A three-gene expression signature model to predict clinical outcome of clear cell renal carcinoma

Abstract

Renal cell carcinomas (RCCs) are morphologically and genetically heterogeneous tumors and present diverse clinical courses. We developed a scoring system using levels of gene expression to predict the outcome for clear cell RCC patients. We selected differentially expressed genes from the DNA microarray data of 27 clear cell RCCs; 16 were metastasis phenotypes and 11 were not. We compared the selected gene set with previously published data and identified 33 overlapping genes closely associated with patient outcome. We selected the 12 top-ranked genes and confirmed the level of expression using quantitative reverse transcriptase PCR. Multivariate Cox analysis revealed that 3 genes—vascular cell adhesion molecule 1 (VCAM1), endothelin receptor type B (EDNRB), and regulator of G-protein signaling 5 (RGS5)—were the most tightly associated with cancer-specific survival and that higher expression of the 3 genes correlated with better outcome. A formula for an outcome predictor was generated from integration of the measurements of the expression levels of the 3 genes. Multivariate Cox models combined with a split-sample cross-validation method in a cohort of 386 clear cell RCC patients demonstrated that the derived score for outcome prediction was an independent predictor in cancer-specific survival tests. The accuracy of the prediction of cancer death after nephrectomy was improved by the inclusion of this score in receiver operating characteristic analysis from multivariate logistic regression models, suggesting that a scoring system based on the expression levels of these 3 genes is useful in the prediction of survival for patients with clear cell RCC. © 2008 Wiley-Liss, Inc.

Renal cell carcinoma (RCC) is the most common malignant tumor of the adult kidney, accounting for 2–3% of human malignancies, and complete surgical resection is considered to be the only effective treatment for patients with localized RCC.1, 2 Despite recent improvements in diagnostic imaging, 20–30% of patients are diagnosed to be in advanced stage, i.e., with local invasion or distant metastasis, at initial presentation. Moreover, tumors recur postoperatively in up to 50% of patients who undergo potentially curative surgery.1–3 RCC is characterized by its diverse clinical course. Some RCCs are aggressive and resistant to current therapies, while others are slow-growing.1–3 Although a number of traditional clinicopathologic parameters are used in current clinical practice, the ability to predict the outcome after surgical or systemic therapy is still limited, underscoring the need for new approaches or novel prognostic markers to the management of RCC.3–5

Recent studies demonstrated that RCC is a morphologically and genetically heterogeneous tumor.6 The clear cell tumor subtype is the most frequent and accounts for around 80% of all RCCs and more than 90% in advanced metastatic cases.6, 7 Loss of the von Hippel-Lindau tumor suppressor function due to the gene alteration or promoter hypermethylation is the crucial event in the tumorigenesis for this tumor subtype.1, 6, 8, 9

DNA microarray is a powerful tool that permits simultaneous rapid screening of expression levels for a large set of genes that generates enormous amounts of data. Subtype-specific tumor markers, candidate prognostic biomarkers, and novel candidate molecules for therapeutic targets have been identified for a variety of malignancies.10–12 A selective, small number of gene sets can be used to predict tumor behavior of, for example, metastatic and/or aggressive phenotypes, in several malignancies.13–15

In a previous study, we examined microarray gene expression profiles of RCCs with various histologic subtypes,16 and analyzed the gene expression signature tightly linked to the clinical outcome for clear cell subtype RCC patients. Here, we present a formula for an outcome predictor score for patients with clear cell RCC derived from the expression levels of 3 selected genes that shows an improvement in prediction value over current methods for generating survival outcome scores.

Material and methods

Microarray analysis and self-organizing map clustering

We previously examined the gene expression profiles of 33 renal tumors and 9 normal kidney samples using an oligonucleotide microarray, the GeneChip U95A platform (Affymetrix Corp., Santa Clara, CA), containing 12,600 probe sets.16 Among these tumors, a total of 27 RCCs with conventional clear cell histology, including 25 primary tumors and 2 metastases, were selected for the analysis. The clinicopathologic data for the 27 tumors are presented in Table I. Microarray analysis was performed as described in a previous report.16 The study protocol was approved by the institutional review board.

