Multi-analytical test based on serum miRNAs and proteins quantification for ovarian cancer early detection

Abstract

Advanced ovarian cancer is one of the most lethal gynecological tumor, mainly due to late diagnoses and acquired drug resistance. MicroRNAs (miRNAs) are small-non coding RNA acting as tumor suppressor/oncogenes differentially expressed in normal and epithelial ovarian cancer and has been recognized as a new class of tumor early detection biomarkers as they are released in blood fluids since tumor initiation process. Here, we evaluated by droplet digital PCR (ddPCR) circulating miRNAs in serum samples from healthy (N = 105) and untreated ovarian cancer patients (stages I to IV) (N = 72), grouped into a discovery/training and clinical validation set with the goal to identify the best classifier allowing the discrimination between earlier ovarian tumors from health controls women. The selection of 45 candidate miRNAs to be evaluated in the discovery set was based on miRNAs represented in ovarian cancer explorative commercial panels. We found six miRNAs showing increased levels in the blood of early or late-stage ovarian cancer groups compared to healthy controls. The serum levels of miR-320b and miR-141-3p were considered independent markers of malignancy in a multivariate logistic regression analysis. These markers were used to train diagnostic classifiers comprising miRNAs (miR-320b and miR-141-3p) and miRNAs combined with well-established ovarian cancer protein markers (miR-320b, miR-141-3p, CA-125 and HE4). The miRNA-based classifier was able to accurately discriminate early-stage ovarian cancer patients from health-controls in an independent sample set (Sensitivity = 80.0%, Specificity = 70.3%, AUC = 0.789). In addition, the integration of the serum proteins in the model markedly improved the performance (Sensitivity = 88.9%, Specificity = 100%, AUC = 1.000). A cross-study validation was carried out using four data series obtained from Gene Expression Omnibus (GEO), corroborating the performance of the miRNA-based classifier (AUCs ranging from 0.637 to 0.979). The clinical utility of the miRNA model should be validated in a prospective cohort in order to investigate their feasibility as an ovarian cancer early detection tool.

Fig 1. Flowchart showing the study design and main results.

The potential miRNAs (n = 45) and proteins (CA-125 and HE4) serum-markers were selected and their levels were assessed and compared in the serum of OC patients and healthy controls. Both tested proteins and six miRNAs (miR-10b-5p, miR-21-5p, miR-29c-3p, miR-141-3p, miR-222-3p, and miR-320b) were overrepresented in the serum of cancer patients compared to the controls in the Discovery Set. These markers were further tested increasing the sample size in a Training Set, where miR-141-3p, miR-320b, CA-125 and HE4 were considered non-redundant independent markers in a multivariate analysis (logistic regression). Two diagnostic classifiers were designed (CCP method), using only miRNAs and combining miRNAs with proteins. The diagnostic classifiers were applied in an independent group of samples (Validation Set), confirming its diagnostic potential, especially when the miRNAs were associated with proteins. The miRNA-based model was additionally applied in four publicly available data series (External Datasets), demonstrating the diagnostic power of the model. *performed as customized miRCURY LNA plates; #performed as individual assays; OC: ovarian cancer; CTRL: healthy controls; RPM: reads per million; TCGA: The Cancer Genome Atlas; Sens: sensitivity; Spec: Specificity; AUC: area under the ROC curve; CCP: Compound Covariate Predictor.

Fig 2. MicroRNAs and proteins that exhibited distinct levels in the serum of ovarian cancer against control cases (Discovery Set).

The plot displays the mean and 95% confidence intervals of the log₂ normalized relative quantification of the miRNAs (evaluated by miRCURY Custom ddPCR Assay) and proteins levels (measured by CMIA). For visualization purposes, the normalized miRNA levels were median-adjusted. Orange: controls (n = 26); Green: early stage (I-II) ovarian cancer (n = 12); Red: late stage (III-IV) ovarian cancer (n = 13). *P<0.05; **P<0.01; ***P<0.001 (P from Tukey post-hoc with Benjamini-Hochberg correction).

Fig 3. Training and validation of the ovarian cancer diagnostic models.

A. Application of the miRNA-classifier (miR-141-3p and miR-320b) and the miRNA/protein-classifier (miR-320b, miR-141-3p, CA-125 and HE4) in the Training Set (n = 115) and the resulted classification and AUC. B. Application of the same classifier in the Validation Set (n = 62). The dotted line in the dot-plot represents the threshold, above of which a malignant status would be predicted (low-risk prediction in green and high-risk prediction in orange). Sample size: 105 Controls (Training n = 68; Discovery n = 37); 30 Early stages (I-II) ovarian cancer (Training n = 20; Discovery n = 10); 42 Late stages (III-IV) ovarian cancer (Training n = 27; Discovery n = 15).*HE4/CA-125 levels are not available for four OC cases (1 early and 3 late stages). CCP: Compound Covariate Predictor; AUC: area under the ROC curve; CS: clinical stages. The dotted line in the ROC curve represents the random reference (AUC = 0.5).

Fig 4. Performance of the circulating miRNA model in the publicly available datasets.

A. Screening of appropriate datasets to be used in the cross-study validation step. Among eight studies found in the GEO datasets, four were eligible to test the 2-miRNA diagnostic model. B. The forest plot shows the AUCs and 95% confidence interval (CI) obtained for each study, all of them presenting the lower bound higher than 0.5.