GENE EXPRESSION ARRAYS

The development of gene expression microarrays has enabled
investigators to move beyond analysis of single genes and not
only explore patterns or profiles of gene expression in tissues
but also compare these patterns between tissue types. These
gene expression profiles have been shown to be consistent between
institutions, with good comparability of sample characteristics.
96,97 Examining the expression profiles of tumors
allows for novel identification of genes previously not associated
with malignancy. In addition, comparing expression
profiles between groups of patients has enabled investigators
to perform molecular phenotyping and classification of tumors.
Unique and characteristic profiles have been identified
not only for the histologic types of NSCLC, but also for subgroups
within these histologic classes. 98–101
There are other potential uses for data acquired from
microarray-derived gene expression profiles. Correlating these
profiles with defined tumor behavior may help to elucidate specific
processes or pathways involved in carcinogenesis. For example,
profiling invasive and metastatic tumors may reveal novel
genes or pathways of interest that could subsequently be targeted
for anticancer therapeutics. Also, these methods may be used to
investigate a tumor’s response (or nonresponse) to therapy, elucidating
possible mechanisms of drug resistance and providing a
basis for predicting future clinical behavior (see Chapter 47).
Gene expression profiles have also been used to refine and
predict prognosis and survival among patients with identical
TNM staging but differing clinical outcome, 102,103 and is the
focus of this chapter. Other uses include identification of novel
biomarkers associated with and specific to lung tissue and/or
NSCLC. 104 In the current era of molecular therapeutics, there is
potential to use these gene expression profiles and biomarkers not
only to classify tumors with molecular staging techniques, but
also to identify targets for therapy and treatment. As these techniques
continue to develop, it is possible that lung cancer staging,
and therefore treatment, will depend not only on TNM status
but also on the genetic profiles generated by these methods.
Lung Cancer Heterogeneity Clinically, lung cancer is
classified as small cell (SCLC) and non–small cell (NSCLC),
with NSCLC accounting for approximately 80% of all lung
cancers. Within each histologic subtype, there is significant
heterogeneity such that NSCLCs are further classified as adenocarcinoma,
squamous cell carcinoma, large cell carcinoma,
and neuroendocrine carcinoma as well as tumors with mixed
histology such as adenosquamous tumors. 105 Among adenocarcinomas,
further variation is present in acinar, papillary,
bronchoalveolar carcinoma (BAC), and mucinous carcinoma
subtypes. For example, BAC appears to arise from type II
pneumocytes and is generally associated with better prognosis
compared with invasive adenocarcinomas. 106
The heterogeneity among primary lung tumor subtypes
likely reflects the potential cell derivation, and these differences
may be further increased by the diverse genetic alterations observed
in lung cancers. 107–110 Additional tumor heterogeneity
may be a result of alterations in gene expression that affect
diverse processes such as proliferation, apoptosis, and cellular
differentiation, among others. 111,112 To this end, gene expression
profiling methods have been employed to better understand
this heterogeneity and to identify specific pathways or
genes that might distinguish tumors of different cellular origin
or clinical behavior.
One of the first studies of gene profiling for lung cancer
was performed by Petersen et al. 113 Comparing a metastatic
lung adenocarcinoma with human small airway epithelial cells,
cDNA libraries of upregulated and downregulated genes were
constructed. DNA sequencing of over 500 clones revealed 315
unknown and 205 known cDNA fragments. Gene expression
analysis of 167 of these clones was performed using northern
blot techniques, confirming differential expression in 58% of
the clones. In addition, an expression pattern similar to that of
the metastatic adenocarcinoma was observed in lung cancer cell
lines. No primary lung tumors were examined in this study.
Garber et al. 114 performed a more global analysis of gene
expression using 23,100 element cDNA arrays and 12,600 transcript
containing oligonucleotide arrays. Samples in this study
included 41 adenocarcinomas, 16 squamous cell carcinomas,
5 large cell carcinomas, and 5 small cell lung tumors. In addition,
there were 11 tumors with corresponding lymph node or
intrapulmonary metastasis as well as 5 normal lung samples.
