Differential Impact of BMI-Associated Genetic Variants on Breast Cancer Risk: Insights From Mediation Analysis.

1
Introduction
Breast cancer remains one of the most commonly diagnosed malignancies among women, with approximately 2.3 million new cases and 670,000 deaths reported worldwide in 2022 [1]. The etiology of breast cancer is multifactorial. Among established risk factors, obesity has been recognized as a significant contributor, along with advancing age, excessive alcohol consumption, family history, reproductive history (e.g., age at menarche and first pregnancy), postmenopausal hormone therapy, and genetic factors including certain high‐penetrance genetic mutations, such as those in BRCA1, BRCA2, and PALB2 [1].
Obesity is a complex condition characterized by excessive body fat accumulation, which contributes to various chronic diseases [2]. Body Mass Index (BMI), calculated as weight in kilograms divided by the square of height in meters, is a widely used measure for evaluating obesity [3]. Epidemiological studies have demonstrated that higher BMI is associated with an increased risk of breast cancer, particularly in postmenopausal women [4, 5]. This association is thought to result primarily from the significant decline in ovarian function that occurs after menopause, which reduces estrogen production by the ovaries and shifts the main source of estrogen synthesis to adipose tissue [6].
Genetic factors play a crucial role in obesity susceptibility, with heritability estimates suggesting that genetic variation accounts for more than 20% of BMI variation [7, 8, 9]. In our previous study of the general Japanese population, 11 single nucleotide polymorphisms (SNPs) associated with BMI were replicated [10], selected based on a previous Japanese study [11] and GWAS catalog [12]. However, the potential role of these SNPs in breast cancer risk remains unclear. Specifically, it is not well understood whether these SNPs influence breast cancer risk through BMI‐related pathways or via alternative, BMI‐independent mechanisms.
Given that obesity is associated with breast cancer risk, it is important to determine whether BMI‐associated genetic variants contribute to breast cancer susceptibility through BMI‐dependent or independent pathways. Understanding these genetic mechanisms could provide insights into the complex interplay between obesity and breast cancer and potentially identify novel targets for prevention and intervention strategies. Therefore, this study aimed to elucidate the role of BMI‐associated SNPs in breast cancer risk through mediation analysis [13]. This approach would allow us to distinguish between the direct effects of SNPs on breast cancer risk and the indirect effects mediated through BMI.

2
Material and Methods
2.1
Study Population
We conducted a breast cancer case–control study, selecting cases and controls from participants in the Hospital‐based Epidemiologic Research Program at Aichi Cancer Center (HERPACC)‐2 (2001–2005) and HERPACC‐3 (2005–2013), as described in detail elsewhere [14, 15]. Cases were first‐visit outpatients at Aichi Cancer Center Hospital who were diagnosed with breast cancer. Controls were first‐visit outpatients with no history of cancer or neoplasms, selected from the source population [14]—that is, individuals considered to be at potential risk of developing cancer in the future and, if diagnosed, likely to seek care at Aichi Cancer Center Hospital. All participants provided written informed consent, completed a self‐administered questionnaire, and supplied a peripheral blood sample. This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Ethics Committee of Aichi Cancer Center.

