Cornelli, L. | Search bioRxiv

bioRxiv2026

Benchmarking fragmentation-derived artificial cfDNA reference standards

An important step toward clinical implementation of (epi-)genomic assays on liquid biopsies is their validation on identical samples within and across laboratories. For these validation studies, there is a need for cell-free DNA (cfDNA) samples with defined tumor fractions and (epi-)genomic aberrations. However, the amount of circulating cfDNA isolated from patient samples is often limited, especially in pediatric cases. Additionally, patient samples contain a high degree of variability in cfDNA yield and tumor fraction. Several commercial artificial cfDNA products are available for validation studies, however their use is restricted to specific assays, aberrations and/or tumor entities. Alternatively, artificial cfDNA samples can be produced by fragmenting genomic DNA to mimic highly fragmented cfDNA derived from both tumor and healthy blood, followed by mixing artificial tumoral and healthy cfDNA at defined fractions. In this study, we compared native cfDNA with artificial cfDNA generated by three different fragmentation methods, including sonication and two enzymatic digestions using micrococcal nuclease and double-stranded deoxyribonuclease (dsDNase). We assessed fragment length profiles, end motifs and nucleosome occupancy patterns from shallow whole-genome sequencing data, as well as coverage profiles from targeted panel sequencing, together with a small-scale mixing experiment of tumor and healthy cell derived artificial cfDNA. Although sonication remains a convenient high-throughput approach to generate artificial cfDNA for certain downstream applications, enzymatic fragmentation, particularly the dsDNase-based method, more faithfully reproduced native cfDNA characteristics.

genomics↗

bioRxiv2025

A benchmark of DNA methylation deconvolution methods for tumoral fraction estimation using DecoNFlow

In cancer patients, circulating cell-free DNA (cfDNA) is released into body fluids from both healthy and cancer cells. The proportion of tumor-derived cfDNA serves as a surrogate marker of tumor burden allowing disease monitoring. Tumoral cfDNA can be distinguished based on patient specific tumoral mutations or using more general tumor specific DNA methylation patterns, that are preserved on tumoral cfDNA. DNAm profiling of cfDNA thus enables non-invasive cancer detection and monitoring. However, accurately determining tumour fractions remains challenging due to the heterogeneous mixture of cfDNA sources in body fluids. Computational DNAm deconvolution methods address this by inferring cell-type contributions either with or without reference methylomes. While several tools exist and multiple benchmarking studies have been performed, none have specifically evaluated the sensitivity and accuracy of tumour-fraction estimation in cfDNA-focused contexts. Here, we benchmarked 10 reference-based and 2 reference-free DNAm deconvolution tools using 3,690 in silico mixtures spanning multiple tumour types, different bisulfite-based sequencing strategies and several sequencing depths. Overall, CelFiE showed the most accurate tumour-fraction estimation across the different conditions. Interestingly, reference-free methods demonstrated superior sensitivity for tumour detection, but consistent over-estimation of tumoral fraction. We further observed that sequencing depth strongly affects performance until sufficient saturation is achieved. To enable reproducible evaluation and tool selection within this benchmark, we developed DecoNFlow, a scalable Nextflow pipeline integrating 12 deconvolution tools and 3 marker selection methods, making it the most comprehensive pipeline for sequencing-based deconvolution up to date. Together, our findings provide practical guidance for tool selection in cfDNA tumour monitoring and establish DecoNFlow as a robust framework for benchmarking and applying DNAm deconvolution.

bioinformatics↗