Subjects were BRCA1 and BRCA2 mutation carriers recruited by 36 study centres in Europe, North America and Australia (Table 1). All carriers participated in clinical or research studies at the host institutions, which have been approved by local ethics committees (list provided in Additional file 1, Table S1). Each committee granted approval for access and use of the medical records for the present analyses.

The large majority of carriers were recruited through cancer genetics clinics offering genetic testing, and enrolled into national or regional studies. Eligibility to participate in CIMBA is restricted to female carriers of pathogenic BRCA1 or BRCA2 mutations who were 18 years old or older at recruitment. Information collected included the year of birth; mutation description, including nucleotide position and base change; age at last follow-up; ages at breast and ovarian cancer diagnoses; and age or date at bilateral prophylactic mastectomy. Information was also available on the country of residence. Related individuals were identified through a unique family identifier. Women were included in the analysis if they carried mutations that were pathogenic according to generally recognized criteria. Only studies that provided tumour pathology information and had genotype information were included in the analysis. However, to maximise the available information, genotyped mutation carriers within those studies missing information on tumour characteristics were included in the analysis and their disease subtype was assumed to be missing at random (see statistical methods for details). Further details about the CIMBA initiative can be found elsewhere 33 .

Tumour pathology data were amalgamated from a range of sources, specifically patient pathology reports, medical records, pathology review data, tumour registry records and results from tissue microarrays. Estrogen and progesterone receptor status was provided as negative or positive, with supplementary immunohistochemistry scoring data and methodology provided when available. Based on definitions supplied, most centres employed a cut off of ≥10% of tumour cells stained positive to define receptor positivity. To ensure consistency across studies, when information on the proportion of cells stained was available, we used the same cut-off to define ER and PR positive tumours. For a small number of cases where composite scoring methods based on the proportion and intensity of staining were available (Allred score, Remmele score and H-score), widely-accepted cut-offs were used (Additional file 1, Table S2). Consistency checks were performed to validate receptor data against supplementary scoring information if provided.

This analysis included genotype data on 12 SNPs that had been previously assessed for their associations with the overall risk of breast cancer for BRCA1 and BRCA2 mutation carriers in CIMBA. Genotyping was performed using either the iPLEX or Taqman platforms and has been described in detail in the previous reports 28 29 30 31 . To ensure genotyping consistency, all genotyping centres were required to adhere to the CIMBA genotyping quality control criteria which are described in detail online 34 . The 12 SNPs genotyped were rs2981582 in FGFR2, rs3803662 in TOX3/TNRC9, rs889312 in MAP3K1, rs3817198 in LSP1, rs13387042 at 2q35, rs13281615 at 8q24, rs4973768 near SLC4A7/NEK10, rs6504950 in the STXBP4/COX11 region, rs2046210 near ESR1 at 6q25.1 and rs11249433 at 1p11.2. A Taqman assay could not be adequately designed for SNP rs999737 in the RAD51L1 region and studies using this platform genotyped the surrogate SNP rs10483813 (pair-wise r² = 1 with rs999737 based on HapMap CEU data). Data for these two SNPs were combined and treated as a single locus in the analysis of associations.

The aim of this study was to evaluate the associations between each genotype and breast cancer subtypes defined by tumour characteristics in BRCA1 and BRCA2 mutation carriers separately. The phenotype of each individual was defined by the age at diagnosis of breast cancer and its subtype or by age at last follow-up. Individuals were censored at the age of the first breast cancer diagnosis, ovarian cancer diagnosis, or bilateral prophylactic mastectomy or the age at last observation. Mutation carriers censored at ovarian cancer diagnosis were considered unaffected.

The analysis of risk modifiers in BRCA1 and BRCA2 mutation carriers is complicated by the fact that mutation carriers are not randomly sampled with respect to their disease status. Many carriers are sampled through families seen in genetic clinics. The first tested individual in a family is usually someone diagnosed with cancer at a relatively young age. Such study designs, therefore, tend to lead to an over-sampling of affected individuals, and standard analytical methods like Cox regression or case-control analysis may lead to biased estimates of the risk ratios 35 . This can be illustrated by considering an individual affected at age t. In a standard analysis of a cohort study or a case-control analysis, the SNP genotype for the individual will be compared with those of all individuals at risk at age t or in a case-control analysis, with controls randomly sampled from all possible at risk individuals. This analysis leads to consistent estimates of the hazard ratio or odds ratio estimates. However, in the present design, mutation carriers are already selected on the basis of disease status (where affected individuals are over-sampled). If standard cohort analysis were applied to these data, it would lead to affected individuals at age t being compared to unaffected carriers selected on the basis of their future disease status. If the genotype is associated with the disease, the risk estimate will be biased to zero because too many affected individuals (in whom the at-risk genotype is overrepresented) are included in the comparison group. Simulation studies have shown that this effect can be quite marked 35 . To address this, a retrospective likelihood approach was previously proposed, which models the observed genotypes conditional on the disease phenotypes 36 . For the current analyses we have extended this method to model the simultaneous effect of each SNP on more than one tumour subtype. We briefly describe this method for the analysis of associations with ER-positive and ER-negative breast cancer, but the same principles apply for the analysis of associations with other tumour characteristics.