Table I. Clinicopathologic Features and Microarray SOM Clustering for Clear Cell Renal Carcinoma Cases

Tumor ID	Tumor size (cm)	Stage1	Grade1	CRP (mg/dl)	Recurence (Mo)	Survival time2 (Mo)	Outcome3	Clinical phenotype4	SOM clustering according to the gene-set				Gene signature class5
									100 × 2 genes	70 × 2 genes	50 × 2 genes	35 × 2 genes
1 Tumor stage and grade were determined according to the tumor–node–metastasis (TNM) classification.19 2 Survival time from nephrectomy. 3 NED, no evidence of disease; CD, cancer death; AWD, Alive with disease. 4 MN, metastasis negative; MP, metastasis positive. 5 See Results for detail of gene signature class. Go, good outcome; IM, intermediate; PO, poor outcome.
01T	4.0	I	2	0.2	No	200	NED	MN	0	0	0	0	GO
02T	5.0	I	1	0.3	No	185	NED	MN	0	0	0	0	GO
06T	2.3	I	2	0.1	No	170	NED	MN	0	0	0	0	GO
11T	6.0	III	2	0.0	No	108	NED	MN	0	0	0	0	GO
32T	2.0	I	2	0.1	No	111	NED	MN	0	0	0	0	GO
43T	4.5	I	2	0.2	No	90	NED	MN	0	0	0	0	GO
44T	8.0	II	2	0.1	No	81	NED	MN	0	0	0	0	GO
45T	4.5	I	2	0.0	No	88	NED	MN	0	0	0	0	GO
07T	6.5	I	2	0.9	Yes (148)	159	AWD	MP	0	0	0	0	GO
28T	2.9	I	2	0.0	Yes (129)	131	AWD	MP	0	0	0	0	GO
51T	7.0	I	2,1	0.2	Yes (58)	105	CD	MP	0	0	0	0	GO
04T	3.5	III	2	1.0	No	137	NED	MN	1	0	1	0	IM
05T	5.2	I	2	2.3	No	203	NED	MN	0	1	1	1	IM
14T	2.0	I	2	0.4	No	103	NED	MN	1	0	0	0	IM
35T	4.5	I	2	0.0	Yes (50)	103	AWD	MP	0	0	1	1	IM
09T	4.0	I	2	0.0	Yes (90)	104	CD	MP	1	1	1	1	PO
38T	2.5	I	3	5.3	Yes (28)	97	NED	MP	1	1	1	1	PO
52T	7.5	IV	3	3.0	–	30	CD	MP	1	1	1	1	PO
56T	9.0	IV	3	9.0	–	28	CD	MP	1	1	1	1	PO
56M1	2.2	–	3	–	–	–	–	MP	1	1	1	1	PO
56M2	9.5	–	3	–	–	–	–	MP	1	1	1	1	PO
57T	8.0	IV	3	0.3	–	4	CD	MP	1	1	1	1	PO
58T	6.0	I	3	4.1	Yes (3)	8	CD	MP	1	1	1	1	PO
61T	6.5	IV	3	9.7	–	2	CD	MP	1	1	1	1	PO
64T	10.9	IV	3	15.7	–	13	CD	MP	1	1	1	1	PO
66T	6.5	I	2,3	0.3	Yes (24)	42	CD	MP	1	1	1	1	PO
68T	4.0	IV	3	2.0	–	11	CD	MP	1	1	1	1	PO

We compared the 2 cluster-defined tumor subclasses using filtered gene probes in a signal-to-noise (S2N) statistic test. Gene probes that differentiated the subclasses were scored and sorted using the S2N ratio.17 Self-organizing maps (SOM) were applied for the clustering of the tumor cases using GeneCluster 2.0 open source clustering software.18

RCC patients and quantitative reverse transcriptase PCR (qRT-PCR)

