Prediction of Cancer using Network Topological Features

Fernanda Brito Correia

1,3

, Joel P. Arrais

and José Luís Oliveira

Dept. of Electronics, Telecommunications and Informatics (DETI), Inst. of Electronics and Informatics Engineering of

Aveiro (IEETA), University of Aveiro, Aveiro, Portugal

Dept. of Informatics Engineering (DEI), Centre for Informatics and Systems of the University of Coimbra (CISUC),

University of Coimbra, Coimbra, Portugal

Dept. of Informatics and Systems Engineering (DEIS), Polytechnic Institute of Coimbra, Coimbra, Portugal

Keywords: Protein-protein Interaction Networks, Classification, Cancer Prediction.

Abstract: Several data mining methods have been applied to explore biological data and understand the mechanisms

that regulate genetic and metabolic diseases. The underlying hypothesis is that the identification of signatures

can help the clinical identification of diseased tissues. Under this principle many different methodologies have

been tested mostly using unsupervised methods. A common trend consists in combining the information

obtained from gene expression and protein-protein interaction networks analyses or, more recently, building

series of complex networks to model system dynamics. Despite the positive results that these works present,

they typically fail to generalize out of sample datasets. In this paper we describe a supervised classification

approach, with a new methodology for extracting the network topology dynamics embedded in a disease

system, to improve the capacity of cancer prediction, using exclusively the topological properties of biological

networks as features. Four microarrays datasets were used, for testing and validation, three from breast cancer

experiments and one from a liver cancer experiment. The obtained results corroborate the potential of the

proposed methodology to predict a certain type of cancer and the necessity of applying different classification

models to different types of cancer.

1 INTRODUCTION

Cancer is a complex genetic disease that affects an

increasing number of citizens all over the world. In

2015, more than 1.6 million new cancer cases are

expected in the United States, from which around

15% correspond to breast cancer (Siegel et al., 2015).

Understanding the underlying biological mechanisms

behind this disease has been the goal of many and

continuous research initiatives.

One strategy to study cancer is using microarrays,

a high-throughput technology that measures gene

expression, allowing the parallel analysis of genes in

several samples. Different individuals or different

conditions (healthy and non-healthy cells of the same

individual) originate distinct microarray samples with

gene expression values. These different samples can

reveal signatures that help to distinguish cancer from

non-cancer tissues.

In a network-based approach, bio-entities such as

genes and proteins can be represented as nodes and

their relationships as edges. Using this approach we

can model biological processes that can be analysed

using graph and network methods. The construction

of network-based models to study complex

phenomena, like cancer diseases, allows the capture

of the embedded systems dynamics. This dynamics

can be captured by constructing series of different

complex networks through time, through different

stages and through different traits. The study of the

topological properties of these complex networks

allows understanding specific structures, signatures

and similarities.

This paper presents a methodology to construct

protein-protein interaction networks to capture the

existent system dynamics beneath their topology.

These networks, here named Sample-Series

Networks (SSN), are constructed using a group of

cancer and healthy microarray samples. Information

features are exclusively obtained through the analysis

of the topological properties of these networks and

without any other biological information. Using the

obtained set of topological features a supervised

approach is then used to classify between cancer and

non-cancer tissues samples. This work aims to

Correia, F., Arrais, J. and Oliveira, J.

Prediction of Cancer using Network Topological Features.

DOI: 10.5220/0005696202070215

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 3: BIOINFORMATICS, pages 207-215

ISBN: 978-989-758-170-0

207

address several questions, namely: are there

evidences of signatures beneath the SSN that allow us

to classify samples as cancer or non-cancer? Which

topological measures give better results as

classification features among the several groups

considered? Does a classification model distinguish

different types of cancers?

Tests have been made in four gene expression

microarray experiments, three from breast cancer and

one from liver cancer.

2 BACKGROUND

Anomalies in a gene, protein or other bio-entity can

cause diseases and since the arrival of the next-

generation sequencing (NGS), that are found more

evidences of human genes being correlated to

diseases. Data from November 6

, 2015 obtained in

the Online Mendelian Inheritance in Man database

(OMIM) (Amberger et al., 2015) shows that there are

5597 phenotypes for which the molecular basis is

known and 3453 genes with phenotype-causing

mutation. Most disease genes are not essential, being

essential genes typically organized as hubs in a

complex network (Barabási et al., 2011, Jonsson and

Bates, 2006).