Hierarchical clustering was performed on a subset of genes representing
only 918 of the 23,100 transcripts. These methods
distinguished relatively distinct groups of squamous cell, small
cell, and large cell carcinomas; in addition, three groups of adenocarcinomas
were recognized. These results suggested that
the histologic origin of these tumors was reflected in the expression
patterns determined by these 918 genes. Furthermore,
the authors suggested that the division of adenocarcinomas into
three groups reflects tumor heterogeneity within this subgroup
of NSCLC and that cumulative patient survival differed between
these clusters. It should be noted, however, that clinical
factors such as differentiation, grade, or tumor stage were not
used in determining the three groups. In addition, most of the
paired samples of primary and metastatic tumors demonstrated
relatively similar gene expression patterns. Specific genes differentially
expressed in the three adenocarcinoma clusters included
VEGFc (highly expressed in the poor outcome group)
and thyroid transcription factor (highly expressed in the good
outcome group). These results are consistent with known information
regarding VEGF expression, differentiation status, and
prognosis for patients with lung adenocarcinoma. 2,76,99
Bhattacharjee et al. 99 also utilized gene expression profiling
to classify lung tumors into histologic subtypes. Using oligonucleotide
arrays, they examined 125 adenocarcinomas, 21
SCCs, 20 carcinoid tumors, 6 SCLC tumors, and 17 normal
lung samples. Hierarchical and probabilistic clustering of the
3312 most variably expressed transcripts separated the tumors
into distinct clusters, which reflected their histologic subtype.
In addition, primary lung tumors were distinguished from adenocarcinomas
of colonic origin metastatic to the lung based
on their expression profiles.
Specific genes with high expression levels in each histologic
group were identified. Marker genes such as TGF- receptor
type II, tetranectin, and ficolin 3 characterized normal
lung samples. Both carcinoid tumors and small cell tumors
expressed high levels of neuroendocrine genes such as insulinoma-
associated gene 1, gastrin-releasing peptide, and chromogranin
A. However, few other markers were common between
SCLC and carcinoid tumors. For SCLC, high expression of
cell proliferation-associated genes such as PCNA, thymidylate
synthase, MCM2, and MCM6 was demonstrated. Similarly,
carcinoid tumors were defined by a distinct expression profile.
A separate clustering of the adenocarcinoma expression profiles
defined four subclasses, C1 to C4. Consistent with the
Garber et al. 114 study, many of the genes associated with each
of the four clusters appeared to reflect tumor differentiation
status or stage-related differences. In addition, Kaplan-Meier
analysis revealed a significantly worse median overall survival
for patients in the C2 subgroup, comprised of tumors with
high expression levels of neuroendocrine genes.
Nacht et al.114a investigated patterns of gene expression
in NSCLC using serial analysis of gene expression (SAGE).
Libraries were established from normal human bronchiole
epithelial cells, small airway epithelial cells, two squamous cell
carcinomas, two adenocarcinomas, and the A549 lung adenocarcinoma
cell line. From these samples, 18,300 independent
clones were sequenced and 574,634 tags were generated, representing
66,502 distinct transcripts. Adenocarcinomas demonstrated
high levels of the CD74 antigen, major histocompatibility
complex (MHC) class II, and immunoglobulin (Ig) heavy
constant 3, whereas genes highly expressed in squamous cell
tumors included glutathione peroxidase 2 (GPX2) and tumor
necrosis factor receptor subfamily member 18. Consistent with
other studies, high levels of surfactants A and B were observed
in adenocarcinoma; however, this likely reflects the differences
in the cell type derivation of these tumors. 99,114 Additional
investigation was performed on 10 NSCLCs, 4 normal controls
and a larger panel of 32 normal lung and lung tumors
using both real-time reverse transcription polymerase chain
reaction (RT-PCR), and 12,600-element containing (U95)
oligonucleotide arrays. Although only a small number of genes
were compared by both approaches, 21 of 23 exhibited similar
expression patterns as determined by both SAGE and the oligonucleotide-
based techniques. Although the basic molecular
features of adenocarcinoma and squamous cell carcinoma can
be distinguished by gene expression profiling, the authors suggested
that the histological and clinical behavior of the tumors
may depend on more subtle changes in expression levels for a
variety of genes and pathways.
Several studies have used gene expression arrays to determine
differences between matched tumor and normal samples.
