A genome-wide association scan on estrogen receptor-negative breast cancer

Breast cancer is a heterogeneous disease and can be characterized on the basis of estrogen receptor (ER) expression in the tumour cells. The two breast cancer subtypes (ER-positive and ER-negative) are generally considered as biologically distinct diseases and have been associated with remarkably different gene expression profiles [1, 2]. ER status is important clinically, and is used both as a prognostic indicator and treatment predictor since it determines if a patient may benefit from anti-estrogen therapy. Approximately one-third of all breast cancers are ER-negative, and cancers of this ER subtype are highly age-dependent and generally have a more aggressive clinical course than hormone receptor-positive disease.

Estimates show that close to a third of the total risk of breast cancer may be attributed to heritable factors [3]. Several large-scale genome-wide single nucleotide polymorphism (SNP) association studies (GWAS) have identified multiple susceptibility loci for breast cancer [4‐11], but it is estimated that the currently known common risk variants identified by this approach explains only 5.8% of the proportion of familial risk of breast cancer.

Aside from traditional agnostic SNP studies, pathway-based approaches have also emerged in the recent GWAS literature [12‐20]. These novel methods have been developed to mine modest association signals from genome-wide SNP data using prior knowledge on biologically pathways and networks, and have the potential to complement traditional agnostic SNP approaches to provide fertile grounds for follow-up studies of both a genetic and molecular nature. Subtypes of breast cancer, to our knowledge, have not been studied using a pathway-based approach.

Although many of the SNPs identified for breast cancer through GWAS scans have been found to be more strongly associated with ER-positive disease than ER-negative disease [21, 22], there is no quantitative assessment on whether breast cancers of the two different ER subtypes share a polygenic component. In this study, we performed a genome-wide association scan on 617 ER-negative cases and 4,583 controls, the first of its kind, and examined 285,984 SNPs for common variants and biological pathways associated with this unique subtype of breast cancer. We also searched for evidence that ER-negative breast cancer is distinct from ER-positive breast cancer by assessing the amount of shared polygenic variation between the two breast cancer subtypes.

Full methods accompany this paper in Additional file 1.

In this study, we tested the association of 285,984 loci with ER-negative breast cancer in two independent populations consisting of a total of 617 cases and 4,583 controls. It appears that the overall population substructure was adequately accounted for, since a systematic deviation from the expected distribution was not observed in the quantile-quantile plot (Supplementary Figures 2, 3 and 4 in Additional file 2). Quantile-quantile plots generated from the analyses of individual datasets showed that there was no within-study systematic error arising from the use of non-matched population controls or genotyping at different facilities (Supplementary Figures 2 and 3 in Additional file 2). Genotype cluster plots were examined for SNPs with P < 10^-5. Manual reclustering was performed for six SNPs with poor genotype cluster plots. SNPs rs4660646 and rs2462692 were omitted from further analysis as they could not be reclustered. SNPs rs4549482, rs1984492, rs1389545 and rs3748648 were not found to be strongly associated with ER-negative breast cancer after reclustering (Table S1 in Additional file 3).

Figure 3 shows a Manhattan plot summarizing the -log-transformed P-values of 285,984 SNPs analyzed in this study. In a combined analysis of individuals of Swedish and Finnish backgrounds, the strongest association with ER-negative breast cancer below the threshold for genome-wide significance was for a locus marked by rs361147 on chromosome 4 (P trend = 3.13 × 10^-13; OR _{per allele} = 0.60) (Table S2 in Additional file 3). This was the only SNP to achieve statistical significance at the genome-wide level (α = 5 × 10^-8). Overall, no significant signal peak was identified in this study (Figures 4, 5, 6, 7, 8).

