Using Protein Expressions to Predict Survival in Clear Cell Renal Carcinoma

5464 Vol. 10, 5464–5471, August 15, 2004 Clinical Cancer Research

؍ Hyung L. Kim,1 David Seligson,2 Xueli Liu,3 than that of prognostic models based on grade alone (C or the University of ,(0.73 ؍ Nicolette Janzen,1 Matthew H. T. Bui,1 Hong Yu,2 0.65), TNM stage alone (C ؍ Tao Shi,4 Robert A. Figlin,5 Steve Horvath,4 and California Los Angeles integrated staging system (C 0.76). 1 Conclusions: Protein expressions obtained using widely Arie S. Belldegrun available technology can complement standard clinical pre- 1 2 Departments of Urology, Pathology and Laboratory Medicine, dictors such as TNM stage, histological grade, and PS. 3Biostatistics, 4Human Genetics and Biostatistics, and 5Medicine and Urology, University of California School of Medicine, Los Angeles, California INTRODUCTION Clear cell carcinomas of the kidney represent over 80% of ABSTRACT renal cell carcinomas [RCCs (1)]. The treatment of choice for Purpose: An accurate system for predicting survival for localized RCC is a nephrectomy; however, approximately 20% patients with solid tumors will allow for better patient se- of all patients surgically treated with curative intent will ulti- lection for both established and novel therapies. We propose mately experience disease recurrence (2, 3). Furthermore, ap- a staging system for clear cell variants of renal cell carci- proximately 30% of patients will present with metastatic dis- noma (RCC) that includes molecular predictors and stand- ease. RCC is generally resistant to conventional oncologic ard clinical predictors such as tumor-node-metastasis therapies, and patients with systemic disease are treated with (TNM) stage, histological grade, and performance status immunotherapy, which produces response rates of approxi- (PS). mately 15–30% (4–7). As a result of the modest response rates, Experimental Design: A custom tissue array was con- RCC has become a model for testing novel immunotherapeutic structed using clear cell RCC from 318 patients, represent- strategies. An enhanced ability to predict patient survival will ing all stages of localized and metastatic RCC, and immu- allow for better selection of patients most likely to benefit from nohistochemically stained for molecular markers Ki67, p53, systemic therapies and for more accurate comparison of clinical gelsolin, CA9, CA12, PTEN, EpCAM, and vimentin. We trials based on varying inclusion criteria. present a strategy for evaluating individual candidate mark- In general, prior attempts to predict patient survival have ers for prognostic information and integrating informative relied on traditional clinical parameters such as tumor stage and markers into a multivariate prognostic system. grade (3, 8). More recently, methods based on gene arrays, Results: The overall median follow-up and the median which screen for differential expression of thousands of genes, follow-up for surviving patients were 28 and 55 months, have identified large numbers of potential prognostic markers respectively. A prognostic model based primarily on molec- (9, 10). Our study, evaluating protein expression in a high- ular markers included metastasis status, p53, CA9, gelsolin, throughput tissue array, is a natural extension to the efforts for and vimentin as predictors and had high discriminatory molecular staging. All of the markers examined in this study power: its statistically validated concordance index (C- were selected based on previous reports linking the markers to index) was found to be 0.75. A prognostic model based on a the development of malignancies. Ki67 and p53 are related to combination of clinical and molecular predictors included cellular proliferation (11–13). In RCC, Ki67 is an independent metastasis status, T stage, Eastern Cooperative Oncology predictor of survival (11, 12), and a mutated p53 has been Group PS, p53, CA9, and vimentin as predictors and had a shown to be an independent predictor of survival in some C-index of 0.79, which was significantly higher (P < 0.05) studies (14, 15). Gelsolin, EpCAM (epithelial cell adhesion molecule), and vimentin may be involved in cell motility and cancer progres- sion. Gelsolin is the most potent protein known that functions to Received 3/10/04; revised 5/3/04; accepted 5/10/04. sever actin during cell motility (16), and it has been described as Grant support: Supported in part by National Cancer Institute 2 P30 a highly significant indicator of poor prognosis in non-small-cell CA16042-29 through the Jonsson Comprehensive Cancer Center at lung cancer (17). EpCAM is widely expressed on the surface of University of California Los Angeles. many carcinomas (18, 19). Vimentin, an intermediate filament, The costs of publication of this article were defrayed in part by the has previously been identified as an independent predictor of payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to poor prognosis in RCC (20, 21). indicate this fact. CA9, CA12, and PTEN (phosphatase and tensin homo- Note: H. Kim, D. Seligson, S. Horvath, and A. Belldegrun contributed logue deleted from chromosome 10) are critical components of equally to this work. the hypoxia pathway, which allows enlarging tumors to adapt to Requests for reprints: Arie S. Belldegrun, University of California School of Medicine, Department of Urology, 10833 Le Conte Avenue, an oxygen-poor microenvironment. In addition, overexpression Room 66-118 CHS, Los Angeles, CA 90095-1738. Phone: (310) 206- of CA9 and CA12 is a direct consequence of a VHL mutation, 1434; Fax: (310) 206-5343; E-mail: [email protected]. found in over 75% of sporadic clear cell RCC (22, 23). De-

