tO li

a

,

e tu ae gs

with mutations in H3F3A. This gene

encodes the replication-independent his-

These findings build on earlier studies

et al., 2010). Tumors with IDH1mutations

fall primarily into the proneural expression

emphasizes the differences between

ciated with a methylation pattern that is

distinct from both that of H3F3A G34

in IDH1 and H3F3A K27M mutant groups(Noushmehr et al., 2010; Parsons et al.,

2008; Verhaak et al., 2010). Patients

without IDH1 mutation are older and

this tumor into six subgroups, two of

which are primarily pediatric, and the

remainder of which are primarily adult.

to differentiation signals, which are hall-

marks of pre-cancerous cells (Turcan

et al., 2012). Sturm et al. (2012) show inhave more rapidly progressive disease,

whereas those with IDH1 mutation

frequently have an antecedent low grade

Methylation groups largely correlate with

the pattern of expression profiling that

was previously reported, emphasizing

H3F3A G34 mutated glioblastoma a

similar decrease in expression of OLIG2,

an important neuro-developmental gene.that showed that adult GBM could be

subdivided into three epigenetic sub-

groups, one of which correlated with

mutations in the metabolic gene IDH1

adult and pediatric GBM. Sturm et al.

(2012) now show that an expanded anal-

ysis of the methylome of GBM, which

includes pediatric GBM cases, can divide

(Lu et al., 2012). IDH1mutation also leads

to increased neural stem cell marker

expression and decreased expression of

mature differentiation genes in responsetone H3.3, which predominantly binds

transcriptionally active loci and telomeres.

H3.3 is frequentlymethylatedat or near the

residues that are mutated. Alterations in

histonemethylationaffect theaccessibility

of the associated DNA and may promote

changes in DNA methylation (Turcan

et al., 2012). H3F3A mutations therefore

provide a potential mechanism underlying

the global methylation changes observed

in these pediatric GBM subgroups.

profile, and in epigenetic analyses, CIMP+

tumors segregate into that same group

(Noushmehr et al., 2010).

IDH1 is rarely altered in pediatric

GBMs. Indeed, pediatric GBM contain

significantly different genomic alterations

from adult tumors (Paugh et al., 2010).

The recent discovery that mutations in

the histone gene H3F3A occur predomi-

nantly in pediatric high grade gliomas

(Schwartzentruber et al., 2012) further

and the CIMP+ phenotype associated

with IDH1 mutation.

There is also increasing evidence that

alterations in H3F3A and IDH1 interfere

with the normal differentiation of neural

progenitors. The abnormal metabolite

2HG produced by mutant IDH1 leads

to increased repressive methylation at

H3K27 and inhibits immortalized neural

cell differentiation, suggesting a potential

common mechanism for tumorigenesisCancer Cell

Previews

Methylome AlteraNew Therapeutic

Eric H. Raabe1,2,* and Charles G. Eberh1Division of Pediatric Oncology2Division of PathologyJohns Hopkins University School of Medicine*Correspondence: eraabe2@jhmi.eduhttp://dx.doi.org/10.1016/j.ccr.2012.10.001

In this issue of Cancer Cell, Sturmforme divide adult and pediatric tprofiles, and most importantly, difftherapeutics for this dreaded disea

Glioblastoma multiforme (GBM) is the

most aggressive brain tumor and is asso-

ciated with very poor overall survival.

GBM occurs in adults much more fre-

quently than in children or adolescents,

and pediatric GBM has genetic abnormal-

ities that make it distinct from adult

tumors, suggesting that although the

microscopic appearance and grim prog-

nosis are shared (Figure 1), pediatric and

adult GBM have different underlying biol-

ogies (Paugh et al., 2010). In this issue of

Cancer Cell, Sturm et al. (2012) show that

pediatric GBM contains two epigenetic

subgroups that are distinct from those

found in adult tumors. The epigenetic

profiles of these groups correlate tightlyions Markpportunities in G

rt2

Baltimore, MD 21287, USA

t al. report that global DNA methylamors into subgroups that have chrent clinical behaviors. These findine.

glioma, are younger, have more frequent

TP53 mutations, and are less likely to

have receptor tyrosine kinase amplifica-

tion (Parsons et al., 2008). The IDH1

mutation is a gain of function alteration

that creates a novel onco-metabolite, 2-

hydroxyglutarate (2HG), which interferes

with the normal cellular methylation

machinery. This in turn leads to wide-

spread increases in global methylation

known as the CpG-island methylator

phenotype (CIMP) (Turcan et al., 2012).

mRNA expression profiling identified four

subgroups of adult GBM (proneural,

neural, classical, and mesenchymal) and

determined that each subgroup contains

distinct pathway alterations (VerhaakCancer Cell 22oblastoma

ion patterns in glioblastoma multi-racteristic DNA mutations, mRNAs suggest novel opportunities for

the importance of epigenetic regulation

in GBM (Verhaak et al., 2010).

Sturm et al. (2012) show that three of

the epigenetic subgroups (two pediatric

and one adult) correlate with mutations

in H3F3A and IDH1. H3F3A and IDH1

mutations are non-overlapping, sug-

gesting that they represent different paths

to achieve widespread alterations in

genomic methylation. The two mutations

in H3F3A are seven amino acid residues

apart and are associated with dramati-

cally different methylation profiles. The

H3F3A G34 mutation is associated with

a hypomethylation phenotype, which is

most prominent at the ends of chromo-

somes. The H3F3A K27 mutation is asso-, October 16, 2012 2012 Elsevier Inc. 417

Loss of OLIG2 is associated

with increased methylation

at the OLIG2 locus, which

occurs despite the global hy-

pomethylation of the genome

in H3F3A G34 mutated glio-

blastoma. Although there

are intriguing similarities in

the pathways between IDH1

., 2007). It

tification of

nslate into

for these

hanges in

entified by

epigenetic

therapeutic

logy Group

ring chro-

eacetylase

560). How-

have not

pediatric

that the

GBM is

chromosome ends and features the

alternative lengthening of telomeres

(ALT) phenotype. Although the precise

mechanism of ALT remains unknown,

the presence of promyelocytic leukemia

bodies and evidence of heterologous

recombination between telomeres pro-

vide enticing targets for therapy (Heaphy

et al., 2011). The recent increases in our

understanding of the genetics of GBM,

and in particular, the integration of

both pediatric and adult tumors into an

overall epigenomic classification scheme

by Sturm et al. (2012), set the stage for

translation of molecular advances into

improved care for patients with this

disease.

REFERENCES

Heaphy, C.M., Subhawong, A.P.,Hong, S.M., Goggins, M.G., Mont-gomery, E.A., Gabrielson, E., Netto,G.J., Epstein, J.I., Lotan, T.L.,Westra, W.H., et al. (2011). Am. J.Pathol. 179, 16081615.

Louis, D.N., Ohgaki, H., Wiestler,O.D., and Cavenee, W.K. (2007).

wicz-Brice, M., Zhang, J., Bax, D.A., Coyle, B.,Barrow, J., Hcol. 28, 3061

SchwartzentrJones, D.T., Pbasso, A.M.,Nature 482, 2

Sturm, D., WiD.-A., JonesTonjes, M., SiCell 22, this is

Turcan,S.,RoF., Yilmaz, E.Ward, P.S., et

Verhaak, R.GV., Qi, Y., WiGolub, T., MAtlas Researc98110.

Si dOm inh oua tom 0n oat one aM

Cancer Cell

Previewsin the literature (Louis et al

remains to be seen how iden

methylation alterations will tra

improved therapeutic options

patients. The significant c

global methylation patterns id

Sturm et al. (2012) make

altering agents an attractive

modality. The Childrens Onco

is currently investigating alte

matin structure with histone d

inhibitors in GBM (NCT01236

ever demethylating agents

been widely investigated in

GBM.

Sturm et al. (2012) show

G34 subgroup of pediatricand H3F3Amutated glioblas-

toma, the disparate methyla-

tion signatures and clinical

course of these three groups

involving children and young

adults suggest that the

mechanism by which these

alterations lead to transfor-

mation of normal cells will

require extensive study.

Although this study shows

that different GBM meth-

ylation subgroups have

somewhat variable clinical

courses, the overall out-

comes of children and adults

with GBM is extremely poor,

with few long-term survivors reported

Figure 1.DivergentA pediatricleft) and a(right) haveeosin photoissue of Caarise in anand behavbar = 100 m418 Cancer Cell 22, October 16, 2012 2012hypomethylated preferentially at the

milar Appearances, Different Methylomes, anutcomesidline GBM of the pons (a diffuse intrinsic pontemispheric glioblastoma multiforme arising in a ysimilar histologic appearance on high power hemaicrographs. Yet, as demonstrated by Sturm et al. (2cer Cell, these two tumors affect different patient pomically distinct regions, have different methylatidifferently. For both images, magnification = 2003.Elsevier Inc.argrave, D., et al. (2010). J. Clin. On-3068.

uber, J., Korshunov, A., Liu, X.Y.,faff, E., Jacob, K., Sturm, D., Fonte-Quang, D.A., Tonjes, M., et al. (2012).26231.

tt, H., Hovestadt, V., Khuong-Quang,, D.T.W., Konerman, C., Pfaff, E.,ll, M., Bender, S., et al. (2012). Cancersue, 425437.

hle,D.,Goenka,A.,Walsh, L.A., Fang,, Campos, C., Fabius, A.W., Lu, C.,al. (2012). Nature 483, 479483.

., Hoadley, K.A., Purdom, E., Wang,lkerson, M.D., Miller, C.R., Ding, L.,esirov, J.P., et al.; Cancer Genomeh Network. (2010). Cancer Cell 17,WHO Classification of Tumours ofthe Central Nervous System (Lyon,France: IARC Press).

Lu, C., Ward, P.S., Kapoor, G.S.,Rohle, D., Turcan, S., Abdel-Wahab,O., Edwards, C.R., Khanin, R., Fig-ueroa, M.E., Melnick, A., et al.(2012). Nature 483, 474478.

Noushmehr, H., Weisenberger, D.J.,Diefes, K., Phillips, H.S., Pujara, K.,Berman, B.P., Pan, F., Pelloski,C.E., Sulman, E.P., Bhat, K.P.,et al.; Cancer Genome AtlasResearch Network. (2010). CancerCell 17, 510522.

Parsons, D.W., Jones, S., Zhang, X.,Lin, J.C., Leary, R.J., Angenendt, P.,Mankoo, P., Carter, H., Siu, I.M.,Gallia, G.L., et al. (2008). Science321, 18071812.

