Background
Breast cancer predictive and prognostic procedures have a significant impact on current medical care. However, traditional prognostic parameters (lymph node diffusion, tumor size, grading, estrogen receptor expression) cannot adequately predict tumor relapse. As an example, 10–20 % of the patients with the best prognosis, i.e. with small size tumors, expressing estrogen receptors and without lymph node invasion, still experience relapse within 5 years [
1,
2]. At the time of diagnosis, progressing cases cannot be distinguished from those that do not relapse by any conventional prognostic parameter. Therefore, effective markers, with better performance than traditional prognostic indicators, are urgently needed.
By merging biological insight and cluster analysis for experimental immunoistochemistry (IHC) parameters, we have previously succeeded in subgrouping breast cancers with distinct outcomes [
3-
5], and response to therapy [
4,
6], indicating the clinical usefulness of such procedures. DNA methylation regulates gene expression, through the inhibition/activation of gene transcription of methylated/unmethylated genes, respectively [
7,
8]. This largely occurs through methylation of CpG islands, most frequently in the promoter region of the genes [
7,
9-
11]. Broad hypomethylation with focal hypermethylation are frequently found in cancer [
8,
12], thus affecting the expression of tumor suppressor genes, e.g.
TP53, DCC, SOCS2, DLEU7 [
13-
16], and favoring the mutation of oncogenes [
17]. In turn, tumor suppressors have been shown to modulate DNA methylation levels, genome stability and DNA methylation-dependent gene amplification [
18,
19], suggesting key interplays between alterations of DNA methylation and tumor progression. Indeed, DNA methylation-mediated loss of expression has been shown to cause functional ablation of hemizygous alleles at loss of heterozygosity (LOH) loci, encoding transcription factors (TF), e.g.
MOS, TTF-1 [
20,
21], or proteins associated with DNA repair [
22,
23], proteolytic processing [
24], morphogenesis [
25], control of cell cycle, signal transduction or apoptosis [
26]. In breast cancer, CpG-island methylation was shown to inhibit
PTCH1 [
27],
EFEMP1 [
28] and
ESR1 [
29] expression.
DNA methylation patterns can be assessed in formalin-fixed paraffin-embedded tissues (FFPE) tumor samples [
30], allowing to profile gene expression regulatory mechanisms in tumors at the time of surgery, through methylation-sensitive restriction enzyme-analysis over a 56-gene cancer-specific biomarker microarray (MethDet-56) [
31]. Long-term follow-up then permits to dissect correlations between DNA methylation profiles and biological outcome [
32-
35]. In this work we identified CpG-island methylation profiles of cancer biomarker regulatory regions, with a deep impact on prognostic determination in breast cancer, and the ability to distinguish cases with limited or nil risk for progression from those at high risk.
Discussion
In this work, we have identified a gene methylation panel for binary classification of breast cancer progression. Utilizing IHC parameters we had previously succeeded in subgrouping breast cancers [
3-
5] for prognostic and therapeutic use [
6]. Profiles of DNA methylation/regulation of expression of pivotal cancer drivers [
51] were expected to provide additional valuable information, and to critically complement current prognostic procedures. Hence, we went on to identify gene methylation profiles that could bear significant value for prognostic determination [
31].
We first identified gene markers whose methylation patterns differed between progressing and non-progressing breast tumors. PLS k-nearest neigbors radial basis machine identified DAPK1, MDGI, BRCA1, P15, PGK1, PGR, SYK, THBS1, 14-3-3σ, APAF1, CALCA and CCND2, as differentially methylated genes between progressing and non-progressing cancers. These genes included cell cycle regulators, signaling kinases, cytoplasmic scaffold/regulatory molecules and p53 interactors, suggesting direct relevance for breast tumor progression.
