Facile synthesis of 1,1-disubstituted taurines

Article abstract of DOI:10.1016/j.ygeno.2005.11.011

A series of 1,1 -disubstituted taurines was synthesized expeditiously from ketones via the Corey-Chaykovsky epoxidation with dimethylsulfonium methylide, episulfidation with potassium sulfocyanate, ring-opening reaction with ammonia in the presence of sil

Full text of DOI:10.1016/j.ygeno.2005.11.011

Genomics 87 (2006) 315 – 328
www.elsevier.com/locate/ygeno
Comprehensive analysis of pathway or functionally related gene expression
in the National Cancer Institute’s anticancer screen
Ruili Huang a, Anders Wallqvist b, David G. Covell a,
*
a
Laboratory of Computational Technologies, Developmental Therapeutics Program, Screening Technologies Branch, National Cancer Institute at Frederick,
National Institutes of Health, Frederick, MD 21702, USA
b
Science Applications International Corporation, National Cancer Institute at Frederick, National Institutes of Health, Frederick, MD 21702, USA
Received 1 June 2005; accepted 19 November 2005
Available online 4 January 2006
Abstract
We have analyzed the level of gene coregulation, using gene expression patterns measured across the National Cancer Institute’s 60 tumor cell
panels (NCI60), in the context of predefined pathways or functional categories annotated by KEGG (Kyoto Encyclopedia of Genes and Genomes),
BioCarta, and GO (Gene Ontology). Statistical methods were used to evaluate the level of gene expression coherence (coordinated expression) by
comparing intra- and interpathway gene–gene correlations. Our results show that gene expression in pathways, or groups of functionally related
genes, has a significantly higher level of coherence than that of a randomly selected set of genes. Transcriptional-level gene regulation appears to
be on a ‘‘need to be’’ basis, such that pathways comprising genes encoding closely interacting proteins and pathways responsible for vital cellular
processes or processes that are related to growth or proliferation, specifically in cancer cells, such as those engaged in genetic information
processing, cell cycle, energy metabolism, and nucleotide metabolism, tend to be more modular (lower degree of gene sharing) and to have genes
significantly more coherently expressed than most signaling and regular metabolic pathways. Hierarchical clustering of pathways based on their
differential gene expression in the NCI60 further revealed interesting interpathway communications or interactions indicative of a higher level of
pathway regulation. The knowledge of the nature of gene expression regulation and biological pathways can be applied to understanding the
mechanism by which small drug molecules interfere with biological systems.
D 2005 Elsevier Inc. All rights reserved.
Keywords: Pathway; Gene expression; Coregulation; Cancer
The genome encodes two major types of information: genes
encoding proteins that execute biological functions and cis-
control elements of transcription [1]. Proteins may function as
monomers, as complexes, or within networks of interacting
proteins, metabolites, and/or small molecules (pathways). The
cis-control elements, together with transcription factors, form
the linkages and architectures of regulatory networks controlling
expression levels for genes that mediate physiological and
developmental responses. Numerous efforts have been dedicat-
ed to the task of deconvoluting the function and regulation of
biological networks, especially with the increasing availability
of RNA expression microarrays, which provide large amounts of
data for analysis of individual genes within predefined pathways
or for elucidation of hidden gene regulation networks [2–6].
Deriving gene regulation networks or pathways, on the basis of
expression data, is based on the general premise that coregulated
genes function in the same pathway or, in other words,
functionally related genes are coregulated or coexpressed.
Verification of this premise and the extrapolation of these results
toward the identification of causal factors underlying gene
coregulation is currently a very active area of research.
Coexpression of genes has been observed using pathways
annotated by KEGG (Kyoto Encyclopedia of Genes and
Genomes, http://www.genome.jp/kegg/), based on gene ex-
pression data in colon and liver cancer cells and normal tissue
samples [6] and in the Arabidopsis genome [7]. Neighboring
genes, that is, genes that are immediately adjacent on
chromosomes, have been found to be coexpressed in humans
[8,9], Drosophila [10–12], yeast [10], Caenorhabditis elegans
[13], and Arabidopsis [7]. Certain transcription factors, e.g.,
CTCF, serve to insulate adjacent genes from transcriptional
* Corresponding author. Fax: +1 301 846 6978.
E-mail address: covell@ncifcrf.gov (D.G. Covell).
0888-7543/$ - see front matter D 2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.ygeno.2005.11.011

316
R. Huang et al. / Genomics 87 (2006) 315–328
activation. The requirement for coregulation of functionally
related genes has been proposed as a possible cause for the
observed coexpression of genes. In addition, the number of
interactions between proteins has been implicated as an
important predictor for the degree of coexpression between
their corresponding genes, a result that has been offered as an
explanation for particularly high degrees of coexpression in
genes encoding proteins that are known to function in
multicomponent complexes, which often contain a large
number of protein–protein interactions [14,15]. In yeast, genes
encoding interacting proteins tend to be coexpressed
[14,16,17]. In contrast, the degree of coexpression for genes
encoding enzymes in metabolic pathways has been found to be
generally low [6,7], despite the observation of similarities in
gene expression patterns for some metabolic pathways in
embryonic and adult mouse tissues [18]. This finding based on
clustering analysis is, however, mostly observational and not
quantitative. Furthermore, functionally linked interacting pro-
teins have been observed to share higher cis-similarity, defined
as the proportion of shared transcription factor binding sites
regulating the transcription of a gene, than enzymes catalyzing
the conversion of adjacent substrates to products in a collection
of metabolic pathways [19]. This finding has led to the
hypothesis that genes encoding a set of interacting proteins will
be transcribed using a common set of regulatory signals,
whereas substrate concentration and enyzme–substrate inter-
actions may exact regulatory control of metabolic pathways, as
distinct from explicit transcriptional control [20].
Ontology). Comparison of the degree of gene expression
coherence in these three annotation schemes will be used to
postulate rationales that might determine the cohesiveness of a
pathway or a group of functionally related genes. The gene
expression data across the NCI60 will be organized into self-
organizing maps (SOMs) [26], which segregate the data into
nodes, i.e., clusters of genes sharing similar expression
patterns. Meta-clustering of the SOM nodes will be used to
generate a hierarchy of clades. Pathways will be organized
according to similarity of gene occurrence patterns in SOM
clades for each pathway. These results will be used to assess
interactions between pathways either through gene sharing or
through coexpression of genes, which may be indicative of a
higher level of regulation and coherence within cellular
processes. One of our future goals is to utilize the knowledge
obtained about the nature of gene expression regulation and
biological pathways to assess relationships between pathway
gene expressions and drug responses derived from various
cancer cell lines, with the aim of gaining a better understanding
of mechanisms of drug action.
Results
Our measure of pathway cohesiveness is based on correla-
tions of gene expression patterns across the NCI60 for genes
within specific pathways. The cohesiveness of a pathway is
represented by the coherence of its gene expressions as
measured by their correlation strength. Genes in the same
pathway are believed to be regulated in a more coordinated
fashion than a random set of genes, thus expression patterns of
these genes are expected to be more coherent. However, some
pathways appear to be more cohesive than others, due to
factors such as the degree of gene expression regulation within
the pathway, modularity of the pathway, degree of gene sharing
(cross talk) with other pathways, expression data availability,
experimental errors/data quality, etc. In addition, even though
computational methods have been widely used in biological
data analysis and interpretation, a statistically significant
finding does not always translate directly to biological
relevance, due partly to the intricate nature of biological
systems and the inherent complexity of data generated therein.
However, applying rigorous statistical procedures to establish
the significance level of the observations and exclude the
possibility of certain events happening by random chance helps
to improve our confidence for proposing statistically significant
results as a foundation for testable hypothesis generation. We
have followed this principle throughout the course of our
investigations.
The normal network structures of a system, however, may
be perturbed in diseases through genetic mutations and/or by
pathological environmental cues, such as infectious agents or
chemical carcinogens. Cancer is believed to arise from multiple
spontaneous and/or inherited mutations functioning in net-
works that control vital cellular events [21–23], which is partly
reflected by genetic alterations in intracellular signaling path-
ways, which normally orchestrate the execution of develop-
mental programs and the cellular response to extrinsic factors
[24]. The evolving states of certain cancers are reflected in
dynamically changing expression patterns of genes and
proteins within the diseased cells [25]. In support of these
observations, computational methods find that there are more
pathways with coexpressed genes in cancer cells than in normal
cells [6]. System-based analyses of cellular processes that
integrate information on genes, proteins, and metabolites,
combined with technological advances to monitor these
changes, offer the potential for improvements in predictive
medicine [1].
The goal of our present work is to propose novel strategies
to examine at length gene expression regulation patterns within
predefined pathways or functional groups. Our analysis focuses
on the constitutive gene expression data measured across the
National Cancer Institute’s 60-tumor-cell screen (NCI60),
which reflects diverse cell lineages (lung, renal, colorectal,
ovarian, breast, prostate, central nervous system, melanoma,
and hematological malignancies). We will analyze the NCI60
gene expression patterns in terms of pathways annotated by
KEGG and BioCarta and gene categories defined by GO (Gene
The gene expression cohesiveness of pathways defined by
KEGG and BioCarta and annotation categories (terms) defined
by GO can be evaluated by comparing the intra- versus
interpathway gene expression correlations using the Kruskal–
Wallis procedure. The Kruskal–Wallis H statistic (H score) is
computed for each pathway, and the significance ( p value) of
each pathway H score is evaluated by a random gene
permutation procedure. The randomization is run 1000 times;
thus the smallest nonzero p value that can be obtained is 0.001

