genesignaturesofprogressionandmetastasisinrenalcellcancer · university (frankfurt, germany),...

Gene Signatures of Progression andMetastasis in Renal Cell CancerJonJones,1,4 Hasan Otu,1,2 Dimitrios Spentzos,1,2,3 Shakirahmed Kolia,1,2 Mehmet Inan,1

Wolf D. Beecken,4 Christian Fellbaum,5 Xuesong Gu,1Marie Joseph,1

AllanJ. Pantuck,6 DietgerJonas,4 and Towia A. Libermann1,2

Abstract Purpose:To address the progression, metastasis, and clinical heterogeneity of renal cell cancer(RCC).Experimental Design: Transcriptional profiling with oligonucleotide microarrays (22,283genes) was done on 49 RCC tumors, 20 non-RCC renal tumors, and 23 normal kidney samples.Samples were clustered based on gene expression profiles and specific gene sets for each renaltumor type were identified. Gene expression was correlated to disease progression and a metas-tasis gene signature was derived.Results:Gene signatures were identified for each tumor type with 100% accuracy. Differentiallyexpressed genes during early tumor formation and tumor progression to metastatic RCC werefound. Subsets of these genes code for secreted proteins and membrane receptors and are bothpotential therapeutic or diagnostic targets. A gene pattern (‘‘metastatic signature’’) derived fromprimary tumor was very accurate in classifying tumors with and without metastases at the time ofsurgery. A previously described ‘‘global’’ metastatic signature derived by another group from vari-ous non-RCC tumors was validated in RCC.Conclusion: Unlike previous studies, we describe highly accurate and externally validated genesignatures for RCC subtypes and other renal tumors. Interestingly, the gene expression of primarytumors provides us information about the metastatic status in the respective patients and has thepotential, if prospectively validated, to enrich the armamentarium of diagnostic tests in RCC.Wevalidated in RCC, for the first time, a previously described metastatic signature and further showedthe feasibility of applying a gene signature across different microarray platforms. Transcriptionalprofiling allows a better appreciation of the molecular and clinical heterogeneity in RCC.

Renal cell cancer (RCC) has a heterogeneous clinical presen-tation with up to 30% metastatic cases at initial diagnosis andf30% of initially organ-confined cases developing metastasesduring follow-up at variable intervals (1, 2). Although surgery is

highly effective for the treatment of localized RCC, the treat-ment options available for patients with metastatic disease arevery limited (3, 4) and do not take into account the underlyingmolecular differences among different histologic RCC subtypes.Although the response rates to currently available systemictherapies for the treatment of locally advanced and metastaticdisease are in general poor, there is a subset of patients with awell-documented long remission (3). The variability in clinicaloutcome is possibly attributable to the molecular heterogeneityof RCC, which has not yet been fully elucidated.

Transcriptional profiling has emerged as a powerful approachto identify new cancer classes (class discovery), assigning tumorsto known classes (class prediction; refs. 5, 6) and predicting clin-ical outcome solely based on gene expression (7), and has thepotential to affect patient diagnosis, staging, prognosis, and treat-ment. Differentially expressed genes between RCC and normalkidney tissue as well as between RCC samples associated withdifferent outcomes have been described in literature (8–10).

In this large study, we addressed additional clinical andbiological aspects of RCC, focusing on disease progression andmetastasis. We investigated the gene expression profile of ourclear cell RCC samples with respect to tumor progression,identified a clear cell RCC metastatic gene signature, validatedthese gene signatures in independent external data sets, andsuccessfully applied a previously described ‘‘global’’ metastaticgene signature to our clear cell RCC samples. We also identified

Human Cancer Biology

Authors’Affiliations: 1Beth Israel Deaconess Medical Center Genomics Centerand Dana-Farber/Harvard Cancer Center Proteomics Core, 2Bioinformatics Core ofBeth Israel Deaconess Medical Center Genomics Center and Harvard MedicalSchool, and 3Division of Hematology/Oncology, Beth Israel Deaconess MedicalCenter, Boston, Massachusetts; 4Department of Urology and Pediatric Urologyand 5Insititute of Pathology, Johann Wolfgang Goethe University, Frankfurt,Germany; and 6University of California-Los Angeles Kidney Cancer Program,Department of Urology, Jonsson Comprehensive Cancer Center, and David GeffenSchool of Medicine at University of California-Los Angeles, Los Angeles, CaliforniaReceived 11/2/04; revised 3/21/05; accepted 5/16/05.Grant support: NIH grants1RO1CA85467 and P50 CA090381 (T.A. Libermann)and Renal Cell Carcinoma Specialized Programs of Research Excellence NIH P5OCA101942.The costs of publication of this article were defrayed in part by the payment of pagecharges.This article must therefore be hereby marked advertisement in accordancewith18 U.S.C. Section1734 solely to indicate this fact.Note: Supplementary data for this article are available online at http://www.bidmegenomics.org/KidneyCancer/index.html.Requests for reprints: Towia A. Libermann, Beth Israel Deaconess MedicalCenter Genomics Center and Harvard Medical School, Harvard Institutes ofMedicine, 4 Blackfan Circle, Boston, MA 02115. Phone: 617-667-3393; Fax: 617-975-5299; E-mail: [email protected].

F2005 American Association for Cancer Research.doi:10.1158/1078-0432.CCR-04-2225

www.aacrjournals.orgClin Cancer Res 2005;11(16) August15, 2005 5730

Research. on April 21, 2021. © 2005 American Association for Cancerclincancerres.aacrjournals.org Downloaded from

http://clincancerres.aacrjournals.org/

gene signatures that were able to accurately characterize RCCsubtypes and other renal tumors, such as oncocytomas andtransitional cell cancers of the renal pelvis (TCC), both ofwhich are important in the differential diagnosis of RCC andwere underrepresented in previous studies.

