Interrogation of ERG gene rearrangements in prostate cancer identifies a prognostic 10-gene signature with relevant implication to patients' clinical outcome

Interrogation of ERG gene rearrangements inprostate cancer identifies a prognostic 10-genesignature with relevant implication to patients'clinical outcomeTarek A. Bismar*†‡, Mohammed Alshalalfa§, Lars F. Petersen*, Liang Hong Teng*,Travis Gerke¶, Ashraf Bakkar*, Amal Al-Mami*, Shuhong Liu*, Michael Dolph*,Lorelei A. Mucci¶ and Reda Alhajj§

*Department of Pathology and Laboratory Medicine, University of Calgary and Calgary Laboratory Services,†Department of Oncology, University of Calgary, ‡Southern Alberta Cancer Institute and Tom Baker Cancer Center,§Department of Computer Science, University of Calgary, Calgary, AB, Canada, and ¶Department of Epidemiology,Harvard School of Public Health, Boston, MA, USA

L.F.P. and L.H.T contributed equally to this work

Objectives• ERG-gene rearrangement defines a distinct molecular

subtype of PCA with potential biological and clinicalimplications.

• To identify a molecular signature reflective of thedownstream effects of ERG-mediated transcriptionalregulation with prognostic implication in patients withprostate cancer (PCA).

Material and Methods• We used a singular value decomposition (SVD)

bioinformatics approach to re-analyse gene expression datapreviously generated from 46 prostate tumours, andidentified an ERG-like gene signature.

• The signature was validated on several patient cohorts andindividual genes were correlated to ERG expression andPCA progression.

Results• An ERG-like 10-gene signature was identified and

validated in PCA cohorts of the physician health study(p115) (n = 110) in addition to three independent publicdatasets, and was significantly associated with diseaseprogression, biochemical recurrence and PCA-specificmortality.

• Patients with the ERG-like signature were significantlyassociated with disease recurrence on univariate (hazardratio [HR] 2.6; 95% confidence interval [CI]:1.3–5.2; P =0.004) and multivariate analysis (HR 2.3; 95% CI:1.1–4.6,P = 0.016) compared with patients without this signature.

• Within the group of patients with Gleason score (GS) 6 and 7PCA, the signature added prognostic value beyond GS andidentified patients at higher risk of cancer deaths moreaccurately than GS alone or in combination with ERG status.

• Protein expression of the 10 genes were significantlyassociated with ERG and disease progression regardless ofERG status.

Conclusion• The characterized ERG-like signature was reflective of

aggressive features of ERG-mediated transcription and wasprognostically robust.

• The combination of this signature with clinicopathologicalvariables should be validated prospectively to explore itsclinical utility in stratifying patients with PCA and inidentifying those at higher risk of metastatic and lethaldisease.

KeywordsERG rearrangements, Gleason score, molecular pathways,ERG-like gene signature, prostate cancer, prognosis

Introduction

ETS (E26) fusions have been proposed to constitute a uniquemolecular subtype of prostate cancer (PCA), as evidenced by

both their mutual exclusivity with other proposed PCAsubtypes and the fundamentally different chromatin biologyconferred by ETS-positivity [1–6]. These findings have allowedthe conception of a road map associating specific molecular

© 2013 The AuthorsBJU International © 2013 BJU International | doi:10.1111/bju.12262 BJU Int 2013Published by John Wiley & Sons Ltd. www.bjui.org wileyonlinelibrary.com

subtypes of PCA with progression pathways, and pairingtargeted therapies with particular PCA subtypes, similar toalgorithms used in breast cancer and other haematologicalmalignancies [1]. To date, however, the prognostic value of ETSgene rearrangements has remained uncertain and has generatedconflicting results. The majority of published reports showingprognostic value have been documented on cohorts reflective ofthe natural history of PCA, while those not able to confirm anyprognostic significance have been investigated in surgicallytreated cohorts, with few exceptions [7–20]. The vast majority ofETS gene rearrangements involve ERG, a prototypical ETStranscription factor that transcriptionally activates hundreds ofgenes. While not all of these rearrangments have prognosticvalue, several studies have characterized novel genes associatedwith ERG gene rearrangements and their respective signallingpathways [21–29]. Furthermore, other studies have chosen adifferential expression approach between two groups ofERG-positive and -negative tumours to identify downstreamderegulated genes and genes with biological significance inpatient prognosis [30,31]. Taken together, these studies showthat a thorough interrogation and characterization of themolecular mechanisms that result in, or result from, ERG-generearrangements will enable us to identify novel PCA biomarkersas well as characterize new pathways potentially involved indisease progression, allowing a better prediction of patientprognosis.

It is also possible that a subset of these target genes couldcontribute to the ‘ERG-ness’ of a tumour, that is, a measure ofthe strength by which ERG-mediated regulation is active in thedownstream gene pathway, independent of whether the tumouris ERG-positive or -negative; therefore, we hypothesized thatERG-ness, or an ERG-associated gene expression signature,would have more prognostic value and be more clinicallyrelevant than ERG status itself. In the present study, we used adifferential expression microarray approach, combined with abioinformatics interrogation of ERG-negative and ERG-positivetumours using the singular value decomposition (SVD) method[32], to identify novel genes associated with disease progressionand with prognostic implication for cancer recurrence anddisease-specific mortality, regardless of ERG status. TheSVD-based approaches discover genes that have high entropyon the whole expression data, unlike other methods that testsingle genes without considering the overall effect on the data.We further validate the prognostic significance of this signatureacross several cohorts of PCA patients.

