A Genome-Wide Association Study Identifies A New Ovarian Cancer Susceptibility Locus On 9p22.2

Abstract

Epithelial ovarian cancer has a major heritable component, but the known susceptibility genes explain less than half the excess familial risk¹. We performed a genome wide association study (GWAS) to identify common ovarian cancer susceptibility alleles. We evaluated 507,094 SNPs genotyped in 1,817 cases and 2,353 controls from the UK and ~2 million imputed SNPs. We genotyped the 22,790 top ranked SNPs in 4,274 cases and 4,809 controls of European ancestry from Europe, USA and Australia. We identified 12 SNPs at 9p22 associated with disease risk (P<10⁻⁸). The most significant SNP (rs3814113; P = 2.5 × 10⁻¹⁷) was genotyped in a further 2,670 ovarian cancer cases and 4,668 controls confirming its association (combined data odds ratio = 0.82 95% CI 0.79 – 0.86, P_-trend = 5.1 × 10⁻¹⁹). The association differs by histological subtype, being strongest for serous ovarian cancers (OR 0.77 95% CI 0.73 – 0.81, P_trend = 4.1 × 10⁻²¹).

Women with a first-degree relative diagnosed with epithelial ovarian cancer have a three-fold increased risk of developing the disease². Environmental and genetic factors contribute to this increased risk, but studies of twins suggest that genetic factors are more important. BRCA1 and BRCA2 mutations confer high risk of ovarian cancer and are responsible for most families with three or more ovarian cancer cases. They account for less than half the excess familial risk^{3, 4} and it is likely that the residual risk is due to a combination of common and/or rare alleles that confer moderate to low penetrance susceptibility⁵.

Many recent studies have reported the identification of common alleles that confer low-penetrance susceptibility to common cancers including breast, prostate and colorectal cancers and melanoma^6–12. These studies all used a genome wide association study (GWAS) design in which the genotype frequencies of hundreds of thousands of single nucleotide polymorphisms (SNPs) distributed throughout the genome are compared between large numbers of cases and unaffected controls. In the current study, we conducted a three-stage GWAS to identify alleles associated with variation in the risks of invasive epithelial ovarian cancer (Table 1). In the first stage, we used the Illumina Infinium 610K array to genotype 620,901 SNPs that report on genetic variation across the genome in 1,890 ovarian cancer cases recruited throughout the UK (see Supplementary Table 1 online for details). We enriched for likely genetic heritability by including 47 cases from families with two or more ovarian cancer cases that had been screened negative for BRCA1 or BRCA2 mutations. After excluding 73 cases that failed to meet genotyping quality control criteria, the genotype frequencies for 1,817 cases were compared with the genotypes of 2,353 UK controls that had been analysed using the similar 550k array on the same genotyping platform as part of a GWAS for other phenotypes (Supplementary Table 1 online). All the subjects analyzed in stage 1 were of European ancestry. After excluding 80,327 SNPs that were not genotyped on controls or failed to genotype on cases, a total of 507,094 SNPs with a minor allele frequency (MAF) of at least 1% in controls passed genotype quality control criteria. We also evaluated an additional ~ 2million SNPs with genotypes imputed using the phase2 Hapmap data (CEU).

Table 1.

Summary of study design.

Stage	Study population	Genotyping method	No. SN Ps genotyped	No. SNPs passed QC	No. SNPs imputed	Number of Subjects 1
						Cases	Controls
Stage 1	UK	Illumina 610K	620,901	507,094	2,056,878	1,890 (73)	2,353
Stage 2	Europe, USA, Australia	Illumina iSelect	23,590	21,955	0	4,964 (131)	5,379 (142)
Stage 3	Europe, USA, Canada	Taqman or Illumina	1	1	0	3,089	5,340
		Total				9,943 (204)	13,072 (142)

¹The numbers of subjects that were excluded from analysis due to failing genotyping quality control criteria are presented in brackets.