A total of 423 sporadic clear cell RCCs were collected from nonselected patients who had each undergone a standard nephrectomy during the period from March 1986 through June 2004. Of the 423 RCCs, 386 patients who had a postnephrectomy follow-up period of greater than 3 years were enrolled in a survival analysis study. Tumor stage and grade were determined according to the tumor-node-metastasis (TNM) classification after surgical treatment.19 Detailed clinicopathologic characteristics are presented in Supplementary Table I. As of October 2007, the follow-up period ranged from 0.4 to 246.9 months and the median follow-up time after nephrectomy was 79.1 months for all patients and 96.0 months for survivors.

Preparation of total RNA and cDNA from fresh tumor samples and subsequent real-time quantitative PCR with a TaqMan fluorescent probe to measure mRNA expression level were performed essentially as previously described.16, 20, 21 The sequences of all primers and double-dye TaqMan probes with 5′-FAM and 3′-TAMRA or Eclipse™ Dark Quencher (Eurogentec, Belgium) were designed using Primer Express software (Applied Biosystems, Foster City, CA). The sequences of all primers and probes are shown in Supplementary Table II. In each experiment, at least 2 independent PCRs were performed to obtain mean expression values. Obtained signals were normalized by dividing by the mean expression signal of β-actin.

Statistical analysis

Normalized gene expression values detected by qRT-PCR were log-transformed (on a base 2 scale) and subsequently applied to the survival analyses. In the case of gene expression undetectable by qRT-PCR, we assigned one half of the signal value of the minimally expressing tumor for the log-transformed data analyses.

From the 386 clear cell RCC cases subjected to qRT-PCR assay, 164 randomly selected cases, including 15 previously subjected to microarray analysis, were used for the initial selection of the top-ranked genes. The entire patient cohort of 386 was used for the final survival tests. To generate a variable representing outcome based on the qRT-PCR results, we first divided the 386-patient cohort into 2 groups: (i) a preliminary training-set group of 193 patients consisting of the 164 randomly selected patients, an additional 10 cases from the microarray study, and 19 patients treated in the period designated “early” (March 1986–December 1989) and (ii) the validation test-set group consisting of the remaining 193 patients, previously not used in any study(s) (Supplementary Table I). Within the training-set group, the significance of the correlation between outcome (cancer-specific survival) and gene expression data from qRT-PCR was determined using the Cox regression model. We then calculated a gene expression-based outcome predictor score for each patient using Cox regression coefficients.13–15 The scores were divided into 3 groups—high, intermediate and low—with the 2 cut-off points determined by the maximum statistical power of the univariate cancer-specific survival test. We applied our formula and cutoff points to the validation test-set group and all patient groups and performed survival tests. Survival time was defined as the time from nephrectomy or, in cases with recurrence after curative surgery, the time from the discovery of tumor recurrence until the patient's death, or the last time at which the patient was known to be alive. Survival curves were estimated using the Kaplan-Meier method, and the resulting curves were compared using the log-rank test.

To examine the predictive performance of the scoring system, we used logistic regression analyses in the multivariate models to calculate the predictive value for each tumor case. The regression coefficients obtained from the training-set group were applied to the test-set patients. The AUCs of the receiver operating characteristic (ROC) curves for the test-set patient group were then computed from the calculated predictive values. Two parametric models, one including 2 conventional prognostic parameters (tumor stage (I + II, III and IV) and grade (G1, G2 and G3+4)) and the other comprising the 2 conventional parameters plus the outcome predictor score of continuous variables, were applied to the analysis. The presence or absence of renal cancer death within 5 years after nephrectomy was defined as the binary gold standard. Among the 193 test-set cases, 168 patients had received more than 5 years of follow-up and were therefore selected for analysis. All statistical analyses were performed using SPSS software (SPSS, Chicago, IL). All statistical tests were two-sided.

