A1464
Title: Deconvolution of cell free DNA via Bayesian tree-based marker selection and signature estimation
Authors: Christopher McKennan - University of Pittsburgh (United States) [presenting]
Yucheng Wang - University of Pittsburgh (United States)
Abstract: Cell-free DNA (cfDNA) consists of small DNA fragments released into the bloodstream during cell death and has enabled non-invasive diagnostics for many diseases. A key step in cfDNA analysis is cellular deconvolution, which estimates the proportion of cfDNA fragments originating from each major cell type using CpG methylation patterns. Existing methods suffer from major limitations, including poorly curated marker CpGs, inaccurate cell type-specific (CTS) methylation signatures, and slow runtimes. To address these challenges, cf-TREBLE is developed, a statistically rigorous and scalable method for identifying marker CpGs, estimating CTS methylation signatures, and performing deconvolution. cf-TREBLE uses CTS reference data, a hierarchical cell type tree, and a novel Bayesian model to classify CpGs based on whether their methylation is shared across or unique to specific cell types. This enables the identification of highly reliable markers and improves CTS signature estimates by pooling information across related cell types, especially when reference data are sparse. cf-TREBLE then applies a novel, computationally efficient deconvolution algorithm that accounts for inter-subject variability in methylation. cf-TREBLEs superior performance is demonstrated on realistic simulated data, and it is applied to develop risk prediction models for adenomyosis and endometriosis.