Materials andMethods

Tumor samples. A total of 65 frozen primary tumors (23 clear cellRCC, 13 papillary RCC, 7 chromophobe RCC, 10 TCC, and 12oncocytomas), 24 frozen normal kidney tissue samples, and 10 frozenmetastatic tumors (9 distant and 1 nodal metastatic clear cell RCC) werereceived from the Department of Urology, Johann Wolfgang GoetheUniversity (Frankfurt, Germany), Ardais and Collaborative Genomicsand processed with approval from the Institutional Review Board/Human Subject Research Committee at Beth Israel Deaconess MedicalCenter (Boston, MA). All tissues were accompanied by pathologyreports. Part of each sample was frozen in liquid nitrogen immediatelyafter surgery and stored at �80jC. The samples were made anonymousbefore the study. The International Union Against Cancer tumor-node-metastasis classification was used for pathology reports (11).

cDNA synthesis and microarray probe preparation. Total RNA wasisolated using TRIzol reagent (Life Technologies, Rockville, MD)according to the manufacturer’s recommendations. Transcriptionalprofiling was done on HG-U133A Affymetrix GeneChips containing22,283 genes. cDNA was prepared according to the manufacturer’sprotocol. Total RNA (8 Ag) was used in the first-strand cDNA synthesiswith T7-(dT)24 primer (GGCCAGTGAATTGTAATACGACTCACTA-TAGGGAGGCGG-(dT)24) and SuperScript II (Life Technologies). Thesecond-strand cDNA synthesis was carried out according to themanufacturer’s protocol.

Affymetrix GeneChip hybridization. cRNA (20 Ag) was fragmented asdescribed and hybridized with a preequilibrated HG-U133A Affymetrixchip at 45jC for 16 hours. After the hybridization cocktails wereremoved, the chips were washed, stained, and scanned according to themanufacturer’s protocols as described previously (12).

Microarray data preprocessing. Samples were analyzed using dChip(13), where a smoothing spline normalization method was appliedbefore obtaining model-based gene expression values. The samplesused for further analysis had 3V/5V ratios (using glyceraldehyde-3-phosphate dehydrogenase and h-actin probes), present call percentage,or array outlier call percentage within 2 SDs (see Supplementary Data athttp://www.bidmcgenomics.org/KidneyCancer/index.html). The 3V/5Vratio gives an indication of the integrity of starting RNA, efficiency offirst-strand cDNA synthesis, and/or in vitro transcription of cRNA.Present call percentage is the proportion of the mRNA transcripts thatwere expressed (‘‘present’’) in the sample. The percentage of arrayoutlier probes was calculated with the dChip algorithm as describedpreviously (13).

Unsupervised analysis. An average linkage hierarchical clusteringtechnique (14) was used where the metric of similarity was Pearson’scorrelation between the transcription profiles of the samples. Thisrepresented the degree of closeness between samples based on the totalgene expression. We also investigated the projection of samples inthree-dimensional using principal components analysis as a dimension-reduction method to explain the variation in the original data set with areasonable number of variables (15).

Supervised analysis. We used supervised analysis of gene expressionfor finding differentially expressed genes, prediction, and patterndiscovery. To find genes that are significantly differentially expressedbetween two groups of samples, we used combination of two criteria indChip. If the 90% lower confidence bound (LCB) of the fold changebetween the two groups was >2.0 and the t test showed a significantdifference (P < 0.001), the corresponding gene was considered to bedifferentially expressed. Tumor type specificity of gene expression was

assessed using leave-one-out cross-validation with the weighed votingalgorithm and significance of prediction accuracy was assessed byFisher’s test on the confusion matrix and permutation tests (5). For genepattern discovery, we employed a pattern recognition algorithmimplemented by IBM’s Genes@Work software (16). This algorithm isa multivariate gene selection tool where a ‘‘pattern’’ is a subset of genesthat have similar expression values in a subset of the samples in thephenotype set with respect to the control set. We identified geneexpression patterns for clear cell RCC metastasis (P < 10�10) thatdistinguish the metastasis samples from the primary tumor samples.Gene lists for secreted proteins and receptors were derived using ouronline gene annotation tool (http://www.bidmcgenomics.org) thatqueries multiple public databases with >90 identifiers, including celllocation, biological function, protein families, etc. Additional supportwas derived from literature search where necessary. A more detaileddiscussion of the methods along with variables used can be found inSupplementary Data.

Results

Data analysis was carried on 59 primary tumors (22 clear cellRCC, 11 papillary RCC, 6 chromophobe RCC, 8 TCC, and 12oncocytoma), 23 normal kidney tissue samples, and 1 nodaland 9 distant metastatic clear cell RCC samples. The clinicaland pathologic characteristics of the samples are presented inTables 1 and 2.

Transcriptional profiling of different histologies. We used thecomplete array expression values to build a dendrogram(Fig. 1), reflecting global expression similarities across thesamples. Hierarchical clustering of all samples showed a cleardistinction between cancerous and normal samples. Oncocy-tomas and chromophobe RCC grouped together in asubcluster with normal kidney samples. All other malignanttumors clustered in a separate branch in concordance with theknown histologic type. Whereas all clear cell RCC samplesgrouped together, papillary RCC formed a separate subclusterwith TCC.

We derived tumor type–specific profiles by comparingeach of the five renal tumor types to the ensemble ofremaining tumor types. We built 30-gene predictors for eachsuch comparison and assessed their accuracy using leave-one-out cross-validation (5). At each iteration, a sample was leftout and a 30-gene predictor was generated using theremaining samples to predict the left-out sample with theweighed voting method (5). All five subtype-specific signa-tures yielded 100% prediction accuracy. In Fig. 2 (and

Table 1. Clinical and pathologic characteristics of allanalyzed samples

Histology No. samples

Clear cell RCC 32Papillary RCC 11Chromophobe RCC 6TCC 8Oncocytoma 12Normal tissue 23

NOTE: Representation of various renal tumor types.

RCCGene Signatures for Progression andMetastasis

www.aacrjournals.org Clin Cancer Res 2005;11(16) August15, 20055731



Supplementary Data), we show 30-gene signatures created foreach tumor type using all samples (5), which were furtherused for external validation (see below).

Transcriptional profiling of low-stage clear cell renal cellcarcinoma tumor samples. Using dChip, we compared all(eight) clear cell T1 samples with the 23 normal kidney samplesand obtained a list of 1,359 genes that are significantly up-regulated in T1 and 493 genes that are significantly down-regulated in T1.

Among the differentially expressed genes between normalkidney and T1 clear cell RCC, we identified a subset of up-regulated genes coding for secreted proteins (Fig. 3A) and cellsurface receptors (Fig. 3B), two very important classes of targetsfor drug discovery and diagnostics. Many of these genes codingfor secreted proteins are involved in various aspects ofangiogenesis (e.g., vascular endothelial growth factor , angiopoie-tin-like 4 , angiopoietin-like 2 , basic fibroblast growth factor ,placental growth factor , endothelial cell growth factor 1 , endothelin1 , transforming growth factor-b1, transforming growth factor-b2 ,adrenomedullin , and ceruloplasmin). Various migration, inva-sion, and adhesion-related genes coding for secretory proteins(such as parathyroid hormone-like hormone , matrix metalloprotei-nase 14, and gonadotropin releasing hormone 2), chemokines(such as CCL5 , CCL20, and CCL18), chemokine receptors(CX3CR1 , CXCR4 , and CXCR6), or tyrosine kinase receptors(e.g., FLT1 , AXL , and epidermal growth factor receptor) were alsoup-regulated in the early-stage tumors (see Supplementary Datafor complete lists).