R. Huang et al. / Genomics 87 (2006) 315–328
317
(see Data and methods for computation details). A pathway is
considered significantly cohesive at the 95% confidence level
if it has a significant ( p ꢀ 0.05) and positive H score. The
larger the H score, the more coherent the gene expressions are
within a pathway, compared to the expressions of genes not
linked by a known pathway. The number of pathways that can
be studied in this fashion is naturally limited by the availability
of gene expression data, and the results obtained will be
affected consequently. In the present investigation, only path-
ways that have at least three gene expression data vectors
available within our microarray set are included in the
calculations. Summary statistics related to the three pathway
systems under investigation are shown in Table 1.
The KEGG pathways can be further grouped into sub-
categories including metabolism, cellular processes, environ-
mental information processing, genetic information processing,
and human diseases, according to their cellular function (see
Table 3). Among those categories, only one group, genetic
information processing, which includes ribosome, proteasome,
aminoacyl-tRNA biosynthesis, basal transcription factors,
DNA polymerase, ubiquitin-mediated proteolysis, protein
export, and RNA polymerase, shows a significant enrichment
of cohesive pathways (Fisher’s exact p = 0.03; see Table 3).
Five of these eight pathways are cohesive and six of them are
composed of genes that encode large protein complexes. In
contrast, the groups metabolism and environmental information
processing, which include all the signaling pathways, appear to
have a less than average percentage of cohesive pathways (see
Table 3), but neither is statistically significant at the 95%
confidence level, compared to the occurrence of all cohesive
pathways in KEGG (19%). The lack of cohesive pathways in
the metabolic pathway group, however, is statistically signif-
icant at the 90% confidence level ( p = 0.07).
Gene expression coherence in pathways annotated by KEGG
Of the 111 KEGG pathways, each of which has at least three
genes with available expression data, 21 (19%) have signifi-
cantly stronger intra- than interpathway gene–gene correla-
tions. These pathways, sorted in descending order by their level
of intrapathway gene expression cohesiveness, and their H
scores are listed in Table 2. The ribosome pathway is the most
cohesive pathway, with a p value of nearly 0. The ribosome
pathway is composed of genes that encode various proteins of
the ribosomal subunits. These proteins need to interact
physically with each other to form a large protein complex,
the ribosome, and are thereby closely related functionally. The
second most cohesive KEGG pathway is oxidative phosphory-
lation, which is composed of genes that encode protein
complexes (complexes I through V) that participate in the
mitochondrial respiration chain. These include the NADH
dehydrogenases, succinate dehydrogenases, cytochrome c oxi-
dases, cytochrome c reductases, ATPases, and ATP synthases.
In fact, in addition to ribosome and oxidative phosphoryla-
tion, most KEGG pathways that are composed of genes
encoding parts of some large protein complex (labeled with
superscript a in Table 2) also appear to be significantly
cohesive, except for RNA polymerase and protein export.
These include the proteasome, ATP synthesis, DNA poly-
merase, and basal transcription factors pathways, which are
significantly enriched (Fisher’s exact p = 0.01) within the set
of cohesive pathways. The fact that the proteins in these
pathways need to interact physically with each other in the
cell to ensure the proper formation and function of the protein
complexes appears to be reflected in their tightly coordinated
gene expression.
Gene expression coherence in pathways deposited in BioCarta
In the 262 BioCarta functional pathways analyzed herein,
34 (13%) show significant intrapathway gene expression
cohesiveness. These pathways, their cohesiveness H scores,
and the significance levels ( p values) of their H scores are
listed in Table 4. A comparison between the genes annotated
in the BioCarta pathway system and the genes in KEGG
shows an overlap of 281 genes (20.5%). When categorized by
the KEGG pathway system, the intersection appears to be
significantly enriched (Fisher’s exact p ꢀ 0.05) in genes that
belong to, in order of descending significance level, MAPK
signaling pathway, integrin-mediated cell adhesion, apoptosis,
cell cycle, toll-like receptor signaling pathway, TGF-h
signaling pathway, cytokine–cytokine receptor interaction,
Jak–STAT signaling pathway, neurodegenerative disorders,
and prion disease. A closer examination of the cohesive
BioCarta pathway genes shows that they are significantly
enriched (Fisher’s exact p ꢀ 0.05), compared to all genes in
the BioCarta–KEGG intersection, with genes that belong to
two KEGG pathways, cell cycle and oxidative phosphoryla-
tion, which are also the third and second most cohesive
pathways, respectively, in KEGG (see Table 2). This indicates
a good agreement in regard to cohesive pathways obtained
from these two pathway systems. This agreement is also
Table 1
Summary statistics for KEGG and BioCarta pathway systems and gene categories defined by GO
Annotation
scheme
Pathways
defined
Pathways with ꢁ3
Maximum pathway
gene count
Average pathway
gene count
Cohesive pathway
Cohesive
gene data vectors
count ( p ꢀ 0.05)
21
pathways (%)
KEGG
134
314
111
262
123
63
19
10
19
13
BioCarta
34
Terms
Terms with ꢁ3
Maximum term
gene count
Average term
gene count
Cohesive term
Cohesive
terms (%)
defined
gene data vectors
count ( p ꢀ 0.05)
171
GO
3564
787
945
20
22

