Journal of Cancer Therapy, 2009, 1, 28-35
Published Online September 2009 in SciRes (www.SciRP.org/journal/cancer)
Regulatory Network Motifs and Hotspots of Cancer
Genes in a Mammalian Cellular Signaling Network
ABSTRACT
Mutations or overexpression of signaling genes can result in cancer development and metastasis. In this study, we
manually assembled a human cellular Signaling network and developed a robust bioinformatics strategy for extracting
cancer-associated single nucleotide polymorphisms (SNPs) using expressed sequence tags (ESTs). We then investigated
the relationshipsof cancer-associated genes [cancer-associated SNP genes, known as cancer genes (CG) and cell mo-
bility genes (CMGs)] in a signaling network context. Through a graph-theory-based analysis, we found that CGs are
significantly enriched in network hub proteins and cancer-associated genes are significantly enriched or depleted in
some particular network motif types. Furthermore, we identified a substantial number of hotspots, the three- and
four-node network motifs in which all nodes are either CGs or CMGs. More importantly, we uncovered that CGs are
enriched in the convergent target nodes of most network motifs, although CMGs are enriched in the source nodes of
most motifs. These results have implications for the foundations of the regulatory mechanisms of cancer development
and metastasis.
Keywords: background contrast, breast, conformal mesh, microwave imaging.
1. Introduction
Cancer cells are characterised by uncontrolled cell grow-
th, invasion of surrounding tissues and finally metastasis
to distant regions of the human body. Accumulation of
genetic mutations in part triggers tumour development
and progression. Gene mutation or deregulation also pro-
motes cell mobility that is highly correlated with tissue
invasion and distant metastasis. A set of gene mutations
or overexpressions are closely linked to patient clinical
outcomes, suggesting that these genes could be cancer
biomarkers for diagnostics.
Cells use sophisticated communication between pro-
teins to perform a series of tasks such as growth, mainte-
nance of cell survival, proliferation and development.
Signaling pathways, which are used to transmit biological
signals, perform the communication between proteins.
Signaling pathways are crucial in maintaining cellular
homeostasis and determine cell behaviour. Thus, altera-
tions of expression of the genes in cellular signaling
pathways could lead to tumour development or promote
cell migration. Indeed, alterations to genes that encode
signaling proteins are commonly observed in many types
of cancers [1–3]. Therefore, recent systematic screenings
of mutations have focused on gene families involved in
signaling pathways, such as kinases and phosphatases in
breast and other cancers [4,5]. These efforts have
identified mutations in a variety of genes, including
PIK3CA, one of the most commonly mutated oncogenes
in human cancers [6–9]. Systematic identification of gene
mutations that are involved in signaling pathways and
associated with cancer progression and cell mobility has
been proven to be useful in finding cancer biomarkers
and therapeutic targets [1,10–12]. With the development
of automatic DNA sequencing technology, large-scale
genome sequencing projects have generated a vast
amount of DNA sequence information. Expressed se-
quence tag (EST) collections represent partial descrip-
tions of transcribed portions of genomes. So far, more
than two million high- quality ESTs from human cancer
tissues have been posted in the cancer genome anatomy
project (CGAP, http:// cgap.nci.nih.gov/) at National
Cancer Institute. Bioinformatics analysis of ESTs from
normal and cancerous tissues could identify genetic
variations associated to cancer. Single nucleotide poly-
morphisms (SNPs) are the most common genetic varia-
tions in the human genome. More and more experimental
evidence shows that some SNPs are closely linked to
cancer and treated as genotypic markers [13]. Therefore
developing a robust bioinformatics method to identify
cancer-associated SNPs and studying them in a cellular
context such as cellular signaling would help not only in
pinpointing cancer biomarkers but also in providing new

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network
29
insights into molecular mechanisms of carcinogenic and
metastatic processes.
To elucidate the underlying molecular mechanisms of
how Signaling gene mutations or overexpression act on
tumour development and metastasis, it is necessary to
dissect Signaling events that are related to the can-
cer-associated genes. Traditionally scientists treat cellular
Signaling events in view of biological pathways, study
one pathway at a time and then try to gather information
from a few pathways together to understand what is go-
ing oninside cells. However the proteins, which make up
one individual pathway, rarely operate in isolation but
‘cross-talk’ with another pathway’s proteins to process
signal information. A network-level view of Signaling
events emerges as an important concept. In this study, we
first developed a robust bioinformatics strategy to find
cancer-associated SNPs by extracting human ESTs of
normal and cancer tissues. At the same time, we manu-
ally assembled a human cellular Signaling network. We
then mapped the integrated cancer-associated genes,
which include the SNP genes we identified, known as
cancer genes (CGs) and cancer cell mobility genes
(CMGs), onto the Signaling network to study their rela-
tionships in a Signaling network context.
