Liquid biopsy via cell-free DNA methylation profiling offers a promising avenue for non-invasive cancer detection and tissue-of-origin identification. However, detecting tumor-derived signals in blood remains challenging due to low tumor fraction and background noise from hematopoietic cells. Here, we plan to create a neural network-based classifier trained on methylation beta values from 40 pure solid tissues and augmented with synthetic blood-tissue mixtures to predict tissue origin in bulk whole blood samples. Our approach addresses key technical hurdles, including signal dilution (via simulated tumor DNA at 1–10% concentrations) and domain shift between pure tissue and heterogeneous blood data (using adversarial domain adaptation). The model employs an attention-based architecture to prioritize tissue-specific CpG markers, while filtering non-informative probes. This framework will demonstrate the feasibility of methylation-based "epigenetic mapping" in blood, with potential applications in early cancer screening and minimal residual disease monitoring.