Using differential cDNA library screening, McDoniels-Silvers
et al. 115 examined differentially expressed genes between primary
lung adenocarcinoma, SCC, and corresponding normal
lung tissues. Dot-blot hybridization techniques confirmed that
1163 clones were differentially expressed between the normal
and tumor tissues. Using RT-PCR methods, the authors confirmed
that 113 genes were differentially expressed between
normal and tumor tissues, some of them underexpressed or
overexpressed in tumors relative to normal lung, or selectively
expressed in adenocarcinoma versus squamous cell carcinoma.
With this approach, genes that were highly overexpressed were
selected with the greatest frequency. In addition to some genes
with unknown functions, the gene identified as highly overexpressed
included genes involved in glycolysis, cell respiratory
complex, inflammation, and cell adhesion.
Nakamura et al. 116 used 425 element cDNA arrays to
examine stage I tumors, including both adenocarcinoma and
squamous cell carcinoma, for genes that are differentially expressed
between tumors and corresponding noncancerous
lung tissue. In the tumor samples, 74 genes were underexpressed
and 40 genes were overexpressed when compared to
normal controls. Elevated expression of plasminogen activator,
MMPs 1, 3, 7, and 10 and keratins 4, 6B, 8, 13, 14, 19, and
20 were observed in the carcinoma group. Conversely, several
cell adhesion–related genes including cadherin 5, cadherin 6,
protocadherin 2, catenin beta 1, integrin beta 1, and CD31
had decreased expression in the tumor group.
Wikman et al. 117 described the use of arrays containing
1176 genes to compare fourteen lung adenocarcinomas
with normal lung tissue as well as lung tissue from four normal
references. Two statistical methods, principal component
analysis and permutation tests, were used to identify the most
differentially expressed genes between the tumors and normal
tissues. Three main groups of genes were identified to be dysregulated
most frequently: those involved in cell motility and
structure, matrix maintenance and degradation, and cell cycle
regulation. Genes upregulated in tumors relative to normal samples
included known tumor markers such as topoisomerase 2A
(TOP2A), KRT19, KRT8, tenascin, polo-like kinase (PLK) 1 , and
cyclin B1 (CCNB1). In addition, MMP11, MMP12, and tissue
inhibitor of metalloproteinase 1 (TIMP1) were upregulated in
the tumors, whereas TIMP3 was downregulated. For lung cancer,
previous reports have demonstrated high levels of expression
of MMP genes, whereas TIMP3 is subject to epigenetic silencing
by methylation. 118–120 In addition, the elevated mRNA expression
of cytokeratins (KRT8, 18, and 19) was consistent with
the gene profiling studies of others. 102,121 Other genes elevated
in the adenocarcinoma group included CCNB1, macrophage
migration inhibitory factor (MIF), high-mobility–group protein
Y (HMGI) and hepatocyte-derived growth factor (HDGF).
Genes that were downregulated in the tumors included SOCS2
and SOCS3, caveolins 1 and 2, gravin, and the mitogen-responsive
phosphoprotein DOC2/DAB2.
Gene Expression Arrays and Prognostication
Several studies have correlated results from gene expression arrays
with patient prognosis. In a study of 19 stage I and II
adenocarcinomas, Miura et al. 122 used cDNA microarrays to
examine tumors from 14 smokers versus 5 nonsmokers; among
these patients, 6 were 5-year survivors and 12 died of lung
cancer recurrence. Gene expression patterns differed between
the smokers and nonsmokers. Several genes exhibiting lower
levels of expression in smokers were those located in known
regions of genomic imbalance for NSCLC such as chromosome
3p21.3. 123 Other genes with lower expression among
smokers were located in chromosomal regions 4q, 11q23–24,
19p, and19q. The authors suggested that inactivation of these
genes was related to tobacco carcinogenesis. In addition, tobacco-
related carcinoma was associated with high expression
levels of RAB4, DJ1, MCT, and ribosomal protein L22. Of
the genes examined, 27 genes were differentially expressed between
nonsurvivors and survivors. Fourteen genes had high expression
levels in nonsurvivors, including anaphase-promoting
complex 2 (APC2). In contrast, expression levels of the mitotic
spindle checkpoint regulatory genes hBUB3 and hZW10 were
lower in tumors from nonsurvivors, highlighting the importance
of cell cycle regulation in human carcinogenesis.
Wigle et al. 124 utilized 19,200 element-containing cDNA
arrays and examined 39 NSCLCs that showed either cancer
recurrence or no recurrence. The cohort included adenocarcinoma
and squamous cell carcinoma as well as other histologic
subtypes; stage I, II, and III lung cancers were included.