Analysis of clear cell renal cell carcinoma progression andmetastasis. To identify genes whose expression is progres-sively dysregulated during stepwise progression from normalkidney tissue (N) through early tumor stage (T1) to distantmetastasis (M), we compared the corresponding clear cell RCCsamples using dChip: N versus T1 and T1 versus M (Fig. 4). Weidentified 31 genes that are significantly differentiallyexpressed in the same direction in both comparisons, thusshowing a continuous deregulation pattern as the disease

progresses. In Fig. 4, we present the colorgram for the31 genes: 18 genes that are up-regulated in T1 compared withN and further up-regulated in M compared with T1 and 13genes that are down-regulated in T1 compared with N andfurther down-regulated in M compared with T1. Examples ofgenes continuously deregulated on progression includecaveolin 1 , annexin A4 , and lysyl oxidase (see SupplementaryData). Progressively dysregulated genes may play a rolein driving tumors toward increased malignancy. Althoughexpression of these genes may be increased or decreased inmetastases, they are not metastasis specific.

The use of transcriptional profiling to build RCC-specificoutcome signatures has been proposed previously (9). Ournext goal was to find a clear cell RCC metastasis-specific genepattern by comparing all (eight) clear cell T1 samples thathave until now not developed any metastases (median,32 months; range, 28-40 months) with all (nine) samplescollected from distant organ metastases. Using Genes@Work,we identified a 155-gene pattern that predicted metastasis inthis training set with 100% accuracy (Fisher’s P = 0.00004,permutation P < 0.001).

Our next goal was to determine whether this patternpreexists in more advanced primary clear cell RCC tumorsthat develop metastases. For this purpose, we tested the 155-gene metastasis signature on the nine primary clear cell RCCtumors, which had or later developed visceral metastases. Ofthese nine samples, which were not used in developing thesignature, five tumors were predicted as ‘‘ T1’’ and four as‘‘metastatic.’’ Four of the five tumors predicted as T1 did nothave organ metastases at the time of surgery. All four tumorsclassified as metastatic had organ metastases at the time ofsurgery. Thus, this gene signature predicted the presence oforgan metastases at the time of surgery with 88.9% accuracy inthis independent group of tumors (see Supplementary Data),indicating differences in the biology between RCC tumors thatpresent with distant metastases at the time of surgery and RCCtumors that only later develop metastases. We also didprincipal components analysis where samples classified as T1

and metastatic could be divided into two distinct areas withinthe three-dimensional space (see Supplementary Data).This was overall in concordance with our classification resultsas only the samples in the independent data set that hadmetastasis at the time of surgery fell in the same space as themetastatic samples.

External validation of gene signatures. To validate our genesignatures and gene signatures generated by other investigators,we tested several independent external data sets across differentmicroarray platforms in four facets of analysis: validating ourtumor type–specific gene signatures, comparing our clear cellRCC–associated gene lists with the lists available in theliterature, testing our metastasis gene signature, and applicationof an external global metastatic signature to our samples.

Validation of tumor type–specific gene signatures. Higginset al. (17) hybridized 23 clear cell RCC, 4 papillary RCC,3 chromophobe RCC, 2 oncocytoma, 1 angiomyolipoma,5 granular cell carcinomas, and 3 normal samples to a22,648-element cDNA microarray representing 17,083 differ-ent UniGene clusters. Because of the differences in gene contentacross the two microarray platforms, we were able to match64 of our 120 tumor type–specific genes for clear cell RCC,papillary RCC, oncocytoma, and chromophobe RCC, 17, 14,

Table 2. Clear cell RCC samples

Median age (range), y 61.5 (35-79)Gender (male/female) 22/10Stage

T1 8T2 5T3 8T4 1

Metastatic tissue samples 10 (9 distant/1nodal)Grade

1 4 (T1),1 (T2),1 (T3a)2 4 (T1), 3 (T2), 3 (T3a), 3(T3b)3 1 (T3a)4 1 (T2),1 (T4)

Nodal status, positive/negative/unknown

0/12/10

Metastatic status,positive/negative

6/16

Follow-up (T1patients),median (range), mo

32 (28-40)





15, and 18 genes, respectively, on the Higgins et al. data setusing EnsEMBL (version Hs 25.34e) and UniGene (Build Hs177) databases (see Supplementary Data). Based on these 64matching genes, we used dChip to hierarchically cluster thesamples in the Higgins et al. data set (Fig. 5A). The resulting treevalidated our gene signatures separating tumor types success-fully where the clusters with clear cell RCC, papillary RCC,chromophobe RCC, granular cell carcinoma, and normalsamples all show significant associations (P < 0.05) ascalculated by dChip (see Supplementary Data). Similar to theresults obtained by our data set, clear cell RCC samples wereseparated from normal, oncocytoma, chromophobe RCC, andpapillary RCC. Oncocytomas grouped together with chromo-phobe RCC and were closer to normal than papillary RCC orclear cell RCC. We then split the Higgins et al. data set into twogroups of 22, each containing 13 clear cell RCC samples. Usingthe 17 genes that matched from our clear cell RCC gene sig-nature, we built a model on the first group of 22 (training set)and applied it on the second group of 22 (test set). The predic-tion accuracy on the test set was 91% (20 of 22 predictedcorrectly) with a Fisher’s P of 2.4 � 10�4 (see SupplementaryData). These data most vividly show that our tumor type–specific gene signatures hold up in an independent external vali-dation set across two completely different microarray platforms.

Clear cell renal cell carcinoma–associated gene list. Lenburget al. compared 8 normal and 9 clear cell RCC samples usingan unpaired t test (3-fold change and P < 0.03), and on theHG-U133A chip, they found 357 probes up-regulated in clearcell RCC versus normal and 669 probes down-regulated inclear cell RCC versus normal. When compared with our T1

clear cell RCC–associated gene lists of 1,359 up-regulated and493 down-regulated genes, the intersection was 187 and 107,respectively, both of which showed a significant overlap (P <10�10), where the probability of observing such an overlap bychance is calculated using hypergeometric distribution (18,19). In Fig. 5B, we show the overlap between the two studiesand the intersecting gene lists are shown in SupplementaryData. These data show that there is a significant overlapbetween the gene lists generated in two independent micro-array analyses, supporting the relevance of this gene set forclear cell RCC; however, there are also significant differencesthat can be due to a combination of several factors, includingthe difference in algorithms used for deriving the signal values(we used the dChip algorithm and Lenburg et al. applied theMAS5 algorithm) and the difference in sample composition(we used only T1 stage clear cell RCC versus normal kidney toobtain differentially expressed genes and Lenburg et al. usedclear cell RCC samples with undefined tumor stages).