Paugh, B.S., Qu, C., Jones, C., Liu, Z., Adamo-

e glioma,ng adultxylin and12) in thispulations,profiles,nd scale

Ttio E-

s dt res e

CDKs in DNA replication and chromo-

specific and dependent on specific onco-

Because the genetic modification was

obtained using a small-molecule inhibitor

Both lung and breast cancer cells display

display high levels of cyclin D2/D3 andgenic alterations. These studies were per-

formed using germline cyclin D1 knockout

of CDK4/6 (PD 0332991) currently being

studied in clinical trials. Importantly, this

cyclin D3-CDK6 complexes in agree-

ment with the modulation of cyclinsome segregation.

The pioneering work by P. Sicinskis

group in 2001 (Yu et al., 2001) showed

that a specific interphase cyclin, cyclin

D1, was required for RAS- and HER2-

induced mammary tumors, but was

dispensable for WNT- or MYC-induced

mammary tumors. This work showed for

the first time that the therapeutic value

of a cell cycle regulator may be cell-type

present since conception, the remaining

question was whether the acute inhibition

of cyclin D1-CDK complexes would be

effective in already developed breast

tumors. A new study in this issue of

Cancer Cell by Choi et al. (2012) now

shows that conditional genetic ablation

of cyclin D1 in adult mice that bear tumors

inhibits the growth of HER2-positive

mammary carcinomas. A similar effect is

characteristics of cell cycle arrest and

senescence after genetic ablation or

chemical inhibition of cyclin D1-CDK4

complexes (Choi et al., 2012; Puyol

et al., 2010). Similar treatments, however,

result in apoptotic cell death in mouse

and human leukemic cells (Choi et al.,

2012; Sawai et al., 2012). This response

seems to be associated with the

finding that Notch1-induced tumorsCell Cycle-Based

Marcos Malumbres1,*1Cell Division and Cancer Group, Spanish Na*Correspondence: malumbres@cnio.eshttp://dx.doi.org/10.1016/j.ccr.2012.09.024

Targeted therapies directed againIn this issue of Cancer Cell, Choi ecyclin-dependent kinase complextumors and leukemias.

The real problem is to understand

why cancer cells grow when their

normal counterparts would not.

(Hunt, 2008)

Tumor cells invariably display defects in

the machinery that controls the cell divi-

sion cycle. Yet, whether the cell cycle is

a useful therapeutic target is still being

intensely debated mostly due to the

essential role of this process in tissue

homeostasis and the difficulties of finding

a therapeutic window (Malumbres and

Barbacid, 2009). Progression through

the cell cycle is driven by several protein

kinases, including cyclin-dependent

kinases (CDKs). Molecular analysis of

tumor cells has provided strong evidence

that the activity of these kinases is de-

regulated in human tumors suggesting

their potential therapeutic use. However,

the clinical benefit of the first generation

of CDK inhibitors has been limited due to

toxicity and lack of specificity, possibly

derived from the critical function of

Cancer Cell

Previewsmice, which raised some important

questions at that time. Since cyclin D1

knockout mice displayed minor defects

in mammary gland development, it was

argued that the effect observed could be

due to special requirements for cyclin D1herapies Move Fo

nal Cancer Research Centre (CNIO), Madrid

t cell cycle regulators have beenal. and Sawai et al. rekindle the theby demonstrating their requirem

in the cell of origin of these tumors. Two

studies in 2006 suggested that this was

not the case, as a similar therapeutic

benefit was observed using CDK4-defi-

cient mice or knockin mice expressing

a mutant cyclin D1, which does not bind

CDKs but maintains other CDK-indepen-

dent functions (Landis et al., 2006; Yu

et al., 2006). None of these models

displayed defects in mammary gland

development; yet, they were resistant to

HER2-induced mammary gland carci-

nomas. In addition, these studies sug-

gested that CDK4, one of the kinase

partners of cyclin D1, was the critical

enzymatic activity to be targeted in

HER2-positive mammary gland carci-

nomas. Subsequent studies by Barbacid

and colleagues (Puyol et al., 2010)

showed that CDK4, but not CDK2 or

CDK6, was critical for the development

of K-RAS-induced lung tumors, whereas

lack of this interphase kinase did not

affect the development of lungs in germ-

line knockout mice.effect is not accompanied by toxicities

associated with the acute inhibition of cy-

clin D1-CDK complexes in adult individ-

uals (Puyol et al., 2010; Choi et al., 2012;

Sawai et al., 2012), suggesting that side-

effects of such treatments will be minimal.

Cancer Cell 22rward

28029, Spain

ifficult to translate into the clinic.apeutic value of inhibiting specificnts in the maintenance of breast

Based on the previous results, cyclin

D1-CDK4 complexes seem to be critical

targets in the proliferation of epithelial

tumors, at least breast and lung carci-

nomas (Figure 1). What about other

interphase CDK complexes? Previous

studies established a critical role for

cyclin D3 and, at least partially, CDK6

in lymphocytes and in the development

of T cell leukemias (Hu et al., 2009;

Sicinska et al., 2003). The relevance of

these findings in cancer therapy has

now been addressed using conditional

knockouts or acute treatments with

kinase inhibitors. Genetic ablation of

cyclin D3 or inhibition of CDK4/6

complexes in Notch1-induced T cell

acute lymphoblastic leukemia (T-ALL)

results in tumor regression without

causing major abnormalities in other

tissues (Choi et al., 2012; Sawai

et al., 2012; in this issue of Cancer Cell).

Importantly, the response of leukemic

cells to these treatments is dramatically

different from that of epithelial cells.D3 expression by the Notch1 pathway

(Choi et al., 2012; Joshi et al., 2009).

In fact, Notch1-negative T-ALLs or

other types of leukemic cells did not

undergo apoptosis following CDK4/6

inhibition (Choi et al., 2012). Whether the

, October 16, 2012 2012 Elsevier Inc. 419

susceptibility to apoptosis is

provided by the activation of

Notch1 or the specific

inhibition of cyclin D3-(and

possibly CDK6) complexes

is not clear at present. Inter-

estingly, cyclin D2 cannot

compensate for the lack of

cyclin D3 when expressed

from the Ccnd3 locus sug-

gesting intrinsic differences

st tissues

tration of

f mamma-

ts on the

activities

Barbacid,

in breast,

rong case

r contexts

ecific cy-

ave thera-

on these

that many

human tumors may be sensitive to

specific cyclin-CDK complexes, as long

as we identify the specific complexes

that mediate the response to specific

oncogenic pathways in each specific

cell type. A long road in which difficult

questions such as why cancer cells

grow when their normal counterparts

would not (Hunt, 2008) need to be

addressed in a cell-type- and onco-

gene-specific manner. Cancer patients

will undoubtedly benefit from these

studies.

REFERENCES

Choi, Y.J., Li, X., Hydbring, P.,Sanda, T., Stefano, J., Christie,A.L., Signoretti, S., Look, A.T.,Kung, A.L., von Boehmer, H., andSicinski, P. (2012). Cancer Cell 22,this issue, 438451.

Hu, M.G., Deshpande, A., Enos, M.,Mao, D., Hinds, E.A., Hu, G.F.,Chang, R., Guo, Z., Dose, M., Mao,C., et al. (2009). Cancer Res. 69,810818.

Strikoudis, A.M.A., Clark, Missue, 45246

Sicinska, E., Aowski, C., YuGeng, Y., vo(2003). Cance

Yu, Q., Geng411, 101710

Yu, Q., Sicinsgozdzon, A., KHarris, L.N.,Cancer Cell 9

A ncis Cp ied cifiK y

fo orsill c

complexes in similar pathologies induced by other oncogenes (e.g., mammarygland or lung tumors without activation of the RAS pathway) or in othertumor types.

Cancer Cell

Previewsneered mice with cell cycle

mutations. The discovery

that individual interphase

cyclins or CDKs were dis-

pensable for the develop-

ment and homeostasis of mo

was considered a demons

the developmental plasticity o

lian tissues and raised doub

usefulness of inhibiting these

in tumors (Malumbres and

2009). The recent studies

lung, and T-ALL make a st

for the identification of cellula

in which the inhibition of sp

clin-CDK complexes may h

peutic value (Figure 1). Based

results, one could proposein the function of these two

proteins (Sawai et al., 2012).

The molecular basis for the

differential response to

CDK4/6 inhibitionsenes-

cence versus apoptosis

may be considered a new

avenue of high interest in

the therapeutic evaluation of

these cell cycle regulators.

It is been a long road20

yearssince the initial gene-

ration of genetically-engi-

Figure 1.MalignanSince theCDK6 combe discardactivate CDpensableresearch w420 Cancer Cell 22, October 16, 2012 2012n Initial Road Map for the Treatment of Humaes with Specific CDK4/6 Inhibitorsmall-molecule inhibitor PD 0332991 inhibits bothlexes, the participation of CDK4 in human leukemat this time. Similarly, it is not clear which spe

4 in lung tumors, although this information is likelr the selection of small-molecule kinase inhibitbe necessary to evaluate the relevance of specificElsevier Inc.Puyol, M., Martn, A., Dubus, P.,Mulero, F., Pizcueta, P., Khan, G.,Guerra, C., Santamara, D., andBarbacid, M. (2010). Cancer Cell18, 6373.

Sawai, C., Freund, J., Oh, P.,Ndiaye-Lobry, D., Bretz, J.C.,, Genesca, L., Trimarchi, T., Kelliher,., et al. (2012). Cancer Cell 22, this5.

ifantis, I., Le Cam, L., Swat, W., Bor-, Q., Ferrando, A.A., Levin, S.D.,n Boehmer, H., and Sicinski, P.r Cell 4, 451461.

, Y., and Sicinski, P. (2001). Nature21.

ka, E., Geng, Y., Ahnstrom, M., Za-ong, Y., Gardner, H., Kiyokawa, H.,Stal, O., and Sicinski, P. (2006)., 2332.Hunt, T. (2008). Cell Cycle 7, 37893790.

Joshi, I., Minter, L.M., Telfer, J.,Demarest, R.M., Capobianco, A.J.,Aster, J.C., Sicinski, P., Fauq, A.,Golde, T.E., and Osborne, B.A.(2009). Blood 113, 16891698.

Landis, M.W., Pawlyk, B.S., Li, T.,Sicinski, P., and Hinds, P.W.(2006). Cancer Cell 9, 1322.