Biological networks are not random, they have

clustered groups of bio-entities, like genes or

proteins. They are also sparsely connected, which is

considered an evolutionary advantage for preserving

robustness to random failures (Barabasi and Oltvai,

2004). Also, it is known that genes and proteins that

are involved in the same phenotype are network

neighbours (Oti et al., 2006) and that a disease

phenotype can be associated to interactions in a

biological complex network that models these

biological processes (Menche et al., 2015).The

comparison of networks can use global and local

topological measures. Both are used in (Pržulj et al.,

2004) to show that the structure of yeast PPI networks

is closer to the geometric random graph model

relatively to graphlet frequency. In (Pržulj, 2007) a

new network similarity measure is defined based on

the graphlet degree distribution as a generalization of

the degree distribution.

Genomic changes that are translated to proteins

can alter biological functions and a system-based

approach modelled through complex networks can

assist the discovery of signatures related to disease

mechanisms, by analysing their topology (Vidal et al.,

2011, Barabási et al., 2011, Arrais and Oliveira, 2011,

Farkas et al., 2011).

Cliques help to understand the mechanisms

involved in cancer, since they are fully connected

subnetworks more conserved in biological networks.

In cliques, genes are functionally related and highly

expressed. In (Pradhan et al., 2012) it is proposed a

topological and biological feature-based network

approach, integrating the expression data, along with

network topological information and biological

information. Cliques are scored based on these

information and are considered as gene signatures for

the colorectal cancer (CRC).

Sets of biological complex networks can be

constructed across multiple conditions, like species,

time, and evolutionary states, traits or even samples,

as the novel approach used in this paper, building

dynamic models of the studied system.

A systems biology approach can be used to

interpret biological data. The (Trapé and Gonzalez-

Angulo, 2012) review addresses the contributions of

systems biology. DNA, RNA and protein changes

data are integrated to understand breast cancer

metastasis process. (Sonachalam et al., 2012) shows

how to build a PPI network representative of the

colorectal cancer (CRC) where nodes are

genes/proteins obtained from Gene Set Enrichment

Analysis (GSEA). (Barter et al., 2014) compares

single-gene, gene-set and two PPI network-based

methods, using gene expression microarrays data,

applied to melanoma and ovarian cancer. In single-

gene, features are the expression values of

informative genes identified via differential

expression analysis. In the gene-set method, genes are

grouped into sets using biological knowledge, which

are used as features for classification. In the first

network-based method features are the most

informative individual genes selected using the PPI

network, while in the second network-based method,

features are identified or are extracted from them

considering the edges or are sub-networks hub genes.

Three classifiers were used, namely Random Forest

(RF), Diagonal Linear Discriminant Analysis and

Support Vector Machines (SVM), with 5-fold cross

validation. It concludes that including network

information may lead to the identification of more

stable gene expression signatures.

(Dominietto et al., 2015), shows how to integrate

imaging data into networks to define tumor

fingerprints, through both network topology and the

detection of dynamic connectivity patterns.

In (Chuang et al., 2007) PPI subnetwork markers

are found to distinguish between metastatic and non-

metastatic tumors, using a score function. Candidate

subnetworks are built starting with a single protein

and are expanded using the PPI network, until the

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

208

score stops to increase. The activity scores calculated

from the average of the expression levels of each

subnetwork were used as feature values. The

classifiers used were based on logistic regression and

SVM using 5-fold cross validation. In (Chen and

Yang, 2014), normal, benign and malignant states of

breast cancer are differentiated, building a gene

regulatory network representative of each state and

comparing their network topological properties (in

and out-degree, betweenness, cluster coefficient and

closeness). Gene ranking was made selecting 53 hub

genes. (Wang et al., 2015) review describes pathway

and network-based approaches applied to cancer

biomarker discovery, in particular to the liver and

hepatocellular carcinoma (HCC). In (Ou-Yang et al.,

2014) dynamic PPI networks are constructed from

time-course gene expression data and PPI data,

extracting stable and dynamic interactions along time

to predict temporal protein complexes. An approach.

using differential co-expression analysis and PPI

networks for study human HCC progression that uses

subnetworks for each of the five stages of this

carcinoma can be found in (Yu et al., 2013).

Data mining classification techniques have been

used to look for signatures in cancer diseases. A large

number of variables can be used to characterize

cancer and non-cancer biological datasets, so it is

necessary to choose the most relevant. There are

several feature selection algorithms and a review is

presented in (Saeys et al., 2007), including the

ReliefF feature selection method (Kononenko, 1994)

used in this paper. In (Nancy and alias Balamurugan,

2013), the ReliefF feature selection method is

claimed to be the best method among several tested

for cancer classification using gene expression data.

Also ReliefF algorithm is efficient, and adequate

when there is much feature interaction, ranking well

the quality of features when there is a strong

dependency between them (Robnik-Šikonja and

Kononenko, 2003).