Hence, we went on to further refine our breast cancer prognostic model, through procedures of best-model fitting of differentially methylated gene profiles. PLS models were shown to provide the highest AUC and were chosen as the best binary classifiers of progressing versus non-progressing breast cancers. The genes that contributed most to a PLS binary classification model were BRCA1, DAPK1, MSH2, CDKN2A, PGR, PRKCDBP, RANKL. Individual-genes assessments showed 63–79 % sensitivity, 53–84 % specificity, positive predictive values of 59–83 % and negative predictive values of 63–80 %. A 5-fold cross-validation in selected models and PLS analysis through distinct procedures (JMP Genomics, Genespring and R) were used to assess gene clusters versus individual genes. Remarkably, when modelled together, the seven genes reached a sensitivity of 93 %, with 100 % specificity, a positive predictive value of 100 % and a negative predictive value of 93 %. The 97 % estimates for the AUC of the 7-gene panel model supported it as a reliable predictor of breast cancer progression, as did the normality of the log-transformed data distributions. Statistical power analysis supported the strength of our analytical strategies. The majority of the predictors (i.e. BRCA1, DAPK1, MSH2, PGR, PRKCDBP) demonstrated high statistical power, with a threshold of ≥80 % and an alpha significance level of 0.05. CDKN2A and RANKL were close to high statistical power thresholds and were shown to provide non-redundant information to prognosis when combined with BRCA1, DAPK1, MSH2, PGR and PRKCDBP. To assess the overall statistical power of the 7-gene set, an average for the seven prognostic genes was computed for each individual sample. Then, the power calculation was performed, as based on the distance between the mean of relapsing cancer samples versus that on non-relapsing cases. A remarkable power of 1 was obtained, strongly supporting the efficiency of 7-genes panel. Unsupervised hierarchical clustering of the tumor samples, demonstrated close agreement with the accuracy measurements. Of note, tight clustering of 13 relapsing cases was oberved, whereas 6 additional cases distributed among tumors with favourable outcome. These findings suggested heterogeneity in the biological paths that are followed to reach pro-metastatic states, in spite of the sharing of candidate causal genes.
Our model predicted that promoter DNA methylation, with subsequent transcriptional inactivation of the 7-gene set would be detrimental and associated with tumor progression. Individual genes findings fully supported this model.
BRCA1 is a tumor suppressor gene [
52,
53] involved in DNA repair, cell cycle checkpoint control, and maintenance of genomic stability [
54]. Germline mutations in BRCA1 predispose women to breast and ovarian cancers [
55], with a 50–85 % lifetime risk of developing breast cancer [
56]. Promoter hypermethylation was shown to cause loss of BRCA1 expression both in sporadic ovarian cancer [
57] and in hereditary ovarian carcinomas [
58]. Promoter methylation was detected in 31 % of carcinomas but in none of the benign or borderline tumors [
59]. Levels of methylation in ovarian tumors quantitatively correlated with decreased BRCA1 expression [
60,
61]. Hypermethylation of BRCA1 was detected at a significantly higher frequency in serous carcinomas than in tumors of the other histological types [
62], with earlier onset of high-grade serous ovarian cancer. BRCA1 promoter methylation was frequently found in triple negative breast cancers and identified a significant fraction of patients with poor outcomes [
63]. Notably, promoter methylation of BRCA1 was also found in 46 % of pancreatic neoplasms [
64] suggesting a broader impact of this alteration, beyond ovarian and breast cancers.
Death-associated protein kinase (
DAPK) is a pro-apoptotic determinant which is dysregulated in a wide variety of cancers [
65]. Hypermethylation of
DAPK1 is the most frequent molecular alteration identified in immunodeficiency-related lymphomas [
66], and was detected in almost all cases of chronic lymphocytic leukemia [
67]. Hypermethylation patterns of
DAPK were found in head and neck cancers [
68], bladder tumors [
69], and brain metastases of solid tumors [
70], and were associated with poor outcome. The DNA methyltransferase inhibitor 5-Azacytidine (5-Aza) was shown to induce promoter demethylation and to restore mRNA expressions of
DAPK in osteosarcoma cells [
71], confirming DNA methylation as a determinant of transcriptional inactivation of this gene.
PGR (progesterone receptor) is a member of the steroid receptor family and mediates the gene transcription regulatory effects of progesterone. The
PGR status yields prognostic information in patients with node-negative breast cancer [
72]. Lack of expression of
PGR, together with loss of estrogen receptors and of Her-2/neu, identifies ‘triple negative’ breast cancers, which are an aggressive, poor-outcome breast cancer subgroup [
4].
PGR was inactivated by promoter methylation in tamoxifen-resistant breast cancer cells. Following promoter demethylation with 5-Aza, the co-addition of oestradiol (E2) restored gene expression, and inhibited cell proliferation [
73].
PGRß was found hypermethylated in 56 % of melanoma cell lines [
74], and in acute myeloid leukemias [
75].
Protein kinase C δ-binding protein is encoded by the
PRKCDBP (
SRBC) gene. Frequent epigenetic or mutational inactivation of
PRKCDBP was observed in sporadic breast, lung, ovarian, and other types of adult cancers as well as childhood tumors [
76]. The expression of the PRKCDBP protein was down-regulated in about 70 % of breast, lung, and ovarian cancer cell lines, whereas a strong expression of the protein is detected in normal mammary and lung epithelial cells [
76].