2. Materials and Methods
2.1 Datasets Used in This Study
Human ESTs of normal (1.89 million) and cancer (2.24
million) tissues were downloaded from NCBI dbEST
(http://www.ncbi.nlm.nih.gov/dbEST) and CGAP, resp-
ectively. As of May 2005, CGAP had 1870 and 3298
normal and cancerous EST libraries, respectively (sup-
plementary Table 1, supplementary materials are at htt-
p://www.bri. nrc.ca/wang/snp1.html). Protein and mRNA
sequences of human genome were downloaded from
ftp://ftp.ncbi.nlm. nih.gov/genomes/H_sapiens/pro- tein/
and ftp://ftp.ncbi.nlm. nih.gov/genomes/H_sapiens/RNA/,
respectively. We took tumour CMGs from a high-throu-
ghput, small RNA-interfering screening of a few cancer
cell lines including ovarian carcinoma cell line, SKOV-3
and breast cancer cell line, MDA-231 [14]. The screening
identified 532 potential tumour CMGs and a few of these
genes were further validated using other experimental
analyses such as RT-PCR, additional RNA-interfering
and cell invasion assays. We collected known CGs from
NCBI Online Mendelian Inheritance in Man database
(http://www.ncbi.nlm.nih.gov/entrez/quer-y.fcg i?db=OM
IM).
2.2 Signaling Network Construction and Net-
work Motif Detection
To construct the human cellular Signaling network, we
manually curated Signaling pathways from literature. The
Signaling data source for our pathways is the BioCarta
database (http://www.biocarta.com/genes/allpath-
ways.asp), which, so far, is the most comprehensive data-
Figure 1 Signaling network motifs for cancer-associated
genes
base for human cellular signaling pathways. Our curated
pathway database recorded gene names and functions,
cellular locations of each gene and relationships between
genes such as activation, inhibition, translocation, en-
zyme digestion, gene transcription and translation, signal
stimulation and so on. To ensure the accuracy and the
consistency of the database, each referenced pathway was
cross-checked by different researchers and finally all the
documented pathways were checked by one researcher.
In total, 164 Signaling pathways were documented (sup-
plementary Table 2).Furthermore, we merged the curated
data with another literature-mined human cellular signal-
ing network [15]. As a result, the merged network con-
tains nearly 1100 proteins (SupplementaryNetworkFile).
To construct a Signaling network, we considered rela-
tionships of proteins as links (activation or inactivation as
directed links and physical interactions in protein com-
plexes as neutral links) and proteins as nodes. To detect
and extract network motifs, we used mfinder [16]. To
obtain statistically significant inference of distributions of
the cancer-associated genes in network motifs,
re-sampling statistical procedures were used. Briefly, we
randomly assigned the same number of the can-
cer-associated genes as they are in the real network, re-
calculated the distributions of the cancerassociated genes
and compared them to the real distributions of the can-
cer-associated genes of the network. We repeated the
simulation 5000 times and then calculated P values. A
detailed description of the network re-sampling proce-
dures was described previously [17].
2.3 SNP Data Mining Strategy
To assign ESTs to human genes, we used ESTs to per-
form non-gap blast on human mRNA and protein se-
quences using BLASTN and BLASTX programs [18].
E-score cutoff was 1 10220. In each blast, the matched
ESTs to genes and proteins were obtained. If an EST has
the best match to a certain gene and also to the gene’s
coding protein, we assigned the EST to the gene. Other-
wise we discarded the EST. We picked up the ESTs that
were aligned and assigned to the genes in the network.
We observed that some sequencing errors occurred
within 100–150 bps of the end-sequence region of the
ESTs; thus, we removed 200 bps from the end-sequence

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network
30
Table 1: Enrichments of cancer-associated genes in network motifsa
Motif ID 38 204 344 394 2190 2252
CG
23.6%
(153/647)
0.57
11.3%
(170/1505)b
2 × 10-4
36.7%
(1092/2977)
2 × 10-4
26.3%
(735/2795)b
2 × 10-4
33.6%
(44/131)
0.25
26.8%
(66/246)
0.09
CMG
27.3%
(177/647)
8.5 × 10-4
46.9%
(707/1505)
2 × 10-4
33.2%
(989/2977)
2 × 10-4
35.7%
(997/2795)
2 × 10-4
35.1%
(46/131)
0.05
34.9%
(86/246)
0.01
aFor each gene type, the rates of motifs having cancer-associated genes are presented in the first row whereas the corre-
sponding P values are in the second row; bIndicates depletion rather than enrichment
Table 2: Distribution of cancer-associated genes on node positions of network motifsa
Motif ID 38 204 344 394 2190 2252
CG
P1
P2
P3
P4
33.3
29.8
36.8
–
35.3
24.4
20.3
20.0
28.6
21.2
23.4
26.8
25.0
30.0
15.0
30.0
29.7
15.8
7.9
46.5
CMG
P1
P2
P3
P4
37.1
31.0
31.9
–
33.7
22.5
19.1
24.7
25.7
24.4
23.1
26.7
17.5
31.6
22.8
28.1
24.8
30.5
31.4
13.3
aP1, P2, P3 and P4 represent node position of motifs. CG and CMG represent cancer genes and cell mobility genes, respec-
tively. The numbers represent the frequencies of CG or CMG on each node position
regions of ESTs. After cutting off 200 bps from the
end-sequence region of an EST, we scanned the EST and
its alignments to find genetic variants. We assumed that
mutations are not often clustered in a short region, so we
set a 25 bp window to avoid sequencing errors. We
defined a single mutation such that it is the only mutation
and at the middle position of a 25 bp length window. We
counted single mutations, which occurred in at least 30
libraries. To associate SNPs with cancer, we used
Fisher’s exact test for the significance of occurrence of an
SNP in cancerous and normal tissues. To control false
positives of multiple tests, false discovery rate was used.
