R E S EARCH ART I C L E

CANCER

http://stm.sciencem

ag.oD

ownloaded from

Magnetic resonance image features identifyglioblastoma phenotypic subtypes with distinctmolecular pathway activitiesHaruka Itakura,1 Achal S. Achrol,2 Lex A. Mitchell,3 Joshua J. Loya,2 Tiffany Liu,1

Erick M. Westbroek,4 Abdullah H. Feroze,2 Scott Rodriguez,2 Sebastian Echegaray,5

Tej D. Azad,2 Kristen W. Yeom,3 Sandy Napel,3 Daniel L. Rubin,1,3 Steven D. Chang,2

Griffith R. Harsh IV,2* Olivier Gevaert1*†

Glioblastoma (GBM) is the most common and highly lethal primary malignant brain tumor in adults. There is adire need for easily accessible, noninvasive biomarkers that can delineate underlying molecular activities andpredict response to therapy. To this end, we sought to identify subtypes of GBM, differentiated solely by quan-titative magnetic resonance (MR) imaging features, that could be used for better management of GBM patients.Quantitative image features capturing the shape, texture, and edge sharpness of each lesion were extracted fromMR images of 121 single-institution patients with de novo, solitary, unilateral GBM. Three distinct phenotypic“clusters” emerged in the development cohort using consensus clustering with 10,000 iterations on these imagefeatures. These three clusters—pre-multifocal, spherical, and rim-enhancing, names reflecting their imagefeatures—were validated in an independent cohort consisting of 144 multi-institution patients with similar tumorcharacteristics from The Cancer Genome Atlas (TCGA). Each cluster mapped to a unique set of molecular signalingpathways using pathway activity estimates derived from the analysis of TCGA tumor copy number and gene ex-pression data with the PARADIGM (Pathway Recognition Algorithm Using Data Integration on Genomic Models)algorithm. Distinct pathways, such as c-Kit and FOXA, were enriched in each cluster, indicating differential molecularactivities as determined by the image features. Each cluster also demonstrated differential probabilities of sur-vival, indicating prognostic importance. Our imaging method offers a noninvasive approach to stratify GBMpatients and also provides unique sets of molecular signatures to inform targeted therapy and personalizedtreatment of GBM.

rg

by guest on A

ugust 22, 2020/

INTRODUCTION

Glioblastoma (GBM) is the most common and lethal primary ma-lignant brain tumor in adults. Upon patient presentation with subacuteand progressive neurologic signs and symptoms, gadolinium-enhancedcranial magnetic resonance imaging (MRI) is used as the main diag-nostic modality for brain abnormalities (1). Characteristic hypointensityon T1-weighted images and heterogeneous enhancement after contrastinfusion strongly suggest GBM. MR images demonstrate the extent andlocation of tumor involvement, which can determine the feasibilityof, and approach used in surgical intervention. Although recent clin-ical trials are evaluating advanced MRI techniques to improve the as-sessment of treatment response in GBM (2) or to evaluate changes intumor blood flow after treatment (3) in known GBM cases, MR im-ages are not currently being used to subclassify GBM risk groups. More-over, regardless of the imaging findings, a tissue diagnosis is ultimatelyrequired for definitive histopathologic confirmation and to distinguishfrom other primary and metastatic brain tumors. Factors currently knownto be associated with survival include age and Karnofsky performancestatus (KPS) (4), as well as O6-methylguanine–DNA methyltransferase(MGMT) promoter hypermethylation (5) and mutations in isocitrate

1Division of Biomedical Informatics, Department of Medicine, Stanford University,Stanford, CA 94305, USA. 2Department of Neurosurgery, Stanford University, Stanford, CA94305, USA. 3Department of Radiology, Stanford University, Stanford, CA 94305, USA.4Department of Neurosurgery, Johns Hopkins University, Baltimore, MD 21218, USA.5Department of Electrical Engineering, Stanford University, Stanford, CA 94305, USA.*Co-senior authors.†Corresponding author. E-mail: olivier.gevaert@stanford.edu

www.ScienceTr

dehydrogenase 1 (IDH1) or IDH2 (6, 7). Furthermore, gene expression–based molecular classification of GBM (8), epidermal growth factor re-ceptor (EGFR) amplification (9), and CpG island methylator phenotype(CIMP) status (10) have emerged as potential, additional predictors oftreatment response and outcome. Although such genomic characteriza-tion that encompasses descriptions of gene expression profiles, underlyinggenomic abnormalities, and epigenetic modification has improved theclinical assessment of GBM (8, 10–12), there remains an unmet clinicalneed for easily accessible, surrogate biomarkers able to delineate accu-rately underlying molecular activities and predict response to therapy.

Tumor molecular heterogeneity poses a challenge to the accurateunderstanding of the underlying molecular activities in GBM (13, 14).Substantial intratumoral heterogeneity requires analysis of multiple re-gions of a tumor to capture its full clonal history. Recent advances in im-aging analysis permit three-dimensional (3D) quantitative characterizationof the imaging phenotype of GBM tumors (15–18) that includes this het-erogeneity. The emerging field of imaging genomics involves mappingimage features to molecular data. In pioneering work, investigatorshave linked quantitative computed tomography (CT) image featuresto gene expression data of non–small cell lung cancer to predict survival(19, 20). Similarly, a handful of groups have discovered associations be-tween imaging and gene expression modules in GBM (15), and builtmodels predicting survival by correlating qualitative imaging phenotypeswith gene expression data alone (9) or with the addition of microRNAdata (21).

In this study, we sought to establish image-based biomarkers of GBMsubtypes, ultimately enabling imaging to substitute for intensive molecular

anslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 1

R E S EARCH ART I C L E

analysis. Such an image-based approach would avoid the risks of biopsyand more comprehensively assess intratumoral heterogeneity. Here, weidentify three GBM subtypes differentiated solely by quantitative MRIfeatures and show that these subtypes have prognostic relevance andreflect distinct molecular pathways.

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

RESULTS

Three MRI GBM subtypes existMRI data were obtained in 265 GBM patients, split into two differentcohorts: the development cohort for subtype identification and a valida-tion cohort. The development cohort consisted of 121 patients from theStanford University Medical Center with solitary unilateral tumors ev-ident on MRIs. The validation cohort, which was used to validate find-ings from the development cohort, was composed of 144 subjects withsolitary, unilateral brain lesions from The Cancer Imaging Archive(TCIA). Table 1 summarizes the baseline characteristics of the twocohorts. The selection process did not materially alter the clinicalcharacteristics (age, sex, and KPS) of each cohort from those ofthe original cohorts (tables S1 and S2). For each subject in each co-hort, we applied a quantitative imaging pipeline, extracting gray val-ue histogram statistics, textures, sharpness of lesion boundaries, andmetrics of compactness and roughness as described in Materials andMethods, and generated 388 image features representing both 2Dand multislice 2D (aggregated 2D slices) characteristics of each le-sion (table S3).

On the basis of consensus clustering of the patients’ quantitativeimaging features, we chose the solution with k = 3 as optimal, using max-imal consensus matrices, the consensus cumulative distribution function(CDF) curve, and the CDF progression graph when the Stanford and TheCancer Genome Atlas (TCGA) cohorts were used as development andvalidation cohorts, respectively (Fig. 1, A to C), and vice versa (Fig. 1, D toF). The three-cluster solution produced the largest k that induced thesmallest incremental change in the area under the receiver operatingcharacteristic curve (AUC) (Fig. 1, C and F) while still maximizingthe consensus within clusters (Fig. 1, A and D) and minimizing the rateof ambiguity in cluster assignments across 10,000 iterations (Fig. 1, B andE). This resulted in 36 development cohort patients in cluster 1 (30%), 51patients in cluster 2 (42%), and 34 patients in cluster 3 (28%). When thedevelopment and validation cohorts were swapped, the consensusclustering for k = 3 yielded 25 patients in cluster 1 (18%), 107 patientsin cluster 2 (74%), and 12 patients in cluster 3 (8%).