creased expression of CA9 has previously been shown to predict performed with monoclonal mouse antibody GS-2C4 (Sigma, worse survival (24). PTEN is a phosphatase that regulates cel- St. Louis, MO) at 3.8 ␮g/ml. Immunostaining for PTEN was lular migration, proliferation, and apoptosis (25, 26). Although performed using rabbit polyclonal antibody PN37 (Zymed, PTEN mutation may be a rare event in RCC (27, 28), PTEN South San Francisco, CA) at 2 ␮g/ml. Immunostaining for deletion has been shown to correlate with poor prognosis (27). EpCAM was performed using monoclonal mouse antibody With the explosive growth of available genomic and pro- KS1/4BD (PharMingen, San Diego, CA) at 20 ␮g/ml. Immu- teomic data, the number of markers that have been correlated nostaining for p53 was performed with monoclonal mouse an- with prognosis is continuously increasing. We present a strategy tibody D0–7 (DAKO) at a 1:100 dilution. CA9 staining was for evaluating individual candidate markers for prognostic in- performed with monoclonal mouse antibody M75 (a gift from formation and integrating informative markers into a multivari- Dr. Eric Stanbridge, University of California at Irvine, Irvine, ate prognostic system. Levels of marker expression were used to CA) at a 1:25,000 dilution of stock. CA12 staining was per- develop two prognostic models for predicting disease-specific formed with a rabbit polyclonal antibody (a gift from Dr. Mi- survival (DSS) in patients with clear cell RCC. The accuracy of chael Lerman, National Cancer Institute, Bethesda, MD) at a these models was compared with traditional, clinical parameters 1:450 dilution. Vimentin staining was performed with mouse used for predicting survival. monoclonal antibody V9 (DAKO) at a 1:1000 dilution. Immunostaining was scored by recording the percentage of cells staining and scoring the area of maximum staining on a MATERIALS AND METHODS 4-point scale, with 0 representing no staining and 3 representing Patients. The patient cohort included 318 patients treated the highest staining. All three cores that were arrayed per tumor with a partial or radical nephrectomy for clear cell RCC between were scored and pooled to yield the mean, median, minimum, 1989 and 1999. After approval by the institutional review board and maximum values for the three cores. (KCP 99-233), immunohistochemical studies were performed, Statistical Analysis. Length of follow-up was the time and clinical data from an established kidney cancer database from nephrectomy to date of death or last contact. The end point were reviewed. Patients were staged using radiographic studies of interest was DSS. Kaplan-Meier curves were generated to and postoperative pathological data, according to the 1997 visualize survival rates, and we used Cox proportional hazards tumor-node-metastasis (TNM) criteria proposed by the Ameri- models to relate DSS to molecular and clinical predictors. The can Joint Committee on Cancer (29). Performance status (PS) Scho¨enfeld residual test was used to evaluate the proportional was determined using the Eastern Cooperative Oncology Group hazard assumptions. All computations were performed with the (ECOG) PS scale (30). Patients with M stage ϭ 1 or N stage Ͼ R statistical software using the Design, Hmisc, Rpart, and Ran- 0 were considered to have metastatic disease. Tumor grade was domForest libraries. P Ͻ 0.05 was considered significant, and categorized using Fuhrman grade (31). The patients were also P Ͼ 0.10 was the criterion when performing Cox regression categorized according to the University of California Los An- backward step-down variable deletions. geles integrated staging system [UISS (8)]. The predictive accuracy of various Cox regression models After surgery, patients were evaluated for disease recur- was quantified by calculating the concordance index (C-index), rence by physical examination, liver function tests, chest X-ray, which provides the area under the receiver operating character- and computerized tomography of the abdomen/pelvis every istics curve for censored data (34, 35). A C-index of 0.5 indi- 6–12 months. The primary outcome of interest was DSS. Pa- cates that outcomes are completely random, whereas a C-index tients with metastatic RCC were treated with cytoreductive of 1 indicates that the model is a perfect predictor. To protect nephrectomy (n ϭ 155) followed by interleukin-2-based immu- against overfitting during stepwise regression, we used a boot- notherapy (n ϭ 116) and then evaluated radiologically in 3 strap procedure as implemented in the “validate” function of the months to determine the response to therapy. A positive re- Design library (34), which allowed for computation of an un- sponse was defined as a complete or partial response. biased estimate of the C-index. We used 500 bootstrap samples. Tissue Array Construction. Archival tumor specimens To test whether the difference in statistical accuracy be- from the cohort of 318 patients were obtained from the Depart- tween non-nested Cox regression models is significant, we used ment of Pathology. All tumors were reviewed to confirm clear the rcorrp.cens function in the Design library (34). This com- cell histology. Three core tissue biopsies, each 0.6 mm in putes U-statistics for testing whether the predictions of one diameter, were taken from selected, morphologically represent- model are more concordant than those of another model, ex- ative regions of each paraffin-embedded renal tumor and pre- tending the C-index. Specifically, the fraction of pairs was cisely arrayed using a custom-built instrument as described determined for which one model correctly selects the patient previously (32). with the longer DSS when the competing model did not. Before Immunohistochemistry. Array sections were deparaf- calculating the C-index or using the rcorrp.cens function, TNM finized and immunostained using either the DAKO Envision staging system (T stage, N stage, and M stage) and UISS (T Plus (DAKO, Carpinteria, CA) or Vector ABC Elite (Vector stage, N stage, M stage, ECOG PS, and grade) were fitted to our Laboratories, Burlingame, CA) staining systems as described data. previously (33). Antigen retrieval involved pepsin digestion for To account for potential overfitting resulting from selection 10 min for EpCAM staining, treatment in a pressure cooker for of the staining criterion and cutoff used to stratify DSS, we used 3 min for Ki67, and heat treatment for 25 min for all other a prevalidation method described in Liu et al. (36) Briefly, the markers. Ki67 staining was performed with mouse monoclonal prevalidation method is carried out in three steps for each antibody MIB-1 (DAKO) at 0.5 ␮g/ml. Gelsolin staining was marker. First, an intercept only Cox regression model is fit to the

