Gene expression signatures that can
discriminate oral leukoplakia subtypes
and squamous cell carcinoma
Nobuo Kondoha,*, Shuri Ohkuraa, Masaaki Araia, Akiyuki Hadaa,
Toshio Ishikawab, Yutaka Yamazakic, Masanobu Shindohd,
Masayuki Takahashie, Yoshimasa Kitagawac,
Osamu Matsubaraf, Mikio Yamamotoa
aDepartment of Biochemistry II, National Defense Medical College, 3-2 Namiki, Tokorozawa-shi 359-8513, Japan
bIshihara Sangyo Kaisha, LTD., Nishi-Shibukawa, Kusatu-shi 525-0025, Japan
cDepartment of Oral Diagnosis, Division of Oral Pathobiological Science, Hokkaido University Graduate
School of Dental Medicine N13 W7, Kita-ku, Sapporo 060-8586, Japan
dDepartment of Oral Pathology and Biology, Division of Oral Pathobiological Science, Hokkaido University Graduate
School of Dental Medicine N13 W7, Kita-ku, Sapporo 060-8586, Japan
eDepartment of Oral surgery, National Defense Medical College, 3-2 Namiki, Tokorozawa-shi 359-8513, Japan
fDepartment of Pathology II, National Defense Medical College, 3-2 Namiki, Tokorozawa-shi 359-8513, Japan
Received 10 March 2006; accepted 25 April 2006
Available online 18 September 2006
noma (OSCC) and leukoplakias (LPs), and evaluate its diagnostic potential. In order to identify
marker gene candidates, differential gene expression between LPs and OSCCs were examined
by cDNA microarray. The expression of 118 marker gene candidates was further evaluated by
quantitative reverse transcription-PCR (QRT-PCR) analyses of 27 OSCC and 19 LP tissues. We
identified 12 up-regulated and 15 down-regulated marker genes in OSCCs compared to LPs.
Using Fisher’s linear discriminant analysis (LDA), we demonstrated that 11-gene predictors
among this novel marker set could best distinguish OSCCs from LPs (>97% accuracy), whereas
a further seven of these gene predictors could be utilized to distinguish higher grade (higher
than moderate) from lower grade (lower than mild) dysplasias (>95% accuracy). These predictor
The purpose of this study is to generate a classifier for oral squamous cell carci-
Head and neck/
Gene expression profiling;
1368-8375/$ - see front matter ?
c 2006 Elsevier Ltd. All rights reserved.
Abbreviations: OSCC, Oral squamous cell carcinoma; LP, Leukoplakia; LDA, Fisher’s linear discriminant analysis; loo, Leave-one-out cross
validation; QRT-PCR, Quantitative reverse transcription-polymerase chain reaction.
*Corresponding author. Tel.: +81 4 2995 1469; fax: +81 4 2996 5190.
E-mail address: firstname.lastname@example.org (N. Kondoh).
Oral Oncology (2007) 43, 455–462
available at www.sciencedirect.com
journal homepage: http://intl.elsevierhealth.com/journals/oron/
gene sets provide multigene classifiers for the diagnosis of pre-cancerous to cancerous transi-
tion of oral malignancy.
c 2006 Elsevier Ltd. All rights reserved.
Oral leukoplakias (LPs) are white lesions that include hyper-
plasias and dysplasias of the oral mucosa, and often undergo
malignant transformation to oral squamous cell carcinoma
(OSCC).1The histological features associated with LP, and
with OSCC, have been described previously.2The dysplasias
are classified as either mild, moderate or severe, based on
histopathological findings, and these designations are
thought to be the sequential phases of oral carcinogenesis.
However, we sometimes experience a discordance between
the pathological diagnosis of these lesions and the corre-
sponding prognosis in an individual patient. It has been sug-
gested that the detection of clonal genetic changes, such as
loss of heterozygosity or microsatellite instability, in both
primary OSCC and pre-malignant lesions could be a more
informative method of monitoring these cancer patients.3
However, the staging and grading of OSCCs are still, for
the most part, dependent on traditional clinicopathological
In line with our previous approach,4there are several
attempts for the identification of molecular biomarkers
associated with specific phenotypes of head and neck
squamous cell carcinomas.5–7Although, these approaches
are effective to observe the wide variety of genetic alter-
ations in oral malignancy, in order to accurately diagnose
cancer subtypes, supervised learning methods are the
most suitable.8–11In this study, we attempted to generate
a multigene classifier for the diagnosis of oral malignan-
cies. Using both cDNA microarray and QRT-PCR tech-
niques, a comprehensive gene expression profile was
generated and compared among OSCC and leukoplakia tis-
sues. We subsequently defined a list of 27 marker genes
that are either significantly elevated or down-regulated
in OSCCs, compared with LPs. Among these genes, predic-
tor gene sets for OSCC-LP classification were determined
by Fisher’s discriminant analysis (LDA) and validated by
the leave-one-out method.
