Introduction

Breast cancer (BC) is currently diagnosed in 1 in 8 Australian women over their lifetime, making it as the primary cause of female cancer-associated deaths in Australia. Decades of studies have identified candidate prognostic biomarkers for BC. Recent bioinformatic and experimental studies have found that microRNAs (miRNAs) can act as biomarkers in BC as they play a crucial role in transcriptional and post-transcriptional gene regulation. miRNAs belong to a group of small non-coding RNAs with 19–25 nucleotides in length. According to conventional RNA logic, miRNAs inhibit/degrade gene expression binding with miRNA response elements (MREs) of messenger RNAs (mRNAs)1. Salmena, Poliseno2 introduced the competing endogenous RNA (ceRNA) hypothesis revealing the bi-directional regulation mechanism of miRNAs. The ceRNA logic explains that non-coding RNA transcripts such as long non-coding RNAs (lncRNAs) with similar MREs can also bind with the relevant miRNAs modulating gene regulation and protein networks. Due to these reasons, ceRNAs have gained considerable attention in cancer studies.

Previous bioinformatics and experimental studies have identified many ceRNAs associated with BC risk3. Apart from generalising results for BC incidents, different cancer stages can also severely impact response to therapy and mortality. In BC, stage I refers to a tumour or small size confined to the breast; stage II explains the disease that has locally advanced beyond the breast; stage III describes BC has spread to the neighbouring organs, and stage IV refers to distant metastatic disease4. The early stages, I and II, are considered treatable compared to advanced stages, III and IV, that require more radical and active treatment strategies5. Therefore, identifying stage-specific biomarkers in BC will significantly contribute to understanding BC biology under different pathological states.

In this study, we performed a ceRNA network analysis to identify stage-specific ceRNAs in BC, and this approach can be applied for any disease of interest in the future. The identified ceRNAs were further studied using two downstream analyses to identify their molecular functions and prognostic ability.

Materials and methods

Patients and samples collection

The expression data (RNA-seq and miRNA-seq) and clinical data of BC were collected from the cancer genome atlas (TCGA) that contains 1091 cases and 113 controls. The HTSeq-counts data of RNA-seq (including protein-coding and long non-coding) and isoform quantification data of miRNA-seq (in BC) were downloaded to a local computing server using the genomics data commons (GDC) data portal6. For selecting miRNA-seq data for individuals, a manifest file was generated using the GDC Data Portal. Then, the GDC data transfer tool was used to transfer data files listed in the manifest file. The differential expression analysis produced the "group" variable to identify differentially expressed protein-coding genes and long non-coding RNAs.

Differential expression analysis

Firstly, we removed TCGA BC samples with duplicated sample IDs. Then samples that are neither solid tissue normal nor primary tumour were removed as we compared primary tumour and healthy samples in the differential expression analysis. First, we performed the counts per million (CPM) normalisation to correct sample library size differences. The low-expressed genes that log (CPM) < 1 in more than 50% of the samples were removed before the differential expression analysis7. Ignoring low-expressed genes improves the total count of differentially expressed genes and enhances sensitivity and precision. Raw counts expression data were re-normalised using the TMM (trimmed mean of M values) method implemented in the edgeR (3.40.0) R package (https://bioconductor.org/packages/release/bioc/html/edgeR.html) to compare expression levels between samples (excluding low-expressed genes)8. The normalised data were transformed into a standard scale using the voom function in the limma (3.54.0, https://bioconductor.org/packages/release/bioc/html/limma.html) (linear modelling for microarrays) R package9. Previous RNA-seq data analysis-related works have recommended this hybrid technique, TMM normalisation with voom transformation, due to its better performance in data preprocessing10,11. Moreover, Oshlack et al. have shown TMM normalisation is robust and outperforms library size normalisation12. In differential expression analysis, linear models were fitted for each gene using the "lmFit" function implemented in the limma (3.54.0, https://bioconductor.org/packages/release/bioc/html/limma.html) R package9. Then eBayes moderation was applied using information across all the genes to obtain more precise estimates of gene-wise variability. Four differential expression analyses were conducted for stage I to IV-control comparisons. The cancer stage was determined using the "pathologic stage" as it provides more accurate information combining results from clinical examinations and surgeries. We gathered 181 (19% basal-like, 5% HER2+, 62% luminal A, 13% luminal B, and 1% normal-like), 619 (21% basal-like, 13% HER2+, 40% luminal A, 25% luminal B, and 1% normal-like), 247 (12% basal-like, 16% HER2+, 40% luminal A, 29% luminal B, and 3% normal-like), and 20 (17% basal-like, 16% HER2+, 25% luminal A, and 42% luminal B) samples for stages I, II, III, and IV, respectively. In each stage-specific expression analysis, differentially expressed mRNAs, lncRNAs, and miRNAs were defined at |log2-fold change (FC)|> 1 and Benjamini–Hochberg (BH)-adjusted p value (default in limma package) < 0.0513.

