PMCCPMCCPMCC

Search tips
Search criteria 

Advanced

 
Logo of jcoHomeThis ArticleSearchSubmitASCO JCO Homepage
 
J Clin Oncol. 2009 August 20; 27(24): 3929–3937.
Published online 2009 July 20. doi:  10.1200/JCO.2008.18.5744
PMCID: PMC2799152

Generation of a Concise Gene Panel for Outcome Prediction in Urinary Bladder Cancer

Abstract

Purpose

This study sought to determine if alterations in molecular pathways could supplement TNM staging to more accurately predict clinical outcome in patients with urothelial carcinoma (UC).

Patients and Methods

Expressions of 69 genes involved in known cancer pathways were quantified on bladder specimens from 58 patients with UC (stages Ta-T4) and five normal urothelium controls. All tumor transcript values beyond two standard deviations from the normal mean expression were designated as over- or underexpressed. Univariate and multivariable analyses were conducted to obtain a predictive expression signature. A published external data set was used to confirm the potential of the prognostic gene panels.

Results

In univariate analysis, six genes were significantly associated with time to recurrence, and 10 with overall survival. Recursive partitioning identified three genes as significant determinants for recurrence, and three for overall survival. Of all genes identified by either univariate or partitioning analysis, four were found to significantly predict both recurrence and survival (JUN, MAP2K6, STAT3, and ICAM1); overexpression was associated with worse outcome. Comparing the favorable (low or normal) expression of ≥ three of four versus ≤ two of four of these oncogenes showed 5-year recurrence probability of 41% versus 88%, respectively (P < .001), and 5-year overall survival probability of 61% versus 5%, respectively (P < .001). The prognostic potential of this four-gene panel was confirmed in a large independent external cohort (disease-specific survival, P = .039).

Conclusion

We have documented the generation of a concise, biologically relevant four-gene panel that significantly predicts recurrence and survival and may also identify potential therapeutic targets for UC.

INTRODUCTION

Current urothelial carcinoma (UC) management primarily depends on histologic grading and pathologic staging of the tumor.1,2 Although these provide assessment of risk, they are unable to predict outcome in an individual patient. Molecular alterations in tumors precede visually identifiable morphologic changes and are responsible for their biologic behavior,3,4 prognosis, and response to therapy. Therefore, histopathologic staging in UC must be complemented with molecular correlates to accurately predict clinical outcome and therapeutic response.

This study was performed on the basis of growing evidence that multiple alterations in major cancer pathways are responsible for progression of UC.5 We profiled the expression of 69 genes involved in eight crucial cancer pathways (Appendix Table A1, online only) by standardized competitive reverse transcriptase–polymerase chain reaction (StaRT-PCR; Gene Express, Toledo, OH), quantifying absolute expressions in relation to a fixed quantity of the housekeeping gene β-actin.6 The ultimate goal is to identify a concise marker panel that can predict clinical outcome in patients with UC; this study was designed to identify genes that would comprise such a panel.

PATIENTS AND METHODS

Patient Selection

The study cohort comprised 58 patients with UC (mean age, 69.5 years) and five normal controls. Frozen UC tissue was obtained after radical cystectomy from 49 patients at the University of Southern California (Los Angeles, CA) and nine patients at the University of California, San Francisco (San Francisco, CA), between 1991 and 2002. These included patients with invasive (T1-4) tumors and noninvasive (Ta) tumors refractory to bladder-conserving therapies. Patients with distant metastasis at time of diagnosis were excluded. TNM staging was standardized to the 2002 American Joint Committee on Cancer recommendations.1 Controls consisted of normal urothelium from the bladder necks of patients who had undergone radical prostatectomy for localized prostatic adenocarcinoma without bladder involvement and no history of UC.

Eight patients (13.8%) received adjuvant chemotherapy and/or radiotherapy. These included patients with high-grade recurrent noninvasive (n = 1), muscle-invasive (n = 2), extravesically extending (n = 1), and nodal metastasized (n = 4) tumors. Mean follow-up was 3.04 years (range, 0.30 to 10.44 years), during which 29 patients developed recurrent disease, and 38 patients died (Appendix, online only). UC was the cause of death in 30 patients, whereas eight patients died as a result of undocumented causes. Informed consent was obtained from all patients. The study was approved by the respective institutional review boards.

StaRT-PCR and Comparison of Tumor and Normal Gene Expression Levels

After RNA extraction by conventional TRIzol method (Invitrogen, Carlsbad, CA), cDNA was prepared, and quantitative gene expression profiling was performed by StaRT-PCR (Appendix, online only).7,8 Each gene was reported as number of mRNA molecules expressed per 106 β-actin molecules.

After log transformation, each transcript expression level for all patients with UC was compared with the respective mean level in normal urothelium. Any tumor transcript level greater than two standard deviations from the mean expression level in normal urothelium was labeled as overexpressed, whereas any expression level less than two standard deviations from the mean level in normal urothelium was labeled as underexpressed. Tumor transcript expression levels falling between two standard deviations above and below the mean levels in normal urothelium were labeled as normally expressed. Thus, each tumor transcript was assigned an expression value (low, normal, or high) depending on its level compared with that of normal urothelium. Once the significant genes were identified, transcript expressions were dichotomized into favorable and unfavorable categories depending on outcomes associated with respective expression values.

Data Analysis

The clinical outcomes analyzed were time to recurrence, disease-specific survival, and overall survival (Appendix, online only). Time to recurrence was preferred over disease-specific survival, because currently most patients who die as a result of UC have documentation of disease recurrence; overall survival also accounts for cases in which cause of death is unknown and in which the impact of UC treatment may contribute to death, although disease does not recur.

