Efficient and accurate non-invasive tumor monitoring and diagnosis by interpretable deep learning

Detecting tumor-specific DNA methylation in circulating tumor DNA (ctDNA) offers a non-invasive method for tumor detection. The primary challenge lies in identifying the extremely low abundance of ctDNA in cell-free blood plasma (cfDNA). In this study, we present Oncoder, an interpretable deep learning-based tool for economical and accurate non-invasive tumor monitoring and diagnosis. Unlike other methods, Oncoder learns scientifically sound reference methylation atlases from patient blood to provide additional diagnostic insights, fostering trust among clinicians and patients, and continuously improves its accuracy through iterative learning. In simulations, Oncoder reduced prediction errors of tumor signals in blood by at least 30% compared to existing methods and showed the highest prediction correlation, indicating more accurate tumor progression monitoring. We also evaluated Oncoder's performance in various real-world applications. Oncoder sensitively detected changes in ctDNA levels during tumor development and treatment and exhibited superior diagnostic potential even in the earliest stages of cancer.