Pan-Cancer Molecular Classes Transcending Tumor Lineage … · Biology of Human Tumors Pan-Cancer Molecular Classes Transcending Tumor Lineage Across 32 Cancer Types, Multiple Data

Biology of Human Tumors

Pan-Cancer Molecular Classes TranscendingTumor Lineage Across 32 Cancer Types, MultipleData Platforms, and over 10,000 CasesFengju Chen1, Yiqun Zhang1, Don L. Gibbons2,3, Benjamin Deneen4,5,6,7,David J. Kwiatkowski8,9, Michael Ittmann10, and Chad J. Creighton1,11,12,13

Abstract

Purpose: The Cancer Genome Atlas data resources representan opportunity to explore commonalities across cancer typesinvolving multiple molecular levels, but tumor lineage andhistology can represent a barrier in moving beyond differencesrelated to cancer type.

Experimental Design:On the basis of gene expression data, weclassified 10,224 cancers, representing 32 major types, into 10molecular-based "classes." Molecular patterns representing tissueor histologic dominant effects were first removed computation-ally, with the resulting classes representing emergent themesacross tumor lineages.

Results: Key differences involving mRNAs, miRNAs, proteins,and DNA methylation underscored the pan-cancer classes. Oneclass expressing neuroendocrine and cancer-testis antigenmarkersrepresented �4% of cancers surveyed. Basal-like breast cancerssegregated into an exclusive class, distinct from all other cancers.Immune checkpoint pathway markers and molecular signatures

of immune infiltrates were most strongly manifested within aclass representing �13% of cancers. Pathway-level differencesinvolving hypoxia, NRF2-ARE, Wnt, and Notch were manifestedin two additional classes enriched for mesenchymal markers andmiR200 silencing.

Conclusions: All pan-cancer molecular classes uncoveredhere, with the important exception of the basal-like breastcancer class, involve a wide range of cancer types and wouldfacilitate understanding the molecular underpinnings ofcancers beyond tissue-oriented domains. Numerous biolog-ical processes associated with cancer in the laboratorysetting were found here to be coordinately manifestedacross large subsets of human cancers. The number ofcancers manifesting features of neuroendocrine tumors maybe much higher than previously thought, which diseaseis known to occur in many different tissues. Clin Cancer Res;24(9); 2182–93. �2018 AACR.

IntroductionCancer is not a single disease, and at the molecular level

there is widespread heterogeneity that may be observed frompatient to patient. Nevertheless, unsupervised classification oftumors on the basis of molecular profiling data can revealmajor subtypes existing within a given cancer type according totissue of origin (1, 2). Such molecular-based subtypes canreflect altered pathways within different cancer subsets, whichcould have important implications for applying existing ther-apies or for developing new therapeutic approaches (3). TheCancer Genome Atlas (TCGA), a large-scale effort to compre-hensively characterize over 10,000 human cancers at the molec-ular level, provides a common platform for the study of diversecancer types, with multiple levels of data including mRNA,miRNA, protein, DNA methylation, copy number, and muta-tion (2). For most cancer types represented in TCGA, anindividual study of the molecular landscape of that cancertype was carried out (2).

With data generation completed, there is opportunity forsystematic analyses of the entire TCGA pan-cancer cohort (4),including defining molecular subtypes and associated pathwaysrelevant to multiple cancer types. One challenge, in the iden-tification of cancer subsets transcending the tissue of origin,is that widespread molecular patterns are associated withtumor lineage and histology (5–7). Although TCGA datasetshave been harmonized to allow for cross-cancer type

1Dan L. Duncan Comprehensive Cancer Center Division of Biostatistics, BaylorCollege of Medicine, Houston, Texas. 2Department of Thoracic/Head and NeckMedical Oncology, The University of Texas MD Anderson Cancer Center, Hous-ton, Texas. 3Department of Molecular and Cellular Oncology, The University ofTexas MD Anderson Cancer Center, Houston, Texas. 4Center for Cell and GeneTherapy, Baylor College of Medicine, Houston, Texas. 5Department of Neuro-science, Baylor College of Medicine, Houston, Texas. 6Neurological ResearchInstitute at Texas' Children's Hospital, Baylor College of Medicine, Houston,Texas. 7Program in Developmental Biology, Baylor College of Medicine, Hous-ton, Texas. 8The Eli and Edythe L. Broad Institute of Massachusetts Institute ofTechnology and Harvard University, Cambridge, Massachusetts. 9Brigham andWomen's Hospital and Harvard Medical School, Boston, Massachusetts.10Department of Pathology& Immunology, Baylor College ofMedicine, Houston,Texas. 11Department of Bioinformatics and Computational Biology, The Univer-sity of Texas MD Anderson Cancer Center, Houston, Texas. 12Human GenomeSequencingCenter, Baylor College ofMedicine, Houston, Texas. 13Department ofMedicine, Baylor College of Medicine, Houston, Texas.

Note: Supplementary data for this article are available at Clinical CancerResearch Online (http://clincancerres.aacrjournals.org/).

F. Chen and Y. Zhang are co-first authors.

Corresponding Author: Chad J. Creighton, Baylor College of Medicine, OneBaylor Plaza, Houston, TX 77030. Phone: 713 798 2264; Fax: 713 798 2716; E-mail:[email protected]

doi: 10.1158/1078-0432.CCR-17-3378

�2018 American Association for Cancer Research.

ClinicalCancerResearch

Clin Cancer Res; 24(9) May 1, 20182182

on October 31, 2020. © 2018 American Association for Cancer Research. clincancerres.aacrjournals.org Downloaded from

Published OnlineFirst February 9, 2018; DOI: 10.1158/1078-0432.CCR-17-3378

http://crossmark.crossref.org/dialog/?doi=10.1158/1078-0432.CCR-17-3378&domain=pdf&date_stamp=2018-4-6

http://clincancerres.aacrjournals.org/

comparisons—which is useful for addressing a host of ques-tions—an alternative approach to molecular classification onthe basis of these data would be to first computationallysubtract the molecular differences between cancer types (8,9). This alternative approach would have the effect of consol-idating the individual subtypes that might be discoverable inindividual cancer types into super-types or pan-cancer "classes"that transcend tissue or histology distinctions.

The aimof this studywas to define pan-cancermolecular-basedsubtypes or "classes," which would transcend tumor lineageacross theover 10,000human cancers and32 cancer types profiledby TCGA, using data frommultiple molecular profiling platformsto characterize these classes in terms of associated pathways.

Materials and MethodsResults are based upon data generated by TCGA Research

Network (http://cancergenome.nih.gov/). Molecular data wereaggregated frompublic repositories. Tumors spanned 32 differentTCGA projects, each project representing a specific cancer type,listed as follows: LAML, acute myeloid leukemia; ACC, adreno-cortical carcinoma; BLCA, bladder urothelial carcinoma; LGG,lower grade glioma; BRCA, breast invasive carcinoma; CESC,cervical squamous cell carcinoma and endocervical adenocarci-noma; CHOL, cholangiocarcinoma; CRC, colorectal adenocarci-noma (combining COAD and READ projects); ESCA, esophagealcarcinoma; GBM, glioblastoma multiforme; HNSC, head andneck squamous cell carcinoma; KICH, kidney chromophobe;KIRC, kidney renal clear cell carcinoma; KIRP, kidney renalpapillary cell carcinoma; LIHC, liver hepatocellular carcinoma;LUAD, lung adenocarcinoma; LUSC, lung squamous cell carci-noma;DLBC, lymphoid neoplasmdiffuse large B-cell lymphoma;MESO, mesothelioma; OV, ovarian serous cystadenocarcinoma;PAAD, pancreatic adenocarcinoma; PCPG, pheochromocytomaand paraganglioma; PRAD, prostate adenocarcinoma; SARC,sarcoma; SKCM, skin cutaneous melanoma; STAD, stomach ade-nocarcinoma; TGCT, testicular germ cell tumors; THYM, thy-moma; THCA, thyroid carcinoma; UCS, uterine carcinosarcoma;UCEC, uterine corpus endometrial carcinoma; UVM, uveal mel-

anoma. Cancer molecular profiling data were generated throughinformed consent as part of previously published studies andanalyzed in accordance with each original study's data use guide-lines and restrictions.

For the mRNA platform, log2-transformed expression valueswithin each cancer type (as defined by TCGA project) werenormalized to standard deviations from the median. By k-means clustering method (using the "kmeans" R function, with10 clusters and 1000 maximum iterations and 25 random sets),cancer cases were subtyped, based on the top 2,000 featureswith the most variable expression on average by cancer type(using standard deviation of the log2-transformed expressionvalues as a measure of variability, excluding genes on X or Ychromosomes).

We defined the top differential genes associated with eachsubtype, or pan-cancer "class." Taking the top 2,000 mRNAfeatures, we first computed the two-sided t test for each gene andeach class, comparing expression levels of each class with that ofthe rest of the tumors.We then selected the top 100 genes with thelowest P-value for each subtype; however, for the c2 class onlythree geneswere associated, and for the c7 class only 51 geneswereassociated, resulting in 854 top class-specific genes in all. In asimilar manner, top features associated with pan-cancer class forRPPA, miRNA, and DNA methylation platforms were identified.