We modelled the likelihood of the observed genotypes and tumour subtype conditional on the disease status, that is:

∏ i = 1 n P g i , d i | y ( t i ) = ∏ i n P ( y ( t i ) , d i | g i ) P ( g i ) P y ( t i )

Where y(t_i ) is the disease phenotype for individual i at censoring age t_i (breast cancer at age t_i or unaffected at age t_i ), d_i is the ER status (0 = negative, 1 = positive) and g_i the observed genotype of individual i (g_i = 0, 1 or 2 minor alleles) and n the number of subjects in the analysis. To allow for tumour characteristics we assumed that breast cancer consists of different disease subtypes, such that the total breast cancer incidence at age t_i , λ(t_i ), is the sum of the disease incidence for the subtypes, that is λ(t_i ) = v(t_i ) + μ(t_i ), where v(t_i ) is the incidence for ER-negative disease and μ(t_i ) is the incidence of ER-positive disease. We assumed that the subtype-specific incidences depend on the underlying genotype through a Cox-proportional hazards model: ν ( t i ) = ν 0 ( t i ) exp ( β t z g i ) and μ ( t i ) = μ 0 ( t i ) exp ( γ t z g i ) where v ₀(t_i ) and μ ₀(t_i ) and are the baseline incidences for disease subtypes (ER-negative and ER-positive respectively), z g i is the genotype vector for individual i and β and γ are the subtype specific genotype log-risk ratios (for ER-negative and ER-positive breast cancer respectively). The probabilities of developing ER-positive and ER-negative breast cancer conditional on the underlying genotype were assumed to be independent. We further assumed that, if tumour subtype is unknown, the information is missing at random with respect to genotype. Then for each individual:

P ( y ( t i ) , d i | g i ) = = ν 0 ( t i ) exp ( β t z g i ) O i ( 1 - d i ) × μ 0 ( t i ) exp ( γ t z g i ) O i d i × exp - ∑ u = 0 t i - 1 μ 0 ( u ) exp γ t z g i + ν 0 ( u ) exp β t z g i  if  d i = 0 ,  1    = μ 0 ( t i ) exp ( γ t z g i ) + ν 0 ( t ) exp ( β t z g i ) O i × exp - ∑ u = 0 t i - 1 μ 0 ( u ) exp γ t z g i + ν 0 ( u ) exp β t z g i  if  d i =  unknown

were O_i = 0 if unaffected and O_i = 1 if affected. Thus, the above formulation allows use of all mutation carriers irrespective of whether the tumour subtype is observed or not. The baseline incidences for each disease subtype (v₀ (t_i ) and μ ₀(t_i )) are unknown. However, it is possible to solve for those recursively by constraining the overall breast cancer incidence for mutation carriers λ(t), to agree with external estimates as previously demonstrated 37 38 and by imposing a further constraint on the ratio of the observed ER-positive to ER-negative breast cancers in each age group:

P ( ERpositive at age t) P ( ERnegative at age t) = π + ( t ) π - ( t ) = = ∑ g P ( g ) μ 0 ( t ) exp ( γ t z g ) exp - ∑ u = 0 t - 1 ( μ 0 ( u ) exp ( γ t z g ) + ν 0 ( u ) exp ( β t z g ) ) ∑ g P ( g ) v 0 ( t ) exp ( β t z g ) exp - ∑ u = 0 t - 1 ( μ 0 ( u ) exp ( γ t z g ) + ν 0 ( u ) exp ( β t z g ) )

The likelihood in equation 1 can then be maximised jointly over the log-risk ratios β and γ , genotype frequencies P(g) and the age and subtype-specific frequencies π ₊(t) and π _-(t) This likelihood is based on the assumption that the ascertainment of mutation carriers is dependent on the overall disease phenotype (breast cancer) but not on tumour subtypes. This allows the subtype frequencies π ₊(t) and π_- (t) to be estimated within the dataset. Relaxing this assumption and conditioning also on tumour subtype requires external estimates for the age and subtype-specific frequencies π ₊(t) and π _-(t).

