Identification of methylation sites and signature genes with prognostic value for luminal breast cancer

Data on breast cancer methylation and mRNA expression profiles were downloaded from the TCGA data portal (https://​gdc-portal.​nci.​nih.​gov/​). A total of 1241 samples were available, 628 of which were marked as luminal type (type A or B). Luminal type samples with methylation data (Platform: Illumina Infinium Human Methylation 450) and mRNA-Seq data (Platform: Illumina HiSeq 2000 RNA sequencing) were selected for further analysis. From this analysis, 231 samples were obtained, including 21 control (non-cancerous) tissues and 210 luminal breast cancer tissues. Among the 210 breast cancer tissues, 191 had the corresponding survival information and status.

Continuous variables were expressed as mean ± standard deviation (SD), and categorical variables were expressed as sample size (composition ratio) in clinical information statistics. Methylation sites with significantly different methylation levels were obtained by comparing the methylation levels between cancer and control samples using the Wilcoxon rank sum test. The influence of the methylation level of these sites on the overall survival of luminal breast cancer patients was analysed using the Cox model. Sites with high correlation were further analysed using a linear correlation model to assess the relationship between their methylation levels and the corresponding mRNA expression. We obtained a subset of methylation sites, the methylation levels of which were significantly correlated with the corresponding mRNA expression. Lastly, the relationship between mRNA expression levels and luminal breast cancer prognosis was assessed using the Cox model. The resulting genes were considered as key genes in luminal breast cancer, as their methylation levels and mRNA expression levels were significantly correlated with each other and with the prognosis.

Based on the cut-off methylation levels (mean methylation levels) of the key genes, samples were classified as hypo- or hypermethylated. Differences in expression were analysed by comparing mRNA levels of hypo- and hypermethylation groups with those of the control group (21 samples) using the EdgeR package (R3.1.0) [16]. False discovery rate (FDR) was calculated using the multtest R package. Genes with an FDR of < 0.05 and an expression fold change of > 1.5 or < 0.67 were considered to be significantly differentially expressed genes.

Significantly differentially expressed gene mRNAs that significantly correlated with prognosis were screened using Cox regression in the survival R package [17]. P-value and the prognosis-relevant coefficient β were obtained using log-rank test. Risk score was defined as follows:

Where β and expr are the prognosis-relevant coefficient and expression level of the corresponding gene, respectively. The risk score was calculated for each sample, and the median risk score was set as the cut-off for determining which samples were divided into low-and high-risk groups [18‐20].

Clinical information of the corresponding samples, including age, ER status, HER2 status, progesterone receptor (PR) status, pathological stages (M, N and T), radiation therapy and the risk score, were integrated. To identify the clinical features significantly related to prognosis, a prognosis-relevant correlation analysis between the samples and their clinical information was performed using univariate and multivariable Cox regression in the survival R package. The resulting clinical features were analysed using Kaplan–Meier survival curves.

To validate the risk scoring system, the expression profile under the accession number GSE22226 (platform GPL1708) [21] from the Gene Expression Omnibus (GEO) database (http://​www.​ncbi.​nlm.​nih.​gov/​gds/​?​term=​) and the Metabric cohort dataset (http://​www.​cbioportal.​org/​study?​id=​brca_​metabric#summary) were downloaded as independent validation datasets. In the GSE22226, a total of 130 breast cancer samples were included in the dataset, 57 of which were of luminal type with survival information and were used for validation. The expression levels of signature genes were extracted, and the risk score was calculated for each sample. The samples were divided into low- and high-risk groups, and the clinical information was also integrated, as described previously. Validation was performed by comparing the distribution of risk scores and overall survival time (in days) as well as by Kaplan–Meier survival curve analysis. In the Metabric cohort, a total of 1979 patients with good follow up data were included, 1140 of which were luminal disease. The risk score was calculated as described previously. Kaplan–Meier survival analysis based on risk score model system and Luminal subtypes were performed. Low- and high-risk groups were divided by signature genes in Luminal A and Luminal B samples from the Metabric cohort and the TCGA database. Prognosis value of low- and high-risk groups were shown by kaplan–Meier survival analysis.

Using the TCGA dataset, the samples were divided into low- and high-risk groups according to their risk scores. Differences in gene expression levels between low- and high-risk groups were analysed using the limma R package, and FDR was calculated using the multtest R package. Genes with an FDR of < 0.05 were considered to be significantly differentially expressed. A correlation analysis of their expression levels and risk scores was performed, followed by two-way hierarchical clustering (shown as a heat map), Gene Ontology (GO) analysis and KEGG pathway analysis. The entire analysis process is shown in Fig. 1 as a flow chart.

A total of 210 luminal breast cancer samples and 21 control samples were included in our analysis, which were obtained from patients with a mean age of 60.73 ± 13.18 and 56.38 ± 14.59 years, respectively. Other clinical information, including ER status, can be found in the Supplementary materials (Additional file 1: Figure S1).

We compared the methylation levels of the cancer and control issues using the Wilcoxon rank sum test. A total of 550 methylation sites displayed significant methylation differences (Additional file 2: Figure S2 and Additional file 3: Figure S3). The relationship between methylation levels and breast cancer prognosis was analysed using the Cox proportional hazards model, and 122 methylation sites were found to be significantly associated with prognosis (Additional file 4: Figure S4).

Aberrant methylation is considered to be correlated with gene expression. We next evaluated the correlation between methylation levels and mRNA expression. For genes with methylation sites associated with prognosis, their mRNA expression values were extracted from the mRNA-Seq data. The correlation analysis of 122 methylation site–mRNA expression pairs revealed that 59 pairs were significantly correlated (P < 0.05), 42 of which were negatively correlated and 17 were positively correlated (Additional file 5: Figure S5). Further analysis of the correlation between mRNA expression levels of these genes and prognosis revealed that 14 of them, including VIM, EPHX3, ACVR1, ANGPT1, TPM3, ALOX15, DIO1, KCNJ2, RSPH9, SOSTDC1, SYCP2, MACF1, TDRD5 and CELSR3, were significantly related to breast cancer prognosis (Table 1). As no previous reports showed the existence of the methylation sites of these genes in other types of breast cancer, we consider that the methylation sites of these genes are specific to luminal breast cancer.

One identified gene, sclerostin domain-containing protein 1 (SOSTDC1) was of particular interest, because SOSTDC1 showed a higher methylation level in breast cancer tissues than the other 13 genes (Additional file 6: Table S1) and among the three genes with the highest significant levels (KCNJ2, CELSR3 and SOSTDC1), SOSTDC1 was the only gene that has been reported to be associated with metastatic survival of breast cancer [22, 23], which indicates that SOSTDC1 plays a complex role in metastatic breast cancer, so we chose SOSTDC1 for further study.

Extracted data on SOSTDC1 methylation for breast cancer and the control samples indicated that SOSTDC1 methylation levels were significantly higher in cancer tissues than in control tissues (P = 3.05E-36) (Additional file 7: Table S2). The mean methylation β-value for breast cancer tissues was determined to be approximately 0.7 (indicated as a black line) and was set as the cut-off. Based on this cut-off value, cancer tissues were divided into hypo- and hypermethylation groups, which contained 64 and 146 samples, respectively.

mRNAs with low expression levels (expression value < 5) were removed from TCGA dataset, leaving a total of 11,858 mRNAs. The density peak of the expression levels significantly increased after removal of low-expressing mRNAs (Additional file 8: Table S3). We then compared control samples with both hypo- and hypermethylation samples using significant difference analysis. In total, 217 genes in the SOSTDC1 hypomethylated group and 312 in the SOSTDC1 hypermethylated group displayed significantly differential expression. We obtained 315 genes by combining the two groups.

The survival information and status of 191 breast cancer tissues in the TCGA dataset were available and used for survival analysis. Among the differentially expressed genes, 67 were identified using Cox regression analysis, with their mRNA levels significantly related to prognosis (P < 0.05) (Additional file 9: Table S4).

To identify prognosis-associated signature genes, 67 genes were sorted according to their P-values derived from Cox regression log-rank test. Top n genes were used to construct a series of risk scoring systems (Table 2). The corresponding risk scores were calculated, and samples were divided into low- and high-risk groups under each risk scoring system. Their correlation with prognosis and the corresponding area under the curve (AUC) are shown in Table 3. The correlation between low- and high-risk groups and prognosis was at a maximum when the top seven genes were used, while the AUC [20] was maximized when the top eight genes were used. Therefore, we used the top eight genes as signature genes [establishment of cohesion 1 homolog 2 (ESCO2), Protein Kinase C And Casein Kinase Substrate In Neurons 1 (PACSIN1), cell division cycle associated 2 (CDCA2), polymeric immunoglobulin receptor (PIGR), pleiotrophin (PTN), repulsive guidance molecule A (RGMA), kallikrein-related peptidase 4 (KLK4) and centromere protein A (CENPA)].

