Breast cancer subtypes and previously established genetic risk factors: A Bayesian approach

Abstract

Background

Gene expression analyses indicate that breast cancer is a heterogeneous disease with at least 5 immunohistologic subtypes. Despite growing evidence that these subtypes are etiologically and prognostically distinct, few studies have investigated whether they have divergent genetic risk factors. To help fill in this gap in our understanding, we examined associations between breast cancer subtypes and previously established susceptibility loci among white and African-American women in the Carolina Breast Cancer Study.

Methods

We used Bayesian polytomous logistic regression to estimate odds ratios (ORs) and 95% posterior intervals (PIs) for the association between each of 78 single nucleotide polymorphisms (SNPs) and 5 breast cancer subtypes. Subtypes were defined using 5 immunohistochemical markers: estrogen receptors (ER), progesterone receptors (PR), human epidermal growth factor receptors 1 and 2 (HER1/2) and cytokeratin (CK) 5/6.

Results

Several SNPs in TNRC9/TOX3 were associated with luminal A (ER/PR+, HER2−) or basal-like breast cancer (ER−, PR−, HER2−, HER1 or CK 5/6+), and one SNP (rs3104746) was associated with both. SNPs in FGFR2 were associated with luminal A, luminal B (ER/PR+, HER2+), or HER2+/ER− disease, but none were associated with basal-like disease. We also observed subtype differences in the effects of SNPs in 2q35, 4p, TLR1, MAP3K1, ESR1, CDKN2A/B, ANKRD16, and ZM1Z1.

Conclusion and Impact

We found evidence that genetic risk factors for breast cancer vary by subtype and further clarified the role of several key susceptibility genes.

INTRODUCTION

Researchers have long recognized that breast cancer is a heterogeneous disease with variable prognoses and clinical characteristics. Further, epidemiologic investigations have discovered evidence of divergent etiologic processes (1, 2), with some key differences in risk factors across disease subgroups (3–6). While these findings have led to advancements in our understanding of the disease, inconsistent subtype definitions and imprecise estimates have hampered progress. Attempts to identify subtype-specific genetic risk factors have been especially discouraging, with little consistency across study populations (7–10).

Most investigators rely on immunohistochemical (IHC) analysis of estrogen receptors (ER), progesterone receptors (PR) and human epidermal growth factor receptors-2 (HER2) to define breast cancer subtypes. These markers are included in most routine clinical evaluations of breast tumors, as they are predictive of response to targeted therapies such as tamoxifen and trastuzumab. Based on concerns that these three markers did not adequately capture disease heterogeneity, researchers turned to gene expression analysis for more in-depth assessments. In one of the first large-scale gene expression analyses of breast tissue, Perou et al. (11) observed that tumors with similar expression patterns also had similar IHC subtypes. The only major exception was triple-negative tumors (i.e. ER−, PR− and HER2−), which clustered into two separate groups with different cytokeratin (CK) 5/6 and human epidermal growth factor receptor-1 (HER1) expression patterns.

This research led to a new classification system with 5 IHC markers serving as adequate, inexpensive surrogates for more complex gene expression profiles (12–14). Because the CK 5/6 protein is usually present in basal epithelial cells but not in more differentiated luminal epithelial cells, the subtypes were designated as follows: Luminal A (ER or PR+, HER2−), Luminal B (ER/PR+, HER2+), HER2+/ER−, and Basal-like (ER, PR−, HER2−, HER1+ or CK 5/6+).

This subtype classification system has led to insights into racial disparities and furthered understanding of etiologic and prognostic differences between disease subgroups. Luminal A is the most common subtype, but subtype prevalence varies by age and race (14–17). Notably, basal-like and other triple-negative tumors are more common in women of African descent (3, 14, 16–20). For women diagnosed before 2000, those with HER2+/ER− and basal-like breast cancers had the poorest prognoses (14, 15, 17, 21). The development and FDA approval of trastuzumab has since improved survival rates for women with HER2+ disease, but women with basal-like or other types of triple-negative disease still experience high short-term mortality (22–24). This phenomenon likely explains some of the racial disparity in mortality between US African-Americans and whites (30.5 versus 21.6 deaths per 100,000 women with breast cancer per year, 2009) (25).

In previous studies of subtype-specific determinants, luminal A breast cancer was associated with most established breast cancer risk factors, including family history of breast cancer, reproductive factors, decreased physical activity, increased alcohol consumption and high breast density (3, 4, 26–41). In case-only risk ratio analyses, women with a family history of the disease, a younger age at diagnosis, or an earlier age at menarche were more likely to have triple-negative than luminal A tumors. Triple-negative tumors were also relatively more common in African-Americans and in women who had more children but did not breastfeed. Risk factors for luminal B and HER2+/ER− subtypes are less established, but evidence suggests that African-American race, family history of breast cancer, lack of breastfeeding and high alcohol consumption are risk factors for HER2+/ER− disease. Luminal B breast cancers are more common in younger women, but otherwise have similar risk profiles to luminal A tumors.

The ground-breaking discovery of the rare but highly penetrant BRCA1 and BRCA2 genes (42, 43) opened a floodgate of linkage analyses, candidate gene studies, and later, genome-wide association studies (GWAS). To date, 74 single nucleotide polymorphisms (SNPs) have met the criteria for genome-wide “discovery” (44) and variants on six candidate genes (ATM, CASP8, CHEK2, CTLA4, NBN, and TP53) have “cumulative evidence of an association” with breast cancer (45). Of the aforementioned variants, only BRCA1 has been consistently linked to a particular subtype, with numerous studies observing associations between BRCA1 mutations and triple-negative disease (12, 46–48) or increased basal marker expression (12, 49).

In an attempt to elucidate subtype-specific genetic risk factors for breast cancer and further our understanding of disease etiology, we estimated associations between breast cancer subtypes and previously identified candidate gene and GWAS hits using women from the Carolina Breast Cancer Study (CBCS). This population is well-suited to answer this research question, as it is one of the few studies to have both a large proportion of African-American participants and information on basal IHC markers.

This evaluation is further enhanced by the use of Bayesian statistical methods. Specifically, based on evidence that most genetic variants have either null or weak associations with breast cancer (44, 45, 50), we improved the precision and overall accuracy of our effect estimates by shrinking them toward an informative, null-centered prior.

MATERIALS AND METHODS

Study population

The CBCS is a population-based, case-control study of invasive and in situ breast cancer. The study was conducted in 24 North Carolina counties between 1993 and 2001. To be eligible, cases had to be between 20 and 74 years of age at the time of their diagnosis, with no prior history of breast cancer. Women with in situ breast cancer were eligible if they were diagnosed with ductal carcinoma in situ with microinvasion to a depth of 2 mm or lobular carcinoma in situ between 1996 and 2001.

Both invasive and in situ cases were identified using the North Carolina Central Cancer Registry’s rapid case ascertainment program (51). A main objective of the CBCS was to collect information on traditionally under-researched populations. Therefore, cases were randomly sampled at disproportionate rates based on race and age. This sampling strategy ensured approximately equal representation of African-American and non-African-American women, as well as younger (age<50) and older women (age 50+).