Results

Identification of differently expressed genes

We performed microarray gene expression analyses on a total of 27 clear cell RCCs from 25 patients. Of the 25, 14 patients suffered from distant metastases at initial presentation (n = 6) or during the follow-up period (n = 8). In addition, 2 tumors (56M1 and 56M2) were discrete metastases in 1 patient (Table I). We considered the clinical phenotype of these 16 tumors to be metastasis positive. The remaining 11 cases were considered to be metastasis negative because no apparent metastasis presented (Table I).

We initially selected gene probes differentially expressed between metasitasis-positive and metastasis-negative RCCs under the following criterions: (i) ≥1.5-fold difference in mean expression between the 2 classes and (ii) ≥400 mean expression in the class with higher expression. Under these criteria, 492 gene probes, 246 probes each from each classification, were selected. We then scored and ranked the 492 probes using the S2N statistic test to distinguish between the metastasis-positive and -negative phenotypes.

Next, we clustered the 27 tumors by SOM using the selected gene probes to find if these gene sets distinguished the tumor samples accurately. First, we selected the 200 top-ranked gene probes, 100 each from the two class-specific probes, for SOM clustering. In the same manner, we chose the top-ranked 140 (70 probes each), 100 (50 probes each), and 70 (35 probes each) gene probes for clustering analysis. As expected, 2 × 1 SOM clustering demonstrated that 8 metastasis-negative tumors (01T, 02T, 06T, 11T, 32T, 43T, 44T and 45T) and 12 metastasis-positive tumors (09T, 38T, 52T, 56T, 56M1, 56M2, 57T, 58T, 61T, 64T, 66T and 68T) were always subclassified in the same group (Table I). Of note, 3 of the metastasis-positive tumors (07T, 28T and 51T) were invariably clustered with the 8 metastasis-negative tumors of the 4 selected gene sets. These 3 cases presented fairly long survivals and are clinically considered slow-growing despite postoperative metastases. In addition, 4 tumors (04T, 05T, 14T and 35T) migrated among subgroups depending on the gene set selected. Among them, 3 cases (04T, 05T and 14T) did not have metastasis but presented positive plasma C-reactive protein (CRP), which is characterized as a biohumoral marker for aggressive tumor behavior.22 Case 35T suffering postoperative metastases presented relatively slow-growing behavior. These 4 cases were therefore suggested to have intermediate characteristics both in clinical behavior and gene expression signature. We omitted these 4 tumors and attempted to reclassify the rest of the 23 tumors according to gene expression signature and determined 12 to be poor outcome (PO) signature tumors and 11 to be good outcome (GO). We then ranked the alternative S2N ratio scores according to the differently expressed 492 gene probes discriminating the PO and GO signatures (Table II).

Table II. Gene Probes Differentially Expressed Between Metastasis-Positive and Metastasis-Negative Classes and Their Signal to Noise Ratio Score