Next, we used significance analysis of microarrays (SAM) (22) toidentify the quantitative image features significantly associated witheach subtype. In SAM, each feature represented a statistical descriptorof tumor pixel intensities and characteristics, and together, they definedthe multivariate image phenotypes (Fig. 2A and table S4). The top 24 2Dand multislice 2D imaging features are in fig. S1. Cluster 1 was character-ized by quantitative features that described the high irregularity of tumorshapes and many concavities along the tumor outlines. In addition, weobserved that the excluded cases from the development cohort with mul-tifocal lesions (n = 76) appeared to resemble features that characterizedcluster 1. Therefore, we defined the phenotype of cluster 1 as the “pre-multifocal GBM cluster” (Fig. 2A). Cluster 2 was characterized by sphericaltumors with regular edges that were well circumscribed. We defined thephenotype of cluster 2 as the “spherical GBM cluster” (Fig. 2A). Cluster 3was distinguished by prominence of central hypointensity encompassed

www.ScienceTr

by a hyperintense rim. We defined the phenotype of cluster 3 as the“rim-enhancing GBM cluster” (Fig. 2A). The aggregated multislice2D renditions of the three clusters are shown in Fig. 2B.

In a separate analysis, we evaluated fluid-attenuated inversion recov-ery (FLAIR) data that we previously analyzed on 30 subjects (15). Wedetermined that one quantitative FLAIR feature (“histogram kurtosis”),which characterizes image homogeneity, was statistically significantly dif-ferent among the three clusters (Kruskal-Wallis, P = 0.01575).

GBM imaging subtypes are confirmed in a heterogeneousvalidation cohortWe used the in-group proportion (IGP) statistic (23) to validate clusters1 to 3 in an independent, heterogeneous cohort. The IGP is a measure of

Table 1. Clinical and molecular characteristics of the development(Stanford) and validation (TCGA) cohorts. Clinical data for the validationcohort were available on 114 subjects, and molecular data were availa-ble on 107 subjects (missing data are noted). KPS is shown as three catego-ries and a mean value. Tumor location, EGFR amplification, IDH1 mutation,and MGMT promoter hypermethylation are tabulated as number (n) andpercentage.

Characteristic

anslationalMedicine.org 2 September

Stanforddevelopment

cohort(n = 121)

2015 Vol 7 Issue 303

TCGAvalidationcohort

(n = 144)

Age, mean (SD)

64.6 (14.1) 58.5 (15.1)

Sex, n male (%)

69 (57) 73 (64)

KPS, n (%)

>70%

65 (54) 78 (68)

50–70%

44 (37) 19 (17)

<50%

11 (9) 1 (1)

Missing

1 16

Mean ± SD

71.7 ± 17 78.6 ± 12.3

Location, n (%)

Frontal

42 (35)

Parietal

23 (19)

Basal ganglia

3 (3)

Temporal

49 (41)

Occipital

4 (3)

EGFR amplification, n (%)

21 (17) 100 (69)

Missing

74 26

IDH1 mutation, n (%)

4 (4)

Missing

22

MGMT hypermethylation, n (%)

27 (22) 25 (23)

Missing

73 60

Molecular subgroups, n (%)

Proneural

32 (30)

Neural

18 (17)

Mesenchymal

31 (29)

Classical

25 (23)

Missing

1

303ra138 2

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

cluster homogeneity and prediction accuracy and indicates whether acluster in one data set is similar to a cluster in another data set; if the clus-ters are similar, the IGP approaches 100%. In IGP, the Pearson’s cen-tered correlation coefficient between each datum and each centroid iscalculated. The datum is then classified to the group whose centroidhas the highest correlation with the datum. The IGP is defined to be theproportion of data in a group whose nearest neighbors (Pearson’s cen-tered correlation) are also classified to the same group. All three clusterswere statistically significant (P < 0.001, P < 0.001, and P < 0.01 forclusters 1, 2, and 3, respectively), denoting low probabilities that theseclusters originated from a null distribution. The corresponding IGPvalues for clusters 1 to 3 were 91, 80, and 80%, respectively. Only 8of 144 subjects (6 in cluster 1 and 2 in cluster 2) were assigned to acluster in the validation cohort based on the nearest neighbor assign-ment probabilities of less than 0.70.

Next, we used prediction analysis for microarrays (PAM) (24) todefine an image classifier on the three development cohort clusters andassign each sample in the validation cohort to one of the clusters. Thisassigned 41 validation subjects to the pre-multifocal cluster 1 (21%), 63to the spherical cluster 2 (64%), and 40 to the rim-enhancing cluster 3(15%). Repeating these procedures for k = 3 by interchanging the devel-opment and validation cohorts led to poor consensus within clusters(Fig. 1D) and high rates of ambiguity of cluster assignments across

www.ScienceTr

the 10,000 iterations (Fig. 1, D and E). Only cluster 1 could be validatedusing IGP (P < 0.001; IGP, 81%).

Secondary analysis including midline-crossing lesions doesnot change clustersTo address the impact of having excluded midline-crossing lesions, weperformed a secondary analysis that included these lesions. The addi-tion of midline-crossing lesions (n = 19) did not alter the distributionof clusters. Cluster assignment remained grossly unchanged in 116 (97%)of the original 121 subjects in the development cohort (table S5). Ofthe additional 19 subjects, 3 (16%) were assigned to cluster 1, 8 (42%) tocluster 2, and 8 (42%) to cluster 3. The additional samples increasedthe total number of members in cluster 1 to 39 (28% of 140), cluster 2to 59 (41%), and cluster 3 to 43 (31%) from 30, 42, and 29%, respectively.Thus, midline-crossing lesions do not form their own cluster but aresubsumed under the existing three clusters.

GBM imaging subtypes are prognostic of survival,independent of established or putative clinical andmolecular markersWe correlated the imaging-based GBM subtypes with overall sur-vival of patients in the validation cohort treated on the Stupp protocol(25), which is standard first-line therapy consisting of concomitant

Fig. 1. Consensus matrix, CDF curve, and delta curve for all clusters.(A to C) The Stanford cohort as the development cohort and the TCGA

nostic tool used to select the optimal number of clusters in consensus clus-tering. The bottom left of the graph represents sample pairs rarely clustered

cohort as the validation cohort. (D to F) The TCGA cohort as the devel-opment cohort and the Stanford cohort as the validation cohort. (A andD) Consensus matrices represented as heat maps for k = 3 (clusters 1, 2,and 3). Subjects are both rows and columns, and consensus values range from0 (never clustered together, white) to 1 (always clustered together, dark blue).The matrices are ordered by consensus-clustered groups, depicted as adendrogram above the heat map. (B and E) The CDF curve was one diag-

together, whereas the upper right contains those almost always pairedtogether. The middle segment represents sample pairs with ambiguous as-signments across different clustering runs. The goal was to identify the lowestrate of ambiguous assignments (flat middle segment). (C and F) The deltacurve depicts the CDF progression graph, plotting the relative change in areaunder the CDF curve, comparing k with k + 1. The goal was to select thelargest k that induced the smallest incremental change in the AUC.

anslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 3

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

radiotherapy with temozolomide, followed by adjuvant temozolomidealone. We observed significant differences in survival probabilities forthe three subgroups in the TGCA cohort (P = 0.004, log-rank test) (Fig.2C). Pre-multifocal cluster 1 had the poorest survival rate; sphericalcluster 2 had an intermediate rate; and rim-enhancing cluster 3 hadthe best survival rate.

We further examined the correlation of these image-based clusters withestablished or putative clinical and molecular risk factors of survival in thedevelopment cohort (Table 2A). There were no significant differencesamong the three clusters for age, sex, KPS, MGMT hypermethylation,or EGFR amplification. However, tumor volume was significantly dif-ferent among the clusters; the smallest tumors were found in the sphericalcluster 2, whereas the largest tumors were in the rim-enhancing cluster 3(P < 0.001, Kruskal-Wallis test) (Fig. 3). Despite this association, tumorvolume was not a sufficient independent predictor of the three clusters.The correlation coefficient between the image-based subtypes andtumor volumes was 0.317 (95% CI, 0.146 to 0.469). Moreover, in

www.ScienceTranslationalMedicine.org 2 Sep

regression models, tumor volume was nota statistically significant predictor of image-based subtypes (P = 0.923, P = 0.0777, P =0.234 for clusters 1, 2, and 3, respectively).Using tumor volume alone to predictimage-based subtypes led to an accuracyof 43.8%. Also, there was not a significantcorrelation with tumor location (basal gan-glia, frontal, occipital, parietal, or temporal)(P = 0.052, Fisher’s exact test).