The log-rank test9 was used to examine how clinical parameters and gene expression values were associated with clinical outcome. In univariate analysis, the relative risk ratio and associated 95% CI were calculated on the basis of the log-rank test.10 To adjust for multiple comparisons and control the false-positive rate, bootstrap internal validation was performed for all genes identified by univariate analysis, thereby eliminating the possibility of overfitting or biasing conclusions on the basis of a small subset.11 One thousand bootstrap samples of 58 observations each were drawn from the original UC cohort using simple random sampling with replacement. Selected genes were retained if associated P ≤ .050 in more than 500 simulations.12,13 Reported P values are two sided.

Three multivariable approaches were adopted. The first approach used nonparametric classification and regression trees generated by recursive partitioning (RP) to explore gene expression variables and separate patients into prognostic subgroups on the basis of time to recurrence and overall survival (Appendix, online only).14,15 In the second approach, stepwise forward selection was used on the basis of the Cox proportional hazards model, stratified by pathologic stage and lymph-node density. Third, Akaike information criterion (AIC) within a Cox proportional hazards model, stratified by pathologic stage, was used to demonstrate the discriminatory ability of the gene panels.16 A smaller AIC value indicates a more desirable panel for predicting outcome.17 Functional pathway analysis was also conducted using Dijkstra's shortest paths algorithm (Appendix, online only).18

External Validation

For validation purposes, multiple public repositories were searched for expression profiling data from independent external UC cohorts that encompassed all stages and provided publicly available corresponding clinical outcome information. The study by Sanchez-Carbayo et al19 provided such a data set online that also profiled all genes investigated in our cohort. We used the same binary outcome as defined in that study: whether the patient had died as a result of UC or had no evidence of disease at last follow-up. Because true normal urothelium was not used in this study, and adjacent normal urothelium can potentially harbor genetic alterations similar to adjacent tumor tissue,20 the expression profiles of adjacent normal urothelium were disregarded in our analysis. The final validation cohort consisted of expression profiles from primary tumors of 91 patients with UC (mean age, 67.8 years; Appendix, online only).

After log transformation, representative probe sets on U133A GeneChips (Affymetrix, Santa Clara, CA; Appendix, online only)21 for the 11 genes predictive for overall survival from our analysis were chosen for validation in the external data set because survival was the only clinical outcome available for this cohort. Expression of any gene below or above its median expression level in the validation cohort was considered favorable or unfavorable, respectively, in accordance with the findings from our study cohort (Appendix Table A2, online only). Pearson's χ2 test was used to examine associations with clinical outcome.

RESULTS

Clinicopathologic Parameters and Clinical Outcome

Associations of clinicopathologic parameters of the study cohort with outcome are listed in Table 1. Pathologic stage was significantly associated with overall survival but not with time to recurrence. Interestingly, three of 10 patients with TaN− disease had postcystectomy pelvic recurrences, demonstrating an unusually aggressive clinical course. In contrast, nine of 11 patients with T3-4N− disease experienced an unusually indolent clinical course with no recurrence at last follow-up.

Table 1.
Association of Patient Demographics and Clinicopathologic Parameters With Outcome

Individual Genes and Clinical Outcome

By univariate analysis, STAT3 (P = .009), IGF1 (P = .021), JUN (P = .026), SOD1 (P = .033), and MAP2K6 (P = .044) were significantly associated with time to recurrence (Table 2; Data Supplement, online only). BCL2 (P = .055) also showed a trend toward significance for time to recurrence. The consistency of these findings was supported by bootstrap analysis that selected these transcripts in more than half of the bootstrap samples for recurrence.

Table 2.
Genes Predictive of Recurrence and Overall Survival by Univariate Analysis*

MAPK12 (P < .001), JUN (P = .001), TNFSF10 (P = .007), CCNA2 (P = .009), ICAM1 (P = .014), BCL2L1 (P = .015), MAP2K6 (P = .016), TGIF (P = .047), and STAT3 (P = .050) were significantly associated with overall survival (Table 2; Data Supplement, online only). FOSL1 (P = .051) also showed a trend toward significance for overall survival. Bootstrap analysis confirmed the consistency of these findings by selecting these genes in more than half of the bootstrap samples for overall survival.

Interdependent Gene Expressions and Clinical Outcome

RP analysis was performed to identify any gene that may, by itself, not be prognostically important and thus not feature in the univariate analysis but, in association with other genes, may be associated with clinical outcome. The expressions of BMP6, SOD1, and ICAM1 were identified as joint determinants for recurrence (Fig 1A). At the end of the study, 87% of patients with low BMP6 and high SOD1 expressions (group 1a) remained recurrence free, whereas this was seen in only 14% of patients with normal or high BMP6 and high ICAM1 expressions (group 1d); in patients with low BMP6 and low or normal SOD1 expressions (group 1b) and patients with normal or high BMP6 and low or normal ICAM1 expressions (group 1c), intermediate recurrence rates were observed. Log-rank analysis of these four groups showed significant association with time to recurrence (P < .001), with group 1a demonstrating the lowest and group 1d demonstrating the highest probability of recurrence (Fig 1B).

Fig 1.
Recursive partitioning analysis for clinical outcome in urothelial carcinoma. (A) Expression values of BMP6, SOD1, and ICAM1 were used to define four distinct patient groups (1a to 1d) on the basis of time to recurrence with (B) significant differences ...

