Embed Size (px)
Citation preview
Imag
LunAreto N
MatthChristAldenPhilip
Abst
Lunworldhistolmors
Author2DeparResearPatholoPhysiolOncoloNorth8DeparSalt La
Note: SResear
CorresChapelPhone:
doi: 10
©2010
Clin C4864
Clinical
Canceresearching, Diagnosis, Prognosis
g Squamous Cell Carcinoma mRNA Expression SubtypesReproducible, Clinically Important, and Correspond
R
ormal Cell Types
ew D. Wilkerson1, Xiaoying Yin1, Katherine A. Hoadley1,2, Yufeng Liu3,4, Michele C. Hayward1,opher R. Cabanski3, Kenneth Muldrew5, C. Ryan Miller1,5, Scott H. Randell1,6, Mark A. Socinski1,7,
M. Parsons7, William K. Funkhouser1,5, Carrie B. Lee1,7, Patrick J. Roberts1, Leigh Thorne1,5, S. Bernard8, Charles M. Perou1,2, and D. Neil Hayes1,7ractPur
diagnsubclament.patienExp
discov56 SCusingSubtyand bRes
were doutcopatienlogicabasal:system
ogic type(2, 3). SC
s' Affiliatitment of Gch, 4Carolgy and Labogy, and 7Dgy, MultidisCarolina atment of Pake City, Uta
upplementach Online (h
ponding AuHill, 450 W919-966-37
.1158/1078-
American A
ancer Res
pose: Lung squamous cell carcinoma (SCC) is clinically and genetically heterogeneous, and currentostic practices do not adequately substratify this heterogeneity. A robust, biologically based SCCssification may describe this variability and lead to more precise patient prognosis and manage-We sought to determine if SCC mRNA expression subtypes exist, are reproducible across multiplet cohorts, and are clinically relevant.erimental Design: Subtypes were detected by unsupervised consensus clustering in five publishedery cohorts of mRNA microarrays, totaling 382 SCC patients. An independent validation cohort ofC patients was collected and assayed by microarrays. A nearest-centroid subtype predictor was builtdiscovery cohorts. Validation cohort subtypes were predicted and evaluated for confirmation.pe survival outcome, clinical covariates, and biological processes were compared by statisticalioinformatic methods.ults: Four lung SCC mRNA expression subtypes, named primitive, classical, secretory, and basal,etected and independently validated (P < 0.001). The primitive subtype had the worst survival
me (P < 0.05) and is an independent predictor of survival (P < 0.05). Tumor differentiation andt sex were associated with subtype. The expression profiles of the subtypes contained distinct bio-l processes (primitive: proliferation; classical: xenobiotic metabolism; secretory: immune response;cell adhesion) and suggested distinct pharmacologic interventions. Comparison with lung models revealed distinct subtype to cell type correspondence.clusions: Lung SCC consists of four mRNA expression subtypes that have different survival
Conoutcomes, patient populations, and biological processes. The subtypes stratify patients for more preciseprognosis and targeted research. Clin Cancer Res; 16(19); 4864–75. ©2010 AACR.
keratiref. 4)
g cancer is the leading cause of cancer-related deathwide (1). Squamous cell carcinoma (SCC) is a major
and comprises ∼30% of all pulmonary tu-C is defined by the presence of cytoplasmic
and mPatienical ouhad resurvivmorpated ttion ocell, cand clcausevariabicallyRec
riety oprimamia (
ons: 1Lineberger Comprehensive Cancer Center,enetics, 3Department of Statistics and Operationsina Center for Genome Sciences, 5Department oforatory Medicine, 6Department of Cell and Molecularepartment of Internal Medicine, Division of Medicalciplinary Thoracic Oncology Program, University oft Chapel Hill, Chapel Hill, North Carolina andthology, University of Utah Health Sciences Center,h
ry data for this article are available at Clinical Cancerttp://clincancerres.aacrjournals.org/).
thor: D. Neil Hayes, University of North Carolina atest Drive, Campus Box 7295, Chapel Hill, NC 27599.86; Fax: 919-966-1587; E-mail: [email protected].
0432.CCR-10-0199
ssociation for Cancer Research.
; 16(19) October 1, 2010
nization and/or desmosomes (intracellular bridges;. Clinically, SCC tumors occur more often in smokersales compared with the other histologic types (2, 5).ts affectedwith SCC tumors show awide range of clin-tcomes. For instance, 83% of autopsied SCC patientsgionalmetastases (5) and 68% of SCC stage I patientsed beyond 5 years (6). Within SCC, there is noticeablehologic variability, especially among poorly differenti-umors (4, 7). The WHO SCC type includes a stratifica-f this variability with four variants (papillary, smalllear cell, and basaloid; ref. 4), but their prevalenceinical and biological significance remain unclear. Be-there is significant pathologic and clinical outcomeility within the SCC histologic type, a robust, biolog-derived subclassification may be valuable.ent years have seen progress in classification of a va-f malignancies using full-genome molecular assays,
rily those directed at mRNA expression [e.g., leuke-8), breast (9), and lung adenocarcinoma (10)]. A
succesdetectsion sas patin SCCsubtypdiscovbiologdifferesifiedevancthat vintrigution cin lunduciband bsurvivacterinorm
Mate
TumoFro
morsNorthunder0573basedfromthe froby fou
tificatand aproberays wand loas theCentenlm.n
PubliA st
pressipressidata sto pecross–correlto asBild ehttp:/(13),surgiccated.wereweremodesubjeccommlevel
UnsuThe
depicwith hto prereliabgenes0.1%in eacdeviatand r25%
Translational Relevance
Lung squamous cell carcinoma (SCC) has broadclinical, genetic, and morphologic heterogeneity. Cur-rently, there is no subclassification that adequately de-scribes this variability and SCC patients are basicallytreated as though they have the same disease. One ex-planation for SCC variability is that SCC is not a sin-gular disease but a mixture of multiple discretediseases or subtypes defined by innate biological dif-ferences. Using five discovery cohorts and an indepen-dent validation cohort totaling 438 patients, we showthat SCC is composed of four robust mRNA expressionsubtypes (named primitive, classical, secretory, andbasal). The subtypes have significantly different surviv-al outcomes, patient populations, and biological pro-cesses. Using these subtypes as a basis for a futureclinical diagnostic assay, patients could receive a moreprecise prognosis. Additionally, we described modelsystem partners for the subtypes that can be used fortargeted basic research.
