Previous GWAS Hits in Relation to Young-onset Breast Cancer

Abstract

Purpose

Genome-wide association studies (GWAS) have identified dozens of single nucleotide polymorphisms (SNPs) associated with breast cancer. Few studies focused on young-onset breast cancer, which exhibits etiologic and tumor-type differences from older-onset disease. Possible confounding by prenatal effects of the maternal genome has also not been considered.

Methods

Using a family-based design for breast cancer before age 50, we assessed the relationship between breast cancer and 77 GWAS-identified breast cancer risk SNPs. We estimated relative risks (RR) for inherited and maternally-mediated genetic effects. We also used published RR estimates to calculate genetic risk scores and model joint effects.

Results

Seventeen of the candidate SNPs were nominally associated with young-onset breast cancer in our 1,296 non-Hispanic white affected families (uncorrected p-value<0.05). Top-ranked SNPs included rs3803662-A (TOX3, RR=1.39; p=7.0×10⁻⁶), rs12662670-G (ESR1, RR=1.56; p=5.7×10⁻⁴), rs2981579-A (FGFR2, RR=1.24; p=0.002), and rs999737-G (RAD51B, RR=1.37; p=0.003). No maternally-mediated effects were found. A risk score based on all 77 SNPs indicated that their overall relationship to young-onset breast cancer risk was more than additive (additive-fit p=2.2×10⁻⁷) and consistent with a multiplicative joint effect (multiplicative-fit p=0.27). With the multiplicative formulation, the case sister’s genetic risk score exceeded that of her unaffected sister in 59% of families.

Conclusions

The results of this family-based study indicate that no effects of previously-identified risk SNPs were explained by prenatal effects of maternal variants. Many of the known breast cancer risk variants were associated with young-onset breast cancer, with evidence that TOX3, ESR1, FGFR2, and RAD51B are important for young-onset disease.

Introduction

In the search for genetic risk variants associated with breast cancer, dozens of genome-wide association studies (GWAS) have identified more than a hundred single nucleotide polymorphisms (SNPs) [1–33]. Many of these susceptibility loci have been confirmed in other study samples and across racial groups [34–39], but relatively few studies were designed to examine whether these GWAS ‘hits’ are associated with young-onset breast cancer [1, 16, 23, 40–43]. Given evidence that age modifies the association between breast cancer and some non-genetic risk factors [44–47], and that younger cases are more likely to have strong family histories of the disease [48], additional studies of the genetic determinants of young-onset breast cancer are warranted.

Young-onset breast cancer is often defined as breast cancer before the age of 50, as this age is both a pro×y for menopausal status and an inflection point for incidence trends in US women [44]. Young-onset disease tends to be relatively more aggressive and difficult to treat [49, 50], and includes a higher fraction of triple-negative breast cancers [51], which may have distinct genetic risk factors [52–56].

An early GWAS of young-onset breast cancer identified one risk-associated SNP in GLG1 [16]. In a larger and more recent GWAS of young-onset breast cancer, Ahsan et al. [1] identified 96 SNPs with genome-wide significant p-values, all of which were located in six regions previously linked to breast cancer risk unrestricted by age. In total, the authors found that 32 of 83 previously identified GWAS hits were associated with young-onset disease (p<0.05).

Family-based designs are feasible for diseases that occur at young ages and offer certain advantages over case-control studies, including robustness to bias due to population stratification and the ability to assess maternally-mediated or imprinting effects [57, 58]. We previously conducted a family-based GWAS of young-onset breast cancer, in which we identified 9 SNPs with unadjusted p-values <10⁻⁵, including several novel loci [59].

Here, we further investigate the role of 77 known risk variants by reporting their individual associations with young-onset breast cancer and exploring whether their joint effects follow additive or multiplicative risk models. The selected risk variants were previously identified by Mavaddat et al. [60] as having an association with breast cancer at p<5×10⁻⁸. We also examined maternally-mediated effects [58] for these same 77 SNPs to assess the influence of the mother’s genotype acting prenatally on her daughter, controlling for the daughter’s genotype. Because mother’s and daughter’s genotypes are correlated, maternally-mediated genetic effects could confound inherited-gene effects and may have biased some of the previously-observed GWAS associations.

Methods

Study participants

Young-onset breast cancer cases and their families were recruited as part of the Sister Study and Two Sister Study. The Sister Study is a prospective cohort of women who had one or more sisters diagnosed with breast cancer, but had never had breast cancer themselves at enrollment (2003–2009). It includes 50,884 US and Puerto Rican women aged 35–74. Because they all have a first-degree family history of breast cancer, participants have, on average, approximately twice the risk of developing breast cancer compared to women with no family history. Sister Study participants who developed breast cancer before age 50 (n=235) were included as cases for this analysis.

When we enrolled women in the Sister Study, we asked them the age and date of diagnosis of their affected sister. Proband sisters whose breast cancer was diagnosed before age 50 and within the previous four years were eligible for inclusion in the Two Sister Study. We developed a family-based genetic study by asking an unaffected sister to forward a study invitation to her eligible affected sister and then asking all of the participating young-onset cases to forward a letter from us to any living parents, asking them to provide saliva samples. If one or both parents were unavailable, we genotyped DNA from the blood or saliva from unaffected sister(s) participating in the Sister Study. In total, 3,331 individuals from 1,477 families were genotyped.