Supplementary figure 1a shows the quantile-quantile (Q-Q) plot of the distribution of test statistics for comparison of genotype frequencies in cases versus controls (1 degree freedom (d.f) Cochran-Armitage trend test) for SNPs genotyped in stage 1. There was little evidence of any general inflation of the test statistics (estimated inflation factor λ₁₀₀₀ = 1.026 based on the bottom 90% of the distribution¹³). This is consistent with the population structure in the UK and for other GWAS that have used UK populations^{9, 14}.

In a second stage, 23,590 SNPs were genotyped in 10 studies comprising 4,964 cases of invasive epithelial ovarian cancer and 5,379 controls (Table 1) using the Illumina iSelect platform. We selected 22,790 SNPs based on the lowest P-trend in tests for association with ovarian cancer risk from the ~2.5 million genotyped and imputed SNPs, and 800 SNPs that reported on ancestry. Data for 273 subjects were excluded because they did not meet quality control criteria; thus, genotyping data were available on 4,833 cases and 5,237 controls. Five hundred and fifty nine cases and 428 controls in stage 2 were of non-European ancestry and excluded from the main analyses. Supplementary figure 1B shows the Q-Q plot for SNPs genotyped in stage 2. Again, there was little evidence of any general inflation of the test statistics for stage 2 (estimated inflation factor λ₁₀₀₀ = 1.005). In the combined analysis of stage 1 and stage 2 data, we identified 12 SNPs associated with ovarian cancer risk with P < 10⁻⁸ (Supplementary Table 2 online). All 12 SNPs were located in the same region on chromosome 9 (9p22.2). The strongest association was for rs3814113, which was the only SNP retained in a multivariate logistic regression model. We genotyped rs3814113 in an additional 3,089 cases of invasive epithelial ovarian cancer and 5,340 controls from ten studies that are part of the Ovarian Cancer Association Consortium (OCAC) (Supplementary Table 1 online). Four hundred and nineteen cases and 672 controls were of self-reported non-European ancestry and excluded from the analysis. These additional data reinforced the evidence of association for rs3814113; P-trend = 5.1 ×10⁻¹⁹ based on data from all three stages (Table 2).

Table 2.

Estimated odds ratio and tests of association for rs3814113 at chromosome 9 (9p22.2) based on combined data for subjects of European ancestry.

Tumor subtype	Stage	No. of Cases/ Controls	MAFa Case/Control	Per-allele ORb [95% CI]	Heterozygotec OR [95% CI]	Homozygoted OR [95% CI]	P-trend
All Cases	Stage 1	1817/2354	0.27/0.32	0.81 (0.74–0.89)	0.76 (0.67–0.86)	0.73 (0.58–0.91)	2.04×10−5
	Stage 2	4274/4809	0.28/0.33	0.79 (0.74–0.84)	0.80 (0.76–0.93)	0.70 (0.63–0.77)	1.98×10−13
	Stage 3	2670/4668	0.29/0.32	0.88 (0.81–0.94)	0.84 (0.77–0.93)	0.80 (0.68–0.95)	4.00×10−4
	Combined	8761/11831	0.28/0.32	0.82 (0.79–0.86)	0.80 (0.75–0.85)	0.70 (0.63–0.77)	5.10×10−19
Serous Cases	Stage 1	837/2354	0.25/0.32	0.71 (0.62–0.81)	0.70 (0.59–0.82)	0.52 (0.38–0.72)	1.44×10−7
	Stage 2	2483/4809	0.27/0.33	0.75 (0.70–0.81)	0.76 (0.68–0.84)	0.57 (0.47–0.68)	5.60×10−13
	Stage 3	1527/4668	0.28/0.32	0.83 (0.76–0.92)	0.81(0.71–0.92)	0.74 (0.60–0.92)	2.07×10−4
	Combined	4847/11831	0.27/0.32	0.77 (0.73–0.81)	0.76 (0.71–0.82)	0.61 (0.54–0.69)	4.10×10−21