Competing endogenous RNA network analysis

The differentially expressed mRNAs, lncRNAs, and miRNAs in each cancer stage were applied in the stage-specific ceRNA network analysis. The ceRNA network analysis consists of three main steps: (1) identifying lncRNA-mRNA pairs that share the significant number of miRNAs, (2) selecting positively correlated lncRNA-mRNA pairs, and (3) jointly estimating the significance of multiple miRNAs in lncRNA-mRNA pairs. These three steps are described in detail in previous ceRNA papers14,15. The mRNA-miRNA and lncRNA-miRNA interactions are required to perform steps i and iii. We used miRcode and starBase databases for miRNA-target predictions16,17. The miRcode database facilitates mRNA-miRNA and lncRNA-miRNA target predictions using a broad searchable map that contains 10,419 lncRNAs. The starBase includes miRNA-mRNA interactions predicted by analysing 108 CLIP-seq datasets. Steps i and ii were performed using the hypergeometric test and the Pearson correlation test, respectively. These two testing methods have been implemented in the GDCRNATools (1.18.0, https://bioconductor.org/packages/release/bioc/html/GDCRNATools.html) R/Bioconductor package18. The third step, multiple sensitivity correlation (mscor) analysis, was executed using the SPONGE (1.20.0, https://bioconductor.org/packages/release/bioc/html/SPONGE.html) (sparse partial correlation on gene expression) R/Bioconductor package19. The significant ceRNA interactions were filtered by three user-defined thresholds, (1) false discovery rate (FDR) < 0.01 in hypergeometric test, (2) Pearson correlation coefficient > 0.40, and (3) the adjusted p value of mscor in SPONGE method < 0.05. The resulting lncRNA-miRNA-mRNA associations in each BC stage were combined into a single column, as "<lncRNA gene ensemble ID>_<gene ensemble ID>_<miRNA name>". Then set of values in each BC stage were applied into a four-sets (for four stages) Venn diagram representation.

Functional enrichment analysis

A total of unique 47 aberrantly expressed genes (25 from stage I, 42 of stage II, 40 of stage III, and 47 of stage IV) were analysed to understand the biological functions of identified ceRNAs in this study. The Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses were conducted using the R/Bioconductor clusterProfiler (4.6.0, https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html) R package20.

Survival analysis

Survival analysis was performed using the Kaplan–Meier (K–M) survival curves, implemented in the survival (3.4.0, https://cran.r-project.org/web/packages/survival/index.html) R package to explore the impact of the expression level of RNAs/miRNAs on the prognostic survival of patients21. For each gene/lncRNA/miRNA, the tumour samples were divided into two groups (low-expressed and high-expressed) according to the median expression level. The log-rank test (Mantel–Haenszel test) was used as the statistical method for the K–M curves. The log-rank test statistic has a chi-square (χ2) distribution with one degree of freedom. Therefore, significant genes, lncRNAs and miRNAs were chosen under the χ2 test statistic p value < 0.05. The survival-significant mRNAs were checked for tumour-suppressive/oncogenic/cancer-driven roles using the CancerMine database22.

Ethical approval

The study was approved by the Human Research Ethics Committees of the Queensland University of Technology (protocol code: 1900001147, date of approval: 19 December 2019) and the QIMR Berghofer Medical Research Institute (protocol code: P1051, date of approval: 23 August 2019).

Results

Differential expression analysis results

First, we conducted stagewise differential expression analysis to determine which genes, lncRNAs, and miRNAs are expressed at different levels between tumour and healthy groups. The number of up/downregulated lncRNAs, genes and miRNAs are available in Table 1.

Table 1 Counts of differentially expressed (up/down) lncRNAs, mRNAs, and miRNAs in each BC stage.
Full size table

The above-listed mRNAs, lncRNAs, and miRNAs were involved in the ceRNA network analysis.

