Prognostic factor analysis for breast cancer using gene expression profiles

We obtained four independent publicly available breast cancer datasets for survival analysis. One is the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) in the United Kingdom and Canada [10]. These data were accessed through Synapse (synapse.sagebase.org, Synapse ID: syn1688369). The other three datasets were collected from the Gene Expression Omnibus (GSE25066 [11], GSE2034 [12], GSE3494 [13]). The METABRIC data set is used for training, and three GSE datasets (GSE25066, GSE2034, and GSE3494) are used for validation. To investigate genes related to triple-negative cancer, we used three breast cancer gene expression datasets: METABRIC, GSE2109 [14] and GSE25066. In the METABRIC dataset, normalized expression levels of a total of 54,675 probes were measured in 1981 breast tumor samples. Data were transformed to a compatible 24,924-gene format by selecting the median values of each probe expression. In the three independent GEO datasets, the normalized expression levels of 22,283 probes were measured in 508, 286, and 251 breast tumor samples. These data were also transformed to a 13,433-gene expression format by selecting the median value of each probe expression. Table 1 indicates used five microarray datasets in this study.

A total of 24,924 genes in METABRIC dataset were used in this research. To identify high/low expressed genes based on patient’s poor survival, we implemented a log-rank test and used an expression fold-change between patients who separated to first quartile and forth quartile corresponding to each gene expression level. This process was implemented by each gene. Hazard ratio was calculated between first and forth quartile patient groups and adjusted p-value cutoff was determined as 0.001. Therefore, if hazard ratio is greater than one with proper threshold and patients’ expression fold-change (first/fourth) is greater than 2, we selected the gene as a high-expressed gene in poor survival. Similarly, if hazard ratio is less than one with proper p-value cutoff and an expression fold-change (fourth/first) is less than 0.5, we selected the gene as a low-expressed gene in poor survival (Additional file 1: Figure S1). In the log-rank test of every 24,924 gene, we found 413 highly expressed genes associated with poor survival and 411 low-expressed genes associated with poor survival.

To construct the list of candidate genes for predicting patient’s outcome, we initially used over 20,000 genes and we selected a list of prognostic candidate genes by using a survival log-rank test. However, since too many number of genes showed significance in the log-rank test, we proposed an algorithm for minimizing and clustering genes according to their significance and co-expressed pattern. For clustering the two previously defined gene sets, we used the maximal clique algorithm [15] with Pearson correlation coefficient scores. Among the 413 high-expressed genes, we connected genes if two genes had a Pearson correlation coefficient which exceeded 0.4. We then determined the maximal clique in the 413 genes, after which we eliminated these genes and found the next maximal clique. Similarly, for the 411 low-expressed genes associated with poor survival, we also clustered genes with a minimum Pearson correlation coefficient of 0.4. To avoid the cluster which has too small number of genes, we used only two major clusters. Here, we used clusters for high/low expression gene sets which have over 15 independent genes. After clustering, we obtained two high-expressed gene groups associated with poor survival and two low-expressed gene groups. The connections between the high- and the low-expressed genes were also identified with a Pearson correlation coefficient of -0.4 through the maximal bi-clique generation algorithm [16]. Finally, there were four matched gene sets which are strongly connected to each other, as represented by high correlation values from the gene expression data. Each gene set has high-expressed and low-expressed genes associated with poor survival. Thus, we identified four prognosis prediction scores as the ratio between the median of the high-expressed gene level to the low-expressed gene level in the four matched gene sets. We defined the module 1 score as the ratio of the 26 high-expressed genes associated with poor survival to the 17 low-expressed genes associated with poor survival. Similarly, Modules 2, 3 and 4 scores were respectively defined as the ratios between the eight, nine, and four high-expressed genes associated with poor survival to the 10, nine, and eight low-expressed genes associated with poor survival. Because we used the maximal clique algorithm to cluster each gene set, there was a strong correlation between the expression levels of each high-expressed gene and low-expressed gene associated with poor survival (Pearson’s r > 0.4). Between high-expressed genes and low-expressed genes, the maximum Pearson correlation coefficient was found to be -0.4.