Based on unsupervised hierarchical clustering of a subset of
2899 genes, two groups differed significantly in disease-free
survival. Genes associated with a more aggressive NSCLC
behavior included ataxia telengiectasia mutated (ATM), upregulation
of the flt1 VEGF receptor, and phosphoinositide-
3-kinase regulatory subunit ( PIK3R2 ). Using Cox proportional
hazards model testing, investigators determined 22 genes that
were significant for disease-free survival.
Beer et al. 102 examined 86 lung adenocarcinomas using oligonucleotide
arrays (HuGeneFL) containing 6800 transcripts.
Sixty-seven of the samples were stage I tumors, whereas 19 were
stage III; 10 samples of normal lung tissue were also examined.
Three clusters of tumors were identified using hierarchical clustering
and other supervised analytical approaches. Significant
relationships were observed between cluster and tumor differentiation
as well as cluster and tumor stage. Suggesting that the gene
expression profile of some early stage tumors is similar to that of
more clinically aggressive tumors, the authors noted that some of
the stage I adenocarcinomas clustered with higher stage tumors.
To help determine which genes were best related to patient prognosis,
a 50-gene “risk index” based on the top 50 survival-related
genes was devised. Using this approach, low- and high-risk stage
I adenocarcinomas that differed significantly with regard to survival
were correctly identified.
The survival-related genes identified in this study were
broadly grouped into the following categories: cell cycle and
cell signaling related; apoptosis related; transcription and
translation; cell adhesion and structure; genes encoding chaperones,
receptors, enzymes, and transcription factors; and those
with unknown function. Genes of particular interest included
VEGF, keratin 7, cathepsin L, and the CRK oncogene. Other
lung cancer profiling studies had reported elevated expression
of cathepsin L and keratin 7 genes in aggressive tumors. 99,114
In addition, VEGF has been previously identified as being associated
with poor prognosis lung cancer. 2,76
Kikuchi et al. 125 and Inamura et al. 126 identified genes associated
with lymph node metastasis among primary lung ADs,
and Hoang et al. 127 identified genes associated with nonmetastatic
tumors, those with micrometastases, and those with overt
metastasis. Xi et al. 128 used the Bhattacharjee et al. 99 (see previous
discussion) and the Beer et al. 102 (see discussion on prognosis
later) datasets to examine whether gene expression in primary
AD tumors was indicative of lymph node metastases. A 318-gene
signature was able to accurately classify node positive patients in
the training 102 and test 99 sets, but frequently misclassified node
negative patients. The classification as node negative or positive
in the node-negative patients was associated with survival. These
studies suggest that the survival differences observed among
stage I ADs in the Garber et al. 114 and Bhattacharjee et al. 99
datasets might be related to the presence of micrometastases or
metastatic potential. The use of gene expression for “molecular
staging” may enhance the sensitivity of clinical and pathologic
methods for staging tumors, improving treatment decisions and
ultimately outcomes for lung cancer patients.
Several studies using the primary lung tumor to predict
lymph node metastases Kikuchi et al. 125 examined 37 NSCLCs
using cDNA microarrays containing 23,040 genes. The initial
data set was trimmed to 899 genes and investigators used hierarchical
clustering methods to separate the tumors into groups
based on their histologic subtypes. Next, the investigators established
a predictive scoring system based on the expression
profiles of selected genes. When used to calculate the predictive
score, 40 genes provided the best separation of node-positive
and negative adenocarcinomas. Previous studies have reported
an association between tumor metastasis and several of the genes
used to calculate the predictive score: ARHA, DB1, NESH, and
TACSTD1. 129–132 Finally, the authors studied the expression
of the metastasis-related genes after treatment with six different
chemotherapeutic agents: cisplatin, docetaxel, gemcitabine, irinotecan,
paclitaxel, and vinorelbine. Analysis revealed a number
of genes correlating the sensitivity of the adenocarcinomas or
SCC to the six drugs. In particular, YWHAQ gene expression
levels correlated with the sensitivity of lung adenocarcinomas to
cisplatin, docetaxel, gemcitabine, and paclitaxel. Xi et al. 133 used
the Bhattacharjee et al. 99 (see previous discussion) and the Beer
et al. 102 datasets to examine whether gene expression in primary
adenocarcinoma tumors was indicative of lymph node metastases.