Throughout the study period, controls aged 20–64 years were selected from North Carolina Department of Motor Vehicles records and were probability matched to cases based on race and age group (52). Controls aged 65–74 were selected form Health Care Financing Administration records in a similar fashion. Women with a history of breast cancer were excluded.

A study nurse conducted detailed in-home interviews of all cases and controls. During the interview, each participant answered questions about her reproductive, medical, and family history, and her exposure to several known or suspected breast cancer risk factors. Each participant was also asked to confirm her age and race and provide a 30 ml blood sample. All participants provided written informed consent and cases were asked to release their medical records and tumor tissue. The Institutional Review Board at the University of North Carolina (UNC) approved this study.

The overall response rate was 77% for cases and 57% for controls. 90% of controls, or 1816 women, provided sufficient blood samples for inclusion in genotype analyses (1105 whites, 681 African-Americans, 30 other race). 88% of cases provided blood samples (2039 women), but only 55% of cases provided both blood and tumor samples (748 whites, 502 African-Americans, 10 other race). This included 247 in situ cases. Individuals who self-identified as a race other than white or African-American were included in overall analyses but excluded from race-specific assessments.

IHC analysis

Tumor tissue and medical records were collected from area hospitals and sent to UNC. ER and PR status was abstracted from the patient’s medical records, when available. If not available, ER and PR IHC assays were performed at the UNC Immunohistochemistry Core Laboratory. Tumors with more than 5% of cells showing nuclei-specific staining were considered receptor positive (53). Agreement between medical records reports and UNC-run assays in 10% random samples of ER+ and ER− tumors was high (concordance = 81%, kappa = 0.62) (14).

All tumor samples with sufficient tissue were assayed for HER2, HER1 and CK 5/6. A case was considered HER2+ if at least 10% of observed cells showed signs of CB11 monoclonal antibody staining (54). Tissue with any sign of cytoplasmic or membranous staining was considered positive for CK 5/6 or HER1, respectively (3, 13). Due to the limited amount of available tissue, in situ staining techniques were slightly modified (see Livasy et al. [1]).

As described above, these subtypes were classified as follows: luminal A (ER+ and/or PR+, HER2−), luminal B (ER+ and/or PR+, HER2+), HER2+/ER− (ER−, PR−, HER2+), and basal-like (ER−, PR−, HER2−, HER1+ and/or CK 5/6+). Additionally, tumors negative for all 5 markers were grouped together as the ‘unclassified’ subtype.

SNP selection

Single nucleotide polymorphisms (SNPs) from ten early breast cancer GWAS (55–62) or GWAS follow-up studies (63, 64) were selected for inclusion in this subtype evaluation study. We included SNPs from these studies that had genome-wide p-values below 10⁻⁵ in preliminary or pooled analyses. We also retained SNPs in CASP8, ATM, and TP53, some of the key genes identified in a recent comprehensive meta-analysis (45). Lastly, we included a number of SNPs in the same gene as GWAS selected variants, most of which were originally selected to enhance coverage of these regions. In total, this analysis included 22 GWAS hits, 19 other GWAS-identified variants that fell short of genome-wide significance criteria, 21 SNPs from CASP8, ATM, or TP53, and 21 tag SNPs from select GWAS genes.

Each CBCS participant was genotyped at 144 ancestry informative markers. This genotype information was used to estimate each participant’s proportion of African ancestry. When included in regression models, this ancestry proportion estimate should control confounding due to population stratification (65, 66).

Genotype analysis

The included SNPs were genotyped using either a Taqman panel (Applied Biosystems, Inc.) or a Custom GoldenGate Genotyping assay (Illumina, Inc.). The majority of SNPs were genotyped on the Illumina panel, as described previously (67). The Taqman panel (68) included SNPs that had low Illumina design scores, failed the Illumina assay, or were identified as GWAS hits after the Illumina assays were performed. Eighty-one women with poor genotyping quality on the Illumina panel were assigned missing values for those SNPs. All of the SNPs selected for inclusion in this subtype analysis passed quality control tests, including those for call rate, assay intensity, and genotype clustering.

For each SNP, we examined published studies to determine which allele was associated with an increased risk of breast cancer in previous analyses. This allele was designated as the risk allele. For whites, we selected risk alleles for all ATM, CASP8, and TP53 SNPs based on the Zhang et al. meta-analysis (45). For the remaining SNPs, we ascertained the risk allele in the initial GWAS (54–63) and subsequent replication studies (see supplementary references). In each case, if the 95% confidence interval (CI) limits excluded the null, the odds ratio (OR) for the specified allele was in the same direction as the initial study. Despite some minor discrepancies in the direction of the ORs in African-American only studies, we assigned the same risk allele for both racial groups to allow pooling and facilitate comparisons. For novel SNPs and SNPs with no prior statistically significant findings, we designated the minor variant as the risk variant, using the HapMap CEU population as a reference.

Statistical methods

We calculated case-stratified descriptive statistics for age, proportion of African ancestry, and menopausal status, and then repeated these analyses for white and African-American participants separately. We also examined overall and race-stratified distributions of stage at diagnosis, breast cancer subtype, and ER, PR, and HER2 status. Participants were weighted according to their inverse sampling probability. Similarly, all regression models included an offset term to account for the weighted sampling procedures.

For all SNPs, we calculated overall and race-stratified risk allele frequencies (RAFs). We tested for departures from Hardy-Weinberg equilibrium (HWE) separately in white and African-American controls using Pearson’s chi-squared test. If a SNP had a HWE p-value less than 0.05 in either population, we re-inspected the SNP’s genotype clustering images for indications of poor genotype differentiation or other lab error. As many of the SNPs were located in the same gene or gene regions, we also calculated overall and race-stratified correlation coefficients.

We calculated ORs and 95% posterior intervals (PIs) for the association between each subtype and SNP using Bayesian polytomous logistic regression models. We assumed additive genetic models and adjusted for self-reported race (African-American or non- African-American), proportion of African ancestry, and age at diagnosis or selection. We centered age at 50 years and ancestry at its mean value. We also calculated race-specific ORs and 95% PIs, adjusting for age and ancestry.

Previous studies of the association between known susceptibility variants and breast cancer have produced ORs in the range of 1.1–1.3 (44, 45, 50), but subtype-specific associations are less well-characterized. Bearing this in mind, we assigned each SNP log OR a null-centered prior with a mean of 0, but selected a variance of τ² ~ 1/Γ(4, 0.5) to reflect the likely effect size. These parameters correspond to prior SNP-subtype ORs with 95% mass between 0.54 and 1.86 when τ² is equal to the mode of the distribution (0.1). As a full Bayes approach requires priors for all parameters, we also assigned null-centered, lognormal priors for age, ancestry, race and the intercept term. We assigned relatively informative priors to age and ancestry (τ²=0.68), which were both mean-centered variables, but a larger variance to race (τ²=1.0). Because the intercept is difficult to define or interpret in a case-control study with weighted sampling, we assigned a vague prior, with τ²=1000. We assumed that all priors were independent.