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
sful approach is unsupervised class discovery, whichs naturally occurring tumor classes (“mRNA expres-ubtypes”) without prespecified characteristics suchient survival (8). Preliminary efforts have been made, suggesting the existence of SCC mRNA expressiones. In independent analyses, investigators (11–13)ered two mRNA expression subtypes with intriguingical profiles and a corresponding patient survivalnce. These studies show that SCC might be subclas-using mRNA expression into groups with clinical rel-e; however, the studies were not done in a manneralidated either the number or the nature of theseing classes. A validated mRNA expression classifica-ould substantially progress patient care and researchg SCC. In this study, we describe four novel repro-le expression subtypes (primitive, classical, secretory,asal) of lung SCC. The SCC subtypes have differental outcomes, patient demographics, physical char-
stics, biological processes, and correspondence to for clmediin eacalgoralgoripropoagglom1-PeavaluesameagglomEuclid
Subty
al lung cell types.
rials and Methods
r collectionzen, surgically extracted, macrodissected, primary tu-from treatment-naive patients at the University ofCarolina with a lung SCC diagnosis were collectedInstitutional Review Board approved protocols 90-and 07-0120. Morphologic quality control wason a review of a representative H&E-stained sectionparaffin-embedded tissue immediately adjacent to
zen tissue for confirmation of squamous histologyr pathologists (Supplementary Fig. S1) and for quan-Cenarray
acrjournals.org
ion of tumor content. Tumor RNA was extracted (14)ssayed for mRNA expression using Agilent 44,000microarrays for a total of 56 microarrays. Microar-ere processed by normexp background correctioness normalization (15). This data set is referred to“validation cohort” and was deposited at Nationalr for Biotechnology Information (http://www.ncbi.ih.gov/geo/query/acc.cgi?acc=GSE17710).
shed data setsructured search for publicly available SCC mRNA ex-on microarray data sets was conducted via Gene Ex-on Omnibus and PubMed and manually selectingets that have a large number of lung SCC samplesrmit subtype analysis and that have significantdata set gene reliability, as measured by integrativeations (16). This search yielded five data sets (referredthe “discovery cohorts”) from the following studies:t al. (17), Expression Project for Oncology (Expo;/www.intgen.org/expo/), Lee et al. (18), Raponi et al.andRoepman et al. (19). Published cohorts containedal resections from treatment-naive patients if indi-Clinical data and raw or processed microarray dataobtained. Only microarrays with SCC histologyretained. Raw microarrays or gene lists from lungl systems were obtained (20–23). Microarrays wereted to standard quality assessments, mapped to aon transcript database, and processed into gene-expression values (Supplementary Table S1).
pervised subtype discoverysubtype discovery and validation procedure is
ted in a flowchart (Supplementary Fig. S2). Genesigh reliability and variability were selected similarviously described methods (9, 10, 12, 13, 16). Geneility was measured by integrative correlations, andhaving an estimated false discovery rate (FDR) ofwere retained (16). To select variable genes, genesh discovery cohort were ranked by median absoluteion in decreasing order. These ranks were averagederanked to make a single, ranked gene list. The topof this ranked list, totaling 2,307 genes, was usedustering. Before clustering, each data set was genean centered (24, 25). Subtypes were determinedh discovery cohort by the Consensus Clusteringithm via ConsensusClusterPlus (26, 27). Thisthm completed 1,000 microarray subsamples at artion of 80% and clustered these subsamples by anerative average-linkage hierarchical algorithm using
rson correlation coefficient distance. Consensuss, the proportion that two microarrays occupy thecluster, were calculated and then clustered by anerative average-linkage hierarchical algorithm usingean distance.
pe summarization by centroids
troids are median expression profiles of a group ofs and were prepared using methods previouslyClin Cancer Res; 16(19) October 1, 2010 4865
descrigroupand otroidstaking
DifferDif
standatweenGeneMpackaeffectsusedranked
ValidSub
by a npreviowas bgeneserror rdifferetype timprolists,online
SurviThe
analyswere cand nmet thsurvivuntil
ImmuCor
tissueray blsembl(DakoMCM
Comsion 2librarispecif
Resu
UnsusubtyLun
therefanalys
sets todiseasanalyzusingprevioClustean opResult(Suppdenceclustefiles,lowedstructone cTherefoundexpreeachusedlishedtheseture adencesignifthe suamoncharaexprestory,
SCC sAlth
validahorts,coveryvalidatypesTo teusingand hUsingfor mmorsin thprevawhichvalidicharawas cWe cogenesthe duniquconcoacross
Wilkerson et al.
Clin C4866
bed (25, 28). Centroids were determined by taking aof microarrays from a gene median centered cohortbtaining the median of each gene. Multicohort cen-are determined by taking a group of centroids andthe median of each gene.
entially expressed genesferentially expressed genes were determined by ardized mean difference procedure that considers be-cohort and within-cohort variation (29) using theeta Bioconductor library (http://bioconductor.org/
ges/2.2/bioc/html/GeneMeta.html) and a randomoption. Gene set enrichment analysis (GSEA) was
to determine gene sets significantly enriched ingene lists (30).
ation cohort subtype predictiontype status of the validation cohort was predictedearest-centroid classification algorithm followingusly published methods (28). In brief, the predictoruilt, using only the discovery cohorts, by addingto a balanced centroid, assessing subtype predictionates by leave-one-out cross-validation, adding genesntially expressed from the most mispredicted sub-o its centroid, and stopping once accuracy failed tove. Subtype predictor centroids, unsupervised geneand all gene multicohort centroids are available(http://cancer.unc.edu/nhayes/publications/scc/).
val analysisR library survival was used for survival statisticales. Patients dead within 1 month following surgeryonsidered to have procedure-related complicationsot considered in survival analyses. Five patientsis condition all from the UNC cohort. Relapse-freeal (RFS) time was defined as the time from surgeryfirst relapse or death.
nohistochemistryes (1 mm) were taken from available UNC cohortblocks and randomly organized into tissue microar-ocks. Consequal 4-μm array block sections were as-ed on array slides and stained with H&E, MAC387), p63 (Dako), CK7 (Leica Microsystems), and6 (Santa Cruz Biotechnology).
