Originally published as JCO Early Release 10.1200/JCO.2005.03.2375 on January 23 2006

This Article Abstract Full Text (PDF) Supplementary Data Purchase Article View Shopping Cart Alert me when this article is cited Alert me if a correction is posted Services Email this article to a colleague Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Save to my personal folders Download to citation manager Rights & Permissions Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Sanchez-Carbayo, M. Articles by Cordon-Cardo, C. Search for Related Content PubMed PubMed Citation Articles by Sanchez-Carbayo, M. Articles by Cordon-Cardo, C.

Defining Molecular Profiles of Poor Outcome in Patients With Invasive Bladder Cancer Using Oligonucleotide Microarrays

From the Division of Molecular Pathology and Computational Biology Center Memorial Sloan-Kettering Cancer Center, New York, NY; and the Centre de Regulacio Genomica, Barcelona, Spain

Address reprint requests to Marta Sanchez-Carbayo, PhD, Tumor Markers Group, Molecular Pathology Programme, Centro Nacional de Investigaciones Oncológicas, Melchor Fernandez Almagro 3, E-28029 Madrid, Spain; e-mail: marta.sanchez-carbayo{at}cnio.es

	ABSTRACT

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
PURPOSE: Bladder cancer is a common malignancy characterized by a poor clinical outcome when tumors progress into invasive disease. We sought to define genetic signatures characteristic of aggressive clinical behavior in advanced bladder tumors.

METHODS: Oligonucleotide arrays were utilized to analyze the transcript profiles of 105 bladder tumors: 33 superficial, 72 invasive lesions, and 52 normal urothelium. Hierarchical clustering and supervised algorithms were used to classify and stratify bladder tumors on the basis of stage, node metastases, and overall survival. Immunohistochemical analyses on bladder cancer tissue arrays (n = 294 cases) served to validate associations between marker expression, staging and outcome.

RESULTS: Hierarchical clustering classified normal urothelium, superficial, and invasive tumors with 82.2% accuracy, and stratified bladder tumors on the basis of clinical outcome. Predictive algorithms rendered an 89%-correct rate for tumor staging using genes differentially expressed between superficial and invasive tumors. Accuracies of 82% and 90% were obtained for predicting overall survival when considering all patients with bladder cancer or only patients with invasive disease, respectively. A genetic profile consisting of 174 probes was identified in those patients with positive lymph nodes and poor survival. Two independent Global Test runs confirmed the robust association of this profile with lymph node metastases (P = 7.3^–13) and overall survival (P = 1.9^–14) simultaneously. Immunohistochemical analyses on tissue arrays sustained the significant association of synuclein with tumor staging and clinical outcome (P = .002).

CONCLUSION: Gene profiling provides a genomic-based classification scheme of diagnostic and prognostic utility for stratifying advanced bladder cancer. Identification of this poor outcome profile could assist in selecting patients who may benefit from more aggressive therapeutic intervention.

	INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
Bladder cancer is a common malignancy characterized by frequent recurrence and a poor clinical outcome when tumors progress into invasive disease.¹ Transitional cell carcinoma of the bladder constitutes a spectrum of diseases that have been classified into three main groups with distinct clinical behavior and prognosis, management, and reported molecular profiles: superficial (stages Ta-Tis-T1), deeply invasive (stages T2-T4), and metastatic disease (N+/M+).^2,3 Bladder cancer can be described as a genetic disease, driven by the multistep accumulation of genetic and epigenetic factors. These molecular alterations result in uncontrolled cellular proliferation, cell cycle deregulation, decrease in cell death or apoptosis, blockage of differentiation, invasion, and metastatic spread. The particular genetic and protein expression alterations that occur as part of the cross talk between these pathways, will in great part determine the biologic behavior of the tumor, including its ability to grow, recur, progress, and metastasize. The advent of high-throughput methods of molecular analysis can comprehensively survey the genetic profiles characteristic of distinct tumor types and identify targets and pathways that may underlie a particular clinical behavior. Gene expression profiling of tumor cell lines, pools and individual bladder tumor specimens have made progress in their classification, yielding insights into molecular events involved in bladder cancer progression.^4-10 A challenge is how to translate the identification of these targets into potential biomarkers of bladder cancer behavior including molecular diagnosis and outcome prediction. In this study, we focus on characterizing genetic signatures characteristic of invasive bladder tumors, and define clusters, classifiers, and individual targets for patients with advanced disease. Moreover, we identified a genetic signature associated with metastatic potential and poor survival.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
Tumor Samples and RNA Extraction
One hundred fifty-seven frozen tissue samples belonging to 105 patients with bladder cancer were included for this gene expression profiling study. Some of the samples were from the same individual. In these cases, correlations with clinical variables of bladder cancer were performed with only the tumor specimens taken into consideration. Specimens were collected under an institutional review board–approved tissue procurement protocol at Memorial Sloan-Kettering Cancer Center (MSKCC; New York, NY). Tumors were staged and graded according to standard criteria.^2,3 Normal urothelium specimens were obtained at distant sites from the bladder tumors resected by cystectomy or cystoprostatectomy. Bladder tissues embedded in optimal cutting temperature compound were macrodissected to ensure a minimum of 75% of tumor or normal urothelium cells, respectively. Total RNA from bladder tissues was isolated in two steps using TRIzol (Life Technologies, Carlsbad, CA), followed by RNeasy purification. RNA quality was evaluated on the basis of 260:280 ratios of absorbances, and the integrity was also checked by gel electrophoresis analysis using the Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA) as previously reported.¹¹

