Bladder cancer SNP panel predicts susceptibility and survival

Abstract

Bladder cancer is the fourth most common malignancy in men and the eighth most common in women in western countries. Single nucleotide polymorphisms (SNPs) in genes that regulate telomere maintenance, mitosis, inflammation, and apoptosis have not been assessed extensively for this disease. Using a population-based study with 832 bladder cancer cases and 1,191 controls, we assessed genetic variation in relation to cancer susceptibility or survival. Findings included an increased risk associated with variants in the methyl-metabolism gene, MTHFD2 (OR 1.7 95% CI 1.3–2.3), the telomerase TEP1 (OR 1.8 95% CI 1.2–2.6) and decreased risk associated with the inflammatory response gene variant IL8RB (OR 0.6 95% CI 0.5–0.9) compared to wild-type. Shorter survival was associated with apoptotic gene variants, including CASP9 (HR 1.8 95% CI 1.1–3.0). Variants in the detoxification gene EPHX1 experienced longer survival (HR 0.4 (95% CI 0.2–0.8). These genes can now be assessed in multiple study populations to identify and validate SNPs appropriate for clinical use.

Introduction

In western countries, bladder cancer is the fourth most common malignancy in men and the eighth most common in women (Kirkali et al. 2005). The majority of this disease is attributed to cigarette smoking; bladder cancer risk is up to fourfold higher among cigarette smokers compared with non-smokers (Kirkali et al. 2005). A hereditary component is likely since a family history of bladder cancer and variations in genes that detoxify aromatic amines are associated with increased risk (Garcia-Closas et al. 2005; Kantor et al. 1985; Murta-Nascimento et al. 2007b). We and others have also identified associated variations in DNA repair pathway genes, yet single nucleotide polymorphisms (SNPs) in important cancer regulatory pathways such as cell growth and apoptosis have not been comprehensively assessed for this disease (Andrew et al. 2006; Wu et al. 2006).

Bladder cancer generally carries a favorable prognosis, nevertheless, approximately 13,750 deaths occurred in the USA from bladder cancer in 2007 (Jemal et al. 2007). According to the US Statistics Epidemiology and End Results (SEER) program (1996–2002), the 5-year survival rate for localized disease is 94%, whereas for regional staged cancers the survival rate is only 47%. Those with distant metastasis at the time of diagnosis carry the lowest survival rate of 6% (Jemal et al. 2007). Established prognostic factors include tumor characteristics such as degree of invasion or carcinoma in situ (CIS) (Carroll et al. 2001). Higher grade of a tumor is also associated with a worse prognosis (Schned et al. 2007). Although these histopathologic factors can be used to predict prognosis, the behavior of each group is quite heterogeneous [reviewed in (Murta-Nascimento et al. 2007a)]. This inter-individual variation in survival highlights the potential for new genetic markers that can assist physicians in selecting appropriately aggressive treatment regimens.

We have assessed the role of genetic variation in bladder cancer risk and case survival using panel of SNPs in hypothesized cancer regulatory pathways, including cell cycle, cell growth, detoxification, telomerase, and apoptosis. Risk- or survival-associated single gene and multi-gene-combinations were ranked and tested using traditional multivariate statistical methods in a large, population-based case–control study in New Hampshire, USA.

Materials and methods

Study group

We identified all cases of bladder cancer among New Hampshire residents, ages 25–74 years from the State Cancer Registry. The study was conducted in two sequential phases based on cancer diagnosis date (Phase I 1 July 1994–30 June 1998, Phase II 1 July 1998–30 June 2001). Detailed methods have been described previously (Karagas et al. 1998). Briefly, we interviewed a total of 832 bladder cancer cases, which was 85% of the cases confirmed to be eligible for the study. Controls less than 65 years of age were selected using population lists obtained from the New Hampshire Department of Transportation. Controls 65 years of age and older were chosen from data files provided by the Centers for Medicare & Medicaid Services (CMS) of New Hampshire. For efficiency, we shared a control group with a study of non-melanoma skin cancer in New Hampshire covering an overlapping diagnostic period of 1 July 1993–30 June 1995 (Karagas et al. 1998). We selected additional controls for bladder cancer cases diagnosed from 1 July 1995 to 30 June 1998 frequency matched to these cases on age (25–34, 35–44, 45–54, 55–64, 65–69, 70–74 years) and gender. Most (>95%) of the subjects in this study are of Caucasian origin; and thus our analyses were not appreciably altered by restricting to Caucasians. We interviewed a total 1,191 controls (the total shared control group and additional controls), which was 70% of the controls confirmed to be eligible for the study.

Personal interview

Informed consent was obtained from each participant and all procedures and study materials were approved by the Committee for the Protection of Human Subjects at Dartmouth College. Consenting participants underwent a detailed in-person interview, usually at their home. Questions covered sociodemographic information (including level of education), lifestyle factors such as use of tobacco (including frequency, duration and intensity of smoking), family history of cancer, and medical history prior to the diagnosis date of the bladder cancer (cases) or reference date assigned to controls. Recruitment procedures for both the shared controls from the non-melanoma skin cancer study and additional controls were identical and ongoing concomitantly with the case interviews. Case–control status and the main objectives of the study were not disclosed to the interviewers. To ensure consistent quality of the study interviewer, interviews were tape recorded with the consent of the participants and routinely monitored by the interviewer supervisor. To assess comparability of cases and controls, we asked subjects if they currently held a driver's license or a Medicare enrollment card. Subjects were asked to provide a blood sample (buccal sample was requested in the case of a refusal). Samples were maintained at 4°C and sent via courier to the study laboratory at Dartmouth within 24 h for processing and analysis.

