Engineering, 2013, 5, 496-501
http://dx.doi.org/10.4236/eng.2013.510B102 Published Online October 2013 (http://www.scirp.org/journal/eng)
Expres sogram: A Visualization of Cytogenetic Lan dscape
in Cancer Samples Using Gene Expression Microarrays
Peikai Chen, Y. S. Hung
Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China
Received 2013
ABSTRACT
In cancer genomes, there are frequent copy number aberration (CNA) events, some of which are believed to be tumori-
genic. While copy numbers can be detected by a number of technologies, e.g., SNP arrays, their relations with gene
expressions are not well clarified. Here, we describe an approach to visualize the global relations between copy num-
bers and gene expressions using expression microarrays. We mapped the gene expression signals detected by microar-
ray probesets onto a reference human genome, the RefSeq, based on their annotated physical positions, resulting in a
landscape that we called expressogram. To study the expressograms under various conditions and their relations with
cytogenetic events, such as CNAs, we obtained three classes of array samples, namely samples of a cancer (e.g., liver
cancer), normal samples in the same tissue, and normal samples of other tissues. We developed a Bayesian based algo-
rithm to estimate a background signal from the latter two sources for the cancer samples. By subtracting the estimated
background from the raw signals of the cancer samples, and subjecting the differences to a kernel-based smoothing
scheme, we produced an expressogram that shows strong consistency with the copy numbers. This indicates that copy
numbers are on average positively correlated with and have strong impacts on gene expressions. To further explore the
applicability of these findings, we submit the expressograms to the significant CNA detection algorithm GISTIC. The
results strongly indicate that expressogram can also be used to infer copy number events and significant regions of CNA
affected dysregulation.
Keywords: Microarrays; Cytogenetics; Cancer Landscape; Copy Number Aberrations
1. Introduction
The copy numbers of genes in normal somatic chromo-
somes are assumed to be two, i.e., one copy from father
and the other from mother. But in cancer tissues, regions
of the genome can experience copy amplifications or
deletions, called copy number aberrations (CNAs). CNAs
are very frequent in cancer genomes [1] and with th e aid
of recent development in biotechnologies, there are mas-
sive efforts to generate measurements of CNAs in vari-
ous cancers [2,3]. However, there have not been uniform
theories on the cause of CNAs, and their relations with
gene expressions and cancer, although there is a general
speculation that some of these CNAs may have initiating
or driving roles in the formation and development of
cancer [4]. Algorithms such as GISTIC [5] have been
developed to identify CNA regions that potentially har-
bor such events. An immediate question following such
efforts then is, if some of the CNAs are cancer causing
events, what are the remaining CNAs? This question is
important because if the remaining CNAs can be con-
firmed to be either mere consequences or by-products,
the role of the cancer-causing CNAs can be further es-
tablished.
The answer to this question may be found by a general
inspection on the relations between copy numbers and
their immediate effects, the gene expressions, which can
now be readily measured by a plethora of methods, in-
cluding gene expression microarrays. And while indi-
vidual copy number changes may cause a gene to be ei-
ther up or down regulated [6], some studies [7] also sug-
gest that copy numbers do positively affect gene expres-
sions. If the latter holds in the general settings, it means
that we may be able to visualize the gene expression
landscape, or as we called it, the expressogram, of a sam-
ple or a group of samples, with respect to their cytoge-
netic profiles, i.e., the genome-wide copy number mea-
surements.
This visualization is necessary for several reasons.
First, it may clarify the copy number-expression rela-
tion simultaneously across chromosomes and across dif-
ferent samples. Particularly, when comparing the ex-
pressogram with the copy number landscapes by other