All participants provided written or verbal consent and the study was approved by the National Institute of Environmental Health Sciences and the Copernicus Group Institutional Review Boards. The GWAS data is publically accessible (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000678.v1.p1). None of the Sister Study or Two Sister Study participants were included in any previously published pooled or meta-analyses.

In addition to providing DNA samples, participants completed computer-assisted telephone interviews about various health-related and lifestyle factors. All cases were asked about their breast cancer diagnosis and treatment and to authorize release of their medical records. Tumor information, including invasiveness and estrogen receptor (ER) status was extracted from the medical records for most cases (85–90%). We relied on self-report for the remainder after observing that Sister Study participants could recall their status with reasonable accuracy (positive predictive values of 99%, 64%, 99% and 84% for invasive, in situ, ER+ and ER− breast cancer, respectively). We also reviewed medical records for BRCA1/2 findings and asked participants to report the results of any BRCA1/2 mutation tests. A case was assumed to be BRCA1/2 positive if she had a positive test or if her sister had a positive test, but the case had not been tested.

Genotyping analysis

DNA samples were collected from saliva (80%), whole blood (19%) or blood clot (1%) and shipped to the Johns Hopkins University Center for Inherited Disease Research for genotyping, with subsequent quality control carried out by the Genetics Coordinating Center at the University of Washington. After extraction, DNA samples were processed using 96-well plates, with family members assigned to the same plate. We included 76 HapMap controls and 74 duplicate samples, balanced across plates. All samples were analyzed using the Illumina OmniExpress plus HumanExome-8v1-2 array, which included 964,193 SNPs.

Genotyping revealed that 11 of the sister pairs were half-sisters. The 11 unaffected half-sisters were excluded. Because numbers were inadequate for sufficiently well-parameterized analyses of minority categories, we limited analyses to the majority category of 1,296 non-Hispanic white families (Table 1). No additional participants were excluded, as individuals’ missing call rates were all <2%, and duplicate discordance and Mendelian inconsistency rates were low (7.9×10⁻⁶ per SNP per duplicate pair, and 0.0003 per SNP, respectively). Individuals with chromosomal anomalies (n=6) were assigned missing values for affected regions.

Table 1.

Non-Hispanic white participants included in the young-onset breast cancer genotyping analysis.

Group Description	Number of Families	Number of Cases
Trios (affected sister, both parents)	416	416
Sister-pairs (1 affected, 1 unaffected), father	81	81
Sister-pairs	353	353
Affected sister and 1 parent	321	321
Unaffected sister and father	2	0
Parents only	2	0
Affected sister only	108	108
Unaffected sister only	11	0
Mother only	2	0
TOTAL	1296	1279

Forty of the 77 candidate SNPs were directly genotyped on our arrays. All met the inclusion criteria in non-Hispanic whites: call rates ≥97%, ≤1 discordant call in 74 study duplicates, ≤5 Mendelian errors, Hardy-Weinberg equilibrium p-values ≥ 1×10⁻⁶ among founders. Two SNPs with a minor allele frequency <1% were excluded from our assessment of individual SNP effects, but were retained in the genetic risk score analysis.

Imputation analysis

Imputation analyses were conducted at the University of Washington’s Genetics Coordinating Center using data from the 1000 Genomes Project [61]. Haplotypes were pre-phased using SHAPEIT2 [62] to improve efficiency and make use of known family structures. Imputation was then conducted using IMPUTE (v2.3.0) [63]. The imputation was highly accurate, with good concordance between measured and masked but imputed as most-probable genotypes (99.7% and 98.0% when minor allele frequencies were <5% and ≥5%, respectively).

In total, 22 million loci were imputed, including 37 of the non-genotyped SNPs from our list of 77. We imputed based on the most probable imputed genotype. If none of an individual’s estimated genotype probabilities exceeded 90%, the genotype for that locus was considered missing.

Statistical analysis

Individual SNP effects

For each of the 75 typed candidate SNPs (38 genotyped, 37 imputed), we recorded the odds ratio (OR) reported in Mavaddat et al. [60] and noted which allele was associated with increased breast cancer risk. We then tested the association between each of the candidate SNPs and young-onset breast cancer using a likelihood-based log-linear model to assess transmission distortion within families [58]. Briefly, this conditional model examines whether the relative frequencies of particular offspring genotypes at a di-allelic locus are consistent with Mendelian inheritance. Expected frequencies are modeled using a multinomial distribution with six mating type parameters, which impose conditioning on parental genotypes, and fifteen possible case-parent outcomes. Using likelihood ratio tests (χ² distribution), we tested the association between each SNP and young-onset breast cancer by comparing models with and without a term for the relative risk of the offspring’s genetic effect. We estimated the relative risks (RR) using Poisson regression. If either or both parents were unavailable, an expectation-maximization algorithm was applied to maximize the likelihood [64, 65]. Because we selected these SNPs based on evidence that they were associated with breast cancer, we did not correct p-values for multiple comparisons, but used α=0.05 to evaluate statistical significance.

