containing all the admissible solutions. The vector %x* denotes

an optimal solution in F.

The concept of Pareto optimality12,16 is useful in the domain

of multi-objective optimization. A formal definition for Pareto

optimality from the viewpoint of minimization problem may

be given as follows: a decision vector %x* is called Pareto

optimal if and only if there is no %x that dominates %x*, i.e.,

there is no %x such that 8i A {1,2,. . .,k}, fi( %x) r fi( %x) and (i A{1,2,. . .,k}, fi( %x) o fi( %x*). In words, %x* is Pareto optimal if

there exists no feasible vector %x which causes a reduction in

some criterion without a simultaneous increase in at least

another. The Pareto optimum usually admits a set of solutions

called non-dominated solutions. Here we use NSGA-II10,11 as

the underlying multi-objective algorithm.

3 Gene ontology based semantic similarity

The Gene Ontology (GO) project17 is a collaborative effort

to provide consistent description of genes and gene products.

GO provides a collection of well-defined and well-structured

biological terms, called GO terms that are shared across

different organisms. They comprise three categories as the most

general concepts: biological processes, molecular functions and

cellular components. The measurement of semantic similarity

between two concepts can be easily extended to measure the

degree of similarity between terms in the GO structures.18 The

GO terms are structured by the relationships to each other, such

as is-a that represents a specific-to-general relationship between

terms, and part-of that represents a part-to-whole relationship.

Two approaches to gene similarity computation are graph

structure-based (GS), which use the hierarchical structure of GO

in computing gene similarity and information content-based (IC)

measures which additionally consider the a priori probabilities, or

information contents, of GO terms in a reference gene set.19,20

For a detailed description see the ESIw website. Among the

various approaches for computing semantic similarity measure

we use three of them here: Lin,21 Jiang and Conrath22 and

Kappa’s measure23 (equation and a brief description can be

found in the ESIw website) to form the objective functions for

computing our clustering results.

Besides the semantic similarity measure between GO terms

annotating a protein pair, we also find some graphical properties

of a protein interaction network to be used as objectives.

4 Proposed method

Here we describe the PROCOMOSS algorithm for clustering

in a PPI network in detail.

4.1 Chromosome representation

A protein complex is a subgraph of the whole PPI graph. Here

a protein complex is encoded as a chromosome. So in the

resulting population a chromosome of the type: ni, i = 1,. . .,p

is an integer denoting the index of a protein in the unique

protein set, represents a protein complex consisting of p

number of nodes or proteins. All nodes in the chromosome

are not necessarily connected.

4.2 Population initialization

Initially the whole network is broken into several biclusters.24–26

Biclustering is done by applying K-means clustering from both

the dimensions of a PPI matrix and taking intersections of the

clusters formed in these two dimensions. Each bicluster represents

a densely connected region in the network. We sort these

biclusters on the basis of density and pick up first 50 biclusters

and encode these in the initial population. The subsequent

populations are created using the genetic operators of NSGA-II.

4.3 Representation of objective functions

Here we use two types of objective functions: one is totally

dependent on the graphical properties of the protein inter-

action network and another is based on Gene Ontological

annotations of proteins.

4.3.1 Graph based objective. All graph theoretic approaches

for finding protein complexes seek to identify dense subgraphs

by maximizing the density of each subgraph on the basis of

local network topology. The density of a graph is defined as a

ratio of the number of edges present in a graph to the possible

number of edges in a complete graph of the same size. As there

are a large number of interactions (or edges) between proteins

(or nodes) in a protein complex (or subgraph), the density of

each complex is generally very high. So using density as an

objective function and maximizing it for individual subgraphs

will yield much denser complexes.

For choosing the next objective we count the number of

interconnecting nodes for a chromosome that are not present in

the current chromosome/cluster. For example in Fig. 1, the

chromosome is represented as black nodes and the interconnecting

nodes of this chromosome (which are not present in the current

chromosome) are shown in yellow colored nodes. This may be


written as:

NðCÞ ¼ j [i2C

nij; ð1Þ

where C is any cluster in G and ni is the set of nodes which are

connected with node i in C, and are not present in C.

Minimizing this will result in clusters, which have a lesser

number of outward interaction partners and we get compact

clusters.

4.3.2 Semantic similarity based objective. The semantic

similarity measure between two GO terms can be directly

converted to a measurement of the similarity between two

proteins. Since a protein is annotated to multiple GO terms,

the similarity between two proteins can be represented as the

average similarity of the GO term cross pairs, which are

associated with both interacting proteins.18 The package

csbl.go (http://csbi.ltdk.helsinki.fi/csbl.go/) is used for calcu-

lating the similarity between protein pairs. We calculate the

similarity between all pairs of proteins and tabulate this as a

matrix form. For calculating the fitness of a chromosome, the

average similarity of each pair of proteins comprising the

chromosome is computed. For example to compute the fitness

of the chromosome: {n1 n2 . . . np} we compute a submatrix s

with rows and columns comprising these nodes from the

similarity matrix S. The average value of the matrix s serves

as the fitness of this chromosome. This may be written as:

simðsÞ ¼P

i2pP

j2p sði; jÞp

: ð2Þ

Fig. 1 Example of outward interconnecting nodes of a chromosome:

black nodes represent chromosome and yellow nodes are outward

interconnecting nodes to this chromosome.

Fig. 2 Illustration of the mutation process: black nodes represent

chromosome. In each iteration randomly select one node in the

chromosome, delete that node or add those which are direct neighbors

of that node with equal probability. (a) Represents the parent chro-

mosome. After one iteration a child chromosome (shown in (b) or (c))

is produced.