putational procedures were executed using R ver-.7.1 (http://www.r-project.org/) and Bioconductores (http://www.bioconductor.org) unless otherwiseied.
lts
pervised discovery of lung SCC expressionpes in five cohortsg SCCs are a heterogeneous group of tumors, and
ore, we did a common set of mRNA expressiones using five previously published lung SCC datathat tyond
ancer Res; 16(19) October 1, 2010
determine how many distinct subtypes/groups ofe might exist. These five discovery cohorts wereed for the presence of mRNA expression subtypesthe Consensus Clustering methodology (26) asusly described for lung cancer (10). Consensusring is a semiquantitative method for determiningtimal number of mRNA expression clusters/groups.s show that all five cohorts contain four clusterslementary Fig. S3). There is no compelling evi-for a higher number of clusters. To test if the fourrs from each cohort have the same expression pro-a published centroid clustering method was fol-(10). The centroid clustering shows a four-group
ure, where each cohort is in each group, with onlyohort absent in one group (Supplementary Fig. S4).fore, the four clusters (mRNA expression subtypes)in the five discovery cohorts have consistent
ssion profiles. To derive the optimal subtype forpatient, a multicohort centroid classification wasto assign each patient to a subtype, similar to pub-methods (28). A centroid clustering based on
optimal subtypes again shows a four-group struc-nd complete, unambiguous cross-cohort correspon-(Fig. 1). The cross-cohort clustering is statistically
icant [Sigclust (31) P values in Fig. 1]. Interestingly,btypes have approximately the same prevalenceg the discovery cohorts (Table 1). Using biologicalcteristics described below, the lung SCC mRNAsion subtypes are named primitive, classical, secre-and basal.
ubtype independent validationough the four SCC subtypes were “cross-cohort”ted in that they were repeatedly found in five co-this validation was not independent because dis-co-occurred with validation. For an independent
tion, we tested the hypothesis that the SCC sub-will exist in a new discovery-independent cohort.st this hypothesis, a subtype predictor was builtthe discovery cohorts, which consisted of 208 genesad 94% leave-one-out cross-validation accuracy.this predictor, subtype classifications were madeicroarrays from a new cohort of 56 lung SCC tu-collected at UNC. All four subtypes were predictede UNC cohort and in approximately the samelence as the discovery cohorts (Fig. 2; Table 1),supports subtype reproducibility. To confirm the
ty of the predictions, a comparison of expressioncteristics between the discovery and UNC cohortsompleted similar to a recent related study (32).mpiled a large validation gene set of the top 100overexpressed and underexpressed per subtype ofiscovery cohorts (Fig. 2A), which yielded 1,117e genes. Subtype expression patterns are highlyrdant between the discovery and UNC cohortsthe validation gene set (Fig. 2A and B), confirming
he large-scale expression patterns are consistent be-the predictor gene set. In addition, the subtypes of
Clinical Cancer Research
the Uits mRrandofinedtherefTo p
kers cpressesecret
chemcroarrtargeting foclassicmutugroup
Fig. 1.Suppleshadingsmall 1at the texpression subtypes. Statistical significance of the three binary divisions leading to the four subtypes is shown by Sigclust (31) P values in the dendrogramat the c
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
NC cohorts are a statistically significant partition ofNA expression [SWISSMADE (33) subtypes versusm classes; P < 0.001]. We conclude that the prede-SCC subtypes exist in the UNC cohort and are,ore, independently validated.reliminarily evaluate if clinically applicable biomar-an distinguish the subtypes, we selected one overex-
orresponding binary split.
d gene per subtype (basal, S100A8; classical, TP63;ory, KRT7; primitive, MCM6) for immunohisto-
icantlexact
acrjournals.org
ical protein expression comparison using a tissue mi-ay subset of the UNC cohort (n = 38). All antibodiesing these genes, except MCM6, had sufficient stain-r analysis. Protein expression clustering using basal,al, and secretory samples revealed three essentiallyally exclusive groups with one marker defining each(Supplementary Fig. S5). These groups were signif-
Discovery cohort correlation matrix and dendrogram. Cells are labeled by discovery cohort and adjusted centroid, where A to D are frommentary Fig. S4. B, Cells in the matrix represent the 1-Pearson correlation coefficient between two discovery cohort and adjusted centroids byaccording to the scale above. For example, BildA and RoepmanA have highly similar expression profiles, a large Pearson correlation coefficient, a
-Pearson correlation coefficient value, and corresponding cells darkly shaded. A, the matrix is ordered by columns and rows by the dendrogramop of the matrix. The dendrogram is the result of an agglomerative, average-linkage, hierarchical clustering using this correlation matrix. C, four
y associated with tumor subtype (P = 0.007, Fisher'stest). This suggests that SCC subtypes can also be
Clin Cancer Res; 16(19) October 1, 2010 4867
distinworkcal an
SubtyTo
subtyprichmbindincauselapsed
Here,relativThe
type ias mintion fapolymenaturethe m
Tab
No.
Age
Gen
Smo
Stag
Gra
OS
% T
% N% FLym
NOTprimOS
Wilkerson et al.
Clin C4868
guished by immunohistochemistry, and futuremay find the optimal panel of immunohistochemi-tibodies.
pes exhibit distinct biological processesdiscern biological processes associated with eache, subtype mRNA expression was evaluated for en-ent in gene ontology, pathway, transcription factorg site, and cytoband gene sets by GSEA (30). Be-
of the inherent redundancy in biology, we have col-these processes into functional themes (Fig. 3).and itumo
le 1. Clinical characteristics
patients BiLe
Rap
de
ki
e
de
um
percents are Kaplan-Meier estimates. Gray shading indicates data una
ancer Res; 16(19) October 1, 2010
subtypes are described in terms of overexpressione to the other subtypes.distinctive functional theme of the primitive sub-s cellular proliferation, which includes genes suchichromosome maintenance 10 (MCM10), E2F transcrip-ctor 3 (E2F3), thymidylate synthetase (TYMS), andrase α1 (POLA1), and a published proliferation sig-(34). This proliferation theme is overexpressed inost rapidly growing breast cancer cell lines (35)
n the most poorly differentiated, poor survivalrs from various organ sites (34). Complementary toof lung SCC expression subtypes
vailability.