Nevertheless, we selected five SNPs to be validated in a combined dataset of two independent studies (Table S2 in Additional file 3). SNPs rs7039994 and rs12000794, located 106310 base pairs away from each other on chromosome 9, were found to be in high LD (r2 = 0.797; D' = 0.952). The former was kept and validated in the SEARCH dataset as its associated P-value was smaller and it was in closer proximity to coding regions (downstream of INVS|TEX10). SNP rs3777218 was selected over rs11882068 due to a better regional signal peak. Other SNPs selected for validation included rs361147 as mentioned above, rs6993922, rs4726078 (within transcript of PRKAG2), and rs3777218 (within transcript of RHOBTB3). Of the five SNPs forwarded for validation, rs4726078 could not be designed and was replaced by rs10952315 (r2 = 0.977 in Centre d'Etude du Polymorphisme Humain (CEPH) from Utah (CEU) HapMap samples). None of the SNPs was significantly associated at the 5% level in the second stage. The smallest P-value obtained was for the surrogate rs10952315 (OR 1.02; 95% CI: 0.93 to 1.13).

To analyze our GWAS data in a pathway context we conducted a permutation-based analysis using the KEGG database. Pathways defined by SNPs located within transcript of genes that were found to be significantly associated with ER-negative breast cancer after 1,000 phenotype permutations at a threshold of P_{α = 0.05} < 0.05 (uncorrected) were: pentose and glucuronate interconversions (hsa00040) (P = 0.022), starch and sucrose metabolism (hsa00500) (P = 0.042), and gap junction (hsa04540) (P = 0.037) (Table 2).

In addition, we limited the analysis to pathway definitions involving only known regulatory SNPs [48]. The GWAS SNPs were first mapped to genes, and then subsequently to KEGG pathways based on publicly available gene regulatory data from lymphoblastoid cells [48]. Only genes with regulatory functions significant on a genome-wide significant level were selected, resulting in 1,720 SNPs regulating members of 182 KEGG pathways being used in our analysis. Pathways that were found to be significant by SRT after 1,000 phenotype permutations at a threshold of P_{α = 0.05} < 0.05 were: long-term potentiation (hsa04720), glioma (hsa05214), non-small cell lung cancer (hsa05223), pancreatic cancer (hsa05212), and prostate cancer (hsa5215) (Table 3). The focal adhesion pathway (hsa04510) was found to be marginally significant (P_{α = 0.05} = 0.052). Two pathways each tagged by only a single SNP, glyoxylate and dicarboxylate metabolism (hsa00630) and glycosphingolipid biosynthesis - ganglio series (hsa00604), were removed from the evaluation of the final results.

Regulatory SNPs involved in pathways associated with cancer (hsa052*) appeared to be overrepresented by small P-values (Figure 9). To evaluate if the combined effect of these signals was statistically significant as a whole, we next carried out a global test of significance for all unique SNPs in the cancer pathways. The AML analysis performed using an algorithm developed by Tyrer et al. [41], yielded P-values (α = 0.05) of 0.0028 (crude) and 0.052 (adjusted for population stratification).

Figure 10 shows the results of analyses aimed at assessing the shared polygenic component between ER-positive and ER-negative breast cancer. Estimates of variance explained in datasets indicate how important the polygenic component of ER-positive disease is in explaining the overall occurrence of ER-positive and ER-negative diseases. The proportion of variance explained for all categories of P-value cut-offs, with the exception of P < 0.05 in the Swedish ER-positive target sample, was higher in the ER-positive target datasets than the ER-negative target datasets.

We test for association between polygenic score and disease status (ER-positive vs controls/ER-negative vs controls) in the target data, when seven groups of SNPs with different P-values thresholds in the training sets were considered (Figure 10a, b). Due possibly to limited statistical power (Table S3 in Additional file 3), even at the least stringent P-value threshold (P < 0.50), the ER-positive and ER-negative breast cancer target case-control datasets failed to provide statistically significant evidence of a polygenic component for ER-positive cancer, or evidence of a polygenic component shared between the two cancers, when training was based on the ER-positive training case-control datasets (Figure 10a, b). Nevertheless, when we relaxed the P-value cut-off in the training dataset to 0.5, the Swedish ER-positive breast cancer target dataset showed borderline significance for a shared polygenic component with ER-positive breast cancer, based on the Finnish ER-positive training dataset (Figure 10a, P = 0.066).