MAPK12, GSTM3, and ICAM1 were identified as joint determinants for overall survival (Fig 1C). At the end of the study, 89% of patients with low MAPK12 and GSTM3 expressions (group 2a) had survived, whereas 0% of patients with normal or high MAPK12 and high ICAM1 expressions (group 2d) had survived. Patients with low MAPK12 and normal or high GSTM3 expressions (group 2b) and those with normal or high MAPK12 and low or normal ICAM1 expressions (group 2c) had intermediate survival rates. Log-rank analysis of these four groups showed significant association with overall survival (P < .001), with group 2a having the best and group 2d having the worst survival probabilities (Fig 1D).

Combined Analysis of Four Common Genes

We hypothesized that the most biologically relevant genes would predict both recurrence and overall survival by univariate and/or RP analysis. JUN, MAP2K6, and STAT3 were significantly associated with time to recurrence and overall survival by univariate analysis, and ICAM1 was significantly associated with overall survival by univariate analysis and with recurrence and overall survival by RP analysis (Appendix Table A3, online only).

On the basis of comparison of individual gene expression patterns with outcome, low or normal expression was found to be favorable, whereas overexpression was unfavorable (Appendix Table A2, online only). This was consistent with their functions as oncogenes.8,22,23 The study cohort was then divided into two groups: patients with favorable (low or normal) expressions of ≥ three of four genes (n = 35) and patients with favorable expressions of ≤ two of four genes (n = 21). Two patients were excluded from the analysis because they each had two favorable, one unfavorable, and one missing gene expressions and could thus not be confidently classified into either group. The 5-year recurrence probabilities in these groups were 41% and 88%, respectively (P < .001; Fig 2A); the 5-year disease-specific survival probabilities were 68% and 7%, respectively (P < .001; Fig 2B); and the 5-year overall survival probabilities were 61% and 5%, respectively (P < .001; Fig 2C; Table 3). To confirm that these findings were not the results of inherent differences in the pathologic stages, the analysis was repeated, stratified by each stage, and the results and patterns remained consistent (data not shown). In a sensitivity analysis, the gene panel was re-evaluated employing the Cox proportional hazards model, stratified by pathologic stage and lymph-node density. When patients with favorable expressions of ≥ three of four genes were used as the reference group, relative risks of recurrence for patients with favorable expressions of ≤ two of four genes were 3.09 and 2.63, respectively, and relative risks of death were 4.48 and 4.11, respectively. These relative risks remained statistically significant even after excluding the eight patients who received adjuvant treatment, indicating that the predictive value of these four genes was not altered by adjuvant therapy (data not shown).

Fig 2.
Generation of a predictive expression signature in urothelial carcinoma. Groups generated on the basis of low or normal (favorable) expressions of ≥ three of four or ≤ two of four genes (JUN, MAP2K6, STAT3, and ICAM1) showed significant ...
Table 3.
Association of Favorable Expressions of JUN, MAP2K6, STAT3, and ICAM1 With Outcome*

Interestingly, all three patients with TaN− disease with pelvic recurrences had expression profiles consistent with high risk of recurrence (favorable expressions of ≤ two of four genes). Similarly, six of seven patients with TaN− disease and seven of nine patients with T3-4N− disease without recurrences had expression profiles consistent with low risk of recurrence (favorable expressions of ≥ three of four genes).

Relative Predictive Power of Gene Panels

To assess how much predictive power was lost on exclusion of the significant genes that were not common predictors of recurrence and overall survival, AIC within a Cox proportional hazards model, stratified by pathologic stage, was used to compare the four-gene panel with the eight- and 11-gene panels, which contained genes individually predictive for recurrence and overall survival, respectively, by univariate and/or RP analysis. Expressions of these genes were dichotomized into favorable and unfavorable categories on the basis of their association with outcome (Appendix Table A2, online only), and patients with missing expression values for these genes were excluded for this analysis. Although the eight- and 11-gene panels, as expected, performed best in predicting time to recurrence and overall survival, respectively, their performance was not substantially superior to that of the four-gene panel (Fig 2D). In fact, the differences in AIC values between the four-gene panel and the best performing panels for time to recurrence and overall survival were 3.03 and 4.07, respectively. This suggests that the predictive performances of these panels were empirically comparable, because the absolute differences in AIC value were close to or less than 4.17

Validation of Identified Gene Panels

An independent external UC cohort,19 profiled for gene expressions using oligonucleotide microarrays, was used to confirm the prognostic potential of the identified gene panels. Because only disease-specific survival was reported for the cohort, the 11-gene panel predictive for overall survival and the common four-gene panel were chosen for validation. Associations of clinicopathologic parameters with disease-specific survival are listed in Table 4. The cohort was divided into two groups on the basis of the 11-gene panel: patients with favorable expressions of ≥ seven of 11 genes (n = 56) and patients with favorable expressions of ≤ six of 11 genes (n = 35). Using the former as the reference group, relative risk of disease-specific death in patients with favorable expressions of ≤ six of 11 genes was 2.00 (P = .007). To assess the predictive power of the common four-gene panel, the validation cohort was again divided into two groups: patients with favorable expressions of ≥ three of four genes (n = 50) and patients with favorable expressions of ≤ two of four genes (n = 41). Using the former as the reference group, the relative risk of disease-specific death in patients with favorable expressions of ≤ two of four genes was 1.71 (P = .039).