Materials and MethodsPatient Cohorts and Samples

Several cohorts were interrogated in this study. To investigatethe downstream ERG transcriptional regulation signature, weused a cohort of 52 patients with castration-resistant disease(54 samples) which was interrogated for gene expression

profiling and further bioinformatics clustering based on ERGstatus. To assess the relationship of the 10-gene signature topatients’ clinical outcome, we used three publically availablecohorts, the Swedish GSE8402, Taylor and Glinsky [33–35]. Wealso used a case–control subgroup of patients from theHarvard University Physician Health study (n = 115) whowere treated by radical prostatectomy and for whominformation regarding the development of metastasis andlethal disease was available. Protein expression of the 10-genesignature relative to disease progression and ERG status, wasassessed using a progression cohort consisting of 61 patients,with samples representing disease progression (benignprostate tissue, localized PCA and castration-resistant disease),obtained from radical prostatectomy and TURP samples,respectively. Additional information about the cohorts can befound in the Supporting Information (supplementary Methodssection). The study was approved by the institutional reviewboard at the University of Calgary, Calgary, AB, Canada andthe Harvard University School of Public Health.

Assessment of ERG-Gene Rearrangement

ERG status was confirmed using a fluorescent in situhybridization (FISH) break-apart probe assay andimmunohistochemistry (IHC) on a tissue microarray ofthe first Cohort (above) treated by different androgendeprivation therapies as previously described [13]. To omitfalse-positive/-negatives, only cases with consistent FISH andIHC results were included in the SVD approach.

Expression Profiling

We re-analysed gene expression data, which was previouslypublished and generated using a complementaryDNA-mediated annealing selection and ligation assay (DASL;Illumina, San Diego, CA, USA) to regroup tumour samplesinto ERG-positive and -negative cases [27].

Immunohistochemistry

We used the Ventana autostainer system (NEXES IHC model;Ventana Medical Systems Inc., Tuczon, AZ, USA) to assessthe protein expression by IHC, using the antibodies andconditions listed in Table S1, and according to manufacturer’sprotocol outlined in full in the supplementary Methodssection.

Pathological Analysis

The samples used in gene expression profiling were previouslyvalidated by at least two pathologists. Protein intensityexpression was assessed semi-quantitatively by twopathologists (L.H.T. and T.A.B.) using a four-tiered system (0,negative; 1, weak; 2, moderate; and 3, strong) without priorknowledge of clinical information.

Bismar et al.

© 2013 The Authors2 BJU International © 2013 BJU International

Statistical Analysis

One-way ANOVA tests were performed to determine if thescores for each marker were significantly different betweenbenign tumours, localized PCA and castration-resistant PCA.Data for each of the 10 markers from the four-tiered IHCanalyses were used for these tests.

Phi-correlations were used to test each marker’s associationwith ERG status in PCA samples. Marker association withERG expression was categorized into two groups, based onERG intensity (negative and weak) vs (moderate and high)expression. Analyses were performed using SPSS v.16 (IBMCorp., Somers, NY, USA). A two-tailed P-value ≤0.05 wasconsidered to indicate statistical significance.

Multivariate Cox regression analysis of 150 patients fromTaylor data was initially performed, including two presurgicalclinical and pathological variables (prebiopsy serum PSA andbiopsy Gleason score (GS) in addition to the 10-gene signatureto predict biochemical recurrence.

Gene Selection and Computational Analysis

We developed a new technique based on SVD to identifygenes with high entropy between the two cancer classes(ERG-gene rearrangement vs ERG no rearrangement). SVD isa linear transformation of the expression data from n genes bym samples represented by a matrix Am_n to the reduceddiagonal L-eigengenes by L-eigenarrays matrix, where L = min(n; m). and si; i = 1; : : :;L are the singular values. We calculatedthe normalized relative significance pk of the k-th eigengenefor Am_n as follows:

pks

s

k

ii

l=

=∑

2

2

1

and the Shannon entropy of the data represented by Am_n iscalculated as:

E AL

pk pkmxnk

L

( )log( )

log( )= −=

∑1

1

Varshavsky et. al [32]. defined the contribution of the i-th gene,denoted CEi, by considering a leaving out comparison as:

CE E A E Ai mxn m xni= − −( ) ( )1

Where Am xni

−1 is the matrix Am_n with the i-th row deleted. TheSVD-based approaches discover genes that show change in theexpression level across samples. Genes that show a large changeacross samples compared with genes that show a small change aremore likely to have a large SVD value (CEi). Genes with a highSVD value pose large entropy compared with the expression ofthe other genes. To assess the entropy of gene i, we find the

entropy of the whole gene expression E (Am_n), including gene i,and then we find the entropy of the gene expression without genei. Genes are sorted based on CEi value and the top genes areselected as significant. We preprocessed the data by averaginggene expression across the normal and cancer samples. Geneswith high entropy across the two average samples are anticipatedto have a global differential expression pattern. The advantage ofthis method compared with the traditional t-test and SignificantAnalysis of Microarrays is that it considers the effect of each geneon all other genes. Genes that have high entropy on the othergenes are ranked high and identified as significant biomarkers.

ResultsComputational Analysis and Characterization ofSignificant Differentially Expressed Genes

We used the SVD-based gene selection method [32] to discoverhighly differentially expressed genes ranked based on theircontribution to the overall entropy of the expression data.Genes that were highly ranked were identified as potentialbiomarkers discriminating between ERG-rearranged (positive)and ERG-non-rearranged (negative) samples. Only samples withknown and consistent ERG status (based on IHC and FISH)were used in this study. A total of 46 samples, 27 ERG-negative(ERG0) and 19 ERG-positive (ERG1) were included in the finalanalysis.