To facilitate comparisons with Mavaddat et al. [60] and other prior studies, we estimated RRs for each SNP in relation to the previously-established risk allele. We assumed a log-additive model, coding genotype as the number of copies of the risk allele carried by the affected offspring. All analyses used LEM (http://members.home.nl/jeroenvermunt/#Software) and R (v3.2.1). We performed t-tests to compare the risk ratio observed in our study (representative of the young-onset breast cancer effect estimate) and the pooled odds ratio reported by Mavaddat et al. [60] (representative of the overall breast cancer effect estimate).

We also assessed the effects of each SNP in family subsets limited to invasive, ER+, or premenopausal breast cancer and in the subset of families not known to carry BRCA1/2 mutations. Small numbers precluded assessments for the complementary categories. We also estimated RRs for maternally-mediated effects [58].

As secondary analyses, we considered 89 other SNPs known to be associated with breast cancer risk. Additional details of the SNP selection process and results can be found in the Supplementary Materials.

Genetic Risk Scores

In addition to assessing their main effects, we also assessed the efficacy of our 77-SNP set by calculating both multiplicative and additive scores and comparing their model fits and predictive properties. The multiplicative genetic risk score was calculated as: ${\mathrm{\Sigma }}_{i=1}^k{x}_i\mathit{\text{ln}}(O{R}_i)$ where ORⁱ is the previously reported OR for one copy of the risk variant at SNP i and xⁱ is the number of copies of that risk variant carried by the individual. The summation is over k=77 candidate SNPs. Therefore, if 1) the effect size for each SNP in our study is equal to the previously-reported effect size, 2) the joint effect of more than one copy is multiplicative, and 3) the joint effect of multiple SNPs is multiplicative, then the risk score will equal the ln-OR for the 77-SNP set. Thus, the coefficient of ${\mathrm{\Sigma }}_{i=1}^k{x}_i\mathit{\text{ln}}(O{R}_i)$ is 1.0 under a multiplicative joint effects model. To the extent that the multiplicative genetic risk score and the ln-odds of developing young-onset breast cancer differ (as when the β coefficient differs from 1), there is a departure from multiplicity. The multiplicative joint effects model fit can be tested by comparing β to 1.0 using a Wald test (β-1 divided by standard error). One can also test whether the multiplicative genetic risk score is associated with young-onset breast cancer by comparing β to 0, again using a Wald test.

The additive genetic risk score was calculated as: $\mathit{\text{ln}}\left(\right({\mathrm{\Sigma }}_{i=1}^k{x}_i({\mathit{\text{OR}}}_{xi}-1))+1)$ Here, if 1) the effect size for each SNP in our study is equal to the previously-reported effect size, 2) the effect of having more than one copy of the risk allele is additive, and 3) the joint effect of multiple SNPs is additive, then that risk score equals the ln-OR for young-onset breast cancer. Thus, the true coefficient of $\mathit{\text{ln}}\left(\right({\mathrm{\Sigma }}_{i=1}^k{x}_i({\mathit{\text{OR}}}_{xi}-1))+1)$ is 1.0 under a fully additive joint effects model. As with the multiplicative model, the additive model fit can be tested by comparing β to 1.0 using a Wald test, and the association between the additive risk score and young-onset breast cancer can be tested by comparing β to 0.

We calculated the effect estimate (β) per unit increase in risk score using a conditional logistic regression model to compare cases to unaffected sister controls. When both sisters were genotyped, we compared them directly. If the unaffected sister was not genotyped but both parents were, we compared cases to an equally-likely, complementary pseudo-sister whose genotype was defined by her parents’ non-transmitted alleles. Altogether, we included 850 case-sister or case-pseudo-sister pairs in the risk scores analyses. For sporadic missing genotypes, we filled in the expected risk allele counts based on the allele frequencies in the parents. As both sets of risk score models were able to accommodate genotypes for imputed SNPs that were not whole numbers, we assigned individual genotypes for imputed SNPs based on expected allele counts.

The relative risk estimates for the 77 SNPs included in the risk score were taken from Mavaddat et al. [60]. These estimates were calculated using data from a pooled analysis with 33,673 cases. None of the 77 selected SNPs was in high linkage disequilibrium (LD) with another included SNP (all r²<0.80) [60].

To assess model fit and risk score utility, we calculated four p-values for each SNP set: an additive score testing β=0; an additive-fit score testing β=1; a multiplicative score testing β=0; and a multiplicative-fit score testing β=1. Additionally, we compared scores between sister pairs to assess their ability to predict risk. We also tested for maternally-mediated genetic effects by comparing scores of mothers to scores of fathers in families where both parents were genotyped (n=418).

As a natural extension and a means to further examine risk score utility, we also used logistic regression to examine whether the mothers’ scores were associated with their own breast cancer risk. Cases included 119 mothers with breast cancer at any age (cases) versus 599 mothers with no history of breast cancer (controls).

Results

At diagnosis, most of the 1,279 cases from the 1,296 non-Hispanic white families included in our analysis were aged 40–49 (89%) and premenopausal (93%) (Table 2). The proportions of invasive, ER+ and recognized BRCA1/2 mutation positive cases were 84%, 81% and 8%, respectively.

Table 2.

Characteristics of young-onset breast cancer casesincluded in the genotyping analysis (n=1279).