Gene Expression Analysis
Complementary DNA of the 157 analyzed specimens was synthesized from 1.5 µg of total RNA using a T7-oligo(dT) Promoter Primer Kit (Affymetrix, Santa Clara, CA). RNA target was synthesized by in vitro transcription and labeled with biotynilated nucleotides (Enzo Biochem, Farmingdale, NY). Labeled target was hybridized on Test GeneChips (Affymetrix), to assess the quality of the sample before hybridizing onto the U133A human GeneChips containing 22,283 probes representing known genes and expression sequence tags (Affymetrix). These test GeneChips include multiple probes of the actin and GADPH genes, which are used as controls for the quality of the cRNA before hybridization. The ratio between signal intensities of 3' probe sets and 5' probe sets for these two genes were used as a measure of total RNA integrity as previously described.¹¹

Data Analysis
Scanned image files were visually inspected for artifacts and analyzed using the Affymetrix Microarray Suite 5.0 (MAS 5.0). To compare the expression values given by each of the 157 arrays, a standard scaling process is performed for each specific GeneChip. In this case, expression levels were multiplicatively scaled on the basis of the reference value of 500, as recommended for the U133A GeneChip by the manufacturers. Differential expression was evaluated using the signal as the main response measure, extracted from each gene on every sample, as determined by default settings of the MAS 5.0.

Hierarchical Clustering
To cluster the gene profiling data set of these 157 bladder tissues, hierarchical clustering using the Ward linkage method was combined with nonparametric bootstrap resampling techniques. The hierarchical clustering algorithm was used taking the Pearson correlation (p) as the distance metric [distance = (1 –p) ÷ 2] and the average linkage method.¹² The application of bootstrapping techniques to hierarchical clustering using the R statistical package estimated the robustness of the associations among samples and gene expression patterns.^12,13 Data sets were replicated by resampling and replacing the original expression data over 100 times, followed by generation of a clustering tree for each new data set. A consensus tree was then constructed from the bootstrap trees, showing at each node the number of times each subgroup appeared in the 100 trees. The closer this number was to 100, the more robust that subgroup was. All calculations were performed using the R Statistical Package (http://www.r-project.org).

Gene Ranking and Support Vector Machine Algorithms
Several scoring methods were applied to rank genes that could better characterize critical steps along bladder cancer progression and clinical outcome. No filtering stage was performed before ranking the 22,283 probes. The main groups under consideration were associated with the major clinicopathologic features in bladder cancer, namely tumor staging (superficial v invasive tumors) and clinical outcome (no evidence of disease v dead as a result of disease). Lymph node metastases represent one of the critical steps in bladder cancer progression, mainly when dealing with invasive tumors; thus, we aimed at linking this important feature with survival. This sums up a total of three main comparisons under analyses. Initially, the Welch's t statistic was applied to the log of the signal intensity to identify genes differentially expressed between various pairings of clinical and pathologic groupings.¹⁴ Once top ranked targets associated with the clinical variables under comparison were identified, receiver operating curve analyses and prognostic tests were evaluated on these ranked genes. To deal with the multiple-testing problem, the false-discovery rate (FDR) was used.¹⁵ Because of the varying significances found in the different comparisons analyzed, it was not practical to use a fixed FDR cutoff value for all of them. Thus, we opted to simply provide the top "n genes" (being n = 250 an arbitrary cutoff) along with their Welch's P values (from which the FDR can be computed). A supervised method was used to examine how differentially expressed genes determined via Welch's t statistic were able to predict various clinical and pathologic properties of the samples. To lower the chance of false discovery, we performed a leave-one-out cross validation at the first stage when ranking the genes before carrying out a Support Vector Machine (SVM) algorithm.¹⁶ An SVM with linear kernels was used.¹⁶ The process was iterated using from one to n = 50 of the top ranked genes (as ranked by the Welch's t tests). The idea was to find a reduced (minimal) set of genes that best discriminated the categories under evaluation.

Chromosome and Functional Analyses
Independent studies were designed to obtain insights into chromosome alterations and functional pathways involved in bladder tumor progression, including patient outcome. Allelic imbalance associated with lymph node status and overall survival (time to death) were estimated on the basis of gene expression profiles. For these analyses, quantile normalization was initially performed, and the most differentially expressed probes regarding lymph node and overall survival status were identified on the basis of t-test P values adjusted following the Benjamini-Hochberg approach using the Bioconductor package.^15,17,18 The fold change of each probe was calculated on the basis of their median signal in each category (node and overall survival status). Probes with absolute fold changes higher than 1.5 were linked to their pathways and chromosome category annotations available in Kyoto Encyclopedia of Genes and Genomes (KEGG) indexes,¹⁹ extracted using the Expression Analysis Systematic Explorer (EASE) software.²⁰ The statistical significance of the most represented probes in each KEGG annotation regarding lymph node and survival status was estimated by means of the EASE score.^17,20 This score represents a conservative adjustment to the Fisher's exact probability test that weights significance according to the distribution of probes in each KEGG index annotation.^19,20 EASE scores marginally significant (< 0.1) were also considered to generate chromoplot and pathway pie charts, using the plotting utilities included in the Bioconductor package.¹⁷ This approach compensated for the scarcity of KEGG pathways annotations, and the limitation that nonannotated genes would not be included in this analysis.