We used the standalone pMut program [19] to test
whether the identified SNPs affect the protein’s function
and are relevant to diseases. To further support the pre-
diction, we carried out molecular modelling of the pro-
teins to visualise the locations of the mutations in the
three-dimensional structures of the proteins (see supple-
mentary modeling). Crystal structures of the proteins
were used when available; otherwise, homology models
were built. For example, histone deacetylase 2 (HDAC2)
has no crystal structure available; a homology model was
built using the available crystal structure of HDAC8 (pdb
code 1w22) as a template for the analysis (see supple-
mentary modeling). The structures were examined to see
if the mutations were expected to affect the biochemical
function of the protein. We should note that molecular
modeling is a prediction approach, which has limitations
in generating false positives.
3. Results
3.1 Mining of Cancer-Associated SNPs Using
ESTs
The availability of a large number of cancer and normal
tissue ESTs provides an opportunity for screening genetic
variations and identifying genes associated with cancer
through bioinformatics analysis. To detect SNPs, we col-
lected 2.24 million cancer tissue ESTs and 1.9 million
normal tissue ESTs. We assigned ESTs to human genes
by BLASTX and BLASTN. Because we focused on cel-
lular signaling genes, we only took the ESTs, which had
been assigned to the genes in the signaling network.We
assigned 629 signaling genes to 48 993 cancer ESTs and
723 signaling genes to 33 285 normal tissue ESTs. Both
EST pools represent almost 40 human tissues and can-
cerous ESTs, which represent most of the cancer cell
types (supplementary Table 3). Direct link of genes to
cancer could test the association between potential func-
tional variants and cancer phenotypes. This involves the
examination of non-synonymous SNPs (nsSNPs) that
result in an amino acid change. Most of the functional
variants of the genes related to diseases occur within
coding regions. We identified 44 nsSNPs in the coding
regions of 26 genes that are associated with cancer by
applying statistical analysis of SNPs in cancer and nor-
mal tissues (P , 0.05). The assumption is that can-
cer-associated SNPs are over-represented in cancerous
libraries over normal tissue libraries. To further charac-
terise putative functional variants of the identified SNPs,
we evaluated the impact of SNPs on protein structure and
function using both automatic and manual procedures. To
automatically evaluate a SNP’s effect on protein function,
we used pMut program which was developed to associate
human diseases with genetic variation by scanning sin-
gle-point amino acidic mutations. The program allows
fast pinpointing of disease-associated mutations with an
accuracy of nearly 80%. Among the 44 SNPs, we
identified 21 SNPs on 14 genes that affect protein func-
tion and link to cancer (supplementary Table 4). To fur-

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network 31
ther confirm pMut predictions, we manually examined
the SNPs by structural study of available crystal struc-
tures and generating homology models of the proteins.
For example, SNPs in HDAC2 and NFkB might cause
structural changes affecting biochemical function or pro-
tein stability (supplementary modelling).
Among the identified 14 genes which have can-
cer-associated SNPs, four of them have been found to
bear cancer-related mutations: the transmembrane protein
tyrosine kinase ERBB2, HDAC2, histone acetyltrans-
ferase (HAT) P300/CBP, the NFkB/Rel family of tran-
script factor RelA and the α subunit of the stimulatory G
protein (GaS) are related with different types of cancers.
HDACs and HATs are enzymes that catalyse the deace-
tylation and acetylation of lysine residues located in the
N-terminal tails of histones and non-histone proteins.
Emerging evidence demonstrates that perturbation of this
balance is often observed in human cancers, and inhibi-
tion of HDACs is considered to be among the most
promising novel therapeutic strategies against cancer.
The role of P300 as a tumour suppressor was first dem-
onstrated as it was identified as an adenoviral
E1A-binding protein. In breast and colon cancers, P300
expression is extremely low [20,21]. The discovery of
SNPs of these proteins in this study indicates that ex-
tracting from EST datasets is a powerful tool for finding
gene mutations in cancer cells.
