Epigenome-wide DNA methylation and risk of breast cancer: a systematic review

Any observational or intervention study that evaluated the association between DNA methylation and breast cancer risk, whatever the study design, was eligible for inclusion. No restrictions were applied regarding language or publication type (articles, short reports and abstracts).

Women included in the studies before or after breast cancer diagnosis, regardless of age, stage, treatment regimen and menopausal status, were eligible. No participants were excluded based on ethnicity. A special attention was paid to identifying overlapping populations between studies, by comparing study population source, date of start and end of study recruitment, inclusion criteria, follow-up duration and population characteristics. When overlapping populations between studies was encountered, the study with the largest sample size was considered as the reference, and information was supplemented by the other publications as required.

Only studies that measured DNA methylation in human samples (blood, breast tissue, breast fine needle aspiration, ductal fluid, human milk), on a genome-wide scale (epigenome-wide studies) using the HM450k or EPIC BeadChip were eligible. Measures of global DNA methylation across all included probes or a predefined set of probes (subset of CpGs defined by spacial localization or a pre-specified function such as epigenetic clocks algorithms) as well as probe-wise differential methylation analysis were considered appropriate exposure estimations.

Breast cancer risk, measured as breast cancer incidence, prevalence or breast mammographic density (a recognized breast cancer risk factor), or as defined by authors of included studies, was the primary outcome. Comparisons between matched normal and tumor tissue from the same patient were not considered a measure of breast cancer risk and were not included.

The references identified by the search strategy were reviewed independently by two authors (KEI and CD) in a 2-step process. First, the title and abstract of each study were screened to exclude obviously non-eligible studies. Then, the full text of retained articles was examined and subjected to evaluation using the predefined eligibility criteria. Whenever required, a third review author (FD) was consulted. When required, further information was sought from the authors by email.

Data extraction was performed using an exhaustive standardized form designed for this review. Information about study design (inclusion criteria, sample size and methodology), participants and tumors characteristics at diagnosis (age, ethnicity, menopausal status, tumor invasiveness, tumor estrogen receptor (ER) status), exposure assessment (timing, tissue sample, tissue processing, data preprocessing methods), measured outcome and reported results (any reported measure of association, adjustment variables, and statistical model selection procedure) were collected. For observational studies, special attention was paid to distinguishing between adjusted and unadjusted results, and to the variable selection method used in multivariate analyses. The study’s definition of each retained characteristic or variable was recorded. In the case of multiple publications related to the same study, and to avoid the overlap across studies populations, the publication reporting the outcomes of interest to the present review or the one with the longest follow-up of these outcomes or with the largest sample size was considered as the reference, and information was supplemented by secondary publications as required. Abstracts with insufficient information and data to permit inclusion were excluded from the qualitative synthesis. Data were extracted independently by two review authors (KEI, DD) to ensure their consistency.

Assessment of risk of bias was performed for each study and for the overall risk of bias across studies. Based on the “STrengthening the Reporting of OBservational studies in Epidemiology” (STROBE) statements [14], and the rating approach of the “Risk Of Bias in Non-randomized Studies - of Interventions” (ROBINS-I) tool [15], the following domains were evaluated for risk of bias in included studies: selection of participants into the study, exposure measurement, outcome measurement, potential confounding accounted for, missing data, and selective reporting. When required, a second reviewer (CD) was consulted.

Given that high heterogeneity between studies was expected, quantitative synthesis of data was considered not appropriate. Using additional tables, a formal systematic qualitative and narrative synthesis of studies characteristics and results was performed separately for each type of tissue sample. A representative population sample was defined as one that includes at least 80% of postmenopausal patients and at least 80% of ER-positive invasive breast cancers [16]. The results were considered adjusted only when all important confounders were considered for adjustment. Authors should have considered age, body mass index or any other estimation of body fat, breastfeeding or parity, alcohol consumption and smoking as potential confounders. In addition, studies including multiple ethnic groups should have adjusted for ethnicity if no bioinformatics method was used to avoid population stratification bias. If authors, in the context of a particular study, demonstrated that a confounding factor is not associated to intervention or to outcome (i.e. a null association measure), and subsequently did not adjust for this factor, the results were considered adjusted [15]. In this context, a “no statistically significant association” was not considered a “no association” [15]. If authors have considered all important confounders for adjustment, and used an appropriate method for variable selection (i.e. backward selection method based on change in estimate) to reduce the number of adjustment covariates, the results were considered adjusted. A stepwise forward selection method or a selection method based on p-values were considered not appropriate as these methods are prone to introduce selection bias. A method based on change of estimate of odds ratio was considered inappropriate because of the non-collapsibility of such measures [17].