Table 4.
Association of Patient Demographics, Clinicopathologic Parameters, and Prognostic Gene Panels With Disease-Specific Survival in the External Validation Cohort19

DISCUSSION

We used a quantitative, pathway-specific approach to profile genes involved in important cellular pathways that are crucial in UC development. The choice for the final predictive panel was determined on the basis of the hypothesis that the most biologically relevant genes should be able to predict both recurrence and survival. In this study, the four-gene panel (JUN, MAP2K6, STAT3, and ICAM1) was a highly significant predictor of these outcomes, independent of standard prognostic criteria (ie, stage and lymph-node density). In addition, this panel identified high-risk patients; nearly all patients with favorable expressions of ≤ two of four genes experienced recurrence and died. The prognostic potential of this panel was additionally supported by an external data set that profiled genes using a completely different methodology, thereby demonstrating the robustness of this four-gene panel in predicting clinical outcome.

Gene expression profiles are usually generated using microarrays. These studies may involve inconsistencies in results and lack of reproducibility across platforms.24,25 Furthermore, the output often contains more than 20 to 100 genes, which can dilute biologic and clinical relevance while increasing noise and opportunities for random chance. Although our hypothesis-driven exploration limited potential for discovery, the rational choice allowed identification of key genes and associated pathways of prognostic value.

The univariate analysis identified six genes associated with recurrence, and 10 associated with overall survival. The protein products of IGF1, JUN, MAP2K6, BCL2, CCNA2, ICAM1, BCL2L1, TGIF, and FOSL1 have been associated with poor prognosis in several cancers, including bladder cancer.8,22,2631 High expressions of these genes were associated with worse prognosis, consistent with their biologic roles as oncogenes. Our study also demonstrated constitutive activation of the mitogen-activated protein kinase pathway in UC5; low MAPK12 expression was associated with higher probability of overall survival. STAT3 overexpression corresponded with poorer prognosis, consistent with observations that signal transducer and activator of transcription 3 increases the invasiveness of UC cell lines.23 Low SOD1 expression corresponded with decreased probability of recurrence, consistent with findings in acute myelogenous leukemia and lymphoproliferative syndromes.32 Although tumor necrosis factor–related apoptosis-inducing ligand (TRAIL), the protein product of TNFSF10, induces apoptosis, patients with increased TNFSF10 expression had poorer overall survival. This patient subset was probably insensitive to TRAIL-mediated apoptosis, consistent with findings demonstrating that different UC cell lines have varying degrees of susceptibility to TRAIL.33

The RP analysis also selected BMP6 and ICAM1 as joint determinants of recurrence. Bone morphogenetic protein 6 promotes tumor angiogenesis, and elevated ICAM1 expression is associated with increased metastatic potential in UC.8,34 Patients with low BMP6 and high SOD1 expressions had the lowest recurrence rates, whereas those with normal or high BMP6 and high ICAM1 expressions had the highest. GSTM3 was also associated with overall survival in RP analysis. GSMT3 polymorphisms are linked to carcinogenesis, and GSTM3 mutations are associated with increasing risk for UC.35 In patients with low MAPK12 expression, those with low GSTM3 expression as well had the highest survival probability, whereas those with normal or high GSTM3 expression had lower survival probability.

When the interrelationships between proteins transcribed from these genes were examined, nine direct and more than 150 indirect interactions were discovered (Data Supplement, online only), which highlights the importance of their crosstalk. This led us to focus on genes that could predict both recurrence and survival. Obtaining a concise prognostic marker list is crucial in such studies, because clinical applications of such panels are more cost effective and practical. Although such prognostic panels have been previously identified, they have usually featured markers from a single cellular pathway.36,37 The four-gene panel obtained after profiling genes across multiple pathways robustly predicted clinical outcome. Additionally, the ability of this panel to accurately predict recurrence independent of stage is likely to be a useful supplement to routine staging. Furthermore, MAP2K6 and ICAM1 were also previously identified by our group to predict nodal metastasis in UC.8 Validation of the four- and 11-gene panels on the external data set was consistent with AIC observations that although the 11-gene panel could expectedly better predict survival, its performance was not substantially superior to that of the four-gene panel. Moreover, the validation highlighted the robustness of the four-gene panel, independent of the platform used for profiling the genes.

In conclusion, using a multiplexed, biologically driven approach, we have identified a panel comprising JUN, MAP2K6, STAT3, and ICAM1 that can predict clinical outcome in UC independent of conventional prognostic criteria and identify patients with operable UC who will experience recurrence despite undergoing definitive surgery alone. These patients would clearly benefit from additional therapy. Increasing numbers of alterations in these genes predict poorer prognosis. Additional study of this panel is warranted to better characterize its ability to identify patients at higher risk. Although limited transcripts were analyzed, this does suggest that these genes and their associated pathways may serve as promising outcome predictors and potential therapeutic targets in UC.

Supplementary Material

[Data Supplements]

Acknowledgment

We thank Marta Sánchez-Carbayo, PhD, and Carlos Cordon-Cardo, MD, PhD, and their group for allowing open access to their expression profiling data that were used as the validation cohort in this study; and Lillian Young for technical assistance.

Glossary Terms

Standardized competitive reverse transcriptase–polymerase chain reaction:
Quantitative polymerase chain reaction that measures absolute expression levels of multiple genes using competitive templates of the target and reference (β-actin) genes incorporated into standardized mixtures of internal standards. Use of the same standardized mixtures potentially allows comparability of data across experiments and laboratories.
Recursive partitioning:
Multivariable analysis that generates a clinically intuitive decision tree model in which the population is divided into prognostic subgroups. This is achieved through multiple dichotomous divisions on the basis of a set of independent variables.
Akaike information criterion:
Measure of the goodness of fit of a statistical model that discourages overfitting and is used as a tool for model selection. For a given data set, competing models are ranked according to their Akaike information criterion value, and the one with the lowest value is considered the best. However, there is no established value above which a given model is rejected.
Dijkstra's shortest paths algorithm:
Graph search algorithm that finds the path with lowest cost (ie, the shortest path) between a given node (or, in the case of functional biological networks, a given gene) and every other node.