A 318-gene signature was able to accurately classify nodepositive
patients in the training and test 99 sets, but frequently
misclassified node negative patients. The classification as node
negative or positive in the node-negative patients was associated
with survival. These studies suggest that the survival differences
observed among stage I adenocarcinomas might be related to
the presence of micrometastases or metastatic potential.
In 2006, Potti et al. 134 reported the use of gene expression
arrays to develop a risk model of recurrence for early stage lung
cancers. Using an initial set of 89 NSCLCs, which included
both squamous cell carcinoma and adenocarcinoma, investigators
first established a collection of gene expression profiles,
which they termed metagenes . Prognostic models were built from
the metagenes using classification- and regression-tree analysis.
In the initial training cohort, the metagene model predicted
disease recurrence with an accuracy of 93%, compared with
64% as predicted by the prognostic model built with clinical
data alone. These data were supported in Kaplan-Meier survival
analyses. Validation of the metagene model was performed on
two independent cohorts from multicenter cooperative group
trials, the American College of Surgeons Oncology Groups
(ACOSOG) Z0030 study and the Cancer and Leukemia
Group B (CALGB) 9761 trial. With these results, investigators
described a possible role for predicting disease recurrence
for patients with early stage lung cancers, thereby identifying
patients for whom adjuvant chemo therapy might otherwise not
be indicated.
More recently, Chen et al. 135 reported the use of gene expression
arrays to develop a five-gene model for prediction of
relapse-free and overall survival in lung cancer. Expression arrays
were used on 125 tumors. Through Cox regression analysis
and calculation of hazard ratios (HR), 16 genes were correlated
with death from any cause. Risk scores were calculated
for these 16 genes and patients were classified as having a highor
low-risk gene signature. Expression levels of the 16 genes
were confirmed by RT-PCR and further statistical analysis
identified 5 genes that were significantly associated with
patient survival: monocyte-to-macrophage differentiationassociated
protein (MMD), dual-specificity phosphatase 6
(DUSP6), v-erb-b2 avain erythroblastic leukemia viral oncogene
homologue 3 (ERBB3), signal transducer and activator
of transcription 1 (STAT1), and lymphocyte-specific protein
tyrosine kinase (LCK). Some of these have previously been
described as playing a role in carcinogenesis: DUSP6 has been
implicated in tumor suppression and apoptosis; ERBB3 is a
tyrosine kinase and member of the EGFR family; STAT1 has
been implicated in cell growth and apoptosis through induction
of p21 Waf1 and caspase; and LCK is a member of the
Src family of tyrosine kinases and has been shown to regulate
mobility of cancer cells. 136–142 The predictive value of the
five-gene signature was subsequently validated on an independent
cohort of 60 additional patients. Compared to those
with a low-risk gene signature, patients with a high-risk gene
signature had a significantly shorter median survival. Results
were similar when patients with stage I disease were examined
separately; however, there was no correlation between gene
signature and overall survival for patients with stage II disease.
An insightful editorial regarding this work was published by
Herbst and Lippman 143 who pointed out that since the specimens
were not microdissected, the analysis could be misleading
with regard to the importance of invasion-related genes,
which can vary in expression throughout a tumor. Moreover
future studies must analyze molecular epidemiologic, stromal,
and vascular factors that are critical to the metastatic process.
Finally, the choice of the cutoff of expression levels and filtering
of the data could influence how the genes were selected.
Nevertheless, further validation studies on a set of 86 tumors
previously analyzed by another group 102 also demonstrated
a significant risk for death from any cause with the high-risk
gene signature as well as a trend toward significance when
analyzed for survival.
Reproducibility of Data One challenge in the use of
gene expression arrays for prognostication in lung cancer is the
reliability of platforms across institutions and reproducibility of
data or identification of candidate genes involved in prognosis
across institutions. Some investigators have addressed this by
validating their predictive models with cohorts from outside
studies and institutions. 135,144 In an effort to investigate the variability
between laboratories and across institutions, Dobbin et
al. 97 reported both “within-laboratory” and between- laboratory
reproducibility of microarray data across four institutions for a
set of primary tumors, lung cancer cell lines and purified RNA
samples. Although the between-laboratory variation was highest,
the investigators concluded that the reproducibility, and
therefore, the comparability of the data were adequate for the
samples studied.
