Non-Coding RNAs as Prognostic Biomarkers: A miRNA Signature Specific for Aggressive Early-Stage Lung Adenocarcinomas

Abstract

Lung cancer burden can be reduced by adopting primary and secondary prevention strategies such as anti-smoking campaigns and low-dose CT screening for high risk subjects (aged >50 and smokers >30 packs/year). Recent CT screening trials demonstrated a stage-shift towards earlier stage lung cancer and reduction of mortality (~20%). However, a sizable fraction of patients (30-50%) with early stage disease still experience relapse and an adverse prognosis. Thus, the identification of effective prognostic biomarkers in stage I lung cancer is nowadays paramount. Here, we applied a multi-tiered approach relying on coupled RNA-seq and miRNA-seq data analysis of a large cohort of lung cancer patients (TCGA-LUAD, n = 510), which enabled us to identify prognostic miRNA signatures in stage I lung adenocarcinoma. Such signatures showed high accuracy (AUC ranging between 0.79 and 0.85) in scoring aggressive disease. Importantly, using a network-based approach we rewired miRNA-mRNA regulatory networks, identifying a minimal signature of 7 miRNAs, which was validated in a cohort of FFPE lung adenocarcinoma samples (CSS, n = 44) and controls a variety of genes overlapping with cancer relevant pathways. Our results further demonstrate the reliability of miRNA-based biomarkers for lung cancer prognostication and make a step forward to the application of miRNA biomarkers in the clinical routine.

Keywords: biomarkers; gene expression; lung cancer; microRNA; prognosis.

Figure 1

Flow chart of study design with data sets and analysis.

Figure 2

mRNA and miRNA expression profile analysis of the TCGA-LUAD cohort. (a) Hierarchical clustering analysis of the 10-gene expression signature. C1-C4 clusters are colored as per the legend. Age, gender, smoking status and stage are colored as per the legend; Unavailable information is colored in white. (b) Kaplan–Meier curves for 3-years overall survival stratified by C1–C4 clusters. Log-rank p-values are shown for C1 vs. non-C1 clusters (C2-C4) comparison. (c) Receiver operating characteristic (ROC) curves showing the False Positive Fraction and True Positive Fraction of the 19- (in blue) and 14-miRNA (in red) models. The areas under curve (AUC) are reported. (d) Networks of miRNA derived from 19-, 14- and 7-miRNAs model and corresponding target genes. Light blue rectangles represent genes; red rectangles represent miRNA from 19-miRNA model; yellow rectangles represent miRNA from 14-miRNA model; green rectangles represent miRNA from both 14- and 19-miRNA model. (e) Hierarchical clustering of 7-miRNAs in the TCGA-LUAD cohort. C1 and non-C1 tumors (defined according to the 10-gene signature) are colored as per the legend. Predicted C1 and non-C1 tumors (defined according the 7-miRNA logistic model) are colored as per the legend.

Figure 3

Validation of the 7-miRNA model. (a) ROC curve showing the False Positive Fraction and True Positive Fraction of the 7-miRNA model. The AUC is reported. (b) Box-plot for C1 predicted probability in C1 and non-C1 patients. Predicted probabilities are calculated through the 19-, 14- and 7-miRNA models. Wilcoxon–Mann–Whitney test p-values are reported. (c) Bubble plot of top 10 GeneSets found significantly overlapping with gene networks targeted by the 7-miRNA signature. Bubbles size is proportional to statistical significance (-Log of q-value) and color codes refer to number of genes found in the overlap. In X-axis, ratios (k/K) of overlap of the query set of genes (k) with overlapping GeneSet size (K). (d) Heatmap of the 10-gene expression of CSS cohort. C1 and non-C1 tumors are colored as per the legend. Risk scores are calculated based on the 10-gene risk model. (e) ROC curves showing the False Positive Fraction and True Positive Fraction of the 7-miRNA model in the CSS cohort, for all stages (in green) or only stage I tumors (in orange). The AUC are reported. (f) Box-plot for C1 predicted probability in C1 and non-C1 tumors in CSS cohort, for all-stages tumors and stage I tumors. Predicted probabilities are calculated through the 7-miRNA model. Wilcoxon–Mann–Whitney test p-values are reported.