Appendix

Details on Patient Follow-Up

The study cohort included patients with Ta (n = 10), T1 (n = 11), T2 (n = 10), T3 (n = 21), and T4 (n = 6) disease. Although the cohort had a modest follow-up duration (mean, 3.04 years), this was primarily the result of early deaths rather than loss to follow-up. Mean follow-up was longer in patients with no recurrences at end of study (3.69 years) than it was in those who experienced recurrence (2.39 years). Similarly, patients alive at end of study also had longer mean follow-up (4.5 years) than those who died (2.28 years). All patients in our cohort who received adjuvant therapy eventually experienced recurrence and died. Although adjuvant therapy did delay the time of first recurrence (mean, 1.5 years) compared with that in patients who did not receive any therapy besides surgery (mean, 1.07 years), all patients receiving adjuvant therapy were followed up until death; we thereby avoided drawbacks of short follow-up in these patients.

Quantitative Standardized Competitive Reverse Transcriptase–Polymerase Chain Reaction After RNA extraction and cDNA preparation,7,8 quantitative gene expression profiling was performed using standardized competitive reverse transcriptase–polymerase chain reaction (StaRT-PCR; Gene Express, Toledo, OH) analysis. The internal standard mixtures (A to F over six logs of concentration) contained competitive templates (CTs) of 69 transcripts (Table A1) in addition to 600,000 β-actin CT molecules/μL (for each sample, StaRT-PCR analysis was performed using five different CT mixes [B to F]). Thus each sample underwent five separate PCR analyses, each separate reaction containing the ready-to-use master mixture, cDNA sufficient for expression measurements of the 70 transcripts (including β-actin), primers for the 70 transcripts, and one of the five CT mixes (B to F). After PCR, the amplification products were electrophoresed, and image analysis and quantitation of band fluorescence intensities were performed.7 The expression of each gene was reported as number of mRNA molecules per 106 molecules of β-actin.

Definitions of Clinical Outcome Parameters

Time to recurrence was calculated from date of cystectomy to first date of clinical recurrence or progression. Patients without recurrence or progression were censored at time of death or last follow-up. Disease-specific survival was calculated from date of cystectomy to date of death as a result of urothelial carcinoma (UC) or last follow-up date. Overall survival was calculated from date of cystectomy to date of death as a result of any cause; patients who were still alive were censored at date of last follow-up.

Recursive Partitioning Analysis

A nonparametric classification and regression tree generated by recursive partitioning (RP) was used to explore the gene expression variables and categorize patients into prognostic subgroups on the basis of time to recurrence and overall survival.14,15 This included two processes: “tree growing” and “tree pruning.”14 In tree growing, all patients started in one group, and a series of binary recursive splits were made on the basis of an expression value cutoff that yielded subgroups with the greatest dissimilarities in clinical outcome. Tree pruning was then performed to produce simpler subtrees by assessing the misclassification error associated with a particular subtree. RP analysis was performed using the RPART function in the S-Plus library (Insightful, Seattle, WA).15

Functional Pathway Analysis

Functional analysis of pathways affected by the differentially expressed genes was conducted using MetaCore (GeneGo, St Joseph, MI). Dijkstra's shortest paths algorithm was employed to connect the significant genes by the shortest curated network paths.18 The number of steps in the path was restricted to a maximum of two to visualize the nearest directly interacting molecules and most significantly affected pathways.

Patient Selection for the Validation Cohort

The study by Sanchez-Carbayo et al19 used 157 tissue samples from 105 patients with UC (Ta-T4) and included whole genome profiles of primary tumors and adjacent normal urothelium generated using U133A human GeneChips (Affymetrix, Santa Clara, CA). Standard demographic and clinicopathologic information on patients without identifiers, along with their corresponding clinical outcome information, was publicly available. After excluding those patients for whom only expression profiling data on adjacent normal urothelial tissue was available, the final validation cohort included UC tissues from patients with Ta (n = 2), T1 (n = 23), T2 (n = 10), T3 (n = 45), and T4 (n = 11) disease, several of them available in duplicates.

Probe Set Selection and Data Analysis for the Validation Cohort

In choosing representative probe sets for each of the 11 genes predictive for overall survival from our study cohort, non–cross-reacting accurate-type (at) probe sets were preferred over cross-reacting types (x_at, s_at), and a cluster that was supported by DNA or full-length mRNA was preferred.21 The probe sets chosen were 206106_at (MAPK12), 213281_at (JUN), 202688_at (TNFSF10), 208991_at (STAT3), 203418_at (CCNA2), 202637_s_at (ICAM1), 215037_s_at (BCL2L1), 205699_at (MAP2K6), 203313_s_at (TGIF), 204420_at (FOSL1), and 202554_s_at (GSTM3). Expression of any gene below or above its median expression level was considered favorable or unfavorable, respectively, in accordance with the observations in the study cohort (Table A2). In case of duplicate samples, when any gene was overexpressed in any tumor duplicate, the gene was considered overexpressed in the tumor. The expression data were analyzed by the Partek Genomics Suite (Partek Inc, St Louis, MO).

Table A1.