Priors were incorporated into regression models using Bayes’ theorem. Briefly, Bayes’ theorem states the posterior probability distribution for the parameter of interest given the observed data, f(β|D), is proportional to the likelihood of the observed data, L(β;D), multiplied by the prior probability distribution f(β) (69–71). The aforementioned likelihood is identical to the likelihood used to obtain standard, frequentist maximum likelihood estimates (MLEs). Put another way, the posterior odds ratio is an inverse-variance weighted combination of the likelihood and prior distribution. Further, the variance of the resulting, normal posterior distribution is the inverse of the sum of the weights.

We also conducted sensitivity analyses, estimating MLE of ORs and 95% CIs and another set of Bayesian ORs and 95% PIs given a more informative, but still null-centered prior [SNP~N(0, τ²), τ²~1/Γ (3, 0.2), mode=0.05]. For each Bayesian model, we took 50,000 samples, discarding the first 1000 draws as a burn in, and thinning by retaining every tenth draw, such that the results are based on 4990 samples. Autocorrelation, trace, and density plots indicated adequate mixing and model convergence. All analyses were conducted using the SAS procedure MCMC (v9.3, Cary, NC). Example code is provided as supplementary material.

RESULTS

As seen in Table 1, white and African-American participants in the CBCS population differed in a few key ways. African-Americans were more likely to have later stage disease, with 63% presenting at stage II or higher, relative to 48% of whites. African-Americans were also less likely to be postmenopausal at the time of their diagnosis and were more likely to have basal-like (22% vs. 11%), unclassified (14% vs. 8%) or HER2+/ER− disease (8% vs. 6%). Luminal A breast cancer was the most common breast cancer subtype overall (60%). Seven SNPs had HWE p-values<0.05 (Table 2), though no SNPs failed HWE tests in both whites and African-Americans. Upon reinspection of genotype clustering images, we found that six of the seven SNPs showed good differentiation with no overlap between genotypes. We excluded the seventh SNP, rs614367 (MYEOV), after discovering evidence of allelic dropout and disparate clustering within the homozygous rare genotype. We also excluded SNPs with minor allele frequencies less than 1% in our sample. This left us with 78 SNPs in the overall analysis, 76 in the white only analysis and 73 in the African-American only analysis. Many of the FGFR2 SNPs were highly correlated with one another, as were the two COX11 SNPs, the two CASP8 SNPs and some TNRC9/TOX3 and ATM SNPs (Supplementary Tables S1a–S1f).

Table 1.

Descriptive statistics for Carolina Breast Cancer Study participants included in subtype analysis

	Cases				Controls
	Overall (%)a N=1260b	White (%)a N=748	African Americans (%)a N=502	Overall (%)a N=1816b	White (%)a N=1105	African Americans (%)a N=681
Age (years); mean (std)	51.5 (11.6)	52.1 (11.8)	50.8 (11.4)	52.5 (11.3)	53.0 (11.2)	51.9 (11.3)
Proportion African Ancestry; mean (std)	0.35 (0.36)	0.06 (0.06)	0.77 (0.13)	0.33 (0.36)	0.07 (0.09)	0.77 (0.14)
Postmenopausal; N (%)	691 (67)	417 (70)	269 (56)	1032 (38)	640 (37)	377 (41)
Stage of Disease; N (%)
In situ	247 (10)	192 (10)	52 (10)
Stage I	369 (39)	232 (42)	136 (28)
Stage II	492 (42)	250 (40)	237 (50)
Stage III	100 (7)	46 (6)	54 (11)
Stage IV	24 (2)	12 (2)	11 (2)
missing	28	16	12
Subtype; N (%)
Luminal A	700 (60)	453 (64)	242 (49)
Luminal B	122 (11)	82 (11)	38 (8)
HER2+/ER−	98 (6)	59 (6)	39 (8)
Basal-like	207 (13)	94 (11)	112 (22)
Unclassified	133 (9)	60 (8)	71 (14)
ER+; N (%)	745 (66)	497 (70)	243 (50)
PR+; N (%)	543 (60)	341 (64)	197 (45)
missing	247	192	52
HER2+; N (%)	220 (17)	141 (17)	77 (16)

^aPercentages weighted by inverse sampling probability

^bIncludes those who self-identified as a race other than white or African-American

Table 2.

Risk allele frequencies (RAF) by race and case status, African-Americans and non African-Americans in the Carolina Breast Cancer Study