In (Furey et al., 2000), a score is calculated for the

expression values of genes, to select those with

highest scores as features in the classification

withhold-one-out cross validation. Tests were made

and the best results were obtained with 50 genes.

(Ramani and Jacob, 2013) uses a Bayesian Network

Learning (BNL) prediction to classify lung cancer

tumors as Small Cell Lung Cancer (SCLC), Non-

Small Cell Lung Cancer (NSCLC) and COMMON

classes, using the structural and physicochemical

properties of protein sequences obtained from genes

using microarray analysis. Several feature selection

methods were used with different prediction

techniques. Best results were obtained using BNL

with Gain Ratio. A model for predicting the survival

rate of patients affected by lung cancer, applying

different feature selection algorithms, can be found in

(Dezfuly and Sajedi, 2015). The classification

algorithms used were: Decision Tree (DT), BNL and

Neural Network (NN).

A new network-based supervised classification

method to predict cancer, named NBC and using only

gene expression levels is presented in (Ay et al.,

2014). It was applied to different datasets, (lung,

breast, leukaemia and colon cancers) using five

classification algorithms, namely SVM, KNN, NBL,

C4.5 and RF with 10-fold cross validation and with

five feature selection methods. A gene-association

network was created for each class, where nodes are

genes and edges represent the correlation between

their expression levels. High accuracy classification

was obtained with less than 100 genes.

3 METHODS

This paper describes a system-based approach to

classify between cancer and non-cancer tissues and

can contribute to find signatures that distinguish

disease biological processes from healthy biological

processes, using the topological properties of

networks and considering the network topological

dynamics embedded in the disease system.

A network-based method is used, by constructing

a set of PPI networks, one for each sample belonging

to the SSN. The ReliefF algorithm (Kononenko,

1994) is used to rank a subset of genes. In each SSN

network, nodes are proteins coded by a subset of the

most expressed genes of the top ReliefF genes and

edges indicate that the proteins coded by those genes

interact physically. A score was used as a threshold

for the PPI interactions.

Network topological properties were used as

features in the supervised binary classification

methodology and their values were obtained from the

topological analysis of each SSN network.

These classification models were evaluated using

the statistical measures, accuracy, precision, recall,

F1-score and area under the ROC curve.

Four datasets were used, three from breast cancer

microarray experiments and one from a liver cancer

microarray experiment. Three types of tests were

made: using 5 fold cross-validation; using data

obtained from two of the breast cancer datasets as

train set, and data obtained from the other breast

cancer dataset as test set; and using data obtained

from the three breast cancer datasets to train the

Prediction of Cancer using Network Topological Features

209

dataset and, for testing, using data obtained from the

liver cancer dataset.

3.1 Gene Expression Microarray Data

Sets

The experiments were obtained from ArrayExpress

(http://www.ebi.ac.uk/arrayexpress/): E-GEOD-

65194 (178 samples, where 167 are from breast

cancer tissue cells), E-GEOD-54002 (433 samples,

where 417 are from breast cancer tissue cells), E-

GEOD-29044 (124 samples, where 75 are from breast

cancer tissue cells) and E-MTAB-950 (276 samples,

where 179 are from liver cancer tissue cells). To

assure probes and samples uniformity all experiments

share the same array design A-AFFY-44 and all

samples were labelled as belonging to one of the two

classes, Cancer or Healthy. The 54673 genes of the

experiments were sorted by decreasing values of

ReliefF (Kononenko, 1994), an algorithm that can be

applied to continuous and discrete values. For each

experiment the top ReliefF 100 genes were selected

and merged in one matrix of 735 samples and 276

genes for breast cancer and one matrix of 124 samples

and 276 genes for the liver cancer. The PPI SSN was

obtained from the 100 most expressed genes from

each sample of the previous matrixes. The number

100 genes belongs to the typical interval of 50 to 150

number of genes used for binary classifications

studies (Ay et al., 2014).

3.2 Protein-protein Interaction

Networks

DAVID and UNIPROT were primarily used (Dennis

et al., 2003, Consortium, 2014) to obtain the mapping

of identifiers from probeset ids and gene names to

proteins. The human PPI dataset was obtained from

STRING, an online database resource with several

distinct types and sources of PPI information. Using

this dataset, several networks were constructed, one

for each sample, representing the entire set of PPI for

all different samples. Only PPIs with score greater or

equal to 300 were considered. These networks were

constructed as undirected, unweight and with no self-

edges.

3.3 Classification Methods

The set of supervised learning algorithms used were

the KNN classifier, the SVM classifier implemented

using an RBF kernel, and the RF, all with default

parameters.

Classification results were obtained using Orange

through scripting in Python and through visual

programming in Orange Canvas (Demsar et al.,

2007).