Gene	Description	Affymetrix-probe ID	MP/MN Class	Signal to noise ratio score and ranking in each class				qRT-PCR assay	Reference no.
				16MP vs. MN tumors		12PO vs. 11GO tumors
				Score	Ranking	Score	Ranking
SLC9A1	Solute carrier family 9, isoform Al	32681_at	MP	0.85	21	0.99	17	Yes	24
TCOF1	Treacher Collins-Franceschetti syndrome 1	40596_at	MP	0.75	31	0.76	45	Yes	25
DPEP1	Dipeptidase 1	37413_at	MP	0.64	48	0.72	54		25
CD97	CD97 molecule	35625_at	MP	0.56	65	0.55	84		24
ST3GAL4	ST3 beta-galactoside alpha–2,3–sialyltransferase 4	36916_at	MP	0.47	92	0.37	146		28
COL6A2	Collagen, type VI, alpha 2	34802_at	MP	0.46	96	0.38	141		25
DST	Dystonin	32780_at	MP	0.09	201	0.54	87		25
CXCL9	Chemokine, C–X–C motif, ligand 9	37219_at	MP	0.02	232	0.01	253		29
TEK	TEK tyrosine kinase, endothelial	1596_g_at	MN	1.25	1	1.40	17	Yes	26
TGFBR2	Transforming growth factor, beta receptor II	1814_at	MN	1.19	2	1.40	19	Yes	23
LDBI	LIM domain binding 1	36065 at	MN	1.17	3	1.42	16	Yes	23, 26
VCAMI	Vascular cell adhesion molecule 1	583_s_at	MN	1.14	4	1.31	27	Yes	24
EDNRB	endothelin receptor type B	1198_at	MN	1.14	5	1.80	3	Yes	23
SNRK	SNF related kinase	377l8_at	MN	1.13	6	1.49	11	Yes	23
TGFBR2	Transforming growth factor, beta receptor II	1815_g_at	MN	1.07	9	1.51	9	Yes	23
VCAMI	Vascular cell adhesion molecule 1	41433_at	MN	1.05	10	1.22	38	Yes	24
TMEM47	transmembrane protein 47	37958_at	MN	1.00	15	1.80	2	Yes	23
RGS5	Regulator of G-protein signaling 5	33890_at	MN	1.00	16	1.50	10	Yes	23
TM4SF2	Transmembrane 4 superfamily member 2	38408_at	MN	0.90	28	1.40	18	Yes	23
PPAP2B	Phosphatidic acid phosphatase type 2B	33862_at	MN	0.86	35	1.61	4	Yes	23
SP3	Sp3 transcription factor	41573_at	MN	1.03	13	0.97	71		25
PECAM1	Platelet/endothelial cell adhesion molecule	37398_at	MN	0.69	77	1.21	39		23
SPARCL1	SPARC-like 1 (hevin)	36627_at	MN	0.69	78	0.97	73		23
ALDH7A1	Aldehyde dehydrogenase 7 family, member Al	36132_at	MN	0.65	91	0.72	133		25
ENPP2	Ectonucleotide pyrophosphatase/phosphodiesterase 2 (autotaxin)	41123_s_at	MN	0.64	92	1.20	41		23
ENPP2	Ectonucleotide pyrophosphatase/phosphodiesterase 2 (autotaxin)	41124_r_at	MN	0.63	99	1.29	28		23
SPRY1	Sprouty homolog 1	38767_at	MN	0.59	117	0.72	131		23
LMO2	LIM domain only 2	32184_at	MN	0.56	125	0.79	119		23
ADFP	Adipose differentiation-related protein	34378_at	MN	0.55	128	0.83	106		21
ZNF292	Zinc finger protein 292	40838_at	MN	0.53	141	0.61	160		25
AQP1	Aquaporin 1	36156_at	MN	0.50	150	1.36	21		30
TIMP3	TIMP metalloproteinase inhibitor 3	1034_at	MN	0.49	153	0.71	134		23
GPR116	G–protein coupled receptor 116	34235_at	MN	0.46	161	0.68	141		23
IFIT1	Interferon–induced protein with tetratricopeptide repeats 1	915_at	MN	0.40	181	0.53	173		24
CCND1	Cyclin Dl	2017_s_at	MN	0.31	200	0.50	181		31
VEGF	Vascular endothelial growth factor	36101_s_at	MN	0.29	203	0.52	179		32
TIMP3	TIMP metalloproteinase inhibitor 3	1035_g_at	MN	0.27	206	0.57	168		23