The effect of each SNP was modelled either as a per-allele HR (multiplicative model) or as separate HRs for heterozygotes and homozygotes, and these were estimated on the logarithmic scale. Heterogeneity in the hazard ratios between tumour subtypes was examined by fitting models where v(t_i ) = v ₀(t_i ) exp (β _1g ) and μ(t_i ) = μ ₀(t_i ) exp (β ₁ + β ₂)g) with g = 0,1 and 2 (for 0, 1, 2 copies of the minor allele respectively) and testing for β ₂ = 0. Analyses were carried out with the pedigree-analysis software MENDEL 39 . All analyses were stratified by country of residence and used calendar-year- and cohort-specific cancer incidences for BRCA1 and BRCA2 40 . For this purpose, a stratified version of the retrospective likelihood (equation 1) was derived as described previously 36 . Countries with small numbers of mutation carriers were grouped together. We used a robust variance-estimation approach to allow for the non-independence among related mutation carriers 41 .

Based on our results we computed the predicted absolute risk of developing ER-negative and ER-positive breast cancer for BRCA1 and BRCA2 mutation carriers by the combined 12 SNP profile. For each individual we derived an empirical score, based on the per-allele log-relative hazard estimates for each genotype, which was of the form ∑ j = 1 12 β j g j where β_j is the per-allele log-hazard estimate for locus j and g_j is the genotype at the same locus (taking values 0, 1 and 2). This assumes a multiplicative model for the combined SNP associations. This is a reasonable assumption given that previous analyses found no evidence of departure from the multiplicative model 35 . Scores were calculated for ER-positive and ER-negative disease, separately for BRCA1and BRCA2 mutation carriers. The empirical distribution of the derived score was then used to compute the subtype specific incidence associated with each multilocus genotype as described previously 31 . We reported the absolute risks of developing ER-specific breast cancer at the 5^th, 50^th and 95^th percentiles of the empirical distribution of the SNP profile.

There were significant differences in the HR for ER-positive and ER-negative disease for BRCA1 mutation carriers for two SNPs (Table 2). The FGFR2 SNP rs2981582 exhibited the clearest difference with a strong association for ER-positive disease but not ER-negative disease (per allele HR = 1.35, 95% CI: 1.17 to 1.56, for ER-positive compared with HR = 0.91, 95% CI: 0.85 to 0.98 for ER-negative, P-heterogeneity = 6.5 × 10^-6). The SLC4A7/NEK10 SNP rs4973768 also exhibited a similar pattern (ER-positive: per-allele HR = 1.17, 95% CI: 1.03 to 1.133, compared with ER-negative: per-allele HR = 0.99, 95% CI: 0.93 to 1.06, P-heterogeneity = 0.027). Although there was no significant evidence of differences between the HRs for ER-positive and ER-negative breast cancer, the TOX3/TNRC9 SNP rs3803662 was significantly associated with the risk of ER-positive disease (per-allele HR = 1.25, 95% CI: 1.06 to .46, P-trend = 0.0062) but not with risk for ER-negative breast cancer (per-allele HR = 1.05, 95% CI: 0.97 to 1.13, P-trend = 0.21). LSP1 SNP rs3817198 was associated with the risk of ER-negative breast cancer (per-allele HR = 1.07, 95% CI: 1.00 to 1.07, P-trend = 0.047) but not with risk of ER-positive breast cancer (per-allele HR = 1.07, 95% CI: 0.93 to 1.22, P-trend = 0.33, P-het = 0.98). The 6q25.1 SNP rs2046210 near ESR1 was associated with the risk for both ER-negative (per-allele HR = 1.19, 95% CI: 1.11 to 1.27, P-trend = 2.4 × 10^-7) and ER-positive (per-allele HR = 1.14, 95% CI: 1.01 to 1.30, P-trend = 0.043) breast cancer. There was no significant evidence of association with either ER-negative or ER-positive breast cancer for any of the other SNPs, although the HR estimates tended to be higher for ER-positive breast cancer (for example, SNPs rs13387042 at 2q35 and rs13281615 at 8q24).

Only SNP rs2046210 at 6q25.1 exhibited differential associations between ER-positive and ER-negative breast cancer for BRCA2 mutation carriers (P-heterogeneity = 0.045, Table 3). The per-allele HR for ER-negative disease was estimated to be 1.17 (95% CI: 0.99 to 1.38) whereas the per-allele HR for ER-positive breast cancer was 0.97 (95% CI: 0.89 to 1.05). Although there were no significant differences in the associations between the two types of disease for BRCA2 mutation carriers, the HR estimates for ER-positive disease tended to be larger compared to ER-negative breast cancer. SNPs at/near FGFR2, TOX3/TNRC9, MAP3K1, LSP1, 2q35, SLC4A7/NEK10, 5p12 and 1p11.2 were associated with ER-positive breast cancer for BRCA2 mutation carriers (using either a per-allele or 2 df genotype test). The strongest associations were for the FGFR2 rs2981582 SNP (HR for ER-positive breast cancer = 1.35, 95% CI: 1.23 to 1.48, P-trend = 1.4 × 10^-10) and TOX3/TNRC9 SNP rs3803662 (HR for ER-positive breast cancer = 1.28. 95% CI: 1.16 to 1.41, P-trend = 1.5 × 10^-6). Only SNPs at or near MAP3K1, STXBP4/COX11 and 6q25.1 were associated with the risk of ER-negative breast cancer for BRCA2 mutation carriers.

