To identify SNPs associated with prognosis, risk and tumour characteristics, we selected 1,001 women from the Prospective study of Outcomes in Sporadic versus Hereditary breast cancer (POSH) cohort for genotyping of 30 genes with prior evidence for biological relevance [see Additional data file 1]. Full details of the POSH study can be obtained from the published protocol 20. The subset of 1,001 POSH patients are all incident cases who were aged 40 years or younger at diagnosis of invasive carcinoma between January 2000 and December 2007 and have no family history of breast cancer or ovarian cancer (all recruits had submitted family history details, collected by questionnaire). These cases are therefore least likely to carry a high risk susceptibility allele such as BRCA1 or BRCA2. To date, comprehensive BRCA1 and BRCA2 mutation screening (using a combination of conformation sensitive capillary electrophoresis and multiplex ligation-dependent probe amplification) has been carried out in 120 patients from the POSH cohort, 39 of which reported no family history of breast or ovarian cancer. For those without family history, one in 39 (2.5%) had a pathogenic BRCA1 mutation and no pathogenic BRCA2 mutations were found (unpublished data).

This subset of POSH cases all presented symptomatically, which means the tumour biology and clinical features have not been modified by earlier detection in a screening programme. The effect of tumour biology on prognosis is reflected in the known clinical prognostic factors (tumour size and grade, ER status and pathological nodal [pN] status). These parameters are used to make decisions about adjuvant therapy 21 and have been recorded for all POSH cases. On the basis of these factors, the subset of POSH cases were stratified into subgroups and compared to identify common variants that affect tumour biology.

A total of 206 tag single nucleotide polymorphisms (SNPs) across 30 genes were selected for genotyping, ensuring coverage of each gene and its promoter region with r² ≥ 0.9 (r² between adjacent SNPs in candidate genes) using data from the HapMap project 22. Genotyping was carried out by Sequenom (San Diego, CA, USA) using the iPLEX service, which is based on mass spectrometry. A little more than one-third of these SNPs (35%, 73) were removed by a screening procedure that rejected SNPs with ≥10% missing genotypes (60) or allele frequencies less than 5% (12). Typically, deviations from Hardy-Weinberg equilibrium in controls would be used to exclude SNPs. However, because only cases were genotyped, Hardy-Weinberg deviations may reflect association with breast cancer, rather than poor genotyping. We therefore verified genotyping calls by examining the Sequenom cluster plots for SNPs with significant deviations from Hardy-Weinberg equilibrium (χ² ≥ 10). On this basis, one SNP was removed leaving 133 (64%) for analysis, which is consistent with the genotyping tier [see Additional data file 1]. A high proportion of missing genotypes per individual DNA sample may suggest relatively poor DNA quality. Most samples had very low rates of missing data, and we removed 58 individuals with ≥ 10% missing genotypes, leaving 943 for further analysis.

Because population stratification may cause false positives in association studies, and the majority of individuals in the POSH cohort have Western European ancestries, we removed individuals with apparently different ancestries. We used the PLINK program 23 to perform multidimensional scaling on genome-wide average identities by state. To avoid confounding of the multidimensional scaling by including any nonindependent SNPs due to extended linkage disequilibrium, we thinned the data to 82 autosomal SNPs with maximum pair-wise r² ≤ 0.2. These data were merged with data from the founders or unrelated individuals in the HapMap sample 22: 60 Western European (CEU), 60 Nigerian (YRI), 45 Han Chinese (CHB) and 45 Japanese (JPT). Before merging the genotypic data from POSH and HapMap, we ensured that all genotypes were configured to the positive strand of the reference sequence (National Cancer for Biotechnology Information [NCBI] build 36.1). Although more than 10,000 autosomal SNPs are recommended for this type of analysis 23, plotting the first two components from the multidimensional scaling analysis of 82 SNPs, which represent geographic and genetic variation, clearly identifies three distinct clusters that correspond to African, Asian, and Western European ancestries (Figure 1). Given this, we were able to exclude 44 samples (35 African and 9 Asian) that did not align with the Western European cluster, leaving 899 for analysis. The genotyping call rate for the 133 SNPs and 899 individuals passing our quality control procedure is 99%.