Malumbres, M., and Barbacid, M.(2009).Nat.Rev.Cancer9,153166.

DK4 andas cannotc cyclinsto be dis-. Furtheryclin-CDK

yil e

a wd me n

protein 4 gene (Fabp4, also called aP2)

The authors validated tumors as true

et al. (2012) was shown to be active in

of SmoM in myogenic lineages, Hatley

given that most Hedgehog pathway inhib-

develop into both muscle and brown fateRMS by histology and immunohisto-

chemistry (Desmin, MyoD, and Myoge-

et al. (2012) activated the SmoM2 allele in

early muscle development, employing

tissue via PRDM16 and PPARg signaling

(Seale et al., 2008). Furthermore, satelliteto drive Cre. Eighty percent of aP2-Cre;

SmoM2/+ mice developed eRMS in the

head and ventral neck by 2 months of

age. aP2-Cre Cdkn2aFlox/Flox mice did

not develop eRMS, whereas deletion of

Cdkn2a in aP2-Cre;SmoM2/+ mice de-

creased latency and increased tumor

penetrance, with all mice having tumors

by 55 days. This finding implicates

Cdkn2a locus loss as a secondary factor

(disease modifier) of eRMS progression.

adipose tissue but not in skeletal muscle,

at least as evidenced from recombination

in the sternocleidomastoid muscle (SCM)

of aP2-Cre;R26-LacZ reporter mice and

aP2-Cre;R26-YFP mice and whole-tissue

examination. Interestingly, it was found

that eRMS tumors in situ were completely

surrounded by non-neoplastic adipose

tissue adjacent to and clearly separated

from the SCM at P14.

To compare the tumorigenic potential

itors currently in clinical trials are Smooth-

ened antogonists, this model may have

highest value as a genetic model of

eRMS for the time being and as a preclin-

ical therapeutic model only when GLI

inhibitors emerge as clinical candidate

agents.

Taking a look onto mesenchymal

progenitor cell biology, the plasticity of

these stem cells may not be so surprising.

It is known that Myf5 expressing cell canThe Not-so-Skinn

Ken Kikuchi1 and Charles Keller1,*1Pediatric Cancer Biology Program, Pape FamPortland, OR 97239 USA*Correspondence: keller@ohsu.eduhttp://dx.doi.org/10.1016/j.ccr.2012.09.018

The childhood cancer embryonal rhthis issue of Cancer Cell, Hatley anorigin for Sonic Hedgehog-drivenmouse model.

One of the most intriguing clinical and

scientific questions in pediatric oncology

is how rhabdomyosarcoma, a tumor with

a myogenic phenotype, can arise in tissue

without hypaxial-like skeletal muscle

elements. This conundrum is especially

evident for the embryonal subtype of

rhabdomyosarcoma (eRMS), which can

arise from the salivary glands, skull base

(parameninges), biliary tree, and genito-

urinary tract (bladder/prostate) (Gurney

et al., 1999; Shapiro and Bhattacharyya,

2006). The work by Hatley et al. (2012) in

this issue of Cancer Cell begins to

address this conundrum by highlighting

the plasticity of cells traditionally thought

to be in the adipogenic lineage.

In this paper, Hatley et al. (2012) condi-

tionally activate a mutant Smoothened

(SmoM2) allele to drive Hedgehog sig-

naling in the adipogenic lineage using

a 5.4 kb promoter/enhancer fragment

from the adipocyte fatty acid binding

Cancer Cell

Previewsnin). Similarly, an embryonal muscle

gene signature (MyoD1, Myogenin, Pax7,

Myf5, Myh3, and Myh8) was evident by

RT-PCR assay. To further affirm the diag-

nosis, Hatley et al. (2012) performed gene

expression profiling of aP2-Cre;SmoM2/+on Muscle Cance

y Pediatric Research Institute, Department of P

bdomyosarcoma can arise in tissuecolleagues report that non-skeletalmbryonal rhabdomyosarcoma in a

tumors in comparison to human eRMS

and tumors of previously reported mouse

eRMS models. Overall, 67% of the probe

pairs and 58% of the ortholog gene

pairs showed agreement in gene expres-

sion between their mouse tumors and

previously published eRMS models (Ru-

bin et al., 2011) or human tumors, respec-

tively. Hatley et al. (2012) concluded that,

despite being of an adipogenic lineage of

origin, aP2-Cre;SmoM2/+ mouse tumors

are an accurate preclinical eRMS model

based on these histological and transcrip-

tome parameters.

The most interesting aspect of this

paper is that these tumors originated

from aP2 expressing cells. Previously,

activity of the aP2 promoter in transgenic

mice has been documented as specific

for the adipocyte lineage and to be

distinct from any skeletal muscle lineage

(Urs et al., 2006). Moreover, the specific

aP2-Cre transgenic line used by HatleyPax3-, Myf5-, or MyoD1-Cre. All of these

crosses resulted in embryonic lethality

without tumor formation. However, acti-

vation of SmoM2 with Myogenin-Cre,

which is specific for myoblast-stage

muscle differentiation, resulted in tongue

Cancer Cell 22r

diatrics, Oregon Health & Science University,

ithout skeletal muscle elements. Inuscle progenitors can be a cell ofadipocyte-restricted conditional

tumors in 100% of the mice. Mice with

activation of SmoM2 in terminally differen-

tiated skeletal muscle (MCK-Cre) were

viable with no evidence of tumorigenesis.

This result is the best by far in asking

whether differentiated versus differenti-

ating myofibers can transform into rhab-

domyosarcoma. To address the tumori-

genic potential of SmoM2 in postnatal

muscle stem cells (satellite cells), a Pax7-

CreERT2 was employed, and mice were

aged until 150 dayswithout develop-

ment of tumors. It would have been

intriguing to see if prior reports of non-

myogenic sarcomas arising from a satel-

lite cell lineage (Rubin et al., 2011) would

have been observed if the mice had

been aged longer. Another caveat of

these findings is that the promoter of

SmoM2 in this system was Rosa26, not

the native Smo promoter, and it is there-

fore difficult to say that this is amolecularly

physiological eRMS model. Furthermore,cells can undergo adipogenesis under

certain experimental conditions (Asakura

et al., 2001; Shefer et al., 2004) or may

apparently undergo transdifferentiation

into fibroblasts (Brack et al., 2007).

Hedgehog activation can also reduce fat

, October 16, 2012 2012 Elsevier Inc. 421

formation via PPARg (Suh

et al., 2006). Such reports

lly in rhab-

p53 status

was unex-

of interest

s of these

969 histor-

Fraumeni

syndrome

to germline

S tumors

and neck,

lso in the

and retro-

is cranial-

of specific

and neck

eRMS patients it models, and closer

examination of the preneoplastic and

early neoplastic lesions will invariably

help us understand the microenvironment

in which these tumor initiating cells

transform.

Overall, the heterogeneity in human

rhabdomyosarcoma phenotypes may

result from the balance between genetic

factors (mutational profile of the tumor

including the initial mutation[s] and modi-

fiers) and epigenetic factors (the cell of

origin). In an era of precision medicine,

this transgenic model representing the

subset of human head and neck ERMS

not derived from myogenic precursors

may be particularly valuable in defining

therapeutic targets for select

patients. Applying knowledge

Graff, J., GaCancer Cell 2

Rubin, B.P., ND.P., Pal, R.,B.R., Chen,177191.

Seale, P., BjoS., Kuang, S.,H.M., Erdjum454, 961967

Shapiro, N.LOtolaryngol. H

Shefer, G.,Reuveni, Z. (2

Suh, J.M., Gaand Graff, J.M

Urs, S., Harr(2006). Transg

Figure 1. Model of Skeletal Myogenesis and Possible CellularOrigins of eRMSPostnatal muscle maintenance and regeneration are regulated by Pax3, Pax7,and muscle regulatory factors (MyoD, Myf5, Myogenin, and Myf6). Studiesdescribed in the text suggest that the cells of origin for eRMS include differen-tiating myoblasts and adipocytes.

Cancer Cell

Previewsof Hedgehog signaling was

evident by RT-PCR studies

of aP2-Cre;SmoM2/+ tumors,

yet, by contrast, less than

30% of human eRMS exhibit

a gene expression signature

consistent with a Hedgehog

pathway on overdrive state

(Rubin et al., 2011), and only

rarely is Hedgehog overdrive

a solitary signaling abnor-

mality. Instead, p53 loss of

function was most frequent

(Rubin et al., 2011)implying

that p53 loss precedes

Hedgehog overdrive tempora

domyosarcomagenesis. The

of aP2-Cre;SmoM2/+ tumors

plored, but it would have been

to have known the p53 statu

tumors, particularly given the 1

ical precedent of Li and

describing a familial eRMS

now known to be attributable

p53 mutation.

In the Hatley model, eRM

occurred from only the head

yet human eRMS occur a

urogenital tract, extremities,

peritoneum. Nevertheless, th

oriented eRMS model may be

interest for the subset of headlend one to speculate that

the origin of aP2-Cre;

SmoM2/+ mice tumors might

be cells that have undergone

a transdifferentiation event

from the adipocyte to the

myogenic lineage (Figure 1).

Hatley et al. (2012) recapit-

ulate another important find-

ing by gene expression pro-

filing that aP2-Cre;SmoM2/+

tumors show an activated

satellite cell phenotype re-

flective of the eRMS tumor

pathology rather than the

lineage of origin for the

tumor. Certainly, activation422 Cancer Cell 22, October 16, 2012 2012 Elsevier Inc.Brack, A.S., Conboy, M.J., Roy, S.,Lee, M., Kuo, C.J., Keller, C., andRando, T.A. (2007). Science 317,807810.

Gurney, J.G., Young, J.L., Jr.,Roffers, S.D., Smith, M.A., andBunin, G.R. (1999). Soft TissueSarcomas. In Cancer Incidence andSurvival among Children andAdolescents: United States SEERProgram 19751995, NationalCancer Institute, SEER Program.NIH Pub. No. 994649. Ries LAG,SM, J.G. Gurney, M. Linet, T. Tamra,J.L. Young, and G.R. Bunin, eds.(Bethesda, MD: National CancerInstitute).

Hatley, M.E., Tang, W., Garcia, M.R.,Finkelstein, D., Millay, D.P., Liu, N.,lindo, R.L., and Olson, E.N. (2012).2, this issue, 536546.

ishijo, K., Chen, H.I., Yi, X., Schuetze,Prajapati, S.I., Abraham, J., Arenkiel,Q.R., et al. (2011). Cancer Cell 19,

rk, B., Yang, W., Kajimura, S., Chin,Scime`, A., Devarakonda, S., Conroe,ent-Bromage, H., et al. (2008). Nature.