^aMinor allele frequency in populations with European ancestry in the studies

^bOdds ratios (OR) and 95% confidence interval per copy of minor allele

^cOR for heterozygotes compared with common homozygotes

^dOR for rare homozygotes compared with common homozygote. Data highlighted with bold text are GWAS significant (P<5.0×10⁻⁸)

rs3814113 was associated with a decrease in the risk of ovarian cancer in carriers of the minor allele (per minor allele odds ratio (OR) = 0.82 95% confidence interval (CI) 0.79- 0.86). The effect size was similar in stages 1 and 2, but slightly smaller in stage 3. There was no heterogeneity in the OR estimates amongst studies for any stages (Table 2 and in Figure 1A). Based on an odds ratio of 0.82 and minor allele frequency of 0.32, this locus explains approximately 0.7% of the polygenic component of ovarian cancer risk. This estimate is also based on the assumption that the known high penetrance genes explain 40% of the excess familial risk and the unexplained component is polygenic. There was no significant difference between the risk of ovarian cancer in subjects of European and non-European ancestry (P = 0.83). However, the per-minor allele risk was slightly attenuated and not significant in the subjects of non-European ancestry (OR = 0.89, 95% CI 0.78–1.01, P = 0.077) (Supplementary Table 3 online).

Figure 1.

Genotype-specific risks of SNP rs3814113 for ovarian cancer by stage and by study - (a) all epithelial ovarian cancer cases, (b) serous ovarian cancer cases only. Results are based on analyses restricted to subjects of European ancestry.

We also evaluated the association for rs3814113 with ovarian cancer risk after stratifying cases by histological subtype. The strength of association increased when serous cases (n = 4,847), the most common histological subtype, were considered alone (OR = 0.77 95% CI 0.73–0.81, P-trend = 4.1×10^-21) (Supplementary Table 4 online and Figure 1B). When the analysis was restricted to serous cases the effects were similar between European and non-European subject (non-European OR = 0.79, 95% CI 0.66–0.94, P = 0.007). We only detected marginal evidence of association for rs3814113 in 1,320 cases diagnosed with endometrioid ovarian cancer (OR = 0.86 95% CI 0.79–0.94, P-trend = 0.001), and no association for patients with mucinous (n = 626) or clear cell (n = 628) ovarian cancer (Supplementary Table 4 online). However, the small numbers of mucinous or clear cell cases limited the power to detect modest effects. Compared to the non-serous subtypes, the ovarian cancer risk in serous cases was significantly lower (P= 7.8× 10⁻⁵). There was also some suggestion of a bigger effect in older women (P-trend = 0.006 for ovarian cancer overall and P-trend = 0.044 for serous type ovarian cancer) (Supplementary Table 5 online). There was no significant difference in genotype frequency for cases reporting a family history of ovarian cancer compared with cases with negative family history of ovarian cancer (P = 0.59)

All 12 SNPs in the region that were identified after stage 2 were in the same LD block and were associated with a decreased risk of ovarian cancer, perhaps suggesting that susceptibility is driven by a single correlated variant within the region (Figure 2), although it is possible that there are multiple independent SNPs all correlated with the best markers. rs3814113 may be the causal SNP on 9p22.2 or it may be a marker in linkage disequilibrium (LD) with the true functional variant or haplotype. Neither rs3814113 nor the highly correlated SNP, rs4445329, (r² > 0.99) are located within an open reading frame or an intronic region of any gene. The nearest genes are BNC2 (basonuclin 2), CNTLN (centlein, a centrosomal protein) and hypothetical gene LOC648570. rs3814113 is ~44kb upstream of BNC2, ~128kb upstream of LOC648570 and ~220kb downstream of CNTLN. Eight of the associated SNPs in the region are located within intron 2 of BNC2 (Figure 2). BNC2 encodes DNA-binding zinc-finger protein that is highly conserved across vertebrates suggesting it is an important regulatory protein for DNA transcription¹⁵. The gene exhibits extensive transcriptional variability; it has six promoters and has the potential to generate up to 90,000 mRNA isoforms encoding more than 2,000 different proteins¹⁶. The Genevar project provides data on gene expression of BNC2 in lymphocyte derived cell lines from the CEU population based on nine probes in the region¹⁷. There was no association for any of these probes and genotype at the top 12 SNPs. Also none of the top 12 SNPs appear to be near predictable or known enhancer binding sites or splice sites using PupaSNP (http://www.pupasnp.org/). BNC2 is highly expressed in reproductive tissues (ovary and testis) and may play a role in the differentiation of spermatoza and oocytes¹⁸. There is little evidence of a role for BNC2 in cancer development, although there is a report of 7~9 fold up-regulated expression in basal cell carcinoma compared to normal basal cells¹⁹. Resequencing of the 9p22.2 region and further genotyping in ovarian cancer cases and controls will be needed to clarify the likely causal variant(s).