		N (%)
Age at diagnosis
<40		136 (11)
40–49		1143 (89)
Menopausal Status at Diagnosis
Premenopausal		1186 (93)
Postmenopausal		86 (7)
Missing		7
Invasive Status
Ductal carcinoma in situ		206 (16)
Invasive		1057 (84)
Missing		16
Estrogen Receptor Status
Positive		1013 (81)
Negative		239 (19)
Missing		27
BRCA1/2 statusa
Case carries BRCA1/2 mutation		100 (8)
Case not known to have		1179 (92)
BRCA1/2 mutation

^aFamilies were categorized as BRCA1 or BRCA2 mutation positive if (1) the case sister reported that she had had a positive test or (2) the case sister was not tested but the unaffected sister reported that she had had a positive test.

Seventeen of the candidate SNPs were associated with increased risk of young-onset breast cancer (p-value<0.05; Table 3). As demonstrated by the quantile-quantile plot (Figure 1), the distribution of p-values for the candidate SNPs was markedly shifted relative to the uniform distribution expected under a global null for the set. The SNP with the smallest p-value was rs3803662-A, which is located upstream of TOX3 (RR=1.39, 95% CI:1.20–1.60; p=7.0×10⁻⁶). The next smallest p-values were for rs12662670-G in ESR1 (RR=1.56, 95% CI: 1.20–2.03; p=5.7×10⁻⁴), rs2981579-A in FGFR2 (RR=1.24, 95% CI: 1.08–1.42; p=0.002), and rs999737-G in RAD51B (RR=1.27, 95% CI: 1.09–1.48; p=0.003).

Table 3.

Relative risks and 95% confidence intervals for the association between candidate SNPs and young-onset breast cancer

