Fig. 3

Architecture and characteristics of DeepTrace model. A The pre-training, fine-tuning, and prediction process of DeepTrace model using sequencing reads. The NEEM-seq reads from human cfDNA were used to pre-train the DeepTrace model to capture global and transferrable understanding of human genome methyl-seq data. Then the pre-trained model was fine-tuned using tumor reads (i.e., HCC-derived reads) from HCC tumor tissue DNA after noise reduction, and non-tumor reads from non-tumor cfDNA. The fine-tuned model was subsequently used to predict the probability that a read is derived from HCC tumor DNA (i.e., ctDNA). B The architecture details of DeepTrace model. The steps indicated by the dashed lines were performed only in pre-training