The databases searched in our study are available online. We were free to use TCGA breast cancer databases from The Cancer Genome Atlas (https://tcga-data.nci.nih.gov) by meeting its freedom-to-publish criteria. All the Gene Expression Omnibus datasets are also accessible to public from the GEO website (www.ncbi.nlm.nih.gov/geo) and the ROCK breast cancer functional genomics (http://www.rock.icr.ac.uk), an integrated online breast cancer knowledgebase which provides and updates detailed microarray databases and information from reports on breast cancer (11).

Five breast cancer microarray databases from published studies were used to identify TNBC specific genes. Normalized gene expression profiling data were downloaded from TCGA and GEO websites, while the clinical information was extracted from corresponding reports, TCGA and ROCK websites. Samples were deposited into TNBC, NTNBC and normal groups according to molecular subtyping and immunohistochemical staining results, as presented in Table I.

The two further online clinical datasets were chosen to analyze the association of TNBC specific genes with breast cancer prognosis. NKI data were extracted from a previous study of van de Vijver et al (12), while 2006CC data from the study of Chin et al (13). The median age of patients in NKI and 2006CC data are 44 and 51, respectively, and each dataset contains patients of different lymph status and molecular subtypes. Except for microarray data, NKI dataset contains overall survival (OS), distant metastasis-free survival (DMFS), disease-free survival (DFS) and Nottingham Grading Score (NGS) information, while 2006CC dataset contains OS, DMFS, relapse-free survival (RFS) and TNM staging information.

Triple-negative subtype constitutes only a small fraction of all breast cancer cases, hence to investigate as many breast cancer patients as possible while avoiding cross-study batch effects, we separately carried out statistical analysis in each of these 5 databases (14). Significant analysis of microarray (SAM) method (p<0.05, q<0.25) and independent sample t-test (p<0.05) were used to compare the expression level of each gene between patients and healthy persons. Since the platform and corresponding gene sums of the five datasets were not in concert, to avoid batch effects between experiments while excluding as many heterogenetic genes as possible, statistical analysis was performed for each dataset, and the overlapping part was output. Overlapping genes upregulated >2-fold in patients while with statistical significance (p<0.05) were deposited into the primary set.

To ensure that members of this gene set were specifically overexpressed in TNBC, we compared corresponding TNBC expression data of genes in the primary set with NTNBC in each database via SAM method or independent sample t-test with the judgment criteria of p<0.05. A cohort of genes statistically upregulated in triple-negative breast cancer was generated and then annotated by DAVID (http://david.abcc.ncifcrf.gov/). The expression intensity of microarray data was demonstrated as heatmap by Mev 4.0 software.

In the five datasets applied to generate a 45-gene cohort, four GEO datasets contain no clinical follow-ups, while a large fraction of TCGA survival information were censored or still upgrading. For further analysis, we chose NKI and 2006CC clinical sets to investigate the association between TNBC cohort genes and breast cancer prognosis. The median of TNBC-upregulated cohort gene levels across samples were calculated to define the attribution of each sample. For each gene, patients with an expression level higher than median was classified as ‘high’, while those with a level lower than median was classified as ‘low’. The relationship between the level of each gene and patient survival [including disease-free survival (DFS), distant metastasis-free survival (DMFS), relapse-free survival (RFS) and overall survival (OS)] was described via Kaplan-Meier analysis in both clinical sets. The survival curves were drawn via GraphPad Prism 5, patients of ‘high’ or ‘low’ categories were displayed by curves of different colors, while censored value were displayed as small bars. Genes whose expression level statistically correlated with breast cancer survival in both clinical sets were defined as hub genes.

To investigate the role of hub genes in survival within different tumor periods, we divided the patients into subgroups according to tumor grade and stage. In two clinical sets applied in this study, NKI contains Nottingham grading score (NGS) information, patients in this set were classified as grade 1–3, while 2006CC contains TNM staging information, patients in this set were classified as stage I–II and stage III–IV. Expression level of each hub gene in each sample were defined as ‘high’ or ‘low’ by using median as the threshold as mentioned above. Kaplan-Meier analysis of hub gene-related survival was then performed.

To find out whether identified hub genes were indispensable factors influencing patient survival, we classified samples of both clinical sets into groups according to the level of hub genes. Samples with all hub genes highly expressed were defined as ‘triple-positive’, samples with all hub genes downregulated were defined as ‘triple-negative’, while those in the middle were further grouped. Kaplan-Meier analysis was made for each group.

To weigh the relative influence of the genes on prognosis, Cox proportional hazards regression analysis was performed in two clinical sets using OS and DMFS. Analysis was performed by SPSS 19.0. All the genes of interest were used as covariates, and the ‘Forward: LR’ method was used.

In order to further investigate the role of these survival-related genes in breast cancer, we constructed an interaction network using GNCpro online tools (15). Genes and pathways shown in this network were analyzed in NCBI and KEGG, and related reports were downloaded from PubMed.