Analyses Related to the Definition of the Poor Outcome Profile
Further analyses were designed to identify common probes differentially expressed in invasive patients with positive lymph nodes and poor outcome simultaneously, using a restricted arbitrary threshold of fold change higher than two. An overlapping genetic signature of 174 probes was identified in patients with poor outcome and positive lymph nodes. The subset of 100 probes representing known genes of this 174-probe genetic profile was also plotted in a color-grade spectrum using Array File Maker software.²¹ The Global Test from the Bioconductor package was utilized to test the impact of these 174 probes associated with aggressive behavior in bladder cancer.²² The global test is based on an empirical Bayesian generalized linear model, where the regression coefficients between gene expression data and clinical outcome are random variables. A goodness of fit test is applied on the basis of this model. The global test is optimal when all coefficients show an effect, even if none of the individual coefficients shows a significant relation between expression and the clinical variables under study. The global test method computes a statistic Q and a P value to measure the influence of a group of genes (in this case associated with lymph node metastases) on the clinical outcome variable (overall survival status in this study).²² For each probe, the influence (Q) in predicting clinical outcome is estimated against the expected value, and ranked among the probes under study. The weight of each probe is also assessed by the z-score considering the standard deviation of each probe in the invasive cases under analyses. The Global Test method is designed to identify groups of targets associated with various binary variables such as overall survival (no evidence of disease v dead as a result of disease).²² Log-rank tests were utilized to associate the expression of individual ranked targets with time to death.

The Ingenuity tool (www.ingenuity.com) was also utilized to link the most differentially expressed genes in invasive tumors regarding their lymph node status and overall survival with the reported signaling networks of these genes. Further analyses were also performed to test the gene expression levels of the profile associated with aggressive behavior in paired normal urothelium of the invasive cases under study, as well as in the superficial cases evaluated. This approach was conducted comparing the gene expression levels of each probe versus the median levels observed in patients with and without positive lymph nodes, and overall survival. Among normal urothelium specimens belonging to patients with lymph node metastases and dead as a result of disease, we calculated the number of patients presenting more altered expression of the probes included in the poor outcome profile. For probes overexpressed in this profile, we estimated the number of normal urothelium specimens displaying higher expression levels than the median of the invasive cases with positive nodes and dead as a result of disease. On the contrary, for probes underexpressed in this profile, we estimated the number of normal urothelium specimens displaying lower expression levels than the median in the invasive cases with positive nodes and dead of disease. The specificity of the profile was tested on the superficial cases under study, and the normal urothelium without lymph node metastases and good outcome.

Diagnostic and Prognostic Properties of Transcript Levels of Clusters of Genes and Top-Ranked and Selected Targets
The diagnostic properties of the transcript levels given by gene expression arrays were evaluated based on the distribution of patients within these clusters using standard contingency tables.¹⁴ Histopathologic reports were considered as the gold standard for classification of bladder tumors as superficial (Ta-T1) or invasive (T2+). Receiver operating characteristics (ROC) curve analyses were performed to evaluate their discriminatory properties between superficial and invasive tumors for the top-ranked and the selected target submitted for immunohistochemical validation analyses. The areas under these curves (AUCs) were utilized to compare their diagnostic properties.¹⁴ Cutoffs for targets of interest were selected for further analyses based on providing the highest areas under the curves given by ROC analyses. SEs of these AUCs using the nonparametric assumption, asymptotic significance taken as the null hypothesis that the true area is equal to 0.5, as well as the asymptomatic 95% CIs, were calculated for the probes under analysis. Additionally, the associations of clusters, top ranked and selected markers identified in the DNA microarray analysis to outcome were also evaluated at the transcript level. Overall-survival time was defined as the months elapsed between transurethral resection or cystectomy and death as a result of disease (or the last follow-up date). Patients who were alive at the last follow-up cystoscopy or lost to follow-up were censored. For survival analysis, the classification given by cluster bootstrapping analyses and individual transcript targets were analyzed as categoric variables. Cutoffs were optimized using the ROC analyses mentioned herein. The association of the molecular target transcript levels with overall survival was analyzed using the log-rank test applying these specific cutoffs for each target under evaluation.¹⁴ Survival curves were plotted using the standard Kaplan-Meier methodology.¹⁴ Associations between transcript molecular targets were analyzed using Kendall's ß test. Statistical analyses were performed using the SPSS statistical package, version 8.0 (SPSS Inc, Chicago, IL).

Tissue Samples and Tissue Microarrays
Four different bladder cancer tissue microarrays were used in this study, including 294 primary tumors obtained under an institutional review board–approved protocol. A total of 93 superficial and 201 invasive tumors were analyzed in these tissue microarrays. These tumors corresponded to 34 grade 1, 69 grade 2, and 191 grade 3 lesions. One of these tissue microarrays comprised a cohort of four superficial lesions and 91 invasive tumors with annotated follow-up allowing clinical outcome assessment. Thus, among the 294 tumors spotted onto the tissue arrays, 95 patients had follow-up available for clinical outcome analyses. Only these patients with available follow-up (either "dead as a result of disease" or "alive with no evidence of disease") were considered for overall survival analyses. Cases with unknown follow-up were excluded from these analyses. Known status of p53/pRB and other targets previously described altered along bladder cancer progression, allowed evaluation of the association of synuclein with other targets previously evaluated on these specimens.