Clinica
l Cancer R esearPrimit
Disc
ive Classi
overy coh
cal Secre
orts
tory Basal
Total PrimitValida
tion cohor t (UNC)ive Classi
cal Secre tory Basal Totalld et al. 7
20 15 10 52 e et al. 14 30 22 9 75 Expo 4 15 11 6 36 oni et al. 20man et al. 154135
3219
3423
12792
RoepTotal 60
141 99 82 382 9 21 14 12 56M
edian 68 64 66 67 66 65 68 64 72 67r %%
Female 36Male 641981
2674
2971
2674
6733
3367
4357
4258
4357
ng % NMean
onsmoker 8pack-years 64
172
360
268
366
043
074
062
046
060
% I 66
58 66 55 61 56 57 57 75 61 % II 25 26 20 37 26 44 33 36 25 34 % III 5% IV 3170
140
90
131
00
100
70
00
50
%
Poor 39 15 21 16 21 56 24 43 33 36 % M%
oderate 58Well 3822
763
768
754
440
760
570
670
640
No.
patients 42 96 66 67 271 8 19 13 11 51 % 1% 3-y survival 64-y survival 47
89 84 88 84 88
100388248
9060
9241
63 59 71 62 15or MInterqu
edianartile range
90
804 60-10800 70-9
930 73-96
9060-95
75-9rosis M
edian 5 5 5 4 5 ecibrosis Median 15 13 18 10 10phocytes % Marked 33 31 40 40 36E: Percent values indicate the proportion of the samples in a particular subtype with a particular variable (e.g., 36% of theitive subtype samples came from female patients in the discovery cohort). Some percents may not total 100% due to rounding.
ch
the cethe E2lator (membtive suDNAeratioThe
themeeign cthemetomes
(37).signatretteInterecentrathe suAKR1pressein refSCC
Fig. 2.normaldiscovery cohort hierarchical clustering. The normal lung centroid is scaled to the validation cohort for visualization. Manually selected, lung-relevant,validati
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
llular proliferation functional theme, target genes ofF transcription factor, a known proliferation modu-36), are overexpressed in this subtype as well as twoers of the E2F family, E2F3 and E2F8. Other primi-btype functional themes are RNA processing andrepair, which could be a consequence of the prolif-n theme or an independent process.classical subtype exhibits the distinctive functionalof xenobiotic metabolism, which detoxifies for-
hemicals. One study showed overexpression of this
on genes are displayed separately for viewability.
in smokers' versus nonsmokers' airway transcrip-, including genes such as GPX2 and ALDH3A1
pressesquam
acrjournals.org
Furthermore, this subtype is enriched with a geneure derived from lung cell lines exposed to ciga-smoke, including genes such as AKR1C3 (38).stingly, the classical subtype has the greatest con-tion of smokers and the heaviest smokers amongbtypes. This theme, including genes such as GPX2,C1, TXNRD1, and GSTM3, was noted as overex-d in one head and neck SCC subtype (group 4. 39), suggesting a possible relative to the lungclassical subtype. The classical subtype overex-
Independent validation of lung SCC expression subtypes. Heat maps depict mRNA expression of discovery cohorts (A), the validation cohort (B), alung centroid (C), and SCC cell lines (D). Microarrays are columns and are labeled with their class. Genes are rows and are ordered by a
s TP63, a transcription factor essential for stratifiedous epithelium development (40) that is more
Clin Cancer Res; 16(19) October 1, 2010 4869
commcompgeneDNAtainscroarrmeasube aImm
themeas Rhonecrosthis ththemetype a
muci(SFTPthyroibe hipresseSCCthe gperhacretorvery sover tlectedevalu
Fig. 3. Srepresebindinglocation
Wilkerson et al.
Clin C4870
only overexpressed and amplified in lung SCCared with other histologic types (41). Cytobandoverexpression, a proxy for underlying genomicamplification, suggests that 3q27-28, which con-TP63, is amplified in the classical subtype. The mi-ays of this study do not have enough resolution tore TP63 isoform-specific expression, but this maygoal of future investigations.une response is the major distinctive functionalof the secretory subtype and includes genes suchGDP dissociation inhibitor β (ARHGDIB) and tumor
is factor receptor 14 (TNFRSF14). Consistent witheme, the secretory subtype has a NF-κB regulation
and NF-κB target gene overexpression. This sub- subtypsite refers to gene sets having a predicted transcription factor binding site. Cell. Drug targets are defined as overexpressed in all pairwise subtype comparison
ancer Res; 16(19) October 1, 2010
n (MUC1) and pulmonary surfactant proteinsC, SFTPB, and SFTPD; refs. 7, 42). Interestingly,d transcription factor 1 (NKX2-1/TTF1), known toghly expressed in adenocarcinoma (43), is overex-d in the secretory subtype relative to the othersubtypes. This commonality could be a result oflandular cell structure of adenocarcinoma, whichps has secretory properties similar to the SCC se-y subtype. A UNC normal lung centroid shows aimilar expression pattern to the secretory subtypehe independent validation gene list, which was se-without considering normal samples (Fig. 2C). To
ate any possible difference between the secretory
e samples and normal samples, an unsupervisedlso overexpresses the lung secretory cell markers clustering was completed using only these microarrays
ubtype biological functional themes. Significantly enriched gene sets that are overexpressed in a subtype (GSEA preranked, FDR < 0.05) and genesntative of the set are shown. Pathways and biological processes are organized into functional themes, indicated by italics. Transcription factor
ular component refers to gene sets having a particular cellulars (FDR < 0.01).
