Prognostic DNA methylation markers for hormone receptor breast cancer: a systematic review

This systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [19]. No review protocol was previously published.

Eligible articles were original research reports in the English language that had investigated hypo- or hypermethylated biomarkers in relation to patient survival or surrogate endpoints such as disease-free survival in breast cancer populations with oestrogen and/or progesterone receptor-positive breast cancer cases. We excluded in vitro studies, studies on non-human material, studies that focused on hereditary breast cancer cases, studies that focused on non-CpG DNA methylation and studies that had reported large amounts of data from biomarker arrays without further specification of the data to a single potential biomarker or biomarker panel.

PubMed and EMBASE were searched up to November 2018 for eligible studies using the following keywords and equivalents of these: ‘breast cancer’, ‘DNA based methylation biomarker’, ‘hormone receptor positive’ and ‘prognostic or predictive clinical outcome’ (see Additional file 1: Table S1 for a complete overview of the search terms).

Two reviewers (FH and TR) independently selected studies based on title, abstract and in selected cases full text. Disagreement was resolved by discussion between the reviewers until consensus was reached. References of selected studies were crosschecked for additional studies that were eligible for inclusion.

The following data from all selected studies was collected independently by two reviewers (FH and TR): year of publication, study design, study population, length of follow-up, assay type and cut-off used, sequence of primer or probe, statistical methods used and reported association between marker and patient outcome were collected from all selected studies. When available, both univariate and multivariate outcome measures were collected. Study population information consisted of population size, country of patient selection, age, grade, hormone receptor status, HER2 status and stage according to the reported American Joint Committee on Cancer classification [20]. The level of evidence (LOE) was assessed for each publication according to criteria as defined by Hayes et al. [21] and the OCEBM Levels of Evidence Working Group [22].

For each publication, all study endpoints on outcome were collected and compared with ‘the proposed Standardized definitions for Efficacy EndPoints in adjuvant breast cancer trials’ (STEEP) [23]. Endpoints not defined in accordance to STEEP definitions were converted to STEEP-defined endpoints when sufficient information was provided. All defined biomarkers were checked for aliases in the NCBI Gene database and were reported by their current RefSeq gene names.

All selected articles were scored according to the ‘REporting recommendations for tumour MARKer prognostic studies’ (REMARK) criteria [24, 25]. The REMARK checklist consists of 20 items containing one or multiple sub-items. A single item was scored with one point if all relevant sub-items were reported, half a point if only part of the information was reported or zero points if no information on this item was reported. The REMARK checklist is presented in Additional file 2: Table S2. Scoring was performed by two independent researchers (FH and TR). If the total score per article differed, the differences were discussed until agreement was achieved.

REMARK scores were used to assess the risk of potential selection, measurement and confounding bias. The risk for selection bias was assessed by REMARK item #2 (‘patient characteristics’) and #6 (‘sample selection and follow-up’). Studies obtaining < 1.5 points for these combined items were considered to have an increased risk. Measurement bias regarding the assay method was assessed using REMARK items #5 (‘assay method’) and #11 (‘handling of marker values’). REMARK item #7 (‘clinical endpoint definition’) was employed to assess the risk of measurement bias regarding outcome assessment; incomplete or lack of reporting of this item (score < 1) was considered at risk for measurement bias. Confounding bias was assessed using REMARK criterion #16 (‘multivariable analysis’), as in multivariate analysis (score = 1) potential confounding is taken into account. In order to investigate the effect of study design on marker significance, we compared REMARK scores between studies that found significant results and studies that did not find significant results using a Wilcoxon signed-rank test.

A forest plot was prepared for all methylation markers that were investigated in two or more study populations. When included studies reported results for more than one location per marker or reported results derived from more than one source of DNA, such as primary tumour tissue or blood serum, all reported results were represented in the forest plot. If available, multivariable hazard ratios (HRs), 95% confidence intervals (CI) and p values were used. When studies reported only p values without HRs, these were still included in the forest plot, in order to give a complete overview. The statistical programming language R (version 3.3.1) was used to perform all analyses and generate the figures.

