We analyzed promoter methylation at 432 CpGs in 14 genes giving rise to 37.000 epigenotypes (Figure ). The analysis included the DNA methylation in ABCB1 (40 CpGs), ATM (56 CpGs), BRCA1 (46 CpGs), CDH3 (35 CpGs), CDKN2A (30 CpGs), CXCR4 (19 CpGs), ESR1 (50 CpGs) FBXW7 (31 CpGs), FOXC1 (14 CpGs), GSTP1 (21 CpGs), IGF2 (16 CpGs), MLH1 (24 CpGs), PPP2R2B (51 CpGs), and PTEN (39 CpGs). The six normal samples were unmethylated for all analyzed regions except for the highly methylated upstream region of BRCA1, the differentially methylated region of the imprinted IGF2 and the promoter region of ESR1 (Figure ). Three amplification products in the ABCB1 gene were found to be methylated in 70%, 64% and 81% of the tumours. Methylation was found for CDKN2A (34% of the samples), FOXC1 (50%), PPP2R2B (56% and 65%), HMLH1 (14%), PTEN (22% and 76%) and GSTP1 (65% and 83%). All samples were unmethylated for the transcription start site of BRCA1, ATM, CDH3, CXCR4 and FBXW7. 10% of the samples exhibited a significant hypomethylation in the far upstream region of the BRCA1 CpG island. Some methylation was found around the transcription start site for ESR1 but also within the normal breast samples. None of the genes displayed an age-dependent variation of DNA methylation at the analyzed loci.
Figure 1 Summary of the methylation data. A) Summary of the average DNA methylation values in percentage for the analysis of the fourteen genes (x-axis) in the 75 breast cancer samples, six normal breast tissues (on top) and the six breast cancer cell lines (bottom (more ...)
Restricting the analysis to the loci with variable DNA methylation levels, no sample showed a completely normal methylation profile, 3/75 tumours (4%) showed abnormal methylation at one locus, 8 (11%) at two loci, 4 (5%) at three, 14 (19%) at four and five loci, respectively, 16 (21%) at six, 9 (12%) at seven, 5 (7%) at eight loci and two tumours (3%) displayed aberrant methylation at nine loci. On average, five loci were thus aberrantly methylated in any sample. Methylation events at the different loci were not randomly distributed and independent from each other (Figure ). As expected, the methylation degrees of the different regions of the same gene were highly correlated if methylation was detected in all amplification products. Less expected, concomitant methylation was often found at different genes such as the ABCB1, FOXC1, GSTP1, PPP2R2B, PTEN promoters identifying thus strongly correlated methylation events on different chromosomes (Figure ). Methylation at the estrogen receptor promoter did not correlate with any other gene. Pyrosequencing provides the advantage of yielding highly quantitative data on consecutive CpGs permitting analysis of the homogeneity of the methylation profiles. We identified for most genes some "core"-regions where DNA methylation levels correlated best with molecular and clinical parameters (see below). For most genes, these regions spanned - as expected - the transcription start sites.
Correlation to expression profiles
The observed methylation patterns were compared to the tumour subclasses as defined by microarray expression profiling [21
]. Basal-like tumours generally showed a lower degree of methylation than the other subclasses (luminal A, luminal B, ERBB2 and normal-like). There was a trend for the absence of methylation at ABCB1
, and GSTP1
in both the basal-like and normal-like tumours, while IGF2
were hypomethylated in the basal-like tumours but not in the normal-like tumours. When analyzing the correlation between the expression level and the DNA methylation status of individual genes, genes with methylated promoters were almost exclusively not expressed, while unmethylated genes could be expressed as well as not be expressed weakening the correlation. The only significant correlation was obtained for GSTP1
(p = 0.003, Correlation coefficient -0.47). Because of their association with survival (see below) we analyzed the expression levels of GSTP1
by qRT-PCR (Additional File 1
). qRT-PCR analysis correlated well with the microarray data (GSTP1
: Pearson Corr. 0.661, p = 0.003; FOXC1
: Pearson Corr. 0.788, p < 0.001; ABCB1
: Pearson Corr. 0.739, p = 0.015). Consequently a significant negative correlation between expression as measured by TaqMan and methylation was found for GSTP1
(Spearman Correlation -0.567, p = 0.018), while expression and methylation for FOXC1
were not significantly correlated (p = 0.5 and p = 0.368, respectively). Highly expressing genes were unmethylated for the respective promoter region of GSTP1
and methylated promoters correlated with silenced expression. The weak correlation between expression and DNA methylation for FOXC1
was due to the fact that the gene was already silenced in most tumours independent of its methylation status. Four samples were methylated for ABCB1
but displayed high expression. This might be due to alternative usage of an upstream promoter [23
] that is not under the control of the analyzed CpG island.