Clinical Cancer Research
(Supprays csuggeare diThe
hesionmininCOL1(CLDNdevelo(S100the gLAMCSCC ssubtypent orThe bgenesS100Aexpres(44).the lumarkerichedbasemIn p
ternscologiof anthe pway isthe seilarityexprepemelung cof sevprimi
SCC scharaThe
typesciatedprimiferentresentnot anotecreasehortssampltype-scent tinfiltrtypes,the suclass w
basalothat thPat
tion ware ovthe clsion phas thWalli(1%;vation
SCC sOve
ly difsubtypsubtypbasal,comessubtysubtyRFS: Plog-racohorthe ba(0/3).subtytors, umodeicantand Rgrademulticancethe tulate-stRap
typesferent“majogrouptheirshowsand seplemein thiseach othe SCly, thepatien(primof Rapcome
SCC stypes
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
lementary Fig. S6). Secretory and normal microar-lustered with their group in essentially all cases,sting that the secretory subtype and normal lungstinct mRNA expression groups.basal subtype expression profile shows a cell ad-functional theme, including genes such as the la-
s (LAMB3 and LAMC2), collagens (COL11A1 and7A1), integrins (ITGB4 and ITGB5), and claudin 11). Additionally, this subtype has an epidermalpment theme, including keratin 5 (KRT5), psoriasinA7), and gap junction protein β5 (GJB5). Several ofenes of the basal subtype, such as COL17A1,2, and CDH3, are common with a head and neckubtype (group 1 in ref. 39) and a breast cancere (basal-like in ref. 9), suggesting that these differ-gan site subtypes may share biological properties.asal subtype overexpresses several S100 family: S100A2, S100A3, S100A7, S100A8, S100A9,12, and S100A14. S100A8 and S100A9 are highlysed in the basal layer in psoriatic epidermal tissueS100A2 is a marker specific for the basal layer ofng epithelium and SCC (45). KRT5 is a basal layerr in epithelial tissue (46). The basal subtype is en-with genes whose products are localized in theent membrane.arallel to differential biological functions are pat-of mRNA expression with implications for pharma-c intervention (Fig. 3). For example, TYMS, a targettifolates including pemetrexed, is overexpressed inrimitive subtype. The antifolate metabolism path-differentially expressed among SCC subtypes, withcretory subtype showing underexpression and sim-to adenocarcinoma (Supplementary Fig. S7). Over-ssion of TYMS has been shown to be related totrexed resistance in a dose-dependent manner inancer cell culture (47). In addition, PARP1, a targeteral drugs in development, is overexpressed in thetive subtype.
ubtype tumor morphologic and patientcteristicsmorphologic and patient characteristics of the sub-are displayed in Table 1. Grade is significantly asso-with subtype (P = 0.024, Fisher's exact test). The
tive subtype has an overrepresentation of poorly dif-iated tumors, and the basal subtype has an overrep-ation of well-differentiated tumors. Tumor stage isppreciably different among subtypes, although wethat the classical and secretory subtypes have in-d proportions of stage III tumors. The surgical co-oversample early stages, and possibly, greatering of late-stage patients may find additional sub-tage associations. Specimen quality metrics of per-umor, percent necrosis, and percent lymphocyteation are not appreciably different among the sub-arguing against sampling artifacts as the source of
btypes. Two cases of WHO morphologic SCC sub-ere definitively called by pathologist review (oneTorived
acrjournals.org
id in primitive and classical subtypes), suggestingese SCC morphologic subclasses are rare.
ient sex approaches statistically significant associa-ith subtype (P = 0.058, Fisher's exact test). Femaleserrepresented in the primitive subtype and males inassical subtype. Consistent with the smoking expres-rofile of the classical subtype, the classical subtypee greatest mean pack-years (73; P = 0.319, Kruskal-s test) and the lowest proportion of nonsmokersP = 0.214, Fisher's exact test), although these obser-s do not meet statistical significance.
ubtypes have different patient survival outcomesrall survival (OS) and RFS outcomes are significant-ferent among SCC subtypes (Fig. 4). The primitivee has worse OS and RFS compared with the otheres in all stages and in stage I (Fig. 4), whereas thesecretory, and classical seem to have similar out-. Considering the UNC cohort alone, the primitivepe outcome is also worse compared with the otherpes over all stages (OS: P = 0.066, log-rank test;= 0.004, log-rank test) and stage I (OS: P = 0.057,
nk test; RFS: P = 0.007, log-rank test). In the UNCt, 7 of 18 recurrences were extrapulmonary andsal subtype had the lowest number and proportionTo evaluate the independent contribution of SCC
pe to patient risk in light of known prognostic fac-nivariate and multivariate Cox proportional hazardls were constructed (Supplementary Table S2). Signif-univariate predictors were primitive subtype for OSFS and tumor stage for OS. Patient age and tumorwere not significant predictors of either outcome. Inple variable models, only subtype retained signifi-for OS and RFS. The nonsignificant prediction ofmor stage may be due to the underrepresentation ofage patients across the cohorts.oni et al. reported two SCC mRNA expression sub-with a survival difference and provided a list of dif-ially expressed genes, where high expression of therity of the genes were downregulated in the high-risk” (13). Comparison of Raponi et al.'s microarrays bygene list and the subtypes discovered in this studytwo clear subtype groups: underexpression (primitivecretory) and overexpression (basal and classical; Sup-ntary Fig. S8). Therefore, the four subtypes discoveredstudy map to prior results and this study has dividedf the prior subtypes into two new ones and improvesC mRNA expression subtype granularity. Interesting-Raponi et al. poor survival subtype totals 43% of theirts, where the poor survival subtype of this studyitive) is 16% of their patients. It seems that a fractiononi et al.'s high-risk subtype shows poor survival out-relative to the remainder of SCC.
ubtypes are similar to different normal lung celland SCC cell lines
evaluate the hypothesis that SCC subtypes are de-from different cell types present in the normal lung,Clin Cancer Res; 16(19) October 1, 2010 4871
SCC sthreelung dtractesimilabetweel. Thearly-similamodeface cnormwhichand la(22).the eainantshowwhichmodementscontaiof noexpreglandthat a
subtypcific luand ctypesmodesimilaIn a
correslish adtionscorrespressiwerepredicto betypesthe vaare coprediccal anbasal
Discu
. 4.typtheavasa
n thtol., RUNpmalueluatvival and subtype.
Wilkerson et al.