Table 1 Summary of the PPI network data sets used here

Data set # Proteins # Interactions Avg. degree Max. degree Density

DIP 4669 21 621 9.2305 241 0.0020MIPS 3950 11 119 5.5792 233 0.0014

Table 2 Comparisons of results with respect to sensitivity, Positive Predictive Value (PPV) and accuracy

Method

General sensitivity of clustering result General PPV of clustering result Accuracy

DIP MIPS DIP MIPS DIP MIPS

MCODE 0.1168 0.0742 0.4922 0.4709 0.2397 0.1869cluster ONE 0.2135 0.0999 0.4078 0.3890 0.2951 0.1971MCL 0.2605 0.1588 0.4464 0.4135 0.3486 0.2563RNSC 0.2909 0.1922 0.6608 0.6048 0.4384 0.3409PROCOMOSS_Lin_Mf 0.2490 0.1150 0.8186 0.7782 0.4515 0.2992PROCOMOSS_Lin_Bp 0.2726 0.1051 0.8141 0.9425 0.4711 0.3147PROCOMOSS_Lin_Cc 0.2321 0.1187 0.6891 0.7144 0.4000 0.2912Affinity_Lin_Mf 0.1624 0.1391 0.3473 0.3513 0.2375 0.2211Affinity_Lin_Bp 0.1443 0.1311 0.3524 0.3492 0.2255 0.2140Affinity_Lin_Cc 0.1565 0.1331 0.3761 0.3523 0.2426 0.2166PROCOMOSS_Jiang_Mf 0.2215 0.1185 0.7379 0.7643 0.4043 0.3010PROCOMOSS_Jiang_Bp 0.1972 0.1178 0.8596 0.7976 0.4117 0.3066PROCOMOSS_Jiang_Cc 0.2119 0.1095 0.8526 0.8952 0.4250 0.3131Affinity_Jiang_Mf 0.1535 0.1507 0.3668 0.3818 0.2373 0.2399Affinity_Jiang_Bp 0.1436 0.1610 0.3754 0.3861 0.2322 0.2493Affinity_Jiang_Cc 0.1512 0.1571 0.3650 0.3830 0.2349 0.2453PROCOMOSS_Kappa_Mf 0.1716 0.1242 0.6390 0.7763 0.3312 0.3105PROCOMOSS_Kappa_Bp 0.2398 0.1285 0.8261 0.7366 0.4450 0.3076PROCOMOSS_Kappa_Cc 0.1833 0.1141 0.5904 0.8113 0.3290 0.3043Affinity_Kappa_Mf 0.1390 0.1391 0.3494 0.3513 0.2204 0.2242Affinity_Kappa_Bp 0.1420 0.1611 0.3546 0.3887 0.2244 0.2503Affinity_Kappa_Cc 0.1487 0.1610 0.3552 0.3868 0.2298 0.2495


Here we use the semantic similarity measures proposed by Lin,

Jiang and Conrath, and Kappa to compute the similarity

matrices. By maximizing it we can group the functionally

similar proteins.

4.4 Selection and mutation

The popularly used genetic operations are selection, crossover,

and mutation. General crossover operation between two

chromosomes results in many disconnected subgraphs and

produces a large number of isolated nodes. So crossover is

not performed here and instead mutation is performed with

high probability (mutation probability = 0.9). The selection

operation used here is the crowded binary tournament selection

used in NSGA-II. If a chromosome is selected to be mutated

then addition or deletion of nodes in the chromosome is

performed in the following way: for a chromosome a random

node ni is selected and either of the two tasks is performed with

equal probability: delete that node or add the nodes which are

direct neighbors of node ni, and are not included in the parent

chromosome. Fig. 2 illustrates the process. Either of the child

chromosomes shown in Fig. 2(b) and (c) are produced from the

parent chromosome shown in Fig. 2(a). The whole operation is

performed five times to create a new diversified chromosome

from the parent chromosome.

5 Experimental results

We ran the proposed algorithm PROCOMOSS on the PPI

network of Saccharomyces cerevisiae (yeast) dataset downloaded

from the DIP27 and MIPS.28 PROCOMOSS takes 2290.93

seconds and 17 776.90 seconds for population initialization and

takes 492.32 seconds and 463.12 seconds to run on a Core 2 duo

2.26 GHz PC having 2 GB internal memory with Windows 7

installed on it, for the DIP and the MIPS dataset respectively.

For DIP out of 5000 S. cerevisiae proteins we used 4669 proteins

and for MIPS we used 3990 proteins out of 6190 due to the

availability of their annotation data. Subsequently we reduced

the interaction dataset which contains the annotated proteins

only. Our used dataset contains 11119 interactions forMIPS and

21621 interactions for DIP. Table 1 summarizes the PPI network

for DIP and MIPS. We match our clustering result with the

known protein complexes consisting of 491 complexes, down-

loaded from the site http://yeast-complexes.russelllab.org/.

The interaction datasets and the benchmark complexes can

be found in the ESIw website.

5.1 Performance comparisons with the existing method

For comparisons of PROCOMOSS clustering results with that

of some other existing algorithms we have employed some

Fig. 3 Proportion of clusters attaining a specified p-value in DIP dataset: (a) MCODE, (b) MCL, (c) cluster ONE, (d) RNSC,

(e) Affinity_Lin_mf, (f) Affinity_Lin_bp, (g) Affinity_Lin_cc, (h) PROCOMOSS_Lin_mf, (i) PROCOMOSS_Lin_bp, (j) PROCOMOSS_Lin_cc.

Here �log of p-value is given in x-axes and proportion of clusters is represented as y-axes.


matching statistics including sensitivity, positive predictive

value (PPV) and accuracy.

We built a contingency table with rows as protein complexes

and columns as resulting clusters. So, the contingency table T

is an n � m matrix having n complexes and m resulting

clusters, where row i corresponds to the i-th annotated

complex, and column j to the j-th cluster. The value of a cell

Ti,j indicates the number of proteins found in common

between complex i and cluster j. Some proteins belong to

several complexes, and some proteins may be assigned to

multiple clusters or not assigned to any cluster.

5.1.1 Sensitivity. Sensitivity is the fraction of proteins of

complex i found in predicted cluster j: Sni;j ¼ Ti;j

Ni; where Ni is

the number of proteins belonging to complex i. A complex-

wise sensitivity Sncoi may be defined as: Sncoi ¼ maxmj¼1 Sni;j .