318
R. Huang et al. / Genomics 87 (2006) 315–328
Table 2
Contrary to the findings with KEGG, however, the
pathways categorized as metabolic in BioCarta are signifi-
cantly more cohesive than a typical pathway. The reason may
be due, largely, to the fact that while the contents of
metabolic pathways in KEGG and BioCarta might be
expected to be similar, they are largely different. Metabolism
is the largest pathway category in KEGG, which includes 81
pathways, 11 of which are statistically cohesive. On the other
hand, BioCarta defines only 22 metabolic pathways and 8 of
them are statistically cohesive. A closer examination of the
genes involved in the BioCarta metabolic pathways in
comparison to those defined by KEGG reveals that some
of the genes over represented in the BioCarta metabolic
pathways are actually involved in KEGG pathways that are
mostly (12 of 14) not defined as metabolic, and a couple of
them are significantly cohesive, such as the integrin-mediated
cell adhesion and neurodegenerative disorders pathways. At
the same time, the genes involved in most KEGG metabolic
pathways are largely underrepresented in BioCarta. Combin-
ing the two pathway systems together results in a total of 373
pathways, 55 of which are cohesive; and a total of 103
metabolic pathways, 19 of which are cohesive. This means
15% of all pathways are cohesive and 18% of metabolic
pathways are cohesive. This difference is, however, not
statistically significant (Fisher’s exact p = 0.36). Therefore,
the overall conclusion regarding the cohesiveness of meta-
bolic pathways is that they are not significantly different
from a typical pathway. Nevertheless, it would be interesting
to assess which metabolic pathways are the most and least
cohesive. Using the combined pathways, the genes involved
in the cohesive metabolic pathways are compared with the
ones in the noncohesive metabolic pathways in terms of their
functions, as annotated by GO. The genes in cohesive
metabolic pathways are shown to be significantly enriched
in functional categories such as proton transport, mitochon-
drial electron transport, oxidative phosphorylation, DNA-
directed RNA polymerase activity, transcription, DNA repli-
cation/binding/metabolism, nucleoside biosynthesis and
metabolism, and isoprenoid and cholesterol biosynthesis.
Statistically cohesive KEGG pathways ordered in decreasing level of
significance ( p value)
Pathway Pathway title
H score p
hsa03010 Ribosomea
hsa00190 Oxidative phosphorylationa
4328.66 <1 ꢂ 10ꢃ3
399.42 <1 ꢂ 10ꢃ3
221.71 <1 ꢂ 10ꢃ3
69.30 <1 ꢂ 10ꢃ3
27.77 <1 ꢂ 10ꢃ3
26.47 <1 ꢂ 10ꢃ3
22.81 <1 ꢂ 10ꢃ3
18.11 <1 ꢂ 10ꢃ3
16.70 <1 ꢂ 10ꢃ3
9.83 3.50 ꢂ 10ꢃ3
9.16 4.33 ꢂ 10ꢃ3
9.54 4.61 ꢂ 10ꢃ3
8.52 6.29 ꢂ 10ꢃ3
16.62 7.75 ꢂ 10ꢃ3
5.41 1.37 ꢂ 10ꢃ2
5.91 1.52 ꢂ 10ꢃ2
5.44 1.70 ꢂ 10ꢃ2
5.03 1.74 ꢂ 10ꢃ2
6.00 2.22 ꢂ 10ꢃ2
5.59 2.39 ꢂ 10ꢃ2
4.03 3.00 ꢂ 10ꢃ2
hsa04110 Cell cycle
hsa03050 Proteasomea
hsa00193 ATP synthesisa
hsa00230 Purine metabolism
hsa00531 Glycosaminoglycan degradation
hsa00900 Terpenoid biosynthesis
hsa00970 Aminoacyl-tRNA biosynthesis
hsa03022 Basal transcription factorsa
hsa03030 DNA polymerasea
hsa00100 Biosynthesis of steroids
hsa00260 Glycine, serine, and threonine metabolism
hsa00240 Pyrimidine metabolism
hsa05050 Dentatorubropallidoluysian atrophy
hsa00350 Tyrosine metabolism
hsa00450 Selenoamino acid metabolism
hsa01510 Neurodegenerative disorders
hsa04510 Integrin-mediated cell adhesion
hsa04350 TGF-h signaling pathway
hsa00511 N-Glycan degradation
These pathways show significantly stronger (Kruskal-Wallis p < 0.05) intra-
versus interpathway gene expression correlations.
a
Pathway composed of genes encoding parts of large protein complexes.
apparent in that ‘‘cyclins and cell cycle regulation’’ is the
most cohesive BioCarta pathway and ‘‘electron transport
reaction in mitochondria,’’ which corresponds to the oxidative
phosphorylation pathway in KEGG, is one of the top
cohesive pathways in BioCarta as well. BioCarta, however,
has no genes in common with the most cohesive KEGG
pathway, the ribosome pathway.
Annotation of all BioCarta genes in terms of functional
categories defined in GO shows an enrichment of genes
involved in cell cycle, RNA splicing, translation regulation,
nuclear pore and DNA metabolism, signal transduction,
chromatin modification, cell growth and proliferation, and
apoptosis. The genes in the cohesive BioCarta pathways, in
addition, are especially enriched in RNA splicing, nuclear
pore and DNA metabolism, cell cycle, and cell growth.
KEGG, however, shows a significant lack of genes, except for
cell cycle, in those categories. As a result, the cohesiveness of
these pathways cannot be evaluated within the KEGG
pathway system. BioCarta further groups its pathways into
12 different categories, as shown in Table 5. Four of those
categories, adhesion, cell cycle regulation, developmental
biology, and metabolism, have an above-average (>13%)
number of cohesive pathways; but none of them, with the
exception of metabolism ( p = 0.012; see Table 5), is
significantly different from average at the 95% confidence
level ( p ꢀ 0.05). The other 8 categories have less than
average (<13%) numbers of cohesive pathways. Among
those, 2 pathway categories, cell signaling and cytokines/
chemokines, show a significant ( p < 0.05) lack of cohesive
pathways. This is consistent with the results obtained from
the KEGG pathway system, in which most of the signaling
pathways are not significantly cohesive.
Table 3
Summary statistics for KEGG pathway categories and their level of
cohesiveness
Pathway category
Pathway
count
Cohesive
pathway
count
Cohesive
(%)
p (Fisher’s
exact)
Genetic information
processing
8
4
5
2
62.5
50.0
0.03
0.24
Cellular
processes
Human diseases
Metabolism
8
81
9
2
11
1
25.0
13.6
11.1
0.66
0.07
1
Environmental
information processing
The statistical significance of the coherence level within each pathway category
is indicated by a p value yielded from the Fisher’s exact test comparing the
percentage of cohesive pathways that belong to that category with the
percentage of all cohesive pathways in KEGG.