List of Genes Involved in Eight Crucial Tumorigenic Pathways Investigated Using Standardized Competitive Reverse Transcriptase–Polymerase Chain Reaction, Including Gene Identification and Protein Transcribed

GeneOfficial NameGene IDTranscribed Protein
Apoptosis
    ANXA5Annexin A5308Annexin A5
    BADBCL2-antagonist of cell death572BAD
    BCL2B-cell CLL/lymphoma 2596Bcl-2
    BCL2L1BCL2-like 1598Bcl-xL
    CYP1A2cytochrome P450, family 1, subfamily A, polypeptide 21544CYP1A2
    DAPDeath-associated protein1611DAP1
    PTGS2Prostaglandin–endoperoxide synthase 2 (prostaglandin G/H synthase and cyclooxygenase)5743COX-2
    TGFBR2Transforming growth factor, beta receptor II (70/80kDa)7048TGF-beta receptor type II
    TGIFTGFB-induced factor (TALE family homeobox)7050TGIF
    TNFTumor necrosis factor (TNF superfamily, member 2)7124TNF-alpha
    TNFAIP1Tumor necrosis factor, alpha-induced protein 1 (endothelial)7126TNFAIP1
    TNFRSF1ATumor necrosis factor receptor superfamily, member 1A7132TNF-R1
    TNFSF10Tumor necrosis factor (ligand) superfamily, member 108743Apo2L
    TRAF4TNF receptor–associated factor 49618TRAF4
Cell cycle
    CCNA2Cyclin A2890Cyclin A2
    CCND3Cyclin D3896Cyclin D3
    CCNE1Cyclin E1898Cyclin E
    CCNG1Cyclin G1900Cyclin G1
    CDC2Cell division cycle 2, G1 to S and G2 to M983CDK1 (p34)
    CDC25CCell division cycle 25C995CDC25C
    CDK7Cyclin-dependent kinase 7 (MO15 homolog, Xenopus laevis, cdk-activating kinase)1022CDK7
    CDK8Cyclin-dependent kinase 81024CDK8
    PCNAProliferating cell nuclear antigen5111PCNA
Gene regulation
    FOSv-fos FBJ murine osteosarcoma viral oncogene homolog2353c-Fos
    FOSL1FOS-like antigen 18061Fra-1
    HSF1Heat shock transcription factor 13297HSF1
    JUNv-jun sarcoma virus 17 oncogene homolog (avian)3725c-jun
    JUNBjun B proto-oncogene3726jun-B
    MAP3K14Mitogen-activated protein kinase kinase kinase 149020NIK (MAP3K14)
    MYCv-myc myelocytomatosis viral oncogene homolog (avian)4609c-Myc
    NFKB1Nuclear factor of kappa light polypeptide gene enhancer in B-cells 1 (p105)4790NF-κB1
    SP1Sp1 transcription factor6667SP1
Apoptosis plus cell cycle
    CDKN1ACyclin-dependent kinase inhibitor 1A (p21, Cip1)1026p21
    CDKN1BCyclin-dependent kinase inhibitor 1B (p27, Kip1)1027p27Kip1
    CDKN2ACyclin-dependent kinase inhibitor 2A (melanoma, p16, inhibits CDK4)1029p14ARF
    CDKN2CCyclin-dependent kinase inhibitor 2C (p18, inhibits CDK4)1031p18
    GAPDHGlyceraldehyde 3–phosphate dehydrogenase2597GAPDH
    MXD1MAX dimerization protein 14084MAD
    RB1Retinoblastoma 1 (including osteosarcoma)5925Rb protein
    RBL2Retinoblastoma-like 2 (p130)5934p130
    TP53Tumor protein p53 (Li-Fraumeni syndrome)7157p53
Apoptosis plus cell cycle plus gene regulation
    E2F1E2F transcription factor 11869E2F1
    E2F2E2F transcription factor 21870E2F2
    E2F4E2F transcription factor 4, p107/p130 binding1874E2F4
    E2F5E2F transcription factor 5, p130 binding1875E2F5
Antioxidation
    GSTM3Glutathione S-transferase M3 (brain)2947GSTM3
    GSTP1Glutathione S-transferase pi2950GSTP1
    GSTT1Glutathione S-transferase theta 12952GSTT1
    SOD1Superoxide dismutase 1, soluble (amyotrophic lateral sclerosis 1 (adult))6647SOD1
Cell-growth regulation
    IGF1Insulin-like growth factor 1 (somatomedin C)3479IGF1
    IGF2RInsulin-like growth factor 2 receptor3482IGF-2 receptor
    PDGFBPlatelet-derived growth factor beta polypeptide (simian sarcoma viral (v-sis) oncogene homolog)5155PDGF-B
    PDGFRLPlatelet-derived growth factor receptor-like5157PDGFRL
Signal transduction
    ERBB2v-erb-b2 erythroblastic leukemia viral oncogene homolog 2, neuro/glioblastoma derived oncogene homolog (avian)2064ErbB2
    LYNv-yes-1 Yamaguchi sarcoma viral related oncogene homolog4067Lyn
    MAPK8Mitogen-activated protein kinase 85599JNK1 (MAPK8)
    MAPK9Mitogen-activated protein kinase 95601JNK2 (MAPK9)
    MAPK12Mitogen-activated protein kinase 126300p38gamma (MAPK12)
    MAP2K6Mitogen-activated protein kinase kinase 65608MEK6 (MAP2K6)
    STAT3Signal transducer and activator of transcription 3 (acute-phase response factor)6774STAT3
Angiogenesis
    FGF5Fibroblast growth factor 52250FGF5
    FGFR4Fibroblast growth factor receptor 42264FGFR4
    KDRKinase insert domain receptor (a type III receptor tyrosine kinase)3791VEGFR-2
    VEGFVascular endothelial growth factor7422VEGF-A
Invasion
    BMP6Bone morphogenetic protein 6654BMP6
    CDH3Cadherin 3, type 1, P-cadherin (placental)1001P-cadherin
    ICAM1Intercellular adhesion molecule 1 (CD54), human rhinovirus receptor3383ICAM1
    MMP16Matrix metallopeptidase 16 (membrane-inserted)4325MMP-16
    TIMP2TIMP metallopeptidase inhibitor 27077TIMP2
Reference gene
    ACTBActin, beta60β-actin

Table A2.