Validation of the metastasis gene signature. Sultmann et al.hybridized 19 primary RCC samples that had metastasis at thetime of diagnosis (M1) and 17 that did not (M0) on customcDNA microarrays containing 4,207 genes and expressedsequence tags (20). We matched 41 of our 155 metastasis-associated genes across the two different microarray platforms

Fig. 1. Dendrogram of all samples. Hierarchical clustering of all profiled samplesusing average linkage method.The complete array expression values were used tobuild the dendrogram where Pearson’s correlation is used as the metric of similarity.Normal, normal kidney; ONC, oncocytomas; ChrRCC, chromophobe RCC; pRCC,papillary RCC; cRCC, clear cell RCC;T, primary tumor stage; B, Br, L, and N,metastatic tissue (bone, brain, lung, and nodal).





in the Sultmann et al. data set using EnsEMBL (version Hs25.34e) and UniGene (Build Hs 177; see Supplementary Data).Using dChip, we clustered these 36 samples (Fig. 5C) based onthe 41 matched genes. The two groups, M0 and M1, wereseparated into two significantly distinct clusters (P < 0.05 foreach branch as calculated by dChip), independently validatingour notion about the presence of the metastasis signature inprimary clear cell RCC tumors with metastasis at the time ofdiagnosis but not in tumors that did not have metastasis at thetime of surgery. Again, as expected due to the difference inmicroarray platforms and the resulting small overlap of genesacross the two platforms (41 of 155), separation of the twophenotypes was not perfect, but despite these limitations, ourmetastasis gene signature resulted with statistical significance inthis independent data set.

External global metastasis gene signature. Using AffymetrixHG-U95A arrays, Ramaswamy et al. compared 12 metastaticand 64 primary adenocarcinomas from several tumors (lung,breast, prostate, colorectal, uterus, and ovary) to find a 128-gene metastatic signature that correlates with poor clinicaloutcome (21). Because this global metastatic signature has notbeen evaluated on RCC, we decided to determine its applica-bility for RCC.

We used 169 probes (corresponding to those 128 predictorgenes) from the HG-U95A array (that was used in the previouswork) and mapped them to the HG-U133A array usingAffymetrix ‘‘ best matched’’ criteria for same-species comparisonguideline. The resulting 123 analogues were used to build anunsupervised hierarchical tree with our T1 and distantmetastatic clear cell RCC samples as well as our >T1 sampleswith or without metastases at the time of surgery. The T1

samples have until now not developed metastases. Theresulting tree showed that the majority of distant metastasesand the tumors with metastasis at the time of surgery correctlyclustered in their corresponding known classes in a separatebranch from the majority of T1 samples and the T2 and T3 RCCsamples that only later developed metastases (see Fig. 5D).These findings suggest that the previously reported (21) globalmetastatic gene signature can successfully distinguish distantmetastasis from tumors that do not develop metastases in RCCand can further distinguish between RCC with metastasis at thetime of surgery and tumors that later developed metastases.Due to the lack of sufficient follow-up, no conclusion aboutclinical outcome can be drawn at the moment. These data alsoconfirm the results obtained with our metastasis signature thatseparated RCC with metastasis at the time of diagnosis fromtumors that later developed metastases and further support ourhypothesis that these tumors are biologically distinct.

Discussion

Despite great refinements in surgical technique, perioper-ative care, and increased knowledge of immunobiology, thereare major limitations in managing RCC. This becomesparticularly obvious when trying to identify prognosticmarkers of outcome or when treating patients with metastaticdisease. Currently available clinical prognostic markers (suchas stage or performance status) are useful, but substantialoutcome variability exists even within patients stratified forthese prognostic factors. Molecular markers of RCC have alsobeen proposed (22), but it is questionable whether they

Fig. 2. Color map of the signal values for the predictor sets of genes in differentrenal tumors.The color bar above the color map represents the renal tumors. Genesymbols are shown next to the color map. For each gene, red and green indicate highand low expression, respectively.





adequately address the heterogeneity of this disease whenstudied in isolation. Current knowledge of RCC mostly derivesfrom the analysis of the most common subtype (clear cellRCC) and fails to address f30% of the remaining subtypes,as the sample size of the less common subtypes is often thelimiting factor for statistically robust statements in moststudies. Furthermore, even within a subtype, there can bemore and less aggressive molecular variants leading todifferent clinical outcomes as proposed recently for clear cellRCC (9). Our aim was to detect these molecular differences invarious RCC subtypes and during tumor progression. Weexpect that our findings will be useful in individualizingdiagnosis and therapy in the future and in decipheringunderlying biological differences.

In our study, we analyzed global gene expression of the threemost common RCC subtypes (clear cell, papillary, andchromophobe RCC), TCC, and oncocytoma in comparisonwith normal kidney tissue. TCC and oncocytoma are importantin the differential diagnosis of RCC and, at initial radiologicalpresentation, often indistinguishable from it. Whereas TCC is ahighly malignant tumor arising from the urothelium of the renalpelvis, oncocytomas are benign tumors originating from theintercalating cells of the collecting duct (23). Hierarchicalclustering showed a perfect distinction of RCC, TCC, oncocy-toma, and normal kidney tissue. Whereas the clear cell andpapillary RCC subtypes clustered together with TCC, chromo-phobe RCC grouped with normal kidney tissue and oncocy-toma. This finding is in concordance with the more benignnature of chromophobe RCC, which has been shown recently tohave the highest 5-year disease-specific survival among all RCCsubtypes (24). Furthermore, it potentially reflects the commonsite of origin (collecting duct) for oncocytoma and chromo-phobe RCC (25). In contrast, clear cell RCC and papillary RCC

did not group together despite their common origin in theproximal tubule of the renal cortex. Their reported difference inaggressiveness and outcome may be reflected in this finding(24, 26). Given that RCC and TCC are distinct tumor entities,the clustering of papillary RCC and TCC was surprising andunexpected. Examination of a previous microarray publicationwith a smaller number of papillary RCC and TCC reveals asimilar clustering of papillary RCC with TCC but not with clearcell RCC (10). This interesting finding needs to be furtherconfirmed in larger, more specifically designed studies.