The Fantom datasets of gene expression by cell type (10) wereanalyzed using a previously utilized approach (7). Briefly, the top2,000 most variable mRNAs (used above for the clustering anal-yses) were examined in Fantom. Logged expression values foreach gene in the fantom dataset were centered on the median ofsample profiles. For each fantom differential expression profile,the inter-profile correlation (Spearman's) was taken with that ofeach TCGA pan-cancer differential expression profile (with genesnormalized within each TCGA project to standard deviationsfrom the median).

We examined an external gene expression profiling dataset ofmultiple cancer types from the Expression Project for Oncology(expO; GSE2109), classifying each external tumor profile by pan-cancer class as defined by TCGA data. Within each cancer type,log2-transformed genes in the expO dataset were normalized tostandard deviations from themedian. As a classifier, the top set of854 mRNAs distinguishing between the TCGA pan-cancer classeswas used. For each pan-cancer class, the average value for eachgene was computed, based on the centered TCGA expression datamatrix. The Pearson's correlation between each expO profile andeach TCGA pan-cancer class averaged profile was computed. EachexpO casewas assigned to TCGApan-cancer class, based onwhichclass profile showed the highest correlation with the given expOprofile.

All P values were two-sided unless otherwise specified. Scoringof expression profiles for pathway-associated gene signatures wascarried out as previously described (9, 11). Additional methodsdetails are provided in supplemental. Supplementary Fig. S1provides a schematic of the various analyses performed and theirrelation to supplementary data files.

ResultsPan-cancer molecular subtypes primarily driven by tumorlineage

Across cancer types, tumor lineage and squamous histologywere found to represent major determinates of unsupervised

Translational Relevance

Unsupervised molecular classification of tumors can revealmajor subtypes existing within a given cancer type as definedby tissue of origin. Such molecular-based subtypes can reflectdifferent pathways at work within different cancer subsets,which couldhave important implications for applying existingtherapies or for developing new therapeutic approaches. Herewe applied an alternative classification approach, in order toconsolidate the individual subtypes thatmight bediscoverablein individual cancer types into super-types or pan-cancer"classes" that transcend tissue or histology distinctions. Coor-dinate pathways and processes were revealed across the tenpan-cancer classes in our cohort. As reflected in these classes,the tumor microenvironment may influence cancer in differ-ent ways between distinct subsets of human tumors. Ourmolecular class manifesting a differential expression profileof neuroendocrine tumors would have therapeutic implica-tions for an appreciable subset of human cancers.

Pan-Cancer Molecular Classes

www.aacrjournals.org Clin Cancer Res; 24(9) May 1, 2018 2183



http://cancergenome.nih.gov/


molecular classification. In total, our study involved 11,232human cancer cases representing 32 different major types, forwhich TCGA generated data on one or more of the followingmolecular characterization platforms (Supplementary Data S1):whole exome sequencing (WES, n ¼ 10,224 cases), somaticDNA copy by SNP array (n ¼ 10,845), RNA-seq (n ¼ 10,224),miRNA-seq (n ¼ 10,128), DNA methylation (n ¼ 10,959), andreverse phase protein array (RPPA, n ¼ 7,663). Using establishedanalytical approaches (5–7), 10,662 TCGA cases (with dataavailable for at least three platforms) were subtyped accordingto each of the above data platforms excluding WES, with thevarious subtype calls for each sample then being consolidated todefine multiplatform-based molecular subtypes.

With hierarchical clustering of the platform-level subtypeassignments (Supplementary Figs. S2A and S3), the cases segre-gated largely according to cancer type, with higher level branchesof the clustering tree representing major tissue-based categoriesincluding breast, liver, kidney, brain, gastrointestinal, squamous(head and neck, lung squamous, esophageal, bladder, cervical),lung adenocarcinoma, immune-related, uterus, and skin. In thispan-cancer molecular subtyping, basal-like BRCA (breast cancer)was molecularly distinct from other BRCA. K-means clusteringwas applied to the platform-level subtype assignment matrix toformally define a discrete number of subtypes (SupplementaryFigs. S2B, 4A, and 4B, and Supplementary Data S2). With a25-subtype solution, for example (Supplementary Fig. S2B),mostsubtypes were specific to either a single cancer type or multipletypes related by common tissue of origin.

Pan-cancer molecular classes that transcend tumor lineageAlthough the above pan-cancermolecular subtypes were found

to be highly concordant with their tissue-of-origin counterparts,we went on to pursue an alternate analytical strategy to definesubtypes that would not be driven by tumor lineage-specificmarkers or patterns. For expression (mRNA, miRNA, protein)and DNA methylation datasets, values were first normalized orcentered within each cancer type, thereby computationallyremoving any tissue or histology dominant effects. For definingpan-cancer subtypes using the above datasets, we had initiallyconsidered multiplatform-based solutions (e.g., analyzing resultsfrom the five data platforms together to define subtypes; ref. 9),but instead opted to use RNA-seq platform to define subtypesand then examine the other platforms for correlates of these RNA-seq-based subtypes. RNA-seq data would represent the full rangeof features for cancer-relevant pathways, whereas miRNA andprotein data platforms both represent expression-based data butwith amore limited number of features and represented pathwaysas compared to RNA-seq data. In addition, DNA copy alterations(while having a normal control) vary by cancer type (Supple-mentary Fig. S3 and ref. 9), and so would serve to segregatesamples by cancer type.

The RNA-seq platform (with values normalized within cancertype) was used to define 10 different subtypes or "classes" ofcancer (Table 1) across the 10,224 cases in TCGA cohort withavailable data (where with solutions ofmore than 10 subtypes noappreciable increase in clustering consensus was observed; Sup-plementary Fig. S5A), with the other profiling platforms thenbeing used to characterize these classes in terms of associatedpathways as described below. These pan-cancermolecular classes,referred to here as "c1" to "c10," were each characterized bywidespread molecular patterns (Fig. 1A; Supplementary

Figs. S5B–S5F and S6). For each class, the top genes most differ-entially expressed in the given class versus the rest of the tumorswere identified (Fig. 1A; Supplementary Data S2), where the topdifferential patterns—involving 854 genes in all—could beobserved to span cancer type. Notably, very few genes werestrongly associated with the c2 class in particular, which repre-sented one distinguishing feature of this class as compared to theother classes, although this c2 class would represent a basisof comparison. Normalized differential expression patterns forspecific genes of interest—including MYC, MKI67, CTAG1B,HIF1A, CD274, VIM, and ZEB1—could further distinguishbetween the classes and suggested associated pathways as furtherexplored below.

Specific miRNAs and proteins (with values normalized withincancer type) could distinguish between the pan-cancer molecularclasses (Fig. 1B; Supplementary Fig. S5E and Supplementary DataS2), with, for example, miR200 family members showing thelowest differential expression in c8 and c7 classes, and withimmune cell protein markers LCK and SYK being highest in thec3 class. Differential DNA methylation patterns could also dis-tinguish between the subtypes (Fig. 1C; Supplementary Figs. S5Band S5F, and Supplementary Data S2), with distinctive patternsassociated with c5 class in particular. For a subset of genes,significant anticorrelations between methylation and expression

Table 1. Tissue-independent pan-cancer molecular classes in TCGA cohort

Class n (%) Description and notable features

c1 847 (8.3) High differential expression of oxidativephosphorylation genes, glycolysis genes, andpentose phosphate pathway genes.

c2 2202 (21.5) Lack of strong associated expression patterns; canserve as a comparison group for the other classes.

c3 1340 (13.1) Strong association with immune checkpoint pathway;differential expression profile associated withimmune cell infiltration; mesenchymal signature;NRF2/KEAP1 pathway signature; Wnt pathwaysignature.

c4 411 (4) Differential expression profile associated withneuroendocrine tumors and with normal cells andtissues of the CNS; CT antigen expression.

c5 187 (1.8) Represents basal-like breast cancer; TP53-relatedalterations, MYC amplification and expression; YAP1target expression; high expression of pentosephosphate and TCA cycle genes; immune checkpointpathway; CT antigen expression.

c6 1179 (11.5) Epithelial signature; normoxia signature; YAP1 targetexpression.

c7 1153 (11.3) Mesenchymal signature; hypoxia signature; Wntpathway signature; Notch pathway signature; NRF2/KEAP1 pathway signature; lowdifferential expressionof miR-200.

c8 948 (9.3) High differential expression of fatty acid metabolismgenes; mesenchymal signature; hypoxia signature;Wnt pathway signature; Notch pathway signature;NRF2/KEAP1 pathway signature; high differentialDNA methylation and low differential expression ofmiR-200; differential expression profile associatedwith normal cells and tissues of the CNS; immunecheckpoint pathway (observed in TCGAcohort only).

c9 732 (7.2) Wnt pathway signature; Notch pathway signature;NRF2/KEAP1 pathway signature.

c10 1217 (11.9) Immune checkpoint pathway; differential expressionprofile associated with immune cell infiltration; YAP1target expression.

Abbreviation: CNS, central nervous system; CT, cancer-testis.

Chen et al.