Genotyping

DNA was isolated using Qiagen genomic DNA extraction kits (QIAGEN Inc, Valencia, CA) from peripheral circulating blood lymphocyte specimens harvested at the time of interview. Genotyping was performed on all DNA samples of sufficient concentration (893 controls, 617 cases), using the Cancer Panel on the GoldenGate Assay system by Illumina's Custom Genetic Analysis service (Illumina Inc., San Diego, CA). The Cancer Panel contains 1,421 SNPs in approximately 400 hypothesized cancer-related genes from the SNP500 database (http://www.illumina.com). Genotype data were available on 75% of interviewed subjects and the genotyped population had similar characteristics as the overall study. Samples repeated on multiple GoldenGate plates yielded the same call for 99.9% of SNPs and 99.5% of samples submitted were successfully genotyped. Genotype calls were 99% concordant between genotyping platforms.

Statistical analysis

The main goals of the statistical analysis were to identify SNPs and combinations of SNPs that relate to bladder cancer susceptibility or survival. Selection of SNPs with dose–responsive relationships to bladder cancer risk was performed by ranking SNPs based on p-for-trend values associated with a dose response of variant alleles with FDR value ≤0.56.

Kullback–Leibler distance measures were used to compare discrete distributions of feature values for cases and controls (Duda et al. 2001). The Kullback–Leibler divergence provides an information-based measure of the diVerence between two probability distributions. The Kullback–Leibler divergence may be expressed as the number of bits of information required to encode samples based on the two distributions. The Kullback–Leibler divergence differs from a conventional distance metric on the space of probability distributions in that it is not symmetric and does not satisfy the triangle inequality. As an information measure, the Kullback–Leibler divergence may be thought of as a non-linear measure that often accentuates distribution difference properties that are not detected by conventional distance metrics (Kullback 1959, 1987). While the divergence cannot be easily generalized in its performance, it is often a valuable complement to conventional distance metrics in applications where well-defined models are not easily available. Kullback–Leibler divergence has recently been to be used in studies of genetic association [e.g. (Cussenot et al. 2008)]. The SNPs were ranked and the top 20 SNPs with the largest scores were selected.

Two, three, and four-way combinations of p-for-trend and Kullback–Liebler selected potentially interacting SNPs were identified using Multifactor Dimensionality Reduction (MDR) analysis (Ritchie et al. 2001). To allow for epistatic interactions in the absence of main effects, we used a more liberal p-for-trend cutoff for entry into the MDR model (FDR value 0.83). MDR is a data reduction (i.e. constructive induction) approach that seeks to identify combinations of multilocus genotypes and discrete environmental factors that are associated with either high risk or low risk of disease. Thus, MDR defines a single variable that incorporates information from several loci and/or environmental factors that can be divided into high risk and low risk combinations. This new variable can be evaluated for its ability to classify and predict outcome risk status using cross-validation and permutation testing (Ritchie et al. 2001, 2003).

We also used the tuned Relief F as a pre-filtering step to help identify epistatic or non-additive, interacting SNPs in MDR. ReliefF considers SNPs jointly rather than individually. In our experience, this method is better for detecting interactions than chi-square test. ReliefF estimates weights for each SNP “estimated based on whether the nearest neighbors (nearest hit, H) of a randomly selected instance from the same class and the nearest neighbors from the other class (nearest miss, M) have the same or different values” (Moore and White 2007). Tuned ReliefF uses vectors of SNPs, removes those of low quality, re-estimates, and then ranks them based on their ability to discriminate case–control status.

To assess the magnitude of the independent effects of the SNPs that were identified by the filters described above, we conducted logistic regression analyses for individuals with one or two variant alleles in comparison to those homozygous wild type for each individual SNP. These analyses were adjusted for age (less than or greater than median age 64), gender, and smoking status (never, former, current). We evaluated the nature of interactions between bladder cancer risk factors (gender, smoking, age) and/or the SNP allele combinations predicted by MDR, by including interaction terms in a logistic regression model. Statistical significances of the interactions were assessed using likelihood ratio tests comparing the models with and without the interaction terms (e.g. logit(p) = constant + SNP1 + SNP2 vs. logit(p) = constant + SNP1 + SNP2 + interaction). MDR identifies combinations of predictive factors, but does not distinguish between multiplicative and additive effects. Thus, we used the likelihood ratio test to assess the multiplicative nature of the interactions. p values represent two-sided statistical tests with statistical significance at P < 0.05. Our analytic strategy embraces both the algorithmic and traditional statistical approaches [discussed in (Breiman 2001)], allowing us to identify epistatic SNP interactions.

Selection of SNPs related to bladder cancer survival was performed using an FDR threshold <0.5, choosing the four individual SNPs with the lowest log-rank p values (≤0.1). Then we used Gene Set Assessment (GSA) to investigate groups of SNPs that are members of a common gene. The GSA method calculates an enrichment score based on whether the gene-based SNP sets contain a number of individual SNPs that are highly associated with cancer (over-represented), with adjustment for multiple testing (Holden et al. 2008). GSA was used to select the top 10 ranking genes (each containing multiple SNPs) by log-rank p values. Survival plots for bladder cancer cases were generated using the Kaplan-Meier method and differences between genotypes were assessed using the log-rank test. To adjust for additional factors related to patient survival, Cox-proportional hazards regression analysis was performed with age, gender, smoking (never, former, current), as well as AJCC tumor stage (0, 0is, 1–4), grade (1–4), size (<, ≥30 mm), treatment (immunotherapy with surgery; chemotherapy, radiotherapy, or immunotherapy only; chemotherapy or radiotherapy with cystectomy; transurethral resection only; cystectomy only) in the model. Results were not sensitive to removal of subjects with metastatic tumors from the analysis (data not shown). p values represent two-sided statistical tests with statistical significance at P < 0.05.