survival times to arrive at deviance residuals, which are normal- Table 1 Characteristics of patients with clear cell RCC (N ϭ 318) ized transforms of martingale residuals. Second, random regres- Patient characteristics sion forest predictors are fit where the outcome is the deviance Age (yrs) residual and the covariates are the pooled marker stainings (37). Mean 60 Random regression forest predictors work well in the setting of Median 61 many covariates but relatively few observations. Third, the Range 27–88 random forest construction is used to define a tumor marker Sex hazard score M as the out-of-bag estimate of the deviance Male 215 Female 103 residual. The hazard score M is a real number. Mean tumor size (cm) 7.4 Liu et al. (36) have shown that when M is used as a Metastasis covariate in a Cox regression model involving the same survival Yes 155 outcome that was used to construct it, it does not lead to an No 163 1997 TNM stage inflation of the false positive rate of the one-sided Wald test that I90 tests whether coefficient is positive. Therefore, we used the II 19 corresponding M for each marker as the bias-corrected marker III 54 covariate in all univariate and multivariate Cox regression anal- IV 155 yses in computing the C-index and in the rcorrp.cens function. Grade 137 The Cox proportional hazards regression analysis was used 2 156 to construct two predictive models. One model was primarily 3 113 based on molecular markers (marker model). A second predic- 412 tive model was based on a combination of molecular markers ECOG PS 0 113 and clinical variables (clinical/marker model). We showed pre- 1 191 viously (24) that CA9 is an independent predictor for survival in 213 patients with metastatic RCC, but not localized RCC. Therefore, 31 an interaction term for CA9 and metastasis status (Met*CA9) IL2-based therapy* was used. Response 38 No response 78 To visualize the relationship between clinical predictors Deaths 186 and DSS, we constructed a nomogram for a Cox model that only Median follow-up (mo) contains significant terms and dichotomized marker stainings. All patients 28 Although dichotomized marker values were used, only markers Surviving patients 55 UISS shown to contain prognostic information using the prevalidation 172 method were included in the nomogram. This nomogram should 257 be considered as an approximation of the validated models 352 described above. In particular, the nomogram is not used for 4 128 model comparisons and inferences as described above. Rather, 59 the nomogram is a descriptive tool that needs further validation * IL2, interleukin-2; response includes complete and partial re- with new and independent data. sponses. For each molecular marker, the optimal cutoff for the staining scores to stratify DSS was determined using the default settings of the recursive partitioning function (RPART) in the Below, the use of dichotomized staining scores is discussed in freely available R statistical software.6 To avoid cutoffs that are more detail. overly sensitive to assay conditions, the optimal cutoffs were confirmed to be robust before they were used to construct the RESULTS prognostic nomograms. Ki67 was considered positive if there The patient characteristics are summarized in Table 1. The was Ͼ15% nuclear staining at any intensity in the core with mean age is 60 years, and the male:female ratio is approximately median staining, and p53 was positive if there was Ͼ15% 2:1. The mean tumor size is 7.4 cm, and 155 (49%) patients staining in the core with maximum staining. Gelsolin was pos- presented with metastatic disease. The overall median follow-up itive if there was any level of cytoplasmic staining in any core. is 28 months; 186 patients have died, and the median follow-up PTEN was positive if there was Ͼ50% cytoplasmic staining in for surviving patients is 55 months. UISS is a prognostic model the core with lowest staining. EpCAM was considered positive that predicts DSS based on the interaction of TNM stage, if there was any staining in any core. CA9 and CA12 were Fuhrman grade, and ECOG PS (8). The most common UISS considered positive if there was 100% staining in all cores and categories were group 4 (n ϭ 128; 40%) and group 1 (n ϭ 72; the mean percentage staining in the three cores was Ͼ80%, 23%). In a single variable analysis, increased immunohisto- respectively. Vimentin was considered positive if the mean of chemical staining for Ki67, p53, vimentin, and gelsolin corre- the maximum staining intensity score in the three cores was Ͻ2. lated with worse survival, whereas the inverse was true for CA9, PTEN, CA12, and EpCAM, with decreased staining correlating with worse survival. However, only gelsolin, p53, CA9, Ki67, and vimentin were statistically significant predictors of DSS in 6 http://www.r-project.org/. a univariate analysis (Table 2, left).