Figure 2.

Map of associated regions on 9p22.2 (nucl. 16655021-17155021). The top panel shows the LD blocks around rs3814113 (±250 kb) for SNPs with MAF≥ 0.05 based on Hapmap CEU data. Squares on the LD block indicate the correlation (r²) between SNPs on a greyscale (darker squares = higher correlations). Approximate location of BNC2 (nucl. 16860786-16399502), CNTLN (nucl. 17125038-17493915) and LOC648570 (nucl. 16777433-16775576) were inferred from the NCBI reference assembly (solid line). The lower panel shows the highlighted LD block in greater detail. The star indicates the location of the most strongly associated SNP rs3814113. The black circles show the position of the other 11 SNPs that reached genome wide significance. The numbers refer to the SNP identifiers in Supplementary Table 2.

We found no additional susceptibility loci reaching genome wide significance (P<10⁻⁸) in this study. Positive associations have previously been reported in candidate gene studies, but none of these reached genome-wide significance. Supplementary Table 6 shows some of the more notable published associations^20–29 together with the results for the same SNPs from this study. In general the reported associations and our results are consistent, but none of the associations in this study ranked highly enough for the SNP to be included in the Stage 2 genotyping. The power to have identified rs3814113 depends on the true relative risk. Assuming a relative risk per-minor allele of 0.82 (combined data estimate), power was > 90% at genome-wide significance, suggesting that common alleles of larger effect are unlikely to have been missed. However, if the true per-allele relative risk is 0.88 (stage 3 estimate) then the power was only 20% suggesting that susceptibility alleles with more modest effects remain to be identified. Furthermore, power to detect less common alleles will be limited unless the effect on risk is greater. For example, we have 90 percent power at genome-wide significance to detect a risk allele with 5 percent frequency that confers a relative risk of 1.44. This would be consistent with the findings for other common cancer types (e.g. breast cancer and prostate cancers)²⁹. In addition, disease heterogeneity may have limited our power to identify additional susceptibility alleles. In the primary analysis, we considered ovarian cancer as a single disease phenotype, but the effects of association for rs3814113 varied when cases were stratified by histological subtype. Different subtypes of ovarian cancer have different biological properties and this finding supports previous studies that suggest susceptibility due to germline genetic variation may be subtype specific. For example, serous and endometrioid cancer are relatively more common in BRCA1 and BRCA2 mutation carriers whereas mutations in the DNA mismatch repair genes are more frequently associated with mucinous ovarian cancers^30–32. Additional ovarian cancer susceptibility loci may exist that are associated with specific histological or molecular subtypes³³. However, the power to identify alleles for the rarer subtypes (endometrioid, mucinous and clear cell ovarian cancers) was limited by the numbers of cases in the study. It is likely that pooling of data from multiple ovarian cancer GWAS will enable additional susceptibility alleles for invasive ovarian cancer in general and for specific sub-types to be identified.