Results

The characteristics of the subjects who provided a DNA sample were similar to those of the overall study population. The majority of the population was Caucasian. The case group contained a higher percentage of current smokers than the controls (Table 1). A large proportion of the cases were non-invasive bladder cancer, reflecting the population-based nature of the study. Testing showed deviation from Hardy–Weinberg equilibrium for ADH1C_14, and MSH6_01.

Table 1.

Genotyped population characteristics

	Controls n (%)	Cases n (%)	p value
Gender
Female	338 (38)	150 (24)	Ref.
Male	555 (62)	467 (76)	<0.0001
Reference age
<40	43 (5)	17 (3)	Ref.
40–55	152 (17)	102 (17)	0.4
55–70	480 (54)	337 (54)	0.3
>70	218 (24)	161 (26)	0.2
Race
White	889 (99)	607 (98)	Ref.
Non-white	4 (1)	10 (2)	0.06
Smoking status
Never	305 (34)	106 (17)	Ref.
Former	447 (50)	305 (50)	<0.0001
Current	141 (16)	205 (33)	<0.0001
AJCC Stage
0	–	337 (56)
0is	–	24 (4)
1	–	170 (28)
2	–	43 (7)
3, 4	–	33 (5)

Data were missing for smoking on 1 case, stage on 10 cases

Risk models

The p-for-trend method that hypothesizes a dose–response effect by the number of variant alleles was used to select the individual SNPs shown in Table 2. Individuals homozygous for the variant forms of the alcohol dehydrogenase SNP ADH1C_14 had an increased risk of bladder cancer compared with wild type [OR 1.7 95% CI (1.3–2.2)]. Increased risk was also observed among individuals with the homozygous variant form of SNPs in MTHFD2, TEP1, AURKA, FZD7, IGF1, MSH6. Gene variants associated with reduced risk included IL8RB, MET, SLC4A2, MBL2 (Table 2). Of the SNPs selected in the p-for-trend analysis, only the risk associated with ADH1C_14 varied by smoking status. Never smokers had modestly higher bladder cancer risk OR 1.4 (95% CI 0.9–2.3) than current smokers OR 0.9 (95% CI 0.6–1.4) (gene—current smoker interaction p = 0.05). We also analyzed these SNPs separated by study phase (I, II). Results within each phase were consistent with our analysis of the overall population.

Table 2.

Individual SNPs selected for association with bladder cancer risk by p-for-trend <0.004

SNP (function)	Genotype	Controls		Cases		Bladder cancer OR (95% CI)
		N	%	N	%
ADH1C_14 (Alcohol dehydrogenase)	wt	440	51	256	43	1.0 Ref.
rs2241894	het	264	31	181	31	1.2 (0.9–1.5)
g.Ex5+106A>G	var	156	18	152	26	1.7 (1.3–2.2)
p.T151T	any var	420	49	333	57	1.4 (1.1–1.7)
MTHFD2_01 (Methyl-metabolism)	wt	290	34	146	25	1.0 Ref.
rs1667627	het	389	45	292	49	1.5 (1.2–1.9)
g.IVS1+3323T>C	var	183	21	152	26	1.7 (1.3–2.3)
	any var	572	66	444	75	1.6 (1.2–2.0)
TEP1_08 (Telomerase)	wt	452	51	261	42	1.0 Ref.
rs1760897	het	366	41	280	45	1.3 (1.0–1.6)
g.Ex1−222T>C	var	75	8	75	12	1.8 (1.2–2.6)
p.S116P	any var	441	49	355	58	1.4 (1.1–1.7)
IL8RB_01 (IL-8 Receptor, Angiogeneisis, CXCR2)	wt	299	35	250	42	1.0 Ref.
rs1126579	het	414	48	255	43	0.7 (0.6–0.9)
Ex3+1235T>C	var	150	17	84	14	0.6 (0.5–0.9)
	any var	564	65	339	58	0.7 (0.6–0.9)
AURKA_15/STK6_15b (Cell cycle kinase, spindle poles in mitosis)	wt	395	46	224	39	1.0 Ref.
rs8173	het	356	42	252	44	1.2 (1.0–1.6)
g.Ex11−347G>C	var	99	12	92	16	1.6 (1.2–2.3)
	any var	455	54	344	61	1.3 (1.1–1.6)
FZD7_10 (Wnt receptor, decreases function of antigen presenting cells)	wt	453	51	274	44	1.0 Ref.
rs13034206	het	363	41	272	44	1.3 (1.0–1.6)
g.Ex1−1926C>T	var	73	8	70	11	1.7 (1.1–2.4)
	any var	436	49	342	55	1.3 (1.1–1.7)
MET_01 (Oncogene, hepatocyte growth factor receptor)	wt	240	28	198	34	1.0 Ref.
rs41736	het	433	50	291	49	0.8 (0.6–1.0)
g.Ex20+60C>T	var	192	22	102	17	0.6 (0.5–0.9)
p.D1304D	any var	625	78	393	66	0.7 (0.6–0.9)
IGF1_11 (Somatomedin, growth factor)	wt	492	57	294	50	1.0 Ref.
rs5742629	het	325	38	250	42	1.3 (1.0–1.6)
g.IVS2+12158A>G	var	48	6	47	8	1.7 (1.1–2.6)
	any var	373	44	297	50	1.4 (1.1–1.7)
FZD7_16 (Wnt receptor, decreases function of antigen presenting cells)	wt	451	52	269	46	1.0 Ref.
rs12474408	het	352	41	261	44	1.3 (1.0–1.6)
g.*2710A>G	var	62	7	60	10	1.7 (1.1–2.5)
	any var	414	48	321	54	1.3 (1.1–1.6)
MSH6_01 (DNA mismatch recognition)	wt	319	37	192	33	1.0 Ref.
rs3136228	het	376	44	248	42	1.1 (0.9–1.4)
g.−556G>T	var	169	20	150	25	1.6 (1.2–2.1)
	any var	545	64	398	67	1.3 (1.0–1.6)
GSK3B_45 (Kinase energy metabolism)	wt	539	62	319	54	1.0 Ref.
rs3755557	het	266	31	222	38	1.5 (1.2–1.9)
−2675T>A	var	58	7	49	8	1.3 (0.8–2.0)
	any var	324	38	271	46	1.5 (1.2–1.8)
SLC4A2_04 (Solute exchange)	wt	514	60	396	67	1.0 Ref.
rs13240966	het	312	36	177	30	0.8 (0.6–1.0)
g.IVS1−194C>G	var	35	4	18	3	0.6 (0.3–1.1)
	any var	347	40	195	33	0.7 (0.6–0.9)
MBL2_30 (Innate immunity)	wt	457	54	346	59	1.0 Ref.
rs2099902	het	329	38	221	37	0.9 (0.7–1.1)
g.Ex4−710G>A	var	66	8	23	4	0.5 (0.3–0.8)
	any var	395	46	244	41	0.8 (0.6–1.0)
TERT_15 (Telomerase reverse transcriptase)	wt	659	77	422	71	1.0 Ref.
rs13167280	het	182	21	152	26	1.4 (1.1–1.8)
g.IVS3−24T>C	var	16	2	17	3	1.7 (0.8–3.4)
	any var	198	23	169	29	1.4 (1.1–1.8)