Competing endogenous RNA networks of BC stages

We constructed four ceRNA networks for BC stages I–IV. In Table 2, we have included the count of significant ceRNA associations in each BC stage. The number of lncRNAs, genes, and miRNAs involved in stage-specific ceRNA networks is given within brackets.

Table 2 The lncRNA-associated ceRNA networks for different breast cancer stages.
Full size table

According to Table 2, stage II and IV analyses have resulted in a considerably large set of ceRNA networks compared to stages I and III.

The detailed list of significant lncRNA-mRNA-miRNA associations of each BC stage is available in Supplementary Information. The Cytoscape tool version 3.9.1 (https://cytoscape.org/download.html)23 was used to visualise ceRNA networks in each BC stage. Figures 1, 2, 3 and 4 illustrates ceRNA networks for BC stage I, II, II, and IV, respectively. In Figs. 1, 2, 3 and 4, blue-, green-, and yellow-coloured squares represent genes, lncRNAs, and miRNAs, respectively.

Figure 1
figure 1

ceRNA network for stage I BC, was constructed by the Cytoscape tool23.

Full size image
Figure 2
figure 2

ceRNA network for stage II BC, was constructed by the Cytoscape tool23.

Full size image
Figure 3
figure 3

ceRNA network for stage III BC, was constructed by the Cytoscape tool23.

Full size image
Figure 4
figure 4

ceRNA network for stage IV BC, was constructed by the Cytoscape tool23.

Full size image

According to Figs. 1, 2, 3 and 4, hsa-miR-374a-5p and 374b-5p tend to build up separated ceRNA clusters in each BC stage. As shown in Fig. 4, the KLF5 gene and TRAF3IP2-AS1 lncRNA, unique to stage IV, create a unique triplet with hsa-miR-153-3p.

The lncRNAs, genes and miRNAs list of significant ceRNA associations were combined for a single variable as "<lncRNA gene ensemble ID>_<gene ensemble ID>_<miRNA name>" (Ex: ENSG00000234456_ENSG00000125845_hsa-miR-374b-5p). The values of gene ensemble ID, lncRNA ensemble ID, and the derived variable columns were inserted into a four-set Venn diagram representation (indicating four BC stages) as Fig. 5a–c, respectively.

Figure 5
figure 5

Venn diagram representation for mRNAs (a), lncRNAs (b), and competing endogenous RNAs (ceRNAs) (c) included in stage-specific significant ceRNA associations. Blue, Red, green, and yellow-coloured sets represent stage I, stage II, stage III, and stage IV, respectively. Most of the mRNAs and lncRNAs have been shared among four stages. Five genes, KDR, SGCB, PRTG, KLF5, and PCNX1 and one lncRNA, TRAF3IP2-AS1, were observed only in stage IV, metastatic stage.

Full size image

According to Fig. 5, most lncRNAs, mRNAs, and lncRNA-mRNA-miRNA associations found in each BC stage ceRNA network have been shared among more than one stage. One lncRNA (TRAF3IP2-AS1) and five mRNAs (KDR, SGCB, PRTG, KLF5, and PCNX1) have been observed only in the stage-IV-specific ceRNA network. We identified 29 and 15 unique ceRNA associations in stage-III and stage-IV BC, respectively.

Functional enrichment analysis for stage-specific competing endogenous RNA networks

We conducted functional enrichment analyses on 25, 42, 40, and 47 genes obtained from the stage I, stage II, stage III, and stage IV ceRNA networks. None of the stage-specific genes was enriched in KEGG pathways24. Two genes resulting from stage I analysis were enriched in four CoA ligase activity-associated GO-molecular function (GO-MF) pathways. The GO-MF hormone-binding pathway was significant across stages II, III, and IV. In stage IV, six genes were enriched in three GO-cellular components (CC) pathways, membrane raft, membrane microdomain, and membrane region. Figure 6 illustrates GO pathway results for each BC stage.

Figure 6
figure 6

Pathway enrichment analysis results for genes included in stage-specific competing endogenous RNA networks in breast cancer20. Four, one, one, and four gene-ontology pathways were statistically significant in stages I, II, III, and IV, respectively. The CoA ligase activity-related pathways have been significant only in stage I. The hormone-binding pathway was significant among stages II, III, and IV.

