Deconvolution of cell-free DNA in cancer liquid biopsy using a deep AutoEncoder

Abstract

Cell-free DNA is emerging as a promising biomarker for early detection of cancer, offering an avenue for both non-invasive diagnosis and personalised cancer treatment. Tissue deconvolution of cell-free DNA is the process of predicting the proportional contributions of DNA from each tissue and cell type, and thus inferring the health status of each tissue and the patient overall. We introduce a new tissue deconvolution method for genome wide cfDNA methylation data, using an autoencoder deep learning model. This trained model produces biologically meaningful predictions, which agree with the literature consensus, and can be used to accurately classify the cancer status of patient samples using a non-invasive cfDNA liquid biopsy.