Clin C4872
ubtypes were compared by mRNA expression topublished model systems. The first model, “Mouseevelopment,” is a time series of mouse lungs ex-
d from embryonic stages to adult (21). Expressionrity is defined as high positive Pearson correlationen an SCC subtype and time points within the mod-e primitive subtype shows expression similarity tostage mouse lung, and the secretory subtype showsrity to late-stage mouse lung (Fig. 5A). The secondl, “Human bronchial epithelial cell air liquid inter-ulture” (HBEC-ALIC), is a time series of culturedal, healthy, human bronchial epithelial cells, inthe early time points consist of stratified basal cellster time points include secretory and ciliated cellsThe basal subtype showed expression similarity torly time points during which basal cells are predom-(Fig. 5B). The primitive and secretory subtypesexpression similarity to the later time points atthere are secretory and ciliated cells. The third
l system, “Human microdissected lung cell compart-” (HMLCC), was laser capture microdissected cellsned in surface epithelium and in submucosal glandsrmal healthy lung (20). The secretory subtype over-sses genes that are overexpressed in submucosal
s (Fig. 5C). The basal subtype overexpresses genesre overexpressed in surface epithelia. The classicalThethat lu
ancer Res; 16(19) October 1, 2010
e does not show appreciable similarity to any spe-ng model, is the only subtype to have this property,ould be most similar to multiple or unobserved cell. Therefore, by the combination of all three lungls, three of the four SCC subtypes have uniquerities to different, normal lung cell types.ddition to the cell type models, SCC subtypes maypond to different SCC cell lines, which could estab-ditional manipulatable models for future investiga-into subtype biology. To ascertain if SCC cell linespond to different SCC tumor subtypes by mRNA ex-on, four published SCC cell line microarrays (23)given subtype classifications by the nearest-centroidtor. Interestingly, the four cell lines were predicteddifferent subtypes (Fig. 2D). Expression of the sub-between the cell lines and tumors is consistent overlidation gene set (Fig. 2A and D). For example, genesnsistent and mutually exclusive in the cell lines asted (HCC15, primitive and MCM10; HCC95, classi-d AKR1C3; HCC2450, secretory and MUC1; H157,and MMP13).
ssion
FigsubbytheThethadueet aandRoeP vevasur
principal novel hypothesis tng SCC expression subtypes e
Clini
Survival outcomes of SCCes. Survival was estimatedKaplan-Meier method usingilable data of all cohorts.mple sizes (n) are differente overall study sample sizedata availability (OS: Bildaponi et al., Roepman et al.,C cohorts; RFS: Lee et al.,an et al., and UNC cohorts).s are from log-rank testsing the independence of
ested in this study isxist, are reproducible,
cal Cancer Research
are clwith u(primin anby croby indgetherso fouUnitesizes fthe saclassictypestientamondictorcludesuch asix cowouldcohorThe
and shtions.the SCcharaand eand Kwhosesimilailarity
tory suas surtory cmucosubtypworsttientspoorlearlyentiattent wThe pHBECexpresdurinlungcell tybryontureshighegreateTP63,locusThe
ties coferentbe cetion sscenar
Fig. 5.differentime is1,000 gto the sIn A, “eoverexp
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
inically relevant, and exhibit patterns that correlatenique cell types in the normal lung. These subtypesitive, basal, secretory, and classical) were identifiedunbiased and objective manner and are supportedss-cohort validation using five training cohorts andependent validation using a sixth cohort, which to-total 438 patients. The expression subtypes were al-nd in a wide variety of patient populations from thed States, Asia, and Europe, in a wide variety of cohortrom 36 to 127. All cohorts showed approximatelyme subtype proportions (overall: primitive, 16%;al, 37%; secretory, 26%; basal, 21%). These sub-were associated with tumor differentiation and pa-sex. Survival outcomes are significantly differentg the subtypes, and subtype is an independent pre-of survival. Possible limitations of our analysis in-possible sample quality artifacts or patient behavior,s smoking immediately before surgery; however, allhorts showed the same results, so any limitationhave to occur in six large, independently collected
ts.SCC expression subtypes are biologically distinctow similarities to distinct normal lung cell popula-These biological characteristics serve as the basis forC nomenclature. The basal subtype exhibits many
cteristics of lung basal cells, such as cell adhesionpidermal development functional themes, S100A2RT5 basal cell markers, overexpression of genesproducts are localized in the basement membrane,
rity to basal cells in the HBEC-ALIC model, and sim- dram” refers to embryonic day and “p” refers to postnatal day. C, the model HMLCC (20ressed in submucosal glands and in surface epithelium as rows and subtype ce
acrjournals.org
btype has many features of lung secretory cells, suchfactant and mucin overexpression, similarity to secre-ells in the HBEC-ALIC model, and similarity to sub-sal glands in the HMLCC model. The primitivee has a cellular proliferation functional theme, thesurvival outcome, an overabundance of female pa-, the most nonsmokers, and an overabundance ofy differentiated tumors. This subtype is similar toembryonic mouse lungs, where primitive, less differ-ed cells may be predominant and would be consis-ith the poorly differentiated nature of these tumors.rimitive subtype also has similarity to late-stage-ALIC, which could be explained by lung “transientsion” in which differentiation markers are expressedg early lung formation and again in the developed(48). Alternatively, a late-emerging and late-activepe in HBEC-ALIC may be most similar to the em-ic mouse lung. The classical subtype exhibits fea-representative of typical lung SCC, including thest prevalence at 37%, overabundance of males,st patient smoking behavior, overexpression ofand putative amplification of the TP63-containing3q27-28.distinct SCC subtype to cell population similari-uld be explained by the SCC subtypes having dif-ancestor cells. These different ancestor cells couldll types of distinct lineages or cellular differentia-tages such as proposed in breast cancer (49). Thisio provides a reason why the SCC subtypes have
atically different mRNA expression. The subtypesto surface epithelia in the HMLCC model. The secre- could arise by genetic mutation from different ancestors
SCC subtypes compared with lung cell type models. The relationship of relative SCC subtype expression differences to relative expressionces of published lung model systems. A and B, the models Mouse lung development (21) and HBEC-ALIC (22) are microarray time series, whereindicated on the horizontal axis. Points mark Pearson correlation coefficients of SCC subtype centroids to model time points using the topenes having the greatest Pearson correlation coefficient with time. Bars represent 95% confidence intervals. Lines connect points correspondingame subtype. Large positive correlations indicate mRNA expression similarity, whereas large negative correlations indicate dissimilarity.
) is compared with SCC subtypes via a heat map of genes that arentroids in columns.