Full size image

Survival analysis for stage-specific competing endogenous RNA networks

K-M survival analyses and log-rank tests were performed to identify the potential stage-specific differentially expressed genes, lncRNAs and miRNAs strongly correlated with BC patients' prognostic characteristics. The significant genes, lncRNAs and miRNAs, were chosen under p value < 0.05. We found two, three, three, and one gene(s) as prognostic biomarkers in stage I, stage II, stage III, and stage IV, respectively and described in Table 3. None of the lncRNAs was identified as biomarkers in stage-specific BC survival analysis. In miRNA-based survival analyses, two, three, and three miRNAs were statistically significant in stage I, stage II, and III, respectively. Figure 7 illustrates K–M curves for the top significant miRNA of each stage, hsa-miR-106b-5p in stage I, hsa-miR-31-5p in stage II, and hsa-miR-551b-3p in stage III.

Table 3 Statistically significant genes from stage-specific survival analyses.
Full size table
Figure 7
figure 7

Kaplan–Meier survival plots for the top significant miRNAs in stage-specific breast cancer networks21. The high-expressed hsa-miR-106b-5p (a), hsa-miR-31-5p (b), and hsa-miR-551b-3p (c) were prognostic in stage I, II, and III, respectively. None of the microRNAs in the stage IV ceRNA network was significant from survival analysis.

Full size image

Discussion

This study identified stage-specific lncRNA-mRNA-miRNA ceRNA associations in BC. According to ceRNA network analysis results, most ceRNA associations were shared across all four stages. In contrast, one lncRNA (TRAF3IP2-AS1) and five genes (KDR, PRTG, KLF5, SGCB, and PCNX1) were statistically significant only in metastatic BC, i.e., stage IV. The lncRNA TRAF3IP2-AS1 has been previously reported in renal cell carcinoma and glioblastoma but not in BC25.

According to AACR (American Association for Cancer Research) project GENIE (Genomics Evidence Neoplasia Information Exchange), the KDR gene is altered in 1.54% of BC patients. It can play an essential role in mediating endothelial cells proliferation, migration, and permeability26. Endothelial cells are actively involved in cancer metastasis27. The PRTG gene has not been previously described in BC. Nevertheless, PRTG has been identified as an oncogenic protein in gastric carcinogenesis by activating the downstream cGMP/PKG signalling pathway28.

The KLF5 gene is found to play a role in BC, but its precise function remains determined. On the one hand, the KLF5 locus at chromosome 13 is frequently deleted in human BC, and its protein is degraded by the WWP1 oncogenic ubiquitin E3 ligase, which suggests a tumour-suppressor function29. On the other hand, increased expression of KLF5 is associated with expression of the HER2 oncoprotein and shorter survival in BC patients suggesting an oncogenic function of KLF5 in BC30. A recent bioinformatic study has found that SGCB protein is specific in basal A subtyped BC gene regulatory networks31. The PCNX1 has been a potential marker of response to chemotherapy in BC, and therapeutic modulation of its activities could enhance chemotherapy responses32. The BC studies mentioned above have described four out of five genes found from our stage IV-specific ceRNA network, and none of these studies explains their contribution to metastatic BC. Therefore, wet-lab experiments will be carried out in the future to investigate their role in BC metastatic nature.

Two downstream analyses, functional enrichment analysis and survival analysis, have ensured stage-specific ceRNA components found in our study. All statistically significant genes resulting from stagewise survival analyses have been previously reported in cancer studies22. BMP2 from the stage I survival analysis has shown oncogenic function in BC33. Only DUSP6 and ACSL1 in stage II have acted as an oncogene and a tumour suppressor in BC, respectively34, 35. DST and AHNAK genes that were significant from the stage III survival analysis have shown tumour suppressive characteristics in both experimental and computational BC studies36,37,38,39,40. In stage IV, we identified MID1 as the only survival significant gene and it has been linked with invasive lobular carcinoma22. The invasive lobular carcinoma is the second most common type of BC. It originates in the milk-producing gland (lobules) of the breast. Invasive cancer is recognised as the cancer cells have broken out of the lobules where they initiated and are potential to expand to the lymph nodes and other areas of the body, leading to metastasis41. Therefore, MID1 gene should be further investigated to identify its role in metastatic BC (stage IV).