Our novel analysis extends previous microarray studiesaddressing the different histologic subtypes of RCC. Thesestudies reported genes differentially expressed between differentsubtypes but came short of reporting the accuracy of theirspecific gene signatures (10, 17). In our study, we providesubtype-specific signatures with 100% accuracy. We alsovalidated our subtype gene signatures across two differentmicroarray platforms on an independent external data set byHiggins et al. (17). This is one of the few times that this cross-platform validation has been done. Nevertheless, due to thedifference between the microarray platforms used, only a subsetof the genes overlapped between the two data sets. Despite thislimitation of gene overlaps and the difference in the microarrayplatform itself, our gene signature did extremely well andseparated all tumor samples of the external data accurately,further validating our analysis and demonstrating the cross-platform and cross-laboratory performance. Although thetumor specificity of our gene lists might serve for differentialdiagnoses in selected cases, we believe these sets of genesmainly have the potential of elucidating unique biologicalpathways inherent to the different renal tumors that can beexplored for discovery of novel molecular drug targets andprognostic biomarkers.

Fig. 3. A , color map of genes coding forsecreted proteins showing the level ofdifferential expression in the renal tumortypes compared with normal kidney tissue.Up-regulated genes coding for secretedproteins were identified from thecomparison ofT1clear cell RCC and normalkidney tissue according to their LCB. Inaddition, the LCB values for these genesfrom metastatic clear cell RCC, clear cellRCC, papillary RCC,TCC, and oncocytomasamples (when compared with normalkidney tissue) are also assessed in the colormap. Red and green indicate higher andlower LCB values normalized within the[�2.5,2.5] interval as indicated in the colorscale (bottom).The function of the genes isdisplayed next to the color map.





The focus of our further analysis was directed towardmolecular characterization of clear cell RCC tumorigenesis andprogression. We first analyzed the transcriptional profiles withrespect to early and late phenotypic stages of clear cell RCC andsubsequently defined a characteristic set of genes for the T1 stage.

Among the differentially expressed genes in T1 clear cellRCC tumors, we identified a subset of genes coding forsecreted proteins and cell membrane receptors that may beused in the future as biomarkers or molecular targets for drugdevelopment. Many of these genes coding for secretedproteins are involved in various aspects of angiogenesis, suchas vascular endothelial growth factor , angiopoietin-like 2 , angio-poietin-like 4 , basic fibroblast growth factor , placental growthfactor , endothelial cell growth factor 1 , transforming growth factor-b1 , transforming growth factor-b2 , adrenomedullin , and cerulo-plasmin , and have been associated with RCC (27–32). Thismay reflect the high vascularity of clear cell RCC, possibly dueto at least partially mutated von Hippel-Lindau, and is in

contrast to the other subtypes of RCC, oncocytoma, and TCCthat only show sporadic up-regulation of angiogenic factors.Among up-regulated receptors, tyrosine kinase receptors (suchas FLT1 or epidermal growth factor receptor), chemokinereceptors (e.g., CXCR4), and apparent autocrine loops (suchas transforming growth factor-a/epidermal growth factor receptor)are particularly interesting, because they all have beenimplicated in various aspects of development or progressionof RCC or other types of cancer (27, 33–38). Thus, theangiogenic phenotype seems to be already expressed at earlystages of clear cell RCC and is likely to play an important rolein tumor development and progression. Further elucidation ofthe precise identity and role of the various genes in thesenetworks may allow development of new therapeutic modelsfor various clear cell RCC tumor stages. Our study represents afirst step toward this direction by providing evidence for aconcerted up-regulation of several angiogenesis and growthfactor pathway genes in clear cell RCC, suggesting thepotential therapeutic utility of angiogenesis inhibitors, tyro-sine kinase inhibitors, and growth factor or receptor antago-nists. In addition, novel approaches with small-moleculeantagonists of a chemokine receptor, such as CXCR4 , whichhave been successfully applied in animal models for othertumors, carry the potential of being effective in RCC (39).

We further identified within clear cell RCC a subset of genesprogressively dysregulated from normal kidney tissue throughT1 stage to metastasis. These progressively dysregulated genesdo not represent metastasis-specific genes, because they arealready up-regulated or down-regulated at the T1 stagecompared with normal kidney but are of significant biologicalinterest with regard to tumor progression and may be involvedin driving the progression toward increased malignancy. Someof these genes (e.g., caveolin 1 , lysyl oxidase , and annexin A4)have already been associated with RCC and aggressive tumorbehavior (40–42). Genes continuously dysregulated through-out RCC progression might be interesting diagnostically andtherapeutically alike.

Metastatic RCC has a particularly poor prognosis with onlyf9% 5-year survival and available treatment options areassociated with response rates of 15% to 20% with a medianresponse duration of 54 months (43). Our goal was to identify agene pattern that best distinguishes metastatic samples and useit to explore the metastatic potential of primary tumors. Thisnovel signature distinguished primary tumors very accurately intwo groups based on the presence of metastases at the time ofsurgery, in two independent validation sets, one of which fromanother study done on a different microarray platform. Thisissue had not been addressed previously by other importantreports focusing on advanced RCC (9, 44). Additionally, weevaluated, to our knowledge for the first time in RCC, apreviously reported global metastatic signature that had beenderived from various non-RCC solid tumors (21). An unsuper-vised classification of our clear cell RCC samples using thisglobal gene signature correctly separated the majority of distantmetastases from T1 RCC samples, further suggesting that RCCmay share a common mechanism of metastasis with differentcancer types. Similar to our metastasis signature, the globalmetastasis signature classifies most of the primary tumorsamples with metastasis at the time of surgery together withdistant metastases and RCC tumors that only later developedmetastasis together with the T1 tumors, indicating that there is

Fig. 3. (Cont’d) B, color map of genes coding for membrane receptor proteinsshowing the level of differential expression in the renal tumor types compared withnormal kidney tissue. Up-regulated genes coding for membrane receptor proteinswere identified from the comparison ofT1clear cell RCC and normal kidney tissueaccording to their LCB. In addition, the LCB values for these genes from metastaticclear cell RCC, clear cell RCC, papillary RCC,TCC, and oncocytoma samples (whencompared with normal kidney tissue) are also assessed in the color map. Red andgreen indicate higher and lower LCB values normalized within the [�2.5,2.5]interval as indicated in the color scale (bottom).The function of the genes isdisplayed next to the color map.