Our present results suggest that these differentially ex-
pressed genes can not only provide a valuable prognostic
tool for OSCC patients, but could also provide important
information concerning the molecular mechanisms underly-
ing the development of oral squamous cell carcinoma.
Materials and methods
Twenty-seven OSCCs and 19 leukoplakias, including four
hyperplastic, three mild, seven moderate and five severely
dysplastic tissues, were incorporated into this study (Table
1). These samples had been surgically resected in the den-
tal hospital of Hokkaido University from February 1998 to
April 2004. Each of the samples was then macrodissected,
and the stromal tissues were carefully removed as much
as possible by a pathologist and stored at ?100 ?C. All
of these procedures were undertaken following informed
consent from each patient, and adhered to the ethical
guidelines of the Dental hospital of the Hokkaido Univer-
sity School of Dentistry (Sapporo, Japan). The main histo-
clinical characteristics of each patient are summarized in
RNA extraction and microarray analysis
RNA extractions were performed using ISOGEN (Nippon
Gene, Japan), according to the manufacturer’s instructions.
Total RNA isolates extracted from five LPs (one hyperplasia,
and one mild, one moderate and two severe dysplasias), and
from five OSCC tissues, were aliquoted as mixtures. cDNA
microarray analysis was performed using the IntelliGene
HS Human expression chip, containing 16,600 probe sets
(Takara, Japan). Briefly, for each mixture, 4 lg of total
RNA was used for single-stranded cDNA synthesis using a
T7-promoter linked oligo-dT primer. Subsequently, ampli-
fied (a)-RNA synthesis was performed in the presence of
aminoallyl-UTP (Ambion, USA), then coupled with either a
30Cy3- or 30Cy5 N-hydroxysuccinimide ester. After purifica-
tion, the labeled a-RNAs were combined and hybridized to
the IntelliGene HS Human expression chips. After overnight
hybridization at 70 ?C, a fluorescence scan was performed
using a ScanArray Express microarray reader (Perkin–Elmer,
CA, USA), and the signals were quantified using the accom-
panying software. To ensure the reliability of the data,
genes were considered to be differentially expressed if
the P values were <0.001 (Student t-test), and had a fold
For validation of the microarray analysis, QRT-PCR, using
RNA extracts from several OSCC and leukoplakia samples
(Table 1), was performed as previously described.12Each
RNA sample (1 lg) was reverse-transcribed using Superscript
II (Invitrogen). For each gene showing a 3-fold or higher dif-
ference in expression on the microarrays, PCR primer pairs
were designed (see Table 2) using Primer Express software
(Perkin–Elmer). Each amplification reaction was carried
out using Sybr Green Mastermix in an ABI Prism 5700 Se-
quence Detection System (Applied Biosystems, CA, USA)
for 15 min at 95 ?C for the initial denaturing, followed by
40 cycles of 95 ?C for 30 s and 60 ?C for 1 min. The expres-
sion values for each gene were normalized to the expression
levels of the S5 ribosomal protein transcript (Genbank
accession no. NM_001009; primer pairs, 50-GAG CGC CTC
ACT AAC TCC ATG ATG A-30and 50-CAC TGT TGA TGA TGG
CGT TCA CCA-30), which was used as an internal control.
456N. Kondoh et al.
To evaluate the statistical significance of the differentially
expressed genes, a Student t-test (two-sided) was per-
formed. We performed the direct sequencing analyses and
examined that each PCR product was correctly amplified
Unsupervised hierarchical clustering analysis was performed
using DNASIS Pro software (Hitachi Software Engineering,
Tokyo, Japan). In this study, a combination of the clustering
algorithms Pearson correlation (for dissimilarity) and Ward
(for cluster consolidation) was used.