R. Huang et al. / Genomics 87 (2006) 315–328
319
Table 4
Expression coherence in gene categories defined by Gene
Ontology
Statistically cohesive BioCarta pathways ordered in decreasing level of
significance ( p values)
Pathway title
H score
p
GO annotates genes according to their participation in a
biological process, the molecular function of the gene product,
or the specific cellular location of the expressed gene. Each of
the 787 GO terms that have at least three genes with available
expression data was evaluated for its intraterm expression
coherence. Among these gene categories, 171 (22%) show
significant expression cohesiveness within a GO term. Because
of space considerations, only the 45 GO terms that are cohesive
at a significance level of p < 10ꢃ3 and their H scores are listed
in Table 6. These results show a clear agreement with those
derived from the KEGG and BioCarta pathways. For example,
the GO terms that are the most cohesive include genes that are
engaged in processes such as protein biosynthesis and
ribosomal pathways, which correspond to the ribosome
pathway in KEGG; cell cycle, which corresponds to the same
pathways in KEGG and BioCarta; RNA splicing and binding,
which are also cohesive in BioCarta; mitochondrial electron
transport, which corresponds to the same pathway in BioCarta
and the oxidative phosphorylation pathway in KEGG; DNA
binding and metabolism, which are also cohesive in BioCarta;
and other activities that are not sufficiently represented in
KEGG and BioCarta, such as metal ion binding and transport.
The GO category of signal transduction turns out to be the least
cohesive GO term of all. This is, in fact, consistent with the
findings from the analysis of the KEGG and BioCarta
pathways, in both cases of which the signaling pathway
category contains a less than average number of cohesive
pathways. This finding may be of importance in decisions
about efforts to discover agents that target signaling pathways.
In addition, we find that many GO terms representing large
protein complexes are significantly cohesive. There are 32 GO
Cyclins and cell cycle regulation
The PRC2 complex sets long-term
gene silencing through modification
of histone tails
33.12
22.82
<1 ꢂ 10ꢃ3
<1 ꢂ 10ꢃ3
Spliceosomal assembly
22.74
20.84
<1 ꢂ 10ꢃ3
<1 ꢂ 10ꢃ3
Role of Ran in mitotic spindle
regulation
Antigen processing and presentation
Agrin in postsynaptic differentiation
Glycolysis pathway
19.53
14.82
14.65
13.38
<1 ꢂ 10ꢃ3
<1 ꢂ 10ꢃ3
<1 ꢂ 10ꢃ3
<1 ꢂ 10ꢃ3
Apoptotic DNA fragmentation and
tissue homeostasis
Ion channels and their functional role
in vascular endothelium
7.84
<1 ꢂ 10ꢃ3
Cycling of Ran in nucleocytoplasmic
transport
10.29
1.57 ꢂ 10ꢃ3
Multidrug resistance factors
8.49
9.26
9.26
7.55
1.81 ꢂ 10ꢃ3
2.94 ꢂ 10ꢃ3
2.95 ꢂ 10ꢃ3
4.10 ꢂ 10ꢃ3
Electron transport reaction in mitochondria
uCalpain and friends in cell spread
CTL-mediated immune response against
target cells
Granzyme A-mediated apoptosis pathway
mCalpain and friends in cell motility
h-Arrestins in GPCR desensitization
Eukaryotic protein translation
Sonic hedgehog receptor Ptc1 regulates
cell cycle
8.23
8.05
6.92
6.57
6.69
4.53 ꢂ 10ꢃ3
4.75 ꢂ 10ꢃ3
5.55 ꢂ 10ꢃ3
6.87 ꢂ 10ꢃ3
8.29 ꢂ 10ꢃ3
ER-associated degradation pathway
Regulation of MAP kinase pathways
through dual-specificity phosphatases
Mechanism of protein import into the nucleus
cdc25 and chk1 regulatory pathway in
response to DNA damage
6.83
6.03
8.53 ꢂ 10ꢃ3
9.28 ꢂ 10ꢃ3
6.55
5.86
9.40 ꢂ 10ꢃ3
1.03 ꢂ 10ꢃ2
Overview of telomerase RNA component
gene hTerc transcriptional regulation
The IGF-1 receptor and longevity
Internal ribosome entry pathway
Adhesion molecules on lymphocyte
Monocyte and its surface molecules
Activation of PKC through G-protein-
coupled receptor
4.84
1.40 ꢂ 10ꢃ2
5.38
5.37
4.71
4.71
4.83
1.43 ꢂ 10ꢃ2
1.43 ꢂ 10ꢃ2
1.79 ꢂ 10ꢃ2
1.79 ꢂ 10ꢃ2
1.79 ꢂ 10ꢃ2
Table 5
Summary statistics for BioCarta pathway categories and their level of
cohesiveness
1.95 ꢂ 10ꢃ2
Pathway category
Pathway
count
Cohesive
pathway
count
Cohesive
(%)
p (Fisher’s
Attenuation of GPCR signaling
ChREBP regulation by carbohydrates
and cAMP
4.36
4.36
1.95 ꢂ 10ꢃ2
exact)
Ghrelin: regulation of food intake and
energy homeostasis
4.31
3.92
4.36
1.99 ꢂ 10ꢃ2
2.46 ꢂ 10ꢃ2
2.59 ꢂ 10ꢃ2
Metabolism
Adhesion
22
14
25
8
4
5
36.4
28.6
20.0
0.012
0.14
0.36
Activation of cAMP-dependent protein
kinase PKA
Cell cycle
regulation
Stathmin and breast cancer resistance to
anti-microtubule agents
Developmental
biology
23
4
17.4
0.53
Expression
Immunology
Apoptosis
51
50
5
4
9.8
8.0
7.4
6.7
5.6
0
1
These pathways show significantly stronger (Kruskal-Wallis p < 0.05) intra-
1
versus interpathway gene expression correlations.
27
2
1
Cell signaling
Cell activation
Cytokines/
149
18
10
1
0.003
1
Conversely, the genes involved in noncohesive metabolic
pathways are significantly enriched in categories including
Golgi apparatus, metabolism, fatty acid metabolism, endo-
plasmic reticulum, protein serine/threonine kinase activity,
RAB small monomeric GTPase activity, protein amino acid
glycosylation, and lipid catabolism. Clearly, there appears to
be a functional separation between the cohesive and the
noncohesive metabolic pathways.
39
0
0.012
chemokines
Neuroscience
Hematopoiesis
18
8
0
0
0
0
0.24
0.60
The statistical significance of the coherence level within each pathway category
is indicated by a p value yielded from the Fisher’s exact test comparing the
percentage of cohesive pathways belonging to that category with the percentage
of all cohesive pathways in BioCarta.

320
R. Huang et al. / Genomics 87 (2006) 315–328
Table 6
frequency of cohesive GO terms (22%) (Fisher’s exact p =
0.017; see Table 7). The GO terms that contain the word
‘‘complex’’ also show a slightly higher than average frequency
of cohesive terms but the difference is not statistically
significant. In fact, among the three different ontologies,
biological process, molecular function, and cellular component,
the first two contain about 20% significantly cohesive terms;
cellular component, however, has 30% significantly cohesive
terms (see Table 7). This difference is significant (Fisher’s
exact p = 0.043), which indicates that genes expressed in the
same cellular location, which include, but are not dominated
by, genes encoding proteins that form large protein complexes,
appear to be more coherently expressed than genes simply
involved in the same biological process or engaging in the
same molecular function.
Statistically cohesive GO terms ordered in decreasing H values
Term Title
Biological process
H score
GO:0006412
GO:0007067
GO:0000398
GO:0006260
GO:0006397
GO:0006118
GO:0007049
GO:0008380
GO:0000910
GO:0006120
Protein biosynthesis
1396.83
205.04
186.73
124.32
98.59
Mitosis
Nuclear mRNA splicing, via spliceosome
DNA replication
mRNA processing
Electron transport
69.56
Cell cycle
60.92
RNA splicing
53.67
Cytokinesis
43.53
Mitochondrial electron transport,
NADH to ubiquinone
Generation of precursor metabolites and
energy
36.34
GO:0006091
33.66
GO:0006265
GO:0007001
DNA topological change
Chromosome organization and biogenesis
(sensu Eukaryota)
29.66
29.13
Gene expression coherence in pathways (KEGG, BioCarta)
versus functionally related gene groups (GO)
GO:0006406
GO:0000070
mRNA-nucleus export
Mitotic sister chromatid segregation
22.54
18.10
So far we have been analyzing the level of gene expression
coherence using three annotation schemes without making the
distinction between a pathway and a group of functionally
related genes, although the functionally related gene groups
defined by GO are not pathways per se. Even though genes
involved in the same pathway are generally assumed to be
functionally related, it would still be interesting to examine the
gene composition within a pathway in terms of GO functional
categories. For this purpose, we calculated the percentage of
genes in each KEGG or BioCarta pathway that belong to the
same GO term, which turned out to be 50% on average (data
not shown). In fact, 41% of KEGG pathways and 56% of
BioCarta pathways have over half of their genes belonging to
the same GO term. This confirms the assumption that pathways
contain functionally related genes. A further analysis shows
that the pathways with a higher percentage of genes belonging
to the same GO term are not significantly more cohesive than
the ones with a lower percentage. This is consistent with the
earlier finding that GO terms are not significantly more
cohesive than pathways.
Molecular function
Structural constituent of ribosome
GO:0003735
GO:0003723
GO:0003677
GO:0008248
GO:0016491
GO:0008137
GO:0003954
GO:0004129
GO:0005507
GO:0005509
GO:0003918
GO:0004175
GO:0015078
GO:0019843
GO:0046870
2339.68
1388.27
192.12
132.11
90.43
54.69
50.42
45.70
42.23
37.93
27.86
22.87
20.62
18.99
17.99
RNA binding
DNA binding
Pre-mRNA splicing factor activity
Oxidoreductase activity
NADH dehydrogenase (ubiquinone) activity
NADH dehydrogenase activity
Cytochrome c oxidase activity
Copper ion binding
Calcium ion binding
DNA topoisomerase (ATP-hydrolyzing) activity
Endopeptidase activity
Hydrogen ion transporter activity
rRNA binding
Cadmium ion binding
Cellular component
GO:0005840
GO:0005739
GO:0005842
Ribosome
1290.60
995.19
737.65
Mitochondrion
Cytosolic large ribosomal subunit
(sensu Eukaryota)
GO:0005843
Cytosolic small ribosomal subunit
(sensu Eukaryota)
583.62
GO:0005622
GO:0019866
GO:0005643
GO:0005764
GO:0030530
Intracellular
464.34
63.27
51.98
47.80
47.22
Inner membrane
Table 7
Nuclear pore
Summary statistics for the three ontologies (biological process, molecular
function, and cellular component) and multicomponent protein complexes
(defined by terms containing either ‘‘some’’ or ‘‘complex’’) defined by GO and
their level of cohesiveness
Lysosome
Heterogeneous nuclear ribonucleoprotein
complex
GO:0005624
GO:0005694
GO:0005746
GO:0000786
GO:0005839
GO:0005732
Membrane fraction
43.73
25.07
24.99
24.59
22.87
19.99
GO category
Term
count
Cohesive
Cohesive
(%)
p (Fisher’s
Chromosome
term count
exact)
Mitochondrial electron transport chain
Nucleosome
Molecular function
Biological process
Cellular component
Large protein
311
346
130
32
61
71
39
13
19.6
20.5
30.0
40.6
1
Proteasome core complex (sensu Eukaryota)
Small nucleolar ribonucleoprotein complex
1
0.043
0.017
These GO terms show significantly stronger (Kruskal-Wallis p < 0.05) intra-
versus interterm gene expression correlations. Only GO terms significantly
cohesive at p < 1 ꢂ 10ꢃ3 are listed.
complex (‘‘some’’)
Protein complex
(‘‘complex’’)
47
11
23.4
0.86
The statistical significance of the coherence level within each GO category is
indicated by a p value yielded from the Fisher’s exact test comparing the
percentage of cohesive pathways belonging to that category with the percentage
of all cohesive terms in GO.
terms that end with ‘‘some’’ (ribosome, lysosome, chromo-
some, nucleosome, proteasome, etc.) and 13 (41%) of them are
cohesive, which is significantly higher than the average