Clin Cancer Res; 24(9) May 1, 2018 Clinical Cancer Research2184




Cancertype

mR

NA

Diff

eren

tial e

xpre

ssio

nno

rmal

ized

with

in c

ance

r typ

e 854 genes

847 2,202 1,340 411 1,179 948 732n 1,153

HigherLower

MYCMKI67CTAG1BHIF1ACD274VIMZEB1

AC

CB

LCA

BR

CA

CE

SC

CH

OL

CR

CD

LBC

ES

CA

GB

MH

NS

CK

ICH

KIR

CK

IRP

LAM

LLG

GLI

HC

LUA

DLU

SC

ME

SO

OV

PAA

DP

CP

GP

RA

DS

AR

CS

KC

MS

TAD

TGC

TTH

CA

THY

MU

CE

CU

CS

UV

M

c1c2c3c4c5c6c7c8c9

c10

<1 2 5 10 20 30 40% Representation ofclass by cancer type

Pan

-can

cer c

lass

n 79 408

1095 30

4 36 623 48 184

161

520 66 533

290

173

516

371

515

501 87 262

178

179

497

259

469

415

150

503

120

545 57 80

DN

A m

eth.

cent

er w

ithin

canc

er ty

pe

HigherLower

476 loci

miR

Exp

ress

ion

norm

aliz

ed w

ithin

canc

er ty

pe -

50 m

iRs

Pro

tein

exp

ress

ion

norm

. with

in c

ance

rty

pe -

25 p

rote

ins

LCKSYKER-alphaGATA3Claudin 7FibronectinCollagen VI

Caspace 7

miR-200cmiR-141

let-7c

miR-100miR-199ab

HigherLower

Pam50Basal

Luminal BA

1217

c1A E

B

C D

c2 c3 c8 c9c4 c6 c7 c10c5

187

Copy alt.

Qui

et

Noisy

00.10.20.30.40.50.60.70.80.9

1

Sur

viva

l pro

babi

lity

Overall survival (months)0 50 100 150 200 250

Censored

Overall P < 1E-10

c1: n = 845 c2: n = 2,181 c3: n = 1,337c4: n = 409 c5: n = 187 c6: n = 1,172c7: n = 1,148 c8: n = 940 c9: n = 720c10: n = 1,213

% Purity 10020

10,224 Human cancers

n136

6341

10985

176113104

5324606589

256134

6540

876859066

1392830

10270

16919

13925317842965295438672

16665651733506087

2497630786471254215

41543757847277949683559

372

c1 c2 c3 c4 c5 c6 c7 c8 c9 c10

Pan-cancer class

BRCA:BasalLUAD:Prox.-prolif.BLCA:c1 luminalPRAD:iCluster2PRAD:iCluster1BRCA:LumBTHCA:1KIRC:CC-e.1KIRP:P-e.2ACC:COC3UCEC:CN highLUSC:ClassicalLUAD:TRUKIRC:CC-e.2KIRP:P-e.1aOV:DifferentiatedGBM:ClassicalGBM:G-CIMPTHCA:4CRC:CINUCEC:CN lowTHCA:3PRAD:iCluster3GBM:NeuralGBM:ProneuralSKCM:KeratinTGCT:Non-SeminomaLGG:IDHmut-codelACC:COC2KIRC:CC-e.3KIRP:P.CIMP-2BLCA:c3 BasalLUAD:Prox.-inflamOV:MesenchymalCESC:C2GBM:MesenchymalLGG:IDHwtLUSC:BasalTHCA:5KIRP:P-e.1bSKCM:ImmuneBRCA:Her2UCEC:MSIUCEC:POLEACC:COC1THCA:2CRC:InvasiveHNSC:BasalLGG:IDHmut-non-codelOV:ImmunoreactiveSTAD/ESCA:EBVSTAD/ESCA:MSITGCT:SeminomaCRC:MSI/CIMPBRCA:NormalBLCA:c2 Lum. ImmuneBLCA:c4 Immune undiff.BRCA:LumALUSC:SecretoryHNSC:MesenchymalSTAD/ESCA:GSCESC:C1LUSC:PrimitiveOV:ProliferativeHNSC:ClassicalHNSC:AtypicalCESC:C3SKCM:MITF-lowSTAD/ESCA:CIN

0 20 P-value (-log10)10Relativeenrichment

FDR (-log10)0 188.4

FDR>0.1

Figure 1.

Molecular classes of TCGA cancers that transcend tumor lineage or tissue-of-origin. A, Using an alternative molecular classification approach, wherebydifferences between cancer types were first removed computationally prior to classification on the basis of mRNA expression data, ten major pan-cancer"classes" were identified. The first heat map shows differential mRNA expression patterns (values normalized within each main cancer type) for a set of854 genes found to best distinguish between the ten subtypes (see Materials and Methods). The second shows differential expression patterns for a selectset of genes representing pathways of particular interest. Numbers of cases (n¼ 10,224) denote representation onRNA-seqdata platform.B,Molecular features fromother data platforms associating with pan-cancermolecular class. Top heat map shows differential expression patterns (values normalizedwithin each cancer type),representing a top set of 50 miRNA features that distinguish between the ten molecular classes from part A. The second heat map shows differential proteinexpression patterns (by RPPA platform, values normalized within each cancer type), representing a top set of 25 features that distinguish between the 10 subtypes.The third heat map shows differential DNA methylation patterns (values centered within each cancer type) for a top set of features that distinguish a classassociated with basal-like breast cancer. Additional sample-level data tracks denote levels of genome-wide copy number alteration, cancer type (accordingto TCGA project, color coding in part C), BRCA Pam50 subtype, and estimated tumor sample purity (ref. 47; white, �100% purity). C, The percent representationsof each pan-cancer class by cancer type (according to TCGA project) are represented using a colorgram. D, Differences in patient overall survival among thepan-cancer molecular classes. P values by stratified log-rank test incorporating cancer type as a confounder. Overall P-value evaluates for significantdifferences among the groups as defined by pan-cancer class. E, Significance of overlap between the pan-cancer class assignments made in the present study(columns), with molecular-based subtype assignments (rows)made previously for a subset of cases. P values by one-sided Fisher exact test; only P values with FDR<0.1 (48) are represented. See Materials and Methods for TCGA project abbreviations. See also Supplementary Figs. S1 to S7 and Supplementary Data S1 and S2.






could be identified (Supplementary Fig. S5F and SupplementaryData S2). Within the top differentially expressed genes under-scoring each pan-cancer molecular class, specific gene categorieswere overrepresented (Supplementary Fig. S7), includingimmune-related genes being most highly expressed in the c3class, neuron-related genes being highly expressed in the c4 class,and cytoskeleton- and keratin-related genes being highlyexpressed in the c9 class. The c4, c5, and c6 classes tended toshow a higher degree of genome-wide copy alterations (Fig. 1B;Supplementary Fig. S5B). Each of the pan-cancer classes werefound to span cases from multiple cancer types (Figs. 1B and C),with the notable exceptionof c5,which consisted entirely of BRCAcases (n ¼ 187) and represents the basal-like breast cancermolecular subtype (ref. 12; Fig. 1B). As compared to the otherpan-cancer classes, the c5 class was associated with better overallpatient survival (Fig. 1D); however, compared with other breastcancers, c5 has a poorer survival (3). Overall, the pan-cancermolecular classes showed significant concordances with othermolecular subtype designations, which had been previouslymade for a subset of TCGA cases in individual cancer type studies(refs. 6, 12–24; Fig. 1E).

Associations involving somatic alterationsAcross the entire TCGA pan-cancer cohort, assessment of genes

within pathways demonstrated a high number of somatic altera-tions (mutation, copy alteration, or epigenetic silencing) involv-ing p53 (62.7% of 10,224 cases with exome data available),PI3K/AKT/mTOR (44.1%), receptor tyrosine kinase signaling(RTK, 43.9%), chromatin modification (40.8%), SWI/SNF com-plex (32.0%), Wnt/b-catenin (22.3%), MYC (10.9%), NRF2-ARE(9.3%), and Hippo signaling (4.5%; Supplementary Fig. S8A).The above pathways were found to be altered in different waysinvolving different genes in different cancer types (SupplementaryFig. S9A). A number of pathway-level or individual gene-levelalterations surveyed were highly represented within specific can-cer types or pan-cancer classes (Supplementary Figs. S8B andS9B). In particular, TP53 mutations were highly representedwithin both c4 and c5 tumors, and MYC amplifications werehighly represented in c5 tumors (Supplementary Figs. S8B and9B). Furthermore, c10 tumors were enriched for somatic altera-tions involving p53 pathway (including TP53 and ATM muta-tions), RTKs (MET mutations), and numerous genes involvingchromatin modification and SWI/SNF complex (SupplementaryFigs. S8B and S9B). Some of the somatic mutation associationsobserved might be attributable to cancer type- or mutation rate-specific patterns (Supplementary Figs. S10 and S11); for example,TP53 and chromatinmodifiermutations being enrichedwithin c4tumors would reflect in part the types of cancers more highlyrepresented within c4 (BLCA, CESC, HNSC, LUSC, etc.).

