Tissue-specific gene expression templates for accurate molecular characterization of the normal physiological states of multiple human tissues with implication in development and cancer studies.

Abstract

To elucidate the molecular complications in many complex diseases, we argue for the priority to construct a model representing the normal physiological state of a cell/tissue.
By analyzing three independent microarray datasets on normal human tissues, we established a quantitative molecular model GET, which consists of 24 tissue-specific Gene Expression Templates constructed from a set of 56 genes, for predicting 24 distinct tissue types under disease-free condition. 99.2% correctness was reached when a large-scale validation was performed on 61 new datasets to test the tissue-prediction power of GET. Network analysis based on molecular interactions suggests a potential role of these 56 genes in tissue differentiation and carcinogenesis.Applying GET to transcriptomic datasets produced from tissue development studies the results correlated well with developmental stages. Cancerous tissues and cell lines yielded significantly lower correlation with GET than the normal tissues. GET distinguished melanoma from normal skin tissue or benign skin tumor with 96% sensitivity and 89% specificity.
These results strongly suggest that a normal tissue or cell may uphold its normal functioning and morphology by maintaining specific chemical stoichiometry among genes. The state of stoichiometry can be depicted by a compact set of representative genes such as the 56 genes obtained here. A significant deviation from normal stoichiometry may result in malfunction or abnormal growth of the cells.

Full text

Page 1

RESEARCH ARTICLEOpen Access
Tissue-specific gene expression templates for
accurate molecular characterization of the normal
physiological states of multiple human tissues
with implication in development and cancer
studies
Pei-Ing Hwang1, Huan-Bin Wu1, Chin-Di Wang1, Bai-Ling Lin2, Cheng-Tao Chen1, Shinsheng Yuan1, Guani Wu1and
Ker-Chau Li1,2,3*
Abstract
Background: To elucidate the molecular complications in many complex diseases, we argue for the priority to
construct a model representing the normal physiological state of a cell/tissue.
Results: By analyzing three independent microarray datasets on normal human tissues, we established a
quantitative molecular model GET, which consists of 24 tissue-specific Gene Expression Templates constructed from
a set of 56 genes, for predicting 24 distinct tissue types under disease-free condition. 99.2% correctness was
reached when a large-scale validation was performed on 61 new datasets to test the tissue-prediction power of
GET. Network analysis based on molecular interactions suggests a potential role of these 56 genes in tissue
differentiation and carcinogenesis.
Applying GET to transcriptomic datasets produced from tissue development studies the results correlated well with
developmental stages. Cancerous tissues and cell lines yielded significantly lower correlation with GET than the
normal tissues. GET distinguished melanoma from normal skin tissue or benign skin tumor with 96% sensitivity and
89% specificity.
Conclusions: These results strongly suggest that a normal tissue or cell may uphold its normal functioning and
morphology by maintaining specific chemical stoichiometry among genes. The state of stoichiometry can be
depicted by a compact set of representative genes such as the 56 genes obtained here. A significant deviation
from normal stoichiometry may result in malfunction or abnormal growth of the cells.
Background
It has been well-recognized that within a cell, not only
genes participate in cascades of biochemical events
(pathways), but also the pathways themselves cross-talk
with each other as a delicate and intriguing network sys-
tem. Such complexity was reflected in the normal biolo-
gical processes (tissue development, for example) as well
as in the complex disease processes such as autism,
cancer, rheumatoid arthritis and coronary artery disease
[1,2]. In addition, the genetic interactions of oncogenes
and tumor suppressor genes may perturb the normal
network system through a variety of altered molecular
properties of the normal genes, magnifying the difficul-
ties encountered in the cancer biology study [3]. Because
of this, it is important to develop quantitative molecular
models which can represent different physiological or
pathological states of a complex biological system and
can be used to predict the related states, using high
throughput molecular data. In line with this viewpoint,
we argue for the priority to construct models for the
* Correspondence: kcli@stat.sinica.edu.tw
1Institute of Statistical Science, Academia Sinica, Taipei, Taiwan 115, Republic
of China
Full list of author information is available at the end of the article
Hwang et al. BMC Genomics 2011, 12:439
http://www.biomedcentral.com/1471-2164/12/439
© 2011 Hwang et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