The identification of common ovarian cancer susceptibility variants may have clinical implications in the future for identifying patients at greatest risk of the disease. Survival rates in patients diagnosed with ovarian cancer are poor - approximately 70% of patients are diagnosed with late stage disease and less than 40% of these cases survive more than 5 years after their diagnosis. The efficacy of using multi-modal approaches to early detection of the disease are limited, but may be improved by using genetic risk profiling to identify a subset of the population that would benefit most from earlier disease detection. The benefits of a similar approach have recently been modelled in breast cancer⁵. Identifying genetic variants that cause ovarian cancer may also improve our understanding of the underlying biology of ovarian cancer, potentially leading to the development of more effective, individualised therapies. For example, the identification of the highly penetrant susceptibility genes BRCA1 and BRCA2, and their subsequent functional characterisation has since led the to development of a potential novel therapy for patients deficient in BRCA1/2 function based on inhibition of the poly (ADP-ribose) polymerase PARP DNA repair pathway^34,35. The 9p22.2 region is the first common susceptibility locus for ovarian cancer to be established. Understanding the mechanisms by which this susceptibility is mediated should improve our understanding of the biology of ovarian cancer, and may lead to new approaches to treat or prevent the disease.

METHODS

Study design

The ovarian cancer case-control studies that participated in stages 1, 2 and 3 are summarized in Supplementary Table 1 online. Stage 1 comprised invasive epithelial ovarian cancer cases from UK and genotype data of UK controls from GWAS of other phenotypes. Stage 2 comprised 10 studies from the OCAC. Stage 3 comprised 10 additional studies from the OCAC. For all studies we have data on disease status, age at diagnosis and date of blood draw, self-reported ethnic group and histological subtype. All but 5 studies provided information of reported first-degree family history of ovarian cancer.

Genotyping

Genotyping for stage 1 cases was conducted using the Illumina Infinium 610K array at Illumina Corporation. Existing data from two sets of controls, genotyped on the Infinium 550k array, were used in stage 1 analyses: the Welcome Trust Case-Control Consortium 1958 birth cohort ¹⁴, and a national colorectal control study³⁶ using Illumina platform Hap550 array. All cases were from the UK and confirmed as invasive epithelial ovarian cancer. Quality control criteria were applied separately to the cases and each control sets because they were genotyped separately. SNPs were excluded if (1) they deviated from hardy-Weinberg equilibiurm (HWE) at P < 10⁻⁴ or (2) had a MAF < 1%, or (3) MAF was between 1% and 5% and call rate < 99% or (4) MAF > 5% and call rate < 95%. We also rejected SNPs if a test for trend by genotype between the two control sets was significant at P < 10⁻⁴. This led to 33,479 SNPs being excluded and 507,094 SNPs passing QC. Genotyping the 10 studies in stage 2 was conducted using an Illumina iSelect array at Illumina Corportation. We excluded SNPs (n=1,635) for the stage 2 data if the sum of the test statistics for deviation from HWE for the 10 studies was significant at p < 10⁻⁵ or if the SNPs had a call rate of <95% or if the MAF < 0.5%. A total of 21,955 SNPs were available for data analysis in stage 2.

We utilized only samples with called genotypes on at least 80% of SNPs. Seventy-three samples were excluded from Stage 1 and 273 from Stage 2. Nineteen samples were included as duplicates in stage 1 and genotype concordance rate for these pairs was 99.99%. One hundred and twenty-two samples were included as duplicates in stage 2 and duplicate concordance rate was 99.99%. Six studies in stage 3 were genotyped for rs3814113 by Taqman using ABI Prism 7900HT sequence detection system at each laboratory. For three studies (TOR, NCO, MAY) genotype data were available from an independent, ongoing GWAS study that is also used the Illumina Infinium 610K platform. Genotyping and QC was performed at Mayo Clinic genotyping shared resources. For Taqman genotyping quality control, we compared genotype call rates and concordance by study and overall. We used the following criteria as a measure of acceptable genotyping: (1) > 3% sample duplicates included; (2) concordance rate for the duplicates ≥ 98%; (3) overall call rate (by study) > 95%; (4) call rates > 90% for each individual 384-well plate and (5) no deviation from HWE in controls (P>=0.05). Genotyping consistency across laboratories using Taqman was also evaluated by genotyping a common panel of CEPH-Utah trios including 90 individual DNA samples, 5 duplicate samples and 1 negative control (http://ccr.coriell.org/). The concordance of genotyping results between the centres was required to be greater than 98% in order for the genotype data to be included. The genotyping results from all studies in the stage 3 met the above criteria and were included in the final analysis.