PRKCDBP is frequently shut-down in glioblastoma multiforme [
77] and in colorectal cancer [
78] by promoter hypermethylation [
79].
PRKCDBP methylation in neuroblastoma was associated with unfavourable outcome [
80].
PRKCDBP is a proapoptotic tumor suppressor which is activated by NF-κB in response to TNFα, suggesting that
PRKCDBP inactivation may contribute to tumor progression by reducing cellular sensitivity to TNFα. Loss of expression of the PRKCDBP protein was associated to hypermethylation in non-small-cell lung cancers and breast cancer cells; re-expression was observed after treatment with 5-Aza [
76,
81].
p16 is a cyclin-dependent kinase inhibitor and a tumor suppressor protein. Loss of the corresponding locus (
CDKN2A) is among the most frequent cytogenetic alteration events in human cancer [
82]. The frequency of inactivation of p16 by DNA methylation is even higher than that by genetic changes in many cancers, e.g. in gastric carcinomas (32–42 % of cases), where this is an early event and is associated with poor clinical outcome [
83]. Correspondingly, p16 methylation is detected in precancerous and inflammatory lesions of colon, lung, liver, oral cavity [
84], and is associated with malignant progression [
85,
86]. p16 methylation is associated with lower overall survival and disease-free survival in non-small cell lung cancer patients [
87], melanomas [
88] and paragangliomas [
85]. p16 is frequently methylated/inactivated in haematopoietic malignancies, such as acute lymphoid leukaemia (ALL), lymphomas and multiple myeloma [
89]. 5-Aza was shown to restore gene transcription of hypermethylated
CDKN2A genes [
90]. Taken together, these findings have led to the FDA approval of 5-Aza for treatment of patients with myelodysplastic syndromes [
89].
MSH2 is a tumor suppressor protein involved in DNA repair, e.g. base excision, and transcription-coupled homologous recombination [
91-
93]. Heterozygous LOH germline mutations of
MSH2 are causal factors of the Lynch syndrome (hereditary non-polyposis colorectal cancer, HNPCC) [
94]. Heritable transmission of propensity to
MSH2 methylation in a family with HNPCC has been reported [
95]. Aberrant DNA methylation and epigenetic inactivation of
MSH2 play a role in the development of ALL, through induction of cell growth and survival [
96]. CpG island methylation in
MSH2 associates with carcinogenesis in colorectal carcinomas presenting with a conventional adenoma-carcinoma sequence. Therefore, the detection of
MSH2 methylation may have clinical significance in the evaluation of colon cancer patients and in a precision-medicine management of the disease [
97].
RANKL (
TNFSF11,
TRANCE) is a TNF family member, and, together with its receptor RANK, is a key regulator of cell survival. The RANKL/RANK system is modulated by osteoprotegerin (OPG) which binds to RANKL and prevents its interaction with RANK. RANKL activates Akt1 through a signaling complex involving Src and TRAF6 [
98]. RANK is found expressed on cancer cell lines and breast cancer cells in patients [
99]. The RANK/RANKL signaling plays an essential role in progestin-induced breast cancer development [
100] and stimulates breast cancer metastasis [
101]. Corresponding, RANKL triggers the migration of cancer and melanoma cells that express the RANK receptor [
99]. The methylation status of both RANKL and OPG quantitatively controls their levels of expression [
102]. Consistent, RANKL expression in myeloma cells was shown to be driven by TNFα-induced gene demethylation [
103]. Thus, RANKL and OPG act as a cancer/metastasis control module, whose balance is determined by epigenetic regulation.
Network analysis was used to identify functional interrelationship across the tumor progression predictors identified. Notably, most of these prognostic genes and their direct interactors in the network were found to map over key cancer cell-signaling pathways. These close relationships suggested the existence of a functional signaling module, which converged on the
ERBB2,
PRG and
BRCA1 hubs. Remarkably, this network module appeared most related to the p53 signaling pathway (
p-value = 0.0000091).
TP53 is the most frequently mutated gene in cancer [
104-
106] and
TP53 mutations are specifically associated to tumor subgroups with distinct biological features, particularly in breast cancer [
4,
5,
33,
107]. Notably, though, the 7-gene-driven network was shown to be active also in cancer histotypes other than breast (non-small cell lung cancer, bladder, ovary), suggesting an even broader relevance for tumor progression.