Page 2

normal physiological states first. This is because (1) nor-
mal cells/tissues are endowed with the most stable bio-
chemical homeostasis and (2) such models may serve as
general references for contrasting with various patholo-
gical or altered physiological states.
Up to now, in part due to the limitation in sample
availability, few studies on normal human tissues have
been reported. Through the transcriptome study of the
disease-free human samples via microarray analysis,
gross patterns of tissue-gene relationships have been
observed by several teams [4-7]. A recent study which
applied statistical and network analysis to transcriptomic
data from 31 normal human tissue types has resulted in
putative tissue-specific networks for nine tissues. These
putative tissue-specific networks were suggested as
potential drug targets [8]. However, it still awaits a dee-
per investigation to find out what molecular signatures
can best represent the normal state of a specific tissue
and offer the most transparent and systematic elucida-
tion on tissue differences (regarding anatomy, pathology
and development). In this study, by re-analyzing some
of the transcriptomic datasets produced from normal
human tissues in the Gene Expression Omnibus (GEO),
we identified a set of 56 genes whose transcript profiles
are endowed with strong tissue-specific properties for
24 different tissue types under the disease-free condi-
tion. These genes present significant variation of expres-
sion amongst tissues. From the expression level of these
56 genes, we constructed 24 tissue-specific Gene Expres-
sion Templates (GETs), one for each of the 24 tissues.
We first validated that these GETs can differentiate tis-
sue types under the normal physiological condition.
Then we demonstrated how GET can be applied under
other conditions, including development and cancerous
conditions. Our results suggest that homeostasis among
various molecules in a cell/tissue may play a key role in
maintaining its normal functioning and the homeostasis
state can be characterized by the 56 genes.
Results
Characterization of 24 tissue types by the 56 genes
We searched for a set of genes whose expression profile
could best represent normal state of a specific tissue
type. We used three large-scale microarray datasets as
our training datasets to identify a group of 56 genes
with high variation in expression across different tissue
types (Additional file 1: Table S1). Briefly, we selected
the probe sets with coefficient of variation (CV) ranked
within the top 2.5% of the entire transcriptome across
all samples from each of the three training datasets.
After intersecting the three groups of highly variably
expressed probe sets, we removed redundant probe sets
that share similar expression patterns. Our procedure
yielded a set of 56 genes. [see methods for more details].
Function enrichment analysis, based on an ontology
search against the Panther database, showed that these
56 genes were most enriched in encoding cytoskeletal
proteins, calmodulin-related proteins, and neuropeptides
(P value < 1 × 10-5). Gene Ontology search indicated that
the proteins encoded by the 56 genes were mostly loca-
lized to extracellular regions (21/56, P value < 3.1 × 10-5)
and were involved in the processes of response to
wounding (10/56, P value = 6.12 × 10-5), response to
steroid hormone stimulus (6/56, P value = 4.67 × 10-4)
and regulation of homeostatic process (5/56, P value =
5.94 × 10-4). While searching KEGG for the pathways
mapped by the 56 genes, we found focal adhesion, ECM-
receptor interaction, and PPAR signaling pathway over-
represented. (Additional file 1: Table S2) We conducted
hierarchical clustering with the 56 genes from the 24 tis-
sue types shared by our three training datasets. The result
showed that all of the 24 tissues were well grouped
(i.e. the same tissue from three different data sources
were grouped together as a cluster; see the dendrogram
in Figure 1A and the heat map in Additional file 1: Fig.
S1). To confirm that the tissue-clustering result is specific
to the gene set we identified, we randomly-selected 56
probe sets and applied the same hierarchical clustering
analysis to their expression profiles. We were no longer
able to find the tissue clusters; instead, the only visible
clusters were three large classes corresponding to tissues
from each data source. (Figure 1B; also see Additional file
1: Fig. S2 for heat map) In addition, as suggested by one
anonymous reviewer, we also selected the probes with
the highest expression level across all tissues as a contrast
group. This contrast group of genes (including many
house-keeping genes) failed to group tissues properly.
(see Additional file 1: Fig. S3)
Taking a closer look at the outcome of the tissue
classification obtained from the hierarchical clustering
analysis, similar to the previous reports published by
Shyamsundar et al [6] and Ge et al [3] on tissue classi-
fication, our results also showed that the tissues were
grouped largely by their physiological functions, anato-
mical locations or cellular composition. For example,
ovary and uterus, the organs from the female produc-
tive system, were clustered together; the gland tissues/
organs including thyroid, pancreas, and salivary gland
also aggregated and the testis as previously reported
[4,7,9] was grouped with the tissues from the central
nervous system like cerebellum, amygdala, and thala-
mus. However, it should be noted that we obtained the
similar tissue classification result with far smaller set of
genes (56) as the classifiers, in comparison with the
two previous reports where they used 7396 [3] genes
and 5592 [6] genes on 36 tissue types (from 36 sam-
ples) and 45 tissue types (from 115 specimens),
respectively.
Hwang et al. BMC Genomics 2011, 12:439
http://www.biomedcentral.com/1471-2164/12/439
Page 2 of 13