Immunohistochemistry
Protein expression patterns of synuclein, a target belonging to the poor outcome profile identified in the DNA microarray analysis, were assessed at the microanatomical level using both cytospins from cancer cell lines (data not shown), and the tissue microarrays. Standard avidin-biotin immunoperoxidase procedures were applied for immunohistochemistry. Synuclein staining was assessed using a mouse monoclonal antibody diluted at 1:500 (clone 42; BD Biosciences, San Diego, CA). The absence of primary antibody was used as negative control. Scores were recorded for cytoplasmic stainings of synuclein, using normal brain as positive control. Ki67 was assessed using a mouse monoclonal antibody diluted at 1:100 (clone MIB-1; DAKO, Glostrup, Denmark).

Statistical Analysis
All cases (n = 296) were used for the analysis of association among synuclein with clinicopathologic variables and the expression patterns of p53 and retinoblastoma protein (pRB) and other molecular targets previously described altered along bladder cancer progression. The consensus value of the representative cores from each tumor sample arrayed was used for statistical analyses. The association of the expression of synuclein with histopathologic stage and tumor grade was evaluated using the nonparametric Wilcoxon-Mann-Whitney and Kruskall-Wallis tests.¹⁴ There is no consensus on the cutoff of the immunohistochemical expression of synuclein, and thus it was analyzed continuously, taking several cutoffs when considered as a categoric variable.

The associations of synuclein with overall survival was also evaluated at the protein level using a subset of 95 cases for which follow-up was available. Overall survival time was defined as the years elapsed between transurethral resection or cystectomy and death as a result of disease (or the last follow-up date). Patients who were alive at the last follow-up or lost to follow-up were censored. The association of synuclein expression levels with overall survival was analyzed using the Wald test, and the log-rank test was used to examine their relationship when different cutoffs were applied.¹⁴ The survival curve was plotted using the standard Kaplan-Meier methodology. Associations among synuclein with p53/pRB and other molecular targets previously described altered along bladder cancer progression were analyzed using Kendall's ß test. Statistical analyses were performed using the SPSS statistical package, version 8.0.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
The molecular classification of bladder tumors provided by transcript profiling was initially analyzed by means of unsupervised hierarchical clustering combined with nonparametric bootstrap analysis. Primary bladder carcinomas were classified on the basis of their histopathologic criteria (Supplementary Table 1, online only), with an overall concordance of 82.2% (Fig 1). The main clusters grouped normal urothelium versus bladder tumors. The bladder cancer cluster identified two groups of invasive patients and a group with patients with superficial lesions. These subclusters differentiated bladder cancer patients on the basis of their clinical outcome (Fig 2). In accordance with the clustering analyses, the difference in survival between superficial and invasive tumors increased for the patients belonging to the invasive² group included in the subcluster at the bottom. A leave-one-out SVM algorithm (see Methods section) was utilized to test the diagnostic and prognostic abilities of gene profiling. The SVM rendered 89% success rate for predicting tumor stage between superficial and invasive tumors, taking any of the top 250 genes at discriminating these categories given by the Welch's t statistic analyses (Supplementary Table 2, online only). The variance of the FDR values for Supplementary Table 2 ranged from 5.8^–12 to 6.3^–8. SVM algorithms were tested to predict overall survival as well. When considering together patients with superficial lesions and those with invasive bladder tumors, the leave-one-out cross validation rendered an overall accuracy of 82.3% with the top 25 genes (Supplementary Table 3, online only). A 10-fold cross validation averaged over 100 trials predicted correctly 72.7% of the cases with the top 250 genes. When only the patients with invasive disease were considered, the leave-one-out validation predicted overall survival correctly on 90% of the cases taking the 100 top genes (Supplementary Table 4, online only). The 10-fold cross validation correctly predicted 74.2% of the cases taking the 100 top genes.

View larger version (30K): [in this window] [in a new window]   Fig 1. Molecular classification of bladder tumors. The upper cluster grouped the majority of normal urothelium (NU), whereas the lower cluster grouped bladder tumors. Two subclusters are observed in the lower cluster: the upper subcluster included invasive tumors (Inv1) and superficial lesions (Sup), and the bottom subcluster grouped invasive tumors (Inv2). Asterisks indicate duplicated/paired specimens.

View larger version (21K): [in this window] [in a new window]   Fig 2. Kaplan-Meier survival curves of bladder cancer patients stratified by bootstrap clusters. (A) Comparing the three subgroups (NS). (B) Comparing the two main subclusters (P = .042). (C) Comparing the Sup and the Inv groups of patients belonging to the same subcluster (NS). (D) Comparing the Sup and the Inv group of patients belonging to different subclusters (P = .007). Inv, invasive; Sup, superficial; NS, not significant.

View larger version (33K): [in this window] [in a new window]   Fig 3. Diagnostic and prognostic properties of top-ranked genes identified by gene profiling. (A) Receiver operating characteristics curves of eight top-ranked known genes at discriminating superficial versus invasive tumors. Kaplan-Meier curves of four top-ranked probes discriminating survival on superficial and invasive bladder tumors for (B) PPIA (peptidylprolyl isomerase A), (C) NXF1 (nuclear RNA export factor 1), (D) TCC9 (tetratricopeptide repeat domain G), (E) HCLS1 (hematopoietic cell–specific Lyn substrate 1), (F) ANK3 (ankyrin G); (G) BIRC3 (baculoviral IAP repeat-containing 3), (H) ICAM1 (intercellular adhesion molecule 1; CD54), and (I) TP53 AP1 (TP53-activated protein 1).

