采用基于最少剪接连接的血小板RNA面板对结直肠癌进行跨队列稳健检测

Background: Tumor-educated platelets (TEPs) incorporate cancer-derived RNA signals and represent a promising minimally invasive platform for early cancer detection. However, gene-level TEP signatures such as the widely used 921-gene panel often exhibit reduced performance across heterogeneous cohorts. Because platelet RNA predominantly reflects regulated splicing events rather than transcriptional abundance, exon-exon junction features may provide higher biological specificity and improved stability against hematologic variability. In this study, we aimed to identify representative exon-exon junction alterations in platelet RNA that distinguish CRC from healthy controls. Methods: We analyzed public platelet RNA-seq data (132 CRC and 21 healthy samples; NIH BioProject PRJNA737596) and a prospective clinical cohort (44 CRC and 96 healthy samples), both of which contained a substantial proportion of early-stage CRC (stage I-II, ~52%). Junction-level read counts were quantified, normalized, and filtered based on differential expression, reproducibility, and independence from hematologic indices. A 10-junction panel was selected using logistic regression modeling, and support vector machine (SVM) classifiers were trained and validated using stratified, independent subsets. Performance of the 10-junction model was compared with an identically preprocessed model based on the previously reported 921-gene TEP panel. Functional annotation of the 10 junctions was conducted to assess mechanistic relevance. Results: The 10-junction panel demonstrated consistently strong diagnostic performance across cohorts. In the public validation set (n=68), the model achieved a sensitivity of 89.4%, specificity of 85.7%, and an AUC of 0.912. In the clinical validation set (n=75), sensitivity was 87.5%, specificity 93.3%, and AUC 0.959. Detection of early-stage CRC was robust in both datasets (AUC 0.903 and 0.956 in the public and clinical cohorts, respectively). Notably, the junction-based model outperformed the 921-gene panel in the clinical cohort (AUC 0.959 vs. 0.895). Functional enrichment analysis indicated involvement of vesicle trafficking, autophagy, and platelet-immune signaling pathways, consistent with known mechanisms of platelet reprogramming in cancer. Conclusions: A compact junction-based TEP RNA panel enables accurate detection of CRC, including early-stage disease, and demonstrates superior cross-cohort robustness compared with conventional gene-level approaches. Its small feature set, strong biological coherence, and consistent performance highlight its potential as a scalable and cost-efficient liquid biopsy for CRC screening. Multi-institutional prospective validation is warranted.