Analyses were adjusted for gender, age, smoking. Genotype data were missing for ADH1C_14 on 5 controls, 28 cases; MTHFD2_01 3 controls, 27cases; TEP1_08 0 controls, 1 case; IL8RB_01 30 controls, 28 cases; AURKA_15 43 controls, 49 cases; FZD7_10 4 controls, 1 case; MET_01 28 controls, 26 cases; IGF1_11 28 controls, 26 cases; FZD7_16 28 controls, 27 cases; MSH6_01 29 controls; 27 cases; GSK3B_45 30 controls, 27 cases; SLC4A2_04 32 controls, 26 cases; MBL2_30 41 controls, 27 cases; TERT_15 36 controls, 26 cases

wt wildtype, het heterozygous, var variant, any variant heterozygous and variant, g. gene, Ex exon, p. protein

To identify SNPs that do not follow the p-for-trend risk model, we used the Kullback–Leibler distance algorithm to rank SNPs for an association with bladder cancer risk (heterozygous (het.) or variant (var.) vs. wildtype). Of the top 20 SNPs selected by Kullback–Leibler distance, five showed individual gene relationships with bladder cancer incidence when assessed by logistic regression analysis after adjustment for age, gender, and smoking status, GSTZ1_02, AKR1C3_35, TYR_02, SCARB1_03, and SLC23A1_05 (Table 3). The other SNPs that had top ranking Kullback–Leibler distances did not show a significant association with bladder cancer by logistic regression analysis with adjustment for potential confounders, including SCARB1_03, PLA2G6_08, PTGS1_02, EPHX1_17, SLC23A2_31, SLC6A3_14, TGM1_01, AKR1C3_11, KRAS_22, CTNNB1_02, IL10_06, HMGCR_01, BARD1_02, AXIN2_09, PTGS1_02, MSH2_06 (data not shown).

Table 3.

Individual SNPs selected for association with bladder cancer risk by Kullback–Leibler distance

Function	SNP	Genotype	Controls		Cases		Bladder cancer OR (95% CI)
			N	%	N	%
Detoxification	GSTZ1_02	wt	636	86	367	80	1.0 Ref.
	g.Ex5−12G>A	het	99	14	88	19	1.5 (1.1–2.1)
	p.G42R	var	5	1	4	1	1.4 (0.4–5.3)
		any var	104	15	92	20	1.5 (1.1–2.0)
Aldo/keto reductase	AKR1C3_35	wt	509	59	312	53	1.0 Ref.
PAH metabolism	g.*12259G>A	het	315	37	238	41	1.3 (1.0–1.6)
		var	40	4	38	6	1.5 (0.9–2.4)
		any var	355	41	276	47	1.3 (1.0–1.6)
Pigmentation	TYR_02	wt	463	53	299	51	1.0 Ref.
Melanin production	g.IVS3−6895A>G	het	322	37	255	43	1.2 (1.0–1.5)
		var	80	9	36	6	0.7 (0.5–1.2)
		any var	402	46	291	49	1.1 (0.9–1.4)
Lipid metabolism	SCARB1_03	wt	383	44	235	40	1.0 Ref.
	g.IVS1−18462G>A	het	381	44	290	49	1.2 (0.9–1.5)
		var	101	12	66	11	1.0 (0.7–1.4)
		any var	482	56	356	60	1.1 (0.9–1.4)
Metabolism	SLC23A1_05	wt	813	94	545	92	1.0 Ref.
	g.Ex8+22G>A	het	50	6	46	8	1.5 (1.0–2.3)
	p.V264M	var	2	–	0	–	–
		any var	52	6	46	–	1.5 (0.9–2.2)