In a separate case-only analysis, we performed a significance test for difference in scores between ER-positive and ER-negative breast cancer cases in the target data. Significant results show that ER-positive and ER-negative breast cancers are not identical diseases (genetically at polygenic level) (Figures 10c, d). The difference in scores between ER-positive and ER-negative samples was found to be statistically significant for all categories of P-value cut-offs in the Finnish target case-only samples, with the exception of the most associated SNPs (Figure 10d).

Little is known about the genetic predisposition to estrogen receptor-negative breast cancer. This subtype is characterized by lower age of onset, a more aggressive disease and low or no response to selective estrogen receptor modulators or aromatase inhibitors. We have examined our GWAS data on two different levels: single marker and pathway. We also provided evidence that breast cancer is a heterogeneous disease with a polygenic nature, with significant differences between the polygenic component between ER-positive and ER-negative breast cancers. This emphasizes the importance of looking at ER-negative breast cancer separately as a unique breast cancer phenotype.

Overall, no significant signal peak was identified in this study (Figures 4, 5, 6, 7, 8). Only one SNP (rs361147) was found to achieve genome-wide significance after correction for multiple testing in the single marker analysis. However, the other loci exhibiting strong associations were interesting for reasons of biological significance, and were considered to merit further research. The associated region on 9q31.1 tagged by rs7039994 contains two known genes, TEX10 (testis expressed sequence 10) and INVS (inversin). No function has been ascribed to TEX10. INVS is reported to function as a molecular switch between different Wnt signalling pathways [49] and is also pivotal in the establishment of the left-right axis. The RHOBTB3 gene, harbouring SNP rs3777218, was identified as a putative breast cancer anti-estrogen resistance gene [50]. However, none of these single markers most strongly associated with ER-negative breast cancer could be replicated in a larger, independent sample made up of two independent studies (Table 1)

To maximize the information obtained from the GWAS scan, we conducted a permutation-based pathway analysis using the KEGG database to capture the joint actions of multiple SNPs with modest effects. In the analysis using default SRT pathway definition files comprising within-transcript SNPs, metabolic pathways involving pentose and glucuronate interconversions (hsa00040) (P = 0.022) as well as starch and sucrose metabolism (hsa00500) (P = 0.042) were found to be nominally significantly related to the risk of developing ER-negative breast cancer (Table 2). Estrogen-induced breast cancer cell proliferation is often accompanied by an increase in intracellular metabolic activity, resulting in a higher growth rate. The pentose phosphate pathway, which works in tight conjunction with the pentose and glucuronate interconversions and starch and sucrose metabolism pathways, has recently been suggested to be essential for estrogen-dependent cell proliferation [51]. Several pathways that were found to be marginally significant (P < 0.1) have been suggested to have potential roles in ER-negative breast cancer, namely, the TGF-beta signalling pathway [52], the renin-angiotensin system [53], and the Notch signalling pathway [54]. In addition, the insulin signalling pathway has been the focus of targeted therapy for breast cancer [55], and the purine metabolism pathway is also closely related to the pentose phosphate pathway described earlier.

Nevertheless, there is neither a precise biological definition of a pathway, nor a "standard" method to map SNPs to genes, and then genes to pathways. Pathway analyses of GWAS of common diseases have mostly based SNP-to-gene mappings on the chromosomal position of the SNP, whether it occurs within transcript of a certain gene [19, 56]. However, it may be more meaningful to map SNPs that are associated with the expression of a gene to the gene. To elucidate pathways with more biological relevance, we further conducted pathway analysis based on a subset of SNPs with known regulatory functions. Recent studies have observed that whereas stronger effects overlap between different tissues, weak effects on gene regulation are tissue-specific [57, 58]. Since we utilized data on gene regulation from lymphoblasts, we decided to restrict our dataset to only genes regulated on a genome-wide significant level (LOD >6). This minimized the bias of tissue-specific gene regulation, but at the same time, limited us to only a fraction of all possible SNPs genotyped within our GWAS, thus reducing the power of the analysis.