The general pattern of associations with PR-positive and PR-negative breast cancer for BRCA1 mutation carriers (Additional file 1, Table S3) was similar to that seen for ER status. Significant differences in the associations between PR-positive and PR-negative breast cancer were observed for three SNPs. The minor allele of FGFR2 SNP rs2981582 was associated with a significantly higher risk for PR-positive breast cancer for BRCA1 mutation carriers (per-allele HR for PR-positive = 1.29, 95% CI: 1.10 to 1.51, HR for PR-negative = 0.93, 95% CI: 0.87 to 1.00, P-heterogeneity = 7 × 10^-4). Allele "A" of SNP rs13387042 at 2q35 was associated with a significantly higher risk of PR-positive breast cancer for BRCA1 mutation carriers (HR for PR-positive breast cancer = 1.16, 95% CI: 1.01 to 1.33; HR for PR-negative = 0.97, 95% CI: 0.91 to 1.04, P-heterogeneity = 0.034). Although the RAD51L1 SNP showed no differential associations with ER-status, there was evidence that the minor allele of this SNP was associated with a lower risk of PR-positive breast cancer (HR for PR-positive = 0.79, 95% CI: 0.61 to 0.95; HR for PR-negative = 1.02, 95% CI: 0.94 to 1.11; P-heterogeneity = 0.027). The only SNPs associated with risk for PR-negative breast cancer were SNPs at 6q25.1 (per-allele HR = 1.19, 95% CI: 1.11 to 1.27, P-trend = 3.7 × 10^-7), and in LSP1 (per-allele HR = 1.09, 95% CI: 1.01 to 1.16, P-trend = 0.017), but these were not significantly different from the associations with PR-positive breast cancer for BRCA1 mutation carriers.

Only two SNPs demonstrated significant differences in the associations with PR-positive and PR-negative breast cancer for BRCA2 mutation carriers (Additional file 1, Table S4). SNP rs13387042 at 2q35 was associated with a higher risk of PR-positive breast cancer (per-allele HR for PR-positive = 1.14, 95% CI: 1.03 to 1.25; per-allele HR for PR-negative = 0.92, 95% CI: 0.81 to 1.04, P-heterogeneity = 0.009). SNP rs10941679 at 5p12 was also associated with a higher risk of PR-positive breast cancer for BRCA2 mutation carriers (per-allele HR for PR-positive = 1.15. 95% CI: 1.03 to 1.27, HR for PR-negative = 0.94, 95% CI: 0.81 to 1.09, P-heterogeneity = 0.028). SNPs near or at FGFR2, TOX3/TNRC9, LSP1 were associated with both PR-negative and PR-positive breast cancer, whereas the 1p11.2 SNP was associated with risk for PR-negative breast cancer. MAP3K1 and SLC4A7/NEK10 were associated only with risk of PR-positive breast cancer among BRCA2 mutation carriers.

Using the estimated HRs for ER-positive and ER-negative breast cancer for BRCA1 and BRCA2 mutation carriers, we computed the predicted absolute risk of developing ER-negative and ER-positive breast cancer at various percentiles of the combined SNP distribution. The SNP profile distribution is different for each disease subtype and mutation. We note that SNPs for which the per-allele HR estimates are close to 1.0 contribute little to the predicted ER-specific risks. Figure 1 shows the predicted risks of developing ER-negative and ER-positive breast cancer for BRCA1 and BRCA2 mutation carriers at the 5^th, 50^th and 95^th percentiles of the empirical risk distribution of the combined SNP profile. A BRCA1 mutation carrier at the 5^th percentile of the SNP profile distribution would be at 43% risk of developing ER-negative breast cancer by age 80 compared with 60% for BRCA1 mutation carriers at the 95^th percentile of the risk distribution. The risks of developing ER-positive breast cancer would be 18% and 46% by age 80 at the 5^th and 95^th percentiles of the ER-positive breast cancer risk distribution. BRCA2 mutation carriers at the 5^th percentile of the ER-negative breast cancer risk distribution are predicted to have a 22% risk of developing ER-negative breast cancer by age 80 compared with 39% for the 95^th percentile of the risk distribution. The risks of developing ER-positive breast cancer by age 80 for BRCA2 carriers varied from 33% to 70% at the 5^th and 95^th percentiles of the ER-positive risk distribution respectively.