R. Huang et al. / Genomics 87 (2006) 315–328
321
Pathway or functionally related genes are more coherently
expressed than a random set of genes
the random distributions. The mean and median H scores for
the three annotation systems are KEGG 47.29, 0.58;
BioCarta 1.47, 0.28; and GO 17.34, 0.59. The mean H
scores for KEGG and GO are skewed by several extremely
cohesive (H > 1000) pathways or GO terms and appear to
be much larger than the median scores as a result. The
observed differences in the H score distributions generated
from the real systems compared to those within the
randomly selected groups of genes are statistically very
significant (t test: GO, p = 2.61 ꢂ 10ꢃ23; KEGG, p =
7.89 ꢂ 10ꢃ19; BioCarta, p = 1.20 ꢂ 10ꢃ15). This leads to
the conclusion that gene expression cohesiveness observed
within pathways is very unlikely to occur by chance;
therefore, it is evident that functionally related genes or
genes involved in the same pathway are, overall, more
coherently expressed than a random set of genes, and
coordinated regulation of these genes, to a certain level, is
a real component of their biological organization.
Based on our analysis of the level of coordinated gene
expression within various pathway systems or functional
categories we find that some functionally related genes are
more coherently expressed than others. The question still
remains as to whether the coexpression of genes could be
observed just by random chance, that is, whether the
categorization of genes into pathways is meaningful. To test
this hypothesis, pathways (groups of genes) are built by
randomly selecting and assigning genes to each pathway, and
the cohesiveness, measured by an H score, is evaluated for
each of these random pathways (see Data and methods for
details). For each annotation system, genes are randomly
permutated a thousand times and H scores for each pathway
calculated. The distribution of the H scores for these
randomly generated pathways, together with that of the
corresponding real gene annotation system, KEGG, BioCarta,
or GO, is shown in Fig. 1. The difference in the
distributions clearly shows that our observations of intrapath-
way gene expression cohesiveness are not random. The three
distributions for the randomly generated gene groups are
almost completely overlapping, yielding a mean random
pathway cohesiveness H score of close to 0 (0.1) and a
median score of 0. The distributions of the three real
systems are similar to one another, but are clearly right-
shifted, with the peak H score at around 1.0, compared to
BioCarta pathways are less cohesive than KEGG and GO
Additionally, among the three gene annotation schemes
analyzed, the pathways defined by BioCarta seem to be the
least cohesive, with the lowest mean and median H scores,
compared to those of KEGG or GO, and this observation is
significant at the 90% confidence level (t test: BioCarta vs
KEGG, p = 0.07; BioCarta vs GO, p = 0.05). Considering
that the mean H scores for KEGG and GO are probably
Fig. 1. Distributions of Kruskal–Wallis H scores calculated for GO terms and KEGG and BioCarta pathways in comparison to sets of gene groups (random
pathways) generated by randomly assigning genes to each group (see Data and methods for details). A large positive H score indicates a high level of pathway
cohesiveness. The distributions of the three true pathway systems are clearly right-shifted compared to the random distributions. The mean H score for the set of
random pathways is close to 0 (0.1) and the median is 0. The median H scores for the three annotation systems are 0.58 (KEGG), 0.28 (BioCarta), and 0.59 (GO),
which are significantly different from random (t test: GO, p = 2.61 ꢂ 10ꢃ23; KEGG, p = 7.89 ꢂ 10ꢃ19; BioCarta, p = 1.20 ꢂ 10ꢃ15).

322
R. Huang et al. / Genomics 87 (2006) 315–328
skewed by the extremely cohesive ribosome pathway (H >
1000), which is not in the BioCarta pathway system, the
median H scores for KEGG and GO are almost the same,
whereas the median H score for BioCarta is 50% lower. This
can probably be explained by the inherent differences within
the nature of genes annotated by KEGG, BioCarta, and GO.
We have analyzed a total of 3365 genes, each of which is
present in at least one of the three annotation schemes, and
1995 of them are annotated only by GO. Of the three, GO is
the most comprehensive gene function annotation scheme,
which tries to assign an annotation to every gene with no bias
toward any specific type of gene. Approximately 90% of the
genes present in KEGG or BioCarta are also annotated by
GO. However, rather than generating annotations per gene,
KEGG attempts to curate and elaborate on metabolic path-
ways, large molecular assemblies, and regulatory pathways. In
fact, a large majority, 81 of 111 (73%), of the KEGG
pathways that we have analyzed are metabolic pathways. As
we have shown earlier, metabolic pathways are not signifi-
cantly more or less cohesive than a typical pathway. On the
other hand, BioCarta has a large collection of signaling
pathways and a smaller number of pathways in other
categories. Cell signaling is the largest pathway category in
BioCarta, which constitutes 57% (149 of 262) of all BioCarta
pathways we have examined. Signaling pathways, however,
are significantly less cohesive than a typical pathway. This
may provide an explanation for the observation that the
overall cohesiveness of KEGG pathways is similar to that of
GO functional categories but BioCarta pathways appear to be
less cohesive than both.
3, top). Furthermore, cluster 1 is composed solely of
metabolic pathways involved in glycan biosynthesis and
metabolism and cluster 4 solely of lipid metabolism (Fig. 3,
bottom).
Cluster 10 contains all of the six signal transduction
pathways, which belong to the environmental information
processing category. The other two pathway subcategories
that belong to environmental information processing are
ligand–receptor interaction, whose pathways are clustered
together in cluster 14, and immune system, which forms one
cluster by itself in cluster 11. Five of the eight pathways
involved in genetic information processing are grouped
together in cluster 15. The ribosome pathway, which belongs
to the translation subcategory in genetic information proces-
sing, and the immune system subcategory, which has only
one pathway (complement and coagulation cascades), each
form one single cluster by themselves (clusters 5 and 11,
respectively). These results can be used as an indication of
the degree of communication or interaction, and subsequent-
ly the level of coregulation, between different pathways.
Clearly, pathways that participate in the same cellular
processes tend to cluster together and thus show a high
degree of interpathway communication by sharing or
coexpressing part of their genes. Therefore, these pathways
are probably coregulated at a level higher than basic gene
expression regulation. Moreover, the pathways that belong to
different categories or subcategories but are clustered
together, as shown at the bottom of Fig. 3, provide valuable
clues about possible interactions or coregulation patterns
between these pathways.
Hierarchical clustering of KEGG pathways shows a higher
level of gene expression regulation
Pathway gene spread over the SOM and expression coherence
As employed by many researchers, simple clustering of
gene expression patterns can also be used to detect
coexpressed or coregulated genes [27]. However, one needs
to be aware of the fact that clustering detects only one form
of coregulation, that is, positive correlations between gene
expression patterns, since negatively correlated genes are
generally not clustered together. Nonetheless, the degree of
spread (DoS; the number of SOM clades the pathway
occupies divided by the maximum number of genes in the
pathway in one clade) of genes within a pathway on the gene
expression SOM can be used as an alternative measure of
pathway gene expression coherence, especially pathways that
are dominated by positive gene correlations. In fact, DoS
turns out to be significantly correlated with the gene
expression cohesiveness H scores (r = ꢃ0.21, p = 0.027),
that is, the more cohesive a pathway is, the less of a spread it
will have on the gene SOM. The average DoS score of the
KEGG pathways is 4.9. The most cohesive is again the
ribosome pathway, having the smallest DoS (0.17), with 80%
of its genes clustered into one clade (see Fig. 2; the ribosome
pathway has only one white square). Other significantly
cohesive pathways, such as cell cycle, oxidative phosphory-
lation, terpenoid biosynthesis, aminoacyl-tRNA biosynthesis,
and biosynthesis of steroids, also have small DoS scores (<3),
An additional perspective on gene expression and path-
ways can be accomplished by independently clustering gene
expressions across tumor cells and then assessing pathway
coherence within neighborhood clusters. Hierarchical cluster-
ing of the spread pattern (occurrences of genes within a
pathway in SOM clades; see Data and methods for details) of
the 111 KEGG pathways on the gene expression SOM (Fig.
2, cohesive pathways are labeled with red dots) further
segregates these pathways into 15 clusters (see Fig. 3), which
can be used to assess whether pathways involved in similar
biological processes share similar regulation by gene expres-
sion patterns. As mentioned earlier, KEGG groups its path-
ways into five general cellular process categories (Fig. 3, top),
which are further divided into 22 subcategories (Fig. 3,
bottom). Fig. 3 shows the composition of each of the 15
pathway clusters, shown in the same order as the dendrogram
in Fig. 2, i.e., adjacent clusters share similar spread patterns,
according to the five general categories (top) and the 22
subcategories (bottom). One can see that pathways belonging
to the same category usually cluster together, that is, genes
from these pathways are partly coexpressed. The 81 metabolic
pathways are scattered across almost all clusters, but 6 of the
15 clusters are composed solely of metabolic pathways (Fig.