Classification of Individually Prognostic Genes Into Favorable and Unfavorable Expressions on the Basis of Expression Patterns

GenePrognostic Status
Expression Value (compared with that in normal urothelium)
RecurrenceOverall SurvivalLowNormalHigh
MAPK12YFUU
JUN*YYFFU
TNFSF10YFFU
CCNA2YFFU
STAT3*YYFFU
ICAM1*YYFFU
BCL2L1YFFU
MAP2K6*YYFFU
IGF1YFFU
SOD1YUUF
TGIFYFFU
FOSL1YFFU
BCL2YFFU
BMP6YFUU
GSTM3YFUU

Abbreviations: Y, selection of gene by log-rank and/or recursive partitioning analysis; F, favorable expression; U, unfavorable expression.

*Gene selected for the common four-gene panel.
Trend toward statistical significance.

Table A3.

List of Genes of Interest From Univariate and Recursive Partitioning Analyses

GeneLog-Rank P (univariate analysis confirmed by bootstrap)*
Recursive Partitioning Analysis (multivariable)
RecurrenceOverall SurvivalRecurrenceOverall Survival
MAPK12< .001Y
JUN.026.001
TNFSF10.007
STAT3.009.050
CCNA2.009
ICAM1.014YY
BCL2L1.015
MAP2K6.044.016
IGF1.021
SOD1.033Y
TGIF.047
BMP6Y
GSTM3Y

Abbreviation: Y, gene predictive of recurrence and/or overall survival by recursive partitioning analysis.

*Only P ≤ .050 shown.
Able to significantly predict recurrence and overall survival by log-rank and/or recursive partitioning analysis and thus selected for the common four-gene panel.

Footnotes

Supported by National Institutes of Health Grant No. CA-86871 and National Cancer Institute Grant No. CA-14089.

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.

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The author(s) indicated no potential conflicts of interest.

AUTHOR CONTRIBUTIONS

Conception and design: Anirban P. Mitra, Vincenzo Pagliarulo, Ram H. Datar, Susan Groshen, Richard J. Cote

Financial support: Donald G. Skinner, Richard J. Cote

Administrative support: Anirban P. Mitra, Frederic M. Waldman, Ram H. Datar, Donald G. Skinner, Susan Groshen, Richard J. Cote

Provision of study materials or patients: Frederic M. Waldman, Donald G. Skinner, Richard J. Cote

Collection and assembly of data: Anirban P. Mitra, Vincenzo Pagliarulo

Data analysis and interpretation: Anirban P. Mitra, Dongyun Yang, Susan Groshen, Richard J. Cote

Manuscript writing: Anirban P. Mitra, Dongyun Yang, Susan Groshen, Richard J. Cote

Final approval of manuscript: Anirban P. Mitra, Vincenzo Pagliarulo, Dongyun Yang, Frederic M. Waldman, Ram H. Datar, Susan Groshen, Richard J. Cote