Table 2 Cox regression analysis for molecular markers Multivariable analysis† Multivariable analysis‡ Univariate analysis* (markers) (clinical variables) Hazard ratio Hazard ratio Hazard ratio (95% CI) P (95% CI) P (95% CI) P Gelsolin 4.35 (1.95–9.69) Ͻ0.001 2.62 (1.10–6.23) 0.030 1.87 (0.82–4.27) 0.140 p53 3.25 (1.83–5.76) Ͻ0.001 2.28 (1.19–4.35) 0.014 2.03 (1.11–3.68) 0.021 CA9 2.89 (1.38–6.03) 0.006 3.17 (1.43–7.05) 0.004 2.52 (1.18–5.38) 0.017 Ki67 2.78 (1.76–4.39) Ͻ0.001 0.95 (0.55–1.63) 0.850 1.35 (0.82–2.23) 0.240 Vimentin 2.36 (1.10–5.09) 0.028 2.19 (0.91–5.24) 0.080 2.48 (1.12–5.50) 0.026 CA12 2.25 (0.98–5.19) 0.058 0.84 (0.31–2.28) 0.730 2.12 (0.85–5.29) 0.110 EpCAM 1.66 (0.55–5.05) 0.380 0.59 (0.18–1.97) 0.400 0.67 (0.21–2.12) 0.490 PTEN 1.42 (0.53–3.82) 0.480 2.19 (0.79–6.03) 0.130 1.66 (0.60–4.62) 0.330 Abbreviation: CI, confidence interval. * Decreased staining for CA9, CA12, PTEN, and EpCAM is associated with worse survival. Otherwise, increased staining is associated with worse survival. † Multivariable analysis includes all markers and metastasis status as a covariate and as an interaction term for CA9. Metastasis status: M ϭ 1 Ͼ or N 0 versus N0M0. ‡ Multivariable analysis with each marker variable separately modeled with clinical variables that include metastasis status, T stage, ECOG, and Fuhrman grade.