View larger version (41K): [in this window] [in a new window]   Fig 4. Chromosomal and functional annotations of top-genes associated with lymph node metastases and poor survival. (A) Summary of chromosome categories. Amplified and underexpressed loci are indicated with red and green, respectively. Pie chart of over- and under-represented pathways associated with node metastases in (B) most upregulated and (C) most downregulated KEGG pathways. Over- and under-represented pathways associated with survival in (D) most upregulated and (C) most downregulated KEGG pathways. The number of probes appears in parentheses. (*), significant P value.

View larger version (70K): [in this window] [in a new window]   Fig 5. Molecular profile associated with lymph node status and poor survival. Summary of 100 known genes of the genetic signature consisting of 174 overlapping probes differentially expressed in patients with lymph node metastases and poor outcome ranked by their differential fold changes. Overexpressed and underexpressed probes are indicated with red and green, respectively.

As part of the validation studies, one gene coding for a soluble protein was selected from the analyses presented above to evaluate associations with tumor progression and overall survival. This strategy used both the transcript expression levels given by gene profiling and the protein expression patterns obtained by immunohistochemistry on tissue microarrays. Synuclein, the ligand of the cannabinoid or synuclein receptor, was selected as one of the top-ranked genes associated with lymph node invasion and overall survival (Fig 5; Supplementary Table 5). The cannabinoid receptor 1 was also among the top targets (ranked 13th) given by the Welch's t statistic analyses and utilized in the SVM for overall survival, taking only patients with invasive disease obtained through independent analyses (Supplementary Table 4). The variance of the FDR values for Supplementary Table 4 ranged from 0.015 to 0.477. Because further validation would be required to establish the importance of genes with higher FDRs, our target for validation was selected among this list of genes. At the transcript level, not only synuclein and its receptor (CNR1) provided diagnostic information by discriminating superficial versus invasive disease, but also were associated with overall survival (Fig 6A -C). AUCs for synuclein and its receptor, CNR1, were 0.868 and 0.708, respectively. Their asymptotic 95% CIs were 0.800 to 0.936 for synuclein (SNCA) and 0.600 to 0.817 for CNR1, with SEs of 0.035 and 0.056, as well as P values lower than .0005 and .001, respectively. Immunohistochemical analyses on tissue arrays (n = 294) showed that the protein expression levels of synuclein were differentially expressed in superficial versus invasive tumors (Fig 6D-E). A significant correlation between the expression of synuclein with tumor stage and grade was also observed (both < 0.0005). The expression patterns of synuclein were significantly associated with overall survival at the protein level as well (P = .002). Patients displaying a higher expression of synuclein had a shorter survival compared with those with low cytoplasmic expression (Fig 6F). Protein expression of synuclein was significantly associated with pRB inactivation (n = 294; P < .0005). Synuclein was significantly associated with proliferative activity, as assessed by Ki67 (n = 294; P = .007).

View larger version (44K): [in this window] [in a new window]   Fig 6. SNCA is associated with bladder cancer progression and clinical outcome at transcript and protein levels. (A) Receiver operating characteristics curve analysis of CNR1 and SNCA transcript levels. Kaplan-Meier curves of transcript levels of (B) CNR1 and (C) SNCA. Representative immunostaining of synuclein in (D) superficial and (E) invasive bladder tumors. (Magnifications: x400). (F) Kaplan-Meier curve of synuclein protein expression.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

Aiming at defining molecular profiles of poor outcome, further analyses focused on identifying top-ranked genes differentially expressed in invasive cases with lymph node metastases and poor survival. Availability of updated chromosomal and functional annotations of these differentially expressed genes allowed estimation of the allelic imbalances, pathways, and the signaling networks in which these genes participate. In this study, allelic imbalances were estimated on the basis of transcript intensities and chromosomal annotations, analyses frequently performed using chromosomal genomic hybridization approaches.²³ The signaling networks analyses pointed out the relevance of TP53 pathway in bladder cancer progression.^24,25 In line with this argument, one of the top targets identified by the SVM assessing overall survival in patients with invasive tumors (Supplementary Table 4), and also displayed in Figure 2I, was a TP53-activated protein 1.

The SVM algorithm utilized in this study was defined with the purpose of building a predictor using a reduced number of genes. In the original list of 250 genes, the goal was to look for a large set of genes that might play a class distinction role, whereas for prediction purposes we were interested in reducing to a minimal set of genes that will accurately predict the categories under study. We used a method previously described for taking the genes one at time from some ranked list (in our case, the Welch's test list) and seeing how well a given number of genes could predict the correct categories using a leave-one-out validation procedure.^14-16 It turned out that 25 genes did the best in terms of this leave-one-out validation score. In our opinion, none of the supervised learning algorithms described to date should be considered best or perfect. Similar results could have probably been achieved using other available techniques (as has been shown previously¹⁶). We chose to use this specific SVM for three main reasons. First, it has already been successfully used in a number of microarray prediction studies.¹⁶ Second, the generalization properties of SVMs are supposed to work well in the limit of sparse data (ie, few samples with many features [genes]).¹⁶ Finally, we already had experience and the necessary code and tools to work with SVMs. Further development of microarray data analysis tools is required in the field to optimize and standardize the selection process of subsets of genes with the most optimal predictive properties.