The general sensitivity (Sn) is the weighted average of Sncoiover all complexes and defined as:

Sn ¼Pn

i¼1 NiSncoiPni¼1 Ni

: ð3Þ

5.1.2 Positive predictive value. The positive predictive

value is the proportion of members of predicted cluster j which

belong to complex i, relative to the total number of members

of this cluster assigned to all complexes: PPVi;j ¼ Ti;jPn

i¼1 Ti;j¼

Ti;j

Tj; where Tj is the marginal sum of a column j. The cluster-

wise positive predictive value PPVcljrepresents the maximal

fraction of proteins of cluster j found in the same complex:

PPVclj ¼ maxni¼1 PPVi;j : The general PPV (PPV) of a clustering

result is the weighted average of clustering-wise-PPV (PPVclj)

over all predicted clusters:

PPV ¼Pm

j¼1 TjPPVcljPmj¼1 Tj

: ð4Þ

5.1.3 Accuracy. The geometric accuracy (Acc) represents a

tradeoff between sensitivity and the positive predictive value

and is defined as:

Acc ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiSn � PPVp

: ð5Þ

It is the geometrical mean of the Sn and the PPV. The

advantage of taking the geometric mean is that it yields a

low score when either the Sn or the PPV metric is low. High

accuracy values thus require a high performance for both

criteria.

In Table 2 we list PROCOMOSS clustering results obtained

by using Lin, Jiang and Conrath, and Kappa’s similarity

measures based on each of the three orthogonal taxonomies

Fig. 4 Proportion of clusters attaining a specified p-value in MIPS dataset: (a) MCODE, (b) MCL, (c) cluster ONE, (d) RNSC,

(e) Affinity_Kappa_mf, (f) Affinity_Kappa_bp, (g) Affinity_Kappa_cc, (h) PROCOMOSS_Kappa_mf, (i) PROCOMOSS_Kappa_bp,

(j) PROCOMOSS_Kappa_cc. Here �log of p-value is given in x-axes and proportion of clusters is represented as y-axes.


or aspects that hold terms describing the molecular function

(mf), biological process (bp) and cellular component (cc) for a

gene product. We also compute the results integrating the

same similarity measures in the Affinity propagation algorithm in

the same network on both DIP and MIPS datasets. As Affinity

propagation groups data points based on the similarity between

each pair of data points, using a semantic similarity matrix in the

Affinity propagation is likely to produce protein modules which

consist of functionally homogeneous proteins. We see in

Table 2 that PROCOMOSS shows comparatively better results

than Affinity propagation for each of the similarity measures.

PROCOMOSS also performs comparatively well than other

algorithms with respect to sensitivity, PPV, and accuracy.

We also perform aGO-based study for comparing the predicted

clusters obtained by PROCOMOSS and other algorithms. We

use org.Sc.sgd.db and GOstats packages from R for computing

the p-values of predicted clusters. In Tables 3 and 4 we list

the most significant GO-terms and corresponding GO-id and

p-values under the biological process category for our clustering

results. It appears that PROCOMOSS shows a comparatively

better result when we use Lin measure as objective function in

the DIP dataset whereas in the MIPS dataset using Kappa’s

measure produces better accuracy values. So we built six tables,

three for Lin measure in the DIP dataset and other three for

Kappa measure in the MIPS dataset describing the most

significant GO-terms, GO-id and p-values of the resulting

clusters. We have given these in the ESIw website. We plotted a

bar diagram as depicted in Fig. 3 and 4 showing the proportion

of clusters attaining a specified p-value for the DIP and theMIPS

dataset respectively. The figures show that a large proportion of

clusters produced by other algorithms have higher p-values in

comparison with PROCOMOSS in which a significantly larger

Table 3 Protein complexes predicted by PROCOMOSS in DIP dataset using Lin measure (biological process), their p-values, most significantGO-terms, and GO-id

Sl no. Real protein complex Predicted protein complex p-Value % of protein covered Most significant GO-term GO-id

1 mRNA guanylyltransferase complex

Ceg1 Cet1 Ckb2 Fun12Nip1 Spt16 Ssa3 Top2

3.46 � 10�19 61.53 Proteolysis involved incellular protein catabolic process

GO:0051603

2 19/22S regulator Ecm29 Pre1 Pre5 Rad50Rpn1 Rpn10 Rpn11 Rpn12Rpn2 Rpn5 Rpn6 Rpn7Rpn8 Rpt1 Rpt2 Rpt3 Rpt5Rpt6 Ubp6

5.75 � 10�16 61.29 Proteolysis involved incellular protein catabolic process

GO:0051603

3 Ppz1 proteinphosphate complex

Cka2 Ckb2 Ssa3 Ykl088w 4.09 � 10�16 66.67 Primary metabolic process GO:0044238

4 Complex 87 Cdc55 Iml1 Yfr006w 1.77 � 10�22 100 Primary metabolic process GO:00442385 Act1–Sac6 complex Act1 Sac6 1.77 � 10�22 100 Primary metabolic process GO:00442386 Ku complex Yku70 Yku80 4.89 � 10�08 100 Macromolecule glycosylation GO:00434137 Complex 187 Brr1 Hta1 3.87 � 10�51 100 RNA splicing, via transes-

terification reactionsGO:0000375

8 Complex 221 Imd2 Imd4 Kar2 Tdh1Tdh2

3.64 � 10�19 62.5 rRNA processing GO:0006364

9 Small subunitprocessome

Dip2 Mpp10 Nan1 Nop1Nop58 Pwp2 Rok1 Rrp9Sof1 Utp13 Utp15 Utp18Utp21 Utp22 Utp4 Utp6Utp7 Utp8 Utp9


10 Complex 250 Bms1 Ecm16 Utp7Ygr054w


11 Complex 266 Dip2 Mpp10 Nop1 Nop58Pwp2 Rok1 Utp22 Utp30Utp6 Utp7


12 Complex 270 Clu1 Ura7 4.09 � 10�16 100 Primary metabolic process GO:004423813 Complex 279 Gdb1 Nup192 3.26 � 10�14 100 rRNA processing GO:000636414 Complex 307 Nop13 Rlp7 4.09 � 10�16 100 Primary metabolic process GO:004423815 Complex 348 Iml1 Psd2 1.77 � 10�22 66.67 Primary metabolic process GO:004423816 Complex 354 Dbp8 Dip2 Enp1 Ero1