Page 3

Accurate prediction of normal tissues
We speculated that the expression profile of this com-
pact 56-gene set might represent a set of important
molecular features for tissue specification under normal
conditions. To determine whether this gene set could
accurately predict the tissue type of a sample of
unknown origin, the expression profiles of the 56 genes
from the 24 tissues were extracted from our three
microarray datasets to form the tissue-specific gene
expression templates (GET) for tissue prediction. More
Figure 1 Dendrograms showing tissue classification by hierarchical clustering analysis with the 56 genes. The dendrograms shown were
generated with standard two-way hierarchical clustering analysis of GenePattern using the expression values of the (A) 56 signature genes or (B)
randomly selected 56 genes extracted from the 24 tissues under the three GEO datasets. The associated heatmaps are displayed in Additional
file 1: Fig. S1 and Fig. S2, respectively.
Hwang et al. BMC Genomics 2011, 12:439
http://www.biomedcentral.com/1471-2164/12/439
Page 3 of 13

Page 4

precisely, for each of the 24 tissues, we obtained a GET
which consisted of the average of the probe values of
the three datasets for the 56 selected genes. To predict
the tissue type of a new sample, we compared the
expression pattern of the 56 genes in the new sample
with each of the 24 GETs by calculating their correla-
tion coefficients (c.f.). The tissue type of the GET show-
ing the highest degree of similarity as measured by c.f. is
our prediction. As an example, we took a new dataset
from Gene Expression Omnibus, GSE5364[10], which
contained data derived from human liver, lung, and
thyroid, in non-malignant and malignant conditions.
Tissue prediction was performed by calculating the cor-
relation between the expressions of the 56 genes in each
test tissue against each of our 24 GETs. For the non-
malignant tissues, we obtained 100% correct prediction,
strongly endorsing our speculation that the expression
profiles of the 56 genes may be used as molecular signa-
tures for normal human tissues. (Additional file 1: Table
S3) To further validate the use of these tissue-specific
GETs as predictors of normal human tissues, an exten-
sive tissue-prediction analysis was carried out on a total
of 61 microarray datasets consisting of 797 samples
from 16 different human tissues. These datasets were
obtained by searching the entire GEO database using
two criteria: (1) the tissue types must fall into the
category of the claimed 24 distinguishable normal
human tissues (2) they were hybridized on one of two
Affymetrix GeneChips, HG-U133A (GPL96) or HG-
U133plus2.0 (GPL570). Cell lines or specific cell types
from an organ were excluded (see Methods for details).
As it turned out, the overall prediction accuracy of our
method already reached 99.2% (791/797). (Table 1 and
Additional file 1: Table S4). This large-scale validation
demonstrated that the tissue-specific profiles of the 56
genes are essential for the 24 human tissue types. It also
suggests that the tissue-specific GETs very likely repre-
sent the overall biochemical equilibrium reached by all
the molecules of a tissue under normal physiological
condition.
To show how stable and robust the relative expression
level of the 56 genes is, we used Spearman’s rank corre-
lation to replace the standard Pearson correlation in the
tissue-prediction analysis and obtained 96.2% (767/797)
of accuracy (Additional file 1: Table S5). It indicates
that the relative expression of the 56 genes is a robust
feature for characterizing the normal state of a specific
human tissue.
So far, we have chosen Affymetrix as the gene expres-
sion platform to test the performance of GET. Our rea-
sons are (1) Affymetrix is the common platform used in
our three training datasets which ensures the inclusion
of the entire set of our 56 genes; (2) the data preproces-
sing procedure is standardized; (3) existence of many
datasets in the public domain for validation and (4) high
reproducibility. However, it remains questionable
whether GET can be applied to other platforms of mea-
suring gene expression or not. To address this issue, we
searched GEO for datasets containing enough samples
from norm human tissues. We found an ABI array gen-
erated dataset, GSE7905 [8], which contains many nor-
mal human tissue samples. Among them, there are 60
samples (20 tissue types in triplicates) matching our 24
tissue types. We treated the data generated by these ABI
array as if they were from the Affymetrix and applied
GET to make tissue prediction. Strikingly, we found that
our Affymetrix-based GETs yield a perfect result, 100%
(60/60) (Additional file 1: Table S6). This demonstrates
that GET is platform-independent.
Network analysis reveals involvement of the 56 genes in
development and tumorigenesis
The above results indicate that these tissue-specific
GETs may be more than simply a set of biomarkers cap-
able of distinguishing different tissue types. Their pro-
files may represent a “net sum” of the interplays among
the complex gene regulation pathways occurring in a
tissue. We explored the possible biological roles of the
56 genes by performing network analysis using the com-
mercial tool MetaCore from GeneGo Inc. which builds
Table 1 Large-scale prediction of normal human tissues
with GETs
Tissue Datasets Samples Correctly
predicted
%
Accuracy
bone marrow1 1111 100%
fetal liver166 100%
heart3 2828 100%
kidney11108 10699%
liver9 84 84100%
lung11170 16999%
ovary144100%
pancrease2 1714 82%
pituitary gland 111100%
placenta43434100%
prostate310 10100%
skeletal
muscle
8134 134 100%
skin 10 121121 100%
testis166 100%
thyroid436 36 100%
Uterus127 27 100%
Total 6179779199.25%
Samples: number of tested samples derived from normal human tissues.
Correctly Predicted: the number of tested samples for which the tissue source
was correctly predicted. The source data as well as the analysis results from
each dataset can be found at the Additional file 1: Table S4.
Hwang et al. BMC Genomics 2011, 12:439
http://www.biomedcentral.com/1471-2164/12/439
Page 4 of 13