3.2 Distribution of Cancer-Associated Genes in
the Network
To obtain insights into the molecular mechanisms of how
gene mutations or deregulations act on tumour develop-
ment in a cellular Signaling network context, we studied
the relationships of cancer-associated genes in a Signal-
ing network. To do so, we first manually curated human
cellular Signaling information from literature and then
merged the data with another literature-mined human
Signaling network. Most of these pathways represent
central Signaling events in cells. Therefore the network
could be seen as a general signal information centre in
cells. The network is presented as a graph with directed
and neutral links, in which, nodes represent proteins, di-
rected links represent activating and inhibitory relations
and neutral links represent only physical interactions be-
tween proteins. To study the relationships of can-
cer-as-sociated genes on the cellular Signaling network,
we first combined the known CGs and the cancer SNP
genes we identified into a set called CGs. We defined the
CGs and the 532 genome-wide RNAi screened cancer
CMGs as cancer-associated genes and then mapped these
genes onto the network. Ninety-five CGs and 87 CMGs
were mapped onto the network. We first asked if the CGs
and the CMGs are network hub proteins which have
many more links than other proteins in the network. We
ranked network proteins based on their link numbers and
then defined the hub pro- teins as the top 15% of highly
linked proteins. We found that 22% (P = 0.02) and 17%
(P = 0.23) of hub proteins are CGs and CMGs, respec-
tively. These results suggest that CGs but not CMGs are
enriched in hub proteins. Hub proteins are the function-
ally important nodes shared by many signaling pathways.
Therefore mutations or deregulations of these hub genes
may lead to cancer. To discover the distribution of can-
cer-associated genes in the network, we divided the net-
work proteins into three groups based on the cellular lo-
cation of the proteins and signal information flow:
ligand-receptor, intracellular components and nuclear
proteins. We calculated the fractions of the CGs and the
CMGs in each region. We found that downstream net-
work regions are significantly enriched with CGs (P < 2 ×
10-4 ): 7.9%, 9.2% and 18.1% in network ligand-receptor,
intracellular components and nucleus, respectively, in
contrast to 8.6%, the average rate of the CGs of the net-
work proteins. This fact suggests that CGs are more en-
riched in network downstream proteins. On the other
hand, CMGs have no significant enrichment in any re-
gion.
3.3 Regulatory Network Motifs of
Cancer-Associated Genes
Cancer-associated genes One way to study a complex
system is to break down the system into sub-systems that
are independently functional units. Biological networks
can be decomposed into statistically over-represented
subgraphs, which appear recurrently in networks and are
called network motifs [22]. A network motif is a group of
interacting components capable of signal processing and
also known as regulatory loops in biology. Network mo-
tifs have been shown to have distinct regulatory functions
and are robust to resistant internal noise. Integration of
commonly accessible data types such as protein interac-
tion, gene expression profiles and gene ortho logues onto
networks has revealed insights into network motif usages
in different cellular conditions [23–25].We have inte-
grated a dataset of genome-wide mRNA decay rates onto
gene regulatory network motifs and revealed the design
principles of gene regulatory network motifs [17]. Fur-
thermore, the integrative analysis of interactions between
microRNAs and a human cellular Signaling network re-
vealed the microRNA regulation principles of the signal-
ing network [26]. Therefore integration of cancer-assoc-
iated genes onto Signaling network motifs would help to
understand the regulatory mechanisms of how cancer-as-
sociated genes work on cancer development and metasta-
sis. To this end, we first identified all the three- and
four-node motifs in the network. We are interested in
cellular regulation of cancer-associated genes. Therefore
we only picked up the motifs in which all the links are
directed. Using this criterion, we found three- and
four-node statistically significant motifs with the follow-
ing motif IDs (mIDs): 38, 204, 344, 394, 2190 and 2252
(Figure 1). We identified all the members of each motif
type and mapped cancer-associated genes to them. We
defined a motif rate.
As the number of motifs having the CGs or the CMGs
of the motif type divided by the total number of the mo-

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network
32
tifs of that type. We found that CMGs and CGs are
significantly enriched in some particular motif types (Ta-
ble 1), suggesting that perturbation of motif genes has
more chance to lead to cancer and metastasis. Notably,
CGs are not significantly enriched in mIDs 204 and 394
motifs, suggesting that these motifs may buffer gene mu-
tations that prevent cancer development. These results
also hint that carefully studying the relationships of can-
cer-associated genes on network motifs will lead to un-
cover the regulatory mechanisms of cancer-associated
genes. Therefore we further examined the distribution of
cancer-associated genes on node positions for each motif
type (Table 2). CMGs are enriched in source nodes in
most of the motif types, whereas CGs are enriched in the
convergent nodes which are the target nodes receiving
signals from two or more source nodes in most of motif
types except the two less CG enriched motif types (Table
2). These results indicate different regulatory mecha-
nisms between cancer development and metastasis.
