Breast and ovarian cancer are two of the most common causes of cancer in women in the United Kingdom (Office for National Statistics). Together, they account for about a third of all new cancer cases and a quarter of cancer deaths. Positive family history is a well-established risk factor for both diseases: the risk to first-degree relatives of a case is about 2 times the population risk [1–3]. Most of the excess familial risk associated with breast and ovarian cancer is likely to be genetic in origin [3, 4]. However, only a small proportion of this risk is accounted for by known highly predisposing genes, BRCA1 and BRCA2, while the remainder might be explained by a combination of weakly predisposing alleles [5–10]. EMSY (C11orf30) is a novel gene that could be involved in low-level predisposition to breast and ovarian cancer [11]. Its recent discovery generated widespread interest [12, 13]. The gene maps to chromosome 11q13, spans 103.3 kilobases and comprises 20 coding exons. EMSY is amplified in 12% of breast cancers and 17% of high-grade ovarian cancers and its amplification has been associated with an increased risk of relapse as well as decreased survival in breast cancer patients [11, 14]. Furthermore, the EMSY protein silences the transcriptional activation potential of BRCA2 exon 3, a region deleted in a Swedish breast and ovarian cancer family [11, 15].

The case-control study design is well suited to the identification of small-effect genes that are likely to underlie common, complex diseases such as breast or ovarian cancer: a difference in allele frequency is sought between affected individuals and unrelated controls [16]. Two approaches have been proposed. The traditional, hypothesis-driven approach is to investigate single-nucleotide polymorphisms (SNPs) in coding regions, since they are more likely to have a functional role and to influence directly the traits under study [17]. The alternative, indirect approach is to select a set of haplotype-tagging SNPs (htSNPs); htSNPs are informative polymorphisms that best characterize haplotype diversity and therefore genetic variation within the gene [18, 19]. They serve as markers to detect associations between a particular region and diseases, whether or not the SNPs themselves have a functional effect [20, 21]. It is not necessary to genotype all possible polymorphisms because the alleles of SNPs that are physically close to each other tend to be correlated with each other: they are in linkage disequilibrium (LD) [22]. The HapMap online database allows the indirect approach to be applied readily to many genes or regions [23]. By December 2004, the database held the genotypes of 90 individuals with northern and western European ancestry for over 850'000 SNPs.

We used a genetic association study design to determine if variation in EMSY was associated with breast or ovarian cancer risk. In order to have good power to detect small relative risks, we restricted our attention to common SNPs and haplotypes (frequency ≥ 5%). We first selected htSNPs using HapMap data. We also screened the gene promoter and coding regions in order to identify additional polymorphisms that are likely to be functional, as HapMap might be insufficient to tag genetic variation comprehensively. The selected SNPs were then genotyped in two large case-control sets (one breast cancer set and one ovarian cancer set) of white British women.

This is the first association study reporting results on EMSY, a gene of importance through its interaction with BRCA2 and its amplification status in tumours. We found no association between any of the EMSY genotypes or associated haplotypes and risk of breast or ovarian cancer in a white British population. We could have failed to observe a true association because of a Type II statistical error, but the large size of our study gives us high statistical power and strongly reduces the likelihood that our results are influenced by chance fluctuations in the case or control genotype frequency [30].

An alternative reason for failure to observe a true association could be that our set of tagging SNPs are poor markers of a true causal variant, which would either be one of the known SNPs in the gene or an as yet unidentified variant. In HapMap, common EMSY haplotypes were tagged by 5 SNPs. However, an assay for rs7106446 could not be designed and thus our htSNP set was suboptimal. Where a tagging SNP is used as a marker for a true disease-predisposing SNP the effective sample size is proportional to the bivariate correlation coefficient (r²) between the marker and causal SNPs [31]. r² _s is the squared correlation coefficient between multi-locus haplotypes and individuals SNPs and is analogous to r². In order to establish how well we had tagged the known SNPs with our set of tagging SNPs, we calculated r² _s [26] between the 4 selected htSNPs and the 18 remaining HapMap polymorphisms. Seventeen (94%) SNPs were tagged with r² _s > 0.8 but 1 SNP, rs7106446, was tagged with r² _s = 0.46. Loss of power was therefore marginal for all HapMap SNPs except rs7601446; for a SNP tagged with r² _s = 0.85, we had 89% power at the 1% significance level to detect a dominant allele with a frequency of 0.3 that confers a relative risk of ovarian cancer of 1.5.

