Predict ovarian cancer by pairing serum miRNAs: Construct of single sample classifiers

Guini Hong; Fengyuan Luo; Zhihong Chen; Liyuan Ma; Guiyang Lin; Tong Wu; Na Li; Hao Cai; Tao Hu; Haijian Zhong; You Guo; Hongdong Li

doi:10.3389/fmed.2022.923275

Predict ovarian cancer by pairing serum miRNAs: Construct of single sample classifiers

Front Med (Lausanne). 2022 Aug 2:9:923275. doi: 10.3389/fmed.2022.923275. eCollection 2022.

Authors

Guini Hong¹, Fengyuan Luo¹, Zhihong Chen¹, Liyuan Ma¹, Guiyang Lin¹, Tong Wu¹, Na Li¹, Hao Cai², Tao Hu¹, Haijian Zhong¹, You Guo², Hongdong Li¹

Affiliations

¹ School of Medical Information Engineering, Gannan Medical University, Ganzhou, China.
² Medical Big Data and Bioinformatics Research Centre, First Affiliated Hospital of Gannan Medical University, Ganzhou, China.

Abstract

Objective: The accuracy of CA125 or clinical examination in ovarian cancer (OVC) screening is still facing challenges. Serum miRNAs have been considered as promising biomarkers for clinical applications. Here, we propose a single sample classifier (SSC) method based on within-sample relative expression orderings (REOs) of serum miRNAs for OVC diagnosis.

Methods: Based on the stable REOs within 4,965 non-cancer serum samples, we developed the SSC for OVC in the training cohort (GSE106817: OVC = 200, non-cancer = 2,000) by focusing on highly reversed REOs within OVC. The best diagnosis is achieved using a combination of reversed miRNA pairs, considering the largest evaluation index and the lowest number of miRNA pairs possessed according to the voting rule. The SSC was then validated in internal data (GSE106817: OVC = 120, non-cancer = 759) and external data (GSE113486: OVC = 40, non-cancer = 100).

Results: The obtained 13-miRPairs classifier showed high diagnostic accuracy on distinguishing OVC from non-cancer controls in the training set (sensitivity = 98.00%, specificity = 99.60%), which was reproducible in internal data (sensitivity = 98.33%, specificity = 99.21%) and external data (sensitivity = 97.50%, specificity = 100%). Compared with the published models, it stood out in terms of correct positive predictive value (PPV) and negative predictive value (NPV) (PPV = 96.08% and NPV=95.16% in training set, and both above 99% in validation set). In addition, 13-miRPairs demonstrated a classification accuracy of over 97.5% for stage I OVC samples. By integrating other non-OVC serum samples as a control, the obtained 17-miRPairs classifier could distinguish OVC from other cancers (AUC>92% in training and validation set).

Conclusion: The REO-based SSCs performed well in predicting OVC (including early samples) and distinguishing OVC from other cancer types, proving that REOs of serum miRNAs represent a robust and non-invasive biomarker.

Keywords: Serum miRNA; early diagnosis; ovarian cancer; relative expression orderings; single sample classifier.