Identification of potential key genes and pathways predicting pathogenesis and prognosis for triple-negative breast cancer

The mRNA expression profiles and corresponding clinical information of breast cancer patients were downloaded from the Cancer Genome Atlas (TCGA) and gene expression omnibus (GEO) databases. We collected 1109 samples with gene expression data, containing 1109 BC tumour tissues samples and 113 normal tissue samples from TCGA database. After removing patients with incomplete information, we were left with117 TNBC samples and 970 non-TNBC samples. We collected 270 samples with 58 normal breast tissue samples and 212 TNBC tissue samples from the GEO dataset of the NCBI GEO database (GSE31519, GSE9574, GSE20194, GSE20271, GSE45255, and GSE15852).

Gene Ontology (GO) analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of DEGs were performed using Database for Annotation, Visualization and Integration Discovery, DAVID version 6.8 [16]. P < 0.05 was chosen as the cut-off criterion. GO is a set of unified vocabulary to describe molecular functions (MF), biological processes (BP) and cellular components (CC) of biology, whereas KEGG analysis was performed to aid understanding of the signalling pathways involving DEGs.

A PPI network of differential genes was constructed, using STRING version 10.5 to evaluate information on protein–protein interactions [17]. Using the Molecular Complex Detection (MCODE) plug-in in Cytoscape 3.7.0, a visualization tool for integrating many molecular states such as expression level and interaction information into a unified conceptual framework [18], the PPI network module with densely connected regions was obtained (Degree cut-off > 15) [19].

Clinical characteristic information for breast cancer was downloaded from TCGA. After removing samples with incomplete clinical overlapping DEG data, samples from 117 TNBC patients were used for further analysis. Univariate and multivariate Cox model analyses were used to identify candidate genes that were significantly associated with overall survival (OS). Based on the expression level and coefficient (β) of each gene, calculated by multivariate Cox proportional hazards regression analysis, a novel reliable prognostic gene signature was established. These TNBC patient samples were further divided into low- or high-risk groups based on the median risk score as the cut-off point. Kaplan–Meier curves were used to assess the prognostic value of the risk score. In addition, a time-dependent receiver operating characteristic (ROC) curve analysis, using the R package “survivalROC” was constructed to assess the predictive accuracy of the gene signature for time-dependent cancer death [20]. The area under the curve (AUC) was calculated to evaluate the predictive ability of the gene signature for clinical outcomes.

We used the “EDGR” and “Volcano” packages in the R software to identify differentially expressed genes between 1109 breast cancer tissue samples and 113 normal tissue samples from TCGA database (|logFC| ≥ 2 and adjusted P < 0.05), and screened out2816 up-regulated and 1095 down-regulated genes (Fig. 1a). We further analysed the DEGs between 117 TNBC and 970 non-TNBC breast cancer samples (|logFC| > 1 and adjusted P < 0.05), and identified a total of 1557 up-regulated genes and 2972 down-regulated genes (Fig. 1b). In addition, we used the Venn diagram web-tool (http://​bioinformatics.​psb.​ugent.​be/​webtools/​Venn/​) to cross the two sets of differential genes and found 755 overlapped DEGs (Additional file 1: Table S1), including 590 up-regulated genes (Fig. 1c) and 165 down-regulated genes (Fig. 1d).

GO function and KEGG pathway enrichment analysis were performed using DAVID to expound the biological functions of 755 DEGs (Additional file 2: Table S2). The BP results indicated that DEGs were mainly significantly enriched in mitotic nuclear division, sister chromatid cohesion, cell division (Fig. 2a). MF analysis showed that DEGs were significantly enriched in microtubule motor, chemokine and structural molecule activities (Fig. 2b). CC analysis showed that the DEGs were mainly enriched in the extracellular region, chromosome centromeric region and kinetochore (Fig. 2c). In addition, the most enriched KEGG pathways were PPAR signalling, AMPK signalling and oocyte meiosis pathways (Fig. 2d).