Overlapping genes with other microarray experiments and literature

We compared our selected genes with gene lists in published microarray results associated with metastasis or clinical outcome for RCC patients. Takahashi et al. hybridized 29 clear cell RCCs to a 21,632 cDNA microarray and found 51 gene probes (37 unique transcripts) associated with good or poor outcome.23 Vasselli et al. hybridized 58 metastatic RCCs to a 6,400 cDNA microarray and found 45 genes most significantly associated with patient survival time.24 Sültmann et al. hybridized 19 clear cell RCCs that had metastasis at the time of diagnosis and 17 that did not have metastasis to cDNA microarrays containing 4,207 probes. They found 85 gene probes statistically significantly associated with metastasis.25 Kosari et al., using Affymetrix HG-U133 Plus2 arrays, examined 28 clear cell RCCs and found 35 gene probes differentially expressed between nonaggressive and aggressive tumors.26 Jones et al., using Affymetrix HG-U133A chips, identified 31 probes that are differentially expressed during stepwise progression from normal kidney tissue (n = 23) through early tumor stage (n = 8) to distant metastasis (n = 9).27 When we compared our 492 gene probes with the above gene sets, 31 gene probes were a match, representing 27 unique transcripts. In addition, we found that 6 genes of our 492 probes were reportedly associated with clinical outcome for RCC patients (Table II). The 33 unique transcripts, including 8 metastasis-positive (or PO) and 25 metastasis-negative (or GO) associated genes, overlapping between our list and those from other experiments became candidates for further validation (Table II).

Validation of top-ranked genes by qRT-PCR-based survival tests

To validate the significance of the selected genes, we measured the gene expression of 164 nonselective RCC cases using qRT-PCR. Among the 33 overlapping genes, we chose 12 (2 PO and 10 GO) because of their high rank (within the top 50) for differentiations in both PO/GO and metastasis-positive/-negative subclasses (Table II). The qRT-PCR assays and subsequent multivariate Cox models, including 5 conventional prognostic parameters (symptomatic presentation, tumor size, stage, grade and microvascular invasion status) together with 1 of 12 genes, demonstrated that 6 of 12 genes (LDB1, VCAM1, EDNRB, RGS5, TM4SF2 and PPAP2B) were strongly associated with cancer-specific survival (p-values of 0.052, 0.071, 0.037, 0.039, 0.057 and 0.007, respectively) (Supplementary Table IIIA). In addition, multivariate Cox analysis consisting of these 6 genes demonstrated that VCAM1, EDNRB and RGS5 were the top 3 genes for correlation to survival (p-values of 0.102, 0.008 and 0.188, respectively); therefore, these 3 genes, all of which belong to the metastasis-negative or GO subclass, were selected for final evaluation (Supplementary Table IIIB).

Outcome predictor score based on the three gene expression levels

To create a formula for an outcome predictor score based on gene expression, we performed qRT-PCR of the 3 genes with an additional 222 tumors and applied split-sample methods to the total cohort of 386 RCC patients. Initially, we constructed the outcome predictor score using the multivariate Cox model in a training-set group consisting of 193 patients. According to the Cox regression coefficients, the formula for the outcome predictor score is: (−0.0984 × VCAM1 value) + (−0.2287 × EDNRB value) + (−0.1812 × RGS5 value) (Supplementary Table IV). The formula and the cutoff points developed in the training set were subsequently applied to the validation test-set group as well as to the all-patient group.

Univariate survival analyses demonstrated that the formula was statistically significant in cancer-specific survival tests of the test-set and that the group with high scores presented significantly poor outcome for cancer death (Fig. 1a). Further multivariate Cox models together with various clinicopathologic parameters revealed that the outcome predictor score remained an independent prognostic parameter for both groups (Table III).

Figure 1. Kaplan-Meier plots of estimated survival and outcome predictor scores in the test set and all patients with clear cell RCCs. Cancer-specific survival probabilities for (a) all patients who underwent nephrectomy, and (b) patients with advanced metastatic RCC, including both stage IV patients who had undergone non-curative surgery and patients in whom metastasis occurred after presumably curative nephrectomy, with survival probabilities evaluated as a single cohort. p-values: Log-rank test between the two groups.