Imp3 Kre33 Krr1 Mpp10Nan1 Nop1 Pre9 Rok1Rrp9 Utp22 Utp30 Utp6Utp7 Utp8


17 Complex 384 Idh2 Lys12 Rvb2 2.71 � 10�30 75 Proteolysis involved incellular protein catabolic process

GO:0051603

18 Complex 409 Arx1 Fpr4 Mrt4 Nip7 Nog2Nsa2 Sda1 Tif6


19 Complex 435 Chd1 Ckb2 Spt16 Ykl088w 3.76 � 10�13 66.67 Proteasome regulatoryparticle assembly

GO:0070682

20 Complex 440 Brx1 Drs1 Mak21 Nip7Noc2 Noc3 Nop4 Spb4Ytm1


21 Complex 473 Ecm29 Rpn1 Rpn11 Rpn12Rpn3 Rpn6 Rpn8 Rpt2Sro7 Ura7

8.20 � 10�25 71.43 Ubiquitin-dependentprotein catabolic process

GO:0006511

22 Complex 477 Cct8 Tdh2 Tef1 Ufd4 3.26 � 10�14 66.67 rRNA processing GO:0006364


number of clusters tend to have smaller p-values (i.e.

larger �log(p)). This establishes that PROCOMOSS predicts

more functionally homogeneous complexes than the other

algorithms.

5.2 Predicted complexes

We found that some real complexes are recognized by our

PROCOMOSS algorithm. We filtered out those complexes

which have less than sixty percent common proteins with some

of our predicted clusters. Here we give two tables and others

can be found in the ESIw website. In Tables 3 and 4 we list the

protein complexes predicted by our PROCOMOSS algorithm

using Lin measure (biological process annotation) and Kappa’s

measure (molecular function annotation) as objective functions.

Table 3 is built by applying PROCOMOSS on DIP data

whereas Table 4 represents the details of complexes found in

MIPS data. The third column indicates the proteins that are

members of the real protein complex shown in the second

Table 4 Protein complexes predicted by PROCOMOSS in MIPS dataset using Kappa measure (molecular function), their p-values, mostsignificant GO-terms, and GO-id

Sl no. Real protein complexPredicted proteincomplex p-Value % of protein covered Most significant GO-term GO-id

1 mRNA decappingcomplex

Dcp1 Dcp2 Edc3 Kem1 2.17 � 10�18 66.67 mRNA metabolic process GO:0016071

2 Ctf19 protein complex Chl4 Ctf19 Ctf3 Mcm21Mcm22 Nkp1

2.17 � 10�37 66.67 Mitotic cell cycle GO:0000278

3 RNA polymerase IIgeneral transcriptionfactor TFIID

Taf1 Taf10 Taf2 Taf5Taf6 Taf8

7.69 � 10�24 75 RNA polymerase IItranscriptional preinitiationcomplex assembly

GO:0051123

4 Tid3 complex Nuf2 Spc24 Spc25 Tid3 1.14 � 10�22 100 Covalent chromatinmodification

GO:0016569

5 Complex 155 Atg11 Atg17 1.18 � 10�20 66.67 Histone acetylation GO:00165736 Complex 166 Kap104 Mtr10 Nmd5 2.55 � 10�14 60 Histone acetylation GO:00165737 TRAPPII complex Bet3 Bet5 Gsg1 Kre11

Trs120 Trs130 Trs20Trs23 Trs31 Trs33

3.58 � 10�11 83.33 Vesicle-mediated transport GO:0016192

8 SBF, SWI4–SWI6-dependent cell cycle boxbinding factor complex

Swi4 Swi6 1.14 � 10�22 66.67 Covalent chromatinmodification

GO:0016569

9 Clathrin-associatedprotein AP-3 complex

Apl5 Apl6 Apm3 2.05 � 10�16 75 Protein localization GO:0008104

10 Complex 346 Gcn5 Ngg1 Pdr1 Sgf29Spt15 Spt20 Srb2 Taf1Taf12

1.30 � 10�21 60 Histone acetylation GO:0016573

11 Retromer subcomplex Vps29 Vps35 Vps5 2.55 � 10�15 100 Protein acylation GO:004354312 Ric1–Rgp1 complex Rgp1 Ric1 1.14 � 10�22 66.67 Covalent chromatin modification GO:001656913 Gim complexes Gim3 Pac10 Yke2 3.68 � 10�19 100 RNA polymerase II

transcriptional preinitiationcomplex assembly

GO:0051123

14 Complex 479 Msn5 Num1 2.17 � 10�37 66.67 Mitotic cell cycle GO:0000278

Fig. 5 Figure of a cluster predicted by PROCOMOSS using Kappa’s semantic similarity measure with molecular function annotation. Yellow

colored nodes are components of the TFIID complex and violet colored edges signify semantic similarity between two proteins connected with that

edge, which is higher than 0.5.


column and are found in some predicted clusters. The fourth

column represents p-values of the corresponding predicted

clusters which have above 60 percent of common proteins in

real protein complexes. The p-value of a cluster is defined as

the lowest p-values of all the functional groups constituting the

cluster. In row 21 of Table 3 we see that PROCOMOSS_

Lin_bp predicts 10 proteins, of which Rpn1, Rpn11, Rpn12,

Rpn3, Rpn6, Rpn8, and Rpt2 are found in complex 473. These

proteins responsibly act as a regulatory subunit of the 26S

proteasome which is involved in the ATP-dependent degradation

of ubiquitinated proteins. The 26S proteasome is a multi-

subunit enzyme composed of a cylindrical catalytic core

(20S) and a regulatory particle (19S) that together perform the

essential degradation of cellular proteins tagged by ubiquitin.29

In row 9, 19 proteins out of 29 proteins that make the complex

small subunit processome are predicted. The small subunit (SSU)

processome is a ribosome biogenesis intermediate that assembles

from its subcomplexes onto the pre-18S rRNA with yet

unknown order and structure.30 The UtpB subcomplex of

the SSU processome consisting of Utp13, Utp15, Utp18,

Utp21, Utp22, Utp4, Utp6, Utp7, Utp8 and Utp9, which are

involved in nucleolar processing of pre-18S ribosomal RNA

and ribosome assembly, is also predicted.

