Common Genetic Variants Associated with Breast Cancer and Mammographic Density Measures That Predict Disease

Introduction

For women of equivalent age, those whose breasts display greater white or bright areas on a mammogram—that is, have greater mammographic density—are at greater risk of developing breast cancer (1). This association has been shown repeatedly by case-control studies nested within large population screening programs in which cases and controls have been matched for age and adjustments have been made for other measured breast cancer risk factors, in particular body mass index (BMI). The association has been shown to persist up to 8 years after initial screening (2) and remains after excluding cases detected within 1 year of the initial screen, indicating that these associations are not simply due to a “masking” effect caused by difficulty in detecting tumors that are in mammographically dense regions of the breast.

Twin and family studies have reported that, under the assumptions of the classic twin model and after adjusting for age, BMI, and other determinants, the patterns of familial correlations for mammographic density measures are consistent with additive genetic factors explaining ∼60% of their residual variances (3, 4). This raises the question of whether variants associated with breast cancer risk also explain the genetic component of variation in those mammographic density measures that predict breast cancer.

Recently, genome-wide and candidate gene association studies have identified a number of loci in which there exists one or more common single nucleotide polymorphisms associated with small gradients in breast cancer risk (5–7). Table 1 lists the variants and their corresponding per allele associations. For example, the minor allele of rs2981582, located in intron 2 of the gene FGFR2, is associated with about a 20% increase in breast cancer risk per allele. It is not known why these variants are associated with breast cancer risk as they are not necessarily “functional,” let alone causal. Note also that although a gene has been assigned to a locus-specific disease association, this does not imply a direct role on disease risk of the specific variant or even other variants in the nominated gene.

Finding that one or more breast cancer–associated common genetic variants are associated with mammographic density could help elucidate some of the biological reasons why women of the same age and BMI differ so much in mammographic density. In doing so, it could also help unravel why these variants are associated with breast cancer risk. To date, three studies (8–10) have examined this question but have not provided a convincing answer (see Discussion).

We genotyped a total of 2,288 monozygotic and dizygotic twins and their sisters for 12 variants, one from each locus associated with breast cancer. Taking into account family structure, we analyzed associations with mammographic measures cross-sectionally and between- and within-sibships to determine whether there was any evidence of association between the mammographic density measures and these variants overall, as well as for each variant individually.

Results

Subjects missing information on BMI, menopausal status, benign biopsy status, or any of the genotyped variants were excluded. Genotyping revealed that 35 of the 532 twin pairs originally reported to be monozygotic were dizygotic. Genotypes of duplicate samples were 100% identical. The final sample comprised 2,288 subjects: 497 monozygotic pairs, 330 dizygotic pairs, and 634 nontwin sisters. There was no evidence of deviation from Hardy-Weinberg equilibrium for any variant (all P > 0.15). The intraclass correlation coefficient was >0.90 within readers and 0.87 between readers for measurements of percent dense area. Similarly, the intraclass correlation coefficient was >0.82 within readers and 0.84 between readers for measurements of dense area.

Table 2 shows the characteristics of subjects. Although the majority (70%) were postmenopausal, there was a wide range in age at mammogram from 31 to 79 years and the SD was 8.38 years. There were also wide variations in the mammographic density measures, with the coefficients of variation (SD/mean) being 67%, 39%, and 28% for percent dense area, dense area, and nondense area, respectively. Table 3 shows the means for each mammographic measure and all three genotypes for each variant.

After adjusting both dense area and percent dense area for age and BMI, their correlation was 0.8. The monozygotic, dizygotic, and sister correlations (95% confidence intervals in parentheses) for the residual measures were 0.73 (0.69–0.77), 0.30 (0.20–0.30), and 0.17 (0.09–0.25) for dense area; 0.67 (0.63–0.71), 0.37 (0.27–0.47), and 0.22 (0.014–0.30) for percent dense area; and 0.64 (0.58–0.70), 0.20 (0.10–0.30), and 0.13 (0.05–0.21) for nondense area.

