Synergy network based inference for breast cancer metastasis

Abstract

Breast cancer is a world wide leading cancer and it is characterized by its aggressive metastasis. In many patients, microscopic or clinically evident metastases have already occurred by the time the primary tumor is diagnosed. Chemotherapy or hormonal therapy reduces the risk of distant metastasis by one-third, but it is estimated that about 70% to 80% of patients receiving treatment would have survived without it. Therefore, being able to predict breast cancer metastasis can spare a significant number of breast cancer patients from receiving unnecessary adjuvant systemic treatment and its related expensive medical costs. Current studies have demonstrated the potential value of gene expression signatures in assessing the risk of post-surgical disease recurrence. However, most of these studies attempt to develop genetic marker-based prognostic systems to replace the existing clinical criteria, while ignoring the rich information contained in established clinical markers. Clinical markers, such as patient history and laboratory analysis, which are the basis of day-to-day clinical decision support, are often underused to guide the clinical management of cancer in the presence of microarray data. As a result, given the complexity of breast cancer prognosis, we proposed a novel strategy based on synergy network that utilize both clinical and genetic markers to identify the potential hybrid signatures and investigate their interactions which are associated with breast cancer metastasis. In this study, a computational method is performed on publicly available microarray and clinical data. A rigorous experimental protocol is used to estimate the prognostic performance of the hybrid signature and other prognostic approaches. The hybrid signature performs significantly better than other methods, including the 70-gene signature, clinical makers alone and the St. Gallen consensus criterion. At 90% sensitivity level, the hybrid signature achieves 77% specificity, as compared to 53% for the 70-gene signature and 43% for the clinical makers. The predicted results also showed a strong dependence of regulator genes that are related to cell death in cell development process. These significant gene regulators are useful to understand cancer biology and in producing new drug design.

Full text

Procedia Computer Science 00 (2010) 000–000
Procedia
Computer
Science
www.elsevier.com/locate/procedia
WCIT-2010
Synergy network based inference for breast cancer metastasis
Farzana Kabir Ahmad a*, Safaai Derisb and Mohd. Syazwan Abdullahc
a, cGraduate Department of Computer Science, Universiti Utara Malaysia, 06010 Sintok, Kedah, Malaysia.
bFaculty of Computer Science and Information Systems, Universiti Teknologi Malaysia, 81310 Skudai, Johor, Malaysia.
Abstract
Breast cancer is a world wide leading cancer and it is characterized by its aggressive metastasis. In many patients, microscopic or
clinically evident metastases have already occurred by the time the primary tumor is diagnosed. Chemotherapy or hormonal
therapy reduces the risk of distant metastasis by one-third, but it is estimated that about 70% to 80% of patients receiving
treatment would have survived without it. Therefore, being able to predict breast cancer metastasis can spare a significant number
of breast cancer patients from receiving unnecessary adjuvant systemic treatment and its related expensive medical costs. Current
studies have demonstrated the potential value of gene expression signatures in assessing the risk of post-surgical disease
recurrence. However, most of these studies attempt to develop genetic marker-based prognostic systems to replace the existing
clinical criteria, while ignoring the rich information contained in established clinical markers. Clinical markers, such as patient
history and laboratory analysis, which are the basis of day-to-day clinical decision support, are often underused to guide the
clinical management of cancer in the presence of microarray data. As a result, given the complexity of breast cancer prognosis,
we proposed a novel strategy based on synergy network that utilize both clinical and genetic markers to identify the potential
hybrid signatures and investigate their interactions which are associated with breast cancer metastasis. In this study, a
computational method is performed on publicly available microarray and clinical data. A rigorous experimental protocol is used
to estimate the prognostic perfor mance of the hybrid s ignature and other pr ognosti c appr oaches. The hybr id signature perfor ms
significantly better than other methods, including the 70-gene signature, clinical makers alone and the St. Gallen consensus
criterion. At 90% sensitivity level, the hybrid signature achieves 77% specificity, as compared to 53% for the 70-gene signature
and 43% for the clinical makers. The predicted results also showed a strong dependence of regulator genes that are related to cell
death in cell development process. These significant gene regulators are useful to understand cancer biology and in producing
new drug design.
Keywords: Synergy network; Bayesian network; breast cancer metastasis; inference; conditional independence.
1. Introduction
Breast cancer is a leading cause of cancer-related death and among one of the most aggr essive metastasis disease
worldwide. The growing mortality rate, with 410,000 deaths each year has yield more than 1.6% of all women
deaths worldwide [1]. The major clinical problems of breast cancer ar e the recurrence of disseminated disease and
metastatic behavior. In numerous patients, miniature or clinically evident metastases have already occurred by the
time the primary tumor is diagnosed. Although, treatments such as chemotherapy and endocrine therapy could
reduce the risk of distant metastasis by approximately one-third, however it is predicted 80% of patient would have
survived without receiving these treatments. Being prescribed with highly expensive medicines which turn out to be
unnecessary has caused several complications and exacerbates the condition of breast cancer patients. As the results,
the study of tumor progression and breast cancer metastasis has become a great interest in biomedical field.
Procedia Computer Science 3 (2011) 1094–1100
www.elsevier.com/locate/procedia
1877-0509 c
⃝2010 Published by Elsevier Ltd.
doi:10.1016/j.procs.2010.12.178
c
⃝2010 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of the Guest Editor.
Open access under CC BY-NC-ND license.

