We have assessed seven measures of epigenetic aging and three methylation-based predictors of lifestyle for their association with breast cancer risk in a large sample (1,655 cases) of women from Western countries (Australia, Europe and the US). We found overall no associations between measures of epigenetic aging and risk of breast cancer. A positive association was observed for the BMI methylation score, but not for smoking and alcohol consumption.
To our knowledge few studies have investigated the association of epigenetic aging with breast cancer risk. We included in this meta-analysis the samples for which an association was reported previously in EPIC-IARC . Slightly different models were used but the results were very similar. The association previously observed in EPIC-IARC was restricted to postmenopausal women (per 1-year IEAA-Horvath: OR = 1.06, 95%CI: 1.02–1.11) compared with OR = 1.00 for premenopausal women. We found no evidence of an association in our meta-analysis, including when restricted to ages older than 50 years. Our results are overall consistent with the findings from the Sister Study , which reported relatively weak associations: based on 1,566 cases, per 5-year AA-Hannum: hazard ratio (HR) = 1.10, 95%CI, 1.00-1.21, AA-Horvath: HR = 1.08, 95%CI = 1.00-1.17, and AA-PhenoAge: HR = 1.15, 95%CI = 1.07–1.23. In our study, the ORs expressed per 5-year AA were compatible for AA-Hannum and AA-Horvath (HR = 1.02, 95%CI, 0.94–1.10 and HR = 1.01, 95%CI = 0.94–1.08, respectively) and more discrepant for AA-PhenoAge: HR = 1.00, 95%CI = 0.95–1.06. Similar to our findings, the authors did not find substantial heterogeneity by e.g. menopausal or ER-positivity status. For AA-GrimAge, the authors expressed the association per standard deviation  and found HR = 1.06, 95%CI: 0.98, 1.14, which is also similar to our study OR = 1.03, 95%CI: 0.94–1.12. Although AA-GrimAge appeared somewhat more strongly associated with risk for postmenopausal women in the Sister Study, the evidence for heterogeneity was weak and there was no indication of this in our data (HR = 1.03 for women aged ≥ 50 years at blood draw).
The main differences between the cohorts included in our meta-analysis and the Sister Study are that it was enriched for family history of breast cancer and had substantially shorter length of follow-up than ours (for the cases, mean time to diagnosis of 3.9 years, compared with > 6 years for all studies we included). We nevertheless did not observe that OR estimates were larger when blood was collected closer to diagnosis (within 5 years: OR ~ 1.01, 0.98, 1.01, 1.04 for AA-Horvath, AA-Hannum, AA-PhenoAge and AA-GrimAge, respectively). The study of Durso and colleagues  compared Horvath and Hannum age acceleration measures between 233 Italian women who developed breast cancer (mean age at recruitment: 52.4 years, mean time to diagnosis: 3.8 years) and cancer-free controls and found no evidence of an association. A study of multiple health outcomes using Generation Scotland data included 83 incident breast cancer cases, diagnosed over 13 years of follow-up in women aged ~ 51 years at baseline . A tendency for risk associations to be positive was observed: per SD, AA-Horvath, HR = 1.01 (P = 0.95), AA-Hannum: HR = 1.24 (P = 0.07), AA-PhenoAge: HR = 1.36 (P = 0.01), and AA-GrimAge = 1.19 (P = 0.16), respectively, in age-adjusted models. The literature to date therefore includes, to our knowledge, approximately 3,550 breast cancer cases and is consistent with a weak (of roughly 8% increase per 5-year AA for AA-PhenoAge) or null association between epigenetic aging measured in blood and breast cancer risk.
There has not been to our knowledge any study examining methylation-based predictors of lifestyle-related factors with risk of breast cancer. A handful of studies have examined risk of overall mortality , survival from oropharyngeal cancer , and risk of several types of cancer in the Melbourne Collaborative Cohort Study . Another study used the Cancer Genome Atlas datasets to develop lifestyle predictors based on tumour DNA methylation  and found that the BMI-associated methylation signature was predictive of shorter breast cancer survival. For the methylation-based predictors used in our study, the variance explained was somewhat higher than that originally reported by McCartney et al.  for BMI (12%), but somewhat lower for smoking and alcohol consumption (61% and 12%, respectively). For smoking, it may be because it was trained to predict log (pack-years) in current smokers, and our analysis also included former smokers; analyses of the MCCS data showed that the R2 was 66% when former smokers were excluded (not shown). Other methylation-based measures of lifestyle have been developed showing similar accuracy, e.g. for alcohol , or smoking [21, 39], and were not tested in the current study; we chose to use these predictors because they were developed using a large sample size of people of similar ancestry (Scottish) and were well validated. In MCCS analyses of other cancer types, the choice of predictor did not appear to make a substantial difference in the observed associations . In another analysis of the Sister Study data, the authors used as inputs to predict breast cancer risk 36 methylation-based measures of biological aging and physiological characteristics and methylation values at 100 individual CpGs (i.e. using altogether methylation values at thousands of CpGs) and derived a risk score that showed reasonable performance with an area under the curve of 0.63, which was similar to, and independent of, the association observed for the 313-SNP polygenic risk score . We did not attempt to combine methylation scores in our study because most associations were weak, but it is likely that this type of approach may yield improvements to breast cancer risk prediction in the future.
That we observed only weak or null associations may be explained by the fact that none of BMI, alcohol consumption or smoking are strong risk factors for breast cancer. Previous studies have generally found weak to moderate associations [32, 36, 37], except for lung cancer in the MCCS , for which the effect of smoking is dramatic. We had hypothesised that methylation predictors of BMI, alcohol and smoking could contain more information about lifestyle than the measured risk factors - for example exposures accumulated over the lifetime, in particular during sensitive periods such as early life or the periconceptional period, which could be better captured by DNA methylation compared with questionnaires at older ages. BMI has consistently been found to be positively associated with risk of breast cancer for postmenopausal women, and negatively for pre-menopausal women ; we did not observe this using methylation scores as the estimates of associations were similar by age at blood draw (< 50 years: HR = 1.10 [0.93–1.30] and ≥ 50 years: HR = 1.10 [1.02–1.19]). The association we observed for BMI might also reflect the combined effect of several aspects of obesity beyond BMI  that could be captured by changes in DNA methylation. For other breast cancer risk factors, there is to date no convincing evidence that they are strongly associated with blood DNA methylation changes, e.g. for mammographic density  or lifetime estrogen exposure . Additional risk factors for breast cancer were not adjusted for, but this would probably make little difference to the results given their confounding effect on the lifestyle methylation-breast cancer association is likely small.
The main strength of our study is the largest sample size to date of ~ 1,650 cases with long follow-up and comprehensive assessment of epigenetic measures for association with risk. The same analysis method was applied across cohort datasets and participants were representative of the general population. Limitations of our study include the relative heterogeneity of the pooled samples; even though most participants were of European ancestry, there was some variation in terms of age at inclusion, follow-up time or sample processing. All studies used the same pipeline for normalisation of the data, but PLCO used the EPIC assay, which may result in small measurement differences.