The statistical measures used to evaluate the

performance of the binary classification models were,

the classification accuracy (CA), the precision

(Precision), the recall (Recall), the F1-score (F1) and

the area under the ROC Curve (AUC) (Sokolova and

Lapalme, 2009), where TP (true positive) is the

number of correctly predicted samples that belong to

the class, TN (true negative) is the number of

correctly predicted samples that do not belong to the

class, FP (false positive) is the number of wrongly

predicted samples that belong to the class and FN

(false negative) is the number of wrongly predicted

samples that do not belong to the class.

Accuracy (CA) calculates the proximity of

measurement results to the true value and gives the

global efficacy of a classifier.

CA = (TP+TN) / (P+N) (1)

Precision (Precision) specifies the positive labels

given by the classifier that are correct.

Precision = TP / (TP+FP) (2)

Recall or sensitivity shows the efficacy of a

classifier to identify positive labels.

Recall = TP/P = TP / (TP+FN). (3)

F1-score (F1) is the harmonic mean of precision

and recall and is between 0 and 1, being 1 the best

value.

F1 = 2 x (Precision x Recall) /

(Precision + Recall)

(4)

Area under the ROC curve (AUC) is the

classifier’s capacity to avoid false classification.

AUC = 1/2 ((TP / (TP+FN))

+ (TN / (FP + TN))

(5)

Three strategies were used, the first one with

classification results obtained by 5 fold cross-

validation and the others two using a separate test

data, one from the same type of cancer and the other

from a different type of cancer.

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

210

Figure 1: SSN network-based features, where ns is the number of samples and nf is the number of features.

3.4 Network-based Approach

An undirected and unweighted graph G can be

defined as a pair G = (V, E) where V is a set of

vertices representing the nodes and E is a set of edges

representing the connections between the nodes i and

j. The number of nodes of a graph G is denoted by N

and the number of edges of a graph is denoted by L.

Given a graph G = (V, E) the adjacency matrix

representation consists of an N x N matrix A = [a

such that a

= 1 if (i, j) belongs to E or a

= 0

otherwise. For undirected graphs the matrix is

symmetric.

Descriptors obtained from the analysis of the SSN

topologies were calculated using the package

NetworkX from Python (Schult and Swart, 2008), the

R package QuACN (Mueller et al., 2011) and the

gtrieScanner (Ribeiro and Silva, 2014) software.

The following methodology was used to obtain

SSN network-based features (Figure 1):

Step 1: Obtain the e matrixes, for e=1,…, ne,

where ne is the number of microarray experiments,

is the number of samples of the experiment e and

ng is the number of genes.

EXP

= [exp

], i=1,…, ns

; j=1, …, ng (6)

Step 2: Obtain the lists of the top genes ranked by

decreasing order of ReliefF

LTGR

, for e=1,..., ne (7)

Step 3: Obtain the union of the previous lists, for

a threshold value, thr_rf that defines the number of

top elements of the lists to be considered.

LUNION = UNION (LTGRe), for

e=1,...,ne

(8)

Step 4: Obtain the submatrices of EXP

, for e=1,

..., ne, obtained in step 1., for the genes selected in

step 3.

SEXP

= [sexp

], for i=1,..., sum

(ns

);

j=1,…, thr_rf

(9)

Step 5: Obtain the lists of the top thr_me most

expressed genes in SEXP, for each sample from e

experiments, for e=1,…, ne.

LGME= [lgme

], for i=1,..., sum

(ns

);

j=1, .., thr_me

(10)

Step 6: Obtain the lists of the proteins encoded by

the genes of the LGME matrix, P(LGME) for each

sample of the experiments e, for e=1, ..., ne.

LPME= P(LGME)

, for i=1,..., sum

(ns

) (11)

Step 7: Obtain the SSN, the PPI human interaction

sub-networks induced by LPME.

SSN= [ssn

], for i=1,..., sum

(ns

); e=1,…,

(12)

Prediction of Cancer using Network Topological Features

211

Table 1: Group D0 of topological network-based descriptors.

D0.1: Number of nodes D0.9: Size of the largest clique

D0.2: Number of edges D0.10: Number of maximal cliques

D0.3: Density D0.11: Degree assortativity coefficient

D0.4: Number of connected components D0.12: Estrada index

D0.5: Number of nodes of the largest component D0.13: Graph transitivity

D0.6: Number of edges of the largest component D0.14: Average clustering coefficient

D0.7: Diameter of the largest component D0.15: Average shortest path length

D0.8: Global clustering coefficient

Table 2: Groups D1, D2 and D3 of topological network-based descriptors.