One of the main discoveries of the present approach is the identification of an overlapping genetic signature among genes differentially expressed in patients with invasive bladder cancer presenting lymph node metastases and poor survival. Strikingly, the impact of each of these probes as a group at predicting lymph node metastases and overall survival status was revalidated by means of Global Test analyses. Two independent Global Test runs concluded that the same subset of genes was robustly associated with lymph node metastases and overall survival simultaneously. The presence of altered expression of these probes in apparently normal uroepithelial tissues of certain invasive cases with poor outcome is supportive of the controversial concept of multifocal field effect in bladder cancer initiation and progression.²⁶ Moreover, the low penetrance of targets belonging to this profile in paired normal urothelium of invasive patients with good outcome and superficial tumors without evidence of node invasion and progressive disease during the follow-up supported their specificity and the clinical relevance of such profile. Early detection of altered expression of targets associated to this molecular signature may assist in selecting patients who may benefit from more aggressive therapeutic intervention.

The expression patterns of molecular targets identified differentially expressed along bladder cancer progression, such as p53, pRB, and other cell cycle related genes, had already been described as altered in the disease, and also evaluated by immunohistochemistry on the present tissue arrays under study.^5,7 As part of novel validation studies, one gene coding for a soluble protein was selected to evaluate associations with tumor progression and overall survival. This strategy used both the transcript expression levels given by gene profiling and the protein expression patterns obtained by immunohistochemistry on tissue microarrays. Synuclein, the ligand of the synuclein receptor, also known as cannabinoid receptor 1,²⁷ was selected as one of the top-ranked genes associated with lymph node invasion and overall survival, with a 2.60 median higher fold change expression in patients dying as a result of disease as compared with those free of disease (Fig 5; Supplementary Tables 5 and 8). In concordance, the cannabinoid receptor 1 was also among the top targets (ranked 13th) given by the Welch's t statistic analyses and utilized in the SVM for overall survival, taking only patients with invasive disease obtained through independent analyses (Supplementary Table 4). Genes for validation analyses were selected on the basis of combining significant differential expression and highest fold differences. Remarkably, significant associations with tumor staging and clinical outcome were found for these selected targets under evaluation. Although synuclein was detected in patients with bladder cancer using proteomic approaches,²⁸ this is the first report describing the association of synuclein with tumor staging and survival. Synuclein is a component of the hemidesmosomes, which are cell-matrix adhesion structures organized around a core of actin filaments that appears early during cell adhesion. Synuclein colocalizes circularly around F-actin cores together with integrin 3ß1 and other molecules.²⁹ Alteration of actin polymerization and/or remodeling plays pivotal role in regulating the morphologic and phenotypic events of a malignant cell. Gene expression analysis suggests that actin alterations are progressive, and that distinctive actin remodeling profiles are associated with different stages of bladder cancer development and progression.^4-10 These patterns can be used as markers for cancer staging and prognostic indication both at the transcript and protein levels. Overall, gene expression profiling and analysis of protein expression on tissue microarrays rendered complementary information. The detection of specific types of actin-signaling pathway alterations suggests potential alternative targeted therapeutic intervention for patients with bladder cancer.

Large-scale survey transcript profiling of bladder tumors using oligonucleotide microarrays contributes to a biologically oriented classification of bladder cancer. Clusters, classifiers, and individual targets provide novel means for molecular diagnosis and outcome prediction of patients with invasive bladder cancer. Our study differs, in part, from other published studies on application of array technologies in bladder cancer mainly in that it addresses tumor progression issues of advanced bladder cancer.^4-10,30-33 This includes molecular correlates with clinical variables such as lymph node status and overall survival. Lack of information on these two critical features in other published studies^4-10,30-33 precluded identification of appropriate external validation sets. The issue of false discovery as a result of multiple comparisons associated with high-throughput microarray technologies is frequently evaluated by means of external validation. In this regard, we believe that the data analysis methodologies applied in our study, such as leave-one-out cross validation together with SVM algorithms are robust tools to address the false-discovery issue. Furthermore, we provide assessment of the significance of the poor outcome profile by means of the Global Test analyses. In our study, not only was advanced disease discriminated from superficial lesions, but invasive tumors were also stratified on the basis of their lymph node status and patient outcome. As part of the clinical relevance of the present approach, the diagnostic ability of novel top-ranked molecular targets was assessed under standard criteria such as ROC curve analyses, and transcript levels cutoffs associated with overall survival were defined. In addition to target identification, we present an attempt to delineate allelic imbalances, functional molecular pathways, and signaling networks characteristics of patients with a more aggressive clinical behavior, on the basis of those transcripts differentially expressed on patients with poor outcome. Two independent Global Test runs concluded the robust association or a poor outcome profile with lymph node metastases and overall survival simultaneously. The link of lymph node status to overall survival represents a critical relevant step that would be addressed in other solid tumors. Our study supports the concept of multifocal field effect in bladder cancer progression. Gene profiling provides a genomic-based classification scheme of diagnostic and prognostic utility for stratifying advanced bladder cancer. Moreover, the identification of this poor outcome profile could assist in selecting patients who may benefit from more aggressive therapeutic intervention.

	Appendix

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 

View larger version (50K): [in this window] [in a new window]   Fig A1. Chromosomal loci of top-rated genes associated with lymph node status and overall survival. (A) Distribution of the amplified and underexpressed top loci in patients with poor outcome as compared with those with no evidence of disease. (B) Distribution of the amplified and underexpressed top loci characteristic of patients with and without lymph node metastases. Overexpressed loci in patient with lymph node metastases and poor outcome are highlighted in red, and underexpressed probes are highlighted in green.

