Identification of upstream regulators for prognostic expression signature genes in colorectal cancer

Abstract

Background: Gene expression signatures have been commonly used as diagnostic and prognostic markers for cancer subtyping. However, expression signatures frequently include many passengers, which are not directly related to cancer progression. Their upstream regulators such as transcription factors (TFs) may take a more critical role as drivers or master regulators to provide better clues on the underlying regulatory mechanisms and therapeutic applications.

Results: In order to identify prognostic master regulators, we took the known 85 prognostic signature genes for colorectal cancer and inferred their upstream TFs. To this end, a global transcriptional regulatory network was constructed with total >200,000 TF-target links using the ARACNE algorithm. We selected the top 10 TFs as candidate master regulators to show the highest coverage of the signature genes among the total 846 TF-target sub-networks or regulons. The selected TFs showed a comparable or slightly better prognostic performance than the original 85 signature genes in spite of greatly reduced number of marker genes from 85 to 10. Notably, these TFs were selected solely from inferred regulatory links using gene expression profiles and included many TFs regulating tumorigenic processes such as proliferation, metastasis, and differentiation.

Conclusions: Our network approach leads to the identification of the upstream transcription factors for prognostic signature genes to provide leads to their regulatory mechanisms. We demonstrate that our approach could identify upstream biomarkers for a given set of signature genes with markedly smaller size and comparable performances. The utility of our method may be expandable to other types of signatures such as diagnosis and drug response.

Figure 1

Overall pipeline of upstream regulator inference. (A) Global regulatory network modeling using ARACNE. (B) Regulon extraction for each TF. (C) Master regulator analysis (MRA) selects the TFs showing a significant overlap with the prognostic signature genes (D) Extraction of top 10 TFs by the signature coverage of MRA derived regulons (E) Stepwise linear regression (SLR) for edge filtering and extraction of top 10 TFs by the signature coverage of MRS + SLR derived regulons.

Figure 2

The transcriptional network between the top 10 TFs and the signature genes by MRA + SLR method. Node shape is triangular for TFs and circle for target signature genes. Node color represents the log2 ratio of gene expression between the poor and the goop prognostic group in the Moffit cohort (n = 177). Arrow shapes represent regulatory modes determined by the sign(+/−) of Spearman’s rank correlation between a TF and its target gene. Edge color represents the magnitude of correlation.

Figure 3

Correlation between the upstream TFs and their target genes. The average of the Spearman’s rank correlation coefficients was calculated between each of the 13 TFs (union of TF_MRA and TF_MRA+SLR) and the low risk (left) or the high risk (right) signature genes for (A) Moffit cohort and (B) Melbourne cohort.

Figure 4

Expression patterns of the selected marker genes between the good and the poor prognostic group. The distinct expression pattern of (A) 85 signature genes and of (B) 13 TFs (union of TF_MRA and TF_MRA+SLR) are shown in the Moffit court (n = 177, training dataset). Differential expression pattern is observed to be well maintained in an independent test dataset (Melbourne cohort, n = 95) for (C) the 85 signature genes, (D) TF_MRA, and (E) TF_MRA+SLR after class prediction.

Figure 5

The prediction performance of the selected prognostic markers. Kaplan-Meier plots for disease-free survival (DFS) are shown between the good and the poor prognostic group for (A) the 85 signature genes, (B) TF_MRA, and (C) TF_MRA+SLR. P-value for difference between two K-M plots was calculated by log-rank test.