Correlation with clinical parameters
Methylation was analyzed in the discovery and validation cohorts both as a categorical variable, i.e. the presence/absence of methylation at the respective promoter in association with the tumour characteristics, and as a quantitative variable investigating potential associations between the extent or the distribution of DNA methylation values and the analyzed clinical and molecular parameters (Additional File 2
). Promoter methylation of PPP2R2B
in the pre-treatment sample was significantly associated with tumour grade (p = 0.019), whereby high-grade tumours were more frequently unmethylated than grade 1 and 2 tumours in the discovery cohort. The same was observed in the validation cohort of unselected breast cancers (p = 0.008). No association between methylation and tumour size was found. Estrogen receptor status positivity was associated with the presence and increased extent of methylation at the PPP2R2B
promoter in both the discovery (p = 0.004) and the validation cohort (p = 0.006). Samples unmethylated for ABCB1
and those with increased levels of methylation in the differentially methylated region 2 of IGF2
had more often overexpression of the ERBB2
oncogene (p = 0.005 and p = 0.007, respectively), previously analyzed by immunohistochemistry [24
]. No ERBB2 data was available for the validation cohort.
Correlation with TP53 mutations
We compared the observed DNA methylation profiles with the TP53 mutations status and found the lack of ABCB1 and PPP2R2B methylation to be associated with the presence of TP53 mutations in the discovery cohort (p = 0.028 and p = 0.010, respectively) as well as in the validation cohort (p = 0.018 and p = 0.001, respectively). Tumours unmethylated for the middle region of the ABCB1 CpG island were associated with mutations in the loop domains L2/L3 (p = 0.022), a region that has previously been shown to be associated with lack of response to doxorubicin based treatment. PPP2R2B did not show any differential degree of methylation in function of the type of TP53 mutation.
Survival analysis and response to treatment in the doxorubicin treated cohort
The eight genes displaying variable DNA methylation patterns in a significant number of tumours (ABCB1
) within the discovery cohort were tested for association with survival by a logrank test. Breast cancer specific survival was significantly improved in patients with methylated promoters for ABCB1
(p = 0.004, p = 0.004 and p = 0.021 respectively, Figure ). Methylation of ABCB1
did also reach statistical significance after correction for multiple testing (Bonferroni correction, uncorrected p < 0.00625). Consistently, absence of methylation (p = 0.0076, Additional File 2
) in the CpG island of ABCB1
was associated with poor response to doxorubicin (progressive disease) in the patient cohort treated with doxorubicin. In the validation cohort treated with different regimens, a significant difference in survival between methylated and unmethylated samples was confirmed for FOXC1
(p = 0.024) with patients unmethylated for the promoter region having again worse survival. Methylation of GSTP1
did not condition improved survival in the validation cohort of patients (p = 0.331). Similarly, only a trend for improved survival was observed for the methylation status of ABCB1
(p = 0.070). The findings for GSTP1
might indicate a treatment specific effect on survival
Figure 2 Kaplan-Meyer plots of overall survival. Kaplan-Meyer plots of overall survival for patients with either methylated or unmethylated GSTP1, FOXC1 or ABCB1 promoter, respectively (left column). Increased differentiation of patients is obtained through the (more ...)