REFERENCES

1. Greene FL, Page DL, Fleming ID, et al., editors. AJCC Cancer Staging Manual. ed 6. New York, NY: Springer-Verlag; 2002. Urinary bladder; pp. 367–374.
2. Mitra AP, Datar RH, Cote RJ. Molecular staging of bladder cancer. BJU Int. 2005;96:7–12. [PubMed]
3. Brambilla C, Fievet F, Jeanmart M, et al. Early detection of lung cancer: Role of biomarkers. Eur Respir J Suppl. 2003;39:36s–44s. [PubMed]
4. Northup JK, Gadre SA, Ge Y, et al. Do cytogenetic abnormalities precede morphologic abnormalities in a developing malignant condition? Eur J Haematol. 2007;78:152–156. [PubMed]
5. Mitra AP, Datar RH, Cote RJ. Molecular pathways in invasive bladder cancer: New insights into mechanisms, progression, and target identification. J Clin Oncol. 2006;24:5552–5564. [PubMed]
6. Willey JC, Crawford EL, Jackson CM, et al. Expression measurement of many genes simultaneously by quantitative RT-PCR using standardized mixtures of competitive templates. Am J Respir Cell Mol Biol. 1998;19:6–17. [PubMed]
7. Pagliarulo V, George B, Beil SJ, et al. Sensitivity and reproducibility of standardized-competitive RT-PCR for transcript quantification and its comparison with real time RT-PCR. Mol Cancer. 2004;3:5. [PMC free article] [PubMed]
8. Mitra AP, Almal AA, George B, et al. The use of genetic programming in the analysis of quantitative gene expression profiles for identification of nodal status in bladder cancer. BMC Cancer. 2006;6:159. [PMC free article] [PubMed]
9. Miller RG. Survival Analysis: Wiley Series in Probability and Mathematical Statistics. New York, NY: John Wiley & Sons; 1981. pp. 44–102.
10. Berry G, Kitchin RM, Mock PA. A comparison of two simple hazard ratio estimators based on the logrank test. Stat Med. 1991;10:749–755. [PubMed]
11. Harrell FE, Jr, Lee KL, Mark DB. Multivariable prognostic models: Issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15:361–387. [PubMed]
12. Chen CH, George SL. The bootstrap and identification of prognostic factors via Cox's proportional hazards regression model. Stat Med. 1985;4:39–46. [PubMed]
13. Altman DG, Andersen PK. Bootstrap investigation of the stability of a Cox regression model. Stat Med. 1989;8:771–783. [PubMed]
14. Breiman L, Friedman JH, Olshen RA, et al. Classification and Regression Trees. Belmont, CA: Wadsworth International Group; 1984.
15. Therneau TM, Atkinson EJ. An Introduction to Recursive Partitioning Using the RPART Routines: Mayo Clinic Biostatistics Technical Report. Rochester, MN: Mayo Foundation; 1997.
16. Akaike H. A new look at the statistical model identification. IEEE Trans Automat Contr. 1974;AC-19:716–723.
17. Burnham KP, Anderson DR, editors. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. ed 2. New York, NY: Springer-Verlag; 2002. Formal inference from more than one model: Multimodel inference (MMI) pp. 149–205.
18. Dijkstra EW. A note on two problems in connexion with graphs. Numer Math. 1959;1:269–271.
19. Sanchez-Carbayo M, Socci ND, Lozano J, et al. Defining molecular profiles of poor outcome in patients with invasive bladder cancer using oligonucleotide microarrays. J Clin Oncol. 2006;24:778–789. [PubMed]
20. Braakhuis BJ, Tabor MP, Kummer JA, et al. A genetic explanation of Slaughter's concept of field cancerization: Evidence and clinical implications. Cancer Res. 2003;63:1727–1730. [PubMed]
21. Affymetrix. Technical note: Array design for the GeneChip Human Genome U133 Set. http://www.affymetrix.com/support/technical/technotes/hgu133_design_technote.pdf.
22. Tiniakos DG, Mellon K, Anderson JJ, et al. c-jun oncogene expression in transitional cell carcinoma of the urinary bladder. Br J Urol. 1994;74:757–761. [PubMed]
23. Itoh M, Murata T, Suzuki T, et al. Requirement of STAT3 activation for maximal collagenase-1 (MMP-1) induction by epidermal growth factor and malignant characteristics in T24 bladder cancer cells. Oncogene. 2006;25:1195–1204. [PubMed]
24. Schultz IJ, Wester K, Straatman H, et al. Prediction of recurrence in Ta urothelial cell carcinoma by real-time quantitative PCR analysis: A microarray validation study. Int J Cancer. 2006;119:1915–1919. [PubMed]
25. Kuo WP, Jenssen TK, Butte AJ, et al. Analysis of matched mRNA measurements from two different microarray technologies. Bioinformatics. 2002;18:405–412. [PubMed]
26. Zhao H, Grossman HB, Spitz MR, et al. Plasma levels of insulin-like growth factor-1 and binding protein-3, and their association with bladder cancer risk. J Urol. 2003;169:714–717. [PubMed]
27. Hussain SA, Ganesan R, Hiller L, et al. BCL2 expression predicts survival in patients receiving synchronous chemoradiotherapy in advanced transitional cell carcinoma of the bladder. Oncol Rep. 2003;10:571–576. [PubMed]
28. Blaveri E, Simko JP, Korkola JE, et al. Bladder cancer outcome and subtype classification by gene expression. Clin Cancer Res. 2005;11:4044–4055. [PubMed]
29. Korkolopoulou P, Lazaris A, Konstantinidou AE, et al. Differential expression of bcl-2 family proteins in bladder carcinomas: Relationship with apoptotic rate and survival. Eur Urol. 2002;41:274–283. [PubMed]
30. Nakakuki K, Imoto I, Pimkhaokham A, et al. Novel targets for the 18p11.3 amplification frequently observed in esophageal squamous cell carcinomas. Carcinogenesis. 2002;23:19–24. [PubMed]
31. Kraemer K, Schmidt U, Fuessel S, et al. Microarray analyses in bladder cancer cells: Inhibition of hTERT expression down-regulates EGFR. Int J Cancer. 2006;119:1276–1284. [PubMed]
32. Gonzales R, Auclair C, Voisin E, et al. Superoxide dismutase, catalase, and glutathione peroxidase in red blood cells from patients with malignant diseases. Cancer Res. 1984;44:4137–4139. [PubMed]
33. Steele LP, Georgopoulos NT, Southgate J, et al. Differential susceptibility to TRAIL of normal versus malignant human urothelial cells. Cell Death Differ. 2006;13:1564–1576. [PubMed]
34. Ren R, Charles PC, Zhang C, et al. Gene expression profiles identify a role for cyclooxygenase 2-dependent prostanoid generation in BMP6-induced angiogenic responses. Blood. 2007;109:2847–2853. [PubMed]
35. Schnakenberg E, Breuer R, Werdin R, et al. Susceptibility genes: GSTM1 and GSTM3 as genetic risk factors in bladder cancer. Cytogenet Cell Genet. 2000;91:234–238. [PubMed]
36. Chatterjee SJ, Datar R, Youssefzadeh D, et al. Combined effects of p53, p21, and pRb expression in the progression of bladder transitional cell carcinoma. J Clin Oncol. 2004;22:1007–1013. [PubMed]
37. Shariat SF, Tokunaga H, Zhou J, et al. p53, p21, pRB, and p16 expression predict clinical outcome in cystectomy with bladder cancer. J Clin Oncol. 2004;22:1014–1024. [PubMed]

Articles from Journal of Clinical Oncology are provided here courtesy of American Society of Clinical Oncology