Hayes et al. 100 evaluated three cohorts of adenocarcinomas,
each from a different institution and with its own gene array
platform. Statistical analysis identified 2553 genes that were
present and reliable across the three platforms. Adenocarcinoma
tumor subtypes, named bronchioid, squamoid, and magnoid
by the authors, were each distinguished by several hundred
genes. Lists of genes characteristic of tumor subtype in one
cohort were predictive of tumor subtypes in the other two cohorts.
While investigators were able to analyze survival data for
only one cohort of patients, patients with stage I and II tumors
had significantly shorter survival times when their tumors were
classified as squamoid and magnoid. In contrast, for patients
with stage III and IV tumors, there was a trend toward increased
survival for patients with squamoid subtype.
Sun et al. 145 used two lung cancer oligonucleotide microarray
data sets of adenocarcinoma and squamous cell carcinoma
as training sets to select prognostic genes independent of conventional
predictors. The top 50 genes from each set were used
to predict the outcomes of two independent validation data
sets of 84 and 91 NSCLC cases. Adenocarcinomas with the
50-gene signature from adenocarcinoma in both validation data
sets had a 2.4-fold (95% confidence interval [CI], 1.3 to 4.4
and 1.0 to 5.8) increased mortality after adjustment for conventional
predictors. Squamous cell carcinoma with the same
high-risk signature had an adjusted risk of 1.1 (95% CI, 0.4 to
3.2) in one data set and 2.5 (95% CI, 1.1 to 5.8) in another.
Adenocarcinoma with the 50-gene signature from squamous
cell carcinoma had an elevated risk of 3.5 (95% CI, 1.4 to 9.0)
after adjustment for conventional predictors. Squamous cell
carcinoma with this high-risk signature had an adjusted risk
of 1.8 (95% CI, 0.7 to 4.6). The authors thus illustrated that
two nonoverlapping but functionally related gene expression
signatures provided consistently improved survival prediction
for NSCLC regardless of the histologic cell type.
Skrzypski et al. 146 studied the expression of 29 progression
and metastasis genes derived from previous lung cancer microarray
data. Their expression was assessed by reverse transcriptase
quantitative PCR in frozen primary tumor specimens obtained
from 66 SCC patients who had undergone surgical resection.
In a multivariate Cox model, the genes CSF1 (HR, 3.5;
p 0.005), EGFR (HR, 2.7; p 0.02), CA IX (HR, 0.2;
p 0.0001), and tumor size 4 cm (HR, 2.7; p 0.02)
emerged as significant markers for survival and the expression
of the three genes (CSF1, EGFR, and CA IX)as risk factors
was positively validated in a separate cohort of 26 patients in
an independent laboratory (p 0.05)
Raz et al. 147 generated a four-gene model based on expression
of WNT3a, ERBB3, LCK, and RND3 that was used to
generate a risk score. The Gene Risk score predicted mortality
better than clinical stage or tumor size (adjusted HR, 6.7; 95%
CI, 1.6 to 28.9; p 0.001). Among 70 patients with stage I
disease, 5-year overall survival was 87% among patients with
low-risk scores, and 38% among patients with high-risk scores
(p = 0.0002). Among all patients, 5-year overall survival was
62% and 41%, respectively (p 0.0054). Disease-free survival
was also significantly different among low- and high-risk score
patients.
Most recently, Shedden et al. 148 reported results from a
large retrospective multisite, blinded study of 442 lung adenocarcinomas
designed to test the prognostic performance of gene
microarray data alone versus performance with the inclusion of
clinical covariates. Data were generated using a common platform,
and training set data were generated at two of the participating
sites. The results were validated using independent data
generated from two additional sites after a blinded protocol.
Eight different methods, including gene clustering, univariate
testing, and mechanistic groupings, were used to provide prognostic
data. Many genes were identified as important for prognosis
in more than one statistical method, including those involved
in cell proliferation such as cyclins, checkpoint genes, topoisomerases,
and chromosomal and spindle protein genes. The most
successful classifier methods incorporated both gene expression
and clinical data, and the investigators stressed the importance of
coordinating the collection of both clinical and pathologic data
across multiple institutions for future prospective studies.