Table 4 shows the nominal P values for a priori hypothesis testing based on cross-sectional, between-sibships, within-sibships, and combined estimates. These were inconsistent with the uniform distribution for transformed and adjusted dense area (P_u = 0.0005 and 0.001 for combined and cross-sectional analyses, respectively) and for adjusted percent dense area (P_u = 0.007 for combined analyses), but not for transformed and adjusted nondense area (P_u = 0.7 and 0.8 for combined and cross-sectional analyses, respectively). That is, for each of the dense area and percent dense area measures, there is evidence that at least one variant is associated with these predictors of breast cancer risk.

For transformed and adjusted dense area, there were highly significant associations with the number of minor alleles of rs3817198/LSP1 and rs13281615/8q (P_c = 0.001 and 0.003, and P_x = 0.0002 and 0.05, respectively). There was marginal evidence of an association with the number of minor alleles of rs889312/MAP3K1 (P_c = 0.04), rs2107425/H19 (P_x = 0.03), and rs17468277/CASP8 (P_x = 0.04). Examination of Table 3 shows that the means increase or decrease monotonically, consistent with the fitted linear associations (except perhaps rs2107425/H19).

For each of the above five variants, the associations were in the same direction as their associations with breast cancer risk (see Table 1). The cross-sectional residual variance explained by these variants individually were 0.71%, 0.26%, 0.16%, 0.27%, and 0.24%, in the order discussed. When fitted together, they explained 1.5% of variance and the associations with rs2107425/H19 and rs889312/MAP3K1 and were no longer nominally significant. The variants rs3817198/LSP1 and rs13281615/8q together explained 1.0% of variance.

For adjusted percent dense area, there were associations with the number of minor alleles for both rs3817198/LSP1 and rs13281615/8q (all P_x and P_c <0.05). For rs889312/MAP3K1, the combined estimate was marginal (P_c = 0.03). All associations were in the same direction as for dense area, as described above. Examination of Table 3 shows that the means increase or decrease monotonically, consistent with the fitted linear associations (except perhaps rs13281615/8q).

For transformed and adjusted nondense area, the only association was for rs889312/MAP3K1, marginal, and from the combined between- and within-sibship analyses only (P_c = 0.04). The negative within-pair association with rs6504950 (P_w < 0.0001) was in the opposite direction to that predicted by the association between mammographic density measures and breast cancer, and to the marginal between-pair association (P_b = 0.02).

None of the associations above differed by menopausal status (all P > 0.1).

Discussion

Twin and family studies have shown that genetic factors could explain a considerable proportion of variance in the mammographic density measures that predict breast cancer risk; dense area and percent dense area both adjusted for age, BMI, and determinants. Based on a theoretical model of multiplicative risk (17) and the assumptions of the classic twin model, it has been estimated that mammographic density genes might explain ∼10% of the familial aggregation of breast cancer (3, 4). In this article, we found that some of the common breast cancer genetic susceptibility variants recently identified and validated by large-scale case-control studies are also associated with the mammographic density measures that predict breast cancer risk, in particular age- and BMI-adjusted dense area. There was overwhelming evidence that there were too many small P values than would be expected under the null hypothesis that none of the variants are associated with these mammographic density measures (Table 4).

We found strong evidence that two variants, one on chromosome 11p in the region of LSP1 and another on chromosome 8, are associated with the mammographic density measures that predict disease. There was some marginal evidence that another 3 of the 12 established breast cancer–associated common genetic variants might be associated with dense area. These common variants explained 1.5% of residual variance of dense area. Given that all these variants are prespecified candidates, and that the direction of their mammographic density associations are consistent with their breast cancer associations, they warrant genotyping by other large studies to determine more definitively if they are associated with this heritable breast cancer risk factor. The nominally significant associations were in general more apparent for dense area than percent dense area, and the test of uniformity was more significant for dense area.