We analyzed three sets of detailed clinical data from each of the studies used. These were GSE2034, GSE25066, and GSE3494. We used the Disease-Free Survival (DFS) clinical information in GSE2034 and GSE25066, and the Distant Recurrence Free Survival (DRFS) in GSE3494. In a Kaplan Meier survival plots, the median of a measured module’s score was used to dichotomize the data, allowing stratification into high and low groups within each of the three individual datasets.

To investigate genes related to triple negative breast cancer (TNBC), after comparing the three independent expression profiles, we selected 230 up-regulated genes and 237 down-regulated genes in TNBC (Cut off p-value < 0.05, FDR < 0.05, from t-test, log fold change < 0.5) from METABRIC, GSE2109 and GSE25066 datasets.

Patients whose tumors had the highest score among the four modules had the worst prognosis. To validate each score, we used the three datasets of GSE2034, GSE25066, and GSE3494. We selected four matched gene sets from METABRIC data set which consist of about 20,000 genes on Illumina HT 12v3 platform. However, in the test datasets, the expression profiles consist of about 12,000 genes on Affymetrix HG U133A platform. Therefore, all genes obtained from METABRIC were not matched in test datasets. In module 1, among 44 METABRIC genes, 37 genes were used for validation. In module 2, module3 and module 4, we used 11, 27 and 8 matched genes among 17, 36, 14 METABRIC genes, respectively. According to Kaplan Meier survival plots of the three independent sets, high-scoring patients had poor survival rates in the scores of all of the modules. In the GSE2034, GSE25066 and GSE3494 datasets with module 1, patients dichotomized by the prognosis prediction score from 23 high-expressed and 14 low-expressed gene expressions were associated with the worst survival prognosis (Fig. 2a-f). Patients whose tumors had high scores on module 1 had the worst prognosis (P = 0.0036, P < 0.0001, and P < 0.0001, respectively). With modules 2, 3, and 4, similarly, patients whose tumors had high scores had the worst prognosis. In GSE2034, patients whose tumors had high module 1, 2 and 3 scores had the worst. In GSE3494, patients whose tumors had high module 1, 2 and 3 score had the worst prognosis. Only the cases of GSE2034 and GSE3494 with module 4 were not significantly different (Additional file 1: Table S1). We also investigated a lot of possible cases of prediction for differently matched gene set according to manifold threshold. The significance of those results was represented in Additional file 1: Table S1.

Among the four modules in this research, the module 1 signature strongly represented a high hazard ratio with significant p-values (Fig 2a-c). Figure 3a shows the expression pattern of the 28 and 17 prognostic factor genes. The 28 genes in the upper part of the figure represent high-expressed genes associated with poor survival, and these genes are co-expressed relative to each other (Pearson’s r > 0.4). The 17 genes in the lower part of the figure represent low-expressed genes associated with poor survival; these genes are also strongly co-expressed relative to each other (Pearson’s r > 0.4). In METABRIC, when gene expressions were compared across different tumor types without dichotomization, the module 1 gene score was significantly higher in the HER2-enriched group with the basal-like type (Fig. 3b). Similarly, the triple-negative breast cancer type, ER-negative, and the TP53 mutation type had higher scores than the other breast cancer types (Fig. 3c and d). The significance levels of the genes used in module 1 were also confirmed using BreastMark [17], which identifies putative prognostic biomarkers. BreastMark gave hazard ratios which were statistically significant for 41 out of 44 genes including previously defined prognostic genes (Table 2, Additional file 1: Table S2). In 26 high-expressed genes in module 1, 16 genes are previously defined as prognostic genes, and among 17 low-expressed genes, 8 genes are previously defined as prognostic genes (Additional file 1: Table S2). 26 genes out of module 1 were differentially expressed genes (DEGs) in TNBC and non-TNBC (19 upregulated genes and 11 down-regulated genes in TNBC) (Table 2).