View larger version (74K): [in this window] [in a new window]   Fig A2. Signaling networks associated with aggressive behavior in patients with invasive disease. Molecular pathways associated with the genetic signature consisting of 174 probes differentially expressed in patients with lymph node metastases and poor outcome simultaneously. These analyses of the signaling networks in which these genes are involved revealed the relevance of TP53 pathway, highlighted in blue, in late steps along bladder cancer progression. Overexpressed genes in patients with lymph node metastases and poor outcome are highlighted in red (+1.000), and underexpressed probes are highlighted in green (–1.000).

	Authors' Disclosures of Potential Conflicts of Interest

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

	Author Contributions

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 

Conception and design: Marta Sanchez-Carbayo, Carlos Cordon-Cardo Financial support: Carlos Cordon-Cardo Administrative support: Carlos Cordon-Cardo Provision of study materials or patients: Marta Sanchez-Carbayo, Fabien Saint, Carlos Cordon-Cardo Collection and assembly of data: Marta Sanchez-Carbayo, Nicholas D. Socci, Juanjo Lozano, Fabien Saint Data analysis and interpretation: Marta Sanchez-Carbayo, Nicholas D. Socci, Juanjo Lozano, Fabien Saint Manuscript writing: Marta Sanchez-Carbayo, Nicholas D. Socci, Juanjo Lozano, Carlos Cordon-Cardo Final approval of manuscript: Marta Sanchez-Carbayo, Nicholas D. Socci, Juanjo Lozano, Fabien Saint, Carlos Cordon-Cardo

 

	Glossary

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
Bootstrap resampling technique: An analytical tool that evaluates how robust the associations are between the specimens under evaluation on the basis of the gene profiles. The higher the number provided by this method, the more robust the associations.

Hierarchical clustering: An analytical tool used to find the closest associations among gene profiles and specimens under evaluation.

Immunohistochemical analyses: Techniques used to evaluate the levels and patterns of expression of protein on cells or tissue specimens located on flat slides.

Tissue array: Used to analyze the expression of genes of interest simultaneously in multiple tissue samples, tissue microarrays consist of hundreds of individual tissue samples placed on slides ranging from 2 to 3 mm in diameter. Using conventional histochemical and molecular detection techniques, tissue microarrays are powerful tools to evaluate the expression of genes of interest in tissue samples. In cancer research, tissue microarrays are used to analyze the frequency of a molecular alteration in different tumor type, to evaluate prognostic markers, and to test potential diagnostic markers.

	Acknowledgment

 
We thank all members of the Genomics Core Laboratories for their technical support in this study. We thank the Tissue Procurement Core, particularly Cora Mariano, Barbara Kaje-Injejian, Katrina Allen, and Raul Meliton for their support in facilitating tumor tissues.

	NOTES

 
Terms in blue are defined in the glossary, found at the end of this article and online at www.jco.org.

Authors' disclosures of potential conflicts of interest and author contributions are found at the end of this article.

	REFERENCES

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions Glossary REFERENCES

 
1. Jemal A, Tiwari RC, Murray T, et al: Cancer Statistics 2004. CA Cancer J Clin 54:8-29, 2004[Abstract/Free Full Text]

2. American Joint Committee on Cancer: Urinary bladder, in Greene FL (ed): American Joint Committee on Cancer: Cancer Staging Manual (ed 6). New York, NY, Springer, 2002, pp 335-340

3. Reuter VE, Melamed MR: The lower urinary tract, in Sternberg SS (ed): Diagnostic Surgical Pathology. New York, NY, Raven Press, 1989, pp 1355-1392

4. Thykjaer T, Workman C, Kruhoffer M, et al: Identification of gene expression patterns in superficial and invasive human bladder cancer. Cancer Res 61:2492-2499, 2001[Abstract/Free Full Text]

5. Sanchez-Carbayo M, Socci ND, Charytonowicz E, et al: Molecular profiling of bladder cancer using cDNA microarrays: Defining histogenesis and biological phenotypes. Cancer Res 62:6973-6980, 2002[Abstract/Free Full Text]

6. Dyrskjot L, Thykjaer T, Kruhoffer M, et al: Identifying distinct classes of bladder carcinoma using microarrays. Nat Genet 33:90-96, 2003[CrossRef][Medline]

7. Sanchez-Carbayo M, Socci ND, Lozano JJ, et al: Gene discovery in bladder cancer progression using cDNA microarrays. Am J Pathol 163:505-516, 2003[Abstract/Free Full Text]

8. Modlich O, Prisack HB, Pitschke G, et al: Identifying superficial, muscle-invasive, and metastasizing transitional cell carcinoma of the bladder: Use of cDNA array analysis of gene expression profiles. Clin Cancer Res 10:3410-3421, 2004[Abstract/Free Full Text]

9. Dyrsjot L, Kruhoffer M, Thykjaer T, et al: Gene expression in the urinary bladder: A common carcinoma in situ gene expression signature exists disregarding histopathological classification. Cancer Res 64:4040-4048, 2004[Abstract/Free Full Text]

10. Nicholson BE, Frierson HF, Conaway MR, et al: Profiling the evolution of human metastatic bladder cancer. Cancer Res 64:7813-7821, 2004[Abstract/Free Full Text]

11. Sanchez-Carbayo M, Saint F, Lozano JJ, et al: Comparison of gene expression profiles in laser-microdissected, nonembedded, and OCT-embedded tumor samples by oligonucleotide microarray analysis. Clin Chem 49:2096-2100, 2003[Free Full Text]

12. Eisen MB, Spellman PT, Brown PO, et al: Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci U S A 95:14863-14868, 1998[Abstract/Free Full Text]