The supervised classification of the data was examined by
the use of Fisher’s linear discriminant analysis (LDA).13Gi-
ven a set of independent variables (pij), the weight vector
of linear discriminant function is determined to maximize
the ratio between the interclass distance and the sum of
variance within each class:
F ¼ jmA? mBj2=fr2
A;B : classes
i ¼ 1 ? Ni : an index for sample ðcaseÞ
j ¼ 1 ? Nj : an index for variable ðgeneÞ
In the present study, a set of independent variables (a
‘‘model’’) was selected by the use of a stepwise increment
method and genetic algorithm.14The stability of the model
was examined by leave-one-out cross validation (loo). Each
sample was masked one after another. An LDA model was
then build from residual Ni ? 1 samples and the class of
the masked sample was predicted from this model.
linear discriminant function
A¼ Ri2Aðgi? mAÞ2=NA
The identification of potential marker genes
between leukoplakia and squamous cell carcinoma
To identify potential marker genes that are differentially
expressed between LP and OSCC, cDNA microarray analyses
were performed using RNA mixtures of five OSCC and of five
leukoplakias. Among the 16,600 target cDNAs on the chip ar-
rays, 4600 genes showed a detectable signal of which 63
were highly expressed (3-fold or more) in the OSCC mixture,
compared with the LP mixture. In addition, 53 genes were
preferentially expressed (3-fold or more) in the LP mixture
(data not shown). Since the cDNA array tip does not contain
probes for keratin-4 and transglutaminase three genes,
whose differential expression between LPs and OSCCs has
already been verified in our previous study,4the expression
of these genes was additionally examined in the following
QRT-PCR analysis. QRT-PCR analyses using 27 OSCC and 19
Pathoclinical appearances of tissue samples
S dys 25
S dys 26
S dys 27
S dys 29
S dys 49,50e
Mod dys 21
Mod dys 30
Mod dys 36–40d
Mod dys 42–44d
Mod dys 33
Mod dys 52
Mod dys 47,48e
Mild dys 28
Mild dys 32,33e
Mild dys 46
NA, not applicable.
bT, tongue; P, palate; LG/UG, lower/upper gingiva; FM, floor
of mouth; Bu, buccal mucosa; Si, sinus.
cM, male; F, female.
dMixture of three samples derived from different portions in
the same lesion.
eMixture of two samples derived from different portions in the
Gene expression signatures that can discriminate oral LP subtypes and SCC457
Leukoplakia (LP)- and OSCC (SC)-dominant marker genes
Gene name (symbol)Category Accession no.PCR primers (forward/reverse)
Matrix Gla protein (MGP)
Nuclear receptor subfamily 2,
group F, member 2 (NR2F2)
Slit homolog 3 (Drosophila) (SLIT3)
Keratin 1 (KRT1)
Keratin 13 (KRT13)
Proline arginine-rich end leucine-rich
repeat protein (PRELP)
Chromosome 1 open reading frame 10
FBJ murine osteosarcoma viral
oncogene homolog B (FOSB)
Transglutaminase 3 (TG3)
Trans cription factor
LP27B NM_006732 CCAATGCTCCAGCTGTCGTCTG/GCAGGAGCAAGCCCTGCTCAC
AGCAATAACTCCATGGGCTCTGG/CAACGGAACAGGAGTCCTTCAACT LP4A, B FXYD domain containing ion transport
regulator 6 (FXYD6)
Microfibrillar-associated protein 4
Dihydropyrimidinase-like 3 (DPYSL3)
Cysteine-rich protein 1 (intestinal)
SC1AChemokine (C-X-C motif) ligand 10
Laminin, gamma 2 (LAMC2),
transcript variant 1
Follistatin (FST), transcript
Immunoglobulin J polypeptide,
linker protein (IGJ)
Retinol binding protein 1,
Transforming growth factor,
beta-induced, 68 kDa (TGFBI)
Epithelial stromal interaction 1
Interferon-induced protein 44
Interferon-induced protein with
tetratricopeptide repeats 3 (IFIT3)
Ubiquitin specific protease 18 (USP18)
Carbonic anhydrase II (CA2)
SC5A Extracellular matrixNM_005562CTACGGATCACAGCTCCCTTGATG/TGAGATTCCGCAGTAACCTTCGATAC
SC13A Secretory proteinNM_006350ACCGCAATGAATGTGCACTCCTAA/CAGTAGGCATTATTGGTCTGGTCCAC
SC43A, B Secretory proteinNM_144646 AGTGTGCCCGGATTACTTCCAGG/CAATGGTGAGGTGGGATCAGAGAT
SC3 Nuclear receptor NM_002899GATCTGACAGGCATAGATGACCG/CTTCCACTCTCATCTCTAGGTGCA
SC27Secretory protein NM_198966CGGTTAGCCCTGTTCCACGAAC/CAAATCAGAGGCGCTTCCTCTG
aGenes up-regulated in leukoplakia (LP) or oral squamous cell carcinoma (SC).
