A machine learning approach to infer DNase1L3 activity from plasma cell-free DNA fragmentomics

DNase1L3 is an endonuclease that fragments DNA during apoptosis and digests DNA from microparticles in plasma, shaping key features of cell-free DNA (cfDNA). The common missense variant p.Arg206Cys (R206C) affects cfDNA through a non-linear, semidominant allele dosage effect. Therefore, commonly used models trained on fragmentomics from individuals with normal DNase1L3 activity perform poorly in R206C homozygotes. To address this we analyzed cfDNA sequencing data from 129,676 Non-Invasive Prenatal Tests and validated R206C genotypes in a selection of 169 matching plasma samples. Supervised and unsupervised learning were used to infer DNase1L3 activity from cfDNA fragmentation properties. Our models accurately identify R206C homozygotes using as little as 10,000 cfDNA fragments, outperforming genotype imputation in this setting. However, unsupervised analysis reveals a small number of samples that cluster with homozygotes but lack the corresponding genotype, suggesting that our method also identifies other or downstream effects of DNase1L3 impairment. Conversely, some R206C homozygotes initially lacked the aberrant fragmentome, but longitudinal follow-up across subsequent pregnancies suggests that they developed aberrant fragmentomes over time. While reversions of the fragmentome are also observed, they only occur in wildtype or heterozygote samples, suggesting that the aberrant fragmentome may be transient in the latter but represents a stable, time-dependent end-state in R206C homozygotes.