The direction and magnitude of observed associations across different statistical models were compared between studies for average methylation analyses, globally and by genomic regions. All individual differentially methylated CpGs identified from each study were compared to detect any overlapping CpGs. Results were considered consistent when associations were in the same direction across studies (at minimum in two studies) with no study reporting an opposite association. Any discrepancy was analyzed for sources of heterogeneity. A positive association was defined as an observed higher risk with higher methylation levels whereas a negative association was defined as an observed inverse association.

Differences between studies, including study design, participant characteristics (age and menopausal status), tumor characteristics (invasiveness, ER status and treatment received), exposure measurement (timing, type of tissue sample, preprocessing methods), statistical analysis (parametric or not, robust or not, adjusted or not) and different levels of risk of bias were considered to explore possible sources of heterogeneity.

Characteristics of the 17 studies of blood-derived DNA methylation are reported in Table 1 and Table S2. These studies involved between 90 and 228,951 participants (median = 465 participants), including 48 to 122,977 cases (median = 233 cases) drawn from one to four different populations. Most studies were nested case-control studies, were conducted on populations from European countries, spanned from 2 weeks to over 20 years of follow-up and evaluated incident breast cancer risk. All studies used the HM450k beadchip. One study aggregated methylation data of common CpGs retrieved from four different populations, of which one have used the EPIC beadchip [19].

Studies included breast cancer patients between 48 and 64 years of mean age (n = 12 studies), and one study included exclusively patients under 40 years old [28]. Proportion of postmenopausal patients varied from 31 to 100% (n = 7 studies), with only one study including at least 80% of postmenopausal patients [26]. Proportion of patients presenting invasive breast cancers varied from 88 to 100% (n = 7 studies), with five studies including exclusively invasive breast cancers [18, 19, 21, 30, 31]. Proportion of ER-positive breast cancers varied from 0 to 83% (n = 7 studies), with only two studies including at least 80% of ER-positive breast cancers [30, 32], and one study including exclusively ER-negative breast cancers [22] (Table 1 and Table S2).

Most studies measured blood-derived DNA methylation in samples collected before cancer diagnosis, estimated blood cell-type proportions using Houseman’s algorithm and used functional normalization to correct for probe design bias. Few studies mentioned exclusion of cross-hybridizing probes, probes containing SNPs, and probes located on X-chromosomes (Table 1 and Table S2). Most studies reporting global methylation analysis across all included probes or a predefined set of probes used methylation beta-values in conditional or unconditional logistic regression models, with only three out of nine studies considering all important confounders for adjustment [23, 24, 27]. Most studies reporting probe-wise differential methylation analysis used methylation beta-values, conditional or unconditional logistic regression models, Bonferroni’s correction for multiple comparisons, with only three out of 16 studies considering all important confounders for adjustment [23, 24, 27] (Table 1 and Table S2).

Characteristics of the three studies of breast tissue DNA methylation are reported in Table S3 [35‐37]. These studies involved between 96 and 262 participants, including 35 to 210 cases, drawn from hospital and tissue bank registries, with one study using The Cancer Genome Atlas (TCGA) data [35]. These studies were mainly cross-sectional and used samples collected after breast cancer diagnosis, with only one study reporting samples collection before any treatment [36]. All three studies used the HM450k beadchip [35‐37].

Included patients in these studies were between 50 and 61 years of mean age, with one study including 29% of patients under 49 years old [37]. One study included 33% of postmenopausal patients [36], and two studies included more than 80% of invasive breast cancers [35, 37]. Proportions of ER-positive breast cancers varied from 63 to 98%, with two studies including more than 80% of ER-positive tumors [35, 36] (Table S3).

Only one study compared normal breast tissue of cases to normal breast tissue of non-cases [36], whereas one study compared tumor tissue of cases to normal tissue of non-cases [37] and the third one compared tumor tissue collected from breast cancer patients to normal breast tissue collected from a different group of breast cancer patients [35]. No study verified cell composition of collected samples nor considered correction for cell-type proportions. Correction for probe design bias was reported by two studies [36, 37], whereas only one study mentioned exclusion of cross-hybridizing probes and probes containing SNPs [37] and none of them reported exclusion of probes located on sex chromosomes. All three studies used methylation beta-values, nonparametric tests for global methylation and probe-wise differential methylation analyses, and Benjamini-Hochberg’s correction for multiple comparisons. No study performed appropriate adjustment for breast cancer risk and prognostic factors to control for confounding and reverse causation bias.