The variants rs2107425 (H19) and rs3817198 (LSP1) are at the same locus (only 100 kb apart) and the association of rs2107425 (H19) with breast cancer risk was originally posited to be due to its linkage disequilibrium with rs3817198 (LSP1; ref. 6). We found that the cross-sectional association of rs2107425 (H19) with dense area was no longer evident after adjustment for rs3817198 (LSP1). Moreover, within sibships, there was evidence of an association with rs3817198 (LSP1; P_w = 0.02) but not with rs2107425 (H19; P_w = 0.9). This study of a disease-associated genetic marker has provided information that could help tease apart the roles of common variants in linkage disequilibrium on the disease itself.

Lee and colleagues (8) genotyped 516 breast cancer cases for six variants [rs889312 (MAP3K1), rs2981582 (FGFR2), rs3803662 (TOX9), rs3817198 (LSP1), rs13281615 (8q), and rs13387042 (2q)] and found no evidence that any was associated with percent dense area. From subanalyses, they reported marginal evidence that rs3817198 (LSP1) was positively associated with percent dense area when analyzing 221 steroid receptor–positive cases only. Tamimi and colleagues (9) genotyped 1,121 women for 11 variants [rs889312 (MAP3K1), rs2981582 (FGFR2), rs12443621 (TOX9), rs3803662 (TOX9), rs3817198 (LSP1), rs13281615 (8q), rs4666451, rs2107425 (H19), rs981782 (HCN1), rs8051542 (TOX3), and rs30099 (5)] and also found no evidence that any was associated with percent dense area. From subanalyses, they reported marginal evidence of associations with rs12443621 (TOX3), rs3817198 (LSP1), and rs4666451 when analyzing premenopausal women only. Woolcott and colleagues (10) genotyped 825 women (361 with breast cancer) for eight variants [rs2981582 (FGFR2), rs3803662 (TOX3), rs12443621 (TOX3), rs3817198 in LSP1, rs981782 (HCN1/MRPS30), rs10941679 (HCN1/MRPS30), rs889312 (MAP3K1), and rs13387042 (2q)] and found marginal evidence that percent dense area was positively associated with rs12443621/TOX3 and negatively associated with rs10941679 near HCN1/MRPS30. Subanalyses showed that rs3817198/LSP1 was associated with percent dense area when analyzing current users of combined hormone replacement therapy only. These nominally significant subgroup associations with rs3817198/LSP1 support our overall finding.

Based on our cross-sectional findings for percent dense area, for which the strongest associations were ∼1% per allele with a SE of ∼0.5%, and that our effective sample size was ∼2,000 after taking into account the nonindependence of some subjects, a cross-sectional study of 1,000 women would have 30% power to detect the associations we found. Therefore, if our results represent reality, none of the previous studies had more than a 50:50 chance of finding a nominally significant association. It would be of interest to pool raw data from published studies.

The strengths of our study include its size, although larger studies are warranted especially for the variants with equivocal findings, and averaging measurements taken by trained observers with demonstrated high repeatability. Collecting, and analyzing as we have, data from twins and their sisters was another strength. The within-sibship associations are naturally adjusted not only for age and other measured covariates but also for unmeasured factors shared by twins and sisters. The between-sibship analyses summarize data from groups of similar ages and familial, including genetic, background. Finally, the combined test statistic for association makes use of the independence of the between- and within-sibship estimates to gain more power to test the null hypothesis of no association.

Little is known about how the variants we have studied convey susceptibility to breast cancer. They could be in linkage disequilibrium with other, perhaps rarer and as yet unidentified, “causal” variants. These findings suggest that whatever the underlying mechanisms, for at least two loci they are also likely associated with mammographically dense area. This lead could help elucidate why those variants are associated with breast cancer risk.