We conducted a similar analysis for thevalidation cohort using the molecular datafrom the TCGA (Table 2B). We observedno significant differences among the threeclusters for age, sex, KPS, or other knownor putative molecular factors that have beenassociated with survival, including MGMThypermethylation, EGFR amplification, IDHmutation status, and CIMP.

GBM subtypes map to canonicalmolecular pathwaysTo estimate various signaling pathway ac-tivities as they pertain to the three clusters,we used the Pathway Recognition Algo-rithm Using Data Integration on GenomicModels (PARADIGM) (26) to integrategene expression and copy number varia-tion (CNV) data of patients in the TCGAvalidation cohort. All three clusters weresignificantly associated with one or moreregulatory pathways [all false discovery rates(FDRs) < 5%] (Fig. 2D, Table 3, and tableS6). Notably, the pre-multifocal cluster1 was marked by only one association: theup-regulation of the c-Kit stem cell factorreceptor pathway. The spherical cluster 2was characterized by the down-regulationof 21 pathways, including c-Kit, VEGFRsignaling, PDGFR-a signaling, FOXA tran-scriptional networks, and Ang/Tie2. Last,

the rim-enhancing cluster 3 was differentiated by the up-regulation of31 pathways, including canonical WNT and PDGFR-b signaling andmany of the pathways down-regulated in cluster 2, such as VEGFRsignaling, FOXA, and Ang/Tie2 (Table 3 and table S6).

DISCUSSION

We have identified three distinct clusters of unilateral, solitary GBMdefined by quantitative MR image features. The clusters of GBM sub-types were discovered in one cohort and validated in a second cohort.In addition to distinguishing imaging phenotypes, or “clusters”—whichwe named pre-multifocal, spherical, and rim-enhancing—the threeclusters demonstrated significant differences in survival probabilitiesand associations with canonical signaling pathways. Imaging-basedmarkers of disease phenotype may therefore offer actionable knowledgefor clinical decision making and therapeutic targeting.

Fig. 2. GBM subtypes cluster by phenotypic MRI characteristics, correlate with survival, and associatewith molecular pathways. Three distinct image-based subtypes were derived from a development cohort

(Stanford cohort) and validated in an independent validation cohort (TCGA cohort). (A) Imaging phenotypesare illustrated as simplified, representative pictograms for each cluster, although the multivariate combina-tion of quantitative images features that characterize each cluster (table S4) cannot be fully visually exem-plified. (B) Aggregate multislice 2D renditions of the three imaging subtypes (clusters). (C) Kaplan-Meiersurvival curves (solid lines) with 95% confidence intervals (CIs) (dotted lines) derived from TCGA survival dataare shown for each cluster in the TCGA validation cohort. Survival differences across the clusters: P = 0.004,log-rank test (n = 37 subjects across the clusters who underwent the same Stupp protocol treatment regi-men: cluster 1, n = 6; cluster 2, n = 22; cluster 3, n = 9). (D) Molecular changes associated with each cluster.Arrows indicate up- or down-regulation of sample pathways identified using PARADIGM. Table S6 provides acomprehensive list of significant regulatory pathways associated with each cluster at FDR <5%.

tember 2015 Vol 7 Issue 303 303ra138 4

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

The current clinical standard for disease relapse surveillance afterchemoradiation is tracking changes clinically (symptomatic and neu-rologic) and with MRI (27–29). The initial treatment for newly diag-

www.ScienceTr

nosed GBM is maximal surgical resection while preserving neurologicfunction. However, a major challenge in the management of GBM isthe limited number of nonsurgical therapeutic options, such as targeted

Table 2. Clinical and molecular characteristics of the three imaging sub-types in both cohorts. (A) The development (Stanford) cohort. (B) The valida-tion (TCGA) cohort. For continuous values, data are means (SD). P values are

derived from comparisons across the three clusters using the Kruskal-Wallis test(*) for continuous variables or the Fisher’s exact test (†) for categorical variables.Where noted as missing, molecular features were not available for analysis.

Characteristic

Cluster 1, pre-multifocal Cluster 2, spherical

anslationalMedicine.org

Cluster 3, rim-enhancing

2 September 2015 Vol 7 Issue 30

P value

(A) Development cohort

Age, mean (SD)

64.7 (17.7) 66.0 (11.7) 62.0 (13.0) 0.324*

Sex, n (%)

Female

14 24 14

Male

22 27 20 0.736†

Mean KPS (SD)

67.7 (19.2) 73.7 (13.9) 72.9 (18.1) 0.439*

Tumor volume, cm3 (SD)

31.2 (30.1) 22.6 (18.9) 55.9 (32.7) 4.49 × 10−7*

MGMT hypermethylation, n

8 12 7 0.747†

Missing

25 29 19

EGFR amplification, n

5 8 8 0.702†

Missing

24 33 17

Location, n

0.0532†

Basal ganglia

0 2 1

Frontal

9 20 13

Occipital

4 0 0

Parietal

7 6 10

Temporal

16 23 11

(B) Validation cohort

Age, mean (SD)

58.6 (14.5) 58.9 (16.5) 57.6 (13.4) 0.506*

Sex, n (%)

Female

15 (37) 15 (24) 11 (28)

Male

15 (37) 37 (59) 21 (52) 0.171†

Missing

11 11 8

Mean KPS ± SD

78.3 ± 10.3 80 ± 12.6 76.6 ± 13.2 0.488*

Missing

7 6 3

Molecular subgroups, n (%)

Classical

4 (10) 13 (21) 8 (20) 0.075†

Mesenchymal

7 (17) 19 (30) 5 (13)

Neural

7 (17) 3 (5) 8 (20)

Proneural

7 (17) 12 (19) 7 (18)

Missing

16 16 12

CIMP, n (%)

4 (10) 1 (2) 1 (3) 0.121†

Missing

1 2 2

IDH1 mutation, n (%)

3 (10) 1 (2) 0 (0) 0.216†

Missing

2 11 9

MGMT hypermethylation, n (%)

6 (21) 13 (27) 6 (21) 0.496†

Missing

17 28 15

EGFR amplification, n (%)

27 (66) 43 (68) 30 (75) 0.662†

Missing

10 15 8

3 303ra138 5

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

therapies against molecular derangements. Our results show the po-tential of imaging features to infer up-regulated molecular activitiesfor which targeted therapies exist. If the effectiveness of such therapiescan be demonstrated, there may be a role of such agents as adjuvanttherapy in the postoperative setting, primary treatment in surgicallyunresectable cases, and neoadjuvant treatment for the preoperative re-duction of tumor volumes.

On the basis of unique associations with canonical signaling pathways,the imaging subtypes have the potential to direct targeted therapy. Spe-cifically, the pre-multifocal cluster 1 was associated with the up-regulationof the c-Kit pathway, whereas the rim-enhancing cluster 3 was character-ized by the up-regulation of several pathways, including canonical Wnt,VEGFR, PDGFR, and Ang/Tie2. This suggests the potential use for pa-tients in cluster 1 of targeting c-Kit with tyrosine kinase inhibitors, suchas imatinib and dasatinib. Similarly, although it has long been describedthat GBM overexpresses VEGF-A and is thus amenable to treatmentwith a VEGF inhibitor (30–37), there is now a strong rationale for the useof bevacizumab—a VEGF inhibitor—or antiangiogenic multikinase in-hibitors that target both VEGFR and PDGFR, such as sorafenib andsunitinib, in patients in cluster 3. Clinical trials are currently investigatinginhibitors of the Ang/Tie2 pathway, potentially expanding therapeutic op-tions for members of cluster 3 (38). These results provide intriguing datafor retrospective validation and eventually for designing prospective clin-ical trials to confirm these hypotheses.