Clin Cancer Res; 16(19) October 1, 2010 4873
that hmRNAThis pusedventiotypes.cell pcausedreflecthas mand fof thefurthesubtypThe
care athe prly bectype oby uscancemodeSCC s
ies anselectIn
subtypvalidatypica
Discl
D.Nis relatfinancia
Grant
NIHLinebergram (K12-RRthrough
Theof pageaccorda
Refe1. Pa
CA2. Ko
oftom
3. Visnoagndis
4. Traand
5. AurelaCh
6. JocanedW.
7. Chof t
8. Gocanmo
9. Pebre
10. Harevtip
11. Inasqanma
12. Larfinecin
13. Radiclun
14. Hu
Wilkerson et al.
Clin C4874
ave different mRNA expression, and this ancestralexpression could persist in progeny tumor cells.
utative subtype ancestral cell information could bein developing SCC subtype pharmacologic inter-ns that exploit differences in the ancestral cellA caveat to our interpretation of SCC subtype toopulation similarity is that the similarity could beby coincidence and expression similarities couldsimilar biology and not similar origin. The lungultiple proposed cellular development pathways,uture studies that describe the molecular profileslung cell types or lung cancer stem cells would
r clarify the putative ancestral cells of the SCCes (50).SCC subtypes may have applications in patientnd in cancer research. For instance, patients withimitive subtype could be treated more aggressive-ause of the poor survival expectation of this sub-r could be given a more accurate prognosis thaning traditional prognostic factors alone. Basicr research could be conducted using the subtype
bed in this study. The
ting prognosis of squamous cell and adenocarcinomas of theg. Cancer Res 2006;66:7466–72.Z, Troester M, Perou CM. High reproducibility using sodium
hyniq
15. Ritco23
16. GaCrmi
17. Bilhu35
18. LeintegCa
19. Ro72ce
20. FisMcwaBio
21. Maopma
22. RoofRe
23. Zhcleca
24. Go20
25. SotumAc
26. Moresge
ancer Res; 16(19) October 1, 2010
d possibly serve as a foundation for clinical trialion.conclusion, we identified four robust expressiones of lung SCC using a multicohort discovery andtion strategy. The subtypes are clinically and pheno-lly different, suggesting different therapies.
osure of Potential Conflicts of Interest
. Hayes, C.M. Perou, and P.S. Bernard hold a provisional patent thated to work described in this manuscript but there is no currentl interest. All other authors have no conflicts of interest.
Support
/National Cancer Institute grant F32CA142039 (M.D. Wilkerson),ger Comprehensive Cancer Center Translational Small Grants Pro-D.N. Hayes), NIH/National Center for Research Resources grant023248 (D.N. Hayes), and Thomas G. Labrecque FoundationJoan's Legacy Foundation (D.N. Hayes).
costs of publication of this article were defrayed in part by the paymentcharges. This article must therefore be hereby marked advertisement innce with 18 U.S.C. Section 1734 solely to indicate this fact.
l system partners descri
ubtypes could be useful for therapy benefit stud-Received 01/22/2010; revised 05/19/2010; accepted 06/27/2010;published OnlineFirst 07/19/2010.
rencesrkin DM, Bray F, Ferlay J, Pisani P. Global cancer statistics, 2002.Cancer J Clin 2005;55:74–108.yi H, Hillerdal G, Branden E. A prospective study of a total materiallung cancer from a county in Sweden 1997-1999: gender, symp-s, type, stage, and smoking habits. Lung Cancer 2002;36:9–14.bal AL, Williams BA, Nichols FC III, et al. Gender differences inn-small-cell lung cancer survival: an analysis of 4,618 patients di-osed between 1997 and 2002. Ann Thorac Surg 2004;78:209–15,cussion 15.vis WD, World Health Organization. Histological typing of lungpleural tumours. Berlin: Springer; 1999.
erbach O, Garfinkel L, Parks VR. Histologic type of lung cancer intion to smoking habits, year of diagnosis and sites of metastases.est 1975;67:382–7.nes DR, Detterbeck FC. Surgery for stage I non-small cell lungcer. In: Detterbeck FC, Socinski MA, Rivera MP, Rosenman JG,
itors. Diagnosis and treatment of lung cancer. 1st ed. Philadelphia:B. Saunders Company; 2001, p. 177–90.urg AM, Myers JL, Tazelaar HD, Wright JL. Thurlbeck's pathologyhe lung. 3rd ed. New York: Thieme Medical Publishers, Inc.; 2005.lub TR, Slonim DK, Tamayo P, et al. Molecular classification ofcer: class discovery and class prediction by gene expressionnitoring. Science 1999;286:531–7.rou CM, Sorlie T, Eisen MB, et al. Molecular portraits of humanast tumours. Nature 2000;406:747–52.yes DN, Monti S, Parmigiani G, et al. Gene expression profilingeals reproducible human lung adenocarcinoma subtypes in mul-le independent patient cohorts. J Clin Oncol 2006;24:5079–90.mura K, Fujiwara T, Hoshida Y, et al. Two subclasses of lunguamous cell carcinoma with different gene expression profilesd prognosis identified by hierarchical clustering and non-negativetrix factorization. Oncogene 2005;24:7105–13.sen JE, Pavey SJ, Passmore LH, et al. Expression profiling de-s a recurrence signature in lung squamous cell carcinoma. Car-ogenesis 2007;28:760–6.poni M, Zhang Y, Yu J, et al. Gene expression signatures for pre-
droxide-stripped long oligonucleotide DNA microarrays. Biotech-ues 2005;38:121–4.chie ME, Silver J, Oshlack A, et al. A comparison of backgroundrrection methods for two-colour microarrays. Bioinformatics 2007;:2700–7.rrett-Mayer E, Parmigiani G, Zhong X, Cope L, Gabrielson E.oss-study validation and combined analysis of gene expressioncroarray data. Biostatistics 2008;9:333–54.d AH, Yao G, Chang JT, et al. Oncogenic pathway signatures inman cancers as a guide to targeted therapies. Nature 2006;439:3–7.e ES, Son DS, Kim SH, et al. Prediction of recurrence-free survivalpostoperative non-small cell lung cancer patients by using an in-rated model of clinical information and gene expression. Clinncer Res 2008;14:7397–404.epman P, Jassem J, Smit EF, et al. An immune response enriched-gene prognostic profile for early-stage non-small-cell lung can-r. Clin Cancer Res 2009;15:284–90.cher AJ, Goss KL, Scheetz TE, Wohlford-Lenane CL, Snyder JM,Cray PB, Jr. Differential gene expression in human conducting air-y surface epithelia and submucosal glands. Am J Respir Cell Moll 2009;40:189–99.riani TJ, Reed JJ, Shapiro SD. Expression profiling of the devel-ing mouse lung: insights into the establishment of the extracellulartrix. Am J Respir Cell Mol Biol 2002;26:541–8.ss AJ, Dailey LA, Brighton LE, Devlin RB. Transcriptional profilingmucociliary differentiation in human airway epithelial cells. Am Jspir Cell Mol Biol 2007;37:169–85.ou BB, Peyton M, He B, et al. Targeting ADAM-mediated ligandavage to inhibit HER3 and EGFR pathways in non-small cell lungncer. Cancer Cell 2006;10:39–50.llub J, Sherlock G. Clustering microarray data. Methods Enzymol06;411:194–213.rlie T, Tibshirani R, Parker J, et al. Repeated observation of breastor subtypes in independent gene expression data sets. Proc Natl
ad Sci U S A 2003;100:8418–23.