Rank	SNPa	Gene/ Region	Risk Allele	RAF	RR (95% CI)	p-value	Reported ORb
1	rs3803662	TOX3	A	0.30	1.39 (1.20, 1.60)	7.0×10−6	1.23
2	rs12662670	ESR1	G	0.08	1.56 (1.20, 2.03)	5.7×10−4	1.14
3	rs2981579	FGFR2	A	0.44	1.24 (1.08, 1.42)	0.002	1.25
4	rs999737	RAD51B	G	0.77	1.27 (1.09, 1.48)	0.003	1.09
5	rs2046210	ESR1	A	0.35	1.21 (1.05, 1.39)	0.008	1.05
6	rs4849887	2q14.2	C	0.89	1.31 (1.07, 1.60)	0.009	1.09
7	rs7726159	TERT	A	0.32	1.28 (1.05, 1.55)	0.01	1.04
8	rs720475	ARHGEF5	G	0.75	1.20 (1.04, 1.39)	0.01	1.06
9	rs11199914	10q26.12	C	0.69	1.20 (1.04, 1.38)	0.02	1.06
10	rs554219	CCND1	G	0.13	1.27 (1.04, 1.55)	0.02	1.12
11	rs6001930	MKL1	G	0.10	1.29 (1.04, 1.59)	0.02	1.13
12	rs10759243	9q31.2	A	0.27	1.21 (1.03, 1.41)	0.02	1.05
13	rs10941679	MRPS30	G	0.25	1.20 (1.02, 1.40)	0.03	1.12
14	rs13281615	8q24.21	G	0.42	1.16 (1.01, 1.32)	0.04	1.10
15	rs616488	PEX14	A	0.66	1.16 (1.01, 1.33)	0.04	1.06
16	rs11780156	8q24.21	T	0.18	1.19 (1.01, 1.41)	0.04	1.07
17	rs6828523	ADAM29	C	0.90	1.24 (1.00, 1.54)	0.05	1.10
18	rs2943559	HNF4G	G	0.08	1.24 (0.98, 1.57)	0.07	1.13
19	rs12493607	TGFBR2	C	0.34	1.14 (0.99, 1.32)	0.07	1.05
20	rs4973768	SLC4A7	A	0.50	1.13 (0.99, 1.28)	0.07	1.09
21	rs75915166	FGF3	A	0.06	1.29 (0.97, 1.72)	0.07	1.02
22	rs17356907	12q22	A	0.73	1.14 (0.98, 1.32)	0.08	1.10
23	rs2736108	TERT	T	0.73	1.17 (0.98, 1.39)	0.09	1.07
24	rs7072776	MLLT10	A	0.31	0.88 (0.77, 1.02)	0.09	1.06
25	rs3903072	11q13.1	G	0.53	1.12 (0.98, 1.27)	0.09	1.06
26	rs941764	CCDC88C	G	0.35	1.12 (0.98, 1.29)	0.10	1.06
27	rs10472076	RPL5P15	G	0.36	1.11 (0.97, 1.27)	0.13	1.04
28	rs1292011	12q24.21	A	0.60	1.11 (0.97, 1.26)	0.13	1.08
29	rs1011970	CDKN2B	A	0.17	1.14 (0.96, 1.36)	0.13	1.05
30	rs11242675	FOXQ1	A	0.63	1.10 (0.96, 1.26)	0.19	1.06
31	rs78540526	CCND1	T	0.08	1.19 (0.92, 1.55)	0.19	1.18
32	rs1045485	CASP8	C	0.88	1.14 (0.93, 1.39)	0.20	1.04
33	rs12710696	LINC01376	T	0.37	1.09 (0.95, 1.24)	0.20	1.04
34	rs9693444	8p12	A	0.44	1.09 (0.95, 1.25)	0.24	1.07
35	rs11075995	FTO	A	0.24	0.92 (0.79, 1.06)	0.25	1.04
36	rs6762644	ITPR1	G	0.39	0.93 (0.82, 1.06)	0.30	1.07
37	rs527616	18q11.2	G	0.68	1.09 (0.92, 1.29)	0.31	1.04
38	rs17817449	FTO	A	0.61	1.07 (0.94, 1.22)	0.33	1.08
39	rs4245739	MDM4	C	0.27	1.08 (0.93, 1.25)	0.33	1.03
40	rs889312	MAP3K1	C	0.31	1.07 (0.93, 1.23)	0.34	1.12
41	rs1436904	CHST9	A	0.60	1.06 (0.93, 1.21)	0.36	1.06
42	rs11249433	EMBP1	G	0.42	1.06 (0.93, 1.22)	0.37	1.10
43	rs704010	ZMIZ1	A	0.41	1.06 (0.93, 1.21)	0.39	1.07
44	rs7904519	TCF7L2	G	0.47	0.94 (0.82, 1.08)	0.39	1.06
45	rs2236007	PAX9	G	0.88	1.07 (0.92, 1.25)	0.40	1.09
46	rs6472903	8q21.11	A	0.83	1.08 (0.91, 1.28)	0.40	1.10
47	rs11820646	11q24.3	C	0.61	0.95 (0.82, 1.09)	0.44	1.05
48	rs2363956	ANKLE1	A	0.51	0.95 (0.83, 1.08)	0.44	1.03
49	rs3817198	LSP1	G	0.34	1.05 (0.92, 1.21)	0.45	1.07
50	rs2016394	DLX2	G	0.53	1.05 (0.92, 1.20)	0.45	1.05
51	rs865686	RPL31P43	A	0.63	1.05 (0.92, 1.21)	0.46	1.11
52	rs17529111	6q14.1	G	0.22	1.06 (0.90, 1.26)	0.48	1.05
53	rs2823093	CYYR1	G	0.73	1.05 (0.91, 1.21)	0.49	1.08
54	rs132390	EMID1	C	0.02	1.16 (0.74, 1.81)	0.52	1.11
55	rs10771399	12p11.22	G	0.89	1.07 (0.87, 1.31)	0.54	1.16
56	rs9790517	TET2	A	0.24	1.05 (0.90, 1.22)	0.55	1.05
57	rs6504950	STXBP4	G	0.73	1.05 (0.90, 1.22)	0.55	1.07
58	rs3760982	KCNN4	A	0.47	1.04 (0.91, 1.19)	0.57	1.06
59	rs13387042	TNP1	G	0.54	1.04 (0.91, 1.18)	0.57	1.14
60	rs12422552	12p13.1	C	0.28	0.97 (0.83, 1.12)	0.64	1.03
61	rs6678914	LGR6	G	0.59	0.97 (0.85, 1.11)	0.65	1.01
62	rs11552449	AP4B1	A	0.15	0.96 (0.81, 1.14)	0.65	1.08
63	rs1550623	CDCA7	A	0.85	1.04 (0.87, 1.24)	0.66	1.06
64	rs10069690	TERT	A	0.27	1.03 (0.89, 1.19)	0.67	1.02
65	rs11814448	10p12.31	C	0.02	0.92 (0.60, 1.39)	0.68	1.22
66	rs204247	RANBP9	G	0.45	1.03 (0.90, 1.17)	0.68	1.05
67	rs13329835	CDYL2	G	0.23	0.97 (0.83, 1.14)	0.72	1.08
68	rs4808801	ELL	A	0.68	1.02 (0.88, 1.17)	0.80	1.07
69	rs2588809	RAD51B	T	0.17	0.98 (0.82, 1.16)	0.80	1.07
70	rs1432679	EBF1	C	0.46	0.99 (0.86, 1.13)	0.84	1.07
71	rs10995190	ZNF365	G	0.86	1.02 (0.85, 1.23)	0.84	1.17
72	rs16857609	DIRC3	T	0.28	0.99 (0.85, 1.15)	0.87	1.07
73	rs2380205	GDI2	G	0.58	0.99 (0.87, 1.13)	0.88	1.02
74	rs8170	BIBAM1	A	0.17	0.99 (0.83, 1.17)	0.88	1.03
75	rs1353747	PDE4D	T	0.91	1.01 (0.81, 1.25)	0.95	1.09

Abbreviations: CI= Confidence Interval; RR = risk ratio; NR = not reported; RAF= Risk allele frequency

^aImputed SNPs are italicized

^bOdds ratio for the log-additive model reported in Mavaddat et al. [60] Original ORs<1.00 were transformed to reflect the measured effect of the risk allele.

Figure 1.

Quantile-quantile plot for the association between 75 candidate SNPs and young-onset breast cancer in the Two Sister Study (2008–2012).