R. Huang et al. / Genomics 87 (2006) 315–328
323
Fig. 2. Hierarchical clustering of 111 KEGG pathways. The genes are first grouped into 50 clades based on their expression patterns across the NCI60 using the SOM
clustering procedure. The fraction of genes that fall into each SOM clade is then calculated for every pathway, yielding 111 data vectors. These data vectors are
finally clustered hierarchically, grouping similar pathways together. Each row represents a pathway and each column represents a SOM clade. The fraction of genes
in a pathway that fall into a particular clade is shown by the color of the square in which the pathway and the clade intersect. A yellow to white color indicates a high
fraction value, a reddish color indicates a medium fraction value, and black indicates zero. The dendrogram generated from the hierarchical clustering of pathways is
shown on the left and the 111 pathways are arranged in the order in which they appear in the dendrogram such that neighboring pathways are similar to each other
and have many of their genes fall into the same clade. The dendrogram at the top is generated from hierarchical clustering of the SOM nodes, each cluster of which
forms a clade. The 50 clades are also shown in their dendrogram order such that genes in neighboring clades share similar expression patterns. The general categories
each pathway group belongs to are shown on the right (all pathway names can be found in the supplementary information). The cohesive pathways (intrapathway
gene–gene correlations are significantly stronger than interpathway gene–gene correlations at p < 0.05; see Data and methods for details) are labeled with red dots.
These pathways generally have their genes localized in a few clades, whereas the less cohesive pathways tend to have a larger spread of their genes across many
SOM clades.
characterized by occupying few, but hot, white squares in Fig.
2. These pathways are dominated by strong positive intrapath-
way gene–gene correlations. Interestingly, however, some
pathways shown as significantly cohesive by their H scores,
such as the TGF-h signaling pathway and neurodegenerative
disorders, exhibit a large degree of spread over the SOM
(DoS ꢁ 7; see Fig. 2), indicating that these pathways are
enriched in negative intrapathway gene–gene correlations.
Pathways that have smaller than average DoS scores but are
not shown as significantly cohesive, for example sulfur

324
R. Huang et al. / Genomics 87 (2006) 315–328
Fig. 3. Histograms of the 15 KEGG pathway clusters generated from hierarchical clustering (see Fig. 2 legend for details). Each histogram represents the number of
pathways in a cluster and the clusters are ordered as they appear in the dendrogram. Top: Histograms are colored according to the five general pathway categories
(shown in the key) defined by KEGG. Bottom: Histograms are colored according to the 22 KEGG pathway subcategories (shown in the key). Pathways engaged in
similar cellular processes tend to cluster together.
metabolism, ascorbate and aldarate metabolism, chondroitin/
heparan sulfate biosynthesis, keratan sulfate biosynthesis, and
nitrogen metabolism, are themselves characterized by posi-
tive, but weak, intrapathway gene–gene correlations or do not
have enough genes to achieve statistical significance.
that contributes to the cohesiveness of a pathway. Since the
DoS score of a pathway can be used as another indicator of
pathway cohesiveness, we have examined the relationship
between the number of genes in a pathway that are shared
with other pathways and the DoS score of that pathway for
the KEGG pathways. These two attributes are significantly
correlated with each other (r = 0.36, p = 9.37 ꢂ 10ꢃ5), that
is, the more genes that a pathway shares with others, the less
cohesive the pathway. Fig. 4 clearly shows that the genetic
information processing pathway category, which contains the
As we have mentioned in previous sections, pathways may
communicate with each other not only through coexpressed
genes but also by actually sharing genes. In fact, many genes
participate in multiple pathways and the degree of pathway
cross talk, manifested by gene sharing, may be another factor

R. Huang et al. / Genomics 87 (2006) 315–328
325
Fig. 4. Relationship between the number of genes in a pathway that are shared with other pathways and the cohesiveness of the pathway. Each histogram represents a
KEGG pathway category. Top: Each histogram shows the percentage of cohesive pathways in a particular pathway category. Bottom: Each histogram shows the
average number of genes that a pathway in that category shares with others. Pathways that share many genes with others tend to be less cohesive. The genetic
information processing KEGG category has the highest percentage of cohesive pathways, each of which has the lowest number of genes shared on average.
most cohesive pathways, is also the most modular, i.e., it has
the least number of shared genes (12 per pathway).
Environmental information processing, which contains all
the signal transduction pathways, is one of the least cohesive
pathway categories and also has the largest number of genes
shared with other pathways (78 per pathway).
participating in the same biological process or engaging in the
same molecular function.
Pathways involved in genetic information processing
(replication and repair, sorting and degradation, transcription,
translation) and cell cycle tend to be cohesive as well,
probably because they are responsible for the most vital
processes in a biological entity and, therefore, need to be
tightly regulated at the transcriptional level to ensure precisely
synchronized action of their constitutive genes to minimize
error, or probably because these processes are important for
growth and proliferation and therefore are cohesive in
constantly proliferating tumor cells. On the other hand,
pathways involved in environmental information processing
(signaling pathways) and most metabolic pathways are
generally found to be not cohesive, presumably because these
pathways need to be more robust to respond to environmental
changes. The co-presence of all these genes may not be
necessary, and genes are turned on or off sequentially when
needed. Therefore, these pathways do not need to be tightly
regulated at the transcriptional level and they are more likely
to be regulated by substrate concentration and enyzme–
substrate interactions. This is also reflected by the fact that
these pathways are much less modular than the cohesive
pathways. There is a high degree of cross talk (gene sharing)
between the signal transduction pathways as opposed to the
very cohesive pathways responsible for genetic information
processing, which are highly modular. The few metabolic
pathways that are cohesive show, as previously found, that
pathways responsible for vital life processes such as
nucleotide metabolism, energy metabolism, and isoprenoid
and cholesterol biosynthesis need to have more tightly
regulated gene expression than generic metabolic pathways,
such as carbohydrate metabolism, metabolism of cofactors
Discussion
Our study provides a comprehensive evaluation of the
level of coregulation in pathway gene expressions measured
across the NCI60 using three different gene annotation
schemes, KEGG, BioCarta, and GO. We have discovered
that the level of gene coexpression is, overall, significantly
higher in pathways or functionally related gene groups than a
randomly selected set of genes. Approximately 20% of
pathways or functionally related gene groups analyzed have
statistically significant, coherent gene expressions. Based on
the types of pathways found to be cohesive vs noncohesive,
we postulate that pathway gene expression cohesiveness is
probably on a ‘‘need to be’’ basis, that is, genes in the same
pathway are coexpressed only when there is a need for it.
Pathways are probably designed by nature to be robust and
flexible enough to adapt to environmental changes to ensure
cell survival, that is, alternate mechanisms can take over if
parts of a pathway fail to operate. However, pathways with
genes encoding parts of a large protein complex need to be
cohesive probably because the co-presence and close physical
interaction of the proteins required for the proper function of
the protein complex demand the coexpression and tight
regulation of their corresponding genes. The same may be
true for genes that are expressed in the same cellular location/
component, which are found to be more cohesive than genes