2.2
SNP Selection and Genotyping
A previous study conducted within the Japanese Multi‐Institutional Collaborative Cohort (J‐MICC) Study identified 11 SNPs that were strongly associated with body mass index (BMI) in the Japanese population [10]. To avoid redundancy and potential multicollinearity, we restricted our analysis to five independent SNPs from the original set of 11, namely rs939584 (LOC105373352, TMEM18), rs4409766 (BORCS7‐ASMT), rs6265 (BDNF), rs1927790 (HS6ST3), and rs1421085 (FTO) as exposure variables. This selection was based on the evaluation of linkage disequilibrium (LD) between SNPs on the same chromosome using LDlink [16] based on the 1000 Genomes Project phase 3 (JPT). Specifically, rs939584 was in strong LD with rs13021737 (D' = 1.0, R
2 = 0.9068) and rs4854349 (D′ = 1.0, R
2 = 0.9068) on chromosome 2; rs6265 was in strong LD with rs11030100 (D′ = 0.9795, R
2 = 0.9215) and rs11030104 (D′ = 1.0, R
2 = 1.0) on chromosome11; and rs1421085 was in strong LD with rs11642015 (D′ = 0.9671, R
2 = 0.9353) and rs1558902 (D′ = 1.0, R
2 = 0.9673) on chromosome 16. Genotyping was performed using the Illumina platform (HumanCoreExome [HCE] or Infinium Asian Screening Array [ASA]) (Illumina, San Diego, CA, USA). For HCE, details on genotyping, quality control filtering, and genotype imputation are described elsewhere [16]. For ASA, a total of 15,655 subjects from the HERPACC Study were genotyped using the AsianScreeningArray‐24 v1.0 BeadChip array (Illumina, San Diego, CA, USA). Five samples with a genotype call rate of < 0.95 were excluded. An additional 14 samples were excluded due to discrepancies between reported and genetically inferred sex. Using the kinship‐based identity determination method implemented in PLINK 1.9 [17], 85 duplicate samples (pi‐hat > 0.99) were identified, and one individual from each duplicate pair (except for confirmed identical twins) was removed. Four samples showing closely related to multiple other samples were also excluded. Among the 659,184 SNPs genotyped with the array, we excluded non‐autosomal SNPs, along with SNPs that had a call rate < 0.98, a Hardy–Weinberg equilibrium exact test p value < 1 × 10−6, a minor allele frequency < 0.01, or a marked deviation from allele frequencies reported in East Asian populations from the 1000 Genomes Project (phase 3). After quality control filtering, 15,547 subjects and 428,547 SNPs remained for analysis. Genotype imputation was performed with SHAPEIT2 [18] and Minimac3 [19], based on the 1000 Genomes Project all ancestries as a reference panel (phase 3) [20]. After imputation, we extracted target SNPs of interest. Notably, breast cancer cases were included only in the ASA‐genotyped group; therefore, case–control analyses were restricted to subjects genotyped with the ASA array. After excluding individuals with missing data on menopausal status, the final analysis included 1273 breast cancer cases and 4816 controls. SNP–BMI associations were evaluated using a larger sample of 6079 controls, including those genotyped with either the HCE or ASA array, to improve the precision of the estimates.

2.3
Evaluation of Environmental Factors
Information on environmental risk factors was collected using a self‐administered questionnaire, in which participants reported their exposure status before the onset of symptoms leading to their initial hospital visit. Trained interviewers carefully reviewed the responses to ensure completeness and resolve any inconsistencies.
BMI was calculated as the self‐reported body weight in kilograms divided by the square of height in meters, based on data provided at the time of questionnaire completion. Importantly, participants were instructed to report their weight prior to the onset of any illness, and the questionnaire was completed prior to any cancer diagnosis. Participants with extreme values for current BMI (< 12, > 60) (n = 37) were treated as having missing data. Alcohol consumption was assessed based on daily alcohol intake (g/day), calculated using information on drinking frequency and the amount of pure alcohol consumed per session [21]. Cumulative smoking exposure was evaluated as pack‐years, calculated by multiplying the number of packs smoked per day by the number of years of smoking. Physical activity was evaluated as metabolic equivalent (MET) hours per week [22], based on the frequency, intensity and the amount of time per session. Hormone therapy use and lactation history were coded as “yes” or “no,” based on self‐reporting. Age at menarche was categorized into three groups: ≤ 12, 13–14, and ≥ 15 years. Number of livebirths was treated as a continuous variable. Menopausal status was categorized as either pre‐menopausal or post‐menopausal. In the self‐administered questionnaire, participants who reported their menstrual cycles as “Continuing” or “Stopping” were classified as pre‐menopausal, whereas those who reported “Stopped” were classified as post‐menopausal. Frequency of tofu consumption was classified into three categories: < 1, 1–4, and ≥ 5 times per week, based on a validated food frequency questionnaire [23].