We found eight miRNAs from the stage-specific survival analyses, and these miRNAs have shown a tumour-suppressive/oncogenic role in previous BC experimental studies. The hsa-miR-106b-5p and hsa-miR-374b-5p were associated with the prognosis of stage I BC patients. The hsa-miR-106b-5p promoted cell migration, invasion, and proliferation by targeting FUT642. Abnormal hsa-miR-374b-5p expression in luminal-HER2-positive BC cells can be used for classifying clinicopathologic subtypes of BC43. Three miRNAs, hsa-miR-150-5p and hsa-miR-31-5p, are significant in stage II BC survival analysis, and hsa-miR-374b-5p was significant in both stages I and II. The other two miRNAs, hsa-miR-150-5p and hsa-miR-31-5p, have shown an oncogenic and tumour-suppressive role in BC, respectively44,45. We identified three miRNA signatures (hsa-miR-551b-3p, hsa-miR-101-3p, and hsa-miR-26a-5p) as prognostics in stage III BC survival. Among them, hsa-miR-551b-3p have promoted oncogenic features in BC cells46. In previous BC experimental studies, both hsa-miR-101-3p and hsa-miR-26a-5p have shown a tumour-suppressive role47,48. We did not find statistically significant prognostic miRNAs from the stage IV BC survival analysis.

Functional enrichment analyses found molecular pathways associated with protein-coding genes in stage-specific ceRNA networks. Two genes in Acyl-CoA Synthetase Long (ACSL) Chain family, ACSL1 and ACSL4, included only in stage I ceRNA networks, are enriched in four CoA ligase activity-associated pathways. Therefore, wet-lab experiments are required to understand the tumour-suppressive/oncogenic/cancer-driven role of ACSL Chain family members among early-stage BC patients. Four genes found in stages II, III, and IV (NR3C1, AVPR1A, LEPR, and THRB) are associated with hormone binding, which plays a role in BC pathophysiology and defining risk. Six genes in the stage IV ceRNA network are enriched in three components in the GO-CC pathway: membrane raft, membrane microdomain, and membrane region. These membrane domains have shown an important role in cancer metastasis49.

Our study has shared a limited set of ceRNAs with the previous ceRNA network study for overall BC cases by Tuersong et al.3. Four miRNAs (hsa-miR-141, hsa-miR-200a, hsa-miR-204, and hsa-miR-301b) have been identified in ceRNA networks in both studies. Among these four miRNAs, Tuersong et al.3 have demonstrated that hsa-miR-204 was downregulated and hsa-miR-301b was upregulated in patients with BRCA compared with healthy controls and were associated with overall survival. Our previous transcriptome-wide association study also demonstrated hsa-miR-204 as a tumour-suppressive miRNA in prostate cancer with statistically significant low expressed levels in prostate cancer cell lines50. Moreover, we found two genes, SPRY2 and CHL1, involved in Tuersong et al.3 and our works (SPRY2 in stages I–IV and CHL1 in stages II, III, and IV). Observing a smaller number of shared ceRNAs between studies can be occurred due to higher heterogeneity between breast cancer stages, and it will lead to different RNA/gene expression levels. Zhou et al. conducted a ceRNA network analysis on BRCA subtypes, basal-like, HER2+, luminal A, and luminal B51. The authors have identified three lncRNAs, NEAT1, OPI5-AS1, and AC008124.1, among all four subtype-related ceRNA networks. Moreover, three lncRNAs, NEAT1, FAM83H-AS1, and XIST1, were significantly differentially expressed in the basal-like subtype-related network. Nevertheless, we could not find a shared outcome between our study and subtype-related networks. This can be due to our stage-based analyses containing RNA/gene expression levels from multiple subtypes.

This study is limited to ceRNA networks mediated by microRNA expression levels. Other genomic (copy number alteration), transcriptomic (transcription factors), and epigenetic (DNA methylation) factors were not considered in the ceRNA network analysis52. Moreover, other possible ceRNA components such as pseudogenes and lincRNAs were not considered. Therefore, future studies should be extended to address these concerns. Nevertheless, this study elucidates a new level of ceRNA network analysis, stage-specific ceRNA networks, to understand better common/unique ceRNA(s) among/within the stage(s) of a given cancer. Identifying novel stage-level cancer biomarkers will significantly contribute to the knowledge of personalised therapeutics and determining risk.

Conclusions

We conducted ceRNA networks analyses in four stages of BC. Only one lncRNA and five genes were significant in the stage IV BC ceRNA network. Further validation experiments are required to characterise their role in BC metastatic nature. Identifying ceRNA components across cancer stages will advance the diagnosis, risk identification, and therapeutics.