Machine learning separates a DNase1L3 fragmentomic signal from low-coverage cfDNA data
Cell-free DNA sequencing is increasingly used as a proxy for placental, tumor, and other tissue-derived DNA, yet the biology behind its fragment patterns is not uniform. In this study, the strongest signal came from an inherited nuclease defect rather than from tissue of origin. Machine learning was able to identify the fragmentomic signature linked to homozygosity for the DNASE1L3 p.Arg206Cys variant, a genotype that reduces DNase1L3 function and distorts cfDNA fragmentation.
The clinical relevance reaches beyond prenatal testing. DNase1L3 is central to the clearance of extracellular DNA, including DNA bound to microparticles and neutrophil extracellular traps, and impaired activity has already been tied to autoimmunity. When that enzyme is inefficient, plasma cfDNA no longer reflects only the balance of cell death and tissue contribution; it also reflects altered nuclease processing. In practical terms, that means a liquid biopsy can carry a hidden host-factor signature that confounds downstream interpretation.
Why an inherited nuclease defect can distort liquid biopsy readouts
DNase1L3 helps fragment DNA during apoptosis and digest extracellular DNA in plasma. When function drops, the resulting cfDNA carries a distinctive fragmentomic pattern: fewer mononucleosome-sized fragments and relatively more multi-nucleosome fragments. Earlier work linked the p.Arg206Cys variant to this pattern, along with reduced accuracy in cfDNA-based assays such as noninvasive prenatal testing.
The biological logic is straightforward. Fragmentomics is often treated as a readout of tissue contribution, but nuclease activity sits upstream of that signal. If plasma DNA is not cut in the expected way, the resulting size distribution, end motifs, and cleavage-site patterns shift in a direction that can mimic or obscure disease-associated changes. A genotype-based approach alone does not capture all of that variability, especially when the same functional phenotype may arise through other routes, including other DNASE1L3 variants, autoantibodies against DNase1L3, or acquired influences.
129,676 prenatal samples formed the testing ground
Investigators re-analyzed low-coverage whole-genome cfDNA sequencing data from 129,676 non-invasive prenatal tests generated in routine practice in the Netherlands. The broader population came from the TRIDENT-2 study, and samples with aneuploidies or without consent for research use were excluded. To establish labeled training data, maternal genotypes were validated in a smaller set of plasma samples using digital droplet PCR.
The modeling strategy relied on genome-wide fragmentomic features derived from three main readouts: insert-size distributions, cleave-site motifs, and 5′ end patterns. In the initial training set, 63 genotyped plasma samples included wild-type, heterozygous, and homozygous variant cases. The expanded labeled dataset reached 169 samples after adding 106 plasma samples in which genotype imputation and model prediction disagreed, a deliberate design choice to enrich the positive class and sharpen the boundary between phenotypes.
This is not a random population sample, and the authors did not present it as one. The cohort was assembled to test whether a machine-learning model could recognize the fragmentomic fingerprint of R206C homozygosity better than haplotype-based genotype imputation alone.
The model outperformed haplotype imputation and needed surprisingly little sequence data
Across supervised classifiers and unsupervised dimensionality reduction, the fragmentomic signal proved strong enough to recover the R206C homozygote phenotype with better performance than genotype imputation. That matters because imputation is attractive in low-coverage cfDNA workflows, but it is also vulnerable when plasma yield is limited. Earlier work from the same group had already shown a genotyping error rate of roughly 10 percent for haplotype-based inference of this variant.
A particularly useful finding was the sequencing threshold. Accurate predictions were possible with as few as 1,000 sequenced cfDNA fragments, far below the amount routinely generated in many cfDNA assays. In a practical laboratory setting, that shifts the question from whether there is enough material for a meaningful readout to whether the assay can be interpreted in the presence of a nuclease-driven phenotype. For low-coverage testing pipelines, that is a major operational advantage.
Machine learning also worked with both paired and unpaired sequencing data, extending its utility across assay designs. The models were not dependent on a single threshold or one narrow fragment feature; rather, they integrated multiple correlated properties of cfDNA, which likely explains why they were able to capture the phenotype more reliably than a genotype-imputation strategy built around haplotypes alone.
Discordance between genotype and phenotype suggests biology beyond a single variant
Unsupervised analysis added an important caution. A small number of samples clustered with homozygotes even though they lacked the corresponding genotype. That pattern suggests that similar fragmentomic alterations can arise through mechanisms other than R206C homozygosity. In other words, the model was finding a functional state, not just a DNA sequence label.
The reverse discordance was equally notable. Longitