Participants

Women with germline, deleterious mutations in BRCA1 and BRCA2 were identified in 14 centers in the US, one center in Canada, and one center in Austria - including Baylor University Medical Center - Dallas, Beth Israel in Boston, City of Hope, Creighton University, Dana Farber, Fox Chase Cancer Center, Georgetown University, the Mayo Clinic, Medical University of Vienna, North Shore University Health System in Chicago, University of California, Los Angeles, University of California, Irvine, University of Chicago, University of Pennsylvania, and Women's College Hospital. The majority of subjects were recruited from the Medical University in Vienna, Creighton University in Nebraska, the University of Pennsylvania, and the University of California Irvine (previously at the University of Utah). All centers are part of the Modifiers and Genetics in Cancer consortium. All participants were enrolled under Institutional Review Boards or ethics committee approval at each participating site.

Women were participating in research studies or were either physician or self-referred to risk evaluation clinics for genetic testing, generally because of a strong family history of breast cancer and/or ovarian cancer. The current study is composed of a total of 1,665 adult, female mutation carriers, including 1,122 BRCA1 carriers (433 cases and 689 controls) and 543 BRCA2 carriers (238 cases and 305 controls). The BRCA1 and BRCA2 mutation status of all subjects was confirmed by direct mutation testing, with full informed consent under protocols approved by the human subjects review boards at each institution.

Women were eligible for entry into the study cohort if they tested positive for a known deleterious mutation in BRCA1 or BRCA2. Women with BRCA1 and BRCA2 variants of unknown functional significance were excluded. Women were excluded if they were missing information on year of birth, parity, menopausal status, and oral contraceptive use, or had been diagnosed with cancer more than 3 years prior to study entry. Information about invasive breast cancer, ovarian cancer, prophylactic mastectomy and prophylactic oophorectomy was obtained from medical records, and information on reproductive history and lifestyle habits was obtained by questionnaire.

SNP genotyping

The 47 SNPs had been selected and genotyped in a previous case-control study of African-American women. Briefly, a minimal set of informative SNPs (tagging SNPs) had been chosen across each gene to mark the common genetic variation and to minimize the genotyping costs. Tagging SNP sets were selected using the TagSNPs program [21] from genotype data, downloaded directly from the National Institute of Environmental Health Sciences Environmental Genome Project [22]. For the data available at the time, it was not possible to select tagging SNPs for just a Caucasian population.

Genotyping was performed by the MGB Taqman probe Assay from Applied Biosystems Inc. (Foster City, CA USA) or the MGB Eclipse™ probe assay from Nanogen Inc. (San Diego, CA USA) for all SNPs. Primer and probe sequences are available from the authors on request.

Specifically, for the MGB Taqman probe assays, the reaction mix in a final volume of 5 μl included 10 ng genomic DNA, 4.5 pmol each primer, 1.25 pmol each probe, 1 × PCR reaction buffer (Qiagen, Gaithersburg, MD USA), 2 × Q solution (Qiagen), 500 pmol dNTP, and 0.15 units Qiagen DNA polymerase. PCR cycling included 55 cycles of a two-step PCR (95°C for 15 seconds, and 60°C for 1 minute) after an initial 2 minutes at 95°C. PCR plates were read on an ABI PRISM 7900 HT instrument for genotype assignment (Applied Biosystems Inc.).

Specifically, for the MGB Eclipse™ probe assays, the reaction mix in a final volume of 5 μl included 10 ng genomic DNA, 0.5 pmol limiting primer, 5 to 10 pmol excess primer, 1 pmol each probe, 1 × PCR reaction buffer (Qiagen), 2 × Q solution (Qiagen), 500 pmol dNTP, and 0.15 units Qiagen DNA polymerase. PCR cycling included 55 cycles of a three-step PCR (95°C for 10 seconds, 58°C for 20 seconds and 72°C for 20 seconds) after an initial 2 minutes at 95°C.