an intrinsic biological difference between tumors that havedeveloped distant metastasis at the time of surgery and tumorsthat apparently are less aggressive and do not present withmetastases at the time of surgery. These findings suggest thatthere is an inherent gene signature in the primary tumor thatmay be associated with a highly aggressive course. Moreover, itis in partial agreement with the hypothesis generated byRamaswamy et al., suggesting that the metastatic potential ofhuman tumors is encoded in the bulk of the primary tumor(21). However, at least in RCC, our metastatic signature and theglobal metastatic signature by Ramaswamy et al. suggest that asubset of more aggressive RCC tumors (metastasis at the time ofsurgery) indeed encodes the metastatic potential in the bulk ofthe primary tumor, whereas another subset of less aggressivetumors (metastasis developed only later) does not encode themetastatic signature in the primary tumor. From a clinicalperspective, the ability to identify patients at high risk for distantmetastasis may lead to more extensive imaging evaluation ofthese patients at diagnosis and subsequently to initiation offurther therapy. According to some series, a third of patientswith bone metastasis did not have musculoskeletal pain atinitial diagnosis and would have been missed, whereas f50%would have been missed because of normal levels of serumalkaline phosphatase had the indication for a bone scintigraphy

relied on symptoms or elevated serum alkaline phosphatase,respectively, as often suggested (45–47). In addition, potentialintrinsic biological differences between more and less aggressiveRCC tumors may suggest different therapeutic approaches withdifferent molecular targets.

The feasibility of applying microarray data generated on adifferent platform and by a different research group has thepotential to further facilitate translational research by directlyimplementing existing microarray results into new studies. Wehave used four different external data sets using various differentmicroarray platforms to either validate our data or validateexternal data and we have shown that, despite all the limitations,validation on independent external data sets is feasible.

Our clear cell RCC metastasis gene signature contains severalgenes that have been associated with cancer progression,metastasis, or poor outcome in RCC or other cancers (matrixmetalloproteinase 14, topoisomerase IIa, glucosyl ceramide synthase ,and ADAM 20; refs. 48–51). Particularly interesting is the up-regulation of topoisomerase IIa in conjunction with glucosylcer-amide synthase . Topoisomerase IIa has been known to beexpressed in higher-grade RCC (49). Although topoisomeraseIIa inhibitors (such as doxorubicin and etoposide) are effectivein a variety of tumors, their efficacy in RCC is very limited(52–54). Interestingly, the glucosylceramide synthase enzyme,

Fig. 4. Color map of the 31genes that aresignificantly differentially expressed in NversusT1andT1versus M. Expression profilesof 18 genes are positively correlated withdisease progression, whereas expressionprofiles of13 genes are negatively correlatedwith disease progression. Genes arerepresented in rows; samples arerepresented in columns where the title for asample group indicates the beginning of thatparticular group of samples. Red and greenindicate high and low expression,respectively.





which has been implicated in tumor cell resistance to thetopoisomerase inhibitor doxorubicin, is also among the up-regulated genes suggesting a mechanism how clear cell RCCtumors could escape cell death in response to topoisomeraseinhibitors. OGT2378, an imino sugar inhibitor of glucosylcer-amide synthesis, was shown recently to inhibit tumor growth ofmelanoma cells (55). Blocking glucosylceramide synthase mayalso overcome resistance to topoisomerase inhibitors and pavethe way to effective combination therapy. Other less well knowngenes within the metastatic signature, which have not beenassociated with RCC in the past, such as the transcription factorPOU6F1, may deserve evaluation in future studies (56). Wehave used several pathway analysis software tools to integratethe metastatic genes into biological pathways, revealing acomplex interaction between many of the metastatic genesand several nodal points where multiple metastasis genes act on(see Supplementary Data). In particular, the three tumorsuppressor and oncogene pathways involving p53, c-myc , andc-fos as nodal points seem to be targeted by various metastaticgene products. Multiple signal transduction pathways, such as

the mitogen-activated protein kinase, nuclear factor-nB, andphosphatidylinositol 3-kinase/AKT pathways, cell survival,motility, and proliferation seem to be modulated by metastasisgenes.

Historically, RCC was considered a single entity with varioushistologic appearances. Today, RCC is more accurately recog-nized as a family of cancers resulting from distinct geneticabnormalities with unique morphologic features but a commonderivation from the renal tubular epithelium (57). We andothers (8–10) have shown that transcriptional profiling is apowerful tool for better appreciation of the subtle phenotypicdifferences among various renal tumors in general and RCC inparticular. Data generated in these studies, if prospectivelyvalidated, should aid in the development of improved treatmentapproaches for RCC.

Acknowledgments

We thank Vikas Sukhatme and Ananth Karumanchi for comments on the articleand Christa Blumenberg for kind help and assistance.

Fig. 5. Use of external data sets. A , hierarchical clustering of the samples in the Higgins et al. data using our subtype-specific genes for clear cell RCC, oncocytoma,chromophobe RCC, and papillary RCC.The clustering is obtained using the 64 (of120) genes that have matched between the two platforms.The signal values are normalized tothe [�1,1] interval, where red denotes high expression and green denotes low expression as shown by the color scale (bottom). In the horizontal bar, samples are denoted asfollows: c, clear cell; p, papillary; h, chromophobe; O, oncocytoma; a, angiomyolipoma; g, granular; N, normal.B, overlap between the clear cell RCC ^ specific genes identifiedin the two studies. C, hierarchical clustering of M0 and M1samples in the Sultmann et al. data using our metastasis signature.The clustering is obtained using the 41genes(of155) that matched between the two platforms.The signal values are normalized to the [�1,1] interval, where red denotes high expression and green denotes low expressionas shown by the color scale (bottom).D, clustering ofour clear cell RCC samples using the global metastasis signature identified by Ramaswamy et al.The clustering is obtainedusing the 123 (of128) genes that have matched between the two platforms. Red, distant metastasis; pink, primary tumors that have metastasis at the time of surgery; blue,primary tumors that do not have metastasis at the time of surgery but develop later; green,T1.





References1. Levy DA, Slaton JW, Swanson DA, Dinney CP. Stage

specific guidelines for surveillance after radical ne-phrectomy for local renal cell carcinoma. J Urol 1998;159:1163 ^ 7.

2. Uchida K, Miyao N, Masumori N, et al. Recurrence ofrenal cell carcinoma more than 5 years after nephrec-tomy. Int J Urol 2002;9:19 ^ 23.

