From multi-omics to deep learning: advances in cfDNA-based liquid biopsy for multi-cancer screening

Cancer remains a leading cause of mortality worldwide, with early detection being critical for improving survival rates. Traditional diagnostic methods, such as tissue biopsies and imaging, face limitations in invasiveness, cost, and accessibility, making liquid biopsy a compelling non-invasive alternative. Among liquid biopsy approaches, circulating cell-free DNA (cfDNA) analysis has gained prominence for its ability to capture tumor-derived genetic and epigenetic alterations. This review summarizes key cfDNA biomarkers, including gene mutations, copy number variations (CNVs), DNA methylation, fragmentation patterns, and end motifs (EMs), and highlights their utility in cancer detection and monitoring. By integrating these multi-modal cfDNA biomarkers, feature fusion approaches have not only enhanced the performance of cancer classification models but also stabilized low-abundance signals, thus ensuring more reliable cancer detection and monitoring. Furthermore, the diagnostic power of cfDNA analysis has been further amplified by machine learning (ML), with both traditional ML and deep learning (DL) methods demonstrating strong predictive performance in routine clinical liquid biopsy applications. However, challenges remain, including tumor heterogeneity, standardization of data processing, model explainability, and cost constraints. Future advancements should focus on refining multi-modal feature integration, developing explainable AI (XAI) models, and optimizing cost-effective strategies to enhance clinical applicability. As computational methodologies advance, the integration of cfDNA biomarkers with ML frameworks holds great promise to reshape non-invasive cancer detection by enabling earlier diagnostics, more accurate prognostic evaluation and personalized treatment strategies.