bA – 11 predictor genes that can discriminate between LPs and OSCCs; B – 7 predictor genes that can discriminate between mild- and moderate-dysplasias.
N. Kondoh et al.
LP tissues (see Table 1) further demonstrated that among
these differentially expressed marker gene candidates, 33
were significantly overexpressed in OSCC, whereas 30 genes
were up-regulated in LP tissues (P < 0.06). In order to mini-
mize the potential effects of any anatomical differences in
the oral cavity upon gene expression, we confirmed the dif-
ferential expression levels from the microarrays and QRT-
PCR experiments in a further 9 OSCC and 12 LP tissues de-
rived from tongue only (data not shown).
We focused finally on 27 differentially expressed mar-
ker genes, among which 15 (denoted as LP) were overex-
pressed in LPs compared to OSCCs, whilst 12 genes were
up-regulated in OSCCs and in some moderate- to severe-
dysplasias (denoted as SC) (Table 2). Among these, the
expression of 24 marker genes was significantly altered
between OSCC and LP tissues (P < 0.05), irrespective of
any anatomical differences. Although, the expression of
SC 43 was not significantly altered between OSCCs and
LPs (P = 0.98), we also took the expression of this gene
into account, since it was strongly suppressed in all
hyperplasias and mild dysplasias and may enable the clas-
sification of dysplasias into subgroups. We also took the
expression of SC1 and SC44 into account as the SC1 tran-
scripts seem to be significantly elevated in a specific pop-
ulation of OSCC and because both SC44 and SC1 mRNAs
are significantly suppressed in most LPs except for Mo
dys 47, 48. The expression patterns of these 27 marker
genes are summarized in Figure 1.
Unsupervised classification of oral squamous cell
carcinoma and leukoplakia tissue using our newly
identified marker genes
identified, unsupervised hierarchical clustering was under-
taken to classify OSCC and LP tissue samples. As shown in Fig-
ure 1, 27 OSCC and 19 LP sample tissues were separated into
cluster Ia was mainly comprised of LPs of a relatively well
differentiated phenotype, including hyperplasia and mild dys-
plasia, whereas subcluster Ib consisted of a mixture of both
OSCC and moderate- to severe-dysplasias. With the exception
of Mo dys 33, cluster II consisted almost entirely of OSCCs.
These results of the unsupervised clustering demonstrate
that 27 marker genes could harbor principal components
essential for the classification of the dysplasia–carcinoma
The aim of our present study is to establish a clear distinc-
tion between LP and OSCC, and among dysplasia grades,
based upon a molecular classification. In order to identify
marker gene sets that could discriminate between OSCCs
and LPs, a supervised classification approach using LDA13
was performed. The expression of these 27 marker genes
and II are indicated at the bottom. Gene IDs corresponding to Table 2 are listed on the right. The tumor types (on the top)
correspond to the information in Table 1. The relative gene expression levels are represented as color intensity on a logarithmic
scale (high, red; low, green; medium, black). The statistical significance (Student t-test on both sides) of the differential gene
expression levels between OSCCs and LPs is denoted on the right. The asterisk indicates that the OSCC tumors 56 and 57 were
derived from the same patient, before and after 5-FU treatment, respectively.
Unsupervised clustering of OSCC and LP tumors, based on marker gene expression data. The clustered tissue groups Ia, Ib
Gene expression signatures that can discriminate oral LP subtypes and SCC 459
was analyzed among 27 OSCCs and 19 LPs, including hyper-
plasias and dysplasias (Table 1). This approach involves
parameter (gene) selection by the use of a stepwise incre-
ment and genetic algorithm.14In the former method, start-
ing from a null model, the most effective variable is added
to a previous candidate set of variables in each step. When
the score of the model reaches a maximum, the increment
stops. When the Fisher’s ratio was employed as a score, a
model with 11 parameters was selected as the best model.