Gene	Locus	Risk allele	All (1260 cases, 1817 controls)a		Whites (748 cases, 1105 controls)			African-Americans (502 cases, 681 controls)
			RAF casesb	RAF controlsb	RAF casesb	RAF controlsb	HWE p-value	RAF casesb	RAF controlsb	HWE p-value
1p12	rs11249433	G	0.37	0.36	0.44	0.41	0.54	0.14	0.10	0.01
CASP8	rs1045485	G	0.89	0.88	0.88	0.87	0.63	0.94	0.95	0.74
CASP8	rs17468277	C	0.89	0.88	0.88	0.87	0.63	0.95	0.95	0.95
2q35	rs13387042	A	0.59	0.52	0.55	0.47	0.83	0.73	0.73	1.00
2p	rs4666451	G	0.64	0.65	0.60	0.63	0.30	0.78	0.77	0.12
SLC4A7	rs4973768	T	0.44	0.42	0.47	0.42	0.22	0.35	0.40	0.05
4p	rs12505080	C	0.27	0.23	0.30	0.24	0.80	0.18	0.17	0.64
TLR1	rs7696175	T	0.38	0.38	0.46	0.45	0.91	0.09	0.06	0.54
MRPS30	rs4415084	T	0.46	0.45	0.42	0.42	0.18	0.64	0.58	0.70
MRPS30	rs10941679	G	0.27	0.28	0.29	0.30	0.76	0.19	0.19	0.17
5p12	rs981782	T	0.60	0.65	0.51	0.59	0.26	0.91	0.91	0.60
5q	rs30099	T	0.11	0.10	0.10	0.10	0.40	0.15	0.12	0.75
MAP3K1	rs889312	C	0.33	0.35	0.33	0.34	0.85	0.33	0.36	0.08
ESR1	rs2046210	A	0.43	0.40	0.38	0.35	0.48	0.62	0.61	0.15
ESR1	rs851974	G	0.36	0.39	0.41	0.43	0.28	0.18	0.17	0.46
ESR1	rs2077647	A	0.50	0.49	0.50	0.49	0.64	0.51	0.51	0.16
ESR1	rs2234693	T	0.50	0.55	0.51	0.57	0.45	0.46	0.48	0.63
ESR1	rs1801132	C	0.79	0.78	0.76	0.76	0.43	0.90	0.88	0.36
ESR1	rs3020314	C	0.44	0.41	0.38	0.34	0.15	0.68	0.71	0.75
ESR1	rs3798577	T	0.52	0.53	0.50	0.53	0.43	0.58	0.54	0.27
ECHDC1	rs2180341	G	0.26	0.28	0.24	0.27	0.55	0.32	0.33	0.83
RELN	rs17157903	T	0.14	0.12	0.14	0.12	0.06	0.11	0.10	0.08
8q24	rs13281615	G	0.43	0.42	0.43	0.42	0.17	0.43	0.43	0.58
8q24	rs1562430	T	0.59	0.56	0.60	0.57	0.78	0.53	0.53	0.61
CDKN2A/B	rs3731257	T	0.20	0.21	0.23	0.23	0.24	0.08	0.11	0.89
CDKN2A/B	rs3731249	A	0.02	0.02	0.03	0.03	0.90	0.01	0.00	0.95
CDKN2A/B	rs518394	G	0.36	0.41	0.44	0.48	0.17	0.09	0.08	0.06
CDKN2A/B	rs564398	G	0.35	0.40	0.42	0.47	0.29	0.08	0.08	0.02
CDKN2A/B	rs1011970	T	0.22	0.18	0.19	0.15	0.62	0.33	0.34	0.14
CDKN2A/B	rs10757278	A	0.60	0.60	0.55	0.55	0.18	0.80	0.82	0.77
CDKN2A/B	rs10811661	C	0.14	0.18	0.16	0.20	0.02	0.07	0.07	0.24
ANKRD16	rs2380205	C	0.54	0.58	0.57	0.60	0.88	0.41	0.46	0.72
ZNF365	rs10995190	G	0.85	0.83	0.86	0.82	0.76	0.82	0.83	0.90
ZMIZ1	rs704010	T	0.35	0.36	0.42	0.42	0.93	0.11	0.08	0.82
FGFR2	rs1896395	A	0.05	0.04	0.00	0.00	0.96	0.21	0.20	0.04
FGFR2	rs3750817	C	0.71	0.65	0.65	0.60	0.16	0.91	0.88	0.83
FGFR2	rs10736303	G	0.61	0.55	0.54	0.49	0.19	0.86	0.84	0.75
FGFR2	rs11200014	A	0.40	0.38	0.45	0.41	0.65	0.22	0.21	0.75
FGFR2	rs2981579	T	0.49	0.45	0.46	0.41	0.51	0.61	0.61	0.10
FGFR2	rs1078806	G	0.40	0.38	0.45	0.41	0.53	0.22	0.21	0.99
FGFR2	rs2981578	C	0.61	0.55	0.54	0.49	0.09	0.86	0.84	0.45
FGFR2	rs1219648	G	0.44	0.40	0.44	0.39	0.35	0.47	0.41	0.57
FGFR2	rs2912774	A	0.47	0.42	0.44	0.40	0.26	0.58	0.55	0.07
FGFR2	rs2936870	T	0.47	0.43	0.44	0.40	0.25	0.59	0.56	0.14
FGFR2	rs2420946	T	0.45	0.41	0.43	0.39	0.21	0.54	0.52	0.03
FGFR2	rs2162540	G	0.45	0.41	0.42	0.39	0.28	0.54	0.52	0.41
FGFR2	rs2981582	T	0.44	0.40	0.42	0.39	0.30	0.50	0.49	0.96
FGFR2	rs3135718	G	0.46	0.42	0.43	0.39	0.23	0.58	0.54	0.65
10q	rs10510126	C	0.89	0.90	0.89	0.89	0.38	0.89	0.90	0.21
ATM	rs1800054	G	0.02	0.01	0.02	0.02	0.34	0.00	0.00	0.94
ATM	rs4986761	C	0.01	0.01	0.01	0.01	0.68	0.00	0.00	0.98
ATM	rs1800056	C	0.02	0.01	0.02	0.01	0.67	0.00	0.00	0.95
ATM	rs1800057	G	0.03	0.02	0.04	0.02	0.90	0.01	0.01	0.91
ATM	rs1800058	T	0.02	0.02	0.02	0.02	0.06	0.01	0.01	0.91
ATM	rs1801516	A	0.12	0.12	0.15	0.14	0.17	0.03	0.02	0.48
ATM	rs3092992	C	0.04	0.04	0.05	0.04	0.13	0.01	0.01	0.77
ATM	rs664143	C	0.60	0.61	0.58	0.60	0.70	0.67	0.68	0.45
ATM	rs170548	G	0.27	0.33	0.32	0.37	0.88	0.09	0.12	0.07
ATM	rs3092993	A	0.12	0.12	0.15	0.14	0.19	0.03	0.02	0.48
LSP1	rs3817198	C	0.29	0.31	0.32	0.34	0.18	0.17	0.17	0.16
LSP1	rs909116	T	0.57	0.56	0.53	0.52	0.20	0.72	0.72	0.96
MYEOV	rs614367	T	0.16	0.12	0.17	0.11	0.05	0.14	0.15	0.33
H19	rs2107425	C	0.65	0.65	0.70	0.68	0.74	0.50	0.53	0.42
TNRC9/TOX3	rs8049149	T	0.00	0.00	0.00	0.00	0.98	0.01	0.02	0.32
TNRC9/TOX3	rs16951186	T	0.05	0.04	0.01	0.01	0.75	0.17	0.19	0.95
TNRC9/TOX3	rs8051542	T	0.43	0.41	0.46	0.44	0.43	0.33	0.30	0.12
TNRC9/TOX3	rs12443621	G	0.50	0.43	0.51	0.41	0.39	0.48	0.51	1.00
TNRC9/TOX3	rs3803662	T	0.36	0.29	0.32	0.24	0.73	0.52	0.54	0.65
TNRC9/TOX3	rs4784227	T	0.24	0.19	0.29	0.22	0.62	0.09	0.07	0.59
TNRC9/TOX3	rs3104746	A	0.08	0.05	0.03	0.02	0.48	0.26	0.18	0.87
TNRC9/TOX3	rs3112562	G	0.28	0.25	0.22	0.20	0.45	0.51	0.46	0.88
TNRC9/TOX3	rs9940048	A	0.26	0.25	0.25	0.24	0.50	0.31	0.30	0.64
TP53	rs9894946	G	0.84	0.87	0.81	0.86	0.48	0.96	0.96	0.25
TP53	rs1614984	T	0.41	0.39	0.41	0.39	0.22	0.39	0.40	0.03
TP53	rs4968187	T	0.00	0.00	0.00	0.00	0.93	0.01	0.00	0.92
TP53	rs12951053	C	0.08	0.07	0.08	0.06	0.47	0.11	0.11	0.09
TP53	rs17880604	C	0.01	0.01	0.01	0.01	0.21	0.00	0.00	0.95
TP53	rs1800372	G	0.01	0.01	0.02	0.02	0.54	0.01	0.00	0.98
TP53	rs2909430	G	0.17	0.14	0.14	0.13	0.66	0.28	0.24	0.64
TP53	rs1042522	C	0.67	0.71	0.75	0.77	0.64	0.40	0.43	0.77
TP53	rs8079544	C	0.94	0.94	0.96	0.95	1.00	0.89	0.89	0.83
COX11	rs7222197	G	0.70	0.73	0.72	0.75	0.60	0.66	0.65	0.70
COX11	rs6504950	G	0.70	0.73	0.72	0.75	0.59	0.66	0.65	0.66

^aIncludes individuals who identified as a race other than white or African-American

^bWeighted by inverse sampling probability

Several SNPs were associated with luminal A breast cancer, including 13 of 14 evaluated FGFR2 SNPs (ORs≈1.25) and several SNPs in TNRC9/TOX3 (Table 3). The strongest association was seen for rs3104746 on TNRC9/TOX3 (OR=1.58, 95% PI: 1.24, 1.94). Other noteworthy associations included rs13387042 (2q35), rs12505080 (4p), rs7696175 (TLR1), rs889312 (MAP3K1), rs851974 (ESR1), rs1011970 (CDKN2A/B), and rs9894946 (TP53). Many of the previously identified GWAS hits were positively associated with luminal A disease (Figure 1).