Risk scores were calculated using the risk scoring system containing the top eight genes (Additional file 10: Table S5). The samples were divided into low- and high-risk groups based on the median risk score (55.27). Kaplan–Meier survival curve analysis showed that sample grouping by this risk scoring system correlated significantly with prognosis (Fig. 2a). Moreover, the overall survival was significantly longer in the low-risk group than in the high-risk group (P = 0.00128).

We examined the expression levels of the signature genes, which revealed that their expression levels significantly differed between the low- and high-risk groups (Fig. 3a, Additional file 11: Table S6). Significant differences in expression levels were also found between the control samples and the hypo-or hypermethylation groups. (Figure 3b, Additional file 12: Table S7). However, no significant differences were found between the hypo and hypermethylation groups. Moreover, low- and high- risk groups of Luminal A samples divided by the signature genes significantly correlated with the survival ratio. As to Luminal B samples, the survival rate of low risk group was also higher than that of high risk group, although this correlation was not so significant (Fig. 4a). We also evaluated if the prediction analysis of microarray 50 (PAM50) intrinsic subtypes, which are prognostic independent of standard clinicopathologic factors, could well differentiate Luminal A and Luminal B subtypes. As shown in Additional file 13: Table S8, PAM50 was not the key signature genes for splitting Luminal A and Luminal B subtypes in the patient cohort of this study (Additional file 13: Table S8).

To further validate the signature genes, we used the Metabric cohort [6]. The Metabric cohort included 1979 patients with good follow up data of which 1140 were of luminal disease. These luminal samples were also divided into low- and high-risk groups by median risk score based on the signature genes. Low- and high- risk groups of Luminal A samples significantly correlated with the survival ratio. As to Luminal B samples, this correlation also showed no significant difference (Fig. 4b). We combined Luminal A and Luminal B data of the Metabric cohort and made a table (2 × 2) with Luminal A, Luminal B, high score, low score. We found that the distribution of Luminal A and Luminal B samples between low- and high groups were significantly different. Most Luminal A samples fell in the low score group and the majority of Luminal B samples fell in the high score group (chisq test, X-squared = 137.0685, p < 2.2e-16) (Table 4). Kaplan–Meier analysis revealed that high score group was significantly correlated with poor survival of patients with luminal breast cancer in Metabric cohort (Additional file 14: Table S9). The classification of Luminal A and Luminal B also significantly correlated with survival ratio of luminal subtype patients. The patients with Luminal A breast cancer had a longer survival time than the patients with Luminal B breast cancer (Additional file 14: Table S9).

The GSE22226 expression profile dataset was used for validating our risk scoring system. The expression level values of the signature genes were extracted and the risk score for each sample was calculated (Additional file 15: Table S10). Each sample was divided between the low-risk group (29 samples) and high-risk group (28 samples). The survival ratio in both groups was evaluated. Similar to the results seen in the TCGA training dataset, the survival ratio in the low-risk group was significantly higher than that of high-risk group (P = 0.0397) in the validation dataset (Fig. 2b). Additionally, the distribution of the risk scores and overall survival time were similar between the validation and training datasets (Fig. 2c and d). All validation results that we obtained confirmed the robustness and reliability of our risk scoring system.

Clinical information was integrated for prognosis-related correlation analysis (Additional file 16: Table S11). Univariate Cox regression indicated that both the PR status and the risk score were significantly correlated with prognosis, whereas multivariable Cox regression indicated that only the risk score was an independent prognostic factor (Table 3). Further analysis demonstrated that the survival ratio was higher in the low-risk group than in the high-risk group in both PR-positive (PR+) and PR-negative (PR−) patients (Fig. 5; 14 samples vs. 123 samples; P = 0.105 vs. P = 0.00552). However, the scoring system that we used was more sensitive in PR+ than in PR− samples because of the difference in P-values.

We found 121 genes that exhibited significant differential expression (FDR < 0.05) between the low- and high-risk groups. The correlation analysis indicated that 88 of them were positively correlated and 33 were negatively correlated with the risk score (Additional file 17: Table S12). The expression patterns of the top 20 positively and negatively correlated genes are shown as heatmaps using hierarchical clustering analysis in Fig. 6a. Further biological function enrichment analysis revealed that most positively correlated genes were involved in the cell cycle, whereas most negatively correlated genes were involved in development, cell adhesion, ion transport and homeostasis (Fig. 6b, Additional file 18: Table S13). KEGG pathway analysis indicated that these genes were correlated with cancer, cell cycle and signalling pathways (Fig. 6c, Additional file 19: Table S14). The overall results of the KEGG pathway analysis were consistent with those of the biological function enrichment analysis, considering the complex relationship between tumourigenesis and multiple biological processes, such as cell cycle, cell adhesion and development.