Detection of primary cancer types via fragment size selection in circulating cell-free extrachromosomal circular DNA

BACKGROUND: Extrachromosomal circular DNA has emerged as a pivotal factor in tumor biology, contributing to intratumor heterogeneity, oncogene amplification, and tumor evolution. Despite its relevance, the presence and molecular characteristics of plasma-derived cell-free extrachromosomal circular DNA (eccDNA) in cancer patients remain insufficiently explored. METHODS: In this study, we profiled plasma-derived cell-free eccDNA from a multi-cancer cohort consisting of 413 cancer patients and 239 healthy individuals. We analyzed eccDNA fragment size distributions using a patient-derived xenograft (PDX) mouse model and identified fragment size features to distinguish tumor-derived eccDNA from non-tumor-derived eccDNA. We further performed in silico fragment size selection and developed a gene-based annotation method to characterize the gene content carried by cell-free eccDNA across different cancer types. By utilizing these cell-free eccDNA signatures, we developed ScanTecc (screening cancer types with cell-free eccDNA), a machine learning-based approach for cancer detection and tissue-of-origin classification. The classification performance of ScanTecc was further evaluated at the individual sample level using multiple machine learning classifiers, including adaptive boosting and logistic regression. RESULTS: We observed a significantly higher abundance and longer fragment lengths of eccDNA in cancer patients’ plasma. Analysis of the PDX mouse model revealed a distinct fragment size threshold of approximately 1,000 base pairs that effectively differentiates tumor-derived eccDNA from non-tumor-derived counterparts. Following fragment size selection, gene-based annotation of large-sized eccDNA revealed cancer type–specific enrichment of tumor-associated genes. Leveraging these features, ScanTecc achieved an overall AUC of 0.92 for distinguishing cancer patients from healthy individuals, with consistently high performance across disease stages, including stage I (AUC = 0.92) and stage IV (AUC = 0.93). ScanTecc also enabled accurate tissue-of-origin classification and achieved an overall AUC of 0.79 in identifying specific cancer types, with AUC values ranging from 0.70 for gastric cancer to 0.81 for ovarian cancer. CONCLUSIONS: Our study establishes a multi-cancer plasma cell-free eccDNA landscape and introduces a non-invasive cancer screening framework based on cell-free eccDNA features, highlighting the potential of plasma cell-free eccDNA for early cancer detection and tumor classification.