In spite of the limitations, four of the five significantly associated pathways (P < 0.05) in our analysis were found to be annotated as cancer pathways in KEGG (glioma (hsa05214), non-small cell lung cancer (hsa05223), pancreatic cancer (hsa05212), and prostate cancer (hsa05215) (Table 3)), hence confirming the validity of the choice of this subset of regulatory SNPs in pathway definition. In addition, a global test of the SNPs defining the cancer pathways found the aggregate effect to be approaching statistical significance (P_{α = 0.05} = 0.052). Due to the large number of markers evaluated in a genome-wide scan, signals with small effects and modestly significant P-values are likely to be dismissed after the correction of multiple testing. The implementation of a pathway analysis thus serves as a complementation between a hypothesis-driven (prior knowledge of biological pathways) and a hypothesis-free (genome-wide scan) approach to highlight certain markers, such as those found in the cancer pathways, worthy of further study that would not have been examined otherwise. The lack of a concordance between the results of pathway analyses using two different SNP-to-gene mapping approaches emphasizes the need to put in more consideration in choosing appropriate pathway definitions. An excess of small P-values found for SNPs associated with gene expression involved in cancer-related pathways suggests that the SNP-gene mapping via association with gene expression approach is superior to the SNP-gene mapping by location within a transcript approach, and should be explored in greater detail.

Limitations of this study include an overall lack of statistical power, especially for the single marker analysis, and the existence of further heterogeneity among ER-negative tumours. Although genome-wide pathway-based analysis is an interesting approach, a main limitation is that the associations observed in this study are only nominally significant, and would not be significant after correction for multiple testing. However, as many pathways have SNPs in common with other pathways, the stringent significance thresholds of traditional multiple testing correction methods are potentially over-conservative. There is also indirect evidence that corroborates our pathway findings. Gene expression studies have found pathways related to the renin-angiotensin system and focal adhesion to be significantly associated with prognosis of breast cancer [59]. Others have also reported pathways highlighted in our study, which are involved in pentose and glucuronate interconversions, gap junction, TGF-beta signalling, rennin-angiotensin system, B cell receptor signalling, Fc epsilon RI signalling, VEGF signalling, ErbB signalling, and focal adhesion, to be significantly associated with the breast cancer phenotype [59, 60]. Although replication of the pathway results in independent studies would be needed to confirm the associations, the substantial additional sample collection and genotyping required are beyond the scope of this publication.

Although breast cancer has been classified into ER-positive and ER-negative breast cancers, and these two breast cancer subtypes have been documented to show different gene expression patterns, GWAS scans on breast cancer have always been performed on either overall breast cancer (ER-positive, ER-negative and unknown) or ER-positive breast cancer specific risks. In this study, we found evidence to suggest that ER-negative breast cancers only share a fraction of the polygenic component of the disease with ER-positive breast cancers, implying that ER-negative breast cancer should be examined as a distinct breast cancer phenotype. Although the difference between the polygenic components of ER-positive and ER-negative breast cancers was found only to be significant in the Finnish training samples, we observed similar differences for all seven P-value thresholds in the Swedish training samples. However, due to the smaller number of Swedish ER-negative cases (N = 153, approximately 33% of Finnish ER-negative cases), we had less power to detect significant heterogeneity between the two subtypes in the Swedish target samples.

Given the clinical importance of the ER-negative phenotype and the likelihood that the relative genetic effect sizes are small, greater sample sizes and further studies are required to further the knowledge on ER-negative breast cancers. Identification of factors for a predisposition to ER-negative tumours opens the way for understanding the underlying etiology of the disease, and may ultimately result in improvements in prevention, early detection and specific treatment for this tumour subtype. We used a novel approach to pathway analysis, showing that established cancer pathways could be regulated by common variants associated to ER-negative breast cancer. We also provided molecular genetic evidence which suggests that ER-negative breast cancer is a distinct breast cancer subtype that merits independent analyses. In view of the biological relevance of the pathways identified, a genome-wide pathway approach deserves merit, and has good potential in pointing out directions for future research for ER-negative breast cancers.