A four-long non-coding RNA signature in predicting breast cancer survival

Breast cancer gene expression data and corresponding clinical data used in this study were obtained from the publicly available GEO databases. To analyze the correlation of lncRNA expression signatures with survival endpoints for breast cancer as a whole (disease-free survival, metastasis-free survival and overall survival), we selected those data sets that included more than 100 patients with their survival status information. We followed the strategy of using the largest data set (GSE20685) as training set. This training set from GSE20685 [29] was first used to identify the gene expression signature. Another three independent data sets from GSE12276 [30], GSE21653 [31],[32], GSE42568 [33] were included in this study as testing sets. After filtering out samples without clinical survival information, there were a total of 887 samples, including 327 from GSE20685, 252 from GSE21653, 204 from GSE12276, and 104 from GSE42568, respectively. Figure 1 depicts the diagram of the study.

The raw CEL files were downloaded from GEO database and background adjusted using Robust Multichip Average. GATExplorer [22] was used to process microarrays on a local computer for gene expressions of lncRNAs. This GATExplorer provides a series of R packages, designed to be used with BioConductor tools, that allow to apply in a simple way the probe mapping data included in GATExplorer. A type of files called ncRNA Mapper were also obtained from GATExplorer, which include the probes that do not map to any coding region but that were mapped to a database for non-coding RNA of human and mouse (derived from RNAdb [34]). A customized R scripts was used to perform a microarray expression calculation according to the re-mapping data(file ncrnamapperhgu133plus2cdf_3.0). Each LncRNA should include at least a minimum of 3 probes mapping in the corresponding ncRNAs entity. All of the four lncRNAs were verified online in the ncRNA Expression Database (nred.​matticklab.​com) [35], which provides gene expression information for thousands of long ncRNAs in human and mouse (Additional file 1: Table S1). We created a risk-score formula according to the expressions of these four lncRNAs for survival prediction. Patients having higher risk scores are expected to have poorer survival outcomes. The risk scores are calculated as follows: Risk score = (-0.35717× expression value of AK024118) + (0.518242 × expression value of U79277) + (-0.48664 × expression value of BC040204) + (-0.48122 × expression value of AK000974). In addition, the coding potential analysis of the lncRNAs was carried out by CNCI to classify protein-coding or noncoding transcripts [36].

GSEA was performed by the JAVA program (http://​www.​broadinstitute.​org/​gsea) using MSigDB C2 CP: Canonical pathways gene set collection(1320 gene sets available). Gene sets with a false discovery rate(FDR) value <0.05 after performing 1,000 permutations were considered to be significantly enriched [37]. Cytoscape and Enrichment Map were used for visualization of the GSEA results.

The association between the lncRNA gene expression and patient’s survival was assessed by univariable Cox regression analysis along with a permutation test using Biometric Research Branch-Array tools package [38] in the training set. With a parametric test(p ≤ 0.001), we identified a set of 30 lncRNA expressions strongly correlated with survival. Considering that a smaller number of genes in the model would make the model more practical, we then performed the random survival forests-variable hunting (RSFVH) algorithm [39]. We followed Kawaguchi [39],[40] for the parameters in the algorithm. In brief, the number of Monte Carlo iterations (nrep) was set as n_rep = 100 and value controlling the step size used in the forward process (nstep) was set as n_step = 5. A set of 4 lncRNAs genes have been identified in which expressions were strongly and consistently related to patient survival (Figure 2).

Using these four genes selected fitted in a multivariable Cox regression model, we constructed a formula that would predict survival in the training set. Each patient was then assigned a risk score that is a linear combination of the expression levels of the significant lncRNAs weighted by their respective Cox regression coefficients [41],[42]. According to this risk score, patients in the training set were divided into low-risk and high-risk groups using the median risk score as the cut-off. The Kaplan-Meier method was used to estimate survival time for the other three testing groups. Differences in survival times between the low-risk and high-risk groups in each set were then compared using the two-sided log rank test. Furthermore, we also used Cox multivariate analysis to test whether the risk score was independent of patient age and subtype with the available data. Oncotype DX score was implemented in an R package called genefu, available from the Comprehensive R Archive Network. We used receiver operating characteristic (ROC) curves to compare the sensitivity and specificity of the survival prediction of the lncRNA risk score. Area under the curve (AUC) values were calculated from the ROC curves [43]. All the data were analyzed by R program (www.​r-project.​org). The significance was defined as p values being less than 0.05.