Therefore we inquired whether the CGs and the CMGs
share some regulatory network motifs. If a motif contains
both CGs and CMGs, we counted this motif as shared
motifs. We found that only a few shared motifs, indicat-
ing that CGs and CMGs avoid sharing motifs. This result
is consistent with our observation that CGs and CMGs
use distinct motifs and regulatory mechanisms. We fur-
ther speculated about whether some cancer-associated
genes are clustered in the network and become hotspots.
If all the nodes of a motif are the CGs or the CMGs, we
called this motif as a CG or CMG hotspot, which indi-
cates the vital role of this motif in cancer development or
metastasis. We identified 11 three-node and 9 four-node
motifs for CGs and 2 three-node motifs and 10 four-node
motifs for CMGs. Statistical analyses showed that all
these hotspots are not expected by chance ( P < 2 × 10-4 ).
These results suggest that some network regions or regu-
latory network motifs are critical to induce cancer or me-
tastasis and these genes may work together to govern cell
behaviours. These hotspots are potentially biomarker
clusters or drug target clusters for curing cancer.
4. Discussion
Cells use Signaling networks to communicate between
and within cells to control many cellular processes. Bio-
chemical Signaling events, such as phosphorylation, ace-
tylation, ubiquitylation, proteolytic cleavage and so on,
are known to have mechanisms of activating or inacti-
vating Signaling proteins. The relationships among Sig-
naling proteins are thought to determine cell behaviour;
therefore mutations or overexpression of Signaling genes
will affect Signaling relationships of proteins [1,3]. Map-
ping the cancer-associated genes onto a Signaling net-
work could uncover mechanisms of initiation, prolifera-
tion, survival, mobility and invasion of cancer cells. In
this study, we mapped the cancer-associated genes onto
the Signaling network and found that CGs are enriched in
hub proteins and cancer-associated genes are enriched or
less enriched in some particular network motifs; further-
more, CGs and CMGs are enriched in the target and
source nodes, respectively. In addition, we manually cu-
rated a human cellular signaling network, which, thus far,
is the largest constructed cellular signaling network, and
developed a strategy to extract cancer-associated SNPs
from ESTs of normal and cancer tissues.
4.1 Mining of Cancer-Associated SNPs
Genome sequence data including cancerous ESTs in-
crease as novel and cheaper DNA sequencing techniques
are rapidly developing. We developed a more robust
method to extract cancer-associated SNPs using ESTs.
Compared to other reports [27], we paid more attention
on controlling false positives and sequencing errors.We
assigned the ESTs to genes by performing BLASTX and
BLASTN to not only gene sequences but also the protein
sequences. If an EST matches both a gene and its protein
sequences, we assigned that EST to the gene. This could
reduce the chance of wrong gene assignment of ESTs.
ESTs are known as one-pass, partial sequences of cDNAs;
therefore more sequencing errors appear in the end-se-
quencing regions. To control sequencing errors, we cut
off 200 bps from the end sequencing region of ESTs;
furthermore, we defined a single mutation such that it is
the only mutation and at the middle position of a 25 bp
length window. We also used automatic (pMut program)
and protein molecular modeling techniques to examine
the potential impacts of SNPs on protein structure and
function. By doing so, we could remove almost half of
the insignificant SNPs that could not relate to cancer.
Literature validation of the identified cancer-associated
SNPs showed that almost 30% of known CG mutations
are included in our list. For example, among the can-
cer-associated SNP genes we discovered, four of them
have been found to bear cancer-related mutations:
ERBB2, HDAC2, P300/CBP and RelA. Our method hel-
ps reducing false positives; however, it also loses true
cancer-associated SNPs. Furthermore, by combining SNP
discovery, protein structural studies and molecular mod-
eling would help finding out cancer-associated SNPs.
Nevertheless, our major goal here is to find can-
cer-associated SNP genes and integrate them with other
types of data onto a signaling network.
4.2 Network Motifs of Cancer-Associated Genes
Cellular signal information flow initiates from extracel-
lular space, a ligand binds to a cellular membrane recep-
tor to start the signal, which is then transmitted by intra-
cellular Signaling components in cytosol and finally
reaches the Signaling components in the nucleus. In the
process of signal transduction, mutated genes may result
in tumourgenesis and increased cell mobility and inva-
sion. We found that CGs are enriched in hub proteins
which are the information processing centres for different
Signaling pathways. A few examples of such cancer hub
genes can be found in the network: P53, PIK3CA, Ras,
who have many regulatory partners in the network and
have potentials in integrating multiple upstream signals

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network 33
and diverge many downstream signals [28–30]. This re-
sult suggests that mutation or deregulation of hub pro-
teins in Signaling networks could leadcells to a wrong
state and promote cancer development. Furthermore, we
found that CGs are enriched in down-stream regions of
the Signaling network, especially in the nucleus. This
finding supports the notion that downstream network
components determine cell behaviour and evoke biologi-
cal responses whereas upstream network components
maintain homeostasis. Previously we showed that mi-
croRNA, a small, non-coding RNA also predominately
regulates downstream components of the human singling
network [26]. A substantial amount of microRNAs has
been reported to be associated with cancer [31]. Taken
together, one of the mechanisms of cancer development
and progression might be associated with microRNA’s
regulation of Signaling network downstream proteins.