Analyses were adjusted for gender, age, smoking. Genotype data were missing for GSTZ1_02 on 153 controls, 158 cases; AKR1C3_35 on 29 controls, 29 cases; TYR_02 on 28 controls, 27 cases; SCARB1_03 on 28 controls, 26 cases; SLC23A1_05 on 28 controls, 26 cases

wt wildtype, het heterozygous, var variant, any variant heterozygous and variant, g. gene, Ex exon, p. protein

We then identified potential interactions between these selected SNPs using MDR, followed by quantification of the odds ratios associated with the high versus low risk combinations of alleles/characteristics using logistic regression analysis with adjustment for age, gender and smoking status. Of the p-for-trend selected genes using an FDR value cutoff of 0.8, MDR indicated that a combination of the high-risk genotypes for PIN1_21 rs889162 g.IVS3+2592T>C, MET_01 rs41736 g.Ex20+60C>T p.D1304D, and AMACR_17 rs10941112 g.Ex3-29A>G p.G175D was significantly associated with bladder cancer (MDR accuracy 0.56, cross-validation consistency 7/10). The logistic regression adjusted OR was 2.3 [95% CI 1.8–2.9)] for the high risk alleles, versus the low risk genotype combination. Of these PIN1_21 and AMACR_17 indicated the strongest evidence of synergy (likelihood ratio test gene–gene interaction p value = 0.02).

We also entered the top ranking 20 Kullback–Leibler distance selected genes, along with age, gender and smoking status, into MDR. The combination of high-risk genotypes for SLC23A2_31 rs12479919 g.IVS1+1312G>A, AKR1C3_11 rs2275928 g.IVS8+40A>G, SCARB1_03 rs4765621 g.IVS1-18462G>A, and PLA2G6_08 rs2016755 g.IVS3-309T>C was associated with elevated bladder cancer risk (logistic regression adjusted OR 2.6 (95% CI 2.1–3.2), MDR accuracy 0.53, cross-validation consistency 6/10). Within this MDR-selected group, we tested possible combinations of gene-gene interactions and found a significant relationship between SCARB1_03 and AKR1C3_11 (interaction p value = 0.04). This MDR analysis also picked the strongest two-factor interaction from the Kullback–Leibler selected genes, between AKR1C3_11 and IL10_06 (MDR accuracy 0.56, cross-validation consistency 6/10), adjusted OR 1.6 (95% CI 1.3–2.0), gene–gene interaction p value = 0.0004. Using the tuned ReliefF filter in MDR, we also identified CTNNB1_02 rs11564452 g.IVS7-562A>T, SCARB1_03 rs4765621 g.IVS1-18462G>A, SLC23A2_31 rs12479919 g.IVS1+1312G>A, and BARD1_02 rs1129804 g.Ex1+44C>G as a high risk combination (MDR accuracy 0.56, cross-validation consistency 10/10). The logistic regression adjusted OR was 2.3 (95% CI 1.7–2.9) for the high versus low risk allele combination. Of these, SCARB1_03 and SLC23A2_31 showed the strongest evidence of a gene–gene effect (interaction P = 0.02) on bladder cancer risk.

Additionally, we analyzed the data using a χ² filter on the entire dataset. Entering the χ² top-ranked 40 genes into MDR revealed an interaction between SLC19A1_01 rs1051266 g.Ex4-114T>C p.H27R, and IGF2AS_04 rs3741212 g.Ex1+112A>G, MDR accuracy 0.60, cross-validation consistency 10/10. The combined any variant adjusted odds ratio for this gene–gene combination from logistic regression was 1.4 (95% CI 1.1–1.7) (interaction P < 0.0001).

Survival models

We used the log-rank test to identify the top four individual SNPs likely to be related to bladder cancer case survival (Table 4). Cox-regression analysis (adjusted for age, gender, smoking status, stage/grade, and treatment) suggested that lower survival was associated with having SNPs in the T lymphocyte regulator CD80_04, the apoptosis regulator BCL2L1_03, and the ERCC4_01 nucleotide excision repair gene. The polycyclic aromatic hydrocarbon (PAH) metabolism SNP, EPHX1_15 was associated with better prognosis (Fig. 1). The effects did not vary significantly by smoking status (e.g. ERCC4 among never smokers HR 10.6 (95% CI 2.8–38), ever smokers HR 1.9 (95% CI 1.2–3.1), interaction P = 0.2).

Table 4.