., and Bhattacharyya, N. (2006).ead Neck Surg. 134, 631634.

Wleklinski-Lee, M., and Yablonka-004). J. Cell Sci. 117, 53935404.

o, X., McKay, J., McKay, R., Salo, Z.,. (2006). Cell Metab. 3, 2534.

ington, A., Liaw, L., and Small, D.enic Res. 15, 647653.from these specialized pre-

clinical models to clinical use

will no doubt require novel

statistical designs for clinical

trials, given that the overall

annual incidence of pediatric

rhabdomyosarcoma is only in

the hundreds, not thousands.

Personalized rhabdomyosar-

coma science inches ever

closer.

REFERENCES

Asakura, A., Komaki, M., and Rud-nicki, M. (2001). Differentiation 68,245253.

o h

, n

hw

rk r. i tnd c

genes and tumor suppressors have been

inefficient because most of the energy

informatics study on diffuse large B cell

approach revealed three major DLBCL

DLBCLs. PPARg antagonists suppressed

same type. This point should be empha-that could be generated from glucose

is lost when the cell secretes lactate.

subtypes: the OxPhos subset; the B cell

receptor (BCR) subset, characterized by

sized, because it tempers the paradigm

that tumors are monolithic in their prefer-demonstrated to stimulate glycolysis

(Ward and Thompson, 2012).

Cells have two ways to produce adeno-

sine triphosphate (ATP) for energy: gly-

colysis and oxidative phosphorylation

(OxPhos) (Figure 1). In glycolysis, glucose

is converted to pyruvate, generating

NADH from NAD+ and ATP from ADP. If

the pyruvate is reduced to lactate, NAD+

is regenerated and glycolysis continues.

Although glycolysis is rapid, it is deemed

lymphoma (DLBCL) revealed that some

30% of these tumors belonged to a

subset defined by high expression of

genes involved in OxPhos (Monti et al.,

2005). DLBCL is the most common non-

Hodgkins lymphoma in Western popu-

lations and is characterized by rapid

growth. Several attempts have been

made to classify DLBCL by molecular

typing, particularly through shared

patterns of gene expression. One such

fatty acid oxidation and killed OxPhos

cells in culture. Inhibitors of fatty acid

oxidation had the same effect, as did

silencing of an enzyme required to

synthesize the antioxidant glutathione.

None of these treatments affected non-

OxPhos cells.

The work illustrates several key prin-

ciples in cancer metabolism. First,

it demonstrates significant metabolic

heterogeneity among tumors of theA Mitochondrial P

Ralph J. DeBerardinis1,*1Childrens Medical Center Research InstituteDallas, TX 75390-8502, USA*Correspondence: ralph.deberardinis@utsouthttp://dx.doi.org/10.1016/j.ccr.2012.09.023

Deregulated energetics is a hallmais unknown. A study by Caro et alis defined by reliance on mitochoimpaired.

Altered metabolism was among the first

cancer biomarkers (Koppenol et al.,

2011). By 1925, Otto Warburg had ob-

served that tumors rapidly took up

glucose and converted it to lactic acid,

which was released into the milieu. This

was a departure from the other organs

he studied, which imported less glucose,

oxidized a higher fraction of it to carbon

dioxide, and secreted much less lactate.

This remarkable phenotype, called the

Warburg effect, provides the rationale

for two modern imaging technologies in

cancer diagnosis: 18fluoro-2-deoxyglu-

cose positron emission tomography

(FDG-PET) to identify tumors by rapid

glucose import and 1H magnetic reso-

nance spectroscopy to identify areas of

elevated lactate. Over the last two

decades, the Warburg effect has been

mechanistically tied to the molecular

basis of transformation, as tumor-pro-

moting mutations in many different onco-

Cancer Cell

PreviewsIn contrast, OxPhos is highly efficient.

When substrates like pyruvate are

oxidized in the mitochondria, reducing

equivalents are delivered to the elec-

tron transport chain, creating a proton

gradient coupled to ATP synthesiswer Play in Lymp

University of Texas Southwestern Medical Ce

estern.edu

of malignancy, but metabolic heten this issue of Cancer Cell demonsrial energy generation and is sele

(Figure 1). This system yields nearly 20

times as many ATP molecules per

glucose as the Warburg effect. Other

oxidizable substrates, like fatty acids,

produce an even higher energy yield.

Warburg considered tumor glycolysis

to be a metabolic anomaly, famously

postulating that it was the consequence

of irreversible defects in respiration (i.e.,

OxPhos) which were the root cause of

cancer (Warburg, 1956). This model has

been disproven as a general principle,

because most cancer cells do not have

static defects in their respiratory

machinery, and more pointedly, because

nonmalignant cells also display the

Warburg effect if stimulated to grow

(Wang et al., 1978). Nevertheless, the

concept that glycolysis is a universal

feature of aggressive tumor growth, and

that OxPhos is counter-productive to

tumorigenesis still pervades the literature.

It was therefore surprising when a bio-expression of genes involved in B cell

receptor signaling; and the host response

(HR) subset, expressing markers of infil-

trating inflammatory cells (Monti et al.,

2005). Now, a study in this issue of

Cancer Cell (Caro et al., 2012) examines

Cancer Cell 22oma

ter, 5323 Harry Hines Boulevard,

ogeneity among individual tumorsrates that a subset of lymphomastively killed when this pathway is

the metabolic phenotypes of these sub-

types and found that OxPhos cell lines

and primary biopsies contained elevated

expression of many mitochondrial pro-

teins. OxPhos cells displayed enhanced

glucose oxidation relative to lactate

formation, better defenses against oxida-

tive stress, and a robust ability to oxidize

fatty acids in the mitochondria. Providing

exogenous fatty acids stimulated survival

and growth in OxPhos cells, but not in

cells from other subtypes. Overall, cells

from OxPhos tumors generated a higher

fraction of their ATP in the mitochondria,

and this activity was required for cellular

fitness.

Importantly, inhibition of mitochondrial

metabolism selectively killed OxPhos

cells. The nuclear receptor peroxisome

proliferator-activated receptor gamma

(PPARg), which regulates the expression

of many genes encoding mitochondrial

proteins, was overexpressed in OxPhosence for glycolytic energy production.

Second, metabolic heterogeneity is pro-

grammed, and it persists when cells are

adapted to culture. This argues strongly

against the concept that the glycolytic

phenotype of most cancer cell lines is

, October 16, 2012 2012 Elsevier Inc. 423

(Fig-

ority

cted

des

of

hen

ther

glio-

sion

ito-

cur-

factors that stimulate cell growth and

may contribute to tumorigenesis (Wein-

berg et al., 2010).

It remains to be seen whether the

OxPhos phenotype is driven by tumori-

genic mutations or other features of

DLBCL biology, and whether it bestows

some context-specific advantage to this

subset. Regardless of the mechanism, it

is encouraging that molecular phenotyp-

ing led to the discovery of metabolic

R.M., Mashimo, T., Raisanen, J., Hatanpaa,K.J., Jindal, A., Jeffrey, F.M., Choi, C.,Madden, C., et al. (2012). NMR Biomed.

WangScienc

Warbu

Ward,Cell 2

WeinbWeinbman,del, N8788

Zu, X.

ta

CitrateOxaloacetate

ATP

ATPATPATP

ATP

Reducing equivalents

ATP

ATPATP

ElectronTransport Chain

TCACycle

ADP, Pi

Glycolysis OxPhosfraction of ATP generation

Non-OxPhoscells

OxPhoscells

mitochondrion

Figure 1. Variable Reliance on OxidativePhosphorylation in DLBCLCells produce ATP from glycolysis (blue pathway) and oxida-tive phosphorylation (OxPhos, green). Cancer cell lines varyaccording to what fraction of their total ATP comes fromeach pathway, as shown on the spectrum at the bottom. Asubset of DLBCL tumors was defined by high expression ofgenes related to OxPhos. These cells produced most of theirATP from OxPhos and required this pathway for survival andgrowth. Non-OxPhos cells, which did not share this geneexpression signature, produced a higher fraction of their ATPfrom glycolysis, and were resistant to OxPhos inhibition.Abbreviation: TCA, tricarboxylic.

Cancer Cell

Previewsclinical description of the OxPhos

DLBCL is lacking, patients with

these tumors do not appear to

survive any longer than those with

other types (Monti et al., 2005). The

aggressive course of DLBCL as a

whole suggests that a preference

for OxPhos does not prevent rapid

cell growth in the large subset

of tumors using that metabolic plat-

form. Thus, metabolic programming

to use the mitochondria rather than

glycolysis as themajor site of energy

formation is an effective strategy for

cancer cell growth.

The findings also emphasize that

cancer cells, like normal cells, use

multiple pathways concurrently to

produce energy (Figure 1). The

vast majority of cancer cells use

both glycolysis and OxPhos

together, although the balance

between the two varies widely (Zu

and Guppy, 2004). DLBCL cells

are no exception. Despite their

ability to oxidize pyruvate and

fatty acids in the mitochondria,

OxPhos cells still consumed glu-

cose and secreted a modest

amount of lactate. Conversely, the

non-OxPhos cells still produced

half of their ATP in the mitochondria

ure 1). Along these lines, a large maj

of DLBCL tumors are readily dete

by FDG-PET, and this likely inclu

OxPhos tumors. Thus, avid uptake

glucose analogs may occur even w

the mitochondria are fully active. Ano

recent study on metabolic flux in

blastoma reached this same conclu

(Maher et al., 2012). Besides ATP, m

chondria produce biosynthetic prean unavoidable consequence of

culture conditions. Even the

suppression of OxPhos in other

DLBCL subtypes was programmed,

because these cells could be stimu-

lated to activate fatty acid oxidation

simply by silencing Syk, a kinase in

BCR signaling. Third, high rates of

OxPhos can support growth in

cancer cell lines. This may also be

true in vivo. Although a complete

Lacsors, reactive oxygen species, and other

424 Cancer Cell 22, October 16, 2012 2012Glucose

Pyruvate

NAD+

NADH

te

Acetyl-CoA

Fatty Acids

ADP, Pi

ATPvulnerabilities in DLBCL. In a sense, the Res. C

Elsevier Inc.Published online March 15, 2012. http://dx.doi.org/10.1002/nbm.2794.