Errors in signal transduction lead to wrong develop-
ment and behavioural decisions and sometimes result in
uncontrolled growth or cancer. Signaling gene mutation
or overexpression often results in signal transduction
errors. To understand how mutations and overexpression
of cancer-associated genes induce cancer and metastasis
in complex cellular Signaling networks, it is useful to
identify the simplest units of commonly used network
architecture. These simple units, or network motifs, such
as switches [32], gates [33], positive or negative feedback
loops [34] provide specific regulatory capacities and de-
code signal strength and process information. Both theo-
retical and experimental studies have shown that network
motifs bear particular kinetic properties that determine
the temporal program of gene expression [35]. These
motifs can be self-assembled into networks that help ex-
plaining how a complex regulatory network program is
regulated [17]. Therefore the frequencies and types of
network motifs with which cells use reveal the regulatory
strategies that are selected in different cellular conditions
[17, 36]. For example, FFLs are buffers that respond only
to persistent input signals [37] and are suited for en-
dogenous conditions, although the motifs whose key
regulator’s transcripts have fast decay rates are preferen-
tially used for exogenous conditions [17]. Therefore one
starting point in the study of cancer Signaling networks
might be to characterise how cancer-associated genes are
distributed in the regulatory network motifs of the Sig-
naling network. Our results showed that can-
cer-associated genes are enriched in some particular net-
work motif types. This fact suggests that regulatory net-
work motifs are critical for cancer development and me-
tastasis. On the other hand, we found that CGs are not
significantly enriched in two motif types, suggesting that
these motifs provide a buffer mechanism for gene muta-
tions, alternatively, suggesting that for some motif types
having only one gene mutation is not sufficient to induce
cancer. Indeed, we found that 11 and 2 three-node motifs
(hotspots) in which all nodes are CGs and CMGs, respec-
tively. We also identified nine and ten four-node motif
hotspots of CGs and CMGs, respectively. These results
suggest that some regulatory network motifs and network
regions are important for cancer and metastasis develop-
ment. The hotspots are also potentially biomarker clusters
or anticancer drug target clusters. We further examined
the frequencies of cancer-associated genes on node posi-
tions of each motif. Interestingly, we found that CGs are
enriched on the target nodes of most motifs, especially,
the convergent target nodes that receive signal informa-
tion consolidated from two or more source nodes. This
character hints that the convergent nodes of the
CG-enriched motifs are critical nodes that might be
sufficient to activate other network nodes and then induce
cancer development. In the CG-enriched motifs, source
nodes activate the same Signaling target node. It may
suggest that the source nodes could trigger the critical
nodes (the convergent target nodes) for cancer develop-
ment. Signaling networks govern homeostasis or promo-
tion of cellular state changes. In Signaling networks,
multiple information flows could be convergent to pro-
duce a limited set of phenotypic responses [38]. The
convergence provides redundant cellular functions and
robustness. Critical signal-ling nodes fall into two cate-
gories in the network: those that preserve homeostasis
during perturbation and those that evoke phenotypic
changes. Taken together, the convergent nodes in the
CG-enriched motifs could be the key regulators for pre-
serving homeostasis. Therefore perturbation of these
nodes would lead to losing cellular homeostasis and in-
ducing cancer. On the other hand, the source nodes of the
CMG-enriched motifs are the critical nodes for evoking
phenotypic changes. These data suggest that regulatory
mechanisms for cancer development and metastasis are
different.
In conclusion, we developed an approach to study the
relationships of these cancer-associated genes in a Sig-
naling network context. We found that CGs are enriched
in hub proteins, and that cancer-associated genes are
significantly enriched or depleted in some particular net-
work motif types. More importantly, we uncovered that
CGs are enriched in the convergent target nodes of most
motifs, although CMGs are enriched in the source nodes
of motifs. These results have implications for under-
standing the regulatory mechanisms of cancer develop-
ment and metastasis.
5. Acknowledgments
We thank H. Hogue for setting up NCBI BLAST on
computer cluster environment. This work is partially
supported by Genome Health Initiative, Canada. Sup-
plementary materials are accessible at http://www.bri.nr-
c.ca/wang/snp1.html.
REFERENCES
[1] Bianco, R.,Melisi, D., Ciardiello, F., and Tortora, G.:
‘Key cancer cell signal transduction pathways as thera-
peutic targets’, Eur. J. Cancer, 2006, 42, (3), pp. 290–294
[2] Hanahan, D., and Weinberg, R.A.: ‘The hallmarks of

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network
34
cancer’, Cell, 2000, 100, (1), pp. 57–70
[3] Martin, G.S.: ‘Cell Signaling and cancer’, Cancer Cell,
2003, 4, (3), pp. 167–174
[4] Bardelli, A., and Velculescu, V.E.: ‘Mutational analysis
of gene families in human cancer’, Curr. Opin. Genet.
