Cases were selected from the Studies of Epidemiology and Risk factors in Cancer Heredity (SEARCH) breast cancer study, an ongoing population study of women diagnosed with breast cancer in the region of England included in the Eastern Cancer Registration and Information Centre (ECRIC) (formerly the East Anglian Cancer Registry). Eligible participants include women diagnosed with invasive breast cancer who were either under 70 years of age since the beginning of the study on 1 July 1996 (incident cases) or age 55 or younger since 1 January 1991 and who were alive at the start of the study (prevalent cases). Due to boundary changes, some prevalent cases diagnosed before 1995 were identified by the North Thames Cancer Registry.

Of those eligible, 67% returned a comprehensive epidemiological questionnaire and 64% returned a blood sample for genotyping. All participants in the study provided informed consent, and the study was approved by the Eastern Multicentre Research Ethics Committee. DNA is available from 4,470 cases for genotyping; 27% of these participants are prevalent cases.

The samples have been split into two sets in order to save DNA and reduce genotyping costs. Cases with high genomic yield were randomly selected from the first 3,500 recruited to comprise set 1 (n = 2,270), with set 2 comprising the remainder of these plus the next 970 incident cases recruited (n = 2,200). DNA yield was not associated with genotype for those cases included in set 1 or set 2. SNPs showing a positive association with survival after a diagnosis of breast cancer (P trend < 0.05) were genotyped in set 2. Data from both sets were then combined (n = 4,470) to jointly analyze the SNPs with positive associations. This joint analysis approach results in increased power to detect genetic association despite more stringent significance levels with Bonferroni correction 17.

As the prevalent cases were the first recruited, the proportion of prevalent cases was somewhat higher in set 1 than set 2 (33% versus 20%). In total, 1,370 prevalent cases were included in both sets; median time from diagnosis to blood draw was 3.4 years (range: 0.8 to 9.34 years). Median age at diagnosis was similar in the two sets (50 and 53 years old, respectively). Median time from diagnosis to blood draw was slightly longer for set 2 (18 months) than for set 1 (9 months), but the number of deaths in each set was similar (359 in set 1 and 278 in set 2). There was no significant difference in the morphology, histopathological grade, or TNM (tumour, node, metastasis) stage 18 of the cases by set or by prevalent/incident status.

The ECRIC and the North Thames Cancer Registry have active follow-up at years 3 and 5 after diagnosis and then at 5-year intervals. Follow-up information and all-cause mortality are obtained by searching hospital information systems for recent visits. If a patient has not had a recent visit, the patient's general practitioner is contacted to obtain the vital status. Death certificate flagging through the Office of National Statistics also provides the registries with notification of deaths. The lag times with this process are a few weeks for cancer deaths and 2 months to a year for non-cancer deaths. Cause-specific mortality was obtained from part I of the death certificate.

Thirteen genes involved in the G₁ phase of cell cycle regulation were selected as candidate genes for breast cancer survival. A comprehensive SNP tagging approach was used in which tagging SNPs (tagSNPs) were chosen to capture all known common genetic variation in each gene with an estimated correlation coefficient (r²) of greater than 0.8. In some cases, SNPs that were poorly correlated with other single SNPs could be efficiently tagged with a haplotype defined by multiple SNPs. Correlation between these SNPs and the haplotype of tagSNPs (r²_s) was aimed to be greater than 0.8. TagSNPs were identified with the program Tagger 19. Data from the International HapMap Project 20 and resequencing data from the National Institute of Environmental Health Sciences (NIEHS) Environmental Genome Project 21 were used to select tagSNPs. In total, 85 tagSNPs were chosen.

Genotyping was carried out using the TaqMan^® platform (Applied Biosystems, Foster City, CA, USA) according to the manufacturer's instructions. Primers and FAM- and VIC-labeled probes were supplied directly by Applied Biosystems as Assays-by-Design™. All assays were carried out in 384-well plates. Each plate included negative controls (with no DNA) and positive controls duplicated on a separate quality control plate. Plates were read on the ABI Prism 7900 using Sequence Detection Software (Applied Biosystems). Failed genotypes were not repeated. Assays in which the genotypes of duplicate samples did not show greater than 95% concordance were discarded and replaced with alternative assays with the same tagging properties. Call rates for each assay were over 95%.