The remaining 13 SNPs with statistically-significant associations are located in the following loci: ESR1, 2q14.2, TERT, ARHGEF5, 10q26.12, CCND1, MKL1, 9q31.2, MRPS30, 8q24.21, PEX14 and ADAM29. The ESR1 and 8q24.21 regions both had two statistically-significant SNPs which were not in LD with one another. When we compared our observed effect estimates for young-onset breast cancer to those observed for overall breast cancer [60], the distribution of p-values for those tests did not deviate from the expected distribution under the null of no difference (Supplementary Figure S1).

Analyses restricted to families with invasive, ER+, premenopausal cancer or not known to carry risk-related BRCA1/2 mutations showed similar patterns, with small p-values again seen for rs3803662, rs12662670, rs2981579, rs999737, and other top-ranked SNPs from the overall analysis (Supplementary Tables S1–S4). For maternally-mediated effects, we found only five SNPs associated (p<0.05) with young-onset breast cancer, a number compatible with random chance (73 included SNPs, Table 4 and Figure 2). None of those SNPs were also identified as significant in our primary analysis.

Table 4.

Effect estimates for maternally-mediated genetic effects of previous breast cancer genome-wide association study risk variants on the risk of young-onset breast cancer

Rank	SNPa	Gene	RR (95% CI)	p-value
1	rs9790517	TET2	0.78 (0.66, 0.94)	0.007
2	rs10995190	ZNF365	0.75 (0.61, 0.93)	0.008
3	rs9693444	8p12	0.81 (0.70, 0.93)	0.008
4	rs2236007	PAX9	1.21 (1.01, 1.45)	0.03
5	rs2016394	DLX2	0.85 (0.73, 0.99)	0.04
6	rs2588809	RAD51B	0.84 (0.69, 1.02)	0.07
7	rs2943559	HNF4G	0.80 (0.62, 1.04)	0.09
8	rs6504950	STXBP4	1.15 (0.97, 1.35)	0.10
9	rs6828523	ADAM29	1.19 (0.94, 1.49)	0.15
10	rs2981579	FGFR2	0.90 (0.78, 1.04)	0.15
11	rs10771399	12p11.22	0.84 (0.66, 1.08)	0.17
12	rs720475	ARHGEF5	0.89 (0.75, 1.05)	0.18
13	rs8170	BIBAM1	0.87 (0.71, 1.07)	0.18
14	rs2046210	ESR1	1.10 (0.95, 1.28)	0.20
15	rs12422552	12p13.1	0.90 (0.76, 1.06)	0.21
16	rs4245739	MDM4	1.11 (0.94, 1.32)	0.22
17	rs1045485	CASP8	0.87 (0.70, 1.09)	0.23
18	rs7904519	TCF7L2	0.92 (0.79, 1.06)	0.25
19	rs3817198	LSP1	0.91 (0.77, 1.07)	0.26
20	rs865686	RPL31P43	1.09 (0.93, 1.27)	0.28
21	rs13281615	8q24.21	1.09 (0.93, 1.27)	0.28
22	rs3903072	11q13.1	1.09 (0.93, 1.27)	0.29
23	rs11242675	FOXQ1	1.09 (0.93, 1.27)	0.30
24	rs10759243	9q31.2	0.91 (0.76, 1.09)	0.31
25	rs11820646	11q24.3	0.93 (0.80, 1.07)	0.31
26	rs12662670	ESR1	0.88 (0.67, 1.15)	0.34
27	rs2380205	GDI2	0.93 (0.79, 1.08)	0.34
28	rs13387042	TNP1	1.07 (0.92, 1.25)	0.36
29	rs999737	RAD51B	0.93 (0.78, 1.10)	0.38
30	rs4849887	2q14.2	0.92 (0.72, 1.16)	0.47
31	rs17356907	12q22	0.94 (0.80, 1.12)	0.49
32	rs13329835	CDYL2	1.06 (0.88, 1.28)	0.52
33	rs6762644	ITPR1	0.95 (0.82, 1.11)	0.53
34	rs204247	RANBP9	0.95 (0.82, 1.11)	0.54
35	rs1292011	12q24.21	0.95 (0.81, 1.12)	0.55
36	rs616488	PEX14	0.95 (0.82, 1.11)	0.55
37	rs4973768	SLC4A7	0.95 (0.82, 1.11)	0.55
38	rs10069690	TERT	0.95 (0.80, 1.13)	0.56
39	rs11814448	10p12.31	0.90 (0.57, 1.41)	0.64
40	rs10941679	MRPS30	0.96 (0.80, 1.15)	0.65
41	rs16857609	DIRC3	0.96 (0.81, 1.14)	0.66
42	rs1550623	CDCA7	0.96 (0.78, 1.18)	0.67
43	rs2736108	TERT	0.96 (0.79, 1.17)	0.67
44	rs6678914	LGR6	1.03 (0.88, 1.21)	0.68
45	rs527616	18q11.2	1.04 (0.87, 1.25)	0.68
46	rs12493607	TGFBR2	0.97 (0.83, 1.13)	0.71
47	rs704010	ZMIZ1	0.97 (0.83, 1.14)	0.71
48	rs78540526	CCND1	1.05 (0.78, 1.42)	0.73
49	rs3803662	TOX3	0.97 (0.82, 1.15)	0.74
50	rs889312	MAP3K1	0.97 (0.83, 1.15)	0.75
51	rs11075995	FTO	0.97 (0.82, 1.15)	0.76
52	rs75915166	FGF3	1.05 (0.77, 1.42)	0.76
53	rs1436904	CHST9	1.02 (0.87, 1.20)	0.77
54	rs1353747	PDE4D	0.97 (0.75, 1.24)	0.79
55	rs1011970	CDKN2B	0.97 (0.79, 1.19)	0.80
56	rs11199914	10q26.12	0.98 (0.83, 1.16)	0.81
57	rs2363956	ANKLE1	0.98 (0.85, 1.14)	0.81
58	rs4808801	ELL	0.98 (0.84, 1.15)	0.82
59	rs941764	CCDC88C	0.98 (0.84, 1.16)	0.83
60	rs1432679	EBF1	1.01 (0.87, 1.17)	0.89
61	rs11249433	EMBP1	0.99 (0.85, 1.15)	0.89
62	rs12710696	LINC01376	1.01 (0.86, 1.19)	0.90
63	rs17817449	FTO	0.99 (0.85, 1.16)	0.90
64	rs11552449	AP4B1	0.99 (0.81, 1.21)	0.92
65	rs7726159	TERT	0.99 (0.82, 1.20)	0.92
66	rs554219	CCND1	1.01 (0.82, 1.25)	0.92
67	rs10472076	RPL5P15	1.01 (0.87, 1.17)	0.93
68	rs132390	EMID1	1.02 (0.62, 1.69)	0.93
69	rs11780156	8q24.21	0.99 (0.81, 1.22)	0.95
70	rs2823093	CYRR1	1.00 (0.85, 1.17)	0.95
71	rs6001930	MKL1	1.01 (0.79, 1.29)	0.95
72	rs7072776	MLLT10	1.00 (0.85, 1.16)	0.96
73	rs3760982	KCNN4	1.00 (0.86, 1.16)	1.00