After completion of PCR, endpoint dissociation melting curves were generated on the ABI PRISM 7900 HT instrument by monitoring the fluorescence while heating the reactions from 30°C to 80°C at a 10% rate. An EclipseMeltMacro_v2.328 program (Nanogen Inc., San Diego, CA USA) was employed to assign the genotype from the dissociation curve data. Duplicates of 22 DNA samples and water controls were genotyped for quality control. The laboratory technician was blinded as to whether samples were duplicates, cases, or controls. The order of the DNA samples on 384-well plates was randomized in order to ensure balance in study conditions across covariates. Genotyping call rates ranged from 95% to 99% and duplicate concordance rates were higher than 99%.

Determination of linkage disequilibrium blocks

We recalculated the linkage disequilibrium (LD) blocks for this study, primarily because we wanted the LD groups to reflect this predominantly non-Hispanic Caucasian cohort rather than the mixed sample from which the tagging SNPs were originally identified. SNPs were grouped according to their adjacent pairwise LD coefficient (D'). The coefficient was computed between all adjacent marker pairs within each candidate gene. In order to account for within-family correlation, multiple outputation [23] was used to estimate D'. In this case, a single member from each family was randomly sampled to create a single bootstrap sample, from which D' was computed. This process was repeated to obtain 200 bootstrap samples, yielding an empirical distribution of D'. An LD block was defined as a set of contiguous SNPs having D' values exceeding 0.90 between each contiguous pair of SNPs. The boundary of an LD block would be defined by a marker pair with D' ≤ 0.9. The LD blocks for the SNPs within each gene are shown in Additional data file 1.

Statistical analysis

Breast cancer rates were calculated as the observed number of breast cancers per total patient time at risk, and were standardized to the age distribution of the study cohort at the time of interview [24]. Subjects were considered at risk for breast cancer from birth until the first occurrence of breast cancer diagnosis, death, or loss to follow-up. In addition, subjects were censored in the event that they underwent a bilateral prophylactic surgery of the breasts more than 1 year preceding the diagnosis of breast cancer. Bilateral prophylactic surgery of the breasts occurring within 1 year of breast cancer was considered an event in order to avoid potential biases resulting from informative censoring.

Covariates that vary with time (ovarian cancer and prophylactic ovarian surgery) were treated as time dependent in the calculation of rates. A subject who was diagnosed with ovarian cancer therefore contributed time at risk in the non-ovarian cancer group prior to the diagnosis and then time at risk in the ovarian cancer group following the diagnosis. Because subjects were ascertained primarily from high-risk clinics, there was an oversampling of cases. In order to account for potential bias in cumulative risk estimates due to nonrandom sampling from the general population, Kaplan-Meier estimates of the cumulative probability of breast cancer diagnosis were computed using age-specific sampling weights for cases and controls. Sampling weights were obtained from Antoniou and colleagues [1].

Cox proportional hazards regression was used to model the time from birth to diagnosis of breast cancer. In this model, the hazard or instantaneous probability of breast cancer diagnosis is modeled as a function of the predictor covariates. The relative risk or hazard ratio (HR) is then interpreted for each covariate as the proportionate change in the instantaneous probability of diagnosis for two individuals, differing only by a single unit of that covariate. When analyzing LD blocks with multiple SNPs, an additive haplotype effect was assumed where the most common haplotype was used as the referent group for comparisons. When an LD block consisted of a single SNP, however, a general genetic model making no additivity assumption was used. In order to account for phase uncertainty in haplotype analysis, we used a two-step approximation to the semiparametric maximum likelihood estimator of Lin and Zeng [25]. Using this method, the expectation-maximization algorithm was used to compute posterior estimates of the probability of all potential haplotypes for a subject given their known genotype, and these probabilities were used to weight the individual's contribution to the partial likelihood. A similar approach has previously been applied to logistic regression models for analyzing case-control data and was shown to provide robust inference for relatively common haplotypes with little phase ambiguity [26]. In order to account for hierarchical clustering at the individual level (multiple records per individual were analyzed according to the number of potential diplotypes consistent with the individual's genotype) and at the family level (matched controls were often selected from the family of a case), the sandwich estimator of Lin and Wei [27] was used in combination with multiple outputation [23] to obtain robust variance estimates of haplotype associations.