Monti, S., Savage, K.J., Kutok, J.L.,Feuerhake, F., Kurtin, P., Mihm, M., Wu,B., Pasqualucci, L., Neuberg, D., Aguiar,R.C., et al. (2005). Blood 105, 18511861.

, T., Sheppard, J.R., and Foker, J.E. (1978).e 201, 155157.

rg, O. (1956). Science 123, 309314.

P.S., and Thompson, C.B. (2012). Cancer1, 297308.

erg, F., Hamanaka, R., Wheaton, W.W.,erg, S., Joseph, J., Lopez, M., Kalyanara-B., Mutlu, G.M., Budinger, G.R., and Chan-.S. (2010). Proc. Natl. Acad. Sci. USA 107,8793.

L., and Guppy, M. (2004). Biochem. Biophys.currently in use for cancer. New

approaches to image tumor meta-

bolism noninvasively (Kurhanewicz

et al., 2011) may soon make it

possible to stratify patients to

therapeutic regimens solely by

probing metabolism in vivo.

REFERENCES

Caro, P., Kishan, A.U., Norberg, E., Stanley,I.A., Chapuy, B., Ficarro, S.B., Polak, K.,Tondera, D., Gounarides, J., Yin, H.,et al. (2012). Cancer Cell 22, this issue,547560.

Koppenol, W.H., Bounds, P.L., and Dang,C.V. (2011). Nat. Rev. Cancer 11,325337.

Kurhanewicz, J., Vigneron, D.B., Brindle, K.,Chekmenev, E.Y., Comment, A., Cunning-ham, C.H., Deberardinis, R.J., Green, G.G.,Leach, M.O., Rajan, S.S., et al. (2011).Neoplasia 13, 8197.

Maher, E.A., Marin-Valencia, I., Bachoo,new work fortifies Warburgs histor-

ical claim that metabolism is an

actionable biomarker (even if

he would have been surprised by

these particular results). PPARg

antagonism may provide a starting

point for targeting OxPhos tumors,

and it would also be interesting

to determine whether OxPhos pre-

dicts sensitivity to therapiesommun. 313, 459465.

Cancer CellDepartment of Neurosurgery, Medical and Health Science Center,

35Program in Developmental and Stem Cell Biology, Division of Neurosurgery, Arthur and Sonia Labatt Brain Tumour Research Centre,

Hospital for Sick Children, University of Toronto, Toronto, ON M4N 1X8, Canada36Department of Oncogenomics, AMC, University of Amsterdam, Amsterdam 1105 AZ, The Netherlands

37Departments of Neurology and Neurosurgery, Henry Ford Hospital, Detroit, MI 48202, USA38Department of Neuro-Oncology, University of Texas MD Anderson Cancer Center, Houston, TX 77030, USA39Division of Molecular Histopathology, Department of Pathology, University of Cambridge, Cambridge, CB2 0QQ, United KingdomArticle

Hotspot Mutations in H3F3Aand IDH1 Define Distinct Epigeneticand Biological Subgroups of GlioblastomaDominik Sturm,1,42 Hendrik Witt,1,7,42 Volker Hovestadt,2,42 Dong-Anh Khuong-Quang,11,42 David T.W. Jones,1

Carolin Konermann,3 Elke Pfaff,1 Martje Tonjes,2 Martin Sill,4 Sebastian Bender,1 Marcel Kool,1 Marc Zapatka,2

Natalia Becker,4 Manuela Zucknick,4 Thomas Hielscher,4 Xiao-Yang Liu,11 Adam M. Fontebasso,12 Marina Ryzhova,13

Steffen Albrecht,14 Karine Jacob,11 Marietta Wolter,15 Martin Ebinger,16 Martin U. Schuhmann,17 Timothy van Meter,18

Michael C. Fruhwald,19 Holger Hauch,20 Arnulf Pekrun,21 Bernhard Radlwimmer,2 Tim Niehues,22

Gregor von Komorowski,23 Matthias Durken,23 Andreas E. Kulozik,7 Jenny Madden,24 Andrew Donson,24

Nicholas K. Foreman,24 Rachid Drissi,25 Maryam Fouladi,25 Wolfram Scheurlen,26 Andreas von Deimling,5,9

Camelia Monoranu,27 Wolfgang Roggendorf,27 Christel Herold-Mende,8 Andreas Unterberg,8 Christof M. Kramm,28

Jorg Felsberg,15 Christian Hartmann,29 Benedikt Wiestler,10 Wolfgang Wick,10 Till Milde,6,7 Olaf Witt,6,7

Anders M. Lindroth,3 Jeremy Schwartzentruber,30 Damien Faury,11 Adam Fleming,11 Magdalena Zakrzewska,31

Pawel P. Liberski,31 Krzysztof Zakrzewski,32 Peter Hauser,33 Miklos Garami,33 Almos Klekner,34 Laszlo Bognar,34

Sorana Morrissy,35 Florence Cavalli,35 Michael D. Taylor,35 Peter van Sluis,36 Jan Koster,36 Rogier Versteeg,36

Richard Volckmann,36 Tom Mikkelsen,37 Kenneth Aldape,38 Guido Reifenberger,15 V. Peter Collins,39 Jacek Majewski,40

Andrey Korshunov,5 Peter Lichter,2 Christoph Plass,3,41,* Nada Jabado,11,41,* and Stefan M. Pfister1,7,41,*1Division of Pediatric Neurooncology2Division of Molecular Genetics3Division of Epigenetics and Cancer Risk Factors4Division of Biostatistics5Clinical Cooperation Unit Neuropathology6Clinical Cooperation Unit Pediatric Oncology

German Cancer Research Center (DKFZ) Heidelberg, 69120 Heidelberg, Germany7Department of Pediatric Oncology, Hematology, and Immunology8Department of Neurosurgery9Department of Neuropathology10Department of Neurooncology

Heidelberg University Hospital, 69120 Heidelberg, Germany11Departments of Pediatrics and Human Genetics, McGill University12Division of Experimental Medicine, Montreal Childrens Hospital

McGill University Health Center Research Institute, Montreal, QC H3Z 2Z3, Canada13NN Burdenko Neurosurgical Institute, Moscow, 125047, Russia14Department of Pathology, Montreal Childrens Hospital, McGill University Health Center Research Institute,Montreal, QCH3H 1P3, Canada15Department of Neuropathology, Heinrich-Heine-University, 40225 Dusseldorf, Germany16Department of Hematology and Oncology, Childrens University Hospital Tubingen, 72076 Tubingen, Germany17Department of Neurosurgery, University Hospital Tubingen, 76076 Tubingen, Germany18Virginia Commonwealth University, Richmond, VA 23298, USA19Pediatric Hospital, Klinikum Augsburg, 86156 Augsburg, Germany20Pediatric Hospital, Klinikum Heilbronn, 74078 Heilbronn, Germany21Prof.-Hess-Kinderklinik, Klinikum Bremen-Mitte, 28177 Bremen, Germany22Childrens Hospital, Helios Clinics, 47805 Krefeld, Germany23Childrens University Hospital, 68135 Mannheim, Germany24Department of Pediatrics, University of Colorado Denver, Aurora, CO 80045, USA25Division of Oncology, Cincinnati Childrens Hospital Medical Center, Cincinnati, OH 45229, USA26Cnopfsche Kinderklinik, Nurnberg Childrens Hospital, 90419 Nurnberg, Germany27Department of Neuropathology, Institute of Pathology, University Wurzburg, 97080 Wurzburg, Germany28University Childrens Hospital, Martin Luther University Halle-Wittenberg, 06120 Halle, Germany29Institute of Pathology, Department of Neuropathology, Hannover Medical School, 30175 Hannover, Germany30McGill University and Genome Quebec Innovation Centre, Montreal, QC H3A 1A4, Canada31Department of Molecular Pathology and Neuropathology, Medical University of Lodz 92-216 Poland32Department of Neurosurgery, Polish Mothers Memorial Hospital Research Institute, Lodz 93-338 Poland332nd Department of Paediatrics, Semmelweis University, Budapest H-1094 Hungary34 University of Debrecen, H-4032 Debrecen, HungaryCancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc. 425

40Department of Human Genetics, McGill University, Montreal, QC H3Z 2Z3, Canada41These authors contributed equally to this work42These authors contributed equally to this work

*Correspondence: c.plass@dkfz.de (C.P.), nada.jabado@mcgill.ca (N.J.), s.pfister@dkfz.de (S.M.P.)

http://dx.doi.org/10.1016/j.ccr.2012.08.024

is

Htuoioc

.

shown to be associated with a distinct Glioma-CpG-Island

heterogeneity of glioblastoma across all age groups. An over-

view of all GBM samples subjected to various analyses is given

Cancer Cell

Epigenetic and Biological Subgroups of GlioblastomaSignificance

GBM is the most common and also the most devastating brain tumor, with a 5-year survival rate below 10%. We presentstrong evidence that GBM can be subclassified into multiple groups, indistinguishable by histological appearance, butcorrelating with molecular-genetic factors as well as key clinical variables such as patient age and tumor location. We iden-Mutations in a protein complex comprised of H3.3 and ATRX/ in Figure S1A available online.one-third of pediatric GBMs (Schwartzentruber et al., 2012)GBM, which arise from a preceding lower-grade lesion. How-

ever, stepwise transformation from less-malignant gliomas into

GBM rarely occurs in children (Broniscer et al., 2007). Further-

more, IDH1 or IDH2 mutations, which are found in up to 98%

of adult secondary GBM, are very rare in childhood GBM

(

MGB

Col

or s

cale

(bet

a-va

lues

)

0.5

1

ylat

ion

prob

es (n

= 8

,000

)

Met

hyla

ted

Cancer Cell

Epigenetic and Biological Subgroups of GlioblastomaWe investigated a cohort of GBMs from children (n = 59) and

adult patients (n = 77) for genome-wide DNA methylation

patterns using the Illumina 450k methylation array, and comple-

mented our data with unpublished profiles of 74 adult GBM

samples generated by The Cancer Genome Atlas (TCGA)

Consortium (McLendon et al., 2008) (Table S1). Consensus

clustering using the 8,000 most variant probes across the data

set robustly identified six distinct DNA methylation clusters

(Figures 1 and S1B). Based on correlations with mutational

status, DNA copy-number aberrations, and gene expression

signatures, as outlined below, we have labeled these subgroups

IDH, K27, G34, RTK I (PDGFRA), Mesenchymal, and

RTK II (Classic).