Among the nine studies reporting global methylation analysis, six measured average methylation across all included probes, of which one estimated separate associations in three different populations [33] (Table S2). Out of the eight separate association analyses, four identified a global hypomethylation in women who developed breast cancer, with odds ratios ranging from 0.69 [0.50–0.95] [31, 38] to 0.94 [0.85–1.05] [19], and one study reported a trend toward a marginally lower average methylation in breast cancer patients [32] (Table S2). The three other analyses did not identify a difference between cases and controls [25, 29, 30], and no study reported an opposite association. Three studies reported analyses by CpG location [30, 33, 38], of which one reported higher CpGs islands methylation in breast cancer patients [30] whereas the two other studies did not identify an association with breast cancer risk (Table S2). One study also reported higher methylation in CpGs located in functional promoters but lower methylation in CpGs located far from islands and CpGs located outside promoters in association with breast cancer risk [38].

Five studies measured average methylation across a pre-defined set of probes, four of which corresponded to estimations of epigenetic age using different published algorithms such as Horvath (353 CpGs, n = 4 studies) [25, 29, 30, 36], Hannum (71 CpGs, n = 3 studies) [25, 29, 36], Levine (513 CpGs, n = 1 study) [25] and Weidner (3 CpGs, n = 1 study) [29] epigenetic clocks. Higher Horvath’s epigenetic age was associated with 4 to 9% higher risk of breast cancer in three out of four studies (OR = 1.04 [1.01–1.08] [30], HR = 1.08 [1.00–1.17] [25], and HR = 1.09, p-value 6.3 × 10^− 5 [36]), with one study reporting no association [29]. One study reported 10% higher risk of breast cancer with higher Hannum’s epigenetic age (HR = 1.10 [1.00–1.21]) [25], whereas the two other studies reported no association [29, 36]. One study reported 15% higher risk of breast cancer with higher Levine’s epigenetic age (HR = 1.15 [1.07–1.23]) [25] whereas the study that estimated Weidner’s epigenetic age reported no association with breast cancer risk [29]. One study calculated a methylation index based on 31 CpGs associated with estimated lifetime estrogen exposure and reported 43% higher breast cancer risk in the fourth vs first quartile of methylation index (OR = 1.43 [1.05–2.00]) [18].

Sixteen studies performed probe-wise differential methylation analyses, of which one study performed separate association analyses in two different populations [33]. Out of the 17 probe-wise differential methylation analyses, seven did not identify associations with breast cancer risk, of which one study reported two differentially methylated CpGs positions (DMP) when restricting analyses to cases occurring within 2 years of blood draw [19]. The other 10 probe-wise differential methylation analyses identified between one and 806 DMP (median = 24 DMP) with no overlapping DMP between different studies. Five genes overlapped between two different studies but differed in the identified DMP, namely: GRB10 [18, 25], RPH3AL [18, 25], SEMA5A [25, 33], C7orf50 [25, 27] and XYLT1 [22, 27].

The one study that measured average methylation across all included probes reported higher methylation in tumor tissue of cases than in normal tissue of controls [37], globally and in CpGs located in islands and shores whereas CpGs located in shelves and “open sea” were hypomethylated in tumor tissue of cases [37]. One study measured average methylation across a predefined set of probes corresponding to Horvath’s clock and reported higher epigenetic age in normal breast tissue of cases when compared with normal breast tissue of controls [36] (Table S3).

Two studies performed probe-wise differential methylation analyses and reported respectively 550 [35] and 2761 DMP [37] between tumor and normal tissue. Detailed analysis of overlapping DMP was not performed because the list of DMP was not reported in one study.

No overlapping DMP was identified between studies of blood-derived DNA methylation and studies of breast tissue DNA methylation. Thirteen genes (IGF2BP, HIST1H3E, CUBN, ADCY4, ZNF804A, HIST1H1A, NOX4, CYP24A1, GLIPR1L1, CHODL, PLSCR4, CDH26 and RAD54B) overlapped between a study of blood-derived DNA methylation [25] and the study of tumor vs normal breast tissue of different breast cancer patients [35] but differed in the identified DMP.

Overall, patients age was not related to the observed differences between studies results. Insufficient information was available to evaluate the impact of other population characteristics, such as menopausal status, and tumor characteristics, such as tumor invasiveness and ER status. Studies that have identified an association between global methylation and breast cancer risk reported follow-up periods shorter than 10 years, and one study reported stronger associations after restricting analyses to the first 5 years to 10 years after blood draw [38]. However, this observation was not reflected by differences in time to diagnosis for cases and was not evaluated in studies reporting probe-wise differential methylation analyses because of lacking information.