Conversely, absence of correlation between a cluster and a signal-ing pathway suggests that the tumors in that cluster lack the potentialtherapeutic targets of the pathway. For instance, the treatment optionsdescribed for clusters 1 and 3 would likely be ineffective in cluster 2.Such knowledge makes clinical trials more efficient and spares pa-tients the time and risk of undergoing trials of limited likelihood ofsuccess. Recently, Chinot et al. reported the findings of a randomized,placebo-controlled, double-blind, phase 3 trial of bevacizumab as afirst-line therapy for patients with newly diagnosed GBM in concertwith standard chemoradiotherapy (39) and identified no overall survivaldifference in response to bevacizumab (40). Because many new thera-

www.ScienceTr

peutic approaches are often applied to all patients in a diagnosticcategory without individualization or with suboptimal individualization,potential beneficial effects may be masked. We hypothesize that the rim-enhancing cluster (cluster 3), which up-regulates the VEGFR signal-ing pathway, would be more likely to respond to bevacizumab.

The proposed imaging-based subtypes also stratified survival inGBMpatients treated on the Stupp protocol, which entails concurrent chemo-radiotherapy, followed by adjuvant chemotherapy.Moreover, the imag-ing subtypes conferred survival differences independent of establishedrisk factors, thus offering additional independent prognostic information,similar to identifying significant factors while controlling for establishedrisk factors. The quantitativeMR image features used to define the threeclusters thus constitute potential biomarkers for survival in GBM.

1 2 3

020

40

60

80

100

120

140

Tumor volume by image−based cluster

Image−based cluster

Tum

or

volu

me, cm

3

Kruskal−Wallis

P < 0.001

Fig. 3. Tumor volumes by cluster for the development cohort (Stanfordcohort). The largest tumors were in cluster 3, and the smallest tumors were

in cluster 2 (P < 0.001, Kruskal-Wallis test). An overlap in tumor volume isobserved between clusters 1 and 2.

Table 3. Selected pathways associated with each image-based cluster.From the complete set of pathways significantly associated with eachimage-based cluster (table S6) are selected pathways that are either spe-cifically differentially expressed among the three clusters or among the top10 for the cluster by fold change.

Signaling pathway

anslationalMedicine.org 2 September 2015 Vol 7 Issue

Foldchange

303 303ra

q value(%)

Cluster 1

Signaling events mediated by stem cellfactor receptor (c-Kit)

1.033 0

Cluster 2

Signaling events mediated by c-Kit

0.976 0

Signaling events mediated by prolactin (PRL)

0.997 4.5

Platelet-derived growth factor receptor–a(PDGFR-a) signaling pathway

0.997

0

Forkhead box protein A2 (FOXA2) and FOXA3transcription factor networks

0.991

0

Vascular endothelial growth factor receptor 1(VEGFR1)–specific signals

0.996

4.5

Angiopoietin (Ang) receptor Tie2-mediated signaling

0.988 0

Regulation of nuclear mothers againstdecapentaplegic homolog 2/3 (SMAD2/3) signaling

0.995

4.5

Signaling events mediated by protein tyrosinephosphatase 1B (PTP1B)

0.995

4.5

Osteopontin-mediated events

0.995 4.5

Signaling events activated by hepatocyte growthfactor receptor (c-Met)

0.994 4.5

Cluster 3

FOXA2 and FOXA3 transcription factor networks

1.013 0

PDGFR-b signaling pathway

1.013 0

Canonical Wnt signaling pathway

1.003 3.6197

VEGFR1-specific signals

1.004 0

Ang receptor Tie2-mediated signaling

1.009 3.6197

Syndecan-1–mediated signaling events

1.011 0

Interleukin-6 (IL-6)–mediated signaling events

1.010 3.6197

Osteopontin-mediated events

1.010 0

Fc-e receptor I signaling in mast cells

1.009 0

aMb2 integrin signaling

1.009 0

138 6

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

Our image-based clusters have the potential to be used for non-invasive treatment follow-up. The cluster phenotypes could be usedas surrogate markers of underlying molecular activities for diseasemonitoring and lead to new pathophysiologic understanding. In addi-tion, the multivariate pathway activity profile can potentially be linkedto a drug response profile. The association of distinct molecular activ-ities with specific image-based GBM clusters suggests that image fea-tures alone may be used to track changes in disease activity, includingproviding insight into the natural progression of disease. Such under-standing currently requires invasive biopsies and intensive molecularassays. The main technical challenge is integrating our multivariateimage profile into the clinical workflow to assign patients to the three sub-types. Ideally, our methodology and image profile would be incorporatedinto the radiologists’ toolbox. This would require manual image seg-mentation, which is currently labor-intensive.

Our study extended the earlier efforts in imaging genomics analysisof cancers. Previous studies have linked quantitative CT image featuresto gene expression data of lung cancer (19, 20) and MR image featureswith genomic data in GBM (15, 17, 21). Whereas previous work focusedon genomic analysis before linking with an image phenotype, our studyinverted this design by beginning with a characterization of imagingphenotypes of GBM tumors, followed by linking to signaling pathways.We believe that subtypes defined in this manner are more stable andless susceptible to sampling error resulting from assessing a limited por-tion of a molecularly heterogeneous tumor. Thus, our approach attemptedto minimize the bias inherent in molecular sampling of a heterogeneoustumor and maximize the information captured from the entire 3D im-aging phenotype of each tumor.

Here, we meticulously selected homogeneous study populations.We selected only those subjects with solitary, unilateral lesions and ex-cluded patients with multifocal or midline-crossing lesions. We elimi-nated patients with the most advanced tumor growth and imagingcharacteristics overtly indicative of poor prognosis, such as lesions cross-ing the midline or multifocal lesions. Extensive tumor growth was alsoassumed to signify a confluence of molecular activities with less amena-bility to therapy. Instead, we focused on the subset of patients with thegreatest clinical need for imaging biomarkers that could differentiate prog-nostic groups; that is, those with earlier stages of disease and, thus, for whominterventions may have the greatest impact on the outcome and personal-ization of treatmentmight bemost beneficial. Moreover, the homogeneity ofstudy subjects was designed to maximize signal detection. However, wemeasured the impact of having excluded midline-crossing lesions byperforming a secondary analysis that included these lesions. Subjectcluster assignments remained largely unchanged, with a 97% reten-tion of the original assignment (table S5), suggesting that quantitativeimage features for midline-crossing lesions do not differ materiallyenough to generate their own separate cluster. Conversely, the inclusionof multifocal lesions remains infeasible given that their analysis wouldnecessitate biopsies at each tumor focus to overcome intertumoral mo-lecular heterogeneity, and multifocal biopsies and molecular featuresare not routinely acquired or available through TCGA or in local datarepositories.

Next, we used MR images from a heterogeneous validation cohort(TCGA) that were acquired on multiple scanner models from three dif-ferent manufacturers at four different institutions and successfully vali-dated the imaging subtypes in a real-world setting. Swapping thedevelopment and validation cohorts did not result in the identificationof robust cluster generation (fig. S1), indicating that a homogeneous

www.ScienceTr

cohort is necessary for learning subtypes but not for assigning subjectsto imaging subtypes. Nevertheless, the successful validation of theclusters in a more heterogeneous environment highlights the robust-ness of the model.

We also focused only on T1 MRI, the most widely used MR mo-dality for GBM, to allow the broad applicability and use of the model inthe clinical setting. We identified one quantitative FLAIR feature thatcharacterizes image homogeneity as significantly different among thethree clusters. Larger studies will become possible as more FLAIR databecome available in the future. Nevertheless, it is possible that quanti-tative features derived from other MRI modalities may further enhancethe definition of our clusters or elucidate new associations with signalingpathways. The limited availability of multimodal MR data precluded acomprehensive analysis of imaging features at this time.

In summary, the three GBM subtypes discovered here were distin-guished by different expressions of molecular pathways and survivalprobabilities, which open possibilities for target identification and ther-apeutic development unique to each subtype. Quantitative imagingfeatures may therefore serve as potential biomarkers for defining subtypesof GBM and potentially other cancers.