A striking finding of this integrated analysis is that H3F3A K27

and G34 mutations were exclusively distributed to the K27

(18/18) andG34 (18/18) clusters, respectively (p < 0.001; Fishers

exact test) (Figure 1). The IDH group contained 88% of IDH1-

mutated tumors (23/26) (p < 0.001) and displayed concerted,

0

21

Age60

90

PDGFRA ampl.EGFR ampl.

CDKN2A del.Chr. 10 lossChr. 7 gain

TP53H3F3A

IDH1TCGA ExpressionTCGA MethylationMethylation cluster

0

G34R/V

MUT

WTNA

K27M

Mut

atio

nsC

ytog

enet

ics

IDH K27 G34 RTK I 'PDGFRA

DN

A m

eth

G-CIMP+Cluster #2Cluster #3

ProneuralNeuralClassicalMesenchymal

TCGA subgroups

Met

hyla

tion

Exp

ress

ion

Unm

ethy

late

d

Figure 1. Methylation Profiling Reveals the Existence of Six Epigenetic

Heatmap of methylation levels in six GBM subgroups identified by unsupervised

represents a probe; each column represents a sample. The level of DNA methylat

(n = 210), patient age, subgroup association, mutational status, and cytogenetic

See also Figure S1 and Tables S1 and S2.

C samples (n = 210) Controlsglobal hypermethylation (Figures 1, 2A, and 2B), thereby ex-

panding the previously described link between IDH1 mutation

and G-CIMP+ tumors to a pediatric setting (Noushmehr et al.,

2010). In contrast, tumors in the G34 cluster specifically showed

widespread hypomethylation across the whole genome, and

especially in nonpromoter regions, when compared with all other

subgroups (Figures 2A and 2B). This suggests the existence of

a more global version of a CpG hypomethylator phenotype

(CHOP), as proposed for a small number of genes in gastric

cancer (Kaneda et al., 2002). More detailed mapping of differen-

tially methylated regions revealed that the hypomethylation

observed in H3F3A G34-mutated tumors was particularly prom-

inent at chromosome ends (Figures 2C and 2D), potentially link-

ing subtelomeric demethylation to alternative lengthening of

telomeres, which is most frequently observed in this subgroup

(Schwartzentruber et al., 2012).

Of note, all mutations inH3F3A and IDH1were mutually exclu-

sive (p < 0.001) (Figure 1). To further test this observation, we

' Mesenchymal RTK II 'Classic'

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

Feta

l nor

mal

bra

in (n

=4)

Adu

lt no

rmal

bra

in (n

=2)

WG

A-D

NA

(n=2

)M

.Sss

I-DN

A (n

=2)

GBM Subgroups

k-means consensus clustering, and control samples as indicated. Each row

ion (beta-value) is represented with a color scale as depicted. For each sample

aberrations are indicated.

ancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc. 427

AB

0

5

10

0

0.2

0.4

0.0e+00 5.0e+06 1.0e+07 1.5e+07 2.0e+07 2.5e+07 3.0e+07

-0.04

0.00

0.04

Distance to chromosome end [bp]

Num

ber o

f pro

bes

(x 1

,000

)N

orm

aliz

ed D

NA

Met

hyla

tion Fr

actio

n C

pG Is

land

G34IDH

K27RTK IMESRTK II

C D

***

IDH

K27

G34

RTK

I

ME

S

RTK

II

Con

trol

0.35

0.40

0.45

0.50

0.55

0.60

Mea

n m

ethy

latio

n (m

ean

beta

val

ue)

******

**

****

***

IDH

K27

G34

RTK

I

ME

S

RTK

II

Con

trol

0.20

0.25

0.30

0.35

******

******

******

******

****

IDH

K27

G34

RTK

I

ME

S

RTK

II

Con

trol

0.45

0.50

0.55

0.60

0.65

0.70

0.75

*

****

***

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Overall

Pro

babi

lty F

n(pr

obes

) Hypomethylation

Hypermethylation

DNA methylation (beta-value)0.0 0.2 0.4 0.6 0.8 1.0

Non-promoter

0.0

0.2

0.4

0.6

0.8

1.0

DNA methylation (beta-value)

K27IDH

G34RTK IMESRTK IIControl

0.0 0.2 0.4 0.6 0.8 1.0

Promoter

0.0

0.2

0.4

0.6

0.8

1.0

DNA methylation (beta-value)

K27IDH

G34RTK IMESRTK IIControl

-0.04

-0.03

-0.02

-0.01

0.00

0.01

0.02

Chromosome end (4Mb)

IDH

K27

G34

RTK

I

ME

S

RTK

II

*********

*** ***Mea

n m

ethy

latio

n (m

ean

beta

val

ue)

15

Figure 2. Global DNA Methylation Patterns in GBM Subgroups

(A) Distinct patterns of global DNA methylation in GBM subgroups as identified by consensus clustering. The empirical cumulative distribution function for DNA

methylation levels (beta-values) is plotted individually for each subgroup.

(B) Overall DNAmethylation levels (mean beta-values) of individual GBMmethylation subgroups. Significant differences (***p < 0.001; **p < 0.01; *p < 0.05) to IDH

and G34 subgroups are indicated.

(C) Upper panel: Probe density in respect of distance to chromosome end. The fraction of probes located within CpG-Islands (red line) remains stable. Lower

panel: Mean methylation value per subgroup within windows of 500kb, normalized to control samples. Individual samples are normalized by the mean overall

methylation value.

(D) Mean methylation value within 4 Mb to the chromosome end normalized to the mean overall methylation value and to control samples. Significant differences

(***p < 0.001) between subgroups compared to G34 tumors are indicated. MES, Mesenchymal.

See also Figure S2.

Cancer Cell

Epigenetic and Biological Subgroups of Glioblastoma

428 Cancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc.

extended the targeted sequencing analysis of H3F3A and IDH1

to include 460 GBM samples from patients covering a broad

age range (Figure S1C; Table S2). Even in this expanded series,

no co-occurring mutations in H3F3A and IDH1 were detected

(p < 0.001), and the age distribution confirmed reported associ-

ations of certain mutations with GBM in children (H3F3A K27),

adolescent patients (H3F3A G34), and young adult patients

(IDH1) (Khuong-Quang et al., 2012; Schwartzentruber et al.,

2012; Yan et al., 2009) (Figure S1C; Table S2). As we have

shown, TP53 mutations largely overlap with H3F3A mutations

in pediatric GBM (Schwartzentruber et al., 2012), similar to the

association of TP53 and IDH1 mutations in adults (Yan et al.,

2009). This observation also holds true in our larger cohort,

with a high enrichment of TP53 mutations in the G34 (18/18),

IDH (22/24), and K27 (13/18) clusters (p < 0.001) (Figure 1).

Since pediatric GBMs have been shown to display a distinct

spectrum of focal copy-number aberrations (CNAs) compared

with their adult counterparts (Bax et al., 2010; Paugh et al.,

2010; Qu et al., 2010), we integrated DNA methylation clusters

with copy-number data derived from the methylation arrays

(Figures 1 and S1D). Interestingly, PDGFRA amplification was

significantly more common in the RTK I PDGFRA cluster

than any other subgroup (11/33; p < 0.001), hence our pro-

posed name for this group. The RTK II Classic cluster demon-

strated a very high frequency of whole chromosome 7 gain

(50/56; p < 0.001) and whole chromosome 10 loss (56/56;

p < 0.001), as well as frequent homozygous deletion of CDKN2A

(35/56; p < 0.001) and amplification of EGFR (39/56; p < 0.001)

(Figures 1 and S1D)all hallmark CNAs of adult GBM

(Louis et al., 2007), as reflected by the complete absence of pedi-

atric patients in this cluster. Overall, tumors from the IDH, K27,

and G34 clusters were mostly devoid of the detected CNAs

associated with the other GBM subgroups (amplifications of

PDGFRA and EGFR, deletion of CDKN2A, chromosome 7 gain,

and chromosome 10 loss) (Figure 1; Table S1), in keeping

with a previously reported finding in G-CIMP+ tumors (Noush-

mehr et al., 2010).

To additionally place the methylation subgroups proposed

here into the context of previous GBM classification systems,

we used the gene expression signature described by the

TCGA to classify 122 of the above tumors with available tran-

scriptome data into one of four gene expression subtypes:

Proneural, Neural, Mesenchymal, and Classical (Verhaak et al.,

2010) (Figure 1; Table S1). This further confirmed the prototypic

nature of tumors in the RTK II Classic cluster, which was

clearly enriched for Classical expression profiles (p < 0.001).

The RTK I PDGFRA cluster was highly enriched for Proneu-

ral expression (p = 0.01), further substantiating the previously

reported association of PDGFRA amplification with this expres-

sion subtype (Verhaak et al., 2010). As expected, all tumors in

the IDH cluster displayed Proneural expression patterns. Inter-

estingly, the K27 cluster also showed a clear enrichment of

tumors with a Proneural signature (p < 0.01), indicating that

Cancer Cell

Epigenetic and Biological Subgroups of Glioblastomathis expression subtype can be divided into subgroups harboring

distinct genomic aberrations based on methylation profiling and

targeted gene sequencing. Mesenchymal gene expression

was mostly restricted to one methylation subgroup (p < 0.001),

that showed a much lower incidence of typical GBM-related

CNAs, generally fewer copy-number changes, and no character-

Cistic point mutations. We therefore termed this methylation

cluster, which displayed the largest similarity with normal brain

methylation patterns, Mesenchymal. Copy-number aberra-

tions in these samples were, however, observed at a similar

amplitude as in other cases, indicating an absence of excess

stromal contamination.

Our finding of six GBM methylation clusters is different from

a TCGA study using Illumina 27k arrays, which identified three

methylation clusters in an adult GBM cohort (Noushmehr et al.,

2010). Applying their signature to our data set, however, showed

that two clusters (G-CIMP+ and Cluster #3) mapped almost

exactly to two of our subgroups (IDH and RTK II Classic,,

respectively, p < 0.001) (Figure 1). By adding pediatric cases to

the study cohort, we demonstrate that TCGAmethylation Cluster

#2 can be further divided into four biologically distinct sub-

groups, defined by a clear enrichment for mutations (K27,

G34), CNAs (PDGFRA), and/or gene expression signatures

(Mesenchymal). The same DNA methylation clusters were

apparent when restricting our analyses to the pediatric popula-

tion, with the exception of the RTK II Classic cluster, which

is not represented in the pediatric population (Figure S1E).