It is also possible that we have not adequately tagged an unidentified, disease-predisposing SNP. Whole-gene resequencing across a sample population would be required to identify all existing polymorphisms and allow investigators to select htSNPs that tag all genetic variants. The HapMap project does not re-sequence the genome; it validates SNPs from the dbSNP public database, aiming for a density of polymorphisms that cover the whole of genetic variation across the human genome. Comprehensive tagging requires a genotyping density of around 1 SNP per 2.5 kb [32]. The 1 SNP per 5 kb density available for EMSY in HapMap might therefore be insufficient. In order to identify additional SNPs that are most likely to be functional [17], we screened the gene promoter and coding sequence with DHPLC, a technique with an estimated sensitivity of 94% [33]. Two SNPs near exon-intron boundaries were identified but neither was associated with breast or ovarian cancer.

We also assessed how well a set of htSNPs would have tagged any unidentified SNPs using a leave-one-out cross validation procedure on the HapMap data: each of the 22 known SNPs were dropped in turn and htSNPs selected from the remaining 21, thus simulating unidentified polymorphisms [32]. The ability of htSNPs to tag the dropped SNP was then evaluated by calculating r² _s [26]. Mean r² _s was 0.94. Twenty-one (95.4%) out of 22 dropped SNPs were tagged with r² _s > 0.8, and only 1 (4.6%) was tagged with 0.4 < r² _s < 0.8. Over 95% of the common genetic variation in EMSY should therefore be well represented by tagging polymorphisms. We are therefore likely to have tagged any common, functional variants present in our population. After this study was completed and the first version of the manuscript submitted, genotyping data in a white American population for rs3753051, a synonymous coding SNP in exon 19, were released in dbSNP. We were able to assess how this polymorphism was tagged by our set of SNPs as genotypes from the same individuals were also available for 5'up t>g, IVS9 a>g, IVS16 a>g and 3'down c>t; SNP rs3753051 was perfectly tagged (r² = 1) by 5'up t>g.

This study design can not exclude the involvement of a rare allele in predisposition to breast or ovarian cancer; for example, CHEK2*1100delC, has a frequency of around 1% and was recently shown to confer a two-fold increased risk of breast cancer [34]. Our study set would be too small to detect the effect of such an allele if it doubled the risk of ovarian cancer. Some authors have advocated the use of histopathologic or demographic data that subclassify individuals in order to identify homogeneous subsets for analysis [35]. In the absence of any main effect or strong biological rationale, we have not carried out subgroup analyses as much larger sample sizes would be required to obtain reliable results. The number of possible post-hoc, subgroup analyses is large and there is a strong possibility that one or more tests will be statistically significant simply by chance [36].

We are reporting results for a set of htSNPs selected from HapMap. We used genotypes for the CEPH samples to choose htSNPs. According to the HapMap website, it is unclear how accurately the CEPH samples reflect the patterns of genetic variation in people with northern and western European ancestry. Our results suggest that they correctly predict genetic variation in our white British population: allele frequencies in the breast study controls were similar to those obtained from HapMap (P = 0.57, 0.99, 0.88 and 0.85 for 5'up t>g, IVS7 g>a, IVS16 a>g and 3'down c>t, respectively), thus strengthening the argument for a widespread use of the database for htSNPs selection. A predisposing SNP might have a differential effect in another ethnic group via gene-gene or gene-environment interactions, although in a recent study of the genetic effects for 43 validated gene-disease associations across 697 study populations of various descents, Ioannidis et al. concluded that, even if frequencies of polymorphisms varied among populations, their biological impact on the risk for common diseases should be consistent across traditional ethnic boundaries [37, 38].

We saw no association between common SNPs in EMSY or their associated haplotypes with risk of breast or ovarian cancer in two large study sets of white British women.