The Journal of Liquid Biopsy · Published 2026-07-17 · DOI 10.1016/j.jlb.2026.100483
Cong Lai, Yelisudan Mulati, Xin Huang, Zhenhong Chen, Zhuohang Li, Muhammad Usman, Qiliang Zhai, Jiasi Wang, Huasheng Huang, Cheng Liu, Wang He, Kewei Xu
Background: Avoidable prostate biopsies remain a persistent weakness of prostate specific antigen (PSA)- and imaging-led prostate cancer (PCa) diagnosis. The key need is a non-invasive test that improves pre-biopsy risk stratification while remaining potentially translatable to clinical deployment. We developed and validated an end-to-end liquid-biopsy pipeline linking serum miRNA markers, routine clinical variables, machine learning, and CRISPR/Cas13a-based detection. Methods: Candidate miRNAs were prioritized from GSE112264 by differential expression, Logistic Regression, and Least Absolute Shrinkage and Selection Operator analyses, cross-referenced with PCa tissue expression, and measured by qPCR in 712 biopsy-scheduled participants from Sun Yat-sen Memorial Hospital (SYSMH), Houjie Hospital of Dongguan (HHD), and Ganzhou People's Hospital (GPH). A three-miRNA PCa risk score (PCaRS) was trained in SYSMH and tested in internal, external, and prospective cohorts. PCaRS and independent clinical predictors were integrated using six machine-learning algorithms; the optimal model was selected by receiver operator characteristic and DeLong analyses. Finally, serum miRNAs in the prospective SYSMH-Pro cohort were quantified with polydisperse droplet digital CRISPR/Cas13a (PddCas13a) to assess whether a CRISPR/Cas13a readout could support a practical miRNA-based diagnostic workflow. Results: Three serum miRNAs (miR-17-3p, miR-504-3p, and miR-6877-5p) were identified as diagnostic markers. PCaRS achieved stable discrimination across the SYSMH Train, SYSMH Test, HHD, and GPH cohorts [AUCs: 0.836 (0.790 - 0.881), 0.832 (0.773 - 0.907), 0.826 (0.721 - 0.932), and 0.820 (0.702 - 0.938), respectively]. PCaRS, f/tPSA, PSA Density, and Prostate Imaging Reporting and Data System score were independent predictors of PCa. Among six machine-learning models, the Support Vector Machine based composite model (PCaSVM) achieved the best performance, with AUCs of 0.939 (0.912 - 0.966), 0.899 (0.849 - 0.948), 0.886 (0.806 - 0.967), and 0.905 (0.834 - 0.976) in the four retrospective cohorts and 0.873 (0.772-0.975) in the prospective cohort. In the prospective cohort, a PddCas13a-derived score (PCaCas13aS) achieved an AUC of 0.831 (0.783 - 0.872), with no significant difference from the qPCR-based PCaRS. Conclusions: The PCaSVM achieved satisfactory diagnostic performance, suggesting potential utility for non-invasive diagnosis of PCa. The PddCas13a-based quantitative detection of serum miRNAs presents a feasible approach for diagnosing PCa. Larger prospective multicenter studies are warranted to confirm biopsy-sparing clinical utility.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Lai, C., Mulati, Y., Huang, X., et al. (2026). A serum-clinical composite model for prostate cancer diagnosis: multicenter validation and CRISPR/Cas13a-based detection. The Journal of Liquid Biopsy. https://doi.org/10.1016/j.jlb.2026.100483