Protein interactions among overlapping DEGs were predicted with STRING tools. A total of 148 nodes and 477 edges were displayed in the PPI network (Fig. 3) with PPI-enrichment P value < 1.0e−16. A PPI network for DEG subsets with a combined score > 0.9 was constructed to determine the candidate hub genes. Based on the PPI network of the subsets, a module with an MCODE score of 42 and 45 nodes was identified (Fig. 4a), and functional enrichment analyses showed that the genes in this module were mainly associated with the cell cycle and mitosis (Fig. 4b and Table 1). BP analysis showed that these genes were significantly enriched in microtubule-based movement, mitotic sister chromatid segregation, mitotic metaphase plate congression, cell division, and mitotic cytokinesis. For CC analysis, these genes were significantly enriched in the condensed nuclear chromosome outer kinetochore, kinetochore, and spindle midzone. MF analysis showed the genes were significantly enriched in ATP binding, microtubule motor activity, single-stranded DNA binding, and DNA replication origin binding. In addition, the results of KEGG pathway enrichment analysis suggested that the pathways were enriched as follows: cell cycle, progesterone-mediated oocyte maturation, and oocyte meiosis. As a result, the eight genes correlated with cell cycle were selected as hub genes, which were CCNA2, CCNB2, CDC20, BUB1, TTK, CENPF, CENPA and CENPE (Table 2). Their expression levels were validated in 117 TNBC samples and 113 normal controls with breast cancer mRNA data from TCGA. As shown in Fig. 5, the eight mRNAs were significantly increased in TNBC compared with 113 normal control tissues (P < 0.001). We validated on the GEO database that the eight mRNA were also significantly increased compared with normal control tissues in TNBC (P < 0.001) (Additional file 3: Fig. S1).

Using Cox proportional hazards regression model, we analysed the genes in the module, but no significant gene signature was established to predict overall survival.

A total of 16 out of 755 DEGs were significantly correlated with survival time (P < 0.05) and identified by the univariate Cox proportional hazards regression model (Additional file 2: Table S3). Additionally, a prognostic gene signature, composed of six genes, was developed after using the multivariate Cox proportional hazards regression model. The genes include transmembrane protein 252 (TMEM252), collagen type IX alpha 3 chain (COL9A3), proline rich protein BstNI subfamily 2 (PRB2), single-pass membrane protein with coiled-coil domains 1 (SMCO1), involucrin (IVL), and submaxillary gland androgen regulated protein 3B (SMR3B) (Table 3). Patients were divided into low- and high-risk groups by the median risk score (1.070) (risk score = expression of SMR3B × 1.2141 + expression of TMEM252 × 1.6187 + expression of PRB2 × 1.4416 + expression of PRB2 × 2.0147 + expression of SMCO1 × 1.1471 + expression of COL9A3 × − 0.6101). The six-gene-based risk score distribution was presented in Fig. 6a. A highly significant difference in overall survival (OS) was detected between high- and low-risk groups (P < 0.0001) as shown in Fig. 6b. Moreover, the survival rate of the high-risk group was significantly much lower than for the low-risk group as depicted by Kaplan–Meier analysis in Fig. 6c (P < 0.0001). Time dependent ROC curve revealed that the prognostic signature presented a good performance in survival prediction, as shown in Fig. 6d and that the AUC was 0.929for 3 years OS and 0.902 for 5 years. Expression levels of the six genes in low- and high-risk groups are shown in Fig. 6e.

Using univariate and multivariate Cox regression analyses, we investigated whether the prognostic values of the six mRNA were independent of clinicopathological factors. Univariate Cox regression model showed that the risk score, race, TNM stage, N status, M status, tumour status, and radiation were significantly related to the patients’ overall survival in patients with TNBC (Table 4). In addition, multivariate Cox analysis indicated that the risk score and N stage still had remarkable independent prognostic values, with P = 0.005 and 0.025, respectively (Table 4). These results indicate that the 6-mRNA risk score was an independent prognostic indicator that can effectively predict the prognosis of TNBC patients.