3. Fisher RI, Rosenberg SA, Fyfe G. Long-term survivalupdate for high-dose recombinant interleukin-2 inpatients with renal cell carcinoma. Cancer J Sci Am2000;6 Suppl 1:S55 ^ 7.

4. Flanigan RC, Salmon SE, Blumenstein BA, et al.Nephrectomy followed by interferon alfa-2b com-pared with interferon alfa-2b alone for metastatic re-nal-cell cancer. N Engl J Med 2001;345:1655 ^ 9.

5.Golub TR, Slonim DK,Tamayo P, et al. Molecular clas-sification of cancer: class discovery and class predic-tion by gene expression monitoring. Science 1999;286:531 ^ 7.

6. Alizadeh AA, Eisen MB, Davis RE, et al. Distincttypes of diffuse large B-cell lymphoma identified bygene expression profiling. Nature 2000;403:503 ^ 11.

7. van de Vijver MJ, HeYD, van’t Veer LJ, et al. A gene-expression signature as a predictor of survival in breastcancer. N Engl JMed 2002;347:1999 ^ 2009.

8. BoerJM, Huber WK, Sultmann H, et al. Identificationand classification of differentially expressed genes inrenal cell carcinoma by expression profiling on a globalhuman 31,500-element cDNA array. Genome Res2001;11:1861 ^ 70.

9. Takahashi M, Rhodes DR, Furge KA, et al. Gene ex-pression profiling of clear cell renal cell carcinoma:gene identification and prognostic classification. ProcNatl Acad Sci U S A 2001;98:9754 ^9.

10. Takahashi M,Yang XJ, Sugimura J, et al. Molecularsubclassification of kidney tumors and the discoveryof new diagnostic markers. Oncogene 2003;22:6810 ^ 8.

11. Sobin L,Wittekind C, Ackerley W. UICC TNM classi-fication of malignant tumors. 6th ed. NewYork: Wileyand Sons; 2002.

12.Mitsiades N, Mitsiades CS, PoulakiV, et al. Molecularsequelae of proteasome inhibition in human multiplemyeloma cells. Proc Natl Acad Sci U S A 2002;99:14374 ^ 9.

13. Li C, Wong WH. Model-based analysis of oligonu-cleotide arrays: expression index computation andoutlier detection. Proc Natl Acad Sci U S A 2001;98:31 ^ 6.

14. Sneath PHA, Sokal NN. Numerical taxonomy; theprinciples and practice of numerical classification.San Francisco (CA): W.H. Freeman; 1973.

15. Landgrebe J, Wurst W, Welzl G. Permutation-validated principal components analysis of micro-array data. Genome Biol 2002;3:RESEARCH0019.

16. Lepre J, Rice JJ,TuY, Stolovitzky G. Genes@Work:an efficient algorithm for pattern discovery and multi-variate feature selection in gene expression data. Bio-informatics 2004;20:1033 ^ 44.

17. Higgins JP, Shinghal R, Gill H, et al. Gene expressionpatterns in renal cell carcinoma assessed by comple-mentary DNA microarray. Am J Pathol 2003;162:925 ^ 32.

18. Lenburg ME, Liou LS, Gerry NP, et al. Previously un-identified changes in renal cell carcinoma gene expres-sion identified by parametric analysis of microarraydata. BMC Cancer 2003;3:31.

19. Ivanova NB, Dimos JT, Schaniel C, et al. A stem cellmolecular signature. Science 2002;298:601 ^ 4.

20. Sultmann H, von Heydebreck A, Huber W, et al.Gene expression in kidney cancer is associatedwith cytogenetic abnormalities, metastasis forma-

tion, and patient survival. Clin Cancer Res 2005;11:646 ^ 55.

21. Ramaswamy S, Ross KN, Lander ES, Golub TR. Amolecular signature of metastasis in primary solidtumors. Nat Genet 2003;33:49 ^ 54.

22. Bui MH, Zisman A, Pantuck AJ, et al. Prognosticfactors and molecular markers for renal cell carcinoma.Expert RevAnticancerTher 2001;1:565 ^ 75.

23. Bonsib SM, Bray C, Timmerman TG. Renal chro-mophobe cell carcinoma: limitations of paraffin-embedded tissue. Ultrastruct Pathol1993;17:529^36.

24. Amin MB,Tamboli P, Javidan J, et al. Prognostic im-pact of histologic subtyping of adult renal epithelialneoplasms: an experience of 405 cases. Am J SurgPathol 2002;26:281 ^91.

25. Ortmann M, Vierbuchen M, Fischer R. Sialylatedglycoconjugates in chromophobe cell renal carcinomacompared with other renal cell tumors. Indication of itsdevelopment from the collecting duct epithelium.Virchows Arch B Cell Pathol Incl Mol Pathol 1991;61:123 ^ 32.

26.Delahunt B, EbleJN. Papillary renal cell carcinoma: aclinicopathologic and immunohistochemical study of105 tumors. Mod Pathol 1997;10:537^ 44.

27. Fujita Y, Mimata H, Nasu N, et al. Involvement ofadrenomedullin induced by hypoxia in angiogenesisin human renal cell carcinoma. Int J Urol 2002;9:285 ^ 95.

28. Pejovic M, Djordjevic V, Ignjatovic I, Stamenic T,Stefanovic V. Serum levels of some acute phase pro-teins in kidney and urinary tract urothelial cancers.Int Urol Nephrol 1997;29:427 ^ 32.

29. LeJan S, Amy C, Cazes A, et al. Angiopoietin-like 4is a proangiogenic factor produced during ischemiaand in conventional renal cell carcinoma. Am J Pathol2003;162:1521 ^ 8.

30. SlatonJW, Inoue K, Perrotte P, et al. Expression lev-els of genes that regulate metastasis and angiogenesiscorrelate with advanced pathological stage of renalcell carcinoma. AmJPathol 2001;158:735 ^ 43.

31. MizutaniY,Wada H,Yoshida O, et al.The significanceof thymidine phosphorylase/platelet-derived endo-thelial cell growth factor activity in renal cell carcino-ma. Cancer 2003;98:730 ^ 6.

32.Yagasaki H, Kawata N,TakimotoY, Nemoto N. Histo-pathological analysis of angiogenic factors in renal cellcarcinoma. Int J Urol 2003;10:220 ^ 7.

33. Hu YL, Albanese C, Pestell RG, Jaffe RB. Dualmechanisms for lysophosphatidic acid stimulation ofhuman ovarian carcinoma cells. J Natl Cancer Inst2003;95:733 ^ 40.