We sought to examine the effects on pathway activation—asmeasured by mRNA or protein signature—of somatic altera-tions impacting specific pathways noted above. Previously-defined gene expression signatures for p53, k-ras, MTOR,Wnt/b-catenin, MYC, NRF2, and YAP1 (6, 25–27) were appliedto the mRNA or protein expression profiles of the TCGAsamples, whereby each sample profile was scored for each ofthe above pathways. For each pathway considered, relativelevels of the corresponding signature were significantly differ-ent between somatically altered versus unaltered cases for thatpathway, and the differences were in the anticipated direction(e.g. p53-related alterations were associated with decreased

levels of p53 transcriptional targets, and MTOR-related altera-tions were associated with increased MTOR proteomic signal-ing; Supplementary Fig. S8C). These results would demonstratea widespread level of concordance between observed DNA-level alterations and global expression patterns. At the sametime, within the somatically altered and unaltered groups foreach pathway, a wide range of genes signature levels wereevident (Supplementary Fig. S8C), suggesting that other factorsbesides somatic mutation or copy alteration (e.g., tumor micro-environmental influences) may impact pathways.

Pathway-related gene signatures and mesenchymal cellsTo gain insight into pathways that would distinguish between

the various pan-cancer molecular classes, we applied a numberof gene signatures to TCGA expression profiles with values nor-malized within each cancer type. A number of pathways appearedmore or less active for different pan-cancer classes (Fig. 2A),as further explored below. For example, gene signature scoresfor epithelial–mesenchymal transition (EMT), hypoxia, NRF2/KEAP1, Wnt, and Notch all were higher in c3, c7, and c8 classes,as compared to the rest of the tumors. Focusing here on EMT,representing mesenchymal features, we observed a strong nega-tive association between the expression of miR-200 family mem-bers and their promoter methylation levels (Fig. 2B and 2C;Supplementary Fig. S12A), which demonstrates epigenetic regu-lationof thesemiRNAs across a large subset of human cancers. ThemiR200 family suppresses ZEB1, a key transcriptional regulator ofEMT. Across the entire TCGA cohort, we examined differentialexpression values for a set of genes and proteins representingcanonical mesenchymal or epithelial markers (11), as well as formiR200 family and ZEB1 genes (Fig. 2B). Manifestation of mes-enchymal features was highest in the c7 and c8 tumors and lowestin the c6 tumors (which appeared more epithelial), and miR200expression was lowest in c7 and c8 tumors, with associated DNAmethylation being highest in c8 tumors; c3 tumors also showedmesenchymal-associated patterns but less strongly than did c7and c8 tumors, and c9 tumors showed high expression of somemesenchymal and epithelial marker genes, as well as highmiR200. The c7 and c8 pan-cancer classes each involved a widerange of cancer types (Fig. 2D).

For some tumor subsets, the observed associations with mes-enchymal features could conceivably be attributable to surround-ing stromal cells within the tumor sample, as well as to changeswithin the actual cancer cells. For example, invasive lobularcarcinoma (ILC), the second most prevalent histologic subtypeof invasive breast cancer, is characterized by small discohesiveneoplastic cells invading the stroma in a single-file pattern (12).Immunohistochemical analysis in breast ILC has demonstratedthat the neoplastic lobular cells tend to retain their epithelialidentity, with mesenchymal markers being expressed by thefibroblasts in the prominent stromal component of ILC (28).However, for other cancer types, changes within cancer cells fromepithelial to mesenchymal states may occur at the tumor invasivefront, although not within the main tumor bulk (29). Of our10 pan-cancer molecular classes, four were characterized byrelatively lower sample purity: c3, c5 (representing basal-likebreast cancer), c7, and c8 (Fig. 1B; Supplementary Figs. S12B andS12C). Interestingly, of these four classes, only the c8 class wasstrongly associated with the breast ILC cases in TCGA (Supple-mentary Fig. S12B). In contrast, renal cell carcinomas of thepreviously described "CC-e.3" molecular subtype manifesting

Chen et al.





EMT (30) associatedwith the c7 class but notwith the c8 class, andthe EMT-associated "SQ.1" molecular subtype of lung cancer (7)was distributed among c3, c7, and c8 classes (Supplementary Fig.S12B). Three molecular subtypes previously associated with mes-enchymal features for specific cancer types, GBM:mesenchymal,OV:mesenchymal, and HNSC:mesenchymal, were each respec-tively associatedwith c3, c7, and c8 pan-cancer classes (Fig. 1E). Incomparison to c7, c8was further distinguished by high expressionof fatty acid metabolism genes (Fig. 2A), and associated hypoxia-related changes in c7 and c8, for example, suggest an alteredmicroenvironment. All of this would indicate that the molecularclasses associated with lower purity would each represent distinc-tive biology, apart from the technical aspects involved withsample collection.

Immune checkpoints and neuroendocrine tumorsAnalysis of gene expression patterns of normal tissues and

cells can provide meaningful context to the widespread differ-ential patterns observed in cancer (7, 31). We examined the top2,000 differential mRNAs from TCGA cohort (from Supple-mentary Fig. S5B) in a public expression dataset from theFantom consortium (10) of 850 profiles representing varioushuman cell and tissue specimens. Inter-correlations betweenFantom profiles and TCGA profiles revealed each pan-cancermolecular class to manifest distinctive patterns of global sim-ilarity or dissimilarity with specific categories of normal cellsand tissues (Fig. 3A; Supplementary Data S5). In particular, c3,c5, and c10 classes were strongly associated with immune-related cells and tissues; the c4 class was strongly associated

Fatty acid metabolismGlycolysis/GNG

Pentose phosphateTCA CYCLE

OX-PHOSRas

YAP1MYCEMT

HYPOXIANRF2/KEAP1

WNTNOTCH

Cancer type

c1A

B

C

D

c2 c3 c8 c9c4 c6 c7 c10c5

Gene expression(norm. within cancer type)

HigherLower


c1 c2 c3 c8 c9c4 c6 c7 c10c5

cg24702147cg15822328

MIR200AMIR200B

MIR429MIR141

MIR200C

10,224 Human cancersGene expression

(norm. within cancer type)HigherLower

DNA methylation(center within cancer type)

HigherLower

6

4

2

0

-2

-4

-6

-8

MIR

141/

200C

exp

ress

ion

norm

aliz

ed w

ithin

can

cer t

ype

n = 8540Spearman’s r = -0.46P ~ 0

cg24702147 - Differential DNA methylationbeta values centered within cancer type

-0.8 -0.4 0.80 0.4

CESC 44PRAD 47

THCA 50

BLCA 50

HNSC 57

LUSC 74LUAD 74

OV 41STAD 41

SKCM 40ESCA 38

KIRP 30LIHC 27

PAAD 27SARC 24

PCPG 20THYM 13MESO 12LGG 11DLBC 9CHOL 7KICH 7LAML 7TGCT 7UVM 6GBM 5UCS 5ACC 3

UCEC 75 KIRC 78CRC 84

BRCA 140

Pan-cancer c7 class - 1153 cases

Pan-cancer c8 class - 948 cases

PRAD 40

LUSC 56

LUAD 57

BLCA 71HNSC 78

SKCM 39UCEC 38

LIHC 37THCA 33

ESCA 32CESC 31

PAAD 29KIRC 17

THYM 7TGCT 5MESO 4SARC 3UVM 3CHOL 2KICH 2ACC 1DLBC 1LGG 1UCS 1

STAD 85 CRC 98

BRCA 136

10%

12%

7%7%7%

6%6%

5%4%4%4%

4%4%4%

14%

10%9%8%

7%

6%

6%4%

4%4%4%

3% OV 11KIRP 13

LAML 17

ZEB1CDH2

FN1FOXC2

GSCITGB6MMP2MMP3MMP9SNAI1SNAI2

SOX10TWIST1

VIMCDH1

DSPOCLN

Cancer type

Mesenchymal markerEpithelial marker

Fibronectin proteinE-Cadherin protein

Figure 2.

Pathway-associated gene signatures across pan-cancer molecular classes. A, By pan-cancer molecular class, pathway-associated mRNA signatures (usingvalues normalized within each cancer type). See Fig. 1C and D for cancer type color legend. Numbers of cases (n¼ 10,224) denote representation on RNA-seq dataplatform. B, Corresponding to cases from part A, heat maps showing DNA methylation and expression levels for miR200 family members (using valuesnormalized or centered within each cancer type). Representative DNA methylation probes (49) that map to the promoter of each miRNA cluster are shown(miR-141/200c¼ cg24702147, miR-200a/200b/429¼ cg15822328). Normalized expression levels for a set of canonical epithelial or mesenchymal markers (11) arealso shown. C, Scatter plot of differential methylation vs differential expression (using values normalized or centered within each cancer type), for cg24702147versus miR-141/200c (normalized values for the two miRNAs being averaged). Numbers of cases denote representation on all three data platforms for mRNA-seq,miRNA-seq, and 450K DNA methylation. Data points are colored according to pan-cancer class, as represented in A and B. D, For c7 and c8 pan-cancerclasses (associated with mesenchymal cells, along with hypoxia, NRF2/KEAP1, Wnt, and Notch signatures), distributions by cancer type. See alsoSupplementary Figs. S8 to S12 and Supplementary Data S3 and S4.






with cells and tissues related to the central nervous system(CNS), and c8 class was associated with CNS to a somewhatlesser degree; the c7 and c8 classes were associated with fibro-

blasts and other mesenchymal-related categories (reflecting themesenchymal marker patterns observed above; Fig. 2B); andthe c9 class was associated with adipose cells and heart tissues