All estimates were adjusted for birth cohort (to account for frequency matching of cases and controls), race/ethnicity, parity, and region of center (North American (US) vs. European). Ashkenazi Jewish individuals were considered a separate ethnicity because the carriers only had one of three founder mutations. Parity, prophylactic oophorectomy, and ovarian cancer status were treated as time-dependent covariates in the analysis, with these covariates updated at the time of childbirth. Beyond adjustment for birth cohort, no additional weighting for selection was employed. For LD blocks exhibiting significant associations with the time to breast cancer diagnosis, secondary analyses of individual SNPs making up the LD block were conducted. For all analyses, the proportional hazards assumption was examined by considering multiplicative interactions between each haplotype (or SNP) of interest and (log-transformed) time. No significant departures from the proportional hazards assumption were observed.

In total, the current analysis involves testing of 48 LD blocks, which is likely to result in an inflation of the family-wise type I error rate for the study if unadjusted critical values are used for assessing LD block significance. Noting that this analysis represents a first-stage in identifying variants in the IGF pathway that are associated with time to breast cancer diagnosis, we sought to control the family-wise type error rate at 15% in order to minimize the type II error rate, limiting the possibility of ruling out potentially important LD blocks from future investigation. Simulation was used to estimate the family-wise type I error rate, assuming a correlation of 0.75 across tests was assumed. Based upon 100,000 simulations it was estimated that an adjusted P value of 0.016 on any individual LD block test would result in a family-wise type I error rate of 15% for the study. An adjusted P < 0.016 was interpreted as a significant association.

IGF binding proteins IGFBP1, IGFBP2, and IGFBP5

Figure 2 presents the estimated HR for time to diagnosis by LD block within each of the IGFBPs, and the BRCA status after adjustment for covariates (described in Materials and methods). For BRCA1 carriers, no significant associations were observed for the three IGF binding genes. Among BRCA2 carriers, one LD block in IGFBP2 showed significance in the hazard for diagnosis. For IGFBP2 LD block 2 (defined by a single SNP rs9341134), women with at least one variant allele were estimated to experience a 41% lower risk of diagnosis when compared with women with no variant alleles (HR = 0.59; 95% CI = 0.39, 0.90; unadjusted global P = 0.0145). For IGFBP5 LD block 2 (defined by a single SNP rs2241193), women with at least one variant allele were estimated to experience a 29% lower risk of diagnosis when compared with women with no variant alleles (HR = 0.71; 95% CI = 0.53, 0.96; unadjusted global P = 0.0242).

Insulin-like growth factor receptor substrate 1 and insulin-like growth factor 1

Estimated HRs for the haplotypes of IRS1 are shown in Figure 3a. Among BRCA1 carriers, the global LD block test for IRS1 was not significant (unadjusted global P = 0.0551). Relative to the referent haplotype, however, individuals with haplotypes homozygous for the common variant (excluding haplotypes 001 and 100) were estimated to have a 43% (CI = 1.06, 1.95; P = 0.02) higher risk of breast cancer diagnosis.

For IGF1, no significant associations were found for either BRCA1 or BRCA2 carriers (Figure 3b).

Insulin-like growth factor-1 receptor

Figure 4 shows HR estimates for the 12 LD blocks genotyped in IGF1R. For BRCA1 carriers, significant associations were found between LD block 2 (SNP rs2715415) and LD block 11 and the risk of breast cancer diagnosis (unadjusted global P values corresponding to a test of homogeneity of risk within the LD blocks were 0.011 for LD block 2 and 0.012 for LD block 11). While qualitatively consistent associations were also observed among BRCA2 carriers, they were not significant. After investigation in BRCA1 carriers of the individual SNPs within LD block 11 (Table 2), the only SNP that was significantly associated with risk was rs8038415 - in which individuals homozygous for the variant allele were estimated to experience a 40% higher risk of breast cancer diagnosis (unadjusted P = 0.014, with P for trend = 0.015).