As summarized in the workflow (Figure 1), all analyses were performed in the training data set (GSE20685) first and then validated in the test data set (GSE21653, GSE12276, GSE42568). The training set(n = 327) was analyzed for the detection of prognostic lncRNA genes. By subjecting the lncRNA expression data derived from the training set to univariable Cox proportional hazards regression analysis using the BRB-Array Tools, we identified a set of 30 lncRNAs that were strongly correlated with patients’ overall survival (P ≤0.001). On the basis of the random survival forests model (see Materials and methods), four genes were selected as the predictors. Table 1 shows a list of these four genes with their obtained variable importance values. As depicted in Figure 2, from the plot we can see that AK024118 has dramatically larger importance value than other predictors. Of these, a positive coefficient of U79277 indicated that its higher level of expression was associated with shorter survival. The negative coefficients of the other genes (AK024118, BC040204, AK000974) indicated that their higher levels of expression were associated with longer survival. All of the four lncRNAs have been verified in the ncRNA Expression Database (nred.​matticklab.​com) and these four transcripts were classified as ncRNAs in this website [35]. As coding potential analysis is commonly used to classify whether a transcript is of coding potential or not [25], we also used another tool, CNCI, developed by Sun et.al to test those four transcripts [36]. This tool also suggests that all the four transcripts are non-coding transcripts with no coding potential.

With the risk score formula (see Materials and methods), we calculated the four-lncRNA expression signature risk score for each patient in the training set. The patients were then ranked according to their risk scores. Using the median risk score as cut-off in the training set, the patients were divided into low-risk(n = 164) and high-risk (n = 163) groups. Patients in the high-risk group had significantly shorter overall survival than those in the low-risk group (log-rank test P < 0.0001) (Figure 3A). Overall survival in the training set was 96.95% at 3 years, 91.89% at 6 years, 88.08% at 9 years and 84.86% at 12 years in the low risk group, versus 83.99%, 70.27%, 62.02% and 51.56% in the high risk group respectively. The correlation of the four-lncRNA risk score with overall survival was significant when it was analyzed as a continuous variable in the univariable Cox regression model.

In order to confirm our findings, we calculated the risk score for the testing sets including GSE21653(n = 252), GSE12276 (n = 204) and GSE42568(n = 104). By using the same cut-off value as the training set, the patients from each testing set were separately classified into low-risk and high-risk groups and subjected to survival comparison. As overall survival information was unavailable in GSE21653 and GSE12276, disease-free survival(DFS) and metastasis-free survival(MFS) was evaluated, respectively. Similar to the findings obtained from the training set, patients in the high-risk group had shorter survival time than patients in the low-risk group (Figures 3B, 3C, 3D). In consistence with the results described above, patient survival in the low-risk group was better than that in the high-risk group throughout the follow-up. In the univariable Cox regression model, the similar correlation of the risk score with overall survival was noted with the high-risk group having a shorter overall survival than the low-risk group. The distribution of patient risk scores (Z-score transformed), survival status and lncRNA values were analyzed independently for the training set (Figure 4). We found that patients with high-risk scores tended to have higher expression of U79277 and lower expression of the remaining genes(AK024118, BC040204, AK000974). Detail survival information of individual lncRNA in each data set and the gene signature in the context of different tumor subtypes was shown in Additional file 2: Figure S1 and Additional file 3: Figure S2, respectively.

We carried out Cox multivariate analysis to ascertain whether the four-lncRNA expression signature was an independent predictor of breast cancer patient’s survival. Four-lncRNA risk score, age (available in GSE20685, GSE21653 and GSE42568) and subtype(only available in GSE21653) were defined as covariates. The effect of risk score, age and subtype on breast cancer patient survival time was further evaluated by multivariate Cox proportional hazard model. The results showed that risk score is an independent predictor of patient survival when adjusted by age or subtype in every cohort (Table 2).

As GSE21653 was the only data set with disease-free survival information(DFS), we performed receiver operating characteristic (ROC) analysis to compare the sensitivity and specificity of survival prediction between our model and Oncotype DX [13]. The area under receiver operating characteristic (AUROC) was determined and compared between these two gene signatures. As seen in Figure 5, ROC curves indicated that AUROC of four-lncRNA gene signature and Oncotype DX was 0.603 and 0.675, respectively. No significant difference(p=0.0837) was observed between the Oncotype group and the four-lncRNA gene signature group in terms of disease-free survival (DFS) (Figure 3).

Gene Set Enrichment Analysis(GSEA) was carried out to identify associated biological processes and signaling pathway [37]. We compared the gene expression profile of breast cancer patients with high-risk and low-risk group classified by four-lncRNA gene signature in the training set(GSE 20685). The gene sets with significantly different expression (FDR < 0.01, p <0.005) were picked up for Gene set enrichment analysis (GSEA). Several cancer related pathways such as epithelial mesenchymal transition(EMT) [44], cell cycle signaling and DNA replication [45] were enriched in the high risk group, which implies that the signature might be involved in the metastasis related pathways (Figure 6). The associated biological pathway with each lncRNA was shown in Additional file 4: Figure S3.