A plasma 9-microRNA signature for lung cancer early detection: a multicenter analysis

Abstract

Lung cancer remains the leading cause of cancer-related deaths worldwide. Low-dose computed tomography (LD-CT) screening, combined with effective minimally invasive molecular testing such circulating microRNA, has the potential to reduce the burden of lung cancer. However, their clinical application requires further validation, including studies across diverse patient cohorts from different countries. In this study, we propose a signature of 9 circulating miRNAs derived from a robust multi-platform workflow with a multi-center design, for a total of 276 lung cancer and 451 non-cancer controls, based on the data from two European LD-CT screening cohorts (Poland and Italy). The classification performance of the signature was stable in the two screening cohorts, with AUC=0.78 (SE, 76%; SP, 67%; ACC=70%), and AUC=0.75 (SE, 82%; SP, 68%; ACC=71%) in the Polish and Italian cohorts, respectively. The diagnostic accuracy of the signature was remarkably independent of age, gender, smoking (status and intensity), nodule size, and density. Additionally, the signature demonstrated strong performance in detecting stage I lung cancer, with AUC=0.76 (95%CI: 0.68-0.84), and 0.69 (95%CI: 0.49-0.89) in the Polish and Italian cohorts respectively, with a prediction ability of 63-73%. The signature's ability to discriminate benign nodules was satisfactory, with AUC=0.71 (95%CI: 0.58-0.84). The proposed panel of 9 circulating miRNAs provides a robust and precise diagnostic tool to substantially advance the effectiveness of the LD-CT screening program.

Keywords: Early diagnosis; Liquid biopsy; Lung cancer; Machine learning; MicroRNA.

Fig. 1

A Schematic representation of the study. B Hierarchical clustering analysis of the 321 c-miRs commonly identified in all 3 c-miR expression datasets (GSE64591, GSE46729, GSE68951). Data (arrays) were median centered. Colors are as per the legend. C Volcano plot for the 321 c-miRs common to the 3 datasets GSE64591, GSE46729, GSE68951. Log2 fold-change and –log10 proportion of false positive (pfp) are reported, as per RankProd non-parametric method. Each dot represents one miRNA. Selected 45 c-miRs differentially expressed (pfp<0.05) are highlighted in red (upregulated) or in blue (downregulated) in tumor vs. normal samples; D ROC curves and AUC, for the model including 45 c-miRs and 36 c-miRs selected for the further testing (STEP-4), applied to the IARC dataset (GSE64591); E Hierarchical clustering analysis of the 45 c-miRs expression (data were median-array-centered) in the pools (N=6) of samples (N=108) collected at IRCCS Casa Sollievo della Sofferenza Hospital (CSS) and Humanitas Research Hospital (HUM). Colors are as per the legend. On the right, bubbles represent the different criteria applied (as per the legend) to identify the 36 c-miRs. Highlighted in green, the 29 c-miRs selected (reliable detection in Step 2 analysis), and in yellow, the remaining 7 c-miRs selected in Step 3 and GSE64591 analysis. In bold, has-miR-197-3p which is included in the panel of 6 previously identified housekeeping c-miRs

Fig. 2

A Hierarchical clustering analysis of the median 36 c-miRs and 13 c-miRs (external signature) expression on real-word multi-center LD-CT screening cohorts of high-risk subjects (Step 4 analysis). Colors are as per the legend. B ROC curves, AUC, and optimism-adjusted AUC (200 bootstrap) for the 9-c-miR model in the following cohorts: MUG screen-detected lung cancer (LC) and normal controls (N), HUM screen-detected lung cancer (LC) and normal controls (N), and IARC GSE64591 lung cancer (LC) and normal controls (N). C ROC curve and AUC for the 9-c-miR model applied to the HUM cohort: screen-detected lung cancer (LC) and benign (BEN). D Distribution of the probability of having lung cancer (tumor predicted probability) using the 9-c-miR model in MUG and HUM cohorts; green lines represent the mean values. E Distribution of tumor predicted probabilities using the 9-c-miR model in the various LC subtypes (adenocarcinoma, AC; squamous cell carcinoma, SCC; and other subtypes, Other); green lines represent the mean values. F ROC curves and AUC for the 9-c-miR model applied to stage I disease only in MUG and HUM cohorts