Table III. Cox Multivariate Analyses of Cancer-Specific Survival among Patients With Clear Cell RCC in the Test Set and All Patients

Characteristic		Test set (n = 193)			All patients (n = 386)
		P	HR	95%CI	P	HR	95%CI
Age (years)	≤ 59		1.000			1.000
	60 ≤	0.099	1.649	0.910–2.990	0.488	1.142	0.785–1.662
Sex	Female		1.000			1.000
	Male	0.250	1.461	0.766–2.787	0.715	0.931	0.633–1.368
Symptom	Neg		1.000			1.000
	Pos	0.164	1.783	0.789–4.028	0.006	2.125	1.240–3.639
Size (cm)	≤ 40		1.000			1.000
	4.1–7.0	0.769	0.851	0.290–2.497	0.125	1.660	0.868–3.175
	7.1 ≤	0.596	1.353	0.442–4.143	0.335	1.384	0.715–2.681
Stage	I–II		1.000			1.000
	III	0.709	1.239	0.401–3.824	0.175	1.577	0.817–3.045
	IV	0.001	5.438	1.994–14.831	0.000	6.434	3.435–12.049
Grade	G1		1.000			1.000
	G2	0.144	2.279	0.754–6.887	0.092	1.941	0.897–4.200
	G3+4	0.037	3.429	1.079–10.903	0.006	3.027	1.371–6.683
Vein	Neg+x		1.000			1.000
	Pos	0.005	2.647	1.344–5.212	0.003	1.886	1.237–2.877
Outcome	Low		1.000			1.000
predictor	Med	0.640	1.196	0.565–2.530	0.164	1.437	0.862–2.395
score	High	0.003	3.642	1.546–8.580	0.009	2.094	1.202–3.648

We also studied the ability of the score to predict the survival time of advanced metastatic RCC patients with apparent tumor burdens. Patients in stage IV who had received palliative surgery and patients who had developed tumor recurrence were enrolled together for the analysis. In the univariate analyses, we detected significant survival differences between our outcome predictor score for the test set and the all-patient group (Fig. 1b). In further multivariate analysis models with 4 variables (patient age, sex, tumor grade and outcome predictor score), the score was again found to be an independent predictor of survival time for both groups (Table IV).

Table IV. Cox Multivariate Analyses of Cancer-Specific Survival among Patients With Advanced Metastatic Clear Cell RCC in the Set and All Patients

Characteristic		Test set (n = 71)			All patients (n = 166)
		P	HR	95%CI	P	HR	95%CI
Age (years)	≤ 59		1.000			1.000
	60 ≤	0.006	2.451	1.292–4.648	0.015	1.605	1.095–2.354
Sex	Female		1.000			1.000
	Male	0.106	1.716	0.891–3.304	0.685	0.923	0.628–1.357
Grade	G1		1.000
	G2	0.584	1.364	0.449–4.144	0.948	0.974	0.448–2.117
	G3+4	0.220	1.986	0.664–5.946	0.320	1.477	0.685–3.184
Outcome	Low		1.000			1.000
predictor	Med	0.420	1.342	0.657–2.743	0.064	1.595	0.973–2.615
score	High	0.008	2.657	1.292–5.465	0.000	2.546	1.521–4.262

Outcome predictor score improved diagnostic accuracy of patient prognosis

We further examined whether the 3-gene outcome predictor score provides additional information for the prediction of patient outcome with clear cell RCC. The predictive value of renal cancer death within 5 years was calculated using multivariate logistic regression models and validated from ROC analysis. The AUC of the ROC curve with 2 conventional parameters (tumor stage and grade) for the test-set group was 0.904; when our score was included in the model, the AUC was 0.912 (Fig. 2), indicating improved accuracy. Therefore, accuracy in patient prognosis can be improved by the addition of our 3-gene outcome predictor score.

Figure 2. ROC curves for the test set showing prediction of renal cancer death within 5 years of nephrectomy. ROC curves for the test-set patients (n = 168) generated by a logistic regression model using 2 conventional parameters (tumor stage and grade) (dotted line) and the 2 parameters plus the outcome predictor score (solid line). The regression coefficients obtained from the training-set group (n = 193) were applied to the test-set patients.