Notably, by analyzing tumors from patients spanning a broad

age spectrum, we further observed a clear age-dependent

increase in overall DNA methylation levels (Figure S2A), even

after adjusting our analysis to exclude tumors with age-related

mutations in IDH1 or H3F3A (Figure S2B).

GBM Subgroups Show Correlations withClinicopathological VariablesThe DNA methylation clusters described here were closely

associated with specific age groups, pointing toward the biolog-

ical diversity of epigenetic GBM subgroups (Figure 1). While the

K27 cluster predominantly consisted of childhood patients

(median age 10.5 years, range 523 years), patients in the G34

cluster were found mostly around the threshold between the

adolescent and adult populations (median age 18 years, range

942 years), as previously suggested (Schwartzentruber et al.,

2012). The RTK I PDGFRA cluster also harbored a proportion

of pediatric patients (median age 36 years, range 874 years), in

line with reports of PDGFRA CNAs being more prevalent in

childhood high-grade gliomas (Bax et al., 2010; Paugh et al.,

2010; Qu et al., 2010). The Mesenchymal cluster displayed a

widespread age distribution (median age 47, range 285 years).

The IDH and RTK II Classic clusters were mostly comprised of

younger adult (median age 40 years, range 1371 years) and

older adult (median age 58, range 3681 years) patients, respec-

tively, reflecting the established differences in patient age

between IDH1-mutated/G-CIMP+ and IDH1 WT adult GBM

(Noushmehr et al., 2010; Yan et al., 2009).

The epigenetic GBM subgroups identified here also showed

mutation-specific patterns of tumor location in the central

nervous system (Figure 3A). While K27-mutated tumors were

predominantly seen in midline locations, e.g., thalamus, pons,and the spinal cord (21/25 cases with available data), tumors

from all other subgroups almost exclusively arose in the cerebral

hemispheres (86/92, p < 0.001). To further investigate this asso-

ciation of mutation type and location, we investigated the tran-

scriptomic profiles of H3F3A-mutated samples (n = 13). Gene

signatures characteristic for K27 and G34 mutant GBMs were

ancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc. 429

AFrontal lobe

Parietal lobeBasal gangliaapplied to a published series of 1,340 transcriptomic profiles

representing multiple regions of the developing and adult human

brain (Kang et al., 2011; Figure S3). The G34 mutant signature

appeared to be most strongly expressed in early embryonic

regions and early- to mid-fetal stages of neocortex and striatum

development. In contrast, the K27 signature most closely

matched with mid- to late-fetal stages of striatum and thalamus

development. Thus, G34 and K27 mutant GBMs seem to show

expression patterns of early developmental stages correlating

with their subsequent tumor location, possibly indicating

different cellular origins and/or time of tumor initiation for these

two subgroups.

B

Pons

Thalamus

IDH (36)

G34 (22)RTK I 'PDGFRA' (15)Mesenchymal (10)

K27 (27)

RTK II 'Classic' (9)

n = 119

2

2

7

11

Surv

ival

pro

babi

liy

Overall survival (months)

censored

K27 (14)

RTK I 'PDGFRA' (23)Mesenchymal (50)RTK II 'Classic' (43)

IDH (10)

G34 (12)

n = 152

0 12 24 36 48 60 72 84 96 108 1200

0.2

0.4

0.6

0.8

1

Figure 3. Epigenetic Subgroups of GBM Correlate with Distinct Clinica(A) Location of 119GBMs in the human central nervous systemgrouped bymethyl

Circles without numbers represent single cases. Different colors indicate methyla

sagittal view (left panel), tumors occurring in the cerebral and cerebellar hemisph

(B and C) Kaplan-Meier survival curves for GBM subgroups defined by methylati

rank tests between subgroups.

See also Figure S3.

430 Cancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc.Frontal lobe Parietal lobe

2 32

62143

35 42

Cancer Cell

Epigenetic and Biological Subgroups of GlioblastomaCorrelating our proposed methylation clusters with patient

survival indicated differences between mutation-defined sub-

groups, but this was somewhat restricted by the low number

of patients with available survival data in each subgroup (Fig-

ure 3B). We therefore enlarged our survival analysis to include

all tumors with known H3F3A and IDH1 mutation status (Fig-

ure 3C). As expected, patients with IDH1 mutant tumors had a

significantly longer overall survival (OS) than patients with

H3F3A and IDH1 WT tumors (p < 0.001) (Noushmehr et al.,

2010; Parsons et al., 2008; Yan et al., 2009). Notably, G34mutant

GBM patients also showed a trend toward a better OS than WT

tumor patients, with marginal statistical significance (p = 0.05). In

Cerebellum

Brainstem

Temporal lobe

Occipital lobe

Occipital lobe

Spinal cord

3

2

2

3

7

3

2

102

2

3

3

5

= 1 Case

= 5 Cases

C

Overall survival (months)

censored

H3F3A K27 (24)IDH1 MUT (47)

H3F3A G34 (13)H3F3A/IDH1 WT (199)n = 283

0 12 24 36 48 60 72 84 96 108 1200

0.2

0.4

0.6

0.8

1

p=0.05 p

A B

Sig

nific

ance

(-Lo

g 10 p

-val

ue)

2

4

6

8

710475

OLIG1 OLIG2

FOXG1

1

2

3

4

5

87

Sig

nific

ance

(-Lo

g 10 p

-val

ue)

FOXG1

Cancer Cell

Epigenetic and Biological Subgroups of Glioblastomacontrast, patients with K27 mutations tended toward an even

shorter OS than patients with WT tumors, although this did not

reach statistical significance (p = 0.12). Comparing the two

H3F3A mutations, patients harboring G34-mutated tumors

clearly had a longer OS than patients with tumors carrying the

K27 mutation (p < 0.01). While this association may be partly

linked to G34-mutated tumors being more accessible to surgery

than the midline K27-mutated tumors, the better prognosis of

G34 versus K27 was independent of location for those cases

where both mutation type and tumor site information were avail-

able (p = 0.02; HR = 0.20, 95% CI = 0.050.77; Cox proportional

hazards model).

Integrating Methylome and Transcriptome DataIdentifies Marker Genes of GBM SubgroupsA combined analysis of DNA methylation and gene expression

data was used to identify subgroup-specific differentially regu-

D E

OLIG2OLIG1

Chromosome 21

IDH

K27

G34

Mes

ench

ymal

RTK

IIR

TK I

1.0

0.5

0

4

6

8

10

12

IDH

K27

G34

OLI

G1

expr

essi

on (L

og2)

OLI

G2

expr

essi

on (L

og2)

*** ***

4

6

8

10

12 *** ***

DNA methylation difference (beta-values)

0

0.6 0.4 0.2 0.0 0.2 0.4 0.6 -10 -5

0

Gene expression diffe

Figure 4. Identification of Marker Genes Affected by Differential Methy

(A andB) Volcanoplots illustratingdifferences inDNAmethylation (A) andgeneexpr

values (A) and Log2 fold change in gene expression values (B) are plotted on the x ax

(C) Starburst plot integrating DNA methylation (x axis) and gene expression (y ax

(D) Methylation levels at the OLIG1 and OLIG2 loci across all 210 GBM samples in

CpG-site. Light blue bars indicate promoter regions. Methylation levels are repre

(E) Mean gene expression levels ofOLIG1 (upper panel) andOLIG2 (lower panel) a

*p < 0.05) between subgroups compared to G34 tumors are indicated.

(F) Inverse correlation of promoter methylation (x axis) and gene expression (y ax

coefficient, r). MES, Mesenchymal.

See also Figure S4 and Table S3.

CC53

254

OLIG1

OLIG2

53

-4

2

0

2

4

1014

39

Gen

e ex

pres

sion

hang

e >=

0 L

og10

(pv

alue

, BH

), fo

ld c

hang

e =0 Log10 (pvalue, BH), Difference

in2,

m

pr

s

(DBA

DC

48%

IDH1 (2/2)

WT (9/9)

7%15%

G34 (6/6)

26%K27 (8/8)

4%

ATRX+ ATRX-

6

14 8

17

26

3 2

52

OLI

G2+

/FO

XG1-

OLI

G2-

/FO

XG1+

OLI

G2-

/FO

XG1-

OLI

G2+

/FO

XG1+

0%

20%

40%

60%

80%

100%

ALT+ ALT-

OLI

G2+

/FO

XG1-

OLI

G2-

/FO

XG1+

OLI

G2-

/FO

XG1-

OLI

G2+

/FO

XG1+

0%

20%

40%

60%

80%

100%

23

4 3

51

9

13 7

18

OLIG2- / FOXG1+

OLIG2+ / FOXG1-

OLIG2- / FOXG1-

OLIG2+ / FOXG1+

IDH1R132H

n=143

Figure 5. Identification of H3F3A-Mutated GBMs by Differential Prote(A) Classification of 143 pediatric GBMs according to protein expression of OLIG

with known H3F3A and IDH1 mutation status as predicted by immunohistoche

(B) Typical pattern of OLIG2/FOXG1+ cells with concomitant loss of ATRXcontrasting staining results for comparison. Scale bars represent 100 mm unles

(C and D) Correlation of GBMs as classified in (A) with ATRX loss (C), and ALT

See also Figure S5 and Table S4.hypomethylated (1653/1946, 85%, Figure S4B) in this subgroup,

in contrast to the hypermethylator G-CIMP pattern observed in

the IDH subgroup (Figure S4C). Hypermethylation and concur-

rent downregulation of TP73 antisense RNA 1 (TP73-AS1) was

identified as a unique characteristic of this IDH/G-CIMP+ cluster

(Figures S4D and S4F). Interestingly, inactivation of this gene by

promoter methylation has been reported as a common mecha-

nism in a high proportion of oligodendrogliomas, 80% of which

are also known to harbor IDH1 mutations (Pang et al., 2010).

