Fig. 5

The composition and division of cohort and dataset. Whole-genome NEEM-seq was performed on all samples. About 11 million reads from whole-genome NEEM-seq data were used for DeepTrace model pre-training to capture global and transferrable understanding of human genome methyl-seq data. The pre-trained DeepTrace was then fine-tuned using approximate 1.2 million tumor reads and 2.7 million non-tumor reads after data noise reduction. Before fine-tuning, data noise reduction was conducted on the reads from tumor tissues to retain reads from tumor cells, and remove reads from non-tumor cells in tumor tissues. M: million. LD: liver disease