A comprehensive analysis of the androgen receptor gene and risk of breast cancer: results from the National Cancer Institute Breast and Prostate Cancer Cohort Consortium (BPC3)

The BPC3 has been described in detail elsewhere [37]. Briefly, the consortium includes five large well-established cohorts assembled in the U.S. and Europe which have both DNA samples and extensive questionnaire information (the American Cancer Society Cancer Prevention Study II [38], the European Prospective Investigation into Cancer and Nutrition [EPIC] cohort [39], the Harvard Nurses' Health Study [NHS] [40] and Women's Health Study [WHS] [41], and the Hawaii-Los Angeles Multiethnic Cohort [MEC] [42]). Most women in these cohorts, with the exception of the MEC, were Caucasians of U.S. and European descent. Breast cancer cases were identified in each cohort by self-report with subsequent confirmation of the diagnosis from medical records or tumour registries and/or from linkage with population-based tumour registries (method of confirmation varied by cohort). Controls were matched to cases by ethnicity and age and, in some cohorts, additional criteria (such as country of residence in EPIC).

Coding regions of AR were sequenced in a panel of 95 advanced breast cancer cases from the MEC (19 of each ethnic group: African-American, Latino, Japanese, Native Hawaiian, and white). Thirty-two single nucleotide polymorphisms (SNPs) with minor allele frequency greater than 5% in any of the five ethnic groups or greater than 1% overall were selected from this resequencing as well as any SNP available in dbSNP to be used to select haplotype-tagging SNPs (htSNPs). These SNPs were genotyped in a reference panel of 349 healthy women from the MEC populations (including 70 whites) at the Broad Institute (Cambridge, MA, USA) using the Sequenom, Inc. (San Diego, CA, USA) and Illumina, Inc. (San Diego, CA, USA) platforms, and six htSNPs were selected to maximise R² _H (a measure of correlation between SNPs genotyped and the haplotypes they describe) among Caucasians, using the method of Stram et al. [43]. Genotyping of the six htSNPs in the breast cancer cases and controls was performed in three laboratories (University of Southern California, Los Angeles, CA, USA; Harvard School of Public Health, Boston, MA, USA; and International Agency for Research on Cancer, Lyon, France) using a fluorescent 5' endonuclease assay and the ABI-PRISM 7900 for sequence detection (Taqman) (Applied Biosystems, Foster City, CA). Initial quality control checks of the SNP assays were performed at the manufacturer (Applied Biosystems); an additional 500 test reactions were run by the BPC3. Assay characteristics for the six htSNPs for AR are available on a public website [44]. Sequence validation for each SNP assay was performed and 100% concordance was observed [45]. To assess inter-laboratory variation, each genotyping centre ran assays on a designated set of 94 samples from the Coriell Biorepository (Camden, NJ, USA), showing completion and concordance rates of greater than 99% [45]. The internal quality of genotype data at each genotyping centre was assessed by typing 5%–10% blinded samples in duplicate or triplicate (depending on the study); the resulting concordance was greater than 99%. The genotyping success rate was 94% or greater for each of the six SNPs at each genotyping centre. No deviation from Hardy-Weinberg equilibrium was observed among the controls in each cohort (at the p < 0.01 level) for any given assay. An association among the exon 1 CAG repeat in AR, family history of breast cancer, and breast cancer risk was previously reported (1990–96 follow-up in the NHS, 617 cases and 960 controls [46]). The exon 1 CAG repeat was genotyped in an additional 376 cases and 540 controls from the NHS as well as 669 cases and 674 controls from the WHS, as previously described [46]. Given that there is no association between the CAG repeat and breast cancer risk, and the interaction between this polymorphism and family history was not observed in this larger combined sample set (1,662 cases and 2,174 controls), we decided not to expend the resources necessary to genotype the repeat in the remaining data sets.

We used conditional multivariate logistic regression to estimate odds ratios (ORs) for disease in subjects with a linear (additive) scoring for zero, one, or two copies of the minor allele of each SNP. We also used conditional logistic regression with additive scoring and the most common haplotype as the referent to estimate haplotype-specific ORs, using an expectation-substitution approach to assign haplotypes based on the unphased genotype data and to account for uncertainty in assignment [47, 48]. Haplotype frequencies and expected subject-specific haplotype indicators were calculated separately for each cohort (as well as by country within EPIC and race in the MEC). To test the global null hypothesis of no association between variation in AR haplotypes and htSNPs and risk of breast cancer (or subtypes defined by receptor status), we used a likelihood ratio test comparing a model with additive effects for each common haplotype (treating the most common haplotype as the referent) to the intercept-only model. We combined rare haplotypes (those with estimated individual frequencies less than 5% in all cohorts) into a single category that comprised less than 1.5% of the controls.

We considered both unadjusted conditional models and conditional models adjusting for known breast cancer risk factors. The covariates included to account for breast cancer risk factors were age at menarche (≤12 years, 13–14 years, 15+ years), menopausal status (pre-, peri-, and postmenopausal), parity (ever/never full-term pregnancy), body mass index (BMI) (in kg/m² as a continuous variable), and use of postmenopausal hormones (ever/never). Other common risk factors, including family history of breast cancer, personal history of benign breast disease, and age at menopause, were unavailable for large numbers of women and therefore were not included in the models. Because the results remained essentially unchanged regardless of the model used, we present results using the unadjusted conditional model. We also evaluated these covariates, restricting analyses of interaction to only those subjects with information available for variables such as family history, with categorical variables divided into quintiles. Interaction effects were evaluated using likelihood ratio testing, comparing models with the main effects of the genetic and risk variable to the model with these main effects and a multiplicative interaction term. Lastly, we tested whether the association between AR and breast cancer differed by menopausal status at diagnosis and tumour receptor (oestrogen receptor [ER] and progesterone receptor [PR]) status.

The exon 1 CAG repeat was analysed as previously reported [44]. Interaction p values between number of repeats and family history were calculated using likelihood ratio tests comparing the model with main effects for carrying at least one long repeat (cutoffs of ≥22, 23, 25, and 27 repeats) and family history with the model containing these main effects, and an additional interaction term, with homozyotes of the ≥22 allele with no family history as the reference.