Dev., 2005, 15, (1), pp. 5–12
[5] Stephens, P., Edkins, S., Davies, H., Greenman, C., Cox,
C., and Hunter, C.: ‘A screen of the complete protein
kinase gene family identifies diverse patterns of somatic
mutations in human breast cancer’, Nat. Genet., 2005, 37,
(6), pp. 590–592
[6] Bachman, K.E., Argani, P., Samuels, Y., Silliman, N.,
Ptak, J., and Szabo, S.: ‘The PIK3CA gene is mutated
with high frequency in human breast cancers’, Cancer
Biol. Ther., 2004, 3, (8), pp. 772–775
[7] Broderick, D.K., C, Di, Parrett, T.J., Samuels, Y.R.,
Cummins, J.M., and McLendon, R.E.: ‘Mutations of
PIK3CA in anaplastic oligodendrogliomas, high-grade as-
trocytomas, and medulloblastomas’, Cancer Res., 2004,
64, (15), pp. 5048–5050
[8] Samuels, Y., and Velculescu, V.E.: ‘Oncogenic mutations
of PIK3CAin human cancers’, Cell Cycle, 2004, 3, (10),
pp. 1221–1224
[9] Samuels, Y., Wang, Z., Bardelli, A., Silliman, N., Ptak, J.,
and Szabo, S.: ‘High frequency of mutations of the
PIK3CA gene in human cancers’, Science, 2004, 304,
(5670), p. 554
[10] Bild, A.H., Yao, G., Chang, J.T., Wang, Q., Potti, A., and
Chasse, D.: ‘Oncogenic pathway signatures in human
cancers as a guide to targeted therapies’, Nature, 2006,
439, (7074), pp. 353–357
[11] Huang, E., Ishida, S., Pittman, J., Dressman, H., Bild, A.,
and Kloos, M.: ‘Gene expression phenotypic models that
predict the activity of oncogenic pathways’, Nat. Genet.,
2003, 34, (2), pp. 226–230
[12] Downward, J.: ‘Cancer biology: signatures guide drug
choice’,Nature, 2006, 439, (7074), pp. 274–275
[13] 13 Bond, G.L., Hu, W., and Levine, A.: ‘A single nucleo-
tide polymorphism in the MDM2 gene: from a molecular
and cellular explanation to clinical effect’, Cancer Res.,
2005, 65, (13), pp. 5481–5484
[14] Collins, C.S., Hong, J., Sapinoso, L., Zhou, Y., Liu, Z.,
and Micklash, K.: ‘A small interfering RNA screen for
modulators of tumor cell motility identifies MAP4K4 as a
promigratory kinase’, Proc. Natl. Acad. Sci. USA, 2006,
103, (10), pp. 3775–3780
[15] Ma’ayan, A., Jenkins, S.L., Neves, S., Hasseldine, A.,
Grace, E., Dubin-Thaler, B., Eungdamrong, N.J., Weng,
G., Ram, P.T., Rice, J.J., Kershenbaum, A., Stolovitzky,
G.A., Blitzer, R.D., and Iyengar, R.: ‘Formation of regu-
latory patterns during signal propagation in a mammalian
cellular network’, Science, 2005, 309, pp. 1078–1083
[16] Kashtan, N., Itzkovitz, S., Milo, R., and Alon, U.:
‘Efficient sampling algorithm for estimating subgraph
concentrations and detecting network motifs’, Bioinfor-
matics, 2004, 20, (11), pp. 1746–1758
[17] Wang, E, and Purisima, E: ‘Network motifs are enriched
with transcription factors whose transcripts have short
half-lives’, Trends Genet., 2005, 21, pp. 492–495
[18] Altschul, S.F.,Madden, T.L., Schaffer, A.A., Zhang, J.,
Zhang, Z., and Miller, W.: ‘Gapped BLAST and
PSI-BLAST: a new generation of protein database search
programs’, Nucleic Acids Res., 1997, 25,
[19] , pp. 3389–3402 19 Ferrer-Costa, C., Gelpi, J.L., Zama-
kola, L., Parraga, I., de lC, X., and Orozco, M.: ‘PMUT: a
web-based tool for the annotation of pathological muta-
tions on proteins’, Bioinformatics, 2005, 21, (14), pp.
