Cancer Diagnostics using Machine Learning of Tumor- and Tissue- Specific RNA Transcripts

Abstract

Accurate cancer diagnosis and tissue origin identification are crucial for precision oncology. We explored the potential of tumor-specific RNA transcripts (Tumor-SRTs) and tissue-specific RNA transcripts (Tissue-SRTs) as dual biomarkers using machine learning. Tumor-SRTs effectively distinguished malignant from normal tissues across training, test, and validation sets. Classifiers trained on 25 Tissue-SRTs exhibited high performance, validated externally with varying predictive accuracy across tissue types. We developed SRT-based Cancer Diagnostics (SRT-CD), an intelligent diagnostic system, to diagnose primary and metastatic cancers and determine tissue origin of Cancers of Unknown Primary. SRT-CD achieved top-1/top-3 accuracies of 80%/98.1% for primary and 76.9%/92.3% for metastatic cancers in external validation. This study establishes SRT-CD as a robust tool for clinical cancer diagnosis and tissue origin identification, and guiding personalized treatment strategies.