Individual SNPs selected for relationship with bladder cancer survival

Function	SNP	Genotype	Cases		Deaths		Bladder cancer HR (95% CI)	Log rank test p value
			N	%	N	%
Surface antigen	CD80_04	wt	321	77	116	68	1.0 (ref)	0.008
	rs9282638	het	93	22	54	31	1.9 (1.4–2.8)
	g.IVS2−56G>A	var	5	1	1	1	1.5 (0.2–10.9)
		any var	98	23	55	32	1.9 (1.4–2.7)
Apoptosis	BCL2L1_03							0.01
	rs1994251	wt	265	63	85	50	1.0 (ref)
	g.IVS2+22130A>C	het	133	32	76	45	1.5 (1.1–2.1)
		var	21	5	8	5	1.5 (0.7–3.3)
		any var	154	37	84	50	1.5 (1.1–2.1)
Detoxify/activate polycyclic aromatic hydrocarbons	EPHX1_15 rs2854461	wt	164	14	85	50	1.0 (ref)	0.02
	g.−4786A>C	het	198	47	74	43	0.8 (0.5–1.1)
		var	57	14	13	7	0.5 (0.2–0.9)
		any var	255	61	87	50	0.7 (0.5–1.0)
DNA repair	ERCC4_01							0.1
	rs1800067	wt	371	89	143	83	1.0 (ref)
	g.Ex8+31G>A	het	45	11	27	16	1.7 (1.1–2.8)
	p.R415Q	var	3	1	2	1	5.6 (1.4–23.4)
		any var	48	12	29	17	1.8 (1.2–2.9)

Analyses were adjusted for gender, age, smoking, treatment, tumor stage, grade, tumor size. Genotype data were missing for CD80_04 27 cases; BCL2L1_0329 cases; EPHX1_15 26 cases; ERCC4_01 26 cases

wt wildtype, het heterozygous, var variant, any variant heterozygous and variant, g. gene, Ex exon, p. protein

Fig. 1.

Bladder cancer survival in relation to SNPs. Kaplan–Meier plots show survival (y-axis) versus years from diagnosis (x-axis) by a EPHX1_15 genotype, b IL8RB_01 genotype, c PGR_05 genotype, and d CASP9_01 genotype. Black lines are homozygous wildtype, red are homozygous variant, green are heterozygous

Gene Set Assessment allowed us to use all the SNP data from each gene as a group to identify the genes that most strongly predicted survival by log-rank test (Table 5). We then evaluated the individual SNPs within the top 10 ranked genes using adjusted Cox-regression models. The variant form of the transcriptional regulator GATA3_29 was associated with poor survival HR 2.8 (95% CI 1.6–5.0). The same apoptotic gene BCL2L1 selected by this GSA analysis was also picked in the individual SNP log-rank test (BCL2L1_03). An additional SNP in this gene, BCL2L1_01, was also related to poor bladder cancer survival [any variant HR 1.6 (1.1–2.2)]. This GSA analysis also predicted poor survival for individuals with SNPs in the angiogenic gene IL8RB_01 [any variant HR 1.9 (1.3–2.7)] (Fig. 1), the mitotic genes PLK1_15 [any variant HR 1.3 (0.9–1.8)] and STK6_06 [any variant HR 1.4 (0.9–1.9)], the metabolism gene UGT1A1_24 [any variant HR 1.4 (1.0–1.9)], the apoptotic gene CASP9_01 [any variant HR 1.7 (1.2–2.4)] (Fig. 1) and better survival with the steroid receptor PGR_05 [any variant HR 0.7 (0.5–1.0)] (Fig. 1), and the telomerase TERT_02 [any variant HR 0.8 (0.6–1.2)]. The selected SNPs did not diVer significantly by smoking status (data not shown).

Table 5.

SNPs in genes selected by gene set analysis for relationship with bladder cancer survival

Function	SNP	Genotype	Cases		Deaths		Bladder cancer HR (95% CI)	Log rank test p value
			N	%	N	%
Transription factor	GATA3_29							<0.001
Growth control	rs528778	wt	261	62	104	61	1.0 (ref)
Epithelial cell differentiation	g.IVS4+582C>T	het	150	34	51	30	0.9 (0.67–1.3)
		var	8	2	16	9	2.8 (1.6–5.0)
		any var	158	36	67	39	1.1 (0.8–1.5)
Interleukin 8 receptor	IL8RB_01							0.003
	rs1126579	wt	194	47	56	33	1.0 (ref)
	g.Ex3+1235T>C	het	173	42	82	48	1.7 (1.2–2.5)
		var	50	12	34	20	1.7 (1.1–2.8)
		any var	233	34	116	68	1.7 (1.2–2.5)
Mitotic kinase	PLK1_15							0.1
Spindle formation	rs40076	wt	260	62	91	53	1.0 (ref)
	g.IVS3+26A>G	het	138	33	70	41	1.2 (0.8–1.7)
		var	21	5	10	6	1.0 (0.5–2.1)
		any var	159	38	80	47	1.2 (0.8–1.6)
Glucuronidation	UGT1A1_24							0.1
Lipophilic molecules	rs1042640	wt	270	65	93	55	1.0 (ref)
	g.Ex5−402G>C	het	123	30	67	40	1.4 (1.0–2.0)
		var	22	5	8	5	1.4 (0.6–3.1)
		any var	155	35	75	45	1.4 (1.0–2.0)
Progesterone receptor	PGR_05							0.005
	rs613120	wt	120	29	34	20	1.0 (ref)
	g.IVS2−1671T>C	het	207	50	81	47	1.2 (0.8–1.9)
		var	91	22	57	33	1.4 (0.9–2.3)
		any var	298	77	138	80	1.3 (0.8–2.0)
Mitotic kinase	AURKA/STK6_06							0.1
Chromosome distribution	rs6024840	wt	176	56	68	47	1.0 (ref)
	g.IVS7−80T>C	het	123	39	70	48	1.4 (1.0–2.1)
		var	15	5	8	5	1.1 (0.5–2.5)
		any var	138	44	78	53	1.4 (1.0–2.0)
Telomerase	TERT_02							0.1
	rs6024840	wt	158	39	82	48	1.0 (ref)
	g.IVS7−80T>C	het	197	48	75	44	0.8 (0.6–1.1)
		var	55	13	15	8	0.5 (0.3–1.0)
		any var	252	61	90	52	0.8 (0.5–1.0)
Apoptosis	CASP9_01							0.003
	rs1052576	wt	203	50	55	34	1.0 (ref)
	g.Ex5+32A>G	het	148	36	77	46	1.3 (0.9–2.0)
	p.Q221R	var	59	14	35	21	1.4 (0.8–2.4)
		any var	207	50	112	67	1.3 (0.9–2.0)