Discussion

Our current method of seeking gene expression signatures tightly linked to clinical outcome to develop an outcome predictor formula for clear cell RCC patient is comprised of stepwise selection. We began with DNA microarray data from 27 clear cell RCCs, 16 associated with metastasis phenotype and 11 not, and selected genes differentially expressed in these clinical phenotypes. To delimit a tractable number of gene probes, we compared our gene set with previously published microarray data as well as with data of prognostic correlation in the literature. Promising candidates were genes overlapping between our data and independent external data and a total of 33 unique transcripts were found. We then selected 12 of the high-ranked genes and, applying qRT-PCR to validate their significance, measured the expression levels within a large patient cohort with a substantial follow-up period. We found 3 genes (VCAM1, EDNRB and RGS5) presenting a substantial impact on patient survival. Furthermore, we confirmed that the score integrating the expression levels of these 3 genes improved the accuracy of prediction of survival with clear cell RCC. We believe that accurate prediction of outcome and prognosis based on the conventional prognostic factors combined with our gene-expression score will be valuable for patient counseling, follow-up scheduling and adjuvant treatment.

The 3 genes seem to have underlying implications for tumorigenesis as well as a clinical application for clear cell RCC. VCAM1 is known as a cell adhesion molecule and its significance in clear cell RCC patient survival has been reported recently.20, 24 To our knowledge, this is the first report demonstrating a strong association between expression of EDNRB and RGS5 and clear cell RCC patient survival. EDNRB is a cell surface receptor and belongs to the G protein-coupled receptor (GPCR) system.33, 34 EDNR together with its ligands, EDN1, EDN2 and EDN3 endothelins (EDNs) comprising the ET axis, is known to have diverse physiological functions in normal tissue,33–35 and is also implicated in tumor proliferation, survival, invasion, angiogenesis and metastasis.33, 34 Interestingly, a high expression of EDNRB is known to be associated with metastatic or aggressive phenotypes in malignant melanoma and vulvar squamous cell carcinoma.36–38 On the contrary, in bladder cancer, EDNRB overexpression correlates with favorable prognosis,39 as was observed in this study with clear cell RCC. These findings also likely reflect the diverse functions of the ET axis which depend on the tumor histology or cancer cell type. Of note, small molecules selectively modulating EDNRs have been studied for their antitumor activities.40, 41 In RCC, expressions of EDNs and EDNRs in cancerous cells have been reported by several groups.42–44 It will be interesting to further investigate the role of the ET axis, especially EDNRB together with its modulating molecules, in molecular-targeting therapy for clear cell RCC.

RGS5 is a member of the RGS family and a negative regulator of heterotrimeric G-protein signaling pathways.45 In RCC, RGS5 is highly expressed in tumor endothelial cells but not in those of normal renal vasculature.46 More recent data demonstrated that RGS5 is exclusively upregulated in pericytes rather than in endothelial cells, at sites of physiologic and pathologic angiogenesis.47 RGS5 is assumed to be involved in vessel remodeling during tumor-induced neovascularization and therefore one of the novel targets for antiangiogenic therapy. Recently, novel multikinase inhibitors, such as sunitinib and sorafenib, most likely targeting vascular endothelial growth factor (VEGF) receptor-driven endothelial cells as well as platelet-derived growth factor (PDGF) receptor-dependent pericytes, have been introduced in clinical practice for advanced RCCs and are showing great promise.2, 48–50 Detailed investigations into these novel agents against the RGS5-related signaling pathway will be intriguing.

In this study, we evaluated 12 high-ranked genes from preliminary microarray data but the impact on survival for each gene from qRT-PCR-based validations presented some differences among them. For example, SLC9A1 and TCOF1 were ranked the Top 2 in the poor outcome signature and TGFBR2 presented a relatively high rank in the good outcome signature in the preliminary microarray screening. However, the statistical values for these genes were not prominent in the qRT-PCR-based validation test. We agree with Dupuy and Simon51 that confirmation in a large, independent cohort is required. We also think other genes that emerged in this study may be good candidates for further validations. We plan to further evaluate and refine our prototype formula of molecular diagnosis of RCCs.

Acknowledgements

The authors wish to thank Mrs. Takako Yamaki, Yoko Nakamura, and Rie Shimizu for their excellent technical assistance and Frances Ford for proofreading the manuscript.