In row 3 of Table 4, we see that using Kappa’s semantic

similarity measure with molecular function annotation

PROCOMOSS predicts proteins Taf1, Taf10, Taf2, Taf5, Taf6

and Taf8 which function as components of the DNA-binding

general transcription factor complex TFIID. TFIID plays a key

role in the regulation of gene expression by RNA polymerase II

through different activities. In row 4 the predicted four proteins

Nuf2, Spc24, Spc25 and Tid3 of the Tid3 complex act as

components of the essential kinetochore-associated NDC80

complex, which is involved in chromosome segregation and

spindle checkpoint activity. Out of 12 proteins in the TRAPP II

complex, 10 proteins: Bet3, Bet5, Gsg1, Kre11, Trs120, Trs130,

Trs20, Trs23, Trs31 and Trs33 are identified. TRAPP II seems to

play a role in intra-Golgi transport.

Fig. 5 shows a cluster predicted by PROCOMOSS using

Kappa’s semantic similaritymeasure withMF annotation, applying

on the MIPS dataset. Here the edges are colored corresponding

to the similarity between proteins connected by this edge.

Fig. 6 Figure of some portion of a cluster predicted by PROCO-

MOSS using Lin semantic similarity measure with molecular function

annotation. Yellow colored nodes are components of anaphase-

promoting complex/cyclosome.

Fig. 7 Venn diagrams of predicted complexes that have greater than 60 percent of common proteins in some of the resulting clusters. Complexes

are identified by PROCOMOSS using (a) Lin, (b) Jiang and Conrath, and (c) Kappa measures in DIP dataset. In (d)–(f) complexes are identified in

MIPS dataset using Lin, Jiang and Conrath and Kappa’s measure respectively.


The proteins that are connected by light blue colored edge

have similarity less than 0.5 whereas the proteins that have

similarity greater than 0.5 are connected by violet colored

edges. Yellow colored nodes represent proteins that are the

components of the TFIID complex. Some portion of a cluster

predicted by PROCOMOSS using Lin semantic similarity

measure with molecular function annotation is shown in

Fig. 6. Proteins Apc1, Apc2, Apc4, Apc5, Apc9, Cdc16,

Cdc23 and Cdc27 that are components of the anaphase

promoting complex/cyclosome (APC/C) are shown as yellow

colored nodes. This cluster also captures some portion of the

TFIID complex and the GIM protein complex consisting of

prefoldin subunit 2, 3, 4, 5 (GIM2, GIM3, GIM4 and PAC10).

The figure of the whole cluster is given in the ESIw website.

We drew Venn diagrams to show the overlaps of complexes

predicted by our algorithm PROCOMOSS using three different

semantic similarity measures. Fig. 7(a)–(c) show the overlap of

predicted complexes identified by PROCOMOSS in the DIP

dataset whereas (d)–(f) show the same in the MIPS dataset using

Lin, Jiang and Conrath and Kappa’s semantic similarity measures

respectively. We retain those complexes which have 60 percent of

common proteins with some of the resulting clusters. We see that

in the DIP dataset using Lin measure, PROCOMOSS predicts

in total 41 complexes spanned in three different taxonomies:

biological process (bp), molecular function (mf) and cellular

component (cc), each of which is used to predict 22, 18, and 17

complexes with 9 overlaps between bp and mf, 6 overlaps

between cc and mf and 3 overlaps between cc and bp. We also

drew Venn diagrams to detect overlaps between complexes

predicted by PROCOMOSS using Lin, Jiang and Conrath and

Kappa’s measure for each of the taxonomies that hold terms

describing the molecular function (mf), biological process (bp)

and cellular component (cc) for a gene product. These are

shown in Fig. 8.

6 Application to the detection of modules in a

human PPI network affected by a specific disease

Here we proposed a real-life application of PROCOMOSS by

using it for a dataset of differentially expressed genes of gastric

cancer dataset. We extract the differentially expressed genes

from this dataset and also track the interactions of these genes

which conceded interactions with other genes in the whole

human PPI network. We downloaded the database from

www.biolab.si/supp/bi-cancer/projections/info/GSE2685.htm/.

There are 8 examples (26.7%) of normal gastric tissue (Normal)

and 22 examples (73.3%) of advanced gastric cancer tissue

(Tumor). So the dataset contains in total 30 samples and 4522

genes. We performed a t-test on this dataset to extract 1076

differentially expressed genes at a significant threshold level. We

compiled a network that consists of the aforesaid genes as well

as other genes that conceded interaction with them. We only

take the first neighbor of the differentially expressed genes in

the whole interaction dataset, and thus our network is

composed of 3079 proteins and 6049 interactions.

Using PROCOMOSS we find 20 statistically and biologically

significant clusters in this network using Lin and Kappa’s

semantic similarity measure in each case. It takes 492.32 seconds

Fig. 8 Venn diagrams of predicted complexes which have greater than 60 percent of common proteins in some of the resulting clusters.

Complexes are identified by PROCOMOSS using Lin, Jiang and Conrath, and Kappa measures in DIP and MIPS datasets. (a) Represents Venn

diagram of the complexes predicted by Lin, Jiang and Conrath and Kappa measures using biological process annotation for DIP dataset. (b) and

(c) describe the same using cellular component and molecular function respectively. Similarly (d)–(f) represent the same for MIPS dataset.


to initialize the population and 49.06 seconds to run for 5

generations and 50 population size on a Core 2 duo 2.26 GHz

PC having 2 GB internal memory with Windows 7 installed on

it. Tables 5 and 6 show the predicted clusters found by using

Table 5 GO-terms and pathway predicted by PROCOMOSS in gastric cancer dataset using Kappa measure (molecular function) and theirp-values

Clusters GO-terms (BP) GO-terms (CC) GO-terms (MF) KEGG pathways

Cluster 1(63 proteins)

Regulation of transcription,DNA-dependent (5.9 � 10�35)

Nucleoplasm(5.4 � 10�18)

Steroid hormonereceptor activity(4.2 � 10�45)

Pathways in cancer (4.4 � 10�9), non-smallcell lung cancer (9.6 � 10�7), prostate cancer(1.8 � 10�5), thyroid cancer (3.1 � 10�5)


Regulation of neurologicalsystem process (4.5 � 10�6)

Cell junction(1.4 � 10�7)

Glutamate receptoractivity (2.1 � 10�6)

Long-term potentiation (1.1 � 10�6),neurotrophin signaling pathway (3.6 � 10�5),tight junction (6.0 � 10�4)


Regulation of apoptosis(2.1 � 10�13)

Intracellularorganelle lumen(2.7 � 10�7)