Table 3 Reduced multivariable Cox regression analysis for clinical calculating the C-index, which is the area under the receiver variables and markers operating characteristics curve adapted for survival data. (Table % Retained in 4). The C-index was corrected for overfitting by a bootstrap Hazard ratio 95% CI P bootstrap* procedure. A C-index of 0.5 indicates that outcomes are com- Met† 4.30 2.72–6.81 Ͻ0.001 100 pletely random, whereas a C-index of 1 indicates that the model T stage 1.57 1.25–1.98 Ͻ0.001 91 is a perfect predictor. When using clinical variables alone, TNM Ͻ ECOG 1.84 1.37–2.46 0.001 95 stage and UISS had a C-index of 0.73 and 0.75, respectively. Met*CA9 2.45 1.15–5.20 0.020 50 Vimentin 2.30 1.01–5.21 0.047 44 The marker model and the clinical/marker model had a C-index p53 1.75 0.96–3.19 0.070 30 of 0.75 and 0.79, respectively. Abbreviations: CI, confidence interval; Met, metastasis status. Although the C-index allows various models to be ranked * Percent of bootstrapped samples in which the covariate was retained by a backward stepwise elimination. ϭ Ͼ † Metastasis status: M 1orN 0 versus N0M0. Table 4 Ranking of prognostic models Corrected Prognostic models C-index* All markers were combined in a multivariable Cox regression analysis that included metastasis status as a covariate (Met) Grade 0.65 ECOG 0.66 and as an interaction term for CA9. (Table 2, center) Met, TNM stage 0.73 gelsolin, p53, and Met*CA9 remained significant predictors of UISS 0.75 survival and were used to create a prognostic model (marker Molecular staging model† 0.75 model). Covariates included in the marker model were retained Clinical/molecular model‡ 0.79 in over 50% of the samples in a bootstrapping procedure that * C-index corrected based on bootstrap bias. included a backward stepwise selection of candidate predictors. † Covariates include Met, Met*CA9, gelsolin, and p53. Interactions between metastasis status and other markers as well ‡ Covariates include Met, T stage, ECOG PS, Met*CA9, vimentin, and p53. as interactions between markers were examined; however, none of these variables were significant at a level of 0.10. Using a similar approach, a prognostic model was constructed using a combination of clinical variables and marker Table 5 Comparison of prognostic models: Percentage of patient data (clinical/marker model). In a multivariable Cox analysis, pairs in which one model correctly predicts outcome and the other model does not when each of the markers were controlled for T stage, Met, ECOG PS, and grade, we found that CA9, vimentin, and p53 Model 1 vs. model 2 Model 1 Model 2 P were statistically significant predictors independent of the clin- Marker model* vs. stage 14 12 0.152 ical variables. (Table 2, right). These three markers were com- Marker model* vs. UISS 9 11 0.143 Clinical/marker† vs. stage 14 8 Ͻ0.001 bined in a multivariable Cox analysis that included Met, T stage, Clinical/marker† vs. UISS 7 5 0.038 ECOG PS, and grade. Covariates retained in a Cox step-down UISS vs. stage 9 5 Ͻ0.001 analysis (Table 3) were used to construct a second prognostic * Covariates include Met, Met*CA9, gelsolin, and p53. model (clinical/marker model). † Covariates include Met, T stage, ECOG PS, Met*CA9, vimentin, The predictive ability of various models was quantified by and p53.

Fig. 1 A, nomogram for predicting disease-specific survival using molecular markers. To read the nomogram, draw a vertical line from each tick marker indicating the status of a predictor to the top axis labeled Points. Sum the points and find the corresponding number on the axis labeled Total Points. Draw a vertical line down to the axes showing 2- and 4-year DSS rates and median survival. The nomogram has been cali- brated to provide 2- and 4-year survival rates accurate to within 10%. B, Kaplan-Meier survival curves based on stratification using the total points assigned by the nomogram. The P values comparing adjacent curves using the Wilcoxon test and the number of patients in each group are given.

