A ten-microRNA signature identified from a genome-wide microRNA expression profiling in human epithelial ovarian cancer

Abstract

Epithelial ovarian cancer (EOC) is the most common gynecologic malignancy. To identify the micro-ribonucleic acids (miRNAs) expression profile in EOC tissues that may serve as a novel diagnostic biomarker for EOC detection, the expression of 1722 miRNAs from 15 normal ovarian tissue samples and 48 ovarian cancer samples was profiled by using a quantitative real-time polymerase chain reaction (qRT-PCR) assay. A ten-microRNA signature (hsa-miR-1271-5p, hsa-miR-574-3p, hsa-miR-182-5p, hsa-miR-183-5p, hsa-miR-96-5p, hsa-miR-15b-5p, hsa-miR-182-3p, hsa-miR-141-5p, hsa-miR-130b-5p, and hsa-miR-135b-3p) was identified to be able to distinguish human ovarian cancer tissues from normal tissues with 97% sensitivity and 92% specificity. Two miRNA clusters of miR183-96-183 (miR-96-5p, and miR-182, miR183) and miR200 (miR-141-5p, miR200a, b, c and miR429) are significantly up-regulated in ovarian cancer tissue samples compared to those of normal tissue samples, suggesting theses miRNAs may be involved in ovarian cancer development.

Figure 1. Comparison between 45 ovarian cancer (39 epithelial carcinoma and 6 borderline tissues) and 14 normal ovarian tissues.

A) The plot of miRNA assays used to profile compared samples: fold change (y-axis) against normalized Ct measurements. B) Volcano plot of the resulting p-values of the t-test (y-axis) between the C and N groups. 305 miRNAs show FDR-adjusted p-values below 0.1 and fold change above 2 (shown in red). C) Hierarchical clustering (R package pvclust) of the 45 ovarian cancer tissues and 14 ovarian normal tissues based on top 50 most variable miRNA assays. For each cluster in hierarchical clustering, quantities called p-values (approximately unbiased p-value (red) and Bootstrap Probability p-value (green)) are calculated via multi-scale bootstrap resampling. P-value of a cluster is a value between 0 and 1, which indicates how strong the cluster is supported by data. D) 17% of the variance observed in the Ct measurements of top 50 most variable miRNA assays across all samples can be explained by sample pathology status (C or N). The remaining covariates considered here (source = hospital source, survival, tumor histology, FIGO stage, tumor grade, relapse, and stage) explain 24% of the variance.

Figure 2. Determination of error rates by leave-one-out cross validation vs. number of markers and miRNA markers overlapping.

A) Error rate produced by different classification algorithms as a function of the number of prediction markers used. Leave-one-out cross-validation procedure was used to estimate resulting error rates. B) Percent overlapping of predictor miRNA selected from the different training sets of samples used. C) Leave-one-out cross validation results: each sample class probability (y-axis) is estimated based on SVM model learned from all other samples. Tissues (cancer black, normal red) are represented by classification probability of being cancer. D) ROC curve based on leave-one-out cross validation results using SVM method.

Figure 3. Ovarian cancer classification performance for the 10 selected miRNAs ( Table 2 ) using SVM algorithm and leave-one-out cross-validation.

A) Prediction probabilities (ovarian cancer) for each sample used in this study (C = Cancer; N = Normal). B) ROC curve. C) Prediction error across different tissue groups: normal tissue (N), epithelial carcinomas tissues (CE) and borderline tissues (CB). D) Prediction error within tumor grade groups [Increased error in lower grade samples (but not significant)].