CTAG1B - NY-ESO-1MAGEA4SAGE1CD274 - PDL1PDCD1 - PD1CD247 - CD3PDCD1LG2 - PDL2CTLA4 - CD152TNFRSF9 - CD137TNFRSF4 - CD134TLR9

c1 c2 c3 c8 c9c4 c6 c7 c10


HigherLower

Cancer type

c1

c2

c3

c8

c9

c4

c6

c7

c10

c5

Cancercell lineImmune

CNSSquamous

fibroblastAdipose/heart

A

D

B

C10,2

24 H

uman

can

cers

850 Human cell and tissue profiles (fantom)


212

CT

Ant

igen

gen

es

TEX101CXorf61HORMAD1ACTL8DMRT1CTCFLIGF2BP3ATAD2CASC5KIF20BKIF2CCEP55NUF2TTKPBKOIP5

LCK proteinSYK protein

c5

CHGASYPNCAM1ENO2CDX2

HSF2TPH1SLC18A1SSTR4SSTR5PNMA2SSTR1NAP1L1RNF41AKAP8LZFHX3FLJ10357SMARCD3PHF21AGLT8D1FZD7SPATA7SSTR3PLD3ZXDCOAZ2


HigherLower

51 N

ET

gene

s

c1 c2 c3 c8 c9c4 c6 c7 c10c5


-0.1

0 -0

.07

-0.0

3 0

.00

0.0

3 0

.07

0.1

0

Inter-profilecorrelation (r)

Positive

Neg

ativ

e

8%

7%7%6%6%

6%5%

5%4%4%4%

4%4%4%

11%

LUAD 49KIRP 52

LIHC 55

LUSC 60

KIRC 63

CRC 65

PRAD 77

PCPG 49SARC 49

BLCA 42HNSC 42

CESC 37OV 35

GBM 33STAD 32

PAAD 20ACC 19LAML 18KICH 15UVM 15ESCA 12MESO 12THYM 12TGCT 11UCS 8CHOL 6DLBC 2

UCEC 79 LGG 87THCA 89

SKCM 97

BRCA 106

Pan-cancer c3 class - 1,340 cases Pan-cancer c4 class - 411 casesPan-cancer c10 class - 1,217 cases

STAD 26OV 27

BCLA 36

LUSC 44

LUAD 25SKCM 20

UCEC 19ESCA 15

BRCA 10

THCA 8PAAD 6

LIHC 5MESO 5GBM 4PRAD 4ACC 2KIRP 2TGCT 2KICH 1KIRC 1UCS 1

CESC 59

HNSC 6914%

17%11%

9%

7%

6%6%

5%5%4%

7%

CRC 8

SARC 12

10%

7%7%6%6%

6%5%

5%4%4%

4%4%

4%4%

11%

THCA 50

OV 50

KIRC 51

LUSC 64

LUAD 66BRCA 68

PRAD 46TGCT 46

BLCA 45CESC 39

SARC 35SKCM 34

GBM 33PAAD 25

ESCA 23MESO 20KIRP 18LIHC 12THYM 12PCPG 11UCS 10LAML 8UVM 8ACC 4DLBC 2CHOL 1KICH 1

CRC 71STAD 75

HNSC 84

UCEC 88

LGG 117

Figure 3.

Normal tissue and cell-type associations with the pan-cancer molecular classes. A, Inter-profile correlations were computed between TCGA expressionprofiles (with values normalized within each cancer type) and profiles from the Fantom consortium expression dataset of various cell types or tissues from humanspecimens (n¼ 850profiles; ref. 10). Membership of the Fantomprofiles in general categories of "cancer," "cell line," "immune" (immune cell types or blood or relatedtissues), "CNS" (related to central nervous system including brain), "squamous" (including bronchial, trachea, oral regions, throat and esophagus regions,nasal regions, urothelial, cervix, sebocyte, keratin/skin/epidermis), "fibroblast," or "adipocyte/heart" is indicated. Cancer-type color coding in D. B, Heat mapsof differential expression (values normalized within each cancer type), for genes encoding immunotherapeutic targets (top), for LCK and SYK proteins(middle, representing markers for T and B cells, respectively), and for genes encoding cancer-testis (CT) antigens (from the CT Gene Database, http://cancerimmunity.org/resources/ct-gene-database/). C, Heat maps of differential expression (values normalized within each cancer type), for genes encodingcanonical markers of neuroendocrine tumors (top), and for a set of 51 genes in a panel of neuroendocrine tumor (NET) markers (34), as uncoveredpreviously using gene expression profiling (bottom). D, For c3 (immune-associated), c10 (immune-associated), and c4 (CNS- and neuroendocrine-associated)pan-cancer classes, distributions by cancer type. See also Supplementary Fig. S13 and Supplementary Data S5.

Chen et al.




http://cancerimmunity.org/resources/ct-gene-database/

http://cancerimmunity.org/resources/ct-gene-database/


(which may be reflective of the association with fatty acidmetabolism observed above; Fig. 2A). Analysis of the Fantommouse expression dataset yielded similar associations (Supple-mentary Fig. S13A). In previous analyses utilizing TCGA expres-sion profiles as normalized across all cancers, brain cancers andblood cancers associated as a group with fantom CNS andimmune-related profiles, respectively (7).

Focusing here on the immune-related associations found forspecific pan-cancer molecular classes, we went on to surveyTCGA expression profiles for a set of genes representing poten-tial targets for immunotherapy (6), including cancer-testis (CT)antigen genes and genes involved in immune checkpoint path-way (Fig. 3B), such as genes encoding PD1, PDL1, PDL2, andCTLA4. As a group, CT antigen genes were highest in both c4and c5 classes, with a subset of CT antigen genes being partic-ularly higher in c5 class (see Fig. 3B). However, immunecheckpoint genes were most strongly differentially expressedwithin the c3 class, whereas the c8 and c10 classes also showedincreased though relatively lower expression of these genes.Differential protein expression of T-cell marker LCK and B-cellmarker SYK indicated the presence of T cells in c3, c8, and c10classes, and the presence of B cells in c3 and c10 classes(Fig. 3B). Analysis of gene expression signatures from Bindeaand colleagues (32) suggested that levels of infiltrating immunecell types were highest within the c3 class, followed by the c8class, whereas the c10 class showed signatures for some but notall immune cell types (Supplementary Fig. S13B).

The above associations of c4 tumors with CNS tissues and cellssuggested features of neuroendocrine tumors (NET), whicharise from cells of the endocrine and nervous systems and whichoccur inmany different tissues throughout the human body (33).Genes encoding canonical markers of NETs—including CHGA(chromogranin A), SYP (synaptophysin), NCAM1 (CD56),ENO2 (neuron-specific enolase), and CDX2—were all differen-tially higher in c4 tumors as compared with the other pan-cancerclasses (Fig. 3C). In addition, a set of 51 genes in a panel of NETmarkers previously uncovered using gene expressionprofilingwasexamined (34), with the majority of these being differentiallyhigher within the c4 class (Fig. 3C). TCGA LUAD cases previouslydetermined to represent neuroendocrine cancer were alsosignificantly enriched within the c4 class (SupplementaryFig. S12B, 6/14 cases, P < 1E�5, one-sided Fisher exact test). Thec3, c10, and c4 pan-cancer classes each involved a wide range ofcancer types (Fig. 3D), with c4 class in particular being composedof sizable percentages of head and neck, cervical, lung, bladder,ovarian, and gastrointestinal cancers.

Pathway-level differences across pan-cancer molecular classesThe above pathway-associated gene signatures (Fig. 2A) indi-

cated differential activation of specific pathways among the pan-cancer molecular classes. A number of these signatures wererelated to metabolism, including fatty acid metabolism, glycol-ysis/gluconeogenesis, pentose phosphate pathway, TCA cycle,and oxidative phosphorylation. The individual genes that com-prised these signatures can be represented in a pathway diagram(Fig. 4A), whereby the c1, c6, and c8 classes were each observed toshow evidence for the altered utilization of specific metabolicpathways, as compared to the rest of the cancer cases. Both c1 andc6 tumors showed increased expression of genes involved withoxidative phosphorylation, whereas c6 tumors also showedhigher expression of genes involved in the TCA cycle and of genes

involved in fatty acid synthesis. However, c8 tumors appeared todownregulate TCA cycle, oxidative phosphorylation, and fattyacid synthesis pathways, and toupregulate genes involving in fattyacid metabolism.

Deregulated pathways were also reflective of tumor microen-vironmental effects at work in distinct subsets of cancers. Differ-ential expression patterns involving the c3, c7, or c8 pan-cancermolecular classes were consistent with a consensusmodel (35) oftumor-associated macrophage (TAM) roles in the tumor micro-environment (Fig. 4B), whereby monocytes are first recruited tothe tumor microenvironment (e.g., by CSF1 and CCL2). Tumor-secreted cytokines then have the potential to polarize recruitedmonocytes into TAMs, which play vital roles in a number ofprocesses including immune suppression andEMT. EMTmay alsobe induced by either hypoxia or Wnt signaling pathway (36–38),both of which appear increased in c3, c7, and c8 tumors (Figs. 2Aand 4B); NRF2/KEAP1 and Notch pathways also appearedincreased in these tumors. Immune suppression may involvemultiple redundant immune checkpoints, which were most asso-ciated with c3, c5, c8, and c10 classes (Figs. 3B and 4C); thesetumors showed higher expression of ligands such as PDL1and PDL2, along with higher expression of the correspondingreceptors associated with T cells.