3176–3178
[20] Iyer, N.G., Ozdag, H., and Caldas, C.: ‘p300/CBP and
cancer’, Oncogene, 2004, 23, (24), pp. 4225–4231
[21] Iyer, N.G., Chin, S.F., Ozdag, H., Daigo, Y., Hu, D.E.,
and Cariati,M.: ‘p300 regulates p53-dependent apoptosis
after DNA damage in colorectal cancer cells by modula-
tion of PUMA/p21 levels’, Proc. Natl. Acad. Sci. USA,
2004, 101, (19), pp. 7386–7391
[22] Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N.,
Chklovskii, D., and Alon, U.: ‘Network motifs: simple
building blocks of complex networks’, Science, 2002, 298,
(5594), pp. 824–827
[23] Han, J.D., Bertin, N., Hao, T., Goldberg, D.S., Berriz,
G.F., Zhang, L.V., Dupuy, D., Walhout, A.J.M., Cusick,
M.E., Roth, F.P., and Vidal, M.: ‘Evidence for dynami-
cally organized modularity in the yeast protein-protein in-
teraction network’, Nature, 2004, 430, (6995), pp. 88–93
[24] Luscombe, N.M., Madan Babu, M., Yu, H., Snyder, M.,
Teichmann, S.A., and Gerstein, M.: ‘Genomic analysis of
regulatory network dynamics reveals large topological
changes’, Nature, 2004, 431, (7006), pp. 308–312
[25] 25 Zhang, L.V., King, O.D., Wong, S.L., Goldberg, D.S.,
Tong, A.H.Y., Lesage, G., Andrews, B., Bussey, H.,
Boone, C., and Roth, F.P.:‘Motifs, themes and thematic
maps of an integrated Saccharomyces cerevisiae interac-
tion network’, J. Biol., 2005, 4, (2), p. 6
[26] Cui, Q., Yu, Z., Purisima, E.O., and Wang, E.: ‘Principles
of microRNA regulation of a human cellular Signaling
network’, Mol. Syst. Biol., 2006, 2,p.46
[27] Qiu, P., Wang, L., Kostich, M., Ding, W., Simon, J.S.,
and Greene, J.R.: ‘Genome wide in silico SNP-tumor as-
sociation analysis’, BMC Cancer, 2004, 4,p.4
[28] Oikonomou, E., and Pintzas, A.: ‘Cancer genetics of spo-
radic colorectal cancer: BRAF and PI3KCA mutations,
their impact on Signaling and novel targeted therapies’,
Anticancer Res., 2006, 26, (2A), pp. 1077–1084
[29] Rodriguez-Viciana, P., Tetsu, O., Oda, K., Okada, J.,
Rauen, K., and McCormick, F.: ‘Cancer targets in the Ras
pathway’, Cold Spring Harb. Symp. Quant. Biol., 2005,
70, pp. 461–467
[30] Toledo, F., and Wahl, G.M.: ‘Regulating the p53 pathway:
in vitro hypotheses, in vivo veritas’, Nat. Rev. Cancer,
2006, 6, (12),pp. 909–923
[31] Calin, G.A., and Croce, C.M.: ‘MicroRNA-cancer con-
nection: the beginning of a new tale’, Cancer Res., 2006,
66, (15), pp. 7390–7394
[32] Bhalla, U.S., Ram, P.T., and Iyengar, R.: ‘MAP kinase
phosphatase as a locus of flexibility in a mitogen-activated
protein kinase Signaling network’, Science, 2002, 297,
(5583), pp. 1018–1023
[33] Blitzer, R.D., Connor, J.H., Brown, G.P., Wong, T.,
Shenolikar, S., and Iyengar, R.: ‘Gating of CaMKII by
cAMP-regulated proteinphosphatase activity during LTP’,
Science, 1998, 280, (5371), pp. 1940–1943
[34] Angeli, D., Ferrell, Jr. J.E., and Sontag, E.D.: ‘Detection
of multistability, bifurcations, and hysteresis in a large
class of biological positive-feedback systems’, Proc. Natl.
Acad. Sci. USA, 2004, 101, (7), pp. 1822–1827
[35] Mangan, S., Zaslaver, A., and Alon, U.: ‘The coherent feed-
forward loop serves as a sign-sensitive delay element in tran-
scriptionnetworks’, J. Mol. Biol., 2003, 334, (2), pp. 197–204
[36] Balazsi, G., Barabasi, A.L., and Oltvai, Z.N.: ‘Topologi-

Regulatory Network Motifs and Hotspots of Cancer Genes in a Mammalian Cellular Signaling Network 35
cal units ofenvironmental signal processing in the tran-
scriptional regulatorynetwork of Escherichia coli’, Proc.
Natl. Acad. Sci. USA, 2005,102, (22), pp. 7841–7846
[37] Mangan, S., and Alon, U.: ‘Structure and function of the
feed-forwardloop network motif’, Proc. Natl. Acad. Sci.
USA, 2003, 100, (21),pp. 11980–11985
[38] Prinz, A.A., Bucher, D., and Marder, E.: ‘Similar network
activityfrom disparate circuit parameters’, Nat. Neurosci.,
2004, 7, (12),pp. 1345–1352