Page 5

gene networks based on molecular interactions acquired
from experiments-based literature reports [11]. We
firstly applied the basic algorithm “analyze networks” to
our 56 genes using the default parameter settings. The
top scored results (Figure 2) showed MMP9 (matrix
metallopeptidase 9), STAT3 (signal transducer and acti-
vator of transcription 3), and PPARG (peroxisome pro-
liferator-activated receptor gamma) to be the most
connected molecules in the network (z Score = 65) with
15 of our genes included. MMP9 has been known to be
the key regulator for bone remodeling [12,13], STAT3 is
involved in embryogenesis, hematopoietic cell develop-
ment and is a biomarker for embryo stem cells[14-16],
and PPARG is involved in the regulation of neural stem
cells proliferation and differentiation [17,18]. We then
applied the algorithm “Receptor Targets Modeling”
which allows users to identify the important transcrip-
tion factors (TFs) connected to the query genes and the
signaling receptors associated with these TFs”. The best
result produced by this algorithm delivered a network
(Z score = 131) (Figure 3) with half (28) of our genes
connected to the TFs: STAT3, ESR, SRF, CEBPB, E2F1,
PPARG and TP53 and these transcription factors were
regulated by EGFR. All of these molecules are known to
be important in development and/or tumorigenesis. For
example, ESR (estrogen receptor), a ligand-activated tran-
scription factor, is essential for sexual development and
reproductive function, and is involved in breast cancer,
endometrial cancer, and osteoporosis. The signature
genes linked to ESR include ABAT, CD24, GJA, KRT13,
MSMB, PCP4 TF and THBS, among which CD24 and
THBS were responsive to hypoxia, and CD24, GJA,
KRT13 were involved in development of immune system,
neuron projection and ectoderm, respectively. SRF (c-fos
serum response element-binding transcription factor), the
transcription factor which regulates the activity of many
immediate-early genes, such as c-fos, has been known to
participate in numerous significant processes like cell
cycle regulation, apoptosis, and cell differentiation. It is
the downstream target of many pathways including the
mitogen-activated protein kinase pathway (MAPK) and is
implicated in the hepatoma progression [19]. Among the
four signature genes linked to SRF, Gro2 is a chemokine
and an oncoprotein, Desmin an intermediate filament
involved in muscle contraction and cytoskeleton organi-
zation, DLK a transmembrane protein is implicated in
Figure 2 Network analysis on the 56 genes using MetaCore. This network is the highest scored one when constructed on the basic
algorithm “analyze networks” of MetaCore from GeneGo Inc. using the 56 genes as input. The “analyze networks” algorithm of MetaCore builds
on the Dijkstra’s algorithm which takes a list of root nodes to create shortest paths networks for each of the nodes to others. The network is
stopped while reaching a predefined size [11].
Hwang et al. BMC Genomics 2011, 12:439
http://www.biomedcentral.com/1471-2164/12/439
Page 5 of 13