MATERIALS AND METHODS

Study designWe sought to identify novel subtypes of GBM, differentiated solely byquantitative MRI features, and discover signature canonical signalingpathways that are molecularly associated with each subtype. To do this,we separately analyzed two cohorts of subjects with GBM (Table 1).This study analyzed data that had already been collected and did notinvolve new patients randomized to different interventions or treat-ments. Thus, there was no blinding. Also, we analyzed as many subjectsas possible, applying stringent inclusion and exclusion criteria to homog-enize our study cohorts as was most scientifically reasonable. Because thiswas not a prospective study, there were no a priori power calculationsperformed.

In selected patient images, quantitative image features capturingthe shape, texture, and edge sharpness of each lesion for each subject wereextracted from T1 post-contrast MRI. We used consensus clusteringto discover subtypes in the development cohort and performed 10,000iterations to achieve high robustness and to minimize cluster selectionbias resulting from a few iterations. Regions of interest (ROIs) on theMRI study that were too small (<10 pixels) to interpret were excluded.PAM and the IGP statistic were used to validate the reproducibilityand robustness of our subtype classification in a second cohort of sub-jects (validation cohort). To map imaging subtypes to particular mo-lecular signaling pathways, we performed SAM on pathway activityestimates derived from the analysis of TCGA tumor CNV and geneexpression data with the PARADIGM algorithm. We used SAM tocorrect for multiple-hypothesis testing. We mitigated against biases inour analyses by the performance of numerous iterations of consensusclustering to ensure cluster stability, the use of an external validation co-hort to validate our cluster findings, and correction for multiple hypoth-esis testing in SAM. Imaging subtypes were then correlated with survivaland molecular characteristics, as well as traditional risk factors of GBM.

For survival analysis, we focused on the TCGA cohort patientstreated according to the Stupp protocol. Unlike the TCGA cohort,the Stanford cohort did not have a well-characterized subgroup of

anslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 7

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

patients who had undergone a homogeneous treatment plan (Stuppprotocol), making the comparison of survival curves less interpretable.Because the Stanford cohort had undergone heterogeneous treatmentsdue to diagnosis over a long period (2001–2010), meaningful survivalcurves could not be depicted.

Patient selectionThe study was approved by Stanford’s Institutional Review Board. Forthe first cohort, we identified 364 single-institution subjects diagnosedwith and treated for de novo GBM at the Stanford Medical Centerbetween 2001 and 2010 and selected 121 that met the inclusion crite-ria below; this was defined as the development cohort. The second wasa multi-institutional cohort from the TCGA project, a federally fundedcancer-specific repository of clinical, molecular, and imaging data, inwhich we selected 144 subjects of 575 that met the inclusion criteria;this was defined as the validation cohort. Selection criteria for both co-horts included (i) diagnosis of de novo GBM, (ii) minimum age of18 years, (iii) availability of preoperative MRI data, and (iv) presenceof solitary, unilateral tumors. For the development cohort alone, wefurther selected patients on the basis of similarity of their MR imageswith respect to machine model and slice thickness.

Quantitative imaging pipelineMR images for the development cohort were obtained from the Stan-ford University Medical Center, whereas images for the validation cohortwere obtained from TCIA, TCGA’s companion repository of imagingdata for cases donated to the TCGA (www.cancerimagingarchive.net).The images from both cohorts underwent the same preprocessing pro-cedures. Two blinded, board-certified neuroradiologists reached consen-sus regarding a 3D tumor volume by delineating ROIs around areas ofenhancement in each T1 post-contrast MR slice. We computed a set ofquantitative image features, as previously reported in lung cancer andGBM (15, 19), and applied the imaging feature extraction pipeline tothe development and validation cohorts. Briefly, after linear normalizationof image pixel intensities, our quantitative image feature pipeline extractedseveral features, including morphological characteristics, such as tumorshape, edge sharpness, and pixel intensity statistics on areas of enhance-ment from these ROIs. Two-dimensional and multislice 2D feature repre-sentations are described in Supplementary Methods.

Subtype discoveryWe used k-means consensus clustering for unsupervised class discov-ery on the development cohort to define imaging subtypes (41, 42). Weperformed 10,000 bootstraps with 80% item resampling on the quanti-tative image features and used the k-means clustering algorithm withthe Euclidean distance metric to examine all resulting clusters from 2to 10. We selected the number of clusters that yielded the most stableconsensus matrices and the most unambiguous cluster assignmentsacross permuted clustering runs. This established the optimal numberof intrinsic unsupervised classes as defined by image features in thedevelopment cohort. We analyzed the discovered clusters using SAM,where its application on each discovered cluster identified specific imagefeatures defining each subtype. Significant associations were correctedfor multiple testing using an FDR threshold of ≤15%.

Validation of clustersTo validate the reproducibility of the clusters derived from consensusclustering in the development cohort, we used the IGP analysis to

www.ScienceTr

demonstrate the existence of these clusters in the validation cohort,as described in Supplementary Methods.

Mapping canonical pathways to imaging subtypesWe used the matched molecular data in the TCGA validation cohortto assign canonical signaling pathways to the discovered imaging sub-types, as described in Supplementary Methods.

Evaluating imaging subtypes for traditional risk factorsWe examined the subtypes for statistically significant differences in his-torically associated risk factors, as described in Supplementary Methods.

Statistical analysisTo ensure cluster stability in our unsupervised analysis, we performedconsensus clustering with 10,000 iterations and used CDF and CDF pro-gression graphs to select the optimal number of clusters. Statistical anal-ysis for determining the validity and reproducibility of the three clustersidentified in the development cohort entailed the IGP statistic in the val-idation cohort (Supplementary Methods). When comparing across-clusterclinical andmolecular feature differences in each cohort, we performed theKruskal-Wallis test for continuous variables and the Fisher’s exact testfor categorical variables (Tables 1 and 2). Parametric assumptionswere not made for continuous variables. [The overall sample size andcounts as low as 0 in some cells necessitated the use of the Fisher’s exacttest rather than the Pearson c2 test (43, 44)]. For evaluation of statisti-cally significant imaging and signaling pathway features associated witheach cluster, we used SAM, which adjusts for multiple comparisons(tables S4 and S6). We reported results for FDR <15% for imaging fea-tures and FDR <5% for signaling pathways. The Kaplan-Meier survivalcurves were compared using the log-rank test.

SUPPLEMENTARY MATERIALS

www.sciencetranslationalmedicine.org/cgi/content/full/7/303/303ra138/DC1MethodsFig. S1. Relative contributions of the top 24 imaging features in characterizing each of the clusters.Table S1. Clinical characteristics of the Stanford cohort before and after selection of the de-velopment cohort.Table S2. Clinical characteristics of the TCGA cohort before and after selection of the validationcohort.Table S3. All 2D and multislice 2D quantitative MR image features used for analysis.Table S4. Two-dimensional and multislice 2D quantitative MR image features significantly as-sociated with each cluster.Table S5. Cluster assignment by subjects in the development cohort with and without midline-crossing lesions.Table S6. Regulatory signaling pathways significantly associated with each cluster.References (45–47)

REFERENCES AND NOTES

1. S. M. Chang, I. F. Parney, W. Huang, F. A. Anderson Jr., A. L. Asher, M. Bernstein, K. O. Lillehei,H. Brem, M. S. Berger, E. R. Laws; Glioma Outcomes Project Investigators, Patterns of care foradults with newly diagnosed malignant glioma. JAMA 293, 557–564 (2005).

2. Vanderbilt-Ingram Cancer Center, “3T MRI Biomarkers of Glioma Treatment Response,”ClinicalTrials.gov (Vanderbilt-Ingram Cancer Center, National Cancer Institute, 2013).

3. Massachusetts General Hospital, “Bevacizumab with temozolomide PET and vascular MRIfor GBM,” ClinicalTrials.gov (Massachusetts General Hospital, 2013).

