Integrative cfDNA profiling from low-pass whole-genome sequencing enables tissue-of-origin prediction in cancer.
Summary
This study introduces a stacked ensemble machine learning classifier designed to predict the tissue of origin for cancers using circulating tumor DNA (ctDNA). The model integrates 11 multidimensional ctDNA features, encompassing genomic, fragmentomic, methylation/repeat, and microbial signals, derived from low-pass whole-genome sequencing. It demonstrated robust performance, achieving top-1 and top-2 accuracies of 80% and 90% in an independent validation cohort of 1,221 samples. The classifier proved effective even with low tumor fractions and successfully identified the primary site in 73.3% of cancers of unknown primary (CUP) cases. This non-invasive approach holds significant potential for guiding clinical decision-making.
Analysis
CLINICAL: This classifier has immediate clinical impact by offering a non-invasive method to identify the tissue of origin for cancers, particularly valuable for cancers of unknown primary (CUP) and multiple primary cancers (MPC). With a top-1 accuracy of 80% and top-2 accuracy of 90% in an independent validation cohort, and maintained performance even with low tumor fractions (71% top-1), it could justify clinical trials to evaluate its integration into current diagnostic algorithms. Identifying the primary site in 73.3% of CUP cases could directly improve therapeutic management and prognosis. The time to clinical impact could be 3-5 years, following prospective validation. BIOMOL: The discovery lies in integrating multiple ctDNA signals (genomic, fragmentomic, methylation/repeat, microbial) for accurate tissue-of-origin classification. The technology used is low-pass whole-genome sequencing (WGS) of ctDNA, a cost-effective and non-invasive approach. Feature importance analysis identified nucleosome positioning, fragment size distribution, and repeat elements as key contributors, suggesting potential novel biomarkers. The use of ctDNA, a liquid biopsy, enables early detection and non-invasive monitoring, with implications for screening and diagnosis. BIOINFO: The study developed a stacked ensemble machine learning classifier. This model integrates predictions from five base algorithms (Deep Learning, Distributed Random Forest, Gradient Boosting Machine, Generalized Linear Model, XGBoost) trained within a five-fold cross-validation framework. The model was optimized for top-1 accuracy and evaluated on training (n=1,814) and independent validation (n=1,221) cohorts, demonstrating robustness. Feature importance analysis helped identify the most relevant signals. Once clinically deployed, this tool could automate and improve the accuracy of tissue-of-origin diagnosis, complementing or potentially replacing current multimodal diagnostic methods.