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
Despite significant advances in the treatment of primary breast cancer and enormous studies that have been
conducted, the ability to infer the metastatic behavior of tumors remains one of the most clinical challenges in
oncology. The main cause for this setback is the complex interactions in the cancer progression and metastasis
formation. In early days, three commonly used treatment guidelines such as TNM (tumor, lymph nodes and
metastasis) Tumor Staging System, St. Gallen and NIH (National Institute of Health) consensus criteria have been
used to determine the distant metastases. These breast cancer indices are based on clinical markers such as tumor
size, lymph node involvement, patient age and the aggressiveness of the cancer founded on histopathological
parameters. Regardless of the prominent practiced of these indices, it provides inaccurate r esults in predicting
therapy failure with only 10% specificity at 90% sensitivity level. Thus, a more accurate prognostic criterion is
urgently needed to avoid unnecessary treatment in newly diagnosed patients.
Recently, the development of g enetic marker-based prognostic system has become a breakthrough in cancer
progression research and most studies concentr ated their efforts solel y on this approach. Yet, some researchers do
believe th at the application of gen e expression data to infer cancer progression is often overused in the presence of
clinical data [2]. Clinical data which has been used on daily basis has been neglected and the rich information
contained in established clinical markers has been ignored. Given the complexity of breast cancer metastasis, a more
practical and sensible strategy is to incorporate both clinical and genetic markers that may contain complementary
information.
A small number of studies have been conducted to determine the possibility of integrating clinical and genetic
markers to infer breast cancer metastasis [3, 4]. While some of these approaches show a great promise in
incorporating two different markers to infer cancer metastasis, the issue of high dimensionality data has rarely been
discussed. One important characteristic of microarray data is the extremely large amount of data in a very small
sample size. Thus, by integrating two markers to infer cancer metastasis could be computationally complex as large
number of variables may need to be examined. In this paper, we seek to improve the ability to infer breast cancer
metastasis using a novel strategy known as synergy network. This method is solely based on Bayesian network that
apply two different approaches: an information-theoretic approach and conditional independence approach. Our
keen interest is to obtain correctly learnt network in order to examine the two markers, clinical and genetic markers,
in the pr esence of a third variable which represent the state of the cell (metastasis). In addition, we offered scoring
markers interactions that provide insights into the tumor progression and indicate markers that highly regulate breast
cancer metastasis.
The reminder of this paper is organized as follows. Section 2 describes the method used to develop synergy
network based on Ba yesian network to integrate two diverse markers. This section also elaborates the approaches
taken to implement correct learnt structure learning in order to address the issue of high dimensional data. The
empirical results and discussion are presented in Section 3, while Section 4 provides concluding remarks.
2. Methods
2.1 Bayesian network
Bayesian n etwork is a probabilistic graphical model in which vertices represent random variables an d the absence
of an edge between two vertices represents con ditional independence. Consider a finite set Vn= {X1, X2,…, Xn} of
random variables. Bayesian network representation contains two components: a directed acyclic graph (DAG), G =
(Vn, E
G) which vertices correspond to random variables, and conditional probability distributions of the random
variables, given its dependent variables (parents) in G. The joint distribution of these conditional probability
distributions is defined as follows:




ni VX
iin XPaXPXXP |,...,
1(1)
wher e

ii XPaXP | is a set of conditional probabilities for each variables i
Xand


i
XPa is the set of variables
which are the parents of i
X in graph G.
We use an example to illustrate the basic idea of Ba yesian networks. Given a Bayesian network specified in Fig. 1
for 5 genes: X1, X2, X3, X4, and X5, this structure specifies the parents for genes X3, X4, and X5: Pa(X3)={X1, X2},
Pa (X4)={ X1 }, Pa(X5)={ X3}, where Pa(V) represents the parent vertex set for vertex V.
F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100 1095