nti S, Tamayo P, Mesirov J, Golub T. Consensus clustering: aampling-based method for class discovery and visualization ofne expression microarray data. Machine Learning 2003;52:91–118.Clinical Cancer Research
27. Witoo201
28. Hucon
29. Chies1:i8
30. Suanaexp
31. Liuclu20
32. VeanchNF
33. CaClaEle
34. Whpro
35. PesioPro
36. PoTre
37. Sphum200
38. JorsmproCa
39. Chheex
40. Kinsta
41. Maofan
42. GaepGaOx
43. Gapre20
44. BroduCy
45. Smprean91
46. Chpla
47. Ozsy10
48. WWcolin
49. Pra
Lung Squamous Cell Carcinoma mRNA Expression Subtypes
www.a
lkerson MD, Hayes DN. ConsensusClusterPlus: a class discoveryl with confidence assessments and item tracking. Bioinformatics0;26:1572–3.Z, Fan C, Oh DS, et al. The molecular portraits of breast tumors areserved across microarray platforms. BMC Genomics 2006;7:96.oi JK, Yu U, Kim S, Yoo OJ. Combining multiple microarray stud-and modeling interstudy variation. Bioinformatics 2003;19 Suppl4–90.bramanian A, Tamayo P, Mootha VK, et al. Gene set enrichmentlysis: a knowledge-based approach for interpreting genome-wideression profiles. Proc Natl Acad Sci U S A 2005;102:15545–50.YH, Hayes DN, Nobel A, Marron JS. Statistical significance ofstering for high-dimension, low-sample size data. J Am Stat Assoc07:103.rhaak RGW, Hoadley KA, Purdom E, et al. Integrated genomicalysis identifies clinically relevant subtypes of glioblastomaaracterized by abnormalities in PDGFRA, IDH1, EGFR, and1. Cancer Cell 2010;17:98–110.banski CR, Qi Y, Yin X, et al. SWISS MADE: Standardized WithInss Sum of Squares to Evaluate Methodologies and Datasetments. PLoS One 5:e9905.itfield ML, George LK, Grant GD, Perou CM. Common markers ofliferation. Nat Rev Cancer 2006;6:99–106.rou CM, Jeffrey SS, van de Rijn M, et al. Distinctive gene expres-n patterns in human mammary epithelial cells and breast cancers.c Natl Acad Sci U S A 1999;96:9212–7.lager S, Ginsberg D. E2F—at the crossroads of life and death.nds Cell Biol 2008;18:528–35.ira A, Beane J, Shah V, et al. Effects of cigarette smoke on thean airway epithelial cell transcriptome. Proc Natl Acad Sci U S A4;101:10143–8.gensen E, Stinson A, Shan L, Yang J, Gietl D, Albino AP. Cigarette
oke induces endoplasmic reticulum stress and the unfoldedtein response in normal and malignant human lung cells. BMCncer 2008;8:229.Na50. Sn
co
acrjournals.org
ung CH, Parker JS, Karaca G, et al. Molecular classification ofad and neck squamous cell carcinomas using patterns of genepression. Cancer Cell 2004;5:489–500.g KE,WeinbergWC. p63: defining roles inmorphogenesis, homeo-sis, and neoplasia of the epidermis. Mol Carcinog 2007;46:716–24.ssion PP, Taflan PM, Jamshedur Rahman SM, et al. Significancep63 amplification and overexpression in lung cancer developmentd prognosis. Cancer Res 2003;63:7113–21.illard D, Puchelle E. Differentiation and maturation of airwayithelial cells: role of extracellular matrix and growth factors. In:ultier C, Bourbon JR, Post M, editors. Lung development. Oxford:ford University Press; 1999, p. 46–76.rber ME, Troyanskaya OG, Schluens K, et al. Diversity of gene ex-ssion in adenocarcinoma of the lung. Proc Natl Acad Sci U S A01;98:13784–9.ome AM, Ryan D, Eckert RL. S100 protein subcellular localizationring epidermal differentiation and psoriasis. J Histochemtochem 2003;51:675–85.ith SL, Gugger M, Hoban P, et al. S100A2 is strongly ex-ssed in airway basal cells, preneoplastic bronchial lesionsd primary non-small cell lung carcinomas. Br J Cancer 2004;:1515–24.u PG, Weiss LM. Keratin expression in human tissues and neo-sms. Histopathology 2002;40:403–39.asa H, Oguri T, Uemura T, et al. Significance of thymidylatenthase for resistance to pemetrexed in lung cancer. Cancer Sci1:161–6.uenschell CW, Sunday ME, Singh G, Minoo P, Slavkin HC,arburton D. Embryonic mouse lung epithelial progenitor cells-express immunohistochemical markers of diverse mature celleages. J Histochem Cytochem 1996;44:113–23.t A, Perou CM. Mammary development meets cancer genomics.t Med 2009;15:842–4.
yder JC, Teisanu RM, Stripp BR. Endogenous lung stem cells andntribution to disease. J Pathol 2009;217:254–64.Clin Cancer Res; 16(19) October 1, 2010 4875