Based on SVD analysis, a list of 10 candidate genes wasidentified showing significant expression changes betweenERG0 and ERG1 samples (Fig. 1A; Table S2). The 10-gene listcorrelated to ERG status with a 76% accuracy rate using alinear support vector machine. The 10-gene ERG-likesignature relationship to ERG status was validated on threeindependent public cohorts (the Swedish GSE8402, Taylor andGlinsky cohorts) with a 65–83% accuracy rate (data notshown) [34–36].

Association of Protein Levels of Identified Geneswith ERG Protein Expression

To validate the association of the genes identified bycomputational methods with ERG-gene rearrangement, weanalysed the protein expression of those genes in comparisonwith ERG protein expression using a progression tissuemicroarray of 86 samples (320 cores) belonging to 61 patients.In each subset of samples (i.e. ERG0 or ERG1), the intensitylevel for each marker was categorized as absent/weak ormoderate/strong. Seven of 10 markers (CHD5, Ankyrin,MEIS2, FRP-3, LEF1, PLA2G7 and WNT2) were differentiallyexpressed relative to ERG protein at a significant level, whileING3 and ANXA4 showed similar, although nonsignificant,trends. In these samples, Syntenin expression was notsignificantly associated with ERG expression; however, thiswas confirmed using a different cohort with a larger number

ERG-like signature in prostate cancer prognosis

© 2013 The AuthorsBJU International © 2013 BJU International 3

of samples (data not shown). The mean protein expression ofeach of the 10 markers in PCA progression and its relation toERG is shown in Fig. 1B and Table 1. Antibody specificitieswere confirmed by Western blots on protein isolated fromfrozen tissues from patients (data not shown).

Association of Protein Expression Levels of IdentifiedGenes with PCA Progression

To test our hypothesis that the ERG signature is reflective ofan ERG-mediated transcriptional regulation regardless of ERGstatus, we analysed the expression levels of the 10 markers inbenign, localized and castration-resistant prostate cancer tissueusing the progression tissue microarray described above. Themean intensity level for each marker was then plotted relativeto disease progression (Fig. 2A). Protein expression of the 10

markers was differentially expressed between several stagesof PCA progression, with ANAX4 being significantlydownregulated with disease progression (P < 0.001; Table 1).

ERG-Gene Signature in Relation to Prognosis andOverall Survival of Patients with PCA

As PCA is known to be heterogeneous, we hypothesized thatERG status alone might not be very predictive of patientprognosis, consistent with earlier reports, and that integratingthe status of other genes reflective of the downstream effectsof ERG-mediated transcription in each tumour sample mightimprove overall patient prognosis. To validate our signaturein other cohorts, the 10-gene signature was represented asa vector of length 10 (reference vector), where the genehas a value of 1 if it is overexpressed and 0 when it is

Fig. 1 Association of markers with ERG. (A) Heatmap of the 10-gene signature across ERG1 and ERG0 samples in the Swedish cohort. (B) IHC plots of

the protein expression levels of each of the 10 markers relative to ERG protein expression, plotted as negative/weak vs moderate/high intensity. *Refers

to significant correlation, P < 0.05. Red: over expressed; Green: down expression; black: missing data.

3

2

1

0

−1

−2

−3

ERG

0ER

G 0

ERG

0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

1

ERG

1ER

G 1

ERG

1

ERG

1

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0ER

G 0

ERG

0

ERG

1

ERG

1ER

G 1

ERG

1ER

G 1

ERG

1ER

G 1

ERG

1

100%80%60%40%20%

0%

100%80%60%40%20%

0%ERG− ERG+ ERG−

100%80%60%40%20%

0%

100%80%60%40%20%

0%ERG− ERG+ ERG+ERG−

100%80%60%40%20%

0%

100%80%60%40%20%


100%80%60%40%20%

0%

100%80%60%40%20%


100%80%60%40%20%

0%

100%80%60%40%20%


ERG+

CHD5 Syntenin

WNT2 FRZB

ANK3 MEIS2

ING3 PLA2G7

LEF1 ANXA4

Moderate/HighExpression

Negative/WeakExpression

*

*

* *

*

*

*

*

B

% o

f Cas

es%

of C

ases

% o

f Cas

es%

of C

ases

% o

f Cas

es

ANXA4

Syntenin

WNT2

PLA2G7

CHD5

ANK3

FRZB

MEIS2

LEF1

ING3

A

Bismar et al.


downregulated. To stratify samples based on the signature,the expression of each of the genes across all samples wascategorized into high and low expression, and then a vectorof length 10 for each sample was constructed (1 for highexpression and 0 for low expression). The final step was tofind the correlation between the reference vector and thesamples’ vectors; samples with high correlation (r > 0.5) wereconsidered as ERG1-like and the remainder were consideredERG0-like.

We first investigated the 10-gene signature using the PHS(physician Health study). Samples were grouped intoERG-1-like and ERG-0-like samples and we assessed theirassociation with cancer lethality. The 10-gene signature wasassociated with lethal disease with an odds ratio (OR) of 4.33(95% CI: 1.82–10.3; P < 0.001) compared with ERG statusalone: OR = 1.49 (95% CI: 0.68–3.26, P = 0.32). Wheninvestigating net reclassification improvement, the 10-genesignature showed significant increases in sensitivity (P = 0.02)and specificity (P = 0.008) for predicting lethal diseasecompared with GS alone.

