Figure 6.

Applications of methylation pattern analysis in cancer clinical settings. Each DNA methylation pattern in a single sequence read is evaluated using a trained machine learning classifier. The classifier considers the methylation pattern as a mixture of a healthy (normal) methylation pattern and tumor methylation pattern. In the training step, both the normal methylation pattern and the methylation pattern information data for each cancer type are fed into the classifier. According to trained hyperparameters, the classifier scores each read and classifies whether the read (i.e., DNA molecule) originates from tumor or normal DNA. The model collects information for each read in an ensemble manner for each genomic region of interest. The collected signal is utilized to (1) decide whether the sample is cancerous and (2) deconvolute the tumor of origin. (TOO: tumor of origin, f: fraction of tumor).