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
Fig. 1: A simple Bayesian network representation that explicates relationships between five genes
In our context, it can be interpreted that when genes in Z are at fixed expressi on levels, expression levels of genes
in X do not give any information on the expression levels of genes in Y and vice versa. Once the structure of G is
specified for a set of genes, we can interpret a directional edge from X to Y in G as a statement that X is the “cause”
of Y, or the expression level of X has an effect on the expression level of Y. Therefore, obtaining the correct
structure of Bayesian network is essential to perform an efficient inference and correctly represent the dependency
relationship.
Determining the optimal network through Bayesian learning structure has been investigated for many decades. In
conjunction with the invention of micr oarray technology, the problem of sear ching the best fit network given the
datasets have become harder. It is due to the fact that analyzing these high-dimensionality data require a large
number of variables to be analyzed, which yield to exponential growth in sear ching space, known as NP hard.
Generally there are two approaches to learn the structure of Bayesian networks from data: th e search and scorin g
methods and dependency analysis methods.In the first approach, the learning problem is viewed as searching for a
structure that best fits the data. Different scoring methods have been applied to determine the fit between the
network structure an d the data, in cluding Bayesian scoring, entropy-based, and minimum descr iption length, among
others. The dependency analysis appr oach, on the other hand tries to discover from data the depen dencies among
variables and then use these dependencies to construct the network structure.Lately, the dependency analysis
approach is discovered to be more efficient than the search and scoring approach for sparse networks (the number of
edges in the graph is relatively small).In attaining our goal, to infer breast cancer metastasis by integrating clinical
and genetic markers, in this paper we proposed a new strategy to find the optimal structure. In the following section
2.2, we introduce the main steps of the proposed method.
2.2 Synergy network based on information-theoretic approach and conditional independence
Clinical markers and gene expression profiles play an important role in determining breast cancer metastasis.
Integrating and analyzing all th is information to discover factors that r egulate cancer progression require network-
based algorithm. In this study, we employed a novel strategy known as synergy network to achieve the objective.
This method is developed solely based on Bayesian network that rely on two structure learning algorithm:
information-theoretic approach and conditional independence. Our method generally has two different features. In
the first feature, the synergy network is implemented using mutual information that measures the cooperative effect
of two variables on the state of a third. The two variables in this case are genes and clinical markers (Gi and Ci), an d
the third state is the binary state variable representing the occurrence of metastasis (M) [M = 1 (Metastasis present)
and M = 0 (Metastasis not present)]. Mutual information is used to decide which interactions (edges) are more
prominent than the others. Mutual information between random variables X and Y is defined as follows:
 
YXHXHYXI |;  (2)
where

XH is the entropy of X and

YXH | is conditional entropy of X given Y. By applying the same rule to our
study, we formulaically calculated the integration of genes and clinical markers (Gi and Ci) as:
1096 F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
 
MCIMGIMCGI iiii ;;;,  (3)
where

MCGI ii ;, is the cooper ative effect of two variables and

MCIMGI ii ;; is the individual effect.
The proposed algorithm starts from a non-conn ected network, whereby there is still no edge involved between
nodes. Then, we calculate the mutual information for two nodes from the network and based on these mutual
information values the edges is ordered. Sequentially, conn ection between nodes is dr awn according to mutual
information ordered value, which also offer the highest scoring interactions values. The edges which has the mutual
information value less than thr eshold (threshold = 0.1) are excluded as candidates of correct edges. We only choose
edges that have higher values (> threshold) to be the correct edge as it contains better probability of conn ection.
In the second feature, we further examined the learning structure of constructed network (obtained from the first
feature) by using conditional independence approach. Two variables, for instance A and B may have different
structures, ABand AB but carrying the identical mutual information values. Thus, to overcome this issue,
in this algorithm, conditional indepen dence is used to search edges that are incorrect in a triangular structure as
depicted in Fig. 2. Conditional independence is defined as follows:





kjkikji XXPXXPXXXP |||, (4)
Hence, once we detected the edge create a triangle loop and hold the same mutual information value, all three edges
included in the triangle will be run based on equation and we used the result of these test to update the network.
Fig. 2: Triangular structure of three nodes and two edges
2.3 Dataset and Pre-processing
The proposed method is tested and analyzed on van’t Veer et al. [5] dataset, which was obtained from Integrated
Tumor Transcriptome Array and Clinical data Analysis database (ITTACA (2006)). This data set contains
expression profile information derived from 97 lymph n ode negative breast cancer patients, 55 years old or younger
and associated clinical information including age, tumor size, histological grade, angioinvasion, lymphocytic
infiltration, estrogen receptor (ER) an d progesterone receptor (PR) status, which all together form clinical markers.
Prior the implementation of proposed method, the missing values present in this dataset was addressed. Manifold
missing gene expression values is a common problem in microarray dataset. K-nearest neighbors (kNN) imputation
method with k = 10 was used to handle these missing values. The kNN imputation method is utilized as it is the
most robust and sensitive approach to estimate missing values in microarray data set. It is proven to be prominent
and effective method through Troyanskaya et al.’s research [6]. Subsequently, the processed dataset was used as an
input in the proposed method.
3. Results and Discussion
The proposed method was executed on the breast cancer dataset to obtain insights into the cancer development
and how various factors may trigger metastasis progression, producing a synergy net work as shown in Fig 3.
Additionally, the top ten scoring interactions for this particular network ar e given in Table 1. The learn ed network
reveals a group of genes a nd clinical markers which are primarily associated with causing metastasis, M. The larger
nodes in the graph specify th e genes when expressed at different levels lead to a major effect on the status of other
genes (e.g., on or off)/clinical marker, and the light-shaded nodes denote highly regulated genes. Four genes that are
found to regulate the expression levels of other genes are: BBC3, GNAZ, TSPY-like5 (TSPY5), and DCK. Two
genes are highly regulated: FLJ11354 and CCNE2. Meanwhile, angioinvasion has been identified as strong factor in
causing breast cancer metastasis. This network involved 50 genes and 6 clinical mar kers which are closely
associated with breast cancer metastasis, M.
F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100 1097

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
Fig. 3: S ynergy ne twork for breast ca ncer meta stasis
The constr ucted network indicates that the BBC3 gen e has a prominent role in regulating oth ers genes. Eight
genes are correlated with BBC3. The BBC3 gene, also known as PUMA is activated by the tumor suppressor p53,
which is a key regulator of apoptosis and tumorigenesis in breast cancer. On the other hand, there is insufficient
information about whether GNAZ could directly regulate the progression of breast cancer, however we discovered
that it has an essential role in cellular processes of the nervous system [7]. Meanwhile, TSPYL5 has been identified
as a genetic marker for breast cancer in several studies [4, 8]. Lastly, DCK is revealed to be associated with
resistan ce to antiviral and anticancer chemotherapeutic agents, therefore this gene is clinically important because of
its relationship to drug resistance and sensitivity. Outside of these regulator genes, two additional highly regulated
genes have been identified in the analysis of our proposed method: FLJ11354 and CCNE2. The FLJ11354 gene was
discovered by Sun et al. [9], while CCNE2 has been reported to qualify as independent prognostic markers for
lymph node–negative breast cancer patients [10].From the clinical markers point of view, angioinvasion is
identified as critical factor that yield to cancer progression compared to other clinical mar kers
We then further evaluated the performance of constructed synergy network using a receiver operating
characteristic (ROC) curve obtained by varying a decision threshold, which can provide a direct view on how this
inference network performs at the different sensitivity and specificity levels. By following the study of van’t Veer
and colleagues [5], a sensitivity is set equal to 90%. The corresponding specificities are computed and reported in
Table 2. For the purpose of comparison, the specificities of the TNM Tumor Staging System, St. Gallen and NIH
consensus criteria ar e also compar ed.
Table 1: The t op ten scorin g interactions. No. Rel indi cates the numb er of relation involved while Pr ed referred t o predictor genes.
No. Rel Pred Target Score
1 Metastasis Contig63649_RC 0.000487
2 Metastasis CCNE2 0.001081
3 UCH37 Contig40831_RC 0.003260
4 BBC3 PRC1 0.006655
5 BBC3 ORC6L 0.014038
1098 F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
6 WISP1 COL4A2 0.014892
7 DIAPH3 GNAZ 0.015032
8 MCM6 CCNE2 0.015653
9 DCK Contig55377_RC 0.017289
10 HRASLS FLJ22477 0.017314
Table 2: Inference network at sensitivity of 90%
Methods Specificit y AUC std
NIH 2000 0% 0.61905 0.16234
TNM Staging System 18% 0.71429 0.12747
St. Gallen 43% 0.73810 0.36204
Genetic markers 53% 0.79203 0.03245
Clinical and genetic markers
(hybrid signatures)
77% 0.86438 0.02928
We observed that the St. Gallen criterion significantly outperformed both the TMN staging system and the NIH
2000 consensus, whereas the latter approach (the NIH 2000 consensus) was worse than the TNM staging system.
The St. Gallen criterion achieved a specificity of 43%, while the TNM staging system and the NIH 2000 consensus
obtained a specificity of 18% and 0%, respectively. This result is consistent with previous reports in the literature
[11] whereby the specificity of the St. Gallen criterion outperforms the other clinical indices. On the other hand, the
clinical and genetic marker s (hybrid signatures) improve the specificities of the genetic mar kers and the clinical
markers (St. Gallen criterion) approximately by 20%-30%. We point out that our estimation of the specificity of the
70-gene signature is worse than that reported in [5](43% versus 73%), but is consistent with that in the follow-up
validation done on a larger dataset [12] (53%). Furthermore, we measured the area under curve (AUC) for all five
methods, where th e highest AUC suggesting a better inferen ce network. Therefore, our results have shown that
clinical and genetic marker s improved the specificity of inference n etwork compared to network those based on
genetic and clinical marker alone.
4. Conclusion
Understanding the breast cancer progression network structure reveals the inherent biological information flow
and interactions of various factors which will lead to more effective therapies and disease treatments. In this paper,
we applied computation model which was implemented based on synergy network to study the breast cancer
metastasis using genetic and clinical mar kers. Different genes and clinical markers were found to have high
correlation in causing metastasis. For future work, we intend to validate our discovery by using biological
knowledge. This attempt could arm biologist with information regarding up-stream and down-stream of gene
mechanisms, which further enlighten the interactions in tumor progression.
References
[1] P. Boyle and B. Levin, "World cancer report 2008," International Agency for Resear ch on Cancer, World
Health Organization 2008.
[2] P. Edén, C. Ritz, C. Rose, M. Fernö, and C. Peterson, "“Good Old” clinical markers have similar power in
breast cancer progn osis as microarray gene expression profilers," European Journal of Cancer, vol. 40, pp.
1837-1841, 2004.
[3] O. Geva ert, F. D. Smet, D. Timmerman, Y. Moreau, an d B. D. Moor, "Predicting the progn osis of breast
cancer by integrating clinical and microarray data with Bayesian networks," Bioinformatics, vol. 22, pp.
e184–e190, 2006.
[4] Y. Sun, S. Goodison, J. Li, L. Liu, and W. Farmerie, "Improved breast cancer prognosis through the
combination of clinical and genetic markers," Bioinformatics, vol. 23, pp. 30-37, 2007.
[5] L. J. van't Veer, H. Dai, M. J. van De Vijver, Y. D. He, A. M. Hart, M. Mao, et al., "Gene expression
profilin g predicts clinical outcome of breast cancer," Nature, vol. 415, pp. 530 - 536, 2002.
[6] O. Troyanskaya, M. Cantor, G. Sherlock, P. Brown, T. Hastie, R. Tibshirani, et al., "Missing value
estimation methods for dna microarrays," Bioinformatics, vol. 17, pp. 520–525, 2001.
F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100 1099