D1.1: Wiener D2.1: Total adjacency D3.1: Medium articulation

D1.2: Harary D2.2: Zagreb 1 D3.2: Efficiency

D1.3: BalabanJ D2.3: Zagreb 2 D3.3: Graph index complexity

D1.4: Mean distance deviation D2.4: Modified Zagreb D3.4: Off diagonal

D1.5: Compactness D2.5: Augmented Zagreb D3.5: Spanning tree sensitivity STS

D1.6: Product of row sums D2.6: Variable Zagreb D3.6: Spanning tree sensitivity STSD

D1.7: Hyper distance path index D2.7: Randic

D1.8: Dobrynin eccentricity graph D2.8: Complexity index B

D1.9: Dobrynin avgecc of G D2.9: Normalized edge complexity

D1.10: Dobrynin eccentric graph D2.10: Atom bond connectivity

D1.11: Dobrynin graph integration D2.11: Geometric arithmetic 1

D1.12: Dobrynin unipolarity D2.12: Geometric arithmetic 2

D1.13: Dobrynin variation D2.13: Geometric arithmetic 3

D1.14: Dobrynin centralization D2.14: Narumi Katayama

D1.15: Dobrynin average distance

D1.16: Dobrynin mean distance

vertex deviation

Step 8: Calculate the five groups of topological

properties for each ssn belonging to SSN, where the

number of features, nf, is the number of descriptors

used.

FEAT= [feat

], for i=1,…, sum

(ns

);

j=1,..., nf; e=1,..., ne

(13)

Several topological measures were included to

capture the structural complexity of the biological

networks. Hereafter are referred the five groups of

descriptors used.

A first group of 15 descriptors, named D0,

calculated using NetworkX (

Table 1). A second group

of 16 descriptors, named D1 that uses distances

between nodes to capture the structural complexity of

the network, a third group, named D2, of 14

descriptors and a forth group of 6 more recent

descriptors, named D3, all of them calculated using

QuACN (Table 2). A fifth group , named D4, of 58

descriptors, where the first 29 are corresponding to

the relative frequency values of 3 to 5 nodes

subgraphs if they are a motif and zero if they are not

a motif and the last 29 are the correspondent z-score

values, which were calculated using 1000 random

networks (

Table 3).

These descriptors were calculated using the

gtriesScanner software. A motif is a subgraph that is

frequent compared to their frequency in a set of

similar random networks. In this paper a subgraph is

considered a motif (Milo et al., 2002), if the

frequency of the subgraph in the network is superior

to 4, the difference between the frequency in the

network (f) and the average frequency in 1000 similar

random networks (avgfr) is greater or equal to 0.10 of

the average frequency in those random networks, and

|z-score| > 2, where z-score = (f - avgfr) / sd, with sd

as the standard deviation.

Table 3: Group D4 of topological network-based

descriptors.

D4.1_j: 3_j_fr

j=1,… 2

D4.4_j: 3_j_zsc

j=1,… 2

D4.2_j: 4_j_fr

j=1,… 6

D4.5_j: 4_j_zsc

j=1,… 6

D4.3_j: 5_j_fr

j=1,… 21

D4.6_j: 5_j_zsc

j=1,… 21

To build the binary classification models three

different supervised learning algorithms were used,

namely the KNN, SVM with RBF kernel and RF

classifiers, all with default parameters. All of the

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

212

Table 4: Statistical evaluation (CA, Precision, Recall, F1 and AUC) of the binary classification C - Cancer and H- Healthy

for the cases C1, C2, C3 and C4 using the three classifiers KNN, SVM and RF, for all of the network-based features and for

the group of network-based features D4 for the class C.

classifiers used gave similar results, with a slight

advantage in some statistical measures for RF when

using information from breast cancer datasets.

The statistical measures used to evaluate the

performance of the binary classification models were

the CA, the Precision, the Recall, the F1 and the AUC

and values were obtained using three strategies, with

different sets of features groups.

The first strategy, named Case 1 (C1), used 5 fold

cross-validation on the network-based features values

calculated from the three breast cancer microarray

datasets.

The second strategy included two types of tests

that were named Case 2 (C2) and Case 3 (C3) and

here two of the breast cancer datasets were used to

calculate network-based features values for the

training dataset and the remaining one was used to

calculate the network-based features values for the

test dataset. In C2, the training set used E_GEOD-

65194 and E-GEOD-54002 microarray datasets and

the test set used E_GEOD-29044 dataset and in C3,

the training set used E_GEOD-54002 and E-GEOD-

29044 datasets and the test set used E_GEOD-65194

microarray dataset.

The third strategy, named Case 4 (C4), used data

from the three breast cancer microarray datasets for

the training dataset and the liver microarray dataset

was used for the test dataset.