The stability of this model was examined by the leave-one-
out cross validation (loo) method and compared with that of
a stepwise increment method and a genetic algorithm (fit)
(Fig. 2). The LDA score for each sample is given as the fol-
lowing linear discrimination function:
Score = ?0.231 (LP1) + 0.223 (LP4) ? 0.0537 (LP28) ?
0.0734 (LP21) ? 0.892 (LP12) ? 0.0617 (LP29) ? 0.282
(LP8) + 0.0122 (SC1) + 0.0669 (SC13) ? 0.0684 (SC43) ?
The gene expression and LDA scores are summarized in
Figure 3. With the exception of OSCC16 and Mo dys 33,
the sample sets were correctly discriminated by the 11 mar-
ker genes selected by the Fisher’s ratio. The optimal predic-
tion accuracy with this set of 11 genes was 97.8% (fit) and
97.8% (loo), respectively.
Using the same approach, we also attempted to distin-
guish mild dysplasias and hyperplasias from higher grade
(>moderate) dysplasias and OSCCs. As shown in Figure 4,
we reached the optimal prediction using a set of seven
genes with an accuracy of 97.8% (fit) and 95.6% (loo),
respectively. The LDA score for each sample is given as
the following linear discrimination function:
Score = ?0.372
0.0380 (LP28) ? 0.246 (LP17) ? 0.464 (LP27) + 0.0889
(LP5) + 0.495 (LP4) ? 0.576 (LP19) +
The predictor genes are summarized in Table 2.
In order to identify marker gene candidates, we first
screened differential gene expression between OSCCs
and LPs by using cDNA microarray. Although microarrays
are a powerful technology for measuring genome-wide
expression, there is the difficulty in reproducibly quantify-
ing sources. We, therefore, verified the expression of all
118 marker gene candidates by using QRT-PCR of 27 OSCC
and 19 LP tissues. We first identified 27 marker genes rep-
resenting the differential expression between LPs and
OSCCs. To further identify marker gene sets and establish
appropriate algorism that can sufficiently discriminate be-
tween OSCCs and LPs, a supervised classification approach
using LDA was performed. After intensive parameter
selection and cross validation, we reached an optimal pre-
diction with a set of 11 genes. According to our classifica-
tion, however, a LP sample, Mo dys 33, was classified as
an OSCC. Interestingly, Mo dys 33 had been clinically diag-
nosed and treated as an OSCC, because of its cancerous
macroscopic appearances and history (data not shown).
Therefore, a molecular diagnosis based on these 11 genes
may, in part, predict clinical features or genetical back-
ground of the patient rather than histological grades.
Using the same approach, we could also distinguish mild
dysplasias, hyperplasias, and perhaps normal mucosa from
higher grade (>moderate) dysplasias and OSCCs with a fur-
ther set of seven genes.
Using differential display and northern blot analyses, we
previously identified several differentially expressed genes
between LPs and OSCCs.4Among these genes, keratin-13
and TG-3 are also included in the present 27 marker genes,
however in this experiment, the differential expression of
keratin-14 and -17 remained undetectable level in the cDNA
microarray analysis; the expression of keratin-4 did not
show significant difference in the QRT-PCR analysis using
tissues derived from tongue only (not shown). These dis-
crepancies in the results are mainly attributable to the
technical difference, and in part, the increased sample
1 1116 2126
Accuracy of prediction
of both a stepwise increment method and genetic algorithm (fit). Each parameter represents marker gene.
Stability of the model examined by leave-one-out cross validation (loo) methods. The results were compared with those
460 N. Kondoh et al.
number. Anyhow, among the verified 27 marker genes, 11-
and 7-predictor sets are optimal to keep the accuracy of
diagnosis, demonstrating that the combination usage, but
not simply the number, of marker genes is essential to con-
struct appropriate predictor sets.
