cfDecon: Accurate and Interpretable methylation-based cell type deconvolution for cell-free DNA

Abstract

Cell-free DNA (cfDNA) analysis has emerged as a powerful tool for noninvasive diagnostics and monitoring of aberrant methylation. However, computational deconvolution of cfDNA remains challenging due to the complexity of methylation data, the diverse cellular compositions in cfDNA mixtures, and the limited interpretability of current methods. We present cfDecon, an efficient deep-learning framework for cfDNA deconvolution at the resolution of individual reads. cfDecon employs a multichannel autoencoder core module and an iterative refinement process to estimate cell-type proportions and generate condition-aware cell-type-specific methylation profiles. We rigorously evaluate cfDecon through comprehensive simulation experiments, including scenarios with normal cellular compositions, rare cell types, and unknown cell types, where cfDecon consistently outperforms previous state-of-the-art methods in deconvolution. Using strict data separation and careful control for potential data leakage, cfDecon shows scalability and applicability to large-scale studies by raising Lins Concordance Correlation Coefficient (CCC) over 33% on an atlas-level reference dataset. We further validate cfDecons utility in disease diagnosis on two clinical datasets, Amyotrophic Lateral Sclerosis (ALS) and Hepatocellular Carcinoma (HCC). cfDecon raises the performance of disease detection from 0.53 to 0.79 compared to existing methods on the ALS dataset and 0.55 to 0.77 on the HCC dataset. The capacity of cfDecon to predict condition-specific DNA methylation patterns across cell types offers a valuable framework for studying differentially methylated CpGs. Furthermore, integrating cfDecon with enrichment analysis facilitates the exploration of functional variations in CpGs across different conditions. cfDecon represents a significant advancement in cfDNA analysis, offering improved accuracy, interpretability, and potential for personalized medicine applications. The codebase is available at https://github.com/Susanxuan/cfDecon.