Survival analysis in the doxorubicin treated cohort based on the logrank test indicated that TP53 mutation status (p = 0.001), grade (p = 0.001) and the estrogen receptor status (p = 0.002) could slightly better differentiate two survival groups in the analyzed sample collection when compared to the methylation status of the single genes (ABCB1 (p = 0.004), GSTP1 (p = 0.004) and FOXC1 (p = 0.021)), while separation based on the progesterone receptor status and amplification of ERBB2 or TOP2A did not reach statistical significance. However, combination of two of the discovered DNA methylation markers further improved the distinction between doxorubicin treated patients having two, one or none of the genes methylated. No statistical difference on survival in function of the gene was found when comparing patients that had one of the two genes methylated and these were therefore combined for analysis. The best two-gene methylation pair comprised GSTP1 and FOXC1 (p = 8·10-5), followed by GSTP1 and ABCB1 (p = 0.001) and ABCB1 and FOXC1 (p = 0.01). Patients with all three genes methylated (n = 20) had an improved survival compared to patients with all three genes unmethylated (n = 10, p = 0.001). However, the separation lost its statistical significance when patients with mixed methylation patterns for all three genes were added to the analysis. We investigated if expression could be used as an alternative molecular measure to DNA methylation and divided samples in high versus low expression based on the mean expression values. The expression of GSTP1 was significantly associated with survival with patients with low levels of expression having as expected an improved survival (p = 0.048). FOXC1 (p = 0.247) and ABCB1 (p = 0.181) were not significant but again showed improved survival for low expressing patients. When combining DNA methylation and expression, patients with an unmethylated GSTP1 promoter and expressed gene had poorer survival compared to patients with a methylated promoter that did not express GSTP1 (p = 0.047). The same correlation was observed for FOXC1 (p = 0.045) and ABCB1 (p = 0.022).
Cox regression analysis of methylation markers and clinical variables in the doxorubicin treated cohort
To identify significant parameters contributing to the observed differences in survival, Cox regression analysis was performed. The hazard ratio for each of the contributing factors was estimated separately (univariate analysis) or modelled together (multivariate analysis).
Univariate analysis identified the methylation status of ABCB1
1 and GSTP1
as significant predictors of overall survival. Estrogen receptor status as well as TP53
status and grade were also significant predictors of survival in univariate analysis (Table ). To investigate if the methylation markers ABCB1
were independent prognostic markers, we performed multivariate analysis with the methylation markers, grade, estrogen receptor status, TP53
status and stage. This analysis showed that the patients in this cohort with unmethylated GSTP1
(HR: 7.52, CI: 1.76-32.07, p = 0.006) and FOXC1
(HR: 7.32, CI: 1.11-48.31, p = 0.039) showed a higher risk of dying from breast cancer compared with patients methylated for the same genes (Table ). The effect of ABCB1
methylation on survival was no longer significant in the multivariate analysis probably due to its association with other histopathological factors (Additional File 2
). Inclusion of the operation status (HR: 2.1, p = 0.452) in the multivariate analysis did slight reduce the hazard ratio for GSTP1
(HR: 5.8, p = 0.028) while increasing the HR for FOXC1
(HR: 8.3, p = 0.03). The HR for the other parameters remained unchanged.
Univariate survival analysis
Multivariate survival analysis
In order to identify the model with the minimum number of covariates that fitted best the experimental data, we used the Akaike information criterion. The best model with a reduced number of covariates explaining survival included the methylation status of FOXC1
, stage, grade and estrogen receptor status (Additional File 3A
). The best model with a minimum number of covariates where all covariates were independent of each other included ER, grade and the GSTP1
methylation status (Additional File 3B
). Using only a single covariate to model the survival of the patients by the AIC, the methylation status of any of the three different genes performed superior compared to the classical parameters with GSTP1
fitting the model best followed by FOXC1
. To investigate the effect of the combination of the methylation status of two genes on survival, multivariate Cox regression analysis was again performed. Only the GSTP1/FOXC1
pair (p = 0.005 and p = 0.013 for the combination of either one or both genes being unmethylated, respectively) remained significant together with high grade (p = 0.002) and ER status (p = 0.001) (data not shown).