The rationality of some predictor genes is supported by
the following observations. The down-regulation of TG-3
expression is a common event in the development of hu-
man esophageal, and head and neck cancers.8,9,15The loss
of the C1orf10 gene (LP19) product, also known as a mem-
ber of the fused gene family, cornulin,16has also been
suggested to contribute to the development of OSCC.17
Furthermore, the matrix Gla protein (MGP; LP1), encoding
a vitamin K-dependent extracellular matrix (ECM) protein,
indicated in Fig. 1. The LDA score for each sample is given by the linear discrimination function (see results): d – OSCCs; s – LPs.
Expression profiles of 11 predictor genes and discrimination by LDA scores. The relative gene expression levels are as
described for Fig. 3: d – OSCCs and severe- and moderate-dysplasias; s – mild dysplasias and hyperplasias.
Expression profile of seven predictor genes and discrimination by LDA scores. The gene expression and LDA scores are as
Gene expression signatures that can discriminate oral LP subtypes and SCC461
is associated with transformation and de-differentiation of Download full-text
colonic epithelial cells.18The expression of MGP is inver-
sely correlated with tumor size, lymph-node metastasis
and tumor grade in renal cell carcinoma, though the
expression is significantly elevated in the tumor compared
to matched normal tissue.19The expression of MGP is also
elevated in breast cancer.20Therefore, MGP may set up
scaffolding for primary tumor cells, by contrast acquisition
of metastatic phenotype may be correlated with down-
regulation of this ECM molecule. The decreased expression
of dermatopontin (DPT; LP5) is associated with the devel-
opment of leiomyoma and keloids21and FosB (LP27) is
down-regulated in poorly differentiated breast carcino-
mas.22Hence, the reduction of these genes may be an
important manifestation of the pre-malignant to malignant
On the other hand, Laminin gamma 2 (LAMC2; SC5) is
overexpressed in head-and-neck carcinomas.23Follistatin
(FST; SC13) is overexpressed in liver tumors,24and is sug-
gested to be an antagonist of activin-mediated growth inhi-
bition.25It has also been reported that the chemokine ligand
10 (CXCL10; SC1) may confer anti-malignant properties in
breast cancer.Moreover, in response to IFNgamma,
CXCL10 is induced via the NF-kappa B pathway in OSCC
In conclusion, our present results demonstrate that the
expression profile of the marker genes that we have identi-
fied through cDNA microarrays and QRT-PCR could prove to
be a useful diagnostic marker and also provide valuable
information forour further
understanding of oral
1. Lopez M, Aguirre JM, Cuevas N, Anzola M, Videgain J,
Aguirregaviria J, Castro A, de Pancorbo MM. Gene promoter
hypermethylation in oral rinses of leukoplakia patients – a
diagnostic and/or prognostic tool? Eur J Cancer 2003;39:
2. Bloor BK, Seddon SV, Morgan PR. Gene expression of differen-
tiation-specific keratins in oral epithelial dysplasia and squa-
mous cell carcinoma. Oral Oncol 2001;37:251–61.
3. Forastiere A, Koch W, Trotti A, Sidransky D. Head and neck
cancer. N Engl J Med 2001;345:1890–900.
4. Ohkura S, Kondoh N, Hada A, Arai M, Yamazaki Y, Sindoh M,
et al. Differential expression of the keratin-4, -13, -14, -17 and
transglutaminase 3 genes during the development of oral
squamous cellcarcinoma from
5. Schmalbach CE, Chepeha DB, Giordano TJ, Rubin MA, Teknos
CR, Bradford CR, et al. molecular profiling and the identifica-
tion of genes associated with metastatic oral cavity/pharynx
squamous cell carcinoma. Arc Otolaryngol Head Neck Surg
6. Carinci F, Lo Muzio L, Piattelli A, Rubini C, Palmieri A, Stabellini
G, et al. Genetic portrait of mild and severe lingual dysplasia.
Oral Oncol 2005;41:365–74.
7. Ginos MA, Page GP, Michalowicz BS, Patel LJ, Volker SE,
Pambuccian SE, et al. Identification of a gene expression
signature associated with recurrent disease in squamous cell
carcinoma of the head and neck. Cancer Res 2004;64:55–63.
8. Whipple ME, Mendez E, Farwell DG, Agoff SN, Chen C. A
genomic predictor of oral squamous cell carcinoma. The
9. Roepman P, Weesels LFA, Kettelarij N, Kemmeren P, Miles AJ,
Lijinzaad P, et al. An expression profile for diagnosis of lymph
node metastases from primary head and neck squamous cell
carcinoma. Nature Genet 2005;37(2):182–6.
