Cancer detection and classification using circulating tumor DNA from blood samples

Accurate classification of cancers is essential for guiding diagnosis and treatment, yet traditional approaches often rely on invasive biopsies and are limited in their ability to capture tumor heterogeneity. Circulating tumor DNA (ctDNA) provides a non-invasive alternative for real-time tumor profiling through blood-based liquid biopsies; however, the sparse and high-dimensional nature of ctDNA methylation data across multiple cancer types poses substantial challenges for conventional machine learning models. We propose a deep learning framework that leverages transfer learning to enhance cancer classification from ctDNA methylation profiles. The framework integrates Variational Autoencoders (VAEs) and transfer learning-based architectures to extract latent epigenetic features and capture inter-cancer relationships and differential methylation patterns. Knowledge learned from well-characterized cancer types is transferred to improve model performance in detection and classification of common and rare or understudied malignancies from ctDNA datasets. The proposed approach achieves improved classification performance across multiple cancer types within the benchmarking framework comprised of assembled ctDNA dataset used in this study. The transfer learning augmented deep learning model compared favorably with several conventional machine learning baselines on this heterogeneous ctDNA methylation dataset. These results highlight the effectiveness of transfer learning–based deep learning models for multi-cancer classification using ctDNA. Our study provides a proof-of-concept framework for non-invasive cancer detection and classification and further, for biologically informed feature prioritization.