according to accuracy, it cannot be used for hypothesis testing. DISCUSSION Therefore, we performed a test for concordance to hypothesis To our knowledge, this is the first study of any solid tumor test and obtain P values. When the various models were com- to demonstrate that prognostic models based primarily on pro- pared for concordance, the clinical/marker model was signifi- tein expression profiles can perform at least as well as TNM Ͻ ϭ cantly better than TNM stage or UISS (P 0.001 and P stage, histological grade, or PS. An accurate system for predict- 0.038, respectively; Table 5). The differences between the ing survival is useful for patient counseling, planning follow-up, marker model versus TNM stage (P ϭ 0.152) and marker model and selecting patients for additional treatment. For clinically versus UISS (P ϭ 0.143) were not statistically significant. localized RCC, patients at high risk for recurrence can be The factors included in the marker model (Met, gelsolin, selected for adjuvant therapy trials. For metastatic RCC, low- p53, Met*CA9, and vimentin) were used to construct a prognostic nomogram. (Fig. 1A). Fig. 1B illustrates the stratifica- and moderate-risk patients are good candidates for standard tion of the survival curves based on total points assigned by immunotherapy. However, high-risk patients may be selected the nomogram, and P values comparing adjacent survival for immediate enrollment into clinical trials. In the setting of a curves are indicated. These P values should be considered as clinical trial, having accurate prognostic information helps en- descriptive measures of curve separation. Similarly, factors sure a consistent population of study patients, which will facil- included in the clinical/molecular model (Met, T stage, itate interpretation of outcomes. ECOG PS, Met*CA9, p53, and vimentin) were used to con- The predictive accuracy of our marker model for RCC was struct a second nomogram (Fig. 2A). Fig. 2B illustrates the comparable with UISS, which is a model combining standard stratification of the survival curves based on total points clinical predictors; and the clinical/marker model was signifi- assigned by the second nomogram. cantly more accurate than UISS. Molecular information may

Fig. 2 A, nomogram for predicting DSS survival using clinical data and molecular markers. See legend for Fig. 1A. B, Kaplan-Meier survival curves based on stratification using the total points assigned by the nomogram. See legend for Fig. 1C.

reduce some of the subjectivity involved in current clinical approach can be applied to building and evaluating predictive methods for predicting survival. Although models based on gene models for other solid tumors. In this study, the various models expression profiles have been described (9, 38), the technology were ranked based on predictive accuracy as measured by the for gene expression profiling is costly and not yet widely avail- concordance index. The accuracy of the predictive models was able. The hope is that protein expression profiling can be more also compared by examining the concordance of the models for rapidly incorporated into a clinical setting using resources avail- combinations of patient pairs. Using this approach, both the able at many centers. Immunohistochemical staining is already clinical/marker model and the marker model were significantly widely used for pathological analysis of breast and colon cancer more accurate than histological grade, and the clinical/marker tissue. However, careful multicenter clinical studies involving model was significantly more accurate than grade, TNM stage, standardized marker staining protocols should be used to vali- and UISS. date our findings. Overfitting occurs when random patterns in the data are Our study shows that accurate models for molecular stag- incorporated into a prognostic model as meaningful information, ing of a solid tumor can be developed using a very limited resulting in what appears to be a more accurate model. How- number of markers. We demonstrate the use of a statistical ever, any apparent improvement in model performance due to technique for constructing prognostic models and comparing overfitting is lost when the model is applied to an independent them with established systems for predicting survival. This dataset. The complexity of the training dataset used to build the