P. K. CHEN, Y. S. HUNG
497
means of measurements, such as SNP arrays, it can re-
veal how copy numbers are generally affecting gene ex-
pressions.
Second, it may help pinpoint regions of disease-spe-
cific dysregulation. CNAs typically harbor tens, hun-
dreds, or even thousands of genes, all of which have
uniform copy number states. But the impact on individu-
al genes, and regions may be different. For example,
some genes may be physically adjacent and share some
regulating mechanisms [8], as a result of which, these
genes tend to show region specific co-regulations. These
co-regulations by CNAs are oft e n important in cancer.
Third, conventional CNA inferences are mostly based
on array CGH, SNP arrays, etc., but some of them suffer
from errors [9]. This visualization technique may serve
as an independent source of measurements to help con-
firm that certain regions are real CNAs.
Fourth, instead of relying on SNP arrays for detection
of recurrent CNAs for search of potential cancer causing
events, the expressogram signals may be used to search
for genes that are directly and recurrently affected by
copy numbers. These genes may be more directly related
to the cancer process than those candidates uncovered by
SNP arrays.
Toward this end, we propose an approach to visualize
gene expression landscapes, i.e., expressograms, in can-
cers using gene expression microarrays. The following
sections discuss the algorithms and results of this ap-
proach.
2. Algorithm
A direct approach to visualize the expressions in the
chromosome positions is to plot the signals against the
cytogenetic positions, such as in the work by [10], which
may be subject to huge noises and biases. Here, we use a
two-stage approach. First, a background landscape is
estimated. Second, the estimated background is sub-
tracted from the diseased samples under study, before the
difference signals are subjected to a smooth ing filter .
2.1. Background Estimation
Three gene expression datasets are obtained, n amely, the
dataset of samples under study, often from a disease (e.g.,
certain cancers), denoted as

; the dataset of normal
samples in the same tissue as the studying disease, de-
noted as
n

; and the dataset of samples in other normal
tissues that will be used as the prior information for the
background, denoted as
p

.
For a probeset
{1,..,}jJ∈
, where
J
is the total
number of probesets, the objective is to derive a probe-
set-specific background signal
j
s
from
n

and
p

.
The reason for using both
n

and
p

is that very
often, the
n

dataset is small and not really normal
because these normal references are often tissues donated
by patients dying of other reasons, or patients having
other conditions in the s ame tissue. As a result, they may
not truly reflect the normal conditions in the studying
tissue in the population. Also, most of the genes are sup-
posed to be tissue-non-specific, i.e., their expressions are
not tissue-dependent. Therefore, expressions of a gene
(or probeset) from other tissues may be used as a prior
information for a Bayesian inference of the true back-
ground signal
j
s
. Specifically, let the mean signal from
n

be
j
s
, the Bayesian estimate of the true signal
j
s
is given by :
(|) ()(|)/()
jjj jjj
PssPs PssPs=
(2.1)
where
()
j
Ps
is a constant. Assuming a Gaussian dis-
tributions for
()
j
Ps
, the maximum a posteriori (MAP)
[11] estimation of
j
s
,
ˆ
j
s
, is given by:
2
2
(/)
ˆ1 (/)
j jjj
jjj
Ns
sN
µ σσ
σσ
+
=+


(2.2)
where
N
is the number of samples in
n

,
j
µ
is the
mean of probeset
j
in
p

, and
j
σ

is the standard
deviati o n of p r obeset
j
in
n

and
j
σ
that in
p

.
From Equation (2.2), it can be seen that if
N
and/or
j
σ
are small, the estimate
ˆj
s
will largely depend on
the population estimate, i.e.,
j
µ
. This is favorable be-
cause often the normal samples are noisy and have small
sample sizes. Hopefully, this will produce a more stable
estimate of the background signals.
2.2. Subtracting Background and Smoothing the
Signals
Suppose there are
n
samples in