The two sets of features, whose results are shown

in Table 4, are the set of all of the network-based

features (groups D0 to D4) and the set of network-

based features of the group D4. In the case C1, 5-fold

cross validation was used, with results, above 0.95,

for all the statistical measures considered. To test if

the classification obtained in C1 was suffering from

over fitting, different datasets were used as a train set

and as test set, the cases C2 and C3. The results

obtained were, for example for CA, above 0.80 for C2

and above 0.92 for C3, which evidence good

performance of the classifier. The difference between

the values of C2 and C3 may be explained by the

imbalance between cancer and non-cancer samples.

To check if the classification models with datasets

of one type of cancer can be generalized for another

cancer type, the classification model was trained with

data from three breast cancer datasets and tested with

data obtained from a liver cancer dataset, in case C4.

Values obtained and shown in Table 4 are still

positive, probably due to the fact of all being cancer

diseases, but worse than the previous ones.

To analyse which of the network-based features

contributed more for the classification model a

ranking list of features was done. Table 5 shows the

top 5 ranking of the network-based features.

From the analysis of which features are more

informative, it can be stated that the most relevant

features belong mainly to group D0 and group D4.

When all groups of topological features are used as

features variables, it can be seen that the size of the

largest clique and the number of nodes are better

ranked. Motifs of size 4 and 5 are the most

informative motifs.

Table 5: Top five ranking of network-based features.

4 CONCLUSIONS

The statistical evaluation results were obtained using

only topological properties as features variables,

C 1C 2C 3C 4C 1C 2C 3C 4C 1C 2C 3C 4C 1C 2C 3C 4C 1C 2C 3C 4

KNN 0.98 0.80 0.97 0.61 0.98 0.76 0.98 0.64 0.97 0.97 0.99 0.88 0.98 0.85 0.99 0.75 0.96 0.86 0.94 0.55

SVM 0.96 0.81 0.96 0.62 0.99 0.78 0.99 0.70 0.97 0.95 0.96 0.72 0.98 0.86 0.98 0.71 0.98 0.94 0.98 0.64

RF 0.96 0.88 0.98 0.60 0.98 0.88 0.98 0.69 0.98 0.93 1.00 0.69 0.98 0.90 0.99 0.69 0.98 0.94 0.99 0.60

KNN 0.95 0.76 0.92 0.58 0.98 0.73 0.95 0.67 0.97 0.95 0.97 0.68 0.97 0.83 0.96 0.67 0.96 0.82 0.93 0.56

SVM 0.96 0.81 0.96 0.62 0.98 0.79 0.97 0.76 0.97 0.93 0.99 0.60 0.98 0.85 0.98 0.67 0.97 0.87 0.99 0.62

RF 0.96 0.85 0.98 0.57 0.97 0.84 0.98 0.64 0.98 0.93 1.00 0.79 0.98 0.89 0.99 0.71 0.97 0.93 0.99 0.51

D0+D1+D2+ D3+D4

F1 AUC

Cancer

CA Precision Recall

Tests 1st 2nd 3rd 4th 5th

D0+D1+D2+D3+D4 (case 1) D0.9 D0.5 D0.1 D3.2 D4.3_1

D0+D1+D2+D3+D4 (case 2) D4.2_6 D0.1 D0.5 D0.6 D0.2

D0+D1+D2+D3+D4 (case 3) D0.9 D0.4 D4.2_2 D0.13 D0.1

D0+D1+D2+D3+D4 (case 4) D0.9 D0.5 D0.1 D3.2 D4.3_1

D4 (case 1) D4.3_1 D4.1_1 D4.1_6 D4.2_6 D4.5_3

D4 (case 2) D4.2_6 D4.3_17 D4.3_4 D4.3_2 D4.3_1

D4 (case 3) D4.3_18 D4.3_19 D4.2_5 D4.3_17 D4.2_2

D4 (case 4) D4.3_1 D4.1_1 D4.2_1 D4.2_6 D4.5_3

Prediction of Cancer using Network Topological Features

213

measured in the SSN, which are PPI networks built

from the expressed genes without any other biological

information. The results seem to indicate that there

are signatures embedded in the topology dynamics,

modelled through the SSN, which can distinguish

cancer from non-cancer cells for each type of cancer.

This new methodology of creating SSN allows the

capture of the topology dynamics of the system

through the set of samples and allows data to be

reduced and be computationally manageable, keeping

the more informative data, which is supported by the

good results obtained. We consider that this novel

approach is worth and gives different contributions

compared to previous works, namely: the number of

considered topological properties is much higher; the

exclusive use of topological properties (global and

local) with good binary classification results

obtained; the topological dynamics of the system

captured through each sample, different from other

works that use time or states for example, which can

contribute to the capture of different signatures.

