Computational Methods and Challenges in Cell-Free DNA Analysis for Multi-Cancer Early Detection

Multi-cancer early detection (MCED) aims to identify multiple cancer types simultaneously from a single blood draw by analyzing cell-free DNA (cfDNA). As cancer cells undergo apoptosis and necrosis, they release DNA into the bloodstream that carries distinct molecular signatures, such as altered fragmentation patterns and aberrant methylation profiles. This review evaluates computational methods developed between 2022 and 2025, focusing on how these biological signals are processed to detect early-stage malignancies.

The reviewed methods are categorized into classical statistical/machine learning models and deep learning frameworks. Classical approaches, such as those using random forests or gradient boosting, often provide higher interpretability and lower computational overhead. In contrast, deep learning models excel at capturing complex, high-dimensional, and non-linear patterns in genomic data. The current trend favors multimodal ensemble frameworks—such as THEMIS and ARTEMIS—which combine multiple features like fragment size, end motifs, and methylation density to improve detection sensitivity and tissue-of-origin (TOO) identification.

Despite promising performance metrics, the field faces significant hurdles. A primary concern is the lack of standardized reporting and evaluation protocols, which complicates the comparison of different models. Furthermore, many studies suffer from demographic imbalances where cancer cohorts are significantly older than control groups, raising concerns that models may be learning age-related biological noise rather than cancer-specific signals. Future progress requires larger, well-controlled validation cohorts and a shift toward methods that remain robust at lower sequencing depths to reduce clinical costs.