Farzana Kabir Ahmad / Procedia Computer Science 00 (2010) 000–000
[7] M. Kanehisa, S. Goto, M. Furumichi, M. Tanabe, and M. Hirakawa, "KEGG for representation and
analysis of molecular networks in volvin g disea ses and drugs," Nucleic Acids Research, pp. 1-6, 2009.
[8] G. Alexe, S. Alexe, D. E. Axelrod, T. O. Bonates, Lozina, M. Reiss, et al., "Breast cancer prognosis by
combinatorial analysis of gene expression data," Breast Cancer Research, vol. 8, pp. R41, 2006.
[9] Y. Sun, V. Urquidi, and S. Goodison, "Derivation of molecular sign atures for breast cancer recurrence
prediction using a two-way validation approach," in Breast Cancer Research Treatment: Springer
Netherlands, 2009.
[10] A. M. Sieuwerts, M. P. Look, M. E. Gelder, M. Timmermans, A. A. C. Trapman, R. RodriguezGarcia, et
al. "Which Cyclin E prevails as prognostic marker for breast cancer? Results from a retrospective study
involving 635 lymph node negative breast cancer patients," Clinical Cancer Research, vol. 12, pp. 3319-
3328, 2006.
[11] C. Lohrisch, J. Jackson, A. Jones, D. Mates, and I. A. Olivotto, "Relationship between tumor location and
relapse in 6,781 women with early invasive breast cancer," Journal of Clinical Oncology, vol. 18, pp.
2828-2835, 2000.
[12] M. J. van De Vijver, Y. D. He, L. J. van't Veer, H. Dai, A. M. Hart, D. W. Voskuil, and et al., "A gene-
expression signature as a predict of survival in breast can cer," The New England Journal of Medicine, vol.
347, pp. 1999-2009, 2002.
1100 F.K. Ahmad et al. / Procedia Computer Science 3 (2011) 1094–1100