model correlates with the risk of overfitting. In our study, every 9. van de Vijver MJ, He YD, van’t Veer LJ, et al. A gene-expression step of the modeling process was validated and adjusted for signature as a predictor of survival in breast cancer. N Engl J Med overfitting. Therefore, when the various models are ranked 2002;347:1999–2009. according to predictive accuracy or compared to obtain P val- 10. Takahashi M, Sugimura J, Yang X, et al. Gene expression profiling of renal cell carcinoma and its implications in diagnosis, prognosis, and ues, the significance of our results is not overstated. This helps therapeutics. Adv Cancer Res 2003;89:157–81. to ensure that our models will perform similarly in an external 11. Delahunt B, Bethwaite PB, Thornton A, Ribas JL. Proliferation of dataset as it did in our dataset. renal cell carcinoma assessed by fixation-resistant polyclonal Ki-67 When evaluating a marker by immunohistochemistry, it is antibody labeling. Correlation with clinical outcome. Cancer (Phila) common to dichotomize the staining data using a single cutoff to 1995;75:2714–9. determine positive and negative staining. Dichotomizing the 12. Rioux-Leclercq N, Turlin B, Bansard J, et al. Value of immunohis- staining simplifies the staining analysis and enhances interob- tochemical Ki-67 and p53 determinations as predictive factors of out- server reproducibility. However, when performing the statistical come in renal cell carcinoma. Urology 2000;55:501–5. analysis, the selection of cutoff criterion can introduce bias and 13. Maxwell SA, Rivera A. Proline oxidase induces apoptosis in tumor cells, and its expression is frequently absent or reduced in renal carci- result in overfitting. Therefore, we describe a prevalidation nomas. J Biol Chem 2003;278:9784–9. technique for each marker that takes into account all measures 14. Shiina H, Igawa M, Urakami S, et al. Clinical significance of of staining (such as intensity and percentage of staining), cor- immunohistochemically detectable p53 protein in renal cell carcinoma. rects for overfitting, and quantifies the true prognostic informa- Eur Urol 1997;31:73–80. tion contained in the staining. This prevalidation procedure 15. Uchida T, Gao JP, Wang C, et al. Clinical significance of p53, leads to unbiased P values and C-indices. Only markers that mdm2, and bcl-2 proteins in renal cell carcinoma. Urology 2002;59: were significant predictors of survival on prevalidation were 615–20. dichotomized and used to construct nomograms. Finally, to 16. Selden LA, Kinosian HJ, Newman J, et al. Severing of F-actin by the amino-terminal half of gelsolin suggests internal cooperativity in avoid cutoffs that are overly sensitive to the staining conditions, gelsolin. Biophys J 1998;75:3092–100. only robust cutoffs were used. 17. Shieh DB, Godleski J, Herndon JE II, et al. Cell motility as a We propose two nomograms that can be used to predict prognostic factor in stage I nonsmall cell lung carcinoma: the role of DSS. One nomogram is based on metastasis status and three gelsolin expression. Cancer (Phila) 1999;85:47–57. molecular markers (CA9, p53, and vimentin). The second nom- 18. Shetye J, Christensson B, Rubio C, et al. The tumor-associated ogram combines clinical and molecular variables (Met, T stage, antigens BR55-2, GA73-3 and GICA 19-9 in normal and corresponding neoplastic human tissues, especially gastrointestinal tissues. Anticancer ECOG PS, CA9, p53, and vimentin). By including metastasis Res 1989;9:395–404. status in the nomogram, we found that the nomogram accurately 19. Riethmuller G, Schneider-Gadicke E, Schlimok G, et al. Random- predicts DSS in both patients with localized RCC and those with ised trial of monoclonal antibody for adjuvant therapy of resected metastatic RCC. Both nomograms can be used to calculate 2- Dukes’ C colorectal carcinoma. German Cancer Aid 17–1A Study and 4-year DSS rates as well as median survival. Although these Group. Lancet 1994;343:1177–83. nomograms are useful for visualizing our predictive models, 20. Moch H, Schraml P, Bubendorf L, et al. High-throughput tissue they need to be tested on independent patient populations before microarray analysis to evaluate genes uncovered by cDNA microarray screening in renal cell carcinoma. Am J Pathol 1999;154:981–6. being applied to patient care. 21. Sabo E, Miselevich I, Bejar J, et al. The role of vimentin expression in predicting the long-term outcome of patients with localized renal cell REFERENCES carcinoma. Br J Urol 1997;80:864–8. 1. Belldegrun A, deKernion JB. Renal tumors. In: Walsh PC, Retik AB, 22. Gnarra JR, Tory K, Weng Y, et al. Mutations of the VHL tumour Vaughan ED, Wein AJ, editors. Campbell’s urology, 7th ed. Philadel- suppressor gene in renal carcinoma. Nat Genet 1994;7:85–90. phia: W. B. Saunders Co.; 1998. p. 2283–326. 23. Maxwell PH, Wiesener MS, Chang GW, et al. The tumour suppres- 2. Vogelzang NJ, Stadler WM. Kidney cancer. Lancet 1998;352: sor protein VHL targets hypoxia-inducible factors for oxygen-dependent 1691–6. proteolysis. Nature (Lond) 1999;399:271–5. 3. Kattan MW, Reuter V, Motzer RJ, Katz J, Russo P. A postoperative 24. Bui MH, Seligson D, Han KR, et al. Carbonic anhydrase IX is an prognostic nomogram for renal cell carcinoma. J Urol 2001;166:63–7. independent predictor of survival in advanced renal clear cell carci- 4. Pyrhonen S, Salminen E, Ruutu M, et al. Prospective randomized noma: implications for prognosis and therapy. Clin Cancer Res 2003; trial of interferon alfa-2a plus vinblastine versus vinblastine alone in 9:802–11. patients with advanced renal cell cancer. J Clin Oncol 1999;17:2859– 25. Steck PA, Pershouse MA, Jasser SA, et al. Identification of a 67. candidate tumour suppressor gene, MMAC1, at chromosome 10q23.3 5. Fisher RI, Rosenberg SA, Fyfe G. Long-term survival update for that is mutated in multiple advanced cancers. Nat Genet 1997;15:356– high-dose recombinant interleukin-2 in patients with renal cell carci- 62. noma. Cancer J Sci Am 2000;6(Suppl 1):S55–7. 26. Maehama T, Dixon JE. The tumor suppressor, PTEN/MMAC1, 6. Rosenberg SA, Lotze MT, Muul LM, et al. Observations on the dephosphorylates the lipid second messenger, phosphatidylinositol systemic administration of autologous lymphokine-activated killer cells 3,4,5-trisphosphate. J Biol Chem 1998;273:13375–8. and recombinant interleukin-2 to patients with metastatic cancer. N Engl 27. Velickovic M, Delahunt B, McIver B, Grebe SK. Intragenic PTEN/ J Med 1985;313:1485–92. MMAC1 loss of heterozygosity in conventional (clear-cell) renal cell 7. Zisman A, deKernion JB. Historical perspective: past, present and carcinoma is associated with poor patient prognosis. Mod Pathol 2002; future. In: Belldegrun A, Ritchie A, Figlin R, Oliver R, Vaughan ED, 15:479–85. editors. Renal and adrenal tumors: biology and management. New York: 28. Alimov A, Li C, Gizatullin R, et al. Somatic mutation and homozy- Oxford University Press; 2003. p. 3–16. gous deletion of PTEN/MMAC1 gene of 10q23 in renal cell carcinoma. 8. Zisman A, Pantuck AJ, Dorey F, et al. Improved prognostication of Anticancer Res 1999;19:3841–6. renal cell carcinoma using an integrated staging system. J Clin Oncol 29. Guinan P, Sobin LH, Algaba F, et al. TNM staging of renal cell 2001;19:1649–57. carcinoma: Workgroup No. 3. Union International Contre le Cancer