Statistics

Population stratification

For the stage 1 samples, we used the program LAMP³⁷ to assign intercontinental ancestry based on the Hapmap (release #22) genotype frequency data for European, African and Asian populations. Samples with less than 90% European ancestry were excluded from the analysis (n=73). For the stage 2 data, 800 SNPs that are known to be predictive of ancestry (“Ancestry Informative Markers”) (AIMs) were genotyped. We again used LAMP and the Hapmap data (release #23) on European (CEU), African American (ASW), East Asian (JPT-CHB-CHD), Mexican (MEX) and Indian (GIH) populations to estimate ancestry. Subjects with less than 90 percent European ancestry were excluded from the main analyses (n= 987). We then used the AIMs to calculate principal components for the subjects of European ancestry. The first principal component explained 0.42 percent of the variability and was included as a covariate in subsequent association analyses. Subsequent principal components were not included as they and explained less variability and there was little difference in their Eigenvalues.

Imputation

We imputed missing genotype data for all the common variants in the Hapmap data for two reasons. Firstly the stage 1 cases and controls were genotyped using slightly different SNP sets and different SNPs may have failed QA in different sample sets. Secondly, imputing SNPs that have not been genotyped increases genome coverage and may improve power. We used an in-house method that combines the features of fastPHASE ³⁸ and IMPUTE ³⁹ to impute the ungenotyped or missing SNPs, utilising the phase2 Hapmap data (CEU) which contains phased haplotypes for 60 individuals on 2.5 million SNPs. For each imputed genotype the expected number of minor alleles carried was estimated (weights). Genotyped SNPs were assigned weights of 0, 1 or 2 (actual number of minor alleles carried). We estimated the accuracy of imputation by calculating the estimated r² between the imputed and actual SNP ⁴⁰. SNPs with r²<0.64 were excluded (n = 152,401) leaving a total of 2,563,972 SNPs for stage 1 analysis.

Statistical Tests of Association

We used logistic regression to test for association between genotype and case -control status using the imputed weights and the ethnicity estimates as covariates. SNPs were selected for replication in Stage 2 based on the weighted ranked test statistics from the Stage 1 analysis. The weights were based on whether the SNP was genotyped or imputed (I_genotype = 1 if genotyped_” 0 otherwise), the accuracy of imputation (r²), the association statistic (T) and the design score (s) using somewhat arbitrary formula ((1+0.1I_genotype)*T*r²+2s). SNPs that were correlated with a higher ranked SNP with r²>0.8 were excluded except for those correlated with the 1000 highest ranked SNPs. Using this method we took forward 22,790 SNPs as candidates for association of which 5,380 were purely imputed (not-genotyped). For the stage 2 analysis, we tested for association by performing logistic regression using the imputed values from stage 1, combined with the genotyping results for stage 2. We corrected for the first principal component and the ethnicity estimates in the second stage analysis (all subjects in the first stage were selected to be European) and stratified by study using a Wald test. We corrected for ethnicity in stage 3 using self reported ethnicity and stratified by study. A subgroup analysis was used to compare genotype-specific risks by disease subgroup with controls. The effect of age group, family history and population of origin (European and non-European) was assessed similarly. Modification by these sub-groups was tested by fitting a SNP by subgroup interaction term in a logistic regression model.