To identify SNPs influencing prognosis and because the median follow-up time for the 899 patients passing quality control is only 2.4 years (Table 1), we calculated distant disease-free survival (DDFS) times that we define as the time between diagnosis and the first distant metastasis. This is a surrogate for overall survival because the majority of patients with distant metastases will not survive the disease. In seven cases, the date of death was used because the date of distant metastasis was not available. In one of these cases the cause of death was known to be unrelated to breast cancer. Univariate analyses of DDFS were performed using the log rank test to compare Kaplan-Meier DDFS curves for the three genotypes of each SNP.

Data on ER status (positive/negative), histological grade (1 to 3), invasive tumour diameter (not including ductal carcinom in situ component 0.5 to 180 mm), and axillary nodal status (1/0) were available for 828 (92.1%) of the cases passing the quality control procedure. We were therefore able to determine whether SNPs with significant impacts on survival were independent of the known prognostic factors by treating them as covariates in Cox's proportional hazards model. For each SNP genotype, hazard ratios (HRs) and 95% confidence intervals (CIs) were determined using the common allele homozygotes as the reference group. For SNPs with 10 or fewer minor allele homozygotes, the minor allele homozygotes were combined with heterozygotes. Dominant and recessive disease models were also tested, and the assumption of proportional hazards was tested using log-log plots.

To search for SNPs affecting tumour biology, we stratified the 899 cases into major subgroups on the basis of ER status (positive or negative) and grade (1 or 3). Because tumour grade is necessarily a subjective classification and some grade 2 tumours have gene expression signatures similar to those of grade 1 or grade 3 tumours 24, many grade 2 tumours might be reclassified as either grade 1 or 3. We therefore elected to compare the extremes of grade 1 versus grade 3. The 133 SNPs were individually tested for association using Cochran-Armitage trend test and two-by-three tables of disease-by-genotype.

Although our study was not designed primarily to detect SNPs that increase breast cancer risk, we explored risk associations by contrasting POSH cases with genotyping data in controls that were obtained from the Wellcome Trust Case Control Consortium (WTCCC) 25. Multidimensional scaling analysis of the WTCCC and HapMap data involving 60,364 SNPs was used to infer Western European ancestry for 2,980 of the 3,000 WTCCC controls from the National Blood Service and the 1958 British Birth Cohorts. The 28 SNPs that were typed in both POSH and WTCCC datasets [see Additional data file 1] were individually tested for association using Cochran-Armitage trend test and two-by-three tables of disease-by-genotype to contrast POSH cases with WTCCC controls.

To determine accurate levels of significance, it is important to correct for the number of tests performed. However, the most effective approach to correct for multiple tests has yet to be defined. The HapMap Consortium has proposed that adjustments should be made for the total number of independent tests. For tests of prognosis and tumour biology, 133 SNPs were used, and 82 of these in 30 genes were determined to be independent by the multidimensional scaling procedure with maximum pair-wise r² values of less than 0.2. For these tests, the threshold for significance is a P value of 0.0006 (0.05/82). Only 28 SNPs in 17 genes were tested when searching for SNPs associated with risk for breast cancer. A conservative Bonferroni correction requires P values less than 0.003 (0.05/17) to be considered significant for these tests.

We used SPSS 14.0 (SPSS Inc., San Francisco, CA, USA) to perform Kaplan Meier and Cox regressions and PLINK 23 to carry out Cochran-Armitage trend tests and genotypic tests.

Ethics approval for this research was granted by the South West Multi-Centre Research Ethics Committee (MREC/00/6/69). All patients recruited to POSH gave fully informed written consent.

The phenotypic characteristics of this subset of the POSH cohort are based on clinical and pathological assessment performed at the time of diagnosis. The distribution of grade and ER status in 873 cases that passed quality control and have this phenotypic information available are shown in Table 1. Compared with women aged 40 years and older at diagnosis, the subset of POSH study women have nearly 10% more ER-negative tumours (30.9% in POSH versus 21.4% in women aged ≥ 40 years) 26.

Testing the 28 SNPs that were also typed in the WTCCC confirmed previous associations between risk for breast cancer and the same SNPs in TOX3 and ESR1 3527 (Table 2). We also identified an association with rs12693932 in the CASP8 gene that has previously been implicated 1. However, we report an increased risk with rs12693932, whereas the previous study found a protective effect with rs1045485, which failed to pass our quality control procedure. Unfortunately, none of these SNPs remain significant after conservative correction for multiple tests. The SNPs in FGFR2 from our study were not typed as part of the WTCCC, and so these could not be included in this analysis.