4. E. R. Laws, I. F. Parney, W. Huang, F. Anderson, A. M. Morris, A. Asher, K. O. Lillehei, M. Bernstein,H. Brem, A. Sloan, M. S. Berger, S. Chang; Glioma Outcomes Investigators, Survival followingsurgery and prognostic factors for recently diagnosed malignant glioma: Data from theGlioma Outcomes Project. J. Neurosurg. 99, 467–473 (2003).

anslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 8

R E S EARCH ART I C L E

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from

5. M. E. Hegi, L. Liu, J. G. Herman, R. Stupp, W. Wick, M. Weller, M. P. Mehta, M. R. Gilbert, Cor-relation of O6-methylguanine methyltransferase (MGMT) promoter methylation with clinicaloutcomes in glioblastoma and clinical strategies to modulate MGMT activity. J. Clin. Oncol. 26,4189–4199 (2008).

6. H. Yan, D. W. Parsons, G. Jin, R. McLendon, B. A. Rasheed, W. Yuan, I. Kos, I. Batinic-Haberle,S. Jones, G. J. Riggins, H. Friedman, A. Friedman, D. Reardon, J. Herndon, K. W. Kinzler,V. E. Velculescu, B. Vogelstein, D. D. Bigner, IDH1 and IDH2 mutations in gliomas. N. Engl.J. Med. 360, 765–773 (2009).

7. M. Sanson, Y. Marie, S. Paris, A. Idbaih, J. Laffaire, F. Ducray, S. El Hallani, B. Boisselier, K. Mokhtari,K. Hoang-Xuan, J.-Y. Delattre, Isocitrate dehydrogenase 1 codon 132 mutation is an importantprognostic biomarker in gliomas. J. Clin. Oncol. 27, 4150–4154 (2009).

8. R. G. W. Verhaak, K. A. Hoadley, E. Purdom, V. Wang, Y. Qi, M. D. Wilkerson, C. R. Miller, L. Ding,T. Golub, J. P. Mesirov, G. Alexe, M. Lawrence, M. O’Kelly, P. Tamayo, B. A. Weir, S. Gabrie,W. Winckler, S. Gupta, L. Jakkula, H. S. Feiler, J. G. Hodgson, C. D. James, J. N. Sarkaria,C. Brennan, A. Kahn, P. T. Spellman, R. K. Wilson, T. P. Speed, J. W. Gray, M. Meyerson, G. Getz,C. M. Perou, D. N. Hayes; Cancer Genome Atlas Research Network, Integrated genomic analysisidentifies clinically relevant subtypes of glioblastoma characterized by abnormalities inPDGFRA, IDH1, EGFR, and NF1. Cancer Cell 17, 98–110 (2010).

9. D. A. Haas-Kogan, M. D. Prados, T. Tihan, D. A. Eberhard, N. Jelluma, N. D. Arvold, R. Baumber,K. R. Lamborn, A. Kapadia, M. Malec, M. S. Berger, D. Stokoe, Epidermal growthfactor receptor,protein kinase B/Akt, and glioma response to erlotinib. J. Natl. Cancer Inst. 97, 880–887 (2005).

10. H. Noushmehr, D. J. Weisenberger, K. Diefes, H. S. Phillips, K. Pujara, B. P. Berman, F. Pan,C. E. Pelloski, E. P. Sulman, K. P. Bhat, R. G. W. Verhaak, K. A. Hoadley, D. N. Hayes, C. M. Perou,H. K. Schmidt, L. Ding, R. K. Wilson, D. Van Den Berg, H. Shen, H. Bengtsson, P. Neuvial,L. M. Cope, J. Buckley, J. G. Herman, S. B. Baylin, P. W. Laird, K. Aldape; Cancer Genome AtlasResearch Network, Identification of a CpG island methylator phenotype that defines a distinctsubgroup of glioma. Cancer Cell 17, 510–522 (2010).

11. M. Bredel, D. M. Scholtens, G. R. Harsh, C. Bredel, J. P. Chandler, J. J. Renfrow, A. K. Yadav, H. Vogel,A. C. Scheck, R. Tibshirani, B. I. Sikic, A network model of a cooperative genetic landscape inbrain tumors. JAMA 302, 261–275 (2009).

12. D. Sturm, H. Witt, V. Hovestadt, D.-A. Khuong-Quang, D. T. W. Jones, C. Konermann, E. Pfaff,M. Tönjes, M. Sill, S. Bender, M. Kool, M. Zapatka, N. Becker, M. Zucknick, T. Hielscher, X.-Y. Liu,A. M. Fontebasso, M. Ryzhova, S. Albrecht, K. Jacob, M. Wolter, M. Ebinger, M. U. Schuhmann,T. van Meter, M. C. Frühwald, H. Hauch, A. Pekrun, B. Radlwimmer, T. Niehues, G. von Komorowski,M. Dürken, A. E. Kulozik, J. Madden, A. Donson, N. K. Foreman, R. Drissi, M. Fouladi, W. Scheurlen,A. von Deimling, C. Monoranu, W. Roggendorf, C. Herold-Mende, A. Unterberg, C. M. Kramm,J. Felsberg, C. Hartmann, B. Wiestler, W. Wick, T. Milde, O. Witt, A. M. Lindroth, J. Schwartzentruber,D. Faury, A. Fleming, M. Zakrzewska, P. P. Liberski, K. Zakrzewski, P. Hauser, M. Garami, A. Klekner,L. Bognar, S. Morrissy, F. Cavalli, M. D. Taylor, P. van Sluis, J. Koster, R. Versteeg, R. Volckmann,T. Mikkelsen, K. Aldape, G. Reifenberger, V. P. Collins, J. Majewski, A. Korshunov, P. Lichter, C. Plass,N. Jabado, S. M. Pfister, Hotspot mutations in H3F3A and IDH1 define distinct epigenetic andbiological subgroups of glioblastoma. Cancer Cell 22, 425–437 (2012).

13. A. Sottoriva, I. Spiteri, S. G. M. Piccirillo, A. Touloumis, V. P. Collins, J. C. Marioni, C. Curtis, C. Watts,S. Tavaré, Intratumor heterogeneity in human glioblastoma reflects cancer evolutionary dynam-ics. Proc. Natl. Acad. Sci. U.S.A. 110, 4009–4014 (2013).

14. A. P. Patel, I. Tirosh, J. J. Trombetta, A. K. Shalek, S. M. Gillespie, H. Wakimoto, D. P. Cahill,B. V. Nahed, W. T. Curry, R. L. Martuza, D. N. Louis, O. Rozenblatt-Rosen, M. L. Suvà, A. Regev,B. E. Bernstein, Single-cell RNA-seq highlights intratumoral heterogeneity in primary glio-blastoma. Science 344, 1396–1401 (2014).

15. O. Gevaert, L. A. Mitchell, A. S. Achrol, J. Xu, S. Echegaray, G. K. Steinberg, S. H. Cheshier, S. Napel,G. Zaharchuk, S. K. Plevritis, Glioblastoma multiforme: Exploratory radiogenomic analysis byusing quantitative image features. Radiology 276, 313 (2015).

16. D. A. Gutman, L. A. D. Cooper, S. N. Hwang, C. A. Holder, J. Gao, T. D. Aurora, W. D. Dunn Jr.,L. Scarpace, T. Mikkelsen, R. Jain, M. Wintermark, M. Jilwan, P. Raghavan, E. Huang, R. J. Clifford,P. Mongkolwat, V. Kleper, J. Freymann, J. Kirby, P. O. Zinn, C. S. Moreno, C. Jaffe, R. Colen,D. L. Rubin, J. Saltz, A. Flanders, D. J. Brat, MR imaging predictors of molecular profile and survival:Multi-institutional study of the TCGA glioblastoma data set. Radiology 267, 560–569 (2013).

17. M. Diehn, C. Nardini, D. S. Wang, S. McGovern, M. Jayaraman, Y. Liang, K. Aldape, S. Cha,M. D. Kuo, Identification of noninvasive imaging surrogates for brain tumor gene-expressionmodules. Proc. Natl. Acad. Sci. U.S.A. 105, 5213–5218 (2008).