Immunohistochemical Analysis Correctly SubclassifiesMutation-Defined GBM SubgroupsIn an attempt to subgroup GBM samples based on differential

protein expressionamethod which is likely to be more suitable

for possible clinical applicationwe used commercially avail-

able antibodies against OLIG2, FOXG1, and mutated IDH1

(R132H) to stain a tissue-microarray (TMA) with cores from 143

pediatric GBMs, and classified tumors according to their protein

expression patterns (Figures 5A and 5B; Table S4). The resulting

fractions of tumors with predicted mutations in IDH1 (IDH1R132H,

n = 6) andH3F3A (OLIG2+/FOXG1 for K27, n = 37, and OLIG2/FOXG1+ for G34, n = 21) were consistent with the frequency of

each mutation in the pediatric population as detected by tar-

geted gene sequencing (Figure S1C). Our approach correctly

classified GBMs with known H3F3A and IDH1 mutation status,

and revealed a frequent association between OLIG2/FOXG1+

tumors (assumed to be G34-mutated), loss of ATRX protein

expression, and an ALT phenotype (Figures 5BD), as previously

reported for H3F3A G34-mutated tumors (Schwartzentruber

et al., 2012). The putative H3F3A mutant groups also did not

overlap with tumors harboring IDH1 (R132H) mutations, and

only one case with EGFR amplification and homozygous

432 Cancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc.OLIG2 FOXG1

ATRX ALT

5 m 10 m

Expression PatternsFOXG1, andmutated IDH1 (IDH1R132H). Numbers in brackets indicate samples

istry and verified by targeted gene sequencing, respectively.

otein expression and ALT as observed in G34-mutated GBMs. Insets show

indicated differently.

).

Cancer Cell

Epigenetic and Biological Subgroups of GlioblastomaCDKN2A deletion was detected therein (Figure S5A). The corre-

lation with clinicopathological variables, such as tumor location

and patient survival, also reflected our findings from the array-

based analysis (Figures S5B and S5C). Of note, rare tumors

represented on the TMA occurring in the basal ganglia and the

spinal cord were almost always found in the OLIG2+/FOXG1

subgroup (and therefore predicted to harbor the H3F3A K27

mutation), further strengthening our hypothesis of the H3F3A

K27 mutation as a unifying characteristic of midline GBM.

DISCUSSION

We have identified six biological subgroups of GBM based on

global DNA methylation patterns, which correlate with specific

molecular-genetic alterations and key clinical parameters. Our

findings suggest that at least 30%40% of pediatric/young adult

GBMs are likely characterized by disrupted epigenetic regula-

tory mechanisms, associated with recurrent and mutually exclu-

sive mutations in either H3F3A or IDH1, and aberrant DNA

methylation patterns. Placing these subgroups into the context

of previous molecular GBM classification schemes described

by the TCGA (Noushmehr et al., 2010; Verhaak et al., 2010)

revealed a clear correlation with DNA methylation clusters and

a corresponding enrichment for previously established expres-

sion signatures in different epigenetic subgroups. We have

also demonstrated that our proposed classification can refine

that described by the TCGA for adult GBM, to give a stratification

system that is applicable across all ages, and defines additional

biologically meaningful subgroups. A simplified graphical sum-

mary of the key molecular and biological characteristics of the

GBM subgroups as identified by our integrated classification

strategy is given in Figure 6.

AP

loGMutations /Cytogenetics

IDH1mut H3F3Amut K27

TP53mut

H3F3

TTP53mut

Epigenetic and BioIDH K27

Cancer Cell

Epigenetic and Biological Subgroups of GlioblastomaWe and others have recently described a high frequency of

H3F3A K27 mutations in thalamic GBMs and in diffuse intrinsic

pontine gliomas (DIPGs), suggesting that the latter likely repre-

sent an anatomically-defined subset of K27 mutant GBM

(Khuong-Quang et al., 2012; Schwartzentruber et al., 2012; Wu

et al., 2012). We now extend this observation to a larger

subgroup of GBM, characterized by the K27M mutation, which

almost exclusively occurs in midline locations, including rare

tumors in the basal ganglia and the spinal cord. This is in line

with a recent study by Puget et al. (2012) in which gene expres-

sion patterns of brainstem gliomas were found to resemble

midline/thalamic tumors, indicating a closely related origin. The

K27 subgroup also displays markedly lower expression of the

ventral telencephalic marker FOXG1 than other subgroups.

Conversely, non-K27 tumors were restricted to hemispheric

AgeDistribution

(years)

TumorLocation

PatientSurvival(months)

DNAMethylation

GeneExpression

IHC ProteinMarker

Proneural Proneural

IDH1R132H

0 12060 0 12060 0 6

CIMP+

-0.6 0.60 -0.6 0

0 9030 60 0 9030 60 0 30

Mix

OLIG2+/FOXG1- OLIG2-/

-0.6 0.60

Figure 6. Graphical Summary of Key Molecular and Biological Charac

Simplified schematic representation of key genetic and epigenetic findings in si

clinical patient data.

Cmut G34

53mut

PDGFRA ampl.

CDKN2A del.CDKN2A del.

7+

10-CNVlow

EGFR ampl.

gical Subgroups of Glioblastoma34 RTK I RTK IIMESENCHYMALlocations, further underlining the biological divergence of epige-

netic GBM subgroups. While recurrent focal amplification of

PDGFRA has been suggested as a key oncogenic event in pedi-

atric DIPGs in some studies (Paugh et al., 2011; Puget et al.,

2012; Zarghooni et al., 2010), midline-associated tumors in the

K27 or OLIG2+/FOXG1 subgroups (including ten brainstemgliomas with known PDGFRA copy-number status) lacked this

common feature in our series. PDGFRA amplification was,

however, enriched in a subgroup of supratentorial hemispheric

GBMs. In part, this discrepancy may be explained by the use

of autopsy (and therefore post radio/chemotherapy) material

in previous study cohorts of DIPGs, which might have been

confounded by the higher incidence of PDGFRA amplifications

observed in radiation-induced gliomas (Paugh et al., 2010).

Nevertheless, amplifications of PDGFRA have also been

Proneural ClassicalMesenchymal

0 120601200 0 12060 0 12060

CHOP+

0.6

9060 0 9030 60 0 9030 60 0 9030 60

ed

FOXG1+ OLIG2+/FOXG1+ OLIG2+/FOXG1+ OLIG2+/FOXG1+

-0.6 0.60 -0.6 0.60 -0.6 0.60

teristics of GBM Subgroups

x GBM subgroups as identified by methylation profiling and correlations with

ancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc. 433

future work at a basic and translational/targeted therapeuticdetected in small numbers of pretreatment samples (Paugh

et al., 2011; Puget et al., 2012; Zarghooni et al., 2010), and

post-treatment samples were not found to show increased

widespread genomic instability (Paugh et al., 2011). This partic-

ularly clinically challenging subset of tumors clearly warrants

further investigation, underlining the importance of routine

stereotactic biopsy of DIPGs at the time of primary diagnosis.

OLIG2 has previously been reported as a universal marker

for diffuse gliomas (Ligon et al., 2004), and OLIG2-positive

progenitor-like cells of the subventricular zone have been sug-

gested as potential glioma-initiating cells (Wang et al., 2009).

There is also evidence that OLIG2-mediated modification of

p53 function is required for complete inactivation of the latter

in malignant gliomas, which typically show indirect loss of

p53 activity through MDM2 amplification or p14ARF deletion

(Mehta et al., 2011). Here, we describe a distinct subgroup of

GBM, harboring the H3F3A G34 mutation, in which OLIG1

and OLIG2 protein expression is absent. Given the 100%mutation frequency of TP53 in this subgroup, this may indicate

a different pathogenesis of G34-mutated GBM, in which

direct inactivation of p53 is required rather than via an OLIG2-

dependent mechanism.

The previously reported association of H3F3A mutations,

particular the G34 mutation, with loss of ATRX and ALT

(Schwartzentruber et al., 2012) is further expanded upon here.

Interestingly, the global CHOP that we observed in G34 mutants

was particularly pronounced in subtelomeric regions, suggesting

a possible mechanistic link with ALT in these tumors (Gonzalo

et al., 2006). Whether this is a more general phenotype that

can be observed in clinically and etiologically distinct subgroups

of other human cancers, remains to be investigated.

The close link between H3F3A mutation type, tumor location,

and differential expression of key neuronal lineagemarkers leads

us to speculate that there may be differences in the cell of origin

and/or the time of tumor development between these GBM

subgroups. Although supported by the differential expression

of mutant-specific gene signatures at different stages of human

brain development, this remains to be formally shown. Also

requiring further validation in larger, prospective cohorts is the

association of the G34 mutation with better overall survival

compared with H3F3A and IDH1 WT tumors, and that of K27

mutation with potentially poorer prognosis, as observed in our

series and a recent cohort of pediatric DIPGs (Khuong-Quang

et al., 2012).

Given the location of the H3F3A mutations at or near critical

regulatory histone residues, and their distinct methylation pro-

files, we consider it likely that the H3.3 mutations are directly

involved in producing widespread aberrant DNA methylation

and deregulation of gene expression. This has recently been

shown for IDH1 mutations, which alone are sufficient to induce

the global epigenetic reprogramming of the G-CIMP phenotype

in normal astrocytes (Turcan et al., 2012). Overproduction of

the oncometabolite 2-hydroxyglutarate in IDH1-mutated cellsinhibits the TET family of 5-methlycytosine hydroxylases and

H3K27-specific demethylases. This is thought to lead to de-

creased 5-hydroxylmethycytosine and increased H3K27methyl-

ation (Xu et al., 2011), resulting in aberrant DNA and histone

methylation, and a block to differentiation (Christensen et al.,

2011; Dang et al., 2009; Lu et al., 2012).

434 Cancer Cell 22, 425437, October 16, 2012 2012 Elsevier Inc.level, particularly in a pediatric and young adult setting.

EXPERIMENTAL PROCEDURES

Patients and Tumor Samples

Primary tumor samples for methylation (n = 136; Table S1), mutation (n = 460;

Table S2), and gene expression (n = 69) analysis and all clinical data were

collected at the DKFZ (Heidelberg, Germany) and at McGill University (Mon-

treal, Canada). Paraffin-embedded samples (n = 143; Table S4) for TMA anal-

ysis were collected from the Burdenko Neurosurgical Institute (Moscow,

Russia) and from the Department of Neuropathology, University of Wurzburg

(Germany). Patient clinical details can be found in Table S1 for the methylation

analysis data set and in Table S4 for the TMA cohort. All of the tumors were

banked at the time of primary diagnosis between 1994 and 2011 in accordance

with research ethics board approval from the respective institutes. Informed

consent was obtained from all patients included in this study. An overview of

all samples included in different data collections is given in Figure S1A. All of

the samples were independently reviewed by senior pediatric neuropatholo-

gists (S.A. and A.K.) according to the WHO guidelines. Detailed information

about samples provided by TCGA can be found elsewhere (http://

cancergenome.nih.gov).

DNA Methylation Profiling

For genome-wide assessment of DNAmethylationGBM samples (n = 136) and

controls (n = 10; four fetal and two