Table 3.

Odds ratios and 95% posterior intervals for the association between the selected single nucleotide polymorphisms (SNPs) and each breast cancer subtype, relative to controls [SNP log OR~N(0,τ²), τ² ~ Γ⁻¹(4, 0.5) with mode=0.10]^a

Gene	SNP	Luminal A N=700	Luminal B N=122	HER2+/ER− N=98	Unclassified N=133	Basal-like N=207
1p12	rs11249433	1.09 (0.94, 1.25)	1.04 (0.78, 1.33)	1.11 (0.79, 1.45)	1.24 (0.93, 1.58)	1.09 (0.85, 1.35)
CASP8	rs1045485	1.10 (0.88, 1.32)	1.00 (0.67, 1.35)	0.97 (0.62, 1.33)	1.24 (0.80, 1.73)	1.15 (0.80, 1.56)
CASP8	rs17468277	1.11 (0.89, 1.35)	0.96 (0.64, 1.34)	0.94 (0.60, 1.32)	1.29 (0.84, 1.82)	1.10 (0.79, 1.47)
2q35	rs13387042	1.18 (1.04, 1.35)	0.99 (0.76, 1.25)	0.98 (0.69, 1.25)	1.10 (0.85, 1.40)	0.92 (0.73, 1.12)
2p	rs4666451	0.97 (0.85, 1.11)	1.03 (0.76, 1.29)	1.14 (0.84, 1.48)	1.24 (0.93, 1.56)	1.16 (0.91, 1.41)
SLC4A7	rs4973768	0.98 (0.86, 1.09)	1.01 (0.78, 1.27)	0.91 (0.68, 1.17)	0.93 (0.70, 1.15)	0.99 (0.78, 1.19)
4p	rs12505080	1.18 (1.02, 1.36)	1.14 (0.84, 1.45)	1.15 (0.82, 1.54)	0.88 (0.61, 1.13)	0.98 (0.75, 1.21)
TLR1	rs7696175	1.20 (1.02, 1.38)	0.99 (0.74, 1.29)	0.79 (0.54, 1.04)	1.08 (0.77, 1.38)	1.27 (0.96, 1.59)
MRPS30	rs4415084	1.11 (0.96, 1.25)	1.06 (0.80, 1.32)	1.15 (0.85, 1.49)	1.08 (0.84, 1.35)	1.17 (0.93, 1.41)
MRPS30	rs10941679	1.10 (0.93, 1.27)	0.94 (0.69, 1.22)	1.04 (0.73, 1.36)	0.95 (0.69, 1.23)	1.21 (0.95, 1.45)
5p12	rs981782	0.90 (0.78, 1.03)	0.88 (0.64, 1.12)	1.08 (0.78, 1.41)	0.99 (0.72, 1.28)	1.07 (0.82, 1.33)
5q	rs30099	1.09 (0.89, 1.31)	1.03 (0.66, 1.37)	1.13 (0.75, 1.56)	0.91 (0.60, 1.20)	0.99 (0.73, 1.28)
MAP3K1	rs889312	1.17 (1.02, 1.33)	1.04 (0.80, 1.32)	1.00 (0.74, 1.30)	1.08 (0.82, 1.35)	0.89 (0.70, 1.09)
ESR1	rs2046210	1.03 (0.90, 1.16)	1.07 (0.82, 1.36)	1.29 (0.97, 1.67)	1.15 (0.88, 1.44)	1.29 (1.01, 1.55)
ESR1	rs851974	0.89 (0.77, 1.02)	1.13 (0.84, 1.43)	0.86 (0.62, 1.12)	0.88 (0.64, 1.10)	0.96 (0.76, 1.18)
ESR1	rs2077647	0.95 (0.83, 1.08)	1.00 (0.76, 1.25)	0.98 (0.72, 1.26)	0.94 (0.71, 1.16)	0.96 (0.76, 1.17)
ESR1	rs2234693	0.91 (0.79, 1.04)	0.96 (0.75, 1.20)	1.02 (0.75, 1.27)	0.86 (0.67, 1.08)	0.99 (0.79, 1.18)
ESR1	rs1801132	1.09 (0.92, 1.28)	0.91 (0.66, 1.20)	0.98 (0.67, 1.31)	1.02 (0.74, 1.32)	0.85 (0.64, 1.05)
ESR1	rs3020314	1.02 (0.89, 1.17)	1.23 (0.93, 1.54)	1.08 (0.80, 1.37)	0.99 (0.77, 1.23)	1.12 (0.90, 1.34)
ESR1	rs3798577	0.97 (0.86, 1.10)	1.05 (0.79, 1.32)	0.82 (0.62, 1.02)	1.10 (0.85, 1.35)	1.00 (0.82, 1.21)
ECHDC1	rs2180341	0.99 (0.84, 1.13)	1.01 (0.75, 1.31)	1.04 (0.74, 1.36)	0.95 (0.70, 1.20)	1.01 (0.80, 1.23)
RELN	rs17157903	1.00 (0.83, 1.18)	1.07 (0.74, 1.41)	0.77 (0.48, 1.09)	0.99 (0.70, 1.32)	1.09 (0.78, 1.39)
8q24	rs13281615	1.04 (0.90, 1.17)	0.99 (0.75, 1.25)	1.13 (0.85, 1.43)	1.08 (0.85, 1.35)	0.98 (0.79, 1.17)
8q24	rs1562430	1.06 (0.93, 1.19)	0.96 (0.72, 1.22)	1.20 (0.90, 1.55)	1.06 (0.83, 1.32)	1.07 (0.85, 1.28)
CDKN2A/B	rs3731257	0.92 (0.78, 1.06)	0.89 (0.64, 1.17)	0.96 (0.66, 1.26)	0.91 (0.66, 1.19)	0.91 (0.70, 1.16)
CDKN2A/B	rs3731249	0.99 (0.65, 1.33)	1.08 (0.55, 1.68)	1.11 (0.56, 1.76)	0.95 (0.49, 1.46)	0.96 (0.51, 1.48)
CDKN2A/B	rs518394	0.99 (0.86, 1.13)	1.00 (0.73, 1.26)	0.77 (0.52, 1.00)	1.04 (0.76, 1.34)	1.14 (0.88, 1.38)
CDKN2A/B	rs564398	1.01 (0.85, 1.15)	1.01 (0.76, 1.28)	0.81 (0.56, 1.08)	1.08 (0.80, 1.40)	1.09 (0.83, 1.36)
CDKN2A/B	rs1011970	1.12 (0.96, 1.29)	0.91 (0.65, 1.15)	0.90 (0.63, 1.15)	0.87 (0.63, 1.10)	1.12 (0.86, 1.38)
CDKN2A/B	rs10757278	1.05 (0.90, 1.18)	1.06 (0.80, 1.35)	1.02 (0.75, 1.31)	1.22 (0.93, 1.52)	1.10 (0.86, 1.34)
CDKN2A/B	rs10811661	0.95 (0.78, 1.13)	1.01 (0.71, 1.35)	0.82 (0.54, 1.15)	0.84 (0.56, 1.12)	1.10 (0.81, 1.39)
ANKRD16	rs2380205	1.03 (0.90, 1.15)	1.05 (0.79, 1.30)	1.10 (0.82, 1.40)	0.80 (0.61, 0.99)	0.94 (0.77, 1.13)
ZNF365	rs10995190	1.03 (0.84, 1.22)	0.83 (0.59, 1.09)	0.83 (0.59, 1.11)	1.24 (0.88, 1.64)	0.95 (0.74, 1.21)
ZMIZ1	rs704010	1.09 (0.94, 1.25)	1.34 (0.96, 1.70)	1.36 (0.97, 1.77)	0.96 (0.70, 1.25)	1.34 (1.03, 1.66)
FGFR2	rs1896395	1.05 (0.80, 1.30)	0.92 (0.54, 1.33)	1.31 (0.73, 1.89)	1.06 (0.70, 1.44)	1.06 (0.71, 1.40)
FGFR2	rs3750817	1.33 (1.13, 1.53)	1.27 (0.92, 1.64)	1.22 (0.86, 1.60)	1.26 (0.92, 1.64)	1.01 (0.79, 1.25)
FGFR2	rs10736303	1.32 (1.14, 1.52)	1.31 (0.98, 1.65)	1.26 (0.90, 1.64)	1.37 (0.99, 1.77)	0.99 (0.78, 1.22)