Abbreviations: CI= Confidence Interval; RR = risk ratio

^aImputed SNPs are italicized

Figure 2.

Quantile-quantile plot for maternally-mediated genetic effects of 73 candidate SNPs and young-onset breast cancer in the Two Sister Study (2008–2012).

For the polygenic risk score, both the multiplicative and additive genetic risk scores were associated with young-onset breast cancer (p=2.7×10⁻¹⁰ and p=1.3×10⁻¹⁰, respectively, testing β=0). The coefficient for the additive risk score was different from 1.0 (β= 5.1, additive-fit-p=2.2×10⁻⁷), but the coefficient for the multiplicative risk score was not (β=0.85, multiplicative-fit-p=0.27). This indicates that the overall joint effect of these risk SNPs was more than additive and consistent with multiplicative and that each one-unit increase in score corresponds to a relative risk increase of 2.34 (e^0.85). For both scores, the case sister had a higher score than her control/pseudo-control sister in 59% of pairs, suggesting there is within-family predictive power.

There was no evidence that these genetic effects act prenatally through the mother’s phenotype, as there was no difference between mothers’ and fathers’ genetic scores on either the multiplicative or additive scale (p=0.19 and p=0.18, for paired testing, respectively). When we tested the association between the polygenic risk score and breast cancer in mothers, the score preformed similarly to our original analysis, with associations seen for both multiplicative (p=4.6×10⁻⁷) and additive (p=4.0×10⁻⁷) models. The effects were again more than additive and consistent with multiplicative (multiplicative β=0.88, multiplicative-fit p=0.50; additive β=5.4, additive-fit p=3.5×10⁻⁵), with an estimated area under the receiver operating characteristics curve of 0.65 for both formulations.

Discussion

The purpose of this family-based genetic study was to examine whether previously-identified breast cancer risk variants were associated with young-onset disease through inherited or maternally-mediated effects. We focused on 77 common polymorphisms identified in published GWAS [60]. Seventeen of those candidate SNPs were nominally associated with young-onset breast cancer and a genetic risk score consisting of all 77 SNPs was related to the risk of young-onset breast cancer in daughters and to a history of breast cancer at any age in their mothers. There was no evidence that the independent or combined effects of these candidate SNPs was due to a prenatal effect mediated by the maternal genome.

The SNPs with the lowest p-values were located in well-studied susceptibility genes – TOX3, ESR1, FGFR2, and RAD51B. The results were similar when we restricted analyses to include only families with invasive, ER+, or premenopausal young-onset breast cancer, or to families without known deleterious BRCA1/2 variants.

TOX3 encodes a nuclear protein that regulates calcium-dependent transcription in neurons [66]. Its link to breast cancer is unclear, though there is some evidence that it is a tumor suppressor [66, 67]. Our top hit, rs3803662, was first identified by Easton et al., [7], with several subsequent GWAS confirming the region’s importance [11, 18, 22, 23, 30–32]. Presence of this variant allele has been linked to lower TOX3 expression levels in breast tumor tissue [68]. Several candidate gene studies also reported risks of similar magnitude for rs3803662-T and young-onset breast cancer [1, 38, 40, 43, 69]. In Ahsan et al.’s [1] young-onset GWAS, TOX3 was one of the identified susceptibility regions, and rs3803662 had the smallest p-value for that region. For a more direct comparison of our results to those of Ahsan et al., see Supplementary Table S5 (a comparison of p-values for the 83 previously-established susceptibility loci selected by Ahsan et al. and our results for the same SNPs) and Supplementary Table S6 (an assessment of other established susceptibility loci not included in Mavaddat et al. [60], including the novel GWAS hits from Ahsan et al.).