34. Sumitomo M, Asano T, Asakuma J, Horiguchi A,Hayakawa M. ZD1839 modulates paclitaxel responsein renal cancer by blocking paclitaxel-induced activa-tion of the epidermal growth factor receptor-extracel-lular signal-regulated kinase pathway. Clin Cancer Res2004;10:794 ^ 801.

35. Sulzbacher I, Birner P,Trieb K, et al. Expression ofplatelet-derived growth factor-AA is associated withtumor progressioninosteosarcoma. Mod Pathol 2003;16:66 ^ 71.

36. Matei D, Chang DD, Jeng MH. Imatinib mesylate(Gleevec) inhibits ovarian cancer cell growth througha mechanism dependent on platelet-derived growthfactor receptor a and Akt inactivation. Clin CancerRes 2004;10:681 ^ 90.

37. Harris AL, Reusch P, Barleon B, et al. Soluble Tie2and Flt1 extracellular domains in serum of patientswith renal cancer and response to antiangiogenictherapy. Clin Cancer Res 2001;7:1992 ^ 7.

38. Staller P, Sulitkova J, Lisztwan J, et al. Chemokinereceptor CXCR4 downregulated by von Hippel-

Lindau tumour suppressor pVHL. Nature 2003;425:307 ^ 11.

39. RubinJB, Kung AL, Klein RS, et al. A small-moleculeantagonist of CXCR4 inhibits intracranial growth ofprimary brain tumors. Proc Natl Acad Sci U S A2003;100:13513 ^ 8.

40. Campbell L, Gumbleton M, Griffiths DF. Caveolin-1overexpression predicts poor disease-free survival ofpatients with clinically confined renal cell carcinoma.BrJ Cancer 2003;89:1909 ^ 13.

41. Stassar MJ, Devitt G, Brosius M, et al. Identificationof human renal cell carcinoma associated genes bysuppression subtractive hybridization. Br J Cancer2001;85:1372 ^ 82.

42. ShiT, Dong F, Liou LS, et al. Differential protein pro-filing in renal-cell carcinoma. Mol Carcinog 2004;40:47 ^ 61.

43. Pantuck AJ, Zeng G, Belldegrun AS, Figlin RA.Pathobiology, prognosis, and targeted therapy for re-nal cell carcinoma: exploiting the hypoxia-inducedpathway. Clin Cancer Res 2003;9:4641 ^ 52.

44.VasselliJR, ShihJH, Iyengar SR, et al. Predicting sur-vival in patients with metastatic kidney cancer bygene-expression profiling in the primary tumor. ProcNatl Acad Sci U S A 2003;100:6958 ^ 63.

45. Staudenherz A, Steiner B, Puig S, Kainberger F,Leitha T. Is there a diagnostic role for bone scanningof patients with a high pretest probability for meta-static renal cell carcinoma? Cancer 1999;85:153 ^5.

46. Koga S,Tsuda S, Nishikido M, et al. The diagnosticvalue of bone scan in patients with renal cell carcino-ma. J Urol 2001;166:2126 ^ 8.

47. Shvarts O, Lam JS, Kim HL, et al. Eastern Coopera-tive Oncology Group performance status predictsbone metastasis in patients presenting with renal cellcarcinoma: implication for preoperative bone scans.J Urol 2004;172:867 ^ 70.

48. Hagemann T, Gunawan B, Schulz M, Fuzesi L,Binder C. mRNA expression of matrix metallopro-teases and their inhibitors differs in subtypes of renalcell carcinomas. Eur J Cancer 2001;37:1839 ^ 46.

49. ShuinT, Sano K. The activity of topoisomerases isrelated to the grade and stage in human renal cell car-cinoma. Anticancer Res 1994;14:2621 ^ 6.

50. Roemer A, Schwettmann L, Jung M, et al. In-creased mRNA expression of ADAMs in renal cellcarcinoma and their association with clinical out-come. Oncol Rep 2004;11:529^ 36.

51. Liu YY, Han TY, Yu JY, et al. Oligonucleotides block-ing glucosylceramide synthase expression selectivelyreverse drug resistance in cancer cells. J Lipid Res2004;45:933 ^ 40.

52. Droz JP, Theodore C, Ghosn M, et al. Twelve-yearexperience with chemotherapy in adult metastaticrenal cell carcinoma at the Institut Gustave-Roussy.Semin Surg Oncol 1988;4:97 ^ 9.

53. Jekunen A, Pyrhonen S. A combination of vinblas-tine and doxorubicin with interferona. AmJClin Oncol1996;19:384 ^ 5.

54. Skubitz KM. Phase II trial of pegylated-liposomaldoxorubicin (Doxil) in mesothelioma. Cancer Invest2002;20:693^ 9.

55.Weiss M, Hettmer S, Smith P, Ladisch S. Inhibition ofmelanoma tumor growth by a novel inhibitor of gluco-sylceramide synthase. Cancer Res 2003;63:3654 ^ 8.

56. Donahue LM, Reinhart AJ. POU domain genes aredifferentially expressed in the early stages after lineagecommitment of the PNS-derived stem cell line, RT4-AC. Brain Res Dev Brain Res 1998;106:1 ^ 12.

57. Pantuck AJ, Zisman A, Belldegrun AS.The changingnatural history of renal cell carcinoma. J Urol 2001;166:1611 ^ 23.





2005;11:5730-5739. Clin Cancer Res Jon Jones, Hasan Otu, Dimitrios Spentzos, et al. CancerGene Signatures of Progression and Metastasis in Renal Cell

Updated version

http://clincancerres.aacrjournals.org/content/11/16/5730

Access the most recent version of this article at:

Cited articles

http://clincancerres.aacrjournals.org/content/11/16/5730.full#ref-list-1

This article cites 53 articles, 15 of which you can access for free at:

Citing articles

http://clincancerres.aacrjournals.org/content/11/16/5730.full#related-urls

This article has been cited by 26 HighWire-hosted articles. Access the articles at:

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

Subscriptions

Reprints and

[email protected] at

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

Permissions

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

.http://clincancerres.aacrjournals.org/content/11/16/5730To request permission to re-use all or part of this article, use this link



http://clincancerres.aacrjournals.org/content/11/16/5730.full#ref-list-1

http://clincancerres.aacrjournals.org/content/11/16/5730.full#related-urls

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

mailto:[email protected]



genesignaturesofprogressionandmetastasisinrenalcellcancer · university (frankfurt, germany),...

Documents