We then applied the 10-gene signature to various publiclyavailable patient datasets (Swedish, Taylor and Glinskycohorts) [34–36] to investigate the strength of our genesignature in predicting patients’ prognosis compared withERG expression alone. In the Swedish cohort, ERG status wasdetermined by FISH; whereas, in both the Taylor and theGlinsky cohorts it was predicted based on ERG-geneexpression (samples with ERG expression above the thirdquartile range were considered ERG-positive). Based on ERGstatus, we grouped samples into ERG1 (fusion-positive) andERG0 (fusion-negative). We also used our 10-gene model togroup patients into ERG1-like and ERG0-like signatures toinvestigate the significance of the signature in stratifyingpatients into different prognostic groups.

In the Swedish cohort, patients with ERG1 status (n = 46) wereat higher risk of lethal disease than those with ERG0 (n = 226;

Fig. 3A; hazard ratio [HR]: 1.25, 95% CI: 0.99–1.58, P = 0.005).Application of our 10-gene signature separated patients intoERG1-like (n = 22; seven were ERG1) and ERG0-like patients(n = 259), which showed a stronger association with patientprognosis than did ERG status alone (Fig. 3B; HR: 2.38,95%CI: 1.45–3.8, P < 0.001). Univariate analysis using a Coxproportional hazard model confirmed that the 10-gene modelwas more significantly associated with overall survival thanwas ERG status alone (P = 0.003 vs. P = 0.053, respectively).

In the Glinsky cohort, patients with high ERG (n = 32) werenot well separated from patients with low ERG (n = 47;Fig. 3C; HR: 0.76, 95% CI: 0.38–1.47, P = 0.15), so we furthergrouped samples into high risk (n = 21) and low risk (n = 58)as described by Varambally et al. [37] using ERG status, butthe poor separation remained (Fig. 3D; HR: 1.4, 95% CI:0.67–2.67, P = 0.47). We then used our 10-gene signaturewhich was able to separate patients into high-risk (n = 18,11 were ERG1) and low-risk (n = 61) groups based onpost-surgical biochemical recurrence (Fig. 3E; HR: 1.8, 95%CI: 1.2–2.8, P = 0.15). Using univariate analysis, the 10-genemodel was more associated with prognosis, albeit atborderline significance (P = 0.1) compared with ERG statusalone (P = 0.4). Similarly, in the Taylor cohort, ERG expressionwas not effective in separating patients into clinically distinctgroups (Fig. 3F; HR: 0.7, 95%CI: 0.5–1.5, P = 0.33), but the10-gene model successfully separated patients into high-riskERG1-like (n = 23; nine were ERG1) and low-risk ERG0-like(n = 117) groups (Fig. 3G; HR: 3.2, 95%CI: 2.2–5.8, P = 0.0026).

The 10-gene signature was also able to classify samples withvery aggressive forms of PCA (high GS and highly metastatic;i.e. cluster 5 in Taylor data) from the other clusters, slightlymore accurately than ERG status alone (80%; HR: 1.07, 95%CI: 0.4–1.6 vs 77%; HR: 0.87, 95% CI: 0.2–1.4).

We also used Cox regression multivariate analysis to assess thesignificance of the ERG-like signature in stratifying patientsbefore implementing treatments, so we included presurgical

Table 1 Mean protein expression of the 10 markers in relation to PCA progression and their signifcance to ERG.

Markers ProgressionP value

BenignMean ± SD

Localized PCAMean ± SD

CRPCMean ± SD

ERG associationP value

CHD5(cyto) <0.001 1.81 ± 0.430 1.98 ± 0.311 1.27 ± 0.522 0.182Ankyrin(cyto) <0.001 0.34 ± 0.509 1.52 ± 0.517 1.61 ± 0.502 0.203Ankyrin(mem) <0.001 1.55 ± 0.724 2.09 ± 0.684 1.94 ± 0.725 0.010Syntenin(cyto) <0.001 0.59 ± 0.523 2.04 ± 0.528 2.04 ± 0.713 0.378Meis2(cyto) <0.001 2.51 ± 0.901 1.23 ± 0.661 1.50 ± 0.816 0.587FRP-3(cyto) <0.001 1.07 ± 0.596 1.68 ± 0.592 2.53 ± 0.539 0.011LEF1(nuclear) <0.001 1.26 ± 0.822 1.90 ± 0.640 1.00 ± 0.833 0.000PLA2G7(cyto) <0.001 1.04 ± 0.630 1.79 ± 0.629 1.63 ± 0.809 0.000WNT2(cyto) <0.001 0.58 ± 0.597 1.73 ± 0.610 1.96 ± 0.690 0.001ANX4 (cyto) <0.001 1.95 ± 0.590 1.60 ± 0.660 0.98 ± 0.887 0.083ING3(nuclear/cyto) <0.001 1.59 ± 0.496 2.19 ± 0.549 2.17 ± 0.509 0.082

CRPC, castration-resistant prostate cancer.



BEN PCA CRPCA

CHD5

Syntenin

WNT2

FRP-3(FRZB)

ANK3(Ankyrin G)

MEIS2

ING3

PLA2G7

LEF1

ANXA4(AnnexinA4)

BEN

PCA

CRP

CA

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Rela

tive

Inte

nsity

Fig. 2 Association of markers with prostate

cancer progression. (A) Selected images of

tissue microarray elements representing IHC

analysis of protein levels for the 10 marker

genes in benign prostate gland (BEN),

localized prostate cancer (PCA) and

castration-resistant PCA (CRPCA). Relative

levels of proteins as assessed by blinded

pathology analysis of tissue microarrays (n =320 cores) are provided to the right.

