基于血液的mRNA特征谱检测胰腺癌并鉴别IPMN：发现与初步验证研究

Purpose: Currently, carbohydrate antigen 19-9 (CA19-9) is recommended for monitoring treatment response and recurrence in PDAC, but its poor sensitivity and lack of expression in ~5-10% of patients limit its use for screening in current guidelines. To address this gap, we sought to discover and verify a dedicated set of blood-based mRNA biomarkers and an algorithmic model that (i) differentiates PDAC from benign/healthy states and (ii) discriminates among healthy, IPMN (high-risk), and PDAC. Methods: Phase 1 (discovery): Publicly available datasets [NIH Dataset ID GSE68086, GSE28735, GSE18670] including 96 PDACs and 301 controls (including healthy individuals and non-pancreatic cancers) were analyzed. Using machine learning (ML) algorithmic models trained using a quantitative evolutionary-computing platform (Emerge, Liquid Biosciences Inc.), 18 candidate mRNAs with translational feasibility were identified with strict training/selection/test partitioning and cross-dataset validation. Phase 2 (preliminary verification): We performed RNA sequencing on blood PBMCs from 30 individuals (healthy n=15, IPMN n=5, PDAC n=10) sourced from Crown Bioscience Germany GmbH, Hamburg, Germany. Six independent binary classifiers (PDAC vs healthy + IPMN) were trained using distinct subsets of the 18 mRNAs. We assessed diagnostic performance, redundancy, and a simple “voting” scheme (defined as majority vote across our six independent binary models) across the models. Results: In Phase 1, discovery on blood and tissue/CTC across the three datasets, weighted test performance reached ~95% sensitivity and 98% specificity with multiple ≤6-gene subsets achieving perfect test accuracy on individual datasets. The robust signal was confirmed across diverse modalities and supported reagent availability for all 18 mRNAs, with multiple cross-validated subgroups suitable for clinical translation. In Phase 2 (n=30, PBMCs), the weakest binary model distinguishing between PDAC vs healthy + IPMN made 3/30 errors (all false positives; 100% sensitivity, 85% specificity at 90% accuracy), two models had two errors, two had one, and one model had zero errors. To enhance robustness, we aggregated the six independent binary models via our simple majority voting; the ensemble achieved 100% sensitivity and 100% specificity in the 30-subject cohort. The tri-state classifier achieved 100% three-way accuracy, requiring a minority subset of the 18 biomarkers to resolve healthy, IPMN, and PDAC. Conclusions: A cross-validated set of blood mRNAs enables accurate PDAC detection and simultaneous discrimination of IPMN. While Phase 2 findings are compelling, they derive from a limited, banked and retrospective cohort and warrant confirmation in a larger study including samples from individuals with other high-risk profiles for PDAC and IPMN types that include malignant and benign characteristics.