Unfolded proteinbinding (2.4 � 10�8)

Neurotrophin signaling pathway (2.6 � 10�6),amyotrophic lateral sclerosis (ALS) (3.9 �10�5),MAPK signaling pathway (6.4 � 10�4)


Regulation of transcription,DNA-dependent (6.5 � 10�34)

Nucleoplasm part(5.2 � 10�16)


Pathways in cancer (7.0 � 10�9), non-smallcell lung cancer (4.3 � 10�7), PPAR signalingpathway (6.8 � 10�4), small cell lung cancer(9.5 � 10�5), prostate cancer (1.3 � 10�4),thyroid cancer (5.5 � 10�4), pancreatic cancer(7.6 � 10�3)


Positive regulation oftranscription, DNA-dependent(1.2 � 10�23)

Nuclear lumen(1.7 � 10�18)

Transcription regulatoractivity (2.0 � 10�28)

Pathways in cancer (8.2 � 10�7), prostatecancer(1.3 � 10�4), thyroid cancer (5.5 � 10�4),non-small cell lung cancer (3.4 � 10�3)



Organelle envelope(9.3 � 10�15)

Protein hetero-dimerization activity(9.4 � 10�9)

Amyotrophic lateral sclerosis (ALS)(4.7 � 10�11), pathways in cancer(7.0 � 10�11), colorectal cancer (5.3 � 10�9)


Regulation of transcriptionfrom RNA polymerase IIpromoter (5.3 � 10�31)



Pathways in cancer (3.5 � 10�7), non-smallcell lung cancer (1.7 � 10�4), thyroid cancer(4.5 � 10�4), small cell lung cancer (9.1 �10�4)





Pathways in cancer (2.3 � 10�6), non-smallcell lung cancer (1.4 � 10�4), PPAR signalingpathway (2.1 � 10�5), thyroid cancer (4.0 �10�4)


Apoptosis (�10�40) Organelle envelope(1.2 � 10�15)

Protein hetero-dimerization activity(1.7 � 10�8)

Amyotrophic lateral sclerosis (ALS)(3.1 � 10�13), pathways in cancer(0.6 � 10�12), endometrial cancer(9.2 � 10�9), colorectal cancer (6.3 � 10�11)


Response to protein stimulus(6 � 10�5)

Ribonucleoproteincomplex (7.9 � 10�2)

Unfolded proteinbinding (1.8 � 10�2)

Antigen processing and presentation(3.2 � 10�2)


Positive regulation of geneexpression (7.3 � 10�36)


Transcriptionregulator activity(1.1 � 10�28)

Pathways in cancer (1.9 � 10�11), prostatecancer (3.8 � 10�10), chronic myeloid leuke-mia(1.8 � 10�9), acute myeloid leukemia (5.2 �10�9)


Protein folding (4.1 � 10�9) Cytosol (2.3� 10�10) Enzyme binding(1.8 � 10�8)

Neurotrophin signaling pathway (8.9 � 10�3),prostate cancer (2.1 � 10�2)


Regulation of phosphatemetabolic process (2.6 � 10�6)

Plasma membrane(3.3 � 10�7)

Enzyme binding(7.9 � 10�8)

Long-term potentiation (2.1 � 10�6),neurotrophin signaling pathway (5.5 � 10�6)


Glutamate signalingpathway (1.5 � 10�6)

Cytoplasmicmembrane-boundvesicle (2.3 � 10�8)

Protein kinase Cbinding (6.1 � 10�5)

Long-term potentiation (3.2 � 10�6),tight junction (1.5 � 10�4)


Protein folding (9.9 � 10�9) Cytosol (6.1 � 10�8) Enzyme binding(7.4 � 10�6)

Huntington’s disease (1.1 � 10�2),spliceosome (3.2 � 10�2)



Cell junction(1.5 � 10�8)


Long-term potentiation (3.2 � 10�6),neurotrophin signaling pathway (1.0 � 10�4),tight junction (1.5 � 10�5)



Plasma membrane(1.9 � 10�7)


Endocytosis (4.0 � 10�5), tight junction(4.0 � 10�5), chemokine signaling pathway(4.4 � 10�5)



Cytosol (2.1 � 10�5) Unfolded proteinbinding (3.0 � 10�7)

Neurotrophin signaling pathway (3.6 � 10�8),amyotrophic lateral sclerosis (ALS)(3.0 � 10�6)





Pathways in cancer (3.0 � 10�8), prostatecancer (9.4� 10�7), non-small cell lung cancer(1.8 � 10�5), thyroid cancer (7.4 � 10�4)




Neurotrophin signaling pathway (3.6 � 10�8),amyotrophic lateral sclerosis (ALS)(3.0 � 10�6)


Table 6 GO-terms and pathway predicted by PROCOMOSS in gastric cancer dataset using Lin measure (biological function) and their p-values

Cluster GO-terms (BP) GO-terms (CC) GO-terms (MF) KEGG pathway



Chromatinremodeling complex(9.8 � 10�12)

Ligand-dependentnuclear receptoractivity (1.2 � 10�37)

Pathways in cancer (5.7 � 10�9), notchsignaling pathway (5.9 � 10�5)


Cytoskeleton organization(4.2 � 10�20)

Intracellularnon-membrane-bound organelle(1.8 � 10�20)

Actin binding(3.5 � 10�29)

Fc gamma R-mediated phagocytosis(3.5 � 10�12), regulation of actin cytoskeleton(7.9 � 10�9)


Regulation of transcription,DNA-dependent(1.8 � 10�21)



Prostate cancer (8.3 � 10�6), non-small celllung cancer (0.2 � 10�4), glioma (4.0 � 10�4)


Cytoskeleton organization(1.9 � 10�17)

Intracellularnon-membrane-bound organelle(7.6 � 10�11)

Actin binding(5.1 � 10�27)

Fc gamma R-mediated phagocytosis(4.2 � 10�10), regulation of actin cytoskeleton(3.0 � 10�8)


Positive regulation oftranscription from RNApolymerase II promoter(4.1 � 10�16)


Transcription factorbinding (2.1 � 10�15)

Pathways in cancer (1.5 � 10�10), renal cellcarcinoma (1.5 � 10�9), prostate cancer(8.7 � 10�5)