The search in the PubMed and EMBASE databases yielded a total of 788 potential publications. One hundred seventy-eight publications were removed as duplicates. After removal of 183 publications that either were not written in English or did not concern original research, 427 studies remained and were screened for eligibility based on title, abstract or full text. Three hundred seventy-two papers were excluded for not matching our inclusion and exclusion criteria. In addition to the remaining 55 papers, 17 papers were identified during reading and included in this review. This selection procedure resulted in 72 included papers [13, 20, 26‐95]. A flowchart of this selection procedure is provided in Fig. 1.

A summary of all studies, studied markers and study characteristics is provided in Additional file 3: Table S3. The number of included patients ranged between 34 and 1163, with a median study sample size of 119. Median or mean follow-up time per study ranged between 20 and 238 months. In 59 studies (82%), either fresh frozen or formalin-fixed paraffin-embedded primary tumour tissue from surgical resections was used for DNA extraction. Nine studies (13%) used plasma or serum derived from the blood and one study (1%) used serum derived from the bone marrow. Fine needle aspirates of the tumour were used in one study (1%). Two studies (3%) did not report the origin of the DNA samples. None of the studies reported selection of a specific tumour grade or stage. All but three papers studied hypermethylation as opposed to hypomethylation. Ten different techniques were applied to assess methylation status, of which methylation-specific PCR was used most frequently (n = 33, 46%) followed by quantitative methylation-specific PCR (n = 16, 22%).

We could categorise 55 papers as OCEBM level 4 and 19 studies OCEBM level 3. There were no level 1 or 2 studies.

All included publications were scored for adherence to the REMARK criteria. Scores ranged from 4.5 to 19 out of the maximum 20 points; the median REMARK score was 12. Only four studies (6%) scored over 15 points.

Quality of reporting highly differed per REMARK item. Figure 2 depicts the number of papers that provided any information for each REMARK item and the number of papers that provided all information per item as required by the REMARK guidelines. For most items, information was provided by the authors, but often this information was incomplete. For example, all studies provided information on the study population, but in 72% of the papers, clear inclusion and exclusion criteria or key patient characteristics, needed to place study results into a clinical perspective, were lacking. The assay used to analyse DNA methylation was mentioned in all studies; however, only two studies (3%) provided all information required to reproduce the assay. Similarly, all studies provided information on statistical analysis, but information needed to reproduce the analyses such as handling of missing data and selection of variables was missing in over 95% of the included papers. Other items, such as patient treatment (65%), biological origin of samples (83%), handling of cut-off values (92%) and demographics of included patients (75%), were generally well reported. Correlations between markers and classical prognostic factors, as well as univariate outcome results were adequately provided by most authors. However, multivariable analyses adjusting for classical prognostic markers were performed in only 32% of included papers.

Comparison of REMARK scores between studies that produced significant results and studies that did not produce significant results showed that studies with low REMARK scores were less likely to report a statistically significant correlation between the investigated marker and disease outcome (p = 0.007). The risk of bias of each included study is summarised in Additional file 4: Table S4.

The 72 included studies reported on 87 different DNA methylation markers. Of these, 18 single markers were studied in more than one independent population. Forest plots summarising the results of these repeatedly studied markers are depicted in Fig. 3 and Additional file 5: Table S5. Hypermethylation of seven markers and one marker panel consisting of three markers was independently significantly associated with poor disease outcome [28, 32, 38‐41, 48, 50, 52, 58, 61, 63, 66‐69, 78, 79, 88, 92, 94]. Two markers were both significantly associated with poor and improved outcome in separate studies [54, 75, 78, 91]. Five markers showed a significant relation to poor outcome in one study, while other studies looking into the same marker found no correlation [44, 57, 58, 65, 83]. Finally, four markers showed no significant relation to disease outcome in any study [48, 63, 66, 68, 70, 88, 92, 95]. An overview of all markers studied in two or more independent populations is presented in Fig. 4. Of the 87 reported markers, 68 were only studied once in a single population. An overview of these markers is provided in Additional file 6: Figure S6 and Additional file 7: Table S7.