Analyses were adjusted for gender, age, smoking, treatment, tumor stage, grade, tumor size. Genotype data were missing for GATA3_29 27 cases; IL8RB_01 28 cases; PLK1_15 27 cases; UGT1A1_24 34 cases; PGR_05 27 cases; AURKA 157 cases; TERT_02 35 cases; CASP9_01 40 cases

wt wildtype, het heterozygous, var variant, any variant heterozygous and variant, g. gene, Ex exon, p. protein

Discussion

A genetic component for bladder cancer susceptibility is likely since individuals with a family history have a higher incidence. Furthermore, although cigarette smoking is the strongest risk factor, it is not clear why some smokers get cancer and others do not (Ahsan and Thomas 2004). The magnitude of the association between the SNPs that are well accepted to be predictive (such as the N-acetyltransferases and DNA repair genes) and bladder cancer risk is relatively small. Thus, there are likely to be other causal genetic factors that remain to be discovered. Bladder cancer survival also varies widely between individual patients, even within the same stage and grade of the tumor. We hypothesize that this survival is in part genetically controlled, yet reliable predictors have not been identified or validated.

The occurrence and prognosis for bladder cancer and other relatively common diseases is most likely controlled by a combination of multiple genetic factors and exposures, rather than a single polymorphism. Our strategy was to predict the single SNPs that have the strongest relationships with bladder cancer incidence or survival. The magnitude of these relationships was then evaluated using traditional logistic regression analysis procedures. We then identified the combinations of the top ranked SNPs and genes using multi-factor selection algorithms for the most predictive high risk sets of alleles. These models were then assessed statistically by traditional methods. Using this approach, we identified a number of SNPs and SNP combinations that were associated with bladder cancer incidence and survival. Modifying SNPs occurred in biologically plausible genes that regulate metabolism, DNA repair, telomere maintenance, mitosis, inflammation, and apoptosis.

Alcohol dehydrogenase 1C (ADH1C) metabolizes multiple substrates, most notably ethanol. Consistent with our data, a previous study in the Netherlands suggested an increased risk of bladder cancer among variants, but no interaction with alcohol consumption (van Dijk et al. 2001). UGT1A1 is a phase 2 enzyme which metabolizes lipophilic molecules (e.g. steroids, bilirubin, hormones, drugs, and aromatic amines such as benzidine) via glucuronidation for excretion (Zenser et al. 2002) and was related to worse survival. Polymorphisms in the one-carbon methyl-metabolism gene methylenetetrahydrofolate reductase (MTHFR) have been well studied in relation to bladder cancer risk (Karagas et al. 2005; Moore et al. 2004; Ouerhani et al. 2007; Vineis et al. 2007). The methyl-group metabolism pathway regulates methylation of DNA CpG islands, thus controlling expression and silencing of tumor suppressor and oncogenes (Paz et al. 2002). Our study found increased risk for variants in a related methyl-metabolism dehydrogenase enzyme that provides the 5,10 methylenetetrahydrofo-late pool (MTHFD2). This result is consistent with studies showing that MTHFD2 SNPs modify gastric and lung cancer risk (Liu et al. 2008; Wang et al. 2007).

Telomerase activity is critical for chromosomal stability and has been associated with bladder cancer risk (Alvarez and Lokeshwar 2007; Xing et al. 2007). TERT1 protein levels are a rate limiting step in telomerase activity and were associated with better survival in a previous study of 132 bladder cancer cases (P = 0.007) (Ito et al. 1998; Mavrommatis et al. 2005). Our study indicated better survival for TERT variants. Variants of telomerase-associated protein 1 (TEP1), a telomerase that adds new telomeres to the ends of chromosomes, protecting them from degradation, were associated with increased bladder cancer risk in our study, and in other cancers in a prior study (Savage et al. 2007).

Improper distribution of chromosomes during mitosis can result from modified expression of mitotic kinases, including Aurora kinase A (AURKA), and leads to chromosomal instability and aneuploidy in bladder cells (Fraizer et al. 2004). Subjects with the AURKA non-synonymous polymorphism in exon 11 had significantly greater susceptibility to bladder cancer in our study. Allelic variation in AURKA has been associated with AURKA expression levels and increased risk of multiple cancers in a meta-analysis of 9,549 cases (Ewart-Toland et al. 2005; Matarasso et al. 2007). AURKA amplification is very prevalent in bladder tumors (64%) and tumor AURKA gene and protein expression levels are associated with tumor recurrence and shorter survival time (Comperat et al. 2007; Denzinger et al. 2007; Schultz et al. 2007). The previously examined AURKA polymorphism T91A was not associated with bladder cancer prognosis in a study of 135 patients (Schultz et al. 2007). Our results suggest a lower survival rate for variants at another locus in AURKA intron 7 and a possible relationship with polo-like kinase 1 (PLK1). Aurora kinase A activates PLK1, which regulates spindle formation as DNA damage arrested cells move into mitosis (Macurek et al. 2008). High PLK1 expression levels are associated with bladder tumor chromosomal instability and progression (Yamamoto et al. 2006).