18. B. M. Ellingson, A. Lai, R. J. Harris, J. M. Selfridge, W. H. Yong, K. Das, W. B. Pope, P. L. Nghiemphu,H. V. Vinters, L. M. Liau, P. S. Mischel, T. F. Cloughesy, Probabilistic radiographic atlas of glio-blastoma phenotypes. AJNR Am. J. Neuroradiol. 34, 533–540 (2013).

19. O. Gevaert, J. Xu, C. D. Hoang, A. N. Leung, Y. Xu, A. Quon, D. L. Rubin, S. Napel, S. K. Plevritis,Non–small cell lung cancer: Identifying prognostic imaging biomarkers by leveraging publicgene expression microarray data—Methods and preliminary results. Radiology 264, 387–396(2012)

20. V. S. Nair, O. Gevaert, G. Davidzon, S. Napel, E. E. Graves, C. D. Hoang, J. B. Shrager, A. Quon,D. L. Rubin, S. K. Plevritis, Prognostic PET 18F-FDG uptake imaging features are associatedwith major oncogenomic alterations in patients with resected non–small cell lung cancer.Cancer Res. 72, 3725–3734 (2012).

www.ScienceTr

21. P. O. Zinn, B. Majadan, P. Sathyan, S. K. Singh, S. Majumder, F. A. Jolesz, R. R. Colen, Radio-genomic mapping of edema/cellular invasion MRI-phenotypes in Glioblastoma Multiforme.PLOS One 6, e25451 (2011).

22. V. G. Tusher, R. Tibshirani, G. Chu, Significance analysis of microarrays applied to the ionizingradiation response. Proc. Natl. Acad. Sci. U.S.A. 98, 5116–5121 (2001).

23. A. V. Kapp, R. Tibshirani, Are clusters found in one dataset present in another dataset?Biostatistics 8, 9–31 (2007).

24. R. Tibshirani, T. Hastie, B. Narasimhan, G. Chu, Diagnosis of multiple cancer types by shrunkencentroids of gene expression. Proc. Natl. Acad. Sci. U.S.A. 99, 6567–6572 (2002).

25. R. Stupp, M. E. Hegi, W. P. Mason, M. J. van den Bent, M. J. B. Taphoorn, R. C. Janzer, S. K. Ludwin,A. Allgeier, B. Fisher, K. Belanger, P. Hau, A. A. Brandes, J. Gijtenbeek, C. Marosi, C. J. Vecht,K. Mokhtari, P. Wesseling, S. Villa, E. Eisenhauer, T. Gorlia, M. Weller, D. Lacombe, J. G. Cairncross,R.-O. Mirimanoff; European Organisation for Research and Treatment of Cancer Brain Tumourand Radiation Oncology Groups; National Cancer Institute of Canada Clinical Trials Group, Effectsof radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone onsurvival in glioblastoma in a randomised phase III study: 5-Year analysis of the EORTC-NCIC trial.Lancet Oncol. 10, 459–466 (2009).

26. C. J. Vaske, S. C. Benz, J. Z. Sanborn, D. Earl, C. Szeto, J. Zhu, D. Haussler, J. M. Stuart,Inference of patient-specific pathway activities from multi-dimensional cancer genomicsdata using PARADIGM. Bioinformatics 26, i237–i245 (2010).

27. D. R. Macdonald, T. L. Cascino, S. C. Schold Jr., J. G. Cairncross, Response criteria for phase IIstudies of supratentorial malignant glioma. J. Clin. Oncol. 8, 1277–1280 (1990).

28. P. Y. Wen, D. R. Macdonald, D. A. Reardon, T. F. Cloughesy, A. G. Sorensen, E. Galanis, J. DeGroot,W. Wick, M. R. Gilbert, A. B. Lassman, C. Tsien, T. Mikkelsen, E. T. Wong, M. C. Chamberlain,R. Stupp, K. R. Lamborn, M. A. Vogelbaum, M. J. van den Bent, S. M. Chang, Updated responseassessment criteria for high-grade gliomas: Response Assessment in Neuro-Oncology WorkingGroup. J. Clin. Oncol. 28, 1963–1972 (2010).

29. L. B. Nabors, M. Ammirati, P. J. Bierman, H. Brem, N. Butowski, M. C. Chamberlain, L. M. DeAngelis,R. A. Fenstermaker, A. Friedman, M. R. Gilbert, D. Hesser, M. Holdhoff, L. Junck, R. Lawson,J. S. Loeffler, M. H. Maor, P. L. Moots, T. Morrison, M. M. Mrugala, H. B. Newton, J. Portnow,J. J. Raizer, L. Recht, D. C. Shrieve, A. K. Sills Jr., D. Tran, N. Tran, F. D. Vrionis, P. Y. Wen, N. McMillian,M. Ho; National Comprehensive Cancer Network, Central nervous system cancers. J. Natl. Compr.Canc. Netw. 11, 1114–1151 (2013).

30. K. H. Plate, G. Breier, H. A. Weich, W. Risau, Vascular endothelial growth factor is a potentialtumour angiogenesis factor in human gliomas in vivo. Nature 359, 845–848 (1992).

31. D. J. Hicklin, L. M. Ellis, Role of the vascular endothelial growth factor pathway in tumorgrowth and angiogenesis. J. Clin. Oncol. 23, 1011–1027 (2005).

32. R. A. Berkman, M. J. Merrill, W. C. Reinhold, W. T. Monacci, A. Saxena, W. C. Clark, J. T. Robertson,I. U. Ali, E. H. Oldfield, Expression of the vascular permeability factor/vascular endothelialgrowth factor gene in central nervous system neoplasms. J. Clin. Invest. 91, 153–159 (1993).

33. O. L. Chinot, T. de La Motte Rouge, N. Moore, A. Zeaiter, A. Das, H. Phillips, Z. Modrusan, T. Cloughesy,AVAglio: Phase 3 trial of bevacizumab plus temozolomide and radiotherapy in newly diag-nosed glioblastoma multiforme. Adv. Ther. 28, 334–340 (2011).

34. J. J. Vredenburgh, A. Desjardins, J. E. Herndon II, J. Marcello, D. A. Reardon, J. A. Quinn, J. N. Rich,S. Sathornsumetee, S. Gururangan, J. Sampson, M. Wagner, L. Bailey, D. D. Bigner, A. H. Friedman,H. S. Friedman, Bevacizumab plus irinotecan in recurrent glioblastoma multiforme. J. Clin. Oncol.25, 4722–4729 (2007).

35. T. N. Kreisl, L. Kim, K. Moore, P. Duic, C. Royce, I. Stroud, N. Garren, M. Mackey, J. A. Butman,K. Camphausen, J. Park, P. S. Albert, H. A. Fine, Phase II trial of single-agent bevacizumabfollowed by bevacizumab plus irinotecan at tumor progression in recurrent glioblastoma.J. Clin. Oncol. 27, 740–745 (2009).

36. A. Lai, A. Tran, P. L. Nghiemphu, W. B. Pope, O. E. Solis, M. Selch, E. Filka, W. H. Yong, P. S. Mischel,L. M. Liau, S. Phuphanich, K. Black, S. Peak, R. M. Green, C. E. Spier, T. Kolevska, J. Polikoff,L. Fehrenbacher, R. Elashoff, T. Cloughesy, Phase II study of bevacizumab plus temozolomideduring and after radiation therapy for patients with newly diagnosed glioblastoma multi-forme. J. Clin. Oncol. 29, 142–148 (2011).

37. A. Narayana, D. Gruber, S. Kunnakkat, J. G. Golfinos, E. Parker, S. Raza, D. Zagzag, P. Eagan,M. L. Gruber, A clinical trial of bevacizumab, temozolomide, and radiation for newly diagnosedglioblastoma. J. Neurosurg. 116, 341–345 (2012).

38. T. Cascone, J. V. Heymach, Targeting the angiopoietin/Tie2 pathway: Cutting tumor vesselswith a double-edged sword? J. Clin. Oncol. 30, 441–444 (2012).