. Given a sample
{1,..,}in∈
, and the expression for the
j
-th probeset in
i
,
,ij
e
, the difference sign al with bac kground sub trac ted
is given by:
,,
ˆ
ij ij j
fes= −
. Assuming that the probesets
were pre-arranged in such an order that
,
ij
f and
,1ij
f
+
represent the signals of two physically adjacent probesets
in the human genome (e.g., NCBI RefSeq Build 37.1),
then the series
,
{}
ij
f
represents the gene expression
landsca pe of
i
, referred to as the expressogram.
A major issue in visualizing
,
{}
ij
f
is that on top of
cytogenetic factors, it is also regulated by a large number
of other unknown factors, producing tremendous high-
frequency noises across the chromosomal positions. Fur-
ther, the background estimate from previous step is non-
samp l e-specific (i.e., all disease samples use the same
background signals), which may produce some bias for
the studying samples. To reduce these effects, a kernel-
based smoothing scheme using the Nadaraya-Watson
algorithm [12] is adopted. Specifically, given a cytoge-
netic position
x
, the smoothed difference signal
x
f

is
given by:

P. K. CHEN, Y. S. HUNG
498
()
()
hj j
j
xhj
j
K Xxf
fKX x
∈
∈
−
=−
∑
∑


(2.3)
where h
K
is a kernel function with window width
h
,

is the set of points in the window and
j
X
is the
chromosomal position for
j
f
. Wh en the Gau ss i a n k er n el
is used,
h
is the variance parameter
σ
. Equation (2.3)
acts as a low-pass 1-D spatial filter and the resulting sig -
nal
,
{}
ij
f

represents a location-dependent signal that
may also reflect the impact of cytogenetic factors on ex-
pressions.
3. Results
To test the proposed visualization method, we applied it
to microarray measurements of a cancer, hepatocellular
carcinoma (HCC), i.e., liver cancer. The

dataset con-
sists of 90 samples by Chiang et al. [13] (GE O accession
number: GSE9829), and the
n

dataset consists of 58
normal liver microarray expressions collected from six
studies (GEO accession numbers: GSE7117, GSE14951,
GSE19665, GSE23343, GSE29722 and GSE14668). To
construct the dataset
p

for estimating the prior values
of the probesets, we select a number of tissues, including
normal colorectal (GSE9254), pancreatic (GSE22780),
thyroid (GSE3678), ovarian (GSE14407), endometrial
(GSE7305), breast (GSE30010), skin (GSE14905) and
esophageal (GSE26886) tissues from the controls of dis-
ease studies, or from the samples of non-disease studies.
Most of these selected tissues are made up of epithelial
cells and share some common attributes with hepatocytes
(liver cells), such as fast proliferation rates. Finally, 100
samples were collected for
p

. All samples in the th ree
datasets are from a single microarray platform, i.e., the
Human Genome U133 Plus 2.0 array (Affymetrix, CA),
and are pre-processed with the RMA algorithm [14].
Another 90 SNP arrays matching with the 90 expres-
sions arrays in