TCGA patterns observable in external cohortsThe pan-cancer molecular class associations, as observed in

TCGA datasets, were also examined in an external multicancerexpression dataset. From the expO dataset, mRNA expressionprofiles of 2041 cancer cases representing 26 different typeswere obtained. Within each cancer type, expression values foreach gene in the expO dataset were first normalized, and theneach expO tumor profile was classified by pan-cancer molecularclass as defined by TCGA data (Fig. 5A; Supplementary Figs.S14A and S14B), where the top set of 854 mRNAs distinguish-ing between our 10 classes (Fig. 1B) was used as the classifier. Inthe same manner as carried out above for TCGA datasets, expOexpression profiles were also scored for mRNA signaturesrelated to specific pathways or normal cell types (Fig. 5B;Supplementary Fig. S15), where similar overall trends of path-way-level differences between classes as originally observed inTCGA cohort could also be observed in the expO cohort. Inparticular, c5 tumors were composed almost entirely of breastcancer cases; anticipated pathway-level differences involvingmetabolism, YAP1, MYC, EMT, hypoxia, NRF2/KEAP1, WNT,NOTCH, and immune checkpoint were observed in expOcohort; CT antigen genes were higher as a group in c4, withthe same subset also being high in c5 class as observed forTCGA; and neuroendocrine-associated global patterns andmarkers were associated with c4 class.

In a similar manner to that of the expO dataset, cell lineprofiles from the cancer cell line encyclopedia (CCLE) expres-sion dataset (39) were each assigned to a pan-cancer class(Supplementary Fig. S16). Not all patterns of interest observ-able in human tumor data were as apparent in the CCLE results,which could be attributed to a number of factors includinggrowth conditions of cell lines lacking tumor microenviron-mental effects and differences in the types of cancers repre-sented in CCLE; however, a number of associations wereidentifiable in CCLE, including c7 and c8 classes with EMTand hypoxia, and c4 class with neuroendocrine-related patterns(e.g., involving small cell lung cancers).






NADH

FADH2

NADPH

TCA Cycle

The electron transport chain (ETC)

T cell

Tumor cell/professional APC/target cell

LCK

Monocyte recruitment factors

Tumor-associated macrophage

Tissueremodeling

Immunesuppression

Enhanced invasionand metastasis

EMT

Hypoxia

EMT

Fatty acidsynthesis

FASN

ACC

Complex I

Complex II

Complex III Complex IVComplex V

NRF2-ARE

Notch

Wnt/β-catenin

Fatty acidmetabolism

β-oxidationPyruvate Acetyl-CoA

OAA Citrate

Isocitrate

-ketoglutarateα

Malate

Fumarate

Succinyl-CoASuccinate

PDC, PyruvateDehydrogenase Complex

RNA

PDCD1LG2PDL1 PDL2 HVEM CD137L OX40L

PDCD1 CD244PD1 CD137 OX40

TCR

CD247

LAG32B4

CD3

PDL1 PDL2 HVEM CD137L OX40L

c3 v

s. o

ther

s

c10

vs. o

ther

sc5

vs.

oth

ers

protein

CTLA4

CD274 CD80 CD86

LCK

BTLA TNFRSF9 TNFRSF4

TNFSF14TNFSF9

CTAG1B MAGEA4SAGE1

Higher in c3/c7/c8Lower in c3/c7/c8

B

CCL2 CSF1

LYVE1

MMP9

PDGFA

VEGFA

HIF1A

LAG3

PDCD1

CD68

IL10

CCL18

VIM

ZEB1 MIR200

E-cad

HIF1A Hypoxiatargets

Pyruvate

Malate

Oxaloacetate

Acetyl CoA

Malonyl-CoA

Fatty acids

FASN

ACC

FASN

ACC

Citrate

KEAP1 CUL3

NRF2/KEAP1targets

NFE2L2

NOTCH3 NOTCH4

NOTCH1 NOTCH2

JAG1 JAG2 DLL1 DLL3 DLL4

RNA

c1 v

s. o

ther

s

c8 v

s. o

ther

sc6

vs.

oth

ers

protein

Higher in c1/c6/c8Lower in c1/c6/c8

HES1 HES2 HEY1 HEY2 DTX1

AXIN1

APCCTNNB1

Wnttargets

FGF18

RNA

c3 v

s. o

ther

s

c8 v

s. o

ther

sc7

vs.

oth

ers

protein

A

C

CD48TNFRSF14 MHC II

PDP1 PDP2PDPR

PDK1 PDK2PDK3 PDK4

DLAT DLD

PDHA1 PDHA2 PDHB

PCCS

MDH2 ACO2

IDH3BIDH3G

IDH3A IDH2

ME1

MDH1

ACLY

NNT

SDHASDHBSDHCSDHD

SUCLA2SUCLG1 SUCLG2

FH

Nuclear-encodedUQCRsubunits

Nuclear-encodedCOXsubunits

Nuclear-encodedATPAsubunits

Nuclear-encodedNADH-DHsubunits

Fatty Acyl CoAAcycarnitine

Fatty Acyl CoA

CPT1BCPT1A

CPT1CCPT2

ACADL

ACSL1ACSL4

ACSL6ACSL5

SNAI1

Hig

her i

n c3

/c5/

c10

Low

er in

c3/

c5/c

10

<0.0001/0.0001<1E-10/1.2E-10

<0.01/0.01P-value / FDR

CD4

OGDH

<0.0

001/

0.00

01<1

E-1

0/1.

6E-1

0

<0.0

1/0.

01P

-val

ue/F

DR

<0.0

001/

0.00

01<1

E-1

0/1.

3E-1

0

<0.0

1/0.

01P

-val

ue/F

DR

Figure 4.

Differentially active pathways acrosspan-cancer molecular classes.A, Pathway diagram representingcore metabolic pathways, withdifferential expression patternsrepresented (using values normalizedwithin cancer type), comparingtumors in pan-cancer classes c1, c6, orc8 with tumors in the otherseven classes (red, significantly higherin c1/c6/c8). B, Diagram of tumor-associated macrophage roles in thetumor microenvironment (35), andof Notch, NRF2-ARE, andWnt/beta-catenin pathways, withdifferential expression patternsrepresented (using values normalizedwithin cancer type), comparingtumors in pan-cancer classes c3, c7, orc8 with tumors in the other sevenclasses (red, significantly higher in c3/c7/c8). C, Diagram of immunecheckpoint pathway (featuringinteractions between T cells andantigen-presenting cells, includingtumor cells), with differentialexpression patterns represented(using values normalized withincancer type), comparing tumors inpan-cancer classes c3, c5, or c10 withtumors in the other seven classes (red,significantly higher in c3/c5/c10).P values in A–C by Mann–WhitneyU test.

Chen et al.





DiscussionAlthough previous studies have greatly served to elucidate the

molecular landscape of the individual cancer types represented inTCGA, our pan-cancer molecular classes provide an excellentframework for examining pathways or processes that would cutacross these individual types. All but one of these molecularclasses each involves a wide range of cancer types and wouldtherefore be relevant to the study of multiple diseases. It is also

remarkable that basal-like breast cancer forms its own molecularclass—made up only of breast cancers—entirely distinct from allof the other cancer cases examined. Although basal-like breastcancer is already understood to represent a fundamentally differ-ent disease than other types of breast cancer, questions remain asto the origins of the observed molecular differences (40). Thedistinct patterns associated with basal-like breast cancer at theepigenetic and transcriptional levels, in this study, could suggest

Fatty acid metabolismGlycolysis/GNG

Pentose phosphateTCA CYCLE

OX-PHOSRas

YAP1MYCEMT

HYPOXIANRF2/KEAP1

WNTNOTCH

Ac1 c2 c3 c8 c9c4 c6 c7 c10c5

CD274 - PDL1PDCD1 - PD1CD247 - CD3

PDCD1LG2 - PDL2CTLA4 - CD152

TNFRSF9 - CD137TNFRSF4 - CD134

TLR9TEX101CXorf61

HORMAD1ACTL8

DMRT1CTCFL

IGF2BP3ATAD2CASC5KIF20BKIF2C

CEP55NUF2

TTKPBKOIP5

c1c3c4c5c6c7c8c9

c10

Cla

ss-s

peci

fic g

enes

TCGApan-cancer

Higher

Lower G

ene expression(norm

. within cancer type)

External GSE2109 expression dataset (n = 2,041)

Abdominal, Bladder, Bone & Cartilage, Brain, Breast, Cervix, Colorectal, Corpus Uteri,Endometrium, Esophagus, Kidney, Liver, Lung, Omentum, Ovary, Pancreas,Pelvic Tissue, Prostate, Renal Pelvis, Soft Tissue, Stomach, Thyroid, Uterus, Vulva

c1 c2 c3 c8 c9c4 c6 c7 c10c5External GSE2109 expression dataset (n = 2,041)B

161

CT

Ant

igen

gen

esP

athw

ay s

igna

ture

sIm

mun

e ch

eckp

oint

HigherLower

Gene expression(norm. within cancer type) Average inter-profile correlation

PositiveNegative

Immune profilesCNS profiles

Fibroblast profilesAdipocyte profiles

CHGASYP

NCAM1ENO2CDX2

NE

TFa

ntom

c1 c2 c3 c4 c5 c6 c7 c8 c9 c10

HigherLower

t-statistic, differential expressiongiven class versus others

-8.0

0 -5

.33

-2.6

7 0

.00

2.6

7 5

.33

8.0

0

Figure 5.