326
R. Huang et al. / Genomics 87 (2006) 315–328
and vitamins, fatty acid metabolism, protein amino acid
glycosylation, and lipid catabolism.
cholesterol biosynthesis. Transcriptional level gene regulation
appears to be on a ‘‘need to be’’ basis, such that pathways
responsible for vital cellular processes or processes that are
related to growth or proliferation, specifically in cancer cells,
are the most cohesive. Clustering of pathways, in addition,
reveals interesting interpathway communications or interactions
indicative of a higher level of pathway regulation.
No significant distinction is found in the level of cohesive-
ness between a pathway, as defined by KEGG or BioCarta, and
a functional gene category, as defined by GO. This can partially
be attributed to the fact that most genes in a pathway are
functionally related, as reflected by over 50% of pathways
containing over half of their genes belonging to the same GO
category. KEGG is characterized by a large percentage of
metabolic pathways, whereas the largest pathway category in
BioCarta is signaling pathways, which are mostly not cohesive.
As a result, BioCarta contains a smaller percentage of cohesive
pathways than KEGG and GO. In addition, a higher level of
pathway regulation is revealed by hierarchical clustering of
pathway gene spread patterns on the gene expression SOM,
which shows interactions between pathways that engage in
similar cellular processes. A similar phenomenon has been
observed in the metabolic network of the yeast Saccharomyces
cerevisiae [2,3]. Also noteworthy is that most of the
significantly cohesive KEGG pathways (highlighted in red in
Fig. 2) appear in neighboring SOM clades. This shows a higher
level of coherence within the cellular processes, namely genetic
information processing, cell cycle, energy metabolism, and
nucleotide metabolism, in which these pathways are partici-
pating. Clusters that contain pathways from different cellular
processes are, furthermore, indicative of interprocess commu-
nications through coexpression and thus coregulation of these
pathways.
Data and methods
Gene expression data
Constitutive gene expression data from Novartis, measured in triplicate
across the 60 tumor cell lines using the Affymetrix DNA oligonucleotide
microarray technology, were downloaded from the Developmental Thera-
peutics Program Web server at http://www.dtp.nci.nih.gov. This data set
contains 12,626 mRNA expression profiles and is publicly available. In this
data set, each expression measurement (signal intensity) comes with a p
value that indicates the reliability of that measurement. The signals with
small p values are stronger and more reliable. We first filtered the data set
to include only measurements that exhibited the strongest signal intensity
( p < 0.05). The logarithm of each signal was taken to suppress extreme
data values. Replicate measurements for each gene were then averaged by
taking the median. Finally, only gene expression profiles having data
available for at least 40 cell lines were included. This yielded a data set of
4923 genes for our analysis.
Pathway data
Three databases were used for pathway gene analysis: the Kyoto Encyclopedia
of Genes and Genomes (http://www.genome.ad.jp/kegg/), Gene Ontology (http://
www.geneontology.org/), and BioCarta (http://www.biocarta.com/). Annotations
for 134 human pathways containing 2804 genes were downloaded from the
KEGG ftp site (ftp://ftp.genome.ad.jp/pub/kegg/pathways/hsa/). BioCarta anno-
tations for 314 pathways containing 1406 human genes were downloaded
from NCI’s Cancer Genome Anatomy Project (http://cgap.nci.nih.gov/) ftp site
(ftp://ftp1.nci.nih.gov/pub/CGAP). Annotations for 3564 GO terms containing
10,921 human genes were downloaded from the GO ftp site (ftp://
ftp.geneontology.org/pub/go/). In the gene expression data set we are working
with, 1047 genes are present in KEGG, 604 are present in BioCarta, and
3210 are present in GO.
It would be interesting, as a future endeavor, to examine
the pathway gene expression coherence in normal tissues in a
similar fashion and compare the results to those obtained
herein, since the gene expression data we have analyzed are
derived solely from cancer cell lines. It is hoped that this will
provide clues as to which and how pathway regulations have
changed in cancer. Knowledge of the nature of gene
expression regulation and biological pathways can be applied
toward understanding the mechanism by which compound
substances or drug molecules interfere with the biological
system through interactions with gene products and conse-
quently pathways.
Pathway cohesiveness and significance calculation
For each pathway Px, a Pearson correlation coefficient (r) was calculated
for each unique gene pair using their expression patterns from the NCI60. These
values are referred to as intrapathway r values. Then for each gene expression
data vector x in Px, an r value is calculated between x and every gene data
vector y, where gene y belongs to some other pathway Py where x m y. These
values are referred to as interpathway r values. The Pearson correlation
coefficient, r, is defined as
Conclusion
We have analyzed gene expression patterns derived from the
NCI60 in the context of predefined pathways or functional
categories annotated by KEGG, BioCarta, and GO. The degree
of gene coexpression, gene coregulation, or pathway cohesive-
ness in these three schemes demonstrates that expression of
genes in pathways, or functionally related genes, has a
significantly higher level of coherence than that of a random
set of genes. Pathways with genes encoding physically
interacting proteins and pathways involved in genetic informa-
tion processing and cell cycle tend to be cohesive and modular,
whereas signaling pathways are generally not cohesive and have
a high degree of interpathway cross talk. Most metabolic
pathways are not cohesive, except for the ones responsible for
nucleotide metabolism, energy metabolism, and isoprenoid and
n
ꢃ
ꢃ
~
i¼1 ðxi ꢃ x Þðyi ꢃ y Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r ¼
ð1Þ
2
2
n
ꢃ
ꢃ
~
i¼1 ðxi ꢃ x Þ ðyi ꢃ y Þ
where x and y denote averages, and the summation runs over the number of cell
¯ ¯
lines (n). The intra- and interpathway r values are then used as two sample
populations to calculate the Kruskal–Wallis H statistic for pathway Px,
ꢀ
ꢁ
k
12
Rt2
nt
H ¼
~
ꢃ 3ðN þ 1Þ
ð2Þ
NðN þ 1Þ
t¼1
where Rt is the rank sum of sample population t, nt is the size of sample
population t, k is the number of sample populations being compared, and
N ¼ ~kt¼1nt. When each of the k sample populations being compared includes