Cox regression analysis was performed to determine the effect of each tagSNP on survival. The proportional hazards assumption was evaluated by visual inspection of log-log plots as well as tested analytically using Schoenfeld residuals. TagSNPs significantly associated with survival were re-evaluated in a model adjusted for known breast cancer prognostic factors, which included age at diagnosis (<40, 40 to 49, 50 to 59, or >60 years), clinical stage (TNM stage 1, 2, 3, or 4), histopathological grade (well differentiated, moderately differentiated, or poorly differentiated), estrogen receptor (ER) status, and treatment.

Time at risk began on the date of blood sample receipt and ended on the date of death from any cause or, if death did not occur, on 30 November 2006. This allows for the difference in ascertainment of incident and prevalent cases and provides an unbiased estimate of the relative hazard provided that the proportional hazards assumption is correct. Follow-up was censored at 10 years after diagnosis as follow-up became less reliable for each individual after 10 years. A hazard ratio (HR) was estimated for heterozygous and rare homozygous genotypes relative to the common genotype. Primary tests used were a likelihood ratio test (2 degrees of freedom) for heterogeneity of risk among the three genotypes (common homozygote, heterozygote, and rare homozygote) and a trend test (1 degree of freedom) based on the number of rare alleles carried. All analyses were performed with Intercooled Stata, version 8.2 (StataCorp LP, College Station, TX, USA).

The characteristics of the SEARCH breast cancer study participants for whom genotyping and vital status data were available are described in Table 1. More than 99% of the cases were Caucasian. There were 3,263 (73%) cases enrolled as incident cases and 1,207 (27%) as prevalent cases. No significant difference in survival hazard was found between the two groups (P = 0.19). During the 25,049 person-years at risk, there were 637 deaths before 10 years of follow-up. Five hundred forty-two of these deaths were coded as breast cancer-specific.

CCND3 rs2479717 is an a > t intronic alteration approximately 50 bases downstream from exon 3. It lies in a large linkage disequilibrium (LD) block on chromosome 6, which contains seven genes: CCND3, BYSL, TRFP, USP49, C6ofr49, FRS3, and PGC (Figure 2). To further evaluate the effect of these genes on breast cancer survival, microarray gene expression data for seven breast cancer cohorts were taken from existing literature and public databases (Gene Expression Omnibus and ArrayExpress) 22232425262728. The publicly available data comprise raw expression data that have been 'normalized'. This usually involves background correction, quantile normalization, and log transformation. The retrieved datasets were further normalized, if necessary, by transforming them onto a common log2 scale and shifting the median of each array to zero 29.

Study characteristics are included in Table 3. In total, the studies included tumour samples from 1,241 individuals. All of the studies contained clinical outcomes; five of the studies had information on all-cause mortality, and two of the studies had disease recurrence as an outcome. Most of the studies had a mix of untreated and treated individuals, with treatment including both chemotherapy and/or hormonal therapy. Median age at diagnosis ranged from less than 53 years to 65 years. The majority of tumours were ER-positive and medium grade (Supplementary Table 5 found in Additional File 3). Combined, the datasets provided 9,048 years at risk (median of 7.4 years at risk) and a total of 384 events.

To test whether a gene's expression is associated with clinical outcome, we used Cox proportional hazards regression models. In this context, the HR refers to the proportional increase in hazard risk per unit increase on a log2 scale of expression level of the transcript. To attempt to control for study heterogeneity, expression data for each gene were analyzed two ways: a Cox survival model stratified by study (fixed effects) and a random-effects meta-analysis. Elevated expression levels of the BYSL transcript were significantly associated with breast cancer survival in a random-effects model (HR = 1.84, 95% CI: 1.10 to 3.08; P = 0.02) (Figure 3a), but the association did not reach significance in a fixed-effects model (HR = 1.17, 95% CI: 0.99 to 1.38; P = 0.08). Elevated expression levels of the C6orf49 transcript were significantly associated with breast cancer survival in a fixed-effects model (HR = 1.60, 95% CI: 1.18 to 2.16; P = 0.002) and in a random-effects model (HR = 1.84, 95% CI: 1.11 to 3.05; P = 0.02) (Figure 3b). Expression levels of CCND3, TRFP, USP49, FRS3, or PGC transcripts were not significantly associated with breast cancer survival (Supplementary Table 6 found in Additional File 3).