The results obtained show that classification

models should be different according to the cancer

disease type considered. More, the knowledge of

which features are more informative can be used, in

the future, to look for signatures based in these

features that could help in the identification of certain

cancer types. Two of the most discriminative features

obtained were the size of the largest clique and motifs

of size 4 and 5. Cliques being fully connected

subnetworks where genes are functionally related and

highly expressed were considered by some

researchers as gene signatures (Pradhan et al., 2012).

The relative frequency and z-score of some motifs as

local topological properties measures, showed to be

discriminatory features, indicating that there are clues

that some small subnetworks could help to distinguish

cancer samples. Adding more biological information

to the more discriminative features found in the

classification, may reveal important signatures like

subgraphs markers of cancer diseases. This approach

also seems worth to be further explored.

Finally the proposed methodology for creating

SSN is a novel contribution that can be extended to

other types of networks, besides PPIs, adding

information that can differentiate samples and capture

their topological dynamics helping to uncover new

signatures that can be biologically relevant for the

identification of diseases.

ACKNOWLEDGEMENTS

This work has received support from the RD-

CONNECT European project (EC contract number

305444).

REFERENCES

Amberger, J. S., Bocchini, C. A., Schiettecatte, F., Scott, A.

F. & Hamosh, A., 2015. Omim. Org: Online Mendelian

Inheritance In Man (Omim®), An Online Catalog Of

Human Genes And Genetic Disorders. Nucleic Acids

Research, 43, D789-D798.

Arrais, J. P. & Oliveira, J. L., 2011. Using Biomedical

Networks To Prioritize Gene-Disease Associations.

Open Access Bioinformatics, 1, 123-130.

Ay, A., Gong, D. & Kahveci, T., 2014. Network-Based

Prediction Of Cancer Under Genetic Storm. Cancer

Informatics, 13, 15.

Barabási, A.-L., Gulbahce, N. & Loscalzo, J., 2011.

Network Medicine: A Network-Based Approach To

Human Disease. Nature Reviews. Genetics, 12, 56-68.

Barabasi, A.-L. & Oltvai, Z. N., 2004. Network Biology:

Understanding The Cell's Functional Organization.

Nature Reviews Genetics, 5, 101-113.

Barter, R., Schramm, S.-J., Mann, G. & Yang, Y. H., 2014.

Network-Based Biomarkers Enhance Classical

Approaches To Prognostic Gene Expression

Signatures. Bmc Systems Biology, 8, S5.

Chen, D. & Yang, H., 2014. Comparison Of Gene

Regulatory Networks Of Benign And Malignant Breast

Cancer Samples With Normal Samples. Genetics And

Molecular Research: Gmr, 13, 9453.

Chuang, H. Y., Lee, E., Liu, Y. T., Lee, D. & Ideker, T.,

2007. NetworkBased Classification Of Breast Cancer

Metastasis. Molecular Systems Biology, 3, 140.

Consortium, T. U., 2014. Activities At The Universal

Protein Resource (Uniprot). Nucleic Acids Research,

42, D191-D198.

Demsar, J., Zupan, B. & Leban, G., 2007. Orange: From

Experimental Machine Learning To Interactive Data

Mining. White Paper, Faculty Of Computer And

Information Science, University Of Ljubljana (2004).

Dennis, G., Jr., Sherman, B. T., Hosack, D. A., Yang, J.,

Gao, W., Lane, H. C. & Lempicki, R. A., 2003. David:

Database For Annotation, Visualization, And

Integrated Discovery. Genome Biol, 4, P3.

Dezfuly, M. & Sajedi, H,. 2015. Predict Survival Of

Patients With Lung Cancer Using An Ensemble Feature

Selection Algotithm And Classification Methods In

Data Mining. Journal Of Information, 1, 1-11.

Dominietto, M., Tsinoremas, N. & Capobianco, E., 2015.

Integrative Analysis Of Cancer Imaging Readouts By

Networks. Molecular Oncology, 9, 1-16.

Farkas, I. J., Korcsmáros, T., Kovács, I. A., Mihalik, Á.,

Palotai, R., Simkó, G. I., Szalay, K. Z., Szalay-Beko,

M., Vellai, T. & Wang, S., 2011. Network-Based Tools

For The Identification Of Novel Drug Targets. Sci

Signal, 4, Pt3.

Furey, T. S., Cristianini, N., Duffy, N., Bednarski, D. W.,

Schummer, M. & Haussler, D. 2000. Support Vector

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

214

Machine Classification And Validation Of Cancer

Tissue Samples Using Microarray Expression Data.

Bioinformatics, 16, 906-914.

Jonsson, P. F. & Bates, P. A., 2006. Global Topological

Features Of Cancer Proteins In The Human

Interactome. Bioinformatics, 22, 2291-7.