39. O. L. Chinot, W. Wick, W. Mason, R. Henriksson, F. Saran, R. Nishikawa, A. F. Carpentier,K. Hoang-Xuan, P. Kavan, D. Cernea, A. A. Brandes, M. Hilton, L. Abrey, T. Cloughesy, Bevacizumabplus radiotherapy–temozolomide for newly diagnosed glioblastoma. N. Engl. J. Med. 370,709–722 (2014).

40. H. Colman, L. Zhang, E. P. Sulman, J. M. McDonald, N. L. Shooshtari, A. Rivera, S. Popoff,C. L. Nutt, D. N. Louis, J. G. Cairncross, M. R. Gilbert, H. S. Phillips, M. P. Mehta, A. Chakravarti,C. E. Pelloski, K. Bhat, B. G. Feuerstein, R. B. Jenkins, K. Aldape, A multigene predictor ofoutcome in glioblastoma. Neuro Oncol. 12, 49–57 (2010).

41. M. D. Wilkerson, D. N. Hayes, ConsensusClusterPlus: A class discovery tool with confidenceassessments and item tracking. Bioinformatics 26, 1572–1573 (2010).

anslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 9

R E S EARCH ART I C L E

Dow

nloa

42. S. Monti, P. Tamayo, J. Mesirov, T. Golub, Consensus clustering: A resampling-based methodfor class discovery and visualization of gene expression microarray data. Mach. Learn. 52,91–118 (2003).

43. A. Agresti, Categorical Data Analysis (John Wiley & Sons, Hoboken, NJ, ed. 3, 2013).44. J. H. McDonald, Handbook of Biological Statistics (Sparky House Publishing, Baltimore, MD,

ed. 3, 2014).45. K. R. Lamborn, S. M. Chang, M. D. Prados, Prognostic factors for survival of patients with

glioblastoma: Recursive partitioning analysis. Neuro Oncol. 6, 227–235 (2004).46. J. G. Scott, L. Bauchet, T. J. Fraum, L. Nayak, A. R. Cooper, S. T. Chao, J. H. Suh, M. A. Vogelbaum,

D. M. Peereboom, S. Zouaoui, H. Mathieu-Daude, P. Fabbro-Peray, V. Rigau, L. Taillandier,L. E. Abrey, L. M. DeAngelis, J. H. Shih, F. M. Iwamoto, Recursive partitioning analysis of prog-nostic factors for glioblastoma patinets aged 70 years or older. Cancer 118, 5595–5600 (2012).

47. M. E. Hegi, A. C. Diserens, T. Gorlia, M. F. Hamou, N. de Tribolet, M. Weller, J. M. Kros, J. A. Hainfellner,W. Mason, L. Mariani, J. E. Bromberg, P. Hau, R. O. Mirimanoff, J. G. Cairncross, R. C. Janzer, R. Stupp,MGMT gene silencing and benefit from temozolomide in glioblastoma. N. Engl. J. Med. 352,997–1003 (2005).

Acknowledgments:We thank R. Tibshirani for reviewing the statistical methodologies used inour study. Funding: Research reported in this publication was supported by the National In-stitute of Biomedical Imaging and Bioengineering of the NIH under award no. R01EB020527,and by the National Cancer Institute under award no. R01CA160251. The content is solely theresponsibility of the authors and does not necessarily represent the official views of the NIH. H.I.was supported by the Advanced Residency Training at Stanford (ARTS) Fellowship. S.D.C. acknowl-

www.ScienceTra

edges support for this study from C. and K. Darian and C. Bade. Author contributions: H.I. andO.G. conceived and designed the study; G.R.H., S.D.C., and D.L.R. created the Stanford GBMimaging database; A.S.A., L.A.M., J.J.L., T.L., E.M.W., A.H.F., S.R., S.E., T.D.A., and S.D.C. acquiredimaging data and performed annotations and preprocessing; S.N. provided the software forgenerating quantitative image features; K.W.Y., S.D.C., and G.R.H. provided the single-institutionpatient cohort; H.I. and O.G. analyzed and interpreted the data; H.I. wrote the manuscript; H.I.and O.G. edited the manuscript; G.R.H., S.N., D.L.R., L.A.M., K.W.Y., T.D.A., S.E., A.H.F., T.L., and S.D.C.provided edits; and all authors read and approved the manuscript. Competing interests: O.G. andH.I. have filed a provisional patent on the method described in this manuscript. The authors declarethat they have no competing interests. Data and materials availability: The TCGA molecular dataare available at http://wiki.cancerimagingarchive.net/display/Public/TCGA-GBM. Correspondingimaging data are available at www.cancerimagingarchive.net.

Submitted 12 February 2015Accepted 4 August 2015Published 2 September 201510.1126/scitranslmed.aaa7582

Citation: H. Itakura, A. S. Achrol, L. A. Mitchell, J. J. Loya, T. Liu, E. M. Westbroek, A. H. Feroze,S. Rodriguez, S. Echegaray, T. D. Azad, K. W. Yeom, S. Napel, D. L. Rubin, S. D. Chang, G. R. HarshIV, O. Gevaert, Magnetic resonance image features identify glioblastoma phenotypic subtypeswith distinct molecular pathway activities. Sci. Transl. Med. 7, 303ra138 (2015).

de

nslationalMedicine.org 2 September 2015 Vol 7 Issue 303 303ra138 10

by guest on August 22, 2020

http://stm.sciencem

ag.org/d from

distinct molecular pathway activitiesMagnetic resonance image features identify glioblastoma phenotypic subtypes with

Chang, Griffith R. Harsh IV and Olivier GevaertScott Rodriguez, Sebastian Echegaray, Tej D. Azad, Kristen W. Yeom, Sandy Napel, Daniel L. Rubin, Steven D. Haruka Itakura, Achal S. Achrol, Lex A. Mitchell, Joshua J. Loya, Tiffany Liu, Erick M. Westbroek, Abdullah H. Feroze,

DOI: 10.1126/scitranslmed.aaa7582, 303ra138303ra138.7Sci Transl Med

image-based quantitative biomarkers to be used for patient prognosis.to molecular pathways for targeted therapy but also by survival, indicating the potential for such noninvasivevalidated in a separate cohort of 144 patients. These clusters could be used to stratify patients not only according rim-enhancing cluster, with a hypointense center ringed by hyperintensity. The distinct clusters were further''clusters'': pre-multifocal cluster, with highly irregular tumor shapes; spherical cluster, with defined edges; and that could be used to describe each tumor. The tumors could be grouped into three different phenotypes orat solitary, unilateral tumors from 121 glioblastoma patients and then generated nearly 400 unique image features allowing for targeted therapy without such brain invasion. The authors used magnetic resonance imaging to looknoninvasive determinants of tumor phenotype that would potentially correlate with molecular pathways, thus

sought to identifyet al.targets. For patients with glioblastoma, however, it is invasive to biopsy the brain. Itakura When directing therapies toward tumors, a sample of the cancerous tissue is needed to identify molecular

Brain images create cancer clusters

ARTICLE TOOLS http://stm.sciencemag.org/content/7/303/303ra138

MATERIALSSUPPLEMENTARY http://stm.sciencemag.org/content/suppl/2015/08/31/7.303.303ra138.DC1

CONTENTRELATED

http://stm.sciencemag.org/content/scitransmed/11/473/eaao5253.fullhttp://stke.sciencemag.org/content/sigtrans/8/375/ra42.fullhttp://stm.sciencemag.org/content/scitransmed/7/317/317ra199.fullhttp://stm.sciencemag.org/content/scitransmed/7/310/310ra169.fullhttp://stm.sciencemag.org/content/scitransmed/7/309/309ra163.full

REFERENCES

http://stm.sciencemag.org/content/7/303/303ra138#BIBLThis article cites 43 articles, 17 of which you can access for free

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Terms of ServiceUse of this article is subject to the

registered trademark of AAAS. is aScience Translational MedicineScience, 1200 New York Avenue NW, Washington, DC 20005. The title

(ISSN 1946-6242) is published by the American Association for the Advancement ofScience Translational Medicine

Copyright © 2015, American Association for the Advancement of Science

by guest on August 22, 2020

http://stm.sciencem

ag.org/D

ownloaded from