R. Huang et al. / Genomics 87 (2006) 315–328
327
at least five observations, the sampling distribution of H is a very close
approximation of the m2 distribution for k ꢃ 1 degrees of freedom. It is actually
a fairly close approximation even when one or more of the samples include as
few as three observations. In the present study, only pathways that have at least
three gene expression data vectors available are included in the calculations.
Significance levels ( p values) are obtained using the m2 distribution with 1
degree of freedom (two-sample populations, intra- and interpathway correlation
coefficients, are compared here; therefore, k = 2 and k ꢃ 1 = 1).
single-linkage hierarchical clustering of the 111 data vectors was performed
using MATLAB, which segregated the pathways into 15 groups. A dendrogram
showing the neighboring pathways was generated.
Acknowledgments
The authors thank the members of the STB staff, especially
Drs. Robert Shoemaker and Susan Mertins for valuable
contributions during the preparation of the manuscript. This
project has been funded in whole or in part with federal funds
from the National Cancer Institute, National Institutes of
Health, under Contract NO1-CO-12400. The content of this
publication does not necessarily reflect the views or policies of
the Department of Health and Human Services, nor does
mention of trade names, commercial products, or organizations
imply endorsement by the U.S. Government.
A large H score (H > 3.84) indicates that a statistically significant ( p <
0.05) difference exists between the intra- and the interpathway r populations. A
small r is assigned a lower rank and a large r a higher rank. If the average rank
of the interpathway r values is higher than that of the intrapathway r values,
then a negative sign is applied to the H score. Therefore, a large positive H
score indicates a high level of gene expression coherence within the pathway
compared to expression of genes not linked by a known pathway. Either
absolute or real r values can be used when calculating the H scores, and slightly
different results will be obtained, depending on whether the intrapathway r
values are dominated by positive or negative values. Absolute r values are used
in all calculations unless otherwise specified, since both significant positive and
negative correlations between two genes are assumed to be indicative of
coordinated expression.
Appendix A. Supplementary data
Randomization procedures are used to check the probability of getting a
large positive H score by chance compared to the p values obtained directly
from the m2 distribution. Random pathways of sizes approximately 5, 10, 15,
25, 50, and 100 are built by randomly picking and assigning genes in the data
set to each pathway, and their H scores are calculated. This procedure is
repeated 1000 times and the probabilities of getting H scores at various levels
for each pathway size are calculated. Further increase in the number of
randomizations does not appear to affect the outcome. The probabilities
obtained this way agree very well with the p values obtained directly using the
m2 distribution for pathway sizes 10 to 25 (number of genes in each pathway;
the size of the pathways we are working with is 10–25), that is, an H score of
>3.84 is required to get p < 0.05. However, the H score required to reach a
certain significance level (e.g., p < 0.05) shows a slight linear dependency on
pathway size (i.e., sample size). For larger pathways, higher H scores are
needed to get a significant p value. For this reason, we have applied the
randomization procedure to obtain the probability of observing an extreme H
score by chance for each pathway. Therefore, if a pathway in a particular
annotation system has n genes, then n genes are randomly selected from the
pool of m genes annotated by the system and the H score is calculated. This
procedure is repeated 1000 times for each pathway, and the fraction of H scores
that are more extreme than the actual H score of the pathway is then assigned as
the coherence p value for that particular pathway. If none of the 1000 random H
scores is more extreme than the actual H score, then the pathway is
significantly cohesive at p < 0.001. The number of significantly cohesive
( p < 0.05) pathways selected in this fashion, however, does not differ much
from that obtained directly from the m2 distribution because most pathways
(>90%) have fewer than 50 genes and most significant pathways (>60%) have
very large H scores (>7.5). For each gene annotation system, KEGG, BioCarta,
or GO, a distribution of the pathway H scores was calculated for each random
permutation; the mean and standard deviation (shown as error bars) of the
thousand random distributions were then plotted and are shown in Fig. 1. These
three random distributions are almost entirely coincidental.
Supplementary data associated with this article can be found
in the online version at doi:10.1016/j.ygeno.2005.11.011.
References
[1] L. Hood, J.R. Heath, M.E. Phelps, B. Lin, Systems biology and new
technologies enable predictive and preventative medicine, Science 306
(2004) 640–643.
[2] J. Ihmels, S. Bergmann, N. Barkai, Defining transcription modules
using large-scale gene expression data, Bioinformatics 20 (2004)
1993–2003.
[3] J. Ihmels, R. Levy, N. Barkai, Principles of transcriptional control in the
metabolic network of Saccharomyces cerevisiae, Nat. Biotechnol. 22
(2004) 86–92.
[4] Z. Li, C. Chan, Inferring pathways and networks with a Bayesian
framework, FASEB J. 18 (2004) 746–748.
[5] Z. Li, C. Chan, Integrating gene expression and metabolic profiles, J. Biol.
Chem. 279 (2004) 27124–27137.
[6] H.H. Yang, Y. Hu, K.H. Buetow, M.P. Lee, A computational approach to
measuring coherence of gene expression in pathways, Genomics 84
(2004) 211–217.
[7] E.J. Williams, D.J. Bowles, Coexpression of neighboring genes in
the genome of Arabidopsis thaliana, Genome Res. 14 (2004)
1060–1067.
[8] H. Caron, et al., Evidence for two tumor suppressor loci on chromosomal
bands 1p35–36 involved in neuroblastoma: one probably imprinted,
another associated with N-myc amplification, Hum. Mol. Genet. 4 (1995)
535–539.
[9] M.J. Lercher, A.O. Urrutia, L.D. Hurst, Clustering of housekeeping genes
provides a unified model of gene order in the human genome, Nat. Genet.
31 (2002) 180–183.
We chose the H score as an indicator of pathway gene expression
cohesiveness because a significance measure exists for the H score calculated
for a pathway and, by using rank scores in place of actual correlation values, the
method is less sensitive to pathway size and gene outliers. This method is also a
more reliable measure of pathway coherence because not only gene–gene
relationships within a pathway are considered, but also comparisons are made
to check whether the intrapathway gene–gene interactions are in fact stronger
than interactions between genes not connected by some known pathway.
[10] B.A. Cohen, R.D. Mitra, J.D. Hughes, G.M. Church, A computational
analysis of whole-genome expression data reveals chromosomal domains
of gene expression, Nat. Genet. 26 (2000) 183–186.
[11] A.M. Boutanaev, A.I. Kalmykova, Y.Y. Shevelyov, D.I. Nurminsky, Large
clusters of co-expressed genes in the Drosophila genome, Nature
(London) 420 (2002) 666–669.
[12] P.T. Spellman, G.M. Rubin, Evidence for large domains of similarly
expressed genes in the Drosophila genome, J. Biol. 1 (2002) 5.
[13] M.J. Lercher, T. Blumenthal, L.D. Hurst, Coexpression of neighboring
genes in Caenorhabditis elegans is mostly due to operons and duplicate
genes, Genome Res. 13 (2003) 238–243.
Hierarchical clustering of KEGG pathways
The genes were first grouped into 50 clades based on their expression
patterns across the NCI60 using the SOM clustering procedure. The fraction of
genes that fall into each SOM clade was then calculated for every KEGG
pathway, yielding 111 data vectors each containing 50 fractions. Wards-based
[14] H. Ge, Z. Liu, G.M. Church, M. Vidal, Correlation between transcriptome
and interactome mapping data from Saccharomyces cerevisiae, Nat.
Genet. 29 (2001) 482–486.

328
R. Huang et al. / Genomics 87 (2006) 315–328
[15] L.M. Staudt, P.O. Brown, Genomic views of the immune system, Annu.
Rev. Immunol. 18 (2000) 829–859.
[21] C.A. Klein, Gene expression signatures, cancer cell evolution and
metastatic progression, Cell Cycle 3 (2004) 29–31.
[16] A. Grigoriev, A relationship between gene expression and protein
interactions on the proteome scale: analysis of the bacteriophage T7 and
the yeast Saccharomyces cerevisiae, Nucleic Acids Res. 29 (2001)
3513–3519.
[22] P.A. Covitz, Class struggle: expression profiling and categorizing cancer,
Pharmacogenom. J. 3 (2003) 257–260.
[23] D. Hanahan, R.A. Weinberg, The hallmarks of cancer, Cell 100 (2000)
57–70.
[17] R. Jansen, D. Greenbaum, M. Gerstein, Relating whole-genome expression
data with protein–protein interactions, Genome Res. 12 (2002) 37–46.
[18] R. Miki, et al., Delineating developmental and metabolic pathways in vivo
by expression profiling using the RIKEN set of 18,816 full-length enriched
mouse cDNA arrays, Proc. Natl. Acad. Sci. USA 98 (2001) 2199–2204.
[19] S. Hannenhalli, S. Levy, Transcriptional regulation of protein complexes
and biological pathways, Mamm. Genome 14 (2003) 611–619.
[20] M. Ptashne, A. Gann, Imposing specificity by localization: mechanism
and evolvability, Curr. Biol. 8 (1998) R812–R822.
[24] U. Rennefahrt, M. Janakiraman, R. Ollinger, J. Troppmair, Stress kinase
signaling in cancer: fact or fiction? Cancer Lett. 217 (2005) 1–9.
[25] O.J. Halvorsen, et al., Gene expression profiles in prostate cancer:
association with patient subgroups and tumour differentiation, Int. J.
Oncol. 26 (2005) 329–336.
[26] T. Kohonen, Self-Organizing Maps, Springer-Verlag, Berlin, 1995.
[27] M.B. Eisen, P.T. Spellman, P.O. Brown, D. Botstein, Cluster analysis and
display of genome-wide expression patterns, Proc. Natl. Acad. Sci. USA
95 (1998) 14863–14868.