Observation of patterns associated withTCGApan-cancermolecular classes in anexternalmulticancer expression profilingdataset. A, Gene expressionprofiles of 2041 cancer cases of variouspathologically defined cancer types,represented in the expO (GSE2109)dataset (profiles being normalizedwithin their respective cancer type),were classified according to TCGA pan-cancer molecular class. Expressionpatterns for the top set of 854 mRNAsdistinguishing between the 10 TCGAmolecular classes (from Fig. 1A) areshown for both TCGA and GSE2109datasets. Genes in the GSE2109 sampleprofiles sharing similarity with TCGAclass-specific signature pattern arehighlighted. B, In the same manner ascarried out for TCGA datasets, expOexpression profiles were scored forpathway-associated gene signatures(from Fig. 2A), surveyed for immunecheckpoint markers and for CT antigengenes (from Fig. 3B, using thesame gene ordering), surveyed forcanonical NET markers (from Fig. 3C),and scored for similarity to normal cell-type categories represented in thefantom dataset (from Fig. 3A). Pan-cancer class associations of particularinterest (which tend to follow thepatterns first observed in TCGA cohort)are highlighted. The purple-cyan heatmaps off to the right denote t-statisticsfor comparing the given class versus therest of the tumors; dark purple or cyancorresponds approximately to P < 0.01.A and B have the same ordering of expOexpression profiles. See alsoSupplementary Figs. S14 to S16 andSupplementary Data S6.






that this disease would have a different cell-of-origin from that ofother breast cancers.

The landscape of pan-cancer–associated biological processesand pathways as uncovered here would include many that werewell-established as having a functional role in the experimentalsetting, but for which the full extent of their involvement inhuman cancers may have been unclear. Processes and pathwayssuch as metabolism, immune checkpoint, hypoxia, NRF2-ARE,HIPPO, Wnt, EMT, and Notch have been well characterized inthe experimental setting, using models such as cell lines andmice (41). This study finds each of the above to involve largesubsets of human cancers, as observed based on the coordinateexpression patterns involving large numbers of genes. DNAmutation events alone (and tumor evolution by proxy) wouldnot appear to represent the sole driver of these processes incancer, where tumor microenvironment influences likely play amajor role. For example, mutations in VHL or in RTK genes mayaccelerate processes of hypoxia or RTK signaling, respectively,or microenvironmental conditions such as lack of oxygen oravailability of growth factors could conceivably achieve thesame respective results. Individual molecular markers of ourpan-cancer molecular classes that might seem most relevantfrom a therapeutic standpoint—including markers of processesand pathways noted above—could potentially be evaluated inthe setting of patient care in future studies.

Multiple pan-cancer classes manifested patterns that we couldascribe to the noncancer cellular component of the tumor,including immune infiltrates and cancer-associated fibroblasts.Our finding of different stroma-associated tumor subsets, eachwith distinctive features that would distinguish it from the otherpan-cancer classes, suggests various biological roles for the stro-mal component in human cancer. Three of our pan-cancer classes(c3, c5, and c10) showed particularly strong patterns of immunecell infiltrates, whereas one of these (c3) showing stronger pat-terns of immune checkpoint pathway genes andof genes involvedin specific metabolic pathways. Three of our pan-cancer classes(c3, c7, and c8) showed patterns of mesenchymal cells, with twoof these (c7 and c8) showing strong patterns of fibroblasts inparticular, andwith one of these (c8) showing patterns associatedwith fatty acid metabolism. The tumor microenvironment—which consists of a mixture of fibroblasts, myofibroblasts, endo-thelial cells, immune cells, other cells, and altered extracellularmatrix—is understood to play an important role in the initiationand progression of various cancers (42, 43). For example, EMT ofthe cancer cells can occur in areas of fibrosis and may account forup to 40% of tumor-associated fibroblasts (43). As reflected inTCGA data, the various ways in which the tumor microenviron-ment can influence cancermay be involvedwithin distinct subsetsof human tumors.

Our discovery of a molecular class of cancers manifesting adifferential expression profile associatedwith bothNETs andwithnormal cells and tissues of the CNS would potentially haveimplications for a large subset of human cancers, with �4% of

TCGA cases belonging to this "c4" class. Only through ouranalytical approach of subtracting out tissue-level moleculardifferences were we able to uncover this neuroendocrine-associ-ated pan-cancer class (where otherwise such CNS-related patternswould be more strongly associated with TCGA brain cancers).NETs are understood to occur throughout the body, including inbreast, cervical, lung, pancreas, ovarian, and gastrointestinal tis-sues, although it is believed that the incidence of these tumors isrelatively rare, estimated at about 1% of cancers diagnosed in theUnited States (44, 45). If on the order of 4% of human cancerscould be considered as NETs (or at least manifesting a molecularprofile associated with these tumors), this would be well outsidethe range of previous estimates, suggesting that a large proportionof NETs are not diagnosed as such, but rather are grouped in withother cancers sharing the same tissue of origin. The tumor samplesoriginally contributed to TCGA came from many different treat-ment facilities, with no uniform practices for determining neu-roendocrine features being in place. Current strategies for iden-tifying NETs would include staging at surgery, pathological grad-ing, blood Chromogranin A (CgA) measurements, and detectionof circulating tumor cells, with there being a clear need foradditional and better biomarkers (34). The more accurate iden-tification of neuroendocrine-associated cancers in particularwould have important implications regarding the treatment ofthis disease (46).

Disclosure of Potential Conflicts of InterestNo potential conflicts of interest were disclosed.

Authors' ContributionsConception and design: C.J. CreightonDevelopment of methodology: C.J. CreightonAnalysis and interpretation of data (e.g., statistical analysis, biostatistics,computational analysis): F. Chen, Y. Zhang, B. Deneen, D.J. Kwiatkowski,C.J. CreightonWriting, review, and/or revision of the manuscript: D.L. Gibbons,D.J. Kwiatkowski, M. Ittmann, C.J. CreightonAdministrative, technical, or material support (i.e., reporting or organizingdata, constructing databases): C.J. CreightonStudy supervision: C.J. Creighton

AcknowledgmentsThis work was supported in part by Cancer Prevention and Research Institute

of Texas (CPRIT) grants RP120713 C2 (to C.J. Creighton), RP150405 (toD.L. Gibbons), and RP120713 P2 (to D.L. Gibbons), and by the NIH grantP30CA125123 (to C. Creighton).

The costs of publication of this article were defrayed in part by thepayment of page charges. This article must therefore be hereby markedadvertisement in accordance with 18 U.S.C. Section 1734 solely to indicatethis fact.

Received November 13, 2017; revised January 8, 2018; accepted February 2,2018; published first February 9, 2018.

References1. Perou C, Sørlie T, Eisen M, van de Rijn M, Jeffrey S, Rees C, et al. Molecular

portraits of human breast tumours. Nature 2000;406:747–52.2. Cancer_Genome_Atlas_Research_Network, Weinstein J, Collisson E,

Mills G, Shaw K, Ozenberger B, et al. The Cancer Genome Atlas pan-cancer analysis project. Nat Genet 2013;45:1113–20.

3. Creighton C. The molecular profile of luminal B breast cancer. Biologics2012;6:289–97.

4. Zhang Y, Kwok-Shing Ng P, Kucherlapati M, Chen F, Liu Y, Tsang Y, et al.A pan-cancer proteogenomic atlas of PI3K/AKT/mTOR pathway altera-tions. Cancer Cell 2017 [Epub ahead of print].


Chen et al.




5. Hoadley K, Yau C, Wolf D, Cherniack A, Tamborero D, Ng S, et al. Multi-platform analysis of 12 cancer types reveals molecular classification withinand across tissues of origin. Cell 2014;158:929–44.

6. Chen F, Zhang Y, S enbabao�glu Y, Ciriello G, Yang L, Reznik E, et al.Multilevel genomics-based taxonomy of renal cell carcinoma. Cell Rep2016;14:2476–89.

7. Chen F, Zhang Y, Parra E, Rodriguez J, Behrens C, Akbani R, et al. Multi-platform-based molecular subtypes of non-small cell lung cancer. Onco-gene 2016 [Epub ahead of print].

8. Akbani R, Ng P, Werner H, Shahmoradgoli M, Zhang F, Ju Z, et al. A pan-cancer proteomic perspective on The Cancer Genome Atlas. Nat Commun2014 [Epub ahead of print].

9. Chen F, Zhang Y, Boss�e D, Lalani A, Hakimi A, Hsieh J, et al. Pan-urologiccancer genomic subtypes that transcend tissue of origin. Nat Commun2017;8:199.