Kononenko, I. Estimating Attributes: Analysis And

Extensions Of Relief. Machine Learning: Ecml-94,

1994. Springer, 171-182.

Menche, J., Sharma, A., Kitsak, M., Ghiassian, S. D., Vidal,

M., Loscalzo, J. & Barabási, A.-L., 2015. Uncovering

Disease-Disease Relationships Through The

Incomplete Interactome. Science, 347, 1257601.

Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N.,

Chklovskii, D. & Alon, U., 2002. Network Motifs:

Simple Building Blocks Of Complex Networks. Science,

298, 824-827.

Mueller, L., Kugler, K., Graber, A., Emmert-Streib, F. &

Dehmer, M., 2011. Structural Measures For Network

Biology Using Quacn. Bmc Bioinformatics, 12, 492.

Nancy, S. G. & Alias Balamurugan, S. A., 2013. A

Comparative Study Of Feature Selection Methods For

Cancer Classification Using Gene Expression Dataset.

Journal Of Computer Applications (Jca), 6, 2013.

Oti, M., Snel, B., Huynen, M. A. & Brunner, H. G., 2006.

Predicting Disease Genes Using Protein–Protein

Interactions. Journal Of Medical Genetics, 43, 691-

698.

Ou-Yang, L., Dai, D.-Q., Li, X.-L., Wu, M., Zhang, X.-F.

& Yang, P., 2014. Detecting Temporal Protein

Complexes From Dynamic Protein-Protein Interaction

Networks. Bmc Bioinformatics, 15, 335.

Pradhan, M. P., Nagulapalli, K. & Palakal, M. J., 2012.

Cliques For The Identification Of Gene Signatures For

Colorectal Cancer Across Population. Bmc Systems

Biology, 6, S17.

Pržulj, N., 2007. Biological Network Comparison Using

Graphlet Degree Distribution. Bioinformatics, 23,

E177-E183.

Pržulj, N., Corneil, D. G. & Jurisica, I., 2004. Modeling

Interactome: Scale-Free Or Geometric?

Bioinformatics, 20, 3508-3515.

Ramani, R. G. & Jacob, S. G., 2013. Improved

Classification Of Lung Cancer Tumors Based On

Structural And Physicochemical Properties Of Proteins

Using Data Mining Models. Plos One, 8, E58772.

Ribeiro, P. & Silva, F,. 2014. G-Tries: A Data Structure For

Storing And Finding Subgraphs. Data Mining And

Knowledge Discovery, 28, 337-377.

Robnik-Šikonja, M. & Kononenko, I., 2003. Theoretical

And Empirical Analysis Of Relieff And Rrelieff.

Machine Learning, 53, 23-69.

Saeys, Y., Inza, I. & Larrañaga, P., 2007. A Review Of

Feature Selection Techniques In Bioinformatics.

Bioinformatics, 23, 2507-2517.

Schult, D. A. & Swart, P. Exploring Network Structure,

Dynamics, And Function Using Networkx. Proceedings

Of The 7th Python In Science Conferences (Scipy

2008), 2008. 11-16.

Siegel, R. L., Miller, K. D. & Jemal, A., 2015. Cancer

Statistics, 2015. Ca: A Cancer Journal For Clinicians,

65, 5-29.

Sokolova, M. & Lapalme, G., 2009. A Systematic Analysis

Of Performance Measures For Classification Tasks.

Information Processing & Management, 45, 427-437.

Sonachalam, M., Shen, J., Huang, H. & Wu, X., 2012.

Systems Biology Approach To Identify Gene Network

Signatures For Colorectal Cancer. Frontiers In

Genetics, 3.

Trapé, A. P. & Gonzalez-Angulo, A. M., 2012. Breast

Cancer And Metastasis: On The Way Toward

Individualized Therapy. Cancer Genomics -

Proteomics, 9, 297-310.

Vidal, M., Cusick, M. E. & Barabasi, A.-L., 2011.

Interactome Networks And Human Disease. Cell, 144,

986-998.

Wang, J., Zuo, Y., Man, Y.-G., Avital, I., Stojadinovic, A.,

Liu, M., Yang, X., Varghese, R. S., Tadesse, M. G. &

Ressom, H. W., 2015. Pathway And Network

Approaches For Identification Of Cancer Signature

Markers From Omics Data. Journal Of Cancer, 6, 54.

Yu, H., Lin, C.-C., Li, Y.-Y. & Zhao, Z., 2013. Dynamic

Protein Interaction Modules In Human Hepatocellular

Carcinoma Progression. Bmc Systems Biology, 7, S2.

Prediction of Cancer using Network Topological Features

215