Regulation of RNAmetabolic process(6.8 � 10�34)



Pathways in cancer (1.4 � 10�10), non-smallcell lung cancer (6.9 � 10�9), thyroid cancer(7.8 � 10�8), small cell lung cancer(3.7 � 10�6), prostate cancer (5.5 � 10�6)


Regulation of transcription(6.8 � 10�18)



Cell cycle (2.4 � 10�5), prostate cancer(5.9 � 10�5)





Pathways in cancer (1.1 � 10�14), prostatecancer (8.1 � 10�8), renal cell carcinoma(1.8 � 10�7), pancreatic cancer (4.4 � 10�5)


Regulation of RNAmetabolic process(1.4 � 10�41)



Pathways in cancer (2.5 � 10�11), prostatecancer (5.8 � 10�11), chronic myeloidleukemia (3.1 � 10�9)





Cell cycle (9.0 � 10�3), chronic myeloidleukemia (2.8 � 10�3)


Positive regulationof transcription from RNApolymerase II promoter(5.6 � 10�18)



Pathways in cancer (7.8 � 10�11), renal cellcarcinoma (1.0 � 10�6), prostate cancer(7.2 � 10�5)





Pathways in cancer (5.4 � 10�9),non-homologous end-joining (5.5 � 10�6),chronic myeloid leukemia (6.9 � 10�6)





Pathways in cancer (8.0 � 10�8), PPARsignaling pathway (1.4 � 10�5), notchsignaling pathway (5.9 � 10�5)





Thyroid cancer (6.1 � 10�8), pathways incancer (1.1� 10�7), non-small cell lung cancer(1.1 � 10�7)




Neurotrophin signaling pathway (3.6 � 10�6),amyotrophic lateral sclerosis (ALS)(3.0 � 10�6), MAPK signaling pathway(1.3 � 10�5)


Positive regulation of geneexpression (1.9 � 10�49)



Pathways in cancer (3.9 � 10�19), chronicmyeloid leukemia (7.8� 10�13), acute myeloidleukemia (4.8 � 10�10), prostate cancer(2.4 � 10�9)


Positive regulationof transcription, DNA-dependent (1.4 � 10�15)



Prostate cancer (1.8 � 10�6), pathways incancer (2.7 � 10�4)


Positive regulationof transcription, DNA-dependent (4.1 � 10�31)



Pathways in cancer (2.7 � 10�11), cell cycle(8.8 � 10�7), thyroid cancer (8.9 � 10�7),non-small cell lung cancer (2.1 � 10�5)


Positive regulation ofnitrogen compoundmetabolic process(4.1 � 10�35)



Pathways in cancer (1.4 � 10�10), prostatecancer (0.8 � 10�10), chronic myeloidleukemia (1.4 � 10�9), acute myeloidleukemia (4.1 � 10�9), pancreatic cancer(5.1 � 10�7)





Pathways in cancer (1.7 � 10�7), chronicmyeloid leukemia (3.4 � 10�4), cell cycle(5.0 � 10�4), small cell lung cancer(5.7 � 10�4)


Lin and Kappa’s semantic similarity measures respectively.

Here we have listed the most significant GO-terms, GO-id and

the corresponding p-value of three broadly classified GO

categories: biological process, molecular function and cellular

component. We also find significant KEGG pathway for the

human proteins participating in each cluster. In Table 5, the first

cluster consists of 63 proteins which are involved in several

cancer pathways including lung cancer, prostate cancer, and

thyroid cancer. In cluster 2 the 43 proteins belong to the long-

term potentiation (1.1 � 10�6), neurotrophin signaling pathway

(3.6 � 10�5). Hippocampal long-term potentiation (LTP), a

long-lasting increase in synaptic efficacy, is the molecular basis

for learning and memory. Neurotrophins are a family of trophic

factors involved in differentiation and survival of neural cells.

The neurotrophin family consists of nerve growth factor (NGF),

brain derived neurotrophic factor (BDNF), neurotrophin 3

(NT-3), and neurotrophin 4 (NT-4). Neurotrophin/Trk signaling

is regulated by connecting a variety of intracellular signaling

cascades, which include MAPK pathway, PI-3 kinase pathway,

and PLC pathway, transmitting positive signals like enhanced

survival and growth (http://www.genome.jp/kegg/pathway/hsa/

hsa04722.html). On the other hand, p75NTR transmits both

positive and negative signals. These signals play an important

role in neural development and additional higher-order activities

such as learning and memory.

We see that the proteins in a significant number of predicted

clusters are involved in different types of pathways in cancer

viz., small cell and non-small cell lung cancer, prostate cancer,

thyroid cancer, pancreatic cancer, colorectal cancer, endome-

trial cancer, chronic myeloid leukemia and acute myeloid

leukemia.

In Table 6 we see that 47 proteins in cluster 1 belong to

cancer pathways and the notch signaling pathway. The notch

signaling pathway is an evolutionarily conserved, intercellular

signaling mechanism essential for proper embryonic development

in all metazoan organisms in the animal kingdom. Proteins in

cluster 2 belong to the Fc gamma R-mediated phagocytosis and

regulation of the actin cytoskeleton pathway. Phagocytosis plays

an essential role in host-defense mechanisms through the uptake

and destruction of infectious pathogens. Specialized cell types

including macrophages, neutrophils, and monocytes take

part in this process in higher organisms. After opsonization

with antibodies (IgG), foreign extracellular materials are recognized

by Fc gamma receptors. Cross-linking of Fc gamma receptors

initiates a variety of signals mediated by tyrosine phosphorylation

of multiple proteins, which lead through the actin cytoskeleton

rearrangements and membrane remodeling to the formation of

phagosomes (http://www.genome.jp/kegg/pathway/hsa/hsa04666.

html). Besides the different pathways here also the proteins of our

predicted clusters are involved in a significant number of cancerous

pathways. This implies that the protein complexes identified by

PROCOMOSS are highly involved in cancer progression and thus

are possible candidates for further validation.

In Tables 5 and 6 low p-values of the GO-terms signify that

the clusters are statistically significant and the occurrence of

proteins in those clusters is not merely by chance. Hence we

conclude that PROCOMOSS provides statistically and biologically

significant clusters from a human PPI network consisting of

proteins that are affected by some specific disease.