Hypermethylation of the markers RASSF1, BRCA, PITX2, RARB, PGR, CDH1 and PCDH10, and the marker panel consisting of markers GSTP1, RASSF1 and RARB showed a statistically significant correlation with poor disease outcome. RASSF1 is the most extensively studied breast cancer methylation marker and was studied in 12 independent study populations [28, 32, 36, 40, 50, 52, 63, 68, 74, 79, 88, 92]. Eight studies found statistically significant results; all showed RASSF1 methylation to be associated with poor survival (HR ranging from 1.93 to 5.64). The remaining four studies found no statistically significant association. RASSF1 methylation was tested in DNA derived from primary surgical resections, blood serum, bone marrow-derived serum and in fine needle aspirates, and was capable of predicting outcome independently of DNA origin. Eleven independent studies investigated BRCA1 hypermethylation [20, 29, 42, 48, 55, 56, 76, 78, 79, 93, 94]. Statistically significant results, correlating hypermethylation of this gene with poor disease outcome, were found in four studies, both in DNA from surgical resections and in blood serum. PITX2 was studied in five independent studies, though it should be noted these studies were all performed by the same research group [40‐42, 61, 69]. All but one study found a statistically significant association between PITX2 hypermethylation and poor outcome; this correlation seemed to be location specific and was found in primary surgical resections and blood serum but not in serum derived from the bone marrow [40]. Five studies reported on RARB hypermethylation in relation to breast cancer recurrence [50, 56, 67, 79, 88]. A statistically significant correlation was found in two studies and proved to be independent of studied tissue type. PGR was studied in five independent cohorts in four different studies [64, 66, 73, 79]. Two alternative promotor sites were analysed in these studies; for both alternative promotors, significant correlation to poor survival was found. Three studies reported on CDH1 methylation in relation to disease outcome [49, 50, 58]. Two studies identified a statistically significant correlation between CDH1 hypermethylation and poor disease outcome. PCDH10 was studied by two independent studies; both found PCDH10 hypermethylation significantly correlated with poor prognosis [32, 59]. A combined analysis of GSTP1, RASSF1 and RARB hypermethylation was the only gene panel analysed in two independent study populations [38, 39]. Fujita et al. studied this panel in blood serum derived from two independent study populations and found a strong correlation with poor overall survival in both cohorts.

For two methylation markers, significant correlations with both poor and improved disease outcome were reported. GSTP1 was analysed in four studies [27, 54, 78, 79]. Three studies found promoter methylation of this gene to be associated with poor survival in multivariable analysis. However, Klajic et al. found GSTP1 methylation to be strongly associated with better breast cancer-specific survival [54]. All four studies studied GSTP1 promoter methylation in primary tumour resections; Sharma et al. also investigated the relation between GSTP1 methylation detected in blood serum of breast cancer patients with the recurrence-free interval, but found no statistically significant correlation [78]. ESR1 was studied in eight independent cohorts; in six cohorts, no statistically significant correlation with disease outcome was found [56, 62, 66, 68, 75, 79, 91]. Widschwendter et al. analysed the impact of ESR1 in two patient cohorts in a univariate analysis, one consisting of patients treated with tamoxifen and one consisting of patients that had not received tamoxifen. In the non-treated cohort, no statistically significant correlation was found; however, in the tamoxifen-treated cohort, a borderline statistical significance with improved disease-free survival was shown (HR 0.7; 95% CI 0.5–1.0) [91]. Ramos et al. found a strong correlation of ESR1 hypermethylation and metastasis-free survival in a cohort of 67 patients; in this study, no details on treatment of the cohort were reported [75].

The markers CCND2, SFRP2, PTPRO, CDNK2A and LINE-1 showed all a correlation for methylation and patient outcome in one study, but these effects were not validated in other studies reporting on these markers [42, 44, 45, 48, 50, 54, 56, 57, 63, 65, 72, 79, 83, 87].

Hypermethylation of the markers SFN, APC, ID4 and CDH13 was analysed in two independent sub-populations, but did not show any statistically significant correlation with disease outcome [48, 63, 66, 68, 70, 88, 92, 95].