FGFR2	rs11200014	1.26 (1.10, 1.43)	1.05 (0.80, 1.31)	1.19 (0.86, 1.52)	1.41 (1.08, 1.77)	0.88 (0.70, 1.07)
FGFR2	rs2981579	1.26 (1.10, 1.42)	1.11 (0.84, 1.41)	1.34 (1.00, 1.68)	1.44 (1.11, 1.80)	0.92 (0.74, 1.10)
FGFR2	rs1078806	1.24 (1.08, 1.42)	1.07 (0.79, 1.36)	1.19 (0.86, 1.53)	1.40 (1.06, 1.76)	0.87 (0.68, 1.07)
FGFR2	rs2981578	1.33 (1.15, 1.51)	1.31 (1.01, 1.67)	1.34 (0.98, 1.76)	1.38 (1.03, 1.84)	1.01 (0.79, 1.25)
FGFR2	rs1219648	1.29 (1.13, 1.47)	1.10 (0.83, 1.37)	1.24 (0.89, 1.59)	1.62 (1.24, 2.04)	0.95 (0.76, 1.15)
FGFR2	rs2912774	1.26 (1.10, 1.41)	1.11 (0.84, 1.39)	1.42 (1.06, 1.80)	1.47 (1.14, 1.85)	0.92 (0.73, 1.09)
FGFR2	rs2936870	1.26 (1.09, 1.41)	1.13 (0.86, 1.43)	1.38 (1.02, 1.76)	1.50 (1.16, 1.89)	0.91 (0.72, 1.09)
FGFR2	rs2420946	1.22 (1.06, 1.38)	1.06 (0.80, 1.32)	1.40 (1.04, 1.79)	1.46 (1.12, 1.83)	0.89 (0.71, 1.07)
FGFR2	rs2162540	1.28 (1.11, 1.45)	1.08 (0.82, 1.36)	1.42 (1.05, 1.83)	1.52 (1.17, 1.90)	0.91 (0.72, 1.10)
FGFR2	rs2981582	1.27 (1.09, 1.43)	1.10 (0.87, 1.40)	1.39 (1.02, 1.76)	1.28 (1.00, 1.57)	0.92 (0.73, 1.10)
FGFR2	rs3135718	1.26 (1.10, 1.42)	1.13 (0.85, 1.40)	1.35 (1.01, 1.71)	1.51 (1.17, 1.90)	0.91 (0.73, 1.09)
10q	rs10510126	1.09 (0.88, 1.30)	1.07 (0.73, 1.46)	1.08 (0.72, 1.49)	0.94 (0.63, 1.26)	1.14 (0.81, 1.48)
ATM	rs1800054	1.10 (0.66, 1.59)	1.10 (0.54, 1.76)	1.12 (0.51, 1.78)	1.00 (0.49, 1.62)	1.20 (0.58, 1.91)
ATM	rs1800057	1.19 (0.81, 1.64)	0.95 (0.49, 1.48)	1.33 (0.68, 2.13)	1.01 (0.50, 1.60)	1.10 (0.56, 1.67)
ATM	rs1800058	1.05 (0.67, 1.44)	0.94 (0.45, 1.46)	1.03 (0.51, 1.67)	0.98 (0.47, 1.56)	0.98 (0.49, 1.56)
ATM	rs1801516	1.03 (0.82, 1.24)	1.05 (0.70, 1.43)	0.97 (0.64, 1.37)	0.95 (0.63, 1.32)	0.99 (0.70, 1.32)
ATM	rs3092992	0.97 (0.68, 1.30)	0.99 (0.57, 1.49)	1.19 (0.62, 1.85)	1.05 (0.61, 1.57)	1.38 (0.79, 1.97)
ATM	rs664143	1.09 (0.95, 1.23)	1.00 (0.77, 1.25)	1.06 (0.79, 1.35)	0.94 (0.74, 1.17)	0.98 (0.79, 1.19)
ATM	rs170548	0.98 (0.84, 1.13)	0.90 (0.66, 1.16)	1.00 (0.72, 1.31)	0.94 (0.69, 1.22)	1.00 (0.77, 1.24)
ATM	rs3092993	1.03 (0.83, 1.25)	1.06 (0.69, 1.48)	0.94 (0.60, 1.35)	0.96 (0.61, 1.29)	1.03 (0.73, 1.37)
LSP1	rs3817198	1.02 (0.88, 1.18)	0.87 (0.62, 1.10)	1.19 (0.86, 1.52)	1.21 (0.92, 1.55)	1.02 (0.79, 1.26)
LSP1	rs909116	1.09 (0.94, 1.23)	0.92 (0.69, 1.15)	1.23 (0.90, 1.58)	1.03 (0.79, 1.30)	1.09 (0.86, 1.32)
H19	rs2107425	1.03 (0.90, 1.17)	0.93 (0.71, 1.17)	0.99 (0.74, 1.24)	0.94 (0.72, 1.18)	1.00 (0.81, 1.19)
TNRC9/TOX3	rs16951186	1.02 (0.78, 1.25)	1.30 (0.79, 1.88)	0.84 (0.48, 1.21)	0.83 (0.52, 1.12)	0.97 (0.67, 1.31)
TNRC9/TOX3	rs8051542	1.10 (0.97, 1.24)	0.96 (0.74, 1.20)	1.12 (0.84, 1.43)	1.31 (1.02, 1.64)	0.84 (0.66, 1.03)
TNRC9/TOX3	rs12443621	1.06 (0.94, 1.20)	0.93 (0.71, 1.16)	1.21 (0.92, 1.55)	0.95 (0.74, 1.18)	1.00 (0.82, 1.21)
TNRC9/TOX3	rs3803662	1.16 (1.01, 1.33)	1.09 (0.83, 1.35)	1.01 (0.77, 1.29)	1.13 (0.88, 1.41)	0.96 (0.78, 1.16)
TNRC9/TOX3	rs4784227	1.32 (1.13, 1.54)	1.09 (0.76, 1.43)	1.10 (0.76, 1.46)	1.29 (0.94, 1.67)	0.90 (0.66, 1.17)
TNRC9/TOX3	rs3104746	1.58 (1.24, 1.94)	1.05 (0.60, 1.50)	1.31 (0.80, 1.85)	1.12 (0.76, 1.58)	1.49 (1.06, 1.98)
TNRC9/TOX3	rs3112562	1.07 (0.93, 1.22)	0.88 (0.62, 1.14)	1.46 (1.06, 1.87)	0.80 (0.60, 1.03)	1.33 (1.06, 1.62)
TNRC9/TOX3	rs9940048	1.13 (0.97, 1.28)	0.84 (0.61, 1.08)	1.17 (0.86, 1.50)	0.98 (0.72, 1.25)	0.91 (0.72, 1.11)
TP53	rs9894946	0.86 (0.72, 1.02)	0.83 (0.59, 1.12)	1.01 (0.65, 1.36)	1.05 (0.71, 1.43)	1.12 (0.81, 1.49)
TP53	rs1614984	1.01 (0.88, 1.13)	0.94 (0.71, 1.16)	1.01 (0.78, 1.28)	1.04 (0.79, 1.30)	1.13 (0.92, 1.36)
TP53	rs12951053	0.95 (0.75, 1.18)	1.32 (0.86, 1.84)	1.16 (0.71, 1.62)	1.25 (0.82, 1.73)	0.99 (0.70, 1.30)
TP53	rs17880604	0.87 (0.49, 1.24)	0.84 (0.35, 1.38)	1.21 (0.54, 1.94)	0.96 (0.46, 1.57)	1.12 (0.57, 1.76)
TP53	rs1800372	0.97 (0.55, 1.38)	1.24 (0.57, 2.00)	0.93 (0.39, 1.53)	1.34 (0.59, 2.22)	1.04 (0.46, 1.71)
TP53	rs2909430	1.12 (0.96, 1.31)	1.09 (0.78, 1.45)	1.06 (0.73, 1.38)	0.88 (0.63, 1.13)	1.10 (0.85, 1.37)
TP53	rs1042522	1.03 (0.89, 1.18)	0.99 (0.73, 1.26)	0.82 (0.61, 1.06)	0.94 (0.70, 1.17)	0.97 (0.77, 1.18)
TP53	rs8079544	0.98 (0.78, 1.23)	1.38 (0.83, 2.07)	0.76 (0.45, 1.07)	0.87 (0.59, 1.20)	1.05 (0.71, 1.44)
COX11	rs7222197	1.06 (0.92, 1.22)	1.08 (0.80, 1.35)	1.01 (0.74, 1.29)	1.09 (0.82, 1.37)	1.05 (0.85, 1.27)
COX11	rs6504950	1.05 (0.91, 1.20)	1.07 (0.80, 1.37)	1.00 (0.73, 1.28)	1.08 (0.81, 1.35)	1.05 (0.84, 1.28)