10. FANTOM_Consortium_and_the_RIKEN_PMI_and_CLST_(DGT), ForrestA, Kawaji H, Rehli M, Baillie J, de Hoon M, et al. A promoter-levelmammalian expression atlas. Nature 2014;507:462–70.

11. Gibbons D, Creighton C. Pan-cancer survey of epithelial-mesenchymaltransition markers across The Cancer Genome Atlas. Dev Dyn 2017 [Epubahead of print].

12. Ciriello G, Gatza M, Beck A, Wilkerson M, Rhie S, Pastore A, et al.Comprehensive molecular portraits of invasive lobular breast cancer. CellCycle 2015;163:506–19.

13. Cancer_Genome_Atlas_Research_Network, Kandoth C, Schultz N, Cher-niack A, Akbani R, Liu Y, et al. Integrated genomic characterization ofendometrial carcinoma. Nature 2013;497:67–73.

14. Cancer_Genome_Atlas_Research_Network. The molecular taxonomy ofprimary prostate cancer. Cell 2015;163:1011–25.

15. Cancer_Genome_Atlas_Research_Network. Comprehensive molecularcharacterization of urothelial bladder carcinoma. Nature 2014;507:315–22.

16. Cancer_Genome_Atlas_Research_Network. Comprehensive molecularprofiling of lung adenocarcinoma. Nature 2014;511:543–50.

17. Cancer_Genome_Atlas_Research_Network. Comprehensive genomiccharacterization of squamous cell lung cancers. Nature 2012;489:519–25.

18. Cancer_Genome_Atlas_Research_Network. Integrated genomic analysesof ovarian carcinoma. Nature 2011;474:609–15.

19. Cancer_Genome_Atlas_Research_Network. Comprehensive genomiccharacterization defines human glioblastoma genes and core pathways.Nature 2008;455:1061–8.

20. Cancer_Genome_Atlas_Research_Network. Integrated genomic andmolecular characterization of cervical cancer. Nature 2017;543:378–84.

21. Cancer_Genome_Atlas_Research_Network. Integrated genomic character-ization of oesophageal carcinoma. Nature 2017;541:169–75.

22. Cancer_Genome_Atlas_Network. Genomic classification of cutaneousmelanoma. Cell 2015;161:1681–96.

23. Cancer_Genome_Atlas_Network. Comprehensive genomic characteriza-tion of head and neck squamous cell carcinomas. Nature 2015;517:576–82.

24. Cancer_Genome_Atlas_Research_Network. Comprehensive, integrativegenomic analysis of diffuse lower-grade gliomas. N Engl J Med 2015;372:2481–98.

25. Gingras M, Covington K, Chang D, Donehower L, Gill A, Ittmann M, et al.Ampullary cancers harbor ELF3 tumor suppressor gene mutations andexhibit frequent WNT dysregulation. Cell Rep 2016;14:907–19.

26. Singh A, Greninger P, Rhodes D, Koopman L, Violette S, Bardeesy N,et al. A gene expression signature associated with "K-Ras addiction"reveals regulators of EMT and tumor cell survival. Cancer Cell 2009;15:489–500.

27. Creighton C. Multiple oncogenic pathway signatures show coordinateexpression patterns in human prostate tumors. PloS One 2008;3:e1816.

28. McCart-Reed A, Kutasovic J, Lakhani S, Simpson P. Invasive lobularcarcinoma of the breast: morphology, biomarkers and 'omics'. BreastCancer Res 2015;17:12.

29. Nieto M, Huang R, Jackson R, Thiery J. EMT: 2016. Cell 2016;166:21–45.30. Chang J, Wooten E, Tsimelzon A, Hilsenbeck S, Gutierrez M, Tham Y, et al.

Patterns of resistance and incomplete response to docetaxel by geneexpression profiling in breast cancer patients. J Clin Oncol 2005;23:1169–77.

31. Davis C, Ricketts C, Wang M, Yang L, Cherniack A, Shen H, et al. Thesomatic genomic landscape of chromophobe renal cell carcinoma. CancerCell 2014;26:319–30.

32. BindeaG,Mlecnik B, TosoliniM,KirilovskyA,WaldnerM,Obenauf A, et al.Spatiotemporal dynamics of intratumoral immune cells revealthe immune landscape in human cancer. Immunity 2013;39:782–95.

33. Ramage J, Ahmed A, Ardill J, BaxN, BreenD, CaplinM, et al. Guidelines forthe management of gastroenteropancreatic neuroendocrine (includingcarcinoid) tumours (NETs). Gut 2012;61:6–32.

34. Modlin I, Drozdov I, Kidd M. The identification of gut neuroendocrinetumor disease by multiple synchronous transcript analysis in blood. PloSOne 2013;8:e63364.

35. Cook J, Hagemann T. Tumour-associated macrophages and cancer. CurrOpin Pharmacol 2013;13:595–601.

36. Tsai Y, Wu K. Hypoxia-regulated target genes implicated in tumor metas-tasis. J Biomed Sci 2012;19:102.

37. Kao S,WuK, LeeW.Hypoxia, epithelial-mesenchymal transition, and TET-mediated epigenetic changes. J Clin Med 2016;5:E24.

38. Micalizzi D, Farabaugh S, Ford H. Epithelial-mesenchymal transition incancer: parallels between normal development and tumor progression.J Mammary Gland Biol Neoplasia 2010;15:117–34.

39. Barretina J, Caponigro G, Stransky N, Venkatesan K, Margolin A, Kim S,et al. The cancer cell line encyclopedia enables predictive modelling ofanticancer drug sensitivity. Nature 2012;483:603–7.

40. Skibinski A, Kuperwasser C. The origin of breast tumor heterogeneity.Oncogene 2015;34:5309–16.

41. Hanahan D, Weinberg R. Hallmarks of cancer: the next generation. Cell2011;144:646–74.

42. Dakhova O, Ozen M, Creighton C, Li R, Ayala G, Rowley D, et al. Globalgene expression analysis of reactive stroma in prostate cancer. Clin CancerRes 2009;15:3979–89.

43. Franco O, Shaw A, Strand D, Hayward S. Cancer associated fibroblasts incancer pathogenesis. Semin Cell Dev Biol 2010;21:33–9.

44. Yao J, Hassan M, Phan A, Dagohoy C, Leary C, Mares J, et al. One hundredyears after "carcinoid": epidemiology of and prognostic factors for neu-roendocrine tumors in 35,825 cases in the United States. J Clin Oncol2008;26:3063–72.

45. Basuroy R, Srirajaskanthan R, Ramage J. Neuroendocrine tumors. Gastro-enterol Clin North Am 2016;45:487–507.

46. Strosberg J, El-HaddadG,Wolin E, Hendifar A, Yao J, Chasen B, et al. Phase3 Trial of 177Lu-dotatate formidgut neuroendocrine tumors. N Engl JMed2017;376:125–35.

47. AranD, SirotaM, Butte A. Systematic pan-cancer analysis of tumour purity.Nat Commun 2015;6:8971.

48. Storey JD, Tibshirani R. Statistical significance for genomewide studies.Proc Natl Acad Sci USA 2003;100:9440–5.

49. Cherniack A, Shen H, Walter V, Stewart C, Murray B, Bowlby R, et al.Integrated molecular characterization of uterine carcinosarcoma. CancerCell 2017;31:411–23.






2018;24:2182-2193. Published OnlineFirst February 9, 2018.Clin Cancer Res Fengju Chen, Yiqun Zhang, Don L. Gibbons, et al. CasesAcross 32 Cancer Types, Multiple Data Platforms, and over 10,000 Pan-Cancer Molecular Classes Transcending Tumor Lineage

Updated version

10.1158/1078-0432.CCR-17-3378doi:

Access the most recent version of this article at:

Material

Supplementary

http://clincancerres.aacrjournals.org/content/suppl/2018/02/08/1078-0432.CCR-17-3378.DC1

Access the most recent supplemental material at:

Cited articles

http://clincancerres.aacrjournals.org/content/24/9/2182.full#ref-list-1

This article cites 45 articles, 5 of which you can access for free at:

Citing articles

http://clincancerres.aacrjournals.org/content/24/9/2182.full#related-urls

This article has been cited by 4 HighWire-hosted articles. Access the articles at:

E-mail alerts related to this article or journal.Sign up to receive free email-alerts

Subscriptions

Reprints and

[email protected]

To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department at

Permissions

Rightslink site. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC)

.http://clincancerres.aacrjournals.org/content/24/9/2182To request permission to re-use all or part of this article, use this link



http://clincancerres.aacrjournals.org/lookup/doi/10.1158/1078-0432.CCR-17-3378

http://clincancerres.aacrjournals.org/content/suppl/2018/02/08/1078-0432.CCR-17-3378.DC1

http://clincancerres.aacrjournals.org/content/24/9/2182.full#ref-list-1

http://clincancerres.aacrjournals.org/content/24/9/2182.full#related-urls

http://clincancerres.aacrjournals.org/cgi/alerts

mailto:[email protected]

http://clincancerres.aacrjournals.org/content/24/9/2182


Pan-Cancer Molecular Classes Transcending Tumor Lineage … · Biology of Human Tumors Pan-Cancer Molecular Classes Transcending Tumor Lineage Across 32 Cancer Types, Multiple Data

Documents