In addition, unraveling the mechanisms behind the associations of variants with breast cancer risk could help understand what it is about mammographically dense area that makes it a risk factor for breast cancer. Understanding the function of the genes, and in particular the change in function due to the causal variant(s), could lead to new insights about mechanisms that cause breast cancer. Furthermore, the discovery of these causal variant(s) could be aided by studying the continuously distributed risk factor (mammographic density) rather than the disease itself. That is, understanding the genetic link between mammographic density and breast cancer risk could help unravel the etiology of the disease.

Appendix: Decomposition of Exposure-Outcome Association into Independent between- and within-Sibship Regression Associations Using Orthogonal Transformations

Suppose we have data from sibships, with outcome Y_ij and exposure X_ij recorded on the jth member of the ith sibship. Extension to multiple exposures is straightforward and we omit it only to avoid complicating the notation. The outcome and exposure are related through the linear regression equationYij=αi+βXij+ϵij(5)where α_i incorporates a shared familial effect. We therefore assume that the residual error terms ε_ij are i.i.d. normally distributed with mean 0 and variance σ². For a given sibship of n members, the series of n regression equations can be written in vector formY=α+βX+ϵ(6)where Y, X, and ε are all vectors of length n. In addition, α is a constant vector and β is a scalar. If A is an orthogonal matrix of full rank, the estimability of parameters is preserved if we premultiply both sides of Eq. 6 by A,AY=Aα+βAX+Aϵ(7)which we rewrite asY*=α*+βX*+ϵ*(8)where Y* = AY, α* = Aα, X* = AX, and ε* = Aε. The variance-covariance matrix of ε being σ²I, the corresponding matrix for ε* is Aσ²IA^T = σ²I, by the definition of an orthogonal matrix. Our proposed approach constrains the matrix A to have one row of the form (1,…,1) so that any terms common to all members of the sibship are captured in the corresponding transformed outcome and exposure variables, with the relevant regression coefficient representing a cross-sectional or between-sibship effect. The remaining rows of A will then be orthogonal to (1,…,1) and hence to α, implying that only the first component of α* is nonzero. These other rows represent within-sibship contrasts, with n − 1 transformed (X_j, Y_j) pairs contributing to estimating the within-sibship regression effect. In the simplest case of sib-pairs, we useA=12[111−1]in which case the pair of equationsY1=α+βX1+ϵ1(9)Y2=α+βX2+ϵ2(10)transforms to12(Yi1+Yi2)=β12(Xi1+Xi2)+ηi(11)12(Yi1+Yi2)=β12(Xi1+Xi2)+ϵi(12)

If we use only the latter formula, we obtain the standard paired difference regression. Comparing estimates β_b and β_w from the first and second sets, we can test the hypothesis that the effect of X is the same within and between pairs. For sibships of n > 2 members, the process is similar, with each sibship providing n − 1 data points to the “within” analysis and one to the “between.” Incorporating sibships of different sizes into a single regression analysis is straightforward provided a suitable matrix A_k can be found for sibships of size k that represents an orthogonal transformation and can be standardized (the factors of 1/2 in the n = 2 case above) to ensure that the residual error terms are identically distributed for sibships of different sizes, allowing the estimation to proceed using a single linear regression. The mathematical details, including a systematic method for obtaining A_k for arbitrary k, are given in a separate forthcoming article.

For the purposes of this article, the following orthogonal transformations were used for pairs, triples, quartets, quintets, and sextets.A2=12[111−1];A3=13[111−112(1+3)12(1−3)−112(1−3)12(1+3)];A4=12[1111−15/3−1/31/3−1−1/35/3−1/3−1−1/3−1/35/3];A5=15[11111−114(35+1)14(1−5)14(1−5)14(1−5)−114(1−5)14(35+1)14(1−5)14(1−5)−114(1−5)14(1−5)14(35+1)14(1−5)−114(1−5)14(1−5)14(1−5)14(35+1)].A5=15[11111−115(46+1)15(1−6)15(1−6)15(1−6)−115(1−6)15(46+1)15(1−6)15(1−6)−115(1−6)15(1−6)15(46+1)15(1−6)−115(1−6)15(1−6)15(1−6)15(46+1)].

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.