Definitely molecular subtyped TNBC class covers only ∼12–17% of all breast cancer patients, exclusive of basal-like and BRCA1-related subtypes (5). To expand the sample pool as much as possible, we browsed all available public databases, and those with TNBC samples fewer than 5 were discarded. Only 5 databases were chosen, and the total number of TNBC, NTNBC and normal were 211, 2038 and 122, respectively (16) (Table I). The expression level of triple-negative samples versus normal in GSE3744 (17) and GSE47561 (11) data were compared based on SAM methods (18) with p<0.05 and q<0.25, but for TCGA, GSE2607 (19) and GSE1992 (20) data, independent sample t-test (p<0.05) was applied instead of SAM because of the existence of numerous databases. A primary set of 48 genes were found to be expressed ≥2-fold higher in TNBC than in healthy controls across the five databases. A heatmap for the 48-gene primary set from TCGA database is shown as Fig. 1.

With a judgment criteria of p<0.05, in all 48 members, a cohort of 45 genes was identified to be statistically upregulated in TNBC compared to NTNBC, 2 genes were of similar levels in TNBC and NTNBC, and 1 gene was downregulated in NTNBC, not in TNBC. These 45 TNBC specific genes were annotated to be related to important cellular mechanisms such as cell cycle regulation, centrosome function, enzymatic canalization and maintenance of chromosome integrity, all of which were potential breach points of carcinogenesis (Table II).

The poor clinical outcome and short survival of TNBC might be related to the overexpression of some TNBC specific genes (21). The association between cohort genes and patient survival were analyzed using NKI and 2006CC clinical sets according to procedures mentioned in Materials and methods. We found that among the 45 TNBC-upregulated genes, the overexpression of BIRC5, CENPA and FAM64A showed a significant correlation with poor OS and DMFS (p<0.05) despite the molecular subtypes of breast cancer (Fig. 2). High expression of these genes were also related with poor DFS and RFS. Therefore, these three genes were classified as hub genes.

Apart from BIRC5, CENPA and FAM64A, some other genes in the cohort were found to be co-expressed with these three hub members and were potentially involved in breast cancer survival as well. In these candidates, the level of BUB1 and CCNB2 genes were also negatively associated with clinical survival, though their relationship with OS in 2006CC data were not statistically supported by log-rank test as with DFS, DMFS or RFS.

According to available clinical information, patients in 2006CC were divided as stage I–II and stage III–IV, while patients in NKI clinical set were divided as grade 1–3. Kaplan-Meier analysis showed that patients of stage III–IV have poorer survival than stage I–II, while the rank of survival with NGS was grade 3>2>1, these results agreed with clinical observations. Furthermore, high level of BIRC5, CENPA and FAM64A were correlated with poor survival in stage I–II patients (Fig. 3). Since the number of advanced tumor samples was too few to perform convincible statistic analysis, we are not sure whether this relationship exists in stage III–IV patients. Similar correlation was also found in patients classified as grade 1–3 by Nottingham grading system.

To find out whether these three hub genes were indispensable factors influencing patient survival, we defined patients with all hub genes highly expressed as ‘triple-positive’ (TP), two genes highly expressed as ‘double-positive’ (DP), one gene highly expressed as ‘double-negative’ (DN), and all 3 genes of low level as ‘triple-negative’ (TN). Relationships of these 4 categories with survival were described by Kaplan-Meier analysis. We found that in NKI data, the TN group had the best OS, DMFS and DFS, while TP was of the worst OS but with similar level of DMFS/DFS as DP and DN groups. In 2006CC data, due to the sample number in each group being limited, survival of TP was similar with DP and NP, but TN still had the best OS, DMFS and RFS (Fig. 4). These results suggested that BIRC5, CENPA and FAM64A are all important molecules influencing survival in breast cancer. In other words, synchronous lower level of three hub genes indicates the best prognosis in breast cancer patients, while synchronous upregulation of three hub genes indicates the worst. Since three hub genes has been previously validated to be upregulated in TNBC subtypes, these findings suggested that the overexpression of BIRC5, CENPA and FAM64A may be the reason for the poor clinical outcome of TNBC, and they are potential therapeutic targets for triple-negative breast cancer.

To evaluate the independence of five TNBC-specific genes in inducing poor prognosis, Cox proportional hazards regression analysis was performed in NKI and 2006CC clinical sets according to OS and DMFS. For NKI data, BIRC5 and CCNB2 were considered to be independent prognostic markers for OS and DMFS, while for 2006CC data, BIRC5 and CENPA were considered as independent prognostic markers for OS and DMFS (data not shown), giving a hint that BIRC5 is the more important of these genes in the generation of TNBC high malignancy.

Pathway analysis indicated that BIRC5, CENPA and FAM64A were co-expressed with BUB1 and CCNB2 (Fig. 5). Among the 5 genes, the function of BIRC5, CENPA, BUB1 and CCNB2 have been well studied and proved to be involved with cancer development, while FAM64A is a relatively new molecule with fewer details from NCBI and a limited number of reports. FAM64A was reported to be highly expressed in cancer cell lines and associated with cancer cell cycle regulation and proliferation. In addition, the aurora kinase B gene (aka AURKB, whose protein product regulates chromosome segregation as a serine/threonine kinase) on the map has been confirmed to link BIRC5 and CENPA mechanistically, indicating that AURKB plays an important role among our TNBC-specific genes (Fig. 5).