Insulin-like growth factor 1 (IGF1) variants had increased risk of bladder cancer in our study. SNPs in the promoter and intronic regions of IGF1 predict circulating IGF1 levels and increase colorectal cancer risk (Palles et al. 2008; Wong et al. 2008). Insulin-like growth factor 2 antisense (IGF2AS/PEG1) is aberrantly methylated in colorectal cancers and is overexpressed in Wilms' tumors (Nishihara et al. 2000; Okutsu et al. 2000). IGF2AS variants interacted with the folate transporter, SLC19A1, to increase bladder cancer risk in our study. Similarly, SLC19A1 was related to increased bladder cancer risk in the Spanish population (Moore et al. 2007).

Bladder cancer survival was negatively associated with variation in CD80, one of the surface antigens that increase in expression following treatment with the immunotherapy drug bacillus Calmette-Guerin (BCG) that is commonly used to treat bladder cancer patients (Ikeda et al. 2002). Interleukin-8 receptor (IL8RB/CXCR2) variants were at decreased risk of developing bladder cancer, but also experienced poorer survival than wildtype individuals in our study. IL8RB is associated with inflammatory response mediating angiogenesis, and neutrophil migration (Matheson et al. 2006). IL8RB polymorphisms have previously been associated with breast cancer progression (Kamali-Sarvestani et al. 2007).

Our data also suggest an increased risk of bladder cancer among variants for the mutS homolog 6 (MSH6), which recognizes mismatched nucleotides and induces apoptosis if repair fails. Reduced expression, mutation of MSH6 and concomitant loss of mismatch repair activity have been observed in bladder tumors (Mongiat-Artus et al. 2006; Thykjaer et al. 2001). As in our study, MSH6 variants were at higher risk for malignant bladder, an observation made for other cancers as well (Campbell et al. 2008; Sanyal et al. 2007). We observed shorter survival among ERCC4 variants. Although genetic variation in the DNA repair gene ERCC4 was not independently associated with bladder cancer susceptibility, few have investigated its role in survival (Garcia-Closas et al. 2006; Matullo et al. 2005).

Variants in apoptosis genes also modulated bladder cancer survival. The BCL-2 family member, BCL2L1/BCLX regulates the mitochondrial membrane potential controlling release of apoptotic inducers and expression levels have been related to cancer progression (Gazzaniga et al. 1998; Kirsh et al. 1998). Variants of the early apoptosis cascade member cysteine peptidase, Caspase 9 (CASP9) experienced shorter survival in our study. Both BCL2L1 and caspase 9 may be involved in response to some types of chemotherapy treatment in the bladder (Yuan et al. 2002).

In addition to main-gene eVects, we also detected a possible interaction between peptidylprolyl cis/trans isomerase (PIN1) and Alpha-methylacyl-CoA racemase (AMACR) variants. PIN1 is a conditional tumor suppressor, which transduces phosphorylation of P53 and promotes cell death by mediating dissociation from the apoptosis inhibitor iASPP (Mantovani et al. 2007). PIN1 plays a critical role in the DNA damage checkpoint, in part by protecting p53 from the inhibitor MDM2. Thus, Pin1−/− cells do not undergo mitosis in response to DNA damage (Takahashi et al. 2008). AMACR is involved in cellular energy metabolism by the oxidation of branched-chain fatty acids and in the metabolism of lipophilic drugs (Lloyd et al. 2008). AMACR expression is associated with higher bladder tumor histopathologic grade, implying that endogenous fatty acids may be used for energy as tumors grow (Gunia et al. 2008). AMACR variations have previously been associated with prostate cancer risk (Levin et al. 2007). The combination of inhibition of the DNA damage checkpoint by PIN1 with AMACR modulated energy availability for tumor growth and could result in the synergistic effect on tumor risk that we observed.

Another interaction involved scavenger receptor class B (SCARB1), which regulates cellular cholesterol uptake and sequesters chloesteryl esters from high-density lipoproteins (Miquel et al. 2003) with the aldo-keto reductase, AKR1C3, which is strongly expressed in the bladder. AKR1C3 regulates lipophilic ligand access to hormone receptors and activates PAHs, which are known bladder carcinogens (Azzarello et al. 2008). Our finding of modified bladder cancer risk in New Hampshire corroborates the recently reported association between AKR1C3 polymorphisms and bladder cancer risk in Spain (Figueroa et al. 2008).

The applicability of these results to other populations of bladder cancer patients needs to be explored before these markers can be used to guide clinical practice. These data will be used to inspire investigation of the mechanistic impact of these SNPs on tumor biology. Intronic SNPs may be acting through alternative splice sites, miRNAs, conformational changes or linkage disequilibrium with another functional SNP. Based on our results, a thorough investigation into SNPs in the metabolism, telomere maintenance, mitosis, inflammation, and apoptosis pathways is warranted. Nevertheless, these findings highlight a number of genetic variations in known carcinogenic pathway genes that modify bladder cancer risk and prognosis. These results may provide new insight into methods of predicting bladder cancer susceptibility using SNP panels, identify targets for bladder cancer prevention, and guide individualized choice of treatment methodologies based on molecular pathway alterations.