were also downloaded from GSE9829
and preprocessed with the Affymetrix CNAT algorithm
[15]. Figure 1A shows the copy number landscape of
these 90 samples.
The background signal estimation and smoothing were
conducted as described in the previous section. To see
that the two-step approach does result in better distinc-
tion between the expression probesets with CNAs and
those without, we use the SNP array inferred copy num-
bers as ground truth, and assigned each expression pro-
beset a copy number state based on the measurement of
its closest SNP probeset. The three solid-line distribu-
tions (from left, in blue, black and red) in Figure 2 re-
present the difference signals of genes having copy
number losses, normal (i.e., two) and gains, respectively.
It can be seen that there is a clear positive correlation
between the SNP-inferred copy number states and the
gene expressions. The background signals used in these
curves are based on
{}
j
s
, i.e., without Bayesian updates.
The three dashed-line curves in the same colors corres-
pond to the difference signals undergoing the Bayesian
procedure using Equation (2.2). It is very clear that the
Bayesian step greatly increases the contrast among dif-
ference signals with different copy n umber st a t e s .
We then used the difference signals obtained by Equa-
tion (2.3) to produce an expressogram, i.e., a visualiza-
tion of the landscape of gene expression by heatmap
tools. Figure 1 B shows the result. It can be seen that
there is a strong consistency between the copy number
landscapes and the expressogram in Figure 1.
Next, we submitted the difference signals to GISTIC
[5] (provided by genepattern.broadinstitute.org), a tool
used in SNP arrays to identify recurrent CNAs, to find
the recurrent up- or down-regulations. Figure 3 shows
the result. Comparing the results in Figure s 3A and B, it
can be seen that most of the features are very similar.
This suggests that the expressogram can be used to pin-
point regions of recurrent dysregulations that are caused
by CNAs.
4. Conclusions
In summary, we have described a novel visualization of
gene expressions in the cancer genomes. We make use of
extensive information, both from samples of normal tis-
sue under study and those from other normal human tis-
sues to predict the background signal before it is sub-
tracted from the raw signal. The resulting difference signals
are subjected to Kernel-based smoothing. The expresso-
gram provides a wonderful visualization of gene expres-
sion across the genome. This study is meaningful in sev-
eral aspects.
First, the expressogram clearly corresponds to the cy-
togenetic changes, e.g., CNAs. This indicates that copy
numbers of most genes do affect their expressions. But
the effect is marginal, i.e., it becomes obvious only after
the background signals are subtracted. And how some of
these effects go on to produce cancer-driving conse-
quences is yet to be determined.
Second, the recurrent regions of gene expressions ar-
rays are highly consistent with that from SNP arrays.
This suggests that in cases where SNP arrays are not
available, our method provides an alternative to generate
the GISTIC landscape for identification of recurrent
CNAs. Particularly, this method has advantage over the
sample landscape by SNP arrays, as it directly shows the
recurrence at the expression level, which is believed to
have more biological importance.
5. Acknowledgements
The work described in this paper is partially supported by

P. K. CHEN, Y. S. HUNG
499
Figure 1. The copy number landscape of hepatocellular carcinoma (HCC) by SNP arrays. (B) Expressogram of the same
samples by our algorithm using gene expression microarrays. Color codes for (A): red, copy number gains; blue, copy num-
ber losses. Color codes for (B): red, up-regulation; black or dark green, neutral; light green, down-regulation. In both (A)
and (B), the horizontal axes represent the chromosomal positions and the vertical dashed lines represent chromosomal
boundaries. Each row in both plots represents an HCC sample. (C) The raw difference signals (dots) and smoothed signals
(green) of a specific example, B25. Also shown is the copy number profile of B25 by SNP array. Note the effects of CNAs on
the raw and smoothed signals.

P. K. CHEN, Y. S. HUNG
500
Figure 2. The distribution of difference signals
{ }
,

ij
f
with respect to different copy number states as measured by SNP ar-
rays. Blue curves are the distributions of difference sig- nals with copy number losses. Black curves are the distributions of
those with copy number neutral (=2), and Red curves copy number gains. The three solid-line distributions are based on
background signals from the estimate of
n

, while the dashed lines represent those based on background signals updated
with prior val ues.
Figure 3. The GISTIC land scapes (A) using our difference signals, and (B) using SNP arrays for the hepatocellular carcino-
ma samples. The horizontal axes are the chromosomes while the vertical axes are the log10q-values. The four green lines are
the qv significance thresholds at 0.25. The blue peaks correspond to significant down-regulations in (A) and copy number
losses in (B), while the red ones correspond to significant up-regulations in (A) and copy number gains in (B).