^aEstimates generated using polytomous logistic regression adjusting for age at diagnosis/selection, proportion African ancestry and self-reported race.

Figure 1.

Odds ratios and 95% posterior intervals for previous GWAS–identified SNPs: All CBCS participants

HER2+/ER− disease and unclassified disease were also strongly associated with several FGFR2 and TNRC9/TOX3 SNPs. For both subtypes, the OR estimates for the FGFR2 SNPs were high, with many at or near 1.4. Beyond these key genes, HER2+/ER− disease was positively correlated with the designated risk variant at rs2046210 on ESR1, and rs704010 on ZMIZ1, but negatively correlated with the risk variant at rs7696175 on TLR1, rs3798577 on ESR1, and rs518394 on CDKN2A/B. The ‘C’ allele at rs2380205 (ANKRD16) was inversely associated with the risk of unclassified breast cancer.

We identified relatively few susceptibility variants for luminal B breast cancer. All but one FGFR2 SNP was associated with increased disease risk, but the observed effects were weaker than the other non basal-like subtype ORs, and only one had a posterior interval that excluded the null (rs2981578). The risk allele at rs704010 on ZMIZ1 was also associated with luminal B disease (OR=1.34, 95% PI: 0.96, 1.70).

None of the FGFR2 SNPs were associated with an increased risk of basal-like breast cancer. In fact, most of the FGFR2 ORs for basal-like disease were less than one. Risk variants at two TNRC9/TOX3 SNPs (rs3014746 and rs3112562) were positively associated with basal-like disease, as were risk variants at rs704010 on ZMIZ1 and rs2046210 on ESR1. Additionally, rs7696175 on TLR1 and rs10941679 on MRPS30 each had ORs greater than 1.2 for basal-like breast cancer, relative to controls.

Race-stratified subtype analyses revealed a few additional insights (Supplementary Tables S2 and S3). The most striking was for rs10757278 on CDKN2A/B, where the ‘A’ allele was positively associated with luminal A disease in whites (OR=1.19, 95% PI: 1.02, 1.39) but negatively associated with disease in African-Americans (OR=0.75, 95% PI: 0.58, 0.94). Race-specific results for the previous GWAS hits can be seen in Figure 2.

Figure 2.

Odds ratios and 95% posterior intervals for previous GWAS–identified SNPs for CBCS whites (left) and African–Americans (right)

The two 8q24 SNPs were strongly associated with luminal A breast cancer only among whites (OR=1.16, 95% PI: 0.98, 1.35 and OR=1.17, 95% PI: 1.00, 1.37 for rs13281615 and rs1562430, respectively). The same was true for a TNRC9/TOX3 SNP (rs8051542 OR=1.16, 95% PI: 0.99, 1.35) and a LSP1 SNP (rs909116 OR=1.17, 95% PI: 0.99, 1.37). As for the other subtypes, rs3112562 and rs12443621 (TNRC9/TOX3) were strongly associated with luminal B (OR=0.64, 95% PI: 0.39, 0.89) and HER2+/ER− breast cancer (OR=1.53, 95% PI: 1.05, 2.09), respectively, only among whites. We observed no noteworthy findings in the African-American only analyses.