7 Conclusion

In this article we present a Multiobjective Gene Ontology

based Genetic Algorithm for finding protein complexes in the

protein–protein interaction network. Here we were able to

group functionally similar proteins in a cluster by using

semantic similarity measure of GO terms between protein

pairs as an objective function whereas the density of the cluster

is controlled by graph based objective function. Here we use

the similarity measure proposed by Lin, Jiang and Conrath

and Kappa to compute the similarity matrix. PROCOMOSS

shows better performance in the DIP dataset when we use Lin

measure for building the similarity matrix, compared to the other

measures. But for theMIPS dataset using Kappa’s and Jiang and

Conrath measure PROCOMOSS performs comparatively well.

Irrespective of the similarity measure used, it has been found that

the PROCOMOSS provides a greater number of functionally

homogeneous clusters (lower p-values). Moreover the PPV and

accuracy values are consistently better for the clusters provided

by PROCOMOSS compared to the other algorithms.

We observed that the density of a protein interaction network

built from the MIPS dataset is much lower than that of the DIP

dataset. The number of predicted complexes in MIPS is also

much lower than that in the DIP reference dataset. All the

algorithms including PROCOMOSS have higher sensitivity and

accuracy on the DIP dataset than that on the MIPS dataset.

From this we can conclude that it is more difficult to discover

complexes in a low density network.

We have also applied PROCOMOSS in a human PPI

network consisting of differentially expressed genes in gastric cancer

and have been able to extract statistically and biologically significant

gene modules. Gene ontology based study and pathway analysis

of these modules reveal wider applicability of PROCOMOSS

algorithm.

As a future work we plan to use other semantic similarity

measures as objective functions for predicting protein complexes.

PROCOMOSS can also be applied to the protein interaction

network of other species to predict protein complexes.

References

1 A. Wagner, Proc. R. Soc. London, Ser. B, 2004, 457–466.2 L. Mirny and V. Spirin, Proc. Natl. Acad. Sci. U. S. A., 2003,100(21), 12123–12128.

3 M. Altaf-Ul-Amin, Y. Shinbo, K. Mihara, K. Kurokawa andS. Kanaya, BMC Bioinf., 2006, 7, 207.

4 S. Brohe and J. van Helden, BMC Bioinf., 2006, 7, 471–488.5 J. Pereira-Leal, A. Enright and C. Ouzounis, Proteins, 2004, 54,49–57.

6 G. Bader and C. Hogue, BMC Bioinf., 2003, 4, 1471–2105.7 S. Van Dongen, ‘A new cluster algorithm for graphs,’ Center forMathematics and Computer Science (CWI), Amsterdam technicalreport, 2000.

8 N. Przulj and D. Wigle, Bioinformatics, 2003, 20, 340–348.9 T. Nepusz, H. Yu and A. Paccanaro, Nat. Methods, 2012, 9,471–472.

10 K. Deb, A. Pratap, S. Agrawal and T. Meyarivan, IEEE Trans.Evol. Comput., 2002, 6, 182–197.

11 K. Deb, Multi-objective Optimization Using Evolutionary Algorithms,John Wiley and Sons, Ltd., England, 2001.

12 S. Bandyopadhyay, A. Mukhopadhyay and U.Maullik,MultiobjectiveGenetic Algorithms for Clustering, Springer-Verlag, Berlin, Heidelberg,2011.

13 B. J. Frey and D. Dueck, Science, 2007, 315, 972–976.


14 C. A. Coello, Knowl. Inf. Syst., 1999, 1, 129–156.15 C. Coello Coello, D. V. Veldhuizen and G. Lamont, Evolutionary

Algorithms for Solving Multi-Objective Problems, Kluwer Aca-demic Publishers, 2002.

16 C. Coello Coello, IEEE Comput. Intell. Mag., 2006, 1, 28–36.17 M. Ashburner, C. Ball, J. Blake, D. Botstein, H. Butler, J. Cherry,

A. Davis, K. Dolinski and S. Dwight et al.,Nat. Genet, 2000, 25, 25–29.18 H. Wang, F. Azuaje, O. Bodenreider and J. Dopazo, Proc. IEEE

Symp. Comput. Intell. Bioinf. Comput. Biol., 2004, 25–31.19 F. M. Couto, M. J. Silva and P. M. Coutinho, Data Knowl. Eng.,

2007, 61, 137–152.20 P. Lord, R. Stevens, A. Brass and C. Goble, Bioinformatics, 2003,

19, 1275–1283.21 D. Lin, Proceedings of the 15th International Conference on Machine

Learning, 1998, pp. 296–304.22 J. J. Jiang and D. W. Conrath, Proceedings of the International

Conference Research on Computational Linguistics, 1997.

23 D. Huang, B. Sherman, Q. Tan, J. Collins, W. Alvord, J. Roayaei,R. Stephens, M. Baseler, H. Lane and R. Lempicki, Genome Biol.,2007, 8, R183.

24 U. Maulik, A. Mukhopadhyay and S. Bandyopadhyay, IEEETrans. Inf. Technol. Biomed., 2009, 13, 969–975.

25 A. Mukhopadhyay, U. Maulik and S. Bandyopadhyay, PLoS One,2012, 7, e32289.

26 U. Maulik, M. Bhattacharyya, A. Mukhopadhyay andS. Bandyopadhyay, Mol. BioSyst., 2011, 7, 1842–1851.

27 I. Xenarios, L. Salwinski, X. Duan, P. Higney, S. Kim andD. Eisenberg, Nucleic Acids Res., 2002, 30, 303–305.

28 U. Guldener, M. Munsterktter, M. Oesterheld, P. Pagel, A. Ruepp,H. Mewes and V. Stumpflen,Nucleic Acids Res., 2006, 34, 436–441.

29 R. Rosenzweig, P. A. Osmulski, M. Gaczynska and M. H. Glickman,Nat. Struct. Mol. Biol., 2008, 15, 573–580.

30 E. Champion, B. Lane, M. Jackrel, L. Regan and S. Baserga, Mol.Cell. Biol., 2008, 21, 6547–6556.

Detecting protein complexes in a PPI network: a gene ontology based multi-objective evolutionary...

Documents

Transcript of Detecting protein complexes in a PPI network: a gene ontology based multi-objective evolutionary...