13. Felsenstein J: Confidence limits on phylogenies: An approach using the bootstrap. Evolution 39:783-791, 1985[CrossRef]

14. Dawson-Saunders B, Trapp RG: Basic & Clinical Biostatistics (ed 2). Norwalk, CT, Appleton & Lange, 1994

15. Reiner A, Yekutieli D, Benjamini Y: Identifying differentially expressed genes using false discovery rate controlling procedures. Bioinformatics 19:368-375, 2003[Abstract/Free Full Text]

16. Furey TS, Cristianini N, Duffy N, et al: Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics 16:906-914, 2000[Abstract/Free Full Text]

17. Gentleman RC, Carey VJ, Bates DM, et al: Bioconductor: Open software development for computational biology and bioinformatic. Genome Biol 5:R80, 2004[CrossRef][Medline]

18. Benjamini Y, Drai D, Elmer G, et al: Controlling the false discovery in behavior genetics research. Behav Brain Res 125:279-284, 2001[CrossRef][Medline]

19. Kanehisa M, Goto S, Kawashima S, et al: The KEGG databases at GenomeNet. Nucleic Acids Res 30:42-46, 2002[Abstract/Free Full Text]

20. Hosack DA, Dennis G Jr, Sherman BT, et al: Identifying biological themes within lists of genes with EASE. Genome Biol 4:R70, 2003[CrossRef][Medline]

21. Breitkreutz BJ, Jorgensen P, Breitkreutz A, et al: AFM 4.0: A toolbox for DNA microarray analysis. Genome Biol 2:SOFTWARE0001, 2001

22. Goeman JJ, van de Geer SA, de Kort F, et al: A Global Test for groups of genes: Testing association with a clinical outcome. Bioinformatics 20:93-99, 2004[Abstract/Free Full Text]

23. Veltman JA, Fridlyand J, Pejavar S, et al: Array-based comparative genomic hybridization for genome-wide screening of DNA copy number in bladder tumors. Cancer Res 63:2872-2880, 2003[Abstract/Free Full Text]

24. Markl ID, Jones PA: Presence and location of TP53 mutation determines pattern of CDKN2A/ARF pathway inactivation in bladder cancer. Cancer Res 58:5348-5353, 1998[Abstract/Free Full Text]

25. Dalbagni G, Presti J, Reuter V, et al: Genetic alterations in bladder cancer. Lancet 342:469-471, 1993[CrossRef][Medline]

26. Hafner C, Knuechel R, Zanardo L, et al: Evidence for oligoclonality and tumor spread by intraluminal seeding in multifocal urothelial carcinomas of the upper and lower urinary tract. Oncogene 20:4910-4915, 2001[CrossRef][Medline]

27. Pronin AN, Morris AJ, Surguchov A, et al: Synucleins are a novel class of substrates for G protein-coupled receptor kinases. J Biol Chem 275:26515-26522, 2000[Abstract/Free Full Text]

28. Iwaki H: Diagnostic potential in bladder cancer of a panel of tumor markers (calreticulin, gamma-synuclein, and catechol-o-methyltransferase) identified by proteomic analysis. Cancer Sci 95:955-961, 2004[Medline]

29. Spinardi L, Rietdorf J, Nitsch L, et al: A dynamic podosome-like structure of epithelial cells. Exp Cell Res 295:360-374, 2004[CrossRef][Medline]

30. Kim JH, Tuziak T, Hu L, et al: Alterations in transcription clusters underlie development of bladder cancer along papillary and nonpapillary pathways. Lab Invest 85:532-549, 2005[CrossRef][Medline]

31. Dyrskjot L, Zieger K, Kruhoffer M, et al: A molecular signature in superficial bladder carcinoma predicts clinical outcome. Clin Cancer Res 11:4029-4036, 2005[Abstract/Free Full Text]

32. Blaveri E, Simko JP, Korkola JE, et al: Bladder cancer outcome and subtype classification by gene expression. Clin Cancer Res 11:4044-4055, 2005[Abstract/Free Full Text]

33. Titus B, Frierson HF Jr, Conaway M, et al: Endothelin axis is a target of the lung metastasis suppressor gene RhoGDI2. Cancer Res 65:7320-7327, 2005[Abstract/Free Full Text]

Submitted June 22, 2005; accepted October 13, 2005.

This article has been cited by other articles:

				A. Aleman, V. Cebrian, M. Alvarez, V. Lopez, E. Orenes, L. Lopez-Serra, F. Algaba, J. Bellmunt, A. Lopez-Beltran, P. Gonzalez-Peramato, et al. Identification of PMF1 Methylation in Association with Bladder Cancer Progression Clin. Cancer Res., December 15, 2008; 14(24): 8236 - 8243. [Abstract] [Full Text] [PDF]

				A. Holyoake, P. O'Sullivan, R. Pollock, T. Best, J. Watanabe, Y. Kajita, Y. Matsui, M. Ito, H. Nishiyama, N. Kerr, et al. Development of a Multiplex RNA Urine Test for the Detection and Stratification of Transitional Cell Carcinoma of the Bladder Clin. Cancer Res., February 1, 2008; 14(3): 742 - 749. [Abstract] [Full Text] [PDF]

M. Sanchez-Carbayo, N. D. Socci, L. Richstone, M. Corton, N. Behrendt, J. Wulkfuhle, B. Bochner, E. Petricoin, and C. Cordon-Cardo Genomic and Proteomic Profiles Reveal the Association of Gelsolin to TP53 Status and Bladder Cancer Progression Am. J. Pathol., November 1, 2007;