Results from the MLE analysis and alternate Bayes analysis are presented in Supplementary Tables S4 and S5. Compared with the MLE results, the ORs and PIs presented here are attenuated towards the null and are more precise. As expected, the rarer risk alleles had larger discrepancies between their Bayesian and MLE ORs. For example, the MLE and Bayesian ORs for basal-like breast cancer and rs1800054 (ATM, RAF=1%) were 1.57 (95% CI: 0.66, 3.75) and 1.20 (95% PI: 0.58, 1.91), respectively, compared with MLE and Bayesian ORs of 1.33 (95% CI: 1.02, 1.73) and 1.27 (95% PI: 0.96, 1.59) for rs7696175 (TLR1, RAF=38%). The ORs from the Bayesian analysis with more informative priors were further attenuated. The SNP-subtype association patterns were consistent across all methods.

DISCUSSION

In this study of breast cancer subtypes and previously established susceptibility variants, we observed critical differences in subtype-specific genetic risk factors. The most conspicuous differences involved the FGFR2 gene, where most of the 14 highly correlated SNPs were associated with luminal A, HER2+/ER− and unclassified disease, but not basal-like disease. We also found evidence that SNPs on or near TNRC9/TOX3 are differentially related to breast cancer subtype and that rs10757278 (CDKN2A/B) is differentially related to luminal A disease by race. SNPs in 2q35, 4p, TLR1, MRPS30, MAP3K1, ESR1, ANKRD16, ZM1Z1, and TP53 may also be related to subtype-specific etiology.

As few other studies have employed these enhanced subtype definitions, it is difficult to compare our results with previous reports. Most prior investigations of this topic were limited to comparisons of a single hormone receptor, usually ER+ versus ER− disease (61, 64, 72–82). A few have looked at risk factors for combined ER, PR, HER2 status (7, 9, 10, 83, 84), but to our knowledge, only one other study by Broeks et al. (8) has examined genetic risk factors according to all 5 IHC markers. Broeks et al. (8), Stevens et al. (7), and Han et al. (9) examined some of the SNPs included in this analysis, with some consistencies across populations.

The only FGFR2 SNP examined by Broeks et al. (8) was rs2981582. They also observed positive associations between the T allele and luminal A disease and no association between the SNP and basal-like breast cancer. Their luminal B OR was in the same direction we observed, but of much greater magnitude. Stevens et al. and Han et al. also found near-null associations between rs2981582 and triple negative disease. Contrary to our findings, however, rs2981582 was not associated with HER2+/ER− disease in either study, nor was it associated with unclassified disease in Stevens et al. The effect estimates for rs2981582 and luminal A and B disease reported by Han et al. are similar to those seen in our study. We are the first to report subtype-specific estimates for any other FGFR2 SNPs.

Broeks et al., Stevens et al., and Han et al. also evaluated one TNRC9/TOX3 SNP, rs3803662. Both Broeks et al. and Han et al. observed a positive association with the T allele and luminal A breast cancer, as we did. However, these authors also observed associations between the T allele and luminal B and HER2+/ER− disease, where we found only a weak association with Luminal B and a near-null association with HER2+/ER− disease. Lastly, Broeks et al. observed an association between rs3803662 and basal-like disease, which we did not observe.

These three study groups also assessed other SNPs included in our panel. While it is difficult to draw clear inferences from individual SNP-subtype analyses, these studies, together with ours, suggest that some important differences by subtype do exist. In addition to FGFR2 and TNRC9/TOX3, the effects of rs2046210 (ESR1), rs13387042 (2q35), and rs889312 (MAP3K1) seem to vary according to subtype. Additional studies are needed to further clarify the role of these SNPs and the other potentially important genes identified in our investigation.

While this is one of the first studies to look at genetic risk factors for specific subtypes, breast cancer susceptibility loci are a commonly studied topic. Bayesian methods allowed us to use this plethora of prior information to generate more precise estimates. Assuming we selected reasonable priors, the results presented here will also be more accurate, on average, than those produced using frequentist methods that do not incorporate the wealth of information from prior studies. Further, by selecting null-centered, highly informative priors, bias resulting from these methods is likely to be towards the null (85). In this way, this application of Bayesian methods also reduces the probability of observing false positive associations. We believe the priors specified here are reasonable given existing knowledge of breast cancer susceptibility variants, but we also provide alternate analyses that demonstrate the influence of our assumptions.

Due to differences in blood and tumor sample availability by race, African-Americans were underrepresented in genotyping analyses but overrepresented in IHC analyses. Women with advanced disease were more likely to provide tumor tissue. These trends may result in biased effect estimates for SNPs related to race or disease aggressiveness. Controlling for self-reported race and ancestry should alleviate some of this bias. Though not included in this analysis, we could have used inverse-probability of selection weighting or Bayesian imputation methods to further address this issue.

There is some disagreement in the field as to how best to classify breast cancer subtypes. As discussed, the IHC markers used here are only proxies for more complex gene expression profiles, and thus may not sufficiently capture tumor heterogeneity (86–89). While our approach is likely more informative than one using three or fewer markers, poor subtype specification may attenuate effects and underestimate subtype differences. Misclassification due to inaccurate medical records or IHC evaluations could also bias effects. Other potential sources of misclassification include allelic drop-out and other genotyping errors, though thorough quality control checks likely limited the impact of such errors.

We included in situ cases to increase sample size and improve precision. While this could bias effect estimates of SNPs associated with disease aggressiveness or progression, shared risk profiles (3, 90, 91) and subtype distributions (1, 3, 92) suggest this bias would be small.

The diverse composition of the CBCS population is a major strength of this study. By recruiting a large proportion of African-Americans, study investigators generated a population uniquely suited to answer questions about race and subtype differences in risk factors. To date, this is the largest study to evaluate breast cancer subtypes using a five-marker panel and one of the largest population-based studies of breast cancer in African-Americans.

This analysis of previously established breast cancer susceptibility loci provides strong evidence of etiologic heterogeneity across breast cancer subtypes. Though likely only a small part of the carcinogenic process, the risk variants identified here offer valuable clues about the nature of these diverse pathways. In turn, this vital information may help to advance disease prevention and control efforts.

Abbreviations

CBCS

Carolina Breast Cancer Study

CK 5/6

Cytokeratin 5/6

ER

Estrogen Receptor

HER1

Human Epidermal Growth Factor Receptor-1

HER2

Human Epidermal Growth Factor Receptor-2

HWE

Hardy-Weinberg Equilibrium

IHC

Immunohistochemical

MLE

Maximum Likelihood Estimate

OR

Odds Ratio

PI

Posterior Interval

PR

Progesterone Receptor

RAF

Risk allele frequency

SNP

Single Nucleotide Polymorphism

UNC

University of North Carolina