2.4
Statistical Analysis
First, we evaluated the association between BMI‐associated SNPs and BMI using linear regression analysis in the control group. The regression coefficient and significance of the association were estimated while adjusting for age (years), HERPACC version, and SNP array type.
Next, we assessed the relationship between BMI and breast cancer risk using logistic regression analysis. The effect of a 5 kg/m2 increase in BMI on breast cancer risk was estimated, adjusting for age and HERPACC version, as well as potential confounders thought to be associated with breast cancer risk [24, 25, 26], including age at menarche, number of livebirths, physical activity (MET), daily alcohol intake, hormone therapy use, cumulative smoking exposure (pack‐years), lactation history, and frequency of tofu consumption.
Furthermore, we evaluated the interaction between SNPs and BMI in relation to breast cancer risk using both additive and multiplicative scales. To evaluate interactions, we included an interaction term between each SNP and BMI in the logistic regression model and assessed additive interaction using the relative excess risk due to interaction (RERI) [27] and multiplicative interaction using the Wald test. The covariates included in the interaction analysis were the same as those used in the logistic regression analysis assessing the relationship between BMI and breast cancer risk, as described above.
Finally, we conducted a mediation analysis using BMI as a mediator to decompose the total effect of BMI‐associated SNP alternative allele (effect allele) on breast cancer risk into direct and indirect effects (Figure 1) [13]. The direct effect represents the ratio of risk of breast cancer among individuals carrying the SNP effect allele compared to those without the effect allele, assuming that BMI remains as it would be in the absence of the effect allele. This means that the direct effect captures the influence of the SNP on breast cancer risk through pathways independent of BMI changes. The indirect effect represents the risk ratio of breast cancer among individuals carrying the SNP effect allele, comparing the scenario where BMI reflects the presence of the effect allele versus its absence. This quantifies the proportion of the SNP's effect on breast cancer risk that is mediated through changes in BMI. The covariates used in the mediation analysis were the same as those included in the logistic regression analysis investigating the association between BMI and breast cancer risk, as previously described. The mediation analysis was performed using the R function ‘cmest’ in the R package ‘CMAverse’ [28]. Given that the risk of breast cancer due to obesity is thought to differ between pre‐ and post‐menopause [26], we conducted analyses stratified by menopausal status (pre‐menopause and post‐menopause). Some data were missing for age at menarche, number of live births, physical activity, daily alcohol intake, hormone therapy use, cumulative smoking exposure, lactation history, and frequency of tofu consumption (Table 1). To address this, we applied multiple imputation by chained equations (MICE) in a mediation analysis, specifying multimp = TRUE for multiple imputations and args_mice = list (m = 20) to generate 20 imputed datasets. For all other analyses, we performed complete case analysis (5107 out of 6089 participants). Analyses were performed using R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria). A two‐sided p‐value < 0.05 was considered statistically significant.

3
Results
Table 1 shows the baseline characteristics of participants, comprising 1273 breast cancer cases (643 pre‐menopausal and 630 post‐menopausal) and 4816 controls (2575 pre‐menopausal and 2241 post‐menopausal). The allele frequencies and imputation quality scores for the SNPs included in the analysis are presented in Table 2. The associations between each SNP and BMI among controls are shown in Table 3. rs939584 T allele and rs1421085 C allele were associated with increased BMI (β = 0.338, p = 2.69 × 10−4 and β = 0.243, p = 6.19 × 10−4, respectively), while rs4409766 C allele and rs6265 T allele were associated with decreased BMI, (β = −0.137, p = 0.023 and β = −0.205, p = 2.69 × 10−4, respectively). No significant association was observed between rs1927790 and BMI. The effects of rs939584, rs6265, and rs1421085 on BMI were consistent with previous studies, whereas the effect of rs4409766 contradicted earlier findings [10, 11]. In a stratified analysis by menopausal status, the associations in the premenopausal group were consistent with the overall group. Conversely, in the postmenopausal group, a significant association with BMI was observed for only two SNPs: rs6265 and rs1421085. Table S1 shows the effect of BMI on breast cancer risk. An increase in BMI was significantly associated with an increased risk of postmenopausal breast cancer (odds ratio [OR] per 5 kg/m2 increase: 1.46, 95% confidence interval [CI]: 1.26–1.70), whereas no significant association was observed for premenopausal breast cancer (OR per 5 kg/m2 increase: 0.99; 95% CI 0.85–1.17). Table S2 shows the interaction between each SNP and BMI in relation to breast cancer risk. No significant interaction was found for any SNP, except for rs6265 in premenopausal breast cancer.
Table 4 presents the results of the mediation analysis assessing the impact of SNPs on breast cancer risk. For premenopausal breast cancer, none of the SNPs demonstrated a significant effect. For postmenopausal breast cancer, on the other hand, the rs6265 T allele exhibited a positive direct and total effect, alongside a negative indirect effect (direct effect OR = 1.18; 95% CI 1.04–1.34, indirect effect OR = 0.98; 95% CI 0.96–0.99, total effect OR = 1.15; 95% CI 1.02–1.31). The rs1421085 C allele showed a positive indirect effect (indirect effect OR = 1.03; 95% CI 1.00–1.05), while no significant direct or total effects were observed. No significant effects were observed for the other SNPs.

