Benchmarking and optimization of cell-free DNA deconvolution

Abstract

Reference methylomes, used in deconvolution algorithms to determine cell-free DNA tissue sources, were based on driver CpGs from either microarray or sequencing platforms. Cross-validation of these algorithms is important to allow interpretation of data across studies, select optimal sequencing depth, and thus reduce costs of cf-DNA deconvolution assays. Towards this end, we assessed the performance of two reference-based deconvolution algorithms: cfDNAme, sequencing-based methylome signatures, and Meth-Atlas, a microarray-based methylome signatures using a cfDNA bisulfite sequencing. While both algorithms use NNLS model, cfDNAme uses CpG windows, while Meth-Atlas uses individual CpGs as cell or tissue signatures. We determined the optimal the number of informative CpGs signatures, and the best sequencing depths for precise deconvolution. We found that above 5-fold coverage, much lower coverage than what is frequently used, there is little difference between our two chosen algorithms, both identifying the correct tissue make-up with a high accuracy, suggesting that whole genome bisulfite sequencing for tissue of origin identification can be completed in a much more cost-effective manner than previously thought.