Background: Mesothelioma is an aggressive cancer with a poor prognosis, partly due to the difficulty of distinguishing it from other thoracic diseases. Definitive diagnosis typically requires an invasive tissue biopsy, which is often unsuitable because many patients are diagnosed at an advanced age or with advanced disease. Circular RNAs (circRNAs) are stable, abundant in blood, and have strong potential as disease-specific biomarkers. Building on our cell-line findings, we evaluated the diagnostic performance of circRNAs to develop a minimally invasive liquid biopsy for mesothelioma.
Methods: We quantified the expression of eight circRNAs using droplet digital PCR (ddPCR) in plasma samples from 119 patients with mesothelioma, 104 with other cancer types and 54 from healthy donors, and tumour tissue biopsies from 94 patients with mesothelioma and 58 with other cancers. Diagnostic performance was evaluated using receiver operating characteristic (ROC) area under the curve (AUC) analysis for combinations of one to eight circRNA biomarkers. Two complementary approaches were compared: (1) logistic regression (GLM) followed by ROC analysis using the Mann–Whitney statistic, and (2) an exhaustive grid search to identify optimal biomarker threshold combinations. Leave-one-out cross-validation (LOOCV) and bootstrap resampling were used to assess model robustness and estimate uncertainty. Three ddPCR normalisation strategies were evaluated: (1) housekeeping gene, (2) total RNA input, and (3) ratio-based normalisation using other circRNA measurements. UMAP dimensionality reduction analysis was used to explore the biological relationships and discriminative ability of the circRNA signatures.
Results: The GLM AUC approach produced higher AUCs in both plasma and tissue compared to the simpler threshold-based method. Although the threshold-based approach is conceptually simpler for clinical interpretation, our plasma analyses showed that the same biomarker could be associated with either increased or decreased expression depending on the combination of biomarkers. In tissue, AUCs exhibited minimal variability, indicating robust and consistent performance. In plasma, LOOCV AUCs were lower, indicating overfitting due to higher noise levels in the plasma data. UMAP analyses demonstrated clear separation between mesothelioma and other conditions in tissue, but not in plasma. The candidate housekeeping gene itself exhibited diagnostic discrimination and was therefore unsuitable as a normalisation control. Normalisation by total RNA input was strongly influenced by differences between cell-rich tissue and circRNA-poor plasma samples.
Conclusion: Logistic regression is the more suitable method to capture diagnostic information lying in the joint biological relationships between multiple biomarkers. Ratio-based normalisation produced separation comparable to RNA input normalisation, suggesting that either approach is suitable.