4
Discussion
We performed mediation analyses to estimate the direct and indirect effects of BMI‐associated SNPs on breast cancer risk in a case–control study comprising 1273 cases and 4816 controls. Among the five BMI‐associated SNPs, BDNF rs6265 T allele exhibited a carcinogenic direct effect and a protective indirect effect, whereas FTO rs1421085 C allele demonstrated a carcinogenic indirect effect, both on postmenopausal breast cancer. None of the SNPs were significantly associated with premenopausal breast cancer.
A notable finding of our study is the significantly positive direct effect OR observed for BDNF rs6265 in postmenopausal women (Table 4). As no interaction was observed between rs6265 and BMI (Table S2), it is likely that rs6265 influences breast cancer risk through pathways independent of BMI. rs6265 is a nonsynonymous SNP, and its T allele (Val66Met) is predicted to be “probably damaging” by PolyPhen‐2 [29] and “deleterious” by SIFT [30], suggesting a potential adverse effect on BDNF protein function. In addition to its potential impact on protein function, emerging evidence also suggests regulatory effects at the transcript level. According to the Genotype‐Tissue Expression (GTEx) project [31], the T allele of rs6265 is associated with splicing alterations in breast tissue (p = 3.1 × 10−18; normalized effect size = 0.65), suggesting changes in the isoform composition of BDNF‐AS. BDNF‐AS, a long non‐coding RNA, plays a suppressive role in BDNF expression [32], and its dysregulation may lead to increased BDNF levels. Aberrant expression of BDNF has been reported to promote abnormal proliferation of breast cancer cells by activating signaling pathways such as Ras/MAPK and PI3K/Akt via phosphorylation of its receptor TrkB [33]. Notably, BDNF overexpression has been reported to be associated with aggressive tumor behavior and poorer clinical outcomes [34, 35]. The direct effect of rs6265 observed in this study may therefore suggest that this variant contributes to breast carcinogenesis by both impairing BDNF protein function and altering BDNF‐AS splicing which could lead to altered BDNF expression. Further functional studies are warranted to clarify the dual impact of rs6265 on BDNF structure and expression, and its role in breast cancer development.
Regarding indirect effects, rs1421085 in the FTO gene—originally identified as the first obesity‐susceptibility gene and still the locus with the largest effect on BMI across diverse populations [36]—and rs6265 in the BDNF gene each showed a significant indirect effect on postmenopausal breast cancer, suggesting that both variants influence cancer development indirectly through changes in BMI. As shown in Table 3, the rs6265 T allele was associated with a decrease in BMI, which in turn was linked to a reduced risk of postmenopausal breast cancer, whereas the rs1421085 C allele was associated with an increase in BMI, suggesting a potential elevation in breast cancer risk. In contrast, although the rs939584 T allele (LOC105373352, TMEM18) was strongly associated with increased BMI (Table 3), it did not exhibit an indirect effect on breast cancer risk. Previous studies have reported that rs939584 is particularly associated with BMI in early life [10, 37]. In our study, its association with BMI was in fact more pronounced in premenopausal controls (Table 3). Specifically, the rs939584 T allele was significantly associated with increased BMI only in premenopausal women (coefficient: 0.426, 95% CI 0.18–0.67, p = 7.70 × 10−4), whereas no significant association was observed in postmenopausal women (coefficient: 0.251, 95% CI −0.01–0.52, p = 0.064). This suggests that the absence of an indirect effect in postmenopausal individuals may be due to the weaker influence of rs939584 on BMI in later life. In addition, no significant indirect OR was observed in premenopausal women, likely because BMI has a minimal impact on premenopausal breast cancer risk.
This study has several notable strengths. First, by adjusting for a comprehensive set of covariates, we were able to minimize the potential influence of unmeasured confounding factors, thereby enhancing the robustness of our mediation analysis. Second, the associations observed between SNPs and BMI [8, 9], as well as between BMI and breast cancer risk [4], were largely consistent with previous research findings [4, 10, 11], supporting the validity and reproducibility of our results. Nevertheless, several limitations should be acknowledged. Although our findings provide epidemiologic evidence linking genetic variants with breast cancer risk through BMI, they do not directly elucidate the biological functions of BDNF or BDNF‐AS. Therefore, further biological investigations are warranted to confirm and extend our observations. Additionally, the use of self‐administered questionnaires for lifestyle data collection introduces the possibility of recall bias. However, as lifestyle information was gathered prior to the initial examination, the extent of this bias is likely to be limited. This consideration also applies to the reliability of BMI calculated from self‐reported weight and height. It is well recognized that women tend to overestimate their height and underestimate their weight [38]. Nevertheless, a prior study among Japanese women demonstrated that self‐reported anthropometric data are generally accurate [39]. Specifically, in the HERPACC study, the validity of self‐reported height, weight, and BMI was confirmed by comparison with directly measured values. Among women, Pearson's correlation coefficients were 0.978 for height, 0.910 for weight, and 0.913 for BMI, indicating high concordance [40]. Based on these findings, the use of self‐reported BMI is considered acceptable for the present study. Another limitation is that our analysis focused solely on a single, cross‐sectional measure of BMI at the time of study participation, as our primary focus was on current BMI as a modifiable risk factor relevant to cancer prevention. Given that the duration of elevated BMI may be of etiologic relevance, future studies incorporating longitudinal weight trajectories may help clarify the temporal dynamics of adiposity and cancer risk. Finally, the SNPs used in our analysis were identified from a previous Japanese study [11] and GWAS catalog [12]. While our analysis was conducted in women only, we acknowledge that some SNPs may exert sex‐specific effects on BMI. Nevertheless, given that a previous GWAS in a Japanese population reported that SNP‐BMI associations were largely consistent across sexes [11], we believe our approach appropriately captured the key variants.
In summary, our findings demonstrate that some BMI‐associated SNPs may influence breast cancer risk indirectly through changes in BMI, whereas rs6265—a BMI‐associated SNP—was also found to potentially exert a direct effect independent of BMI. These results may provide new insights into the molecular mechanisms of breast cancer development and inform future strategies for personalized prevention.