To determine whether any of the SNPs affect the duration of DDFS, Kaplan Meier DDFS analysis was conducted in the 133 SNPs that passed the quality control procedure. Before conservative correction for multiple tests, seven of these SNPs, representing six genes, were found to have significant associations with DDFS (Figure 2). Carriers of the minor alleles for five of these SNPs had poorer DDFS, and there was improved DDFS for carriers of two of the SNPs. Data on ER status, histological grade, tumour diameter, and pN stage were available for 828 (92.1%) of the cases passing the quality control procedure. We therefore used Cox regression analysis to determine whether SNPs associated with DDFS were independent of these known prognostic factors by treating them as covariates in the hazard models. After adjustment for ER status, histological grade, tumour diameter and pN stage, three of the seven SNPs identified by the Kaplan Meier analysis remained significant. One additional SNP in TOX3 was found to be associated with DDFS by Cox regression (Table 3) but did not reach significance in the Kaplan-Meier analysis. However, these SNPs do not remain significant after conservative correction for the 82 independent tests.

We found a protective effect against the development of distant metastases among cases with the rare C allele of rs1943779 in the MMP7 (matrix metalloproteinase 7) gene (rare allele homozygotes HR = 0.217, 95% CI = 0.069 to 0.689; heterozygote HR = 0.702, 95% CI = 0.490 to 1.005) and significant heterogeneity among genotypes (P = 0.009).

For the TOX3 gene, which has been shown to be associated with an increased risk for breast cancer in published case control studies, the rare allele homozygotes of rs1420542 (TT; n = 41) and heterozygotes (TA; n = 223) showed a 2.1-fold and 1.6-fold risk for distant metastases, respectively, compared with common allele homozygotes. Heterogeneity among genotypes was significant (P = 0.013), but this SNP did not achieve statistical significance in the Kaplan-Meier analysis (P = 0.077).

The SNP rs889312 in MAP3K1 (which encodes MEK kinase) showed a significant improvement in DDFS for heterozygotes of this SNP (HR = 0.624, 95% CI = 0.435 to 0.897) and appears to demonstrate a dominant effect, but DDFS in rare homozygotes (HR = 0.627, 95% CI 0.324 to 1.215) was not significant, despite significant heterogeneity between genotypes (P = 0.027).

For rs13281615 at chromosome location 8q24 (no known genes), the evidence for effects that are independent of the known prognostic factors is much weaker because significant results were only obtained for the recessive disease model (HR = 1.569, 95% CI = 1.060 to 2.323).

For the four remaining SNPs that did not show significant associations after correcting for ER, grade, tumour size and pN stage, further Cox regressions were performed that treated these factors separately to determine which of them are correlated with the SNP's effect on DDFS. The survival effects of SNPs in ESR1 and LSP1 were no longer significant after grade and pN stage were accounted for, whereas the survival effect associated with the DAPK1 SNP was attributable to ER status, using Cox regression analysis.

To confirm the relationships between SNP genotypes and prognostic tumour phenotypes and to search for additional SNPs, we stratified the 899 cases into major subgroups on the basis of ER status (positive or negative) and grade (1 or 3) and tested the 133 SNPs for association using Cochran-Armitage trend test and genotypic tests. Nine SNPs have nominally significant associations, five with ER status and four with grade (Table 4). Of these, three remain significant after Bonferroni correction for the 82 independent SNPs tested (rs2981582 and rs1219648 in FGFR2, and rs1801516 in ATM, with P < 0.0006) and four were in genes shown to have significant effects on the risk for developing distant metastasis (rs1420546 in TOX3, rs2228480 and rs3798577 in ESR1, and rs661348 in LSP1). Comparison of genotypes in patients developing ER-negative tumours with those developing ER-positive tumours demonstrated that SNPs in FGFR2 and TOX3 are associated with the development of ER-positive breast cancers. In contrast, SNPs in ATM were strongly associated with ER-negative breast cancer (Table 4). Comparison of genotypic extremes of grade (grade 1 versus grade 3) demonstrated a significant difference in genotype frequencies for ESR1, IGF1 and LSP1 (Table 4).