P. K. CHEN, Y. S. HUNG
501
the Hong Kong SAR RGC GRF (Project No HKU_
762111M) and CRCG of the University of Hong Kong.
REFERENCES
[1] M. Baudis, “Genomic Imbalances in 5918 Malignant Epi-
thelial Tumors: An Explorative Meta-Analysis of Chro-
mosomal CGH data,” BMC Cancer, Vol. 7, 2007, p. 226.
http://dx.doi.org/10.1186/1471-2407-7-226
[2] J. Li, K. Wang, et al., “DNA Copy Number Aberrations
in Breast Cancer by Array Comparative Genomic Hybri-
dization,” Genomics Proteomics Bioinformatics, Vol. 7,
No. 1-2, 2009, pp. 13-24.
http://dx.doi.org/10.1016/S1672-0229(08)60029-7
[3] F. Rapaport and C. Leslie, “Determining Frequent Pat-
terns of Copy Number Alterations in Cancer,” PLoS One,
Vol. 5, No. 8, 2010, p. e12028.
http://dx.doi.org/10.1371/journal.pone.0012028
[4] M. R. Stratton, P. J. Campbell and P. A. Futreal, “The
Cancer Genome,” Nature, Vol. 458, No. 7239, 2009, pp.
719-724. http://dx.doi.org/10.1038/nature07943
[5] R. Beroukhim, G. Getz, et al., “Assessing the Signifi-
cance of Chromosomal Aberrations in Cancer: Metho-
dology and Application to Glioma,” Proceedings of the
National Academy of Sciences of the United States of
America, Vol. 104, No. 50, 2007, pp. 20007-20012 .
[6] C. N. Henrichsen, E. Chaignat and A. Reymond, “Copy
Number Variants, Diseases and Gene Expression,” Hu-
man Molecular Genetics, Vol. 18, No. R1, 2009, pp. R1-
R8. http://dx.doi.org/10.1093/hmg/ddp011
[7] M. Kool, J. Koster, e t al ., “Integrated Genomics Identifies
Five Medulloblastoma Subtypes with Distinct Genetic
Profiles, Pathway Signatures and Clinicopathological
Features,” PLoS One, Vol. 3, No. 8, 2008, p. e3088.
http://dx.doi.org/10.1371/journal.pone.0003088
[8] P. Michalak, “Coexpression, Coregulation, and Cofunc-
tionality of Neighboring Genes in Eukaryotic Genomes,”
Genomics, Vol. 91, No. 3, 2008, pp. 243-248.
http://dx.doi.org/10.1016/j.ygeno.2007.11.002
[9] S. Colella, C. Yau, et al., “QuantiSNP: An Objective Bayes
Hidden-Markov Model to Detect and Accurately Map
Copy Number Variation Using SNP Genotyping Data,”
Nucleic Acids Research, Vol. 35, No. 6, 2007, pp. 2013-
2025. http://dx.doi.org/10.1093/nar/gkm076
[10] M. A. Sanders, R. G. Verhaak, et al., “SNPexpress: Inte-
grated Visualization of Genome-Wide Genotypes, Copy
Numbers and Gene Expression Levels,” BMC Genomics,
Vol. 9, 2008, p. 41.
http://dx.doi.org/10.1186/1471-2164-9-41
[11] S. Theodoridis and K. Koutroumbas, “Pattern Recogni-
tion,” 3rd Edition, Academic Press , San Diego, 2006.
[12] M. G. Schimek, “Smoothing and Regression: Approaches,
Computation, and Application,” Wiley Series in Probabil-
ity and Statistics Applied Probability and Statistics Sec-
tion, Wiley, New York, 2000.
http://dx.doi.org/10.1002/9781118150658
[13] D. Y. Chiang, A. Villanueva, et al., “Focal Gains of
VEGFA and Molecular Classification of Hepatocellular
Carcinoma,” Cancer Research, Vol. 68, No. 16, 2008, pp.
6779-6788.
http://dx.doi.org/10.1158/0008-5472.CAN-08-0742
[14] R. A. Irizarry, B. Hobbs, e t al., “Exploration, Normaliza-
tion, and Summaries of High Density Oligonucleotide
Array Probe Level Data,” Biostatistics, Vol. 4, No. 2,
2003, pp. 249-264.
http://dx.doi.org/10.1093/biostatistics/4.2.249
[15] “CNAT4.0: Copy Numbers and Loss of Heterozygosity
Estimation Algorithms for the Genechip Human Mapping
10/50/100/250/500k Array Set,” Affymetrix Inc., Tech.
Rep., 2007.