Author Contributions

Yuri Ando: formal analysis, visualization, writing – original draft, writing – review and editing. Yuriko N. Koyanagi: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, visualization, writing – original draft, writing – review and editing. Yuji Iwashita: writing – review and editing. Sayaka Yamamoto: writing – review and editing. Yumiko Kasugai: resources, writing – review and editing. Isao Oze: data curation, resources, writing – review and editing. Masahiro Nakatochi: data curation, resources, writing – review and editing. Fumikata Hara: resources, writing – review and editing. Fumihiko Matsuda: resources, writing – review and editing. Issei Imoto: resources, writing – review and editing. Hidemi Ito: data curation, resources, writing – review and editing. Keitaro Matsuo: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, supervision, writing – review and editing.

Disclosure
The authors have nothing to report.

Ethics Statement

Approval of the research protocol by an Institutional Reviewer Board: The HERPACC study was approved by the ethics committee of Aichi Cancer Centre (Approval numbers: 2020‐2‐24, 2024‐2‐25).

Consent
All participants provided written informed consent.

Conflicts of Interest
Fumikata Hara has received payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Pfizer, Kyowa Kirin, Eli Lilly, Chugai, Daiichi Sankyo, and MSD. Masahiro Nakatochi, Issei Imoto, and Keitaro Matsuo are editorial board members of Cancer Science (as of June 2025). All other authors declare no conflicts of interest.

Supporting information

Table S1: Effect of BMI on breast cancer risk stratified by menopausal status. Covariates are age at menarche, number of livebirths, physical activity (MET), daily alcohol intake, hormone therapy use, cumulative smoking exposure (pack‐years), lactation history, frequency of tofu consumption, and HERPACC version. Significant estimates (p < 0.05) are shown in bold.

Table S2: Interaction between each SNP and BMI in relation to breast cancer risk. SNP‐BMI interaction on both the multiplicative and additive scales was assessed by including the interaction term of a 1‐unit increase in BMI multiplied by SNP genotype into the conditional logistic regression model adjusted for age, age at menarche, number of live births, physical activity (METs), daily alcohol intake, hormone therapy use, cumulative smoking exposure (pack‐years), lactation history, frequency of tofu consumption, and HERPACC version. Multiplicative interaction was assessed using a Wald test of the coefficient of the interaction term, while test for additive interaction was performed using the relative excess risk due to interaction (RERI). Significant estimates (p < 0.05) are shown in bold.