10. Iizuka N, Oka M, Yamada-Okabe H, Mori N, Tamesa T, Okada
T, et al. Comparison of gene expression profiles between
hepatitis B virus- and hepatitis C virus-infected hepatocellular
carcinoma by oligonucleotide microarray data on the basis of
a supervised learning method. Cancer Res 2002;62(14):
11. Ramaswamy S, Tamayo P, Rifkin R, Mukherjee S, Yeang C-H,
Angelo M, et al. Multiclass cancer diagnosis using tumor gene
12. Arai M, Kondoh N, Imazeki N, Hada A, Hatsuse K, Kimura F,
et al. Transformation-associated gene regulation by ATF6
during hepatocarcinogenesis. FEBS Lett 2006;580:184–90.
13. Fisher NA. The use of multiple measurements in taxonomic
problems. Ann Eugen 1936;7(Part II):179–88.
14. Goldberg DE. Genetic algorithm in search, optimization and
machine learning. Reading (MA): Addison-Wesley; 1989., p. 1–
15. Chen BS, Wang MR, Xu X. Transglutaminase-3, an esophageal
cancer-related gene. Int J Cancer 2000;88(6):862–5.
16. Contzler R, Favre B, Huber M, Hohl D. Cornulin, a new member
of the ‘‘Fused gene’’ family, is expressed during epidermal
differentiation. J Invest Dermatol 2005;124:990–7.
17. Imai FL, Uzawa K, Nimura Y, Moriya T, Imai MA, Shiiba M, et al.
Chromosome 1 open reading frame 10 (C1orf10) gene is
frequently down-regulated and inhibits cell proliferation in
oral squamous cell carcinoma. Int J Biochem Cell Biol 2005;37:
18. Fan C, Sheu D, Fan H, Hsu K, Allen Chang C, Chan E.
Down-regulation of matrix Gla protein messenger RNA in
human colorectal adenocarcinomas. Cancer Lett 2001;165:
19. Levedakou EN, Strohmeyer TG, Effert PJ, Liu ET. Expression of
the matrix Gla protein in urogenital malignancies. Int J Cancer
20. Chen L, O’Bryan JP, Smith HS, Liu E. Overexpression of matrix
Gla protein mRNA in malignant human breast cells: isolation by
differential cDNA hybridization. Oncogene 1990;5:1391–5.
21. Catherino WH, Leppert PC, Stenmark MH, Payson M, Potlog-
Nahari C, Nieman LK, et al. Reduced dermatopontin expression
is a molecular link between uterine leiomyomas and keloids.
Genes Chrom. Cancer 2004;40:204–17.
22. Milde-Langosch K, Kappes H, Riethdorf S, Loning T, Bamberger
AM. FosB is highly expressed in normal mammary epithelia, but
down-regulated in poorly differentiated breast carcinomas.
Breast Cancer Res Treat 2003;77:265–75.
23. Patel V, Aldridge K, Ensley JF, Odell E, Boyd A, Jones J, et al.
Laminin-gamma 2 overexpression in head-and-necksquamous
cell carcinoma. Int J Cancer 2002;99:583–8.
24. Rossmanith W, Chabicovsky M, Grasl-Kraupp B, Peter B,
Schausberger E, Schulte-Hermann R. Follistatin overexpression
in rodent liver tumors: a possible mechanism to overcome
activin growth control. Mol Carcinog 2002;35:1–5.
25. Stove C, Vanrobaeys F, Devreese B, Van Beeumen J, Mareel
M, Bracke M. Melanoma cells secrete follistatin, an antago-
26. Goldberg-Bittman L, Neumark E, Sagi-Assif O, Azenshtein E,
Meshel T, Witz IP, et al. The expression of the chemokine
receptor CXCR3 and its ligand, CXCL10, in human breast
adenocarcinoma cell lines. Immunol Lett 2004;92:171–8.
27. Hiroi M, Ohmori Y. Constitutive nuclear factor kappa B activity
is required to elicit interferon-gamma-induced expression of
chemokine CXC ligand 9 (CXCL9) and CXCL10 in human tumour
cell lines. Biochem J 2003;376:393–402.
462N. Kondoh et al.