The SNPs with the second and fifth smallest p-values - rs12662670 and rs2046210 - are located in the 6q25 region just upstream of ESR1. ESR1 encodes Estrogen Receptor α, a ligand-activated transcription factor crucial to sexual development and reproduction [70]. rs12662670 was previously linked to increased risk of triple-negative breast cancer, a subtype more common in younger women [71]. We did not have enough cases of triple-negative cancer for a separate analysis. Carriers of the other SNP, rs2046210-A, have lower ESR1 levels in both tumor and normal breast tissue [72]. The 6q25 region was not one of the regions with genome-wide significant p-values in Ahsan et al. [1], though rs2046210-A was positively associated with young-onset breast cancer at p<0.05.

FGFR2 encodes a fibroblast growth factor receptor, the overexpression of which may contribute to carcinogenesis through increased cell proliferation, migration, and resistance to apoptosis [73]. Our FGFR2 SNP, rs2981579, was first identified by Thomas et al. [31] and independently replicated in two other GWAS [23, 32]. Other GWAS found stronger associations with other SNPs in the same gene [1, 7, 8, 10, 12, 15, 19, 22, 26]. Ahsan et al. [1], for example, reported genome-wide significant p-values for 4 FGFR2 SNPs and young-onset breast cancer, with the smallest p-value seen for rs2981579. Other studies have also observed associations between FGFR2 SNPs and young-onset breast cancer [41, 42, 69, 74].

RAD51B is a homologous recombination repair gene related to BRCA1 and BRCA2 [75]. Thomas et al. [31] were the first to link rs999737-G to breast cancer risk. This was confirmed in a second GWAS [23] and subsequent meta-analysis [76]. Although RAD51B was not one of the genome-wide significant regions identified in Ahsan et al. [1], rs999737 was associated with young-onset breast cancer at p<0.05.

In addition to TOX3, ESR1, FGFR2, and RAD51, we corroborated Ahsan et al.’s findings that SNPs in 8q24.21 and 11q13.3 (CCND1) may be important for young-onset breast cancer. However, we did not see statistically significant effects for SNPs in SLC4A7 or MAP3K1. It is not clear whether these differences are due to chance or heterogeneity across studies. We had limited power to detect small effects, with approximately 80% power to detect RRs of 1.24 or 1.20 for allele frequencies of 20% or 40%, respectively for α=0.05. This modest power limits our ability to detect small effects and age-at-onset interaction effects. Given this caveat, the results of the statistical tests comparing our effect estimates to those of Mavaddat et al. [60] showed no evidence that effect sizes differed by age at breast cancer onset.

Dite et al. [77] also applied the 77-SNP risk score developed by Mavaddat et al. [60] to evaluate young-onset breast cancer. They reported an area under the curve of 0.61 for the risk score alone, which is consistent with the predictive capability we report here. We additionally found that the 77-SNP risk score was predictive of breast cancer in the mothers of cases (which were predominantly older-onset), and that similar coefficients were seen for the two generations.

Our use of previously reported effect measures helped us avoid over-fitting the prediction model to our data [78] and enabled independent replication. Several previous studies have used similar approaches, but most have considered only models with multiplicative joint effects [18, 22, 60, 77, 79]. Like us, Joshi et al. [80] constructed scores for both multiplicative and additive polygenic effects using externally-reported effect estimates. We both found evidence that the joint effects were super-additive, but our results showed consistency with a multiplicative model, while Joshi et al. found that the observed associations were sub-multiplicative. Future studies should consider both additive and multiplicative models to see which mathematical form of the risk score best fits the data and to determine if the same model is appropriate for different subgroups (e.g. younger versus older women). Ideally, investigators will reach a consensus regarding what SNPs to include in the score and how to correctly specify their combined effects. Ultimately, we hope a risk score can be utilized as a tool to classify women’s individual breast cancer risk and identify who should be selected for more or less frequent screening.

A major strength of this study is its family-based design, which is robust to population stratification and amenable to missing-data imputation methods [64, 65]. However, we restricted our analysis to non-Hispanic whites, which somewhat limits the generalizability of our findings. There is also the potential for survival bias if some SNPs are related to survival after breast cancer. However, short-term survival rates for breast cancer are high (91% 5-year relative survival for invasive breast cancer before age 50 [81]), suggesting the effects of such attrition would be minimal.

In this family-based genetic study, we saw little evidence of maternally-mediated genetic effects, but found that many of the known breast cancer risk variants were also associated with young-onset breast cancer. Our analyses provide further evidence that certain loci, including TOX3, ESR1, FGFR2, and RAD51B, are important for the development of breast cancer at any age.

Abbreviations

CI

confidence interval

ER

estrogen receptor

GWAS

genome-wide association study

LD

linkage disequilibrium

OR

odds ratio

RR

relative risk

SNP

single nucleotide polymorphism