(UICC) and the American Joint Committee on Cancer (AJCC). Cancer glandular neoplasia and cervical carcinomas. Am J Pathol 1994;145: (Phila) 1997;80:992–3. 598–609. 30. Oken MM, Creech RH, Tormey DC, et al. Toxicity and response 34. Harrell F. Regression modeling strategies. New York: Springer- criteria of the Eastern Cooperative Oncology Group. Am J Clin Oncol Verlag; 2001. 1982;5:649–55. 35. Hanley JA, McNeil BJ. The meaning and use of the area under a 31. Fuhrman SA, Lasky LC, Limas C. Prognostic significance of mor- receiver operating characteristic (ROC) curve. Radiology 1982;143:29–36. phologic parameters in renal cell carcinoma. Am J Surg Pathol 1982;6: 36. Liu X, Huang Y, Seligson D, Horvath S. Tree- and forest-based 655–63. methods for analyzing tissue microarray data. Presented at the Joint Statis- 32. Kononen J, Bubendorf L, Kallioniemi A, et al. Tissue microarrays tical Meeting, American Statistical Association, San Francisco, 2003. for high-throughput molecular profiling of tumor specimens. Nat Med 37. Breiman L. Random forests. Machine Learning J 2001;45:5–32. 1998;4:844–7. 38. van’t Veer LJ, Dai H, van de Vijver MJ, et al. Gene expression 33. Liao SY, Brewer C, Zavada J, et al. Identification of the MN antigen profiling predicts clinical outcome of breast cancer. Nature (Lond) as a diagnostic biomarker of cervical intraepithelial squamous and 2002;415:530–6.

Hyung L. Kim, David Seligson, Xueli Liu, et al.

Clin Cancer Res 2004;10:5464-5471.

Updated version Access the most recent version of this article at: http://clincancerres.aacrjournals.org/content/10/16/5464

Cited articles This article cites 30 articles, 5 of which you can access for free at: http://clincancerres.aacrjournals.org/content/10/16/5464.full#ref-list-1

Citing articles This article has been cited by 20 HighWire-hosted articles. Access the articles at: http://clincancerres.aacrjournals.org/content/10/16/5464.full#related-urls

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

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Subscriptions Department at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://clincancerres.aacrjournals.org/content/10/16/5464. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.