Bismar et al.


Fig. 3 Application of ERG-like 10-gene signature better predicts patients' outcome than ERG status alone. (A) Kaplan–Meier survival plot showing that

ERG status can separate patients into two groups in the Swedish PCA cohort. (B) Kaplan–Meier plot showing that the 10-gene model has significant

association with the survival outcome in the Swedish PCA cohort, ERG1-like patients are the cluster of patients enriched with ERG1 samples. (C)

Kaplan–Meier plots of patients grouped based on ERG status. (D) Kaplan–Meier plots of patients grouped into high and low risk based on Varambally

et al. [37], and (E) samples grouped based on the 10-gene model in the Glinsky PCA cohort. (F) Kaplan–Meier plots of patients grouped based on

ERG expression. (G) Kaplan–Meier plots of patients, grouped based on 10-gene model in the Taylor PCA cohort. (H) Kaplan–Meier plots of patients

with GS 6 and 7, grouped based on the 10-gene signature in the Swedish PCA cohort.

A B

C D E

F G

Swedish Prostate Cancer Cohort

Glinsky Prostate Cancer Cohort

Taylor Prostate Cancer Cohort

H

n = 46n = 22

n = 32

n = 48

n = 23 n = 13

n = 18n = 21

n = 226 n = 259

n = 47

n = 92

n = 117

n = 187

n = 61n = 58

Gleason Score 6&7 Patients(Swedish PCA Cohort

High-risk patientsLow-risk patients

Cum

ulat

ive

Surv

ival

Pro

babi

lity

Cum

ulat

ive

Surv

ival

Pro

babi

lity

Rela

pse

Free

Sur

viva

lPr

obab

ility

Rela

pse

Free

Sur

viva

lPr

obab

ility

Rela

pse

Free

Sur

viva

lPr

obab

ility

Rela

pse

Free

Sur

viva

lPr

obab

ility

Rela

pse

Free

Sur

viva

lPr

obab

ility

Cum

ulat

ive

Surv

ival

Pro

babi

lity

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

00 50 100 150 200 250 300

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

00 50 100 150 200 250 300

Time to death, months Time to death, months

ERG positive patientsERG negative patients



ERG positive- like patientsERG negative- like patients




Logrank p = 0.005Cox p = 0.053

Logrank p = 0.0006Cox p = 0.0005

1

0.9

0.8

0.7

0.6

0.5

0.4

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.20 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90

Logrank p = 0.15

Cox p = 0.4


Logrank p = 0.0026Cox p = 0.00086

Logrank p = 0.0001Cox p = 0.0001



Time to recurrence, months Time to recurrence, months Time to recurrence, months

1

0.95

0.9

0.85

0.8

0.75

0.7

0.650 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100Time to recurrence, months Time to recurrence, months Time to death, months

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

1

0.8

0.6

0.4

0.2

00 50 100 150 200 250



variables: needle biopsy GS and pre-biopsy serum PSA fromthe Taylor cohort (n = 150). Patients with ERG1-like signaturehad an overall HR of 2.6 (95% CI: 1.3–5.2; P = 0.004) and HR2.3 (95% CI:1.1–4.6; P = 0.016) for predicting PCA relapse onunivariate and multivariate analysis, respectively (Table 2).

Next we assessed the ERG-like signature in terms of its abilityto stratify subgroups of patients with GS 6 and 7 PCA, giventhat these Gleason patterns represent the majority of patientsencountered in clinical practice, where current clinical andpathological variables are not able to differentiate aggressivefrom indolent disease. Focusing on patients with GS 6 and 7from the Swedish cohort (n = 200), our 10-gene signature wasstill able to separate patients into high-risk (n = 13) andlow-risk (n = 187, 13 of which were ERG positive) groups(Fig. 3H; HR: 3.5, 95% CI: 1.8–6.6, P < 0.001), compared withERG status alone, which separated patients into high-risk (n =26) and low-risk (n = 174) groups and showed a lower HR forlethal disease (HR: 2.4, 95% CI: 1.5–3.9, P < 0.001). Using amixed clinical-molecular signature (i.e. GS and ERG-likesignature) we were able to identify higher-risk patients moreaccurately than by using GS or ERG status alone (Fig. 3H;Table 3).

Figure 4 is a flow chart of the multiple analysis and cohortsevaluated in the present study from initial bioinformatics

analysis to developing of the 10-gene signature and validationof the signature using independent and well-annotatedcohorts.

DiscussionPredicting aggressive disease is one of the most important andnecessary steps in cancer management. This is especially truein PCA, where overdiagnosis resulting from PSA screeninghas reached alarming levels. Currently, a reliable distinctionbetween indolent and aggressive PCA is not achievable basedon pathological and clinical variables alone. ERG-generearrangements are one of the most common gene alterationsaffecting PCA [1,9]. Expression profile studies suggest thatERG represents a subset of prostate tumours that sharespecific progression pathways with potential prognostic andtherapeutic implications [9].

As PCA is among the most heterogeneous tumours, it isexpected that predicting tumour progression would be moreachievable and reliable based on a multi-gene model ratherthan individual genes. This is similar to what is being clinicallyimplemented for patients with breast and colon cancer, withthe use of an ‘oncotype DX’ assay test. In the current study, weidentified and characterized a 10-gene signature that shows

Table 2 Univariate and multivariate analysis for cancer recurrence in PCA: taylor data.

Variable N Univariate analysis Multivariate analysis

HR (95% CI) P HR (95% CI) P

Signature 0.004 0.016ERG0-like 114 1(ref) 1(ref)ERG1-like 23 2.6 (1.3–5.2) 2.3 (1.1–4.6)

Biopsy PSA 0.13 0.281–4 27 1(ref) 1(ref)4–10 85 0.9 (0.3–2.3) 0.9 (0.3–2.3)10–40 23 1.45 (0.8–2.4) 1.26(0.7–2.1)

Biopsy GS 5.0e-4 0.011≤7 125 1(ref) 1(ref)>7 12 4.0 (1.8–9) 2.9 (1.2–6.9)

Table 3 Hazard ratio for cancer-specific mortality in the Swedish Cohort using GS alone, GS and ERG-gene rearrangments and GS with ERG-likesignature.

Group Number of samples HR (95%CI) P value/Cox value

GS 7 Patients GS 7 alone GS 7(3+4) GS 7(4+3) 2.23 (1.5–3.5) 2 × 10–4/2.4 × 10–4

n = 79 n = 38GS 7 + ERGstatus

GS 7(3+4) and ERG0 GS 7(4+3) or ERG1 1.8 (1.4–2.5) 9 × 10–4/5.6 × 10–5

n = 61 n = 44GS 7 + ERG-likesignature

GS 7(3+4) and ERG0-like GS 7(4+3) or ERG1-like 2.52 (1.6–3.4) 3 × 10–5/6.2 × 10–6

n = 72 n = 45GS 6,7 Patients GS 6,7 GS 6 +GS 7(3+4) GS 7(4+3) 3 (2–4.5) <10–7/5.3 × 10–8

n = 162 n = 38GS 6+7 + ERGsignature

GS 6 +GS 7(3+4) and ERG0 GS 7(4+3) or ERG1 1.5 (1.2–2) <10–5/5 × 10–5

n = 136 n = 46GS 6+7 + ERG-likesignature

GS 6 +GS 7(3+4) and ERG0-like GS 7(4+3) or ERG1-like 3.2 (2.1–4.1) <10–10/7.5 × 10–11

n = 153 n = 45

Bismar et al.


potential for further development as a signature for aggressiveand indolent PCA. This panel was identified by a combinationof computational analysis and biological assays and validatedon several well-annotated and large cohorts. The signatureidentified was confirmed to be more accurate than ERG geneexpression alone, which at times was found to be unreliable(based on the Glinsky and Taylor surgical charts). The factthat the signature was reflective of ERG status in only 76% ofpatients is worth noting, as it signifies the ability to reflecttranscriptional regulation within a tumour sample regardlessof its ERG expression level. Hence, it is superior to ERG as apotential biomarker for aggressive disease (based on validationcohorts above). Our signature was also robust in multivariateanalysis, which included biopsy GS and prebiopsy serum PSA,two of the most powerful and currently most used biomarkers

in clinical practice. Moreover, the 10-gene signature was ableto stratify patients in the intermediate grade category (GS 6and 7) and identify two distinct classes within this group ofpatients. When implementing a mixed clinical and molecularmodel based on a combination of GS and the 10-genesignature, we were more accurate in identifying aggressivetumours that could have been misclassified if judged by GSalone. This indicates the strength of our signature comparedwith other models, where combining the signature with GSdid not improve the predictive accuracy [30]. Our 10-genesignature, coupled with other preclinical variables, such as GSand serum PSA, could be the initial step towards improvingour ability to stratify patients into different prognostic groupsof aggressive and indolent disease before implementingdefinite therapies. It could enable us to offer selected patients

Fig. 4 Flow chart of the multiple analysis and cohorts evaluated in this study from initial bioinformatics analysis to the development of the 10-gene

signature and validation of the signature using independent and well-annotated cohorts. RFS, recurrence-free survival; OS; overall survival; KM,

Kaplan–Meier, GS Gleason score.

Initial cohort (n=46)

ERG-fusion-positive

ERG-fusion-

negative

Group samples based on ERG status

Gene selection by SVD

10-gene signature

Validating the protein expression of the 10-gene model on 86 samples (320 cores) in relation to disease progression and ERG expression

Biological validationComputational validation

Assess association between 10-gene model and lethal disease using PHS data

Glinsky data Swedish

data

Taylor data

Validation of 10-gene model on independent datasets

Use KM and Cox regression to

associate model with RFS compared

with ERG expression

Use KM and Cox to associate model

with RFS compared with ERG expression

Multivariate regression analysis

Use KM and Cox to associate model

with OS compared with ERG status, GS

Assess the power of the model to stratify high risk samples in

subset (GS6,7) compared with ERG



expectant therapy, based on the signature of their tumours atthe time of prostate biopsy, thus avoiding over-treatments andunnecessary harmful side effects for a subgroup of patientswith indolent disease.

These data collectively confirm that the expression signatureof our 10-gene model is more reflective of the downstreameffects of ERG-mediated transcription in PCA than the ERGgene alone. It further documents that some ERG-negativetumours are molecularly, biologically and prognosticallymore related to other ERG-positive tumours, based on theERG-mediated transcriptional regulation contributing to PCAprogression (the same is also true for ERG-positive tumoursbeing classified as ERG0-like signature). Our model is furtherevidence of the heterogeneous and multifocal nature of PCA.The 10-gene signature model is more robust than otherpublished ERG-related signatures, as it is able to identify moreaccurately patients in the higher-risk group who could havebeen misclassified as having a ‘favourable prognosis’ based onGS alone.

Finally, the significance of this multi-gene model is currentlybeing validated and further refined using quantitativeRT-PCR-based methods which would allow us to betterquantify the expression levels of the 10 proposed genes withina given patient’s sample using appropriate housekeepinggenes. This would be similar to developing a ‘recurrence score’per patient, based on the relative risk for disease recurrence orlethal outcome. By incorporating such tests in PCA, we hopeto be able to more accurately predict those tumours to eitherindolent or aggressive nature, especially those within the GS 6and 7 category. Those patients represent the commonestgroup seeking medical attention and the most difficult topredict based on current clinical and pathological variablesalone.

In summary, we identified and validated a multi-gene model,reflective of the transcriptional regulation of downstream ERGin a given tumour sample, which is more robust than ERGexpression or GS alone in identifying patients at higher risk ofdisease recurrence and lethal outcome. This signature, coupledwith clinicopathological variables in prostate biopsy, couldenable us to separate aggressive from indolent disease and toidentify patients at highest risk for cancer progression andlethal disease. Finally, functional studies investigatingindividual genes within this signature could shed light onnovel potential pathways associated with disease progressionand with therapeutic potential.

AcknowledgementsThis work was supported in part by the Prostate CancerFoundation Young Investigator Award (to T.A.B). This workwas also supported by Prostate Cancer Canada and is proudlyfunded by the Movember Foundation-Grant #B2013-01. Theauthors thank Felix Feng for scientific input and discussion.

Conflict of InterestT. A. B. is a co-inventor on a patent filed by The University ofCalgary covering the 10-gene signature prognostic andtherapeutic implementation in prostate and breast cancer andleukaemia. No other conflict of interest exists.

References1 Tomlins SA, Rhodes DR, Perner S et al. Recurrent fusion of TMPRSS2

and ETS transcription factor genes in prostate cancer. Science 2005; 310:644–8

2 Tomlins SA, Mehra R, Rhodes DR et al. TMPRSS2:ETV4 gene fusionsdefine a third molecular subtype of prostate cancer. Cancer Res 2006; 66:3396–400

3 Helgeson BE, Tomlins SA, Shah N et al. Characterization ofTMPRSS2:ETV5 and SLC45A3:ETV5 gene fusions in prostate cancer.Cancer Res 2008; 68: 73–80

4 Esgueva R, Perner S, J LaFargue C et al. Prevalence of TMPRSS2-ERGand SLC45A3-ERG gene fusions in a large prostatectomy cohort. ModPathol 2010; 23: 539–46

5 Rickman DS, Pflueger D, Moss B et al. SLC45A3-ELK4 is a novel andfrequent erythroblast transformation-specific fusion transcript in prostatecancer. Cancer Res 2009; 69: 2734–8

6 Palanisamy N, Ateeq B, Kalyana-Sundaram S et al. Rearrangements ofthe RAF kinase pathway in prostate cancer, gastric cancer and melanoma.Nat Med 2010; 16: 793–8

7 Hoogland AM, Jenster G, van Weerden WM et al. ERGimmunohistochemistry is not predictive for PSA recurrence, localrecurrence or overall survival after radical prostatectomy for prostatecancer. Mod Pathol 2012; 25: 471–9

8 Rubin MA, Maher CA, Chinnaiyan AM. Common gene rearrangementsin prostate cancer. J Clin Oncol 2011; 29: 3659–68

9 Tomlins SA, Bjartell A, Chinnaiyan AM et al. ETS gene fusions inprostate cancer: from discovery to daily clinical practice. Eur Urol 2009;56: 275–86

10 Minner S, Enodien M, Sirma H et al. ERG status is unrelated to PSArecurrence in radically operated prostate cancer in the absence ofanti-hormonal therapy. Clin Cancer Res 2011; 17: 5878–88

11 Toubaji A, Albadine R, Meeker AK et al. Increased gene copy number ofERG on chromosome 21 but not TMPRSS2-ERG fusion predicts outcomein prostatic adenocarcinomas. Mod Pathol 2011; 24: 1511–20

12 Attard G, Clark J, Ambroisine L et al. Duplication of the fusion ofTMPRSS2 to ERG sequences identifies fatal human prostate cancer.Oncogene 2008; 27: 253–63

13 Darnel AD, Lafargue CJ, Vollmer RT, Corcos J, Bismar TA.TMPRSS2-ERG fusion is frequently observed in Gleason pattern 3prostate cancer in a Canadian cohort. Cancer Biol Ther 2009; 8: 125–30

14 FitzGerald LM, Agalliu I, Johnson K et al. Association of TMPRSS2-ERGgene fusion with clinical characteristics and outcomes: results from apopulation-based study of prostate cancer. BMC Cancer 2008; 8: 230

15 Gopalan A, Leversha MA, Satagopan JM et al. TMPRSS2-ERG genefusion is not associated with outcome in patients treated byprostatectomy. Cancer Research. 2009; 69: 1400–6

16 Hermans KG, Boormans JL, Gasi D et al. Overexpression ofprostate-specific TMPRSS2(exon 0)-ERG fusion transcripts correspondswith favorable prognosis of prostate cancer. Clin Cancer Res 2009; 15:6398–403

17 Nam RK, Sugar L, Wang Z et al. Expression of TMPRSS2:ERG genefusion in prostate cancer cells is an important prognostic factor for cancerprogression. Cancer Biol Ther 2007; 6: 40–5

18 Nam RK, Sugar L, Yang W et al. Expression of the TMPRSS2:ERG fusiongene predicts cancer recurrence after surgery for localised prostate cancer.Br J Cancer 2007; 97: 1690–5

Bismar et al.


19 Rubio-Briones J, Fernandez-Serra A, Calatrava A et al. Clinicalimplications of TMPRSS2-ERG gene fusion expression in patients withprostate cancer treated with radical prostatectomy. J Urol 2010; 183:2054–61

20 Saramaki OR, Harjula AE, Martikainen PM, Vessella RL, Tammela TL,Visakorpi T. TMPRSS2:ERG fusion identifies a subgroup of prostatecancers with a favorable prognosis. Clin Cancer Res 2008; 14: 3395–400

21 Cai J, Kandagatla P, Singareddy R et al. Androgens induce functionalCXCR4 through ERG factor expression in TMPRSS2-ERG fusion-positiveprostate cancer cells. Transl Oncol 2010; 3: 195–203

22 Gupta S, Iljin K, Sara H et al. FZD4 as a mediator of ERGoncogene-induced WNT signaling and epithelial-to-mesenchymaltransition in human prostate cancer cells. Cancer Res 2010; 70:6735–45

23 Haffner MC, Aryee MJ, Toubaji A et al. Androgen-inducedTOP2B-mediated double-strand breaks and prostate cancer generearrangements. Nat Genet 2010; 42: 668–75

24 Hawksworth D, Ravindranath L, Chen Y et al. Overexpression ofC-MYC oncogene in prostate cancer predicts biochemical recurrence.Prostate Cancer Prostatic Dis 2010; 13: 311–5

25 Kunderfranco P, Mello-Grand M, Cangemi R et al. ETS transcriptionfactors control transcription of EZH2 and epigenetic silencing of thetumor suppressor gene Nkx3.1 in prostate cancer. PLoS ONE 2010; 5:e10547

26 Leshem O, Madar S, Kogan-Sakin I et al. TMPRSS2/ERG promotesepithelial to mesenchymal transition through the ZEB1/ZEB2 axis in aprostate cancer model. PLoS ONE 2011; 6: e21650

27 Rickman DS, Chen YB, Banerjee S et al. ERG cooperates with androgenreceptor in regulating trefoil factor 3 in prostate cancer diseaseprogression. Neoplasia 2010; 12: 1031–40

28 Vainio P, Gupta S, Ketola K et al. Arachidonic acid pathway membersPLA2G7, HPGD, EPHX2, and CYP4F8 identified as putative noveltherapeutic targets in prostate cancer. Am J Pathol 2011; 178: 525–36

29 Wang J, Cai Y, Shao LJ et al. Activation of NF-{kappa}B byTMPRSS2/ERG Fusion Isoforms through Toll-Like Receptor-4. CancerRes 2011; 71: 1325–33

30 Barwick BG, Abramovitz M, Kodani M et al. Prostate cancer genesassociated with TMPRSS2-ERG gene fusion and prognostic ofbiochemical recurrence in multiple cohorts. Br J Cancer 2010; 102:570–6

31 Ribeiro FR, Paulo P, Costa VL et al. Cysteine-rich secretory protein-3(CRISP3) is strongly up-regulated in prostate carcinomas with theTMPRSS2-ERG fusion gene. PLoS ONE 2011; 6: e22317

32 Varshavsky R, Gottlieb A, Linial M, Horn D. Novel unsupervised featurefiltering of biological data. Bioinformatics 2006; 22: e507–13

33 Sboner A, Demichelis F, Calza S et al. Molecular sampling of prostatecancer: a dilemma for predicting disease progression. BMC Med Genomics2010; 3: 8

34 Glinsky GV, Glinskii AB, Stephenson AJ, Hoffman RM, Gerald WL.Gene expression profiling predicts clinical outcome of prostate cancer.J Clin Invest 2004; 113: 913–23

35 Taylor BS, Schultz N, Hieronymus H et al. Integrative genomic profilingof human prostate cancer. Cancer Cell. 2010; 18: 11–22

36 Sboner A, Demichelis F, Calza S et al. Molecular sampling of prostatecancer: a dilemma for predicting disease progression. BMC Med Genomics2010; 3: 8

37 Varambally S, Yu J, Laxman B et al. Integrative genomic and proteomicanalysis of prostate cancer reveals signatures of metastatic progression.Cancer Cell 2005; 8: 393–406

Correspondence: Tarek A. Bismar, University of Calgary, Facultyof Medicine, Departments of Pathology and Laboratory Medicineand Oncology, Rockyview General Hospital, 7007–14th StreetSW, Calgary, AB, Canada T2V 1P9.

e-mail: [email protected]

Abbreviations: PCA, prostate cancer; SVD, singular valuedecomposition; HR, hazard ratio; GS, Gleason score; FISH,fluorescent in situ hybridization; IHC, immunohistochemistry.

Supporting InformationAdditional Supporting Information may be found in theonline version of this article at the publisher’s web-site:

Supplementary Materials and MethodsTable S1 Antibody specification used for IHC.Table S2 Fold-change of genes relative to ERG.



Documents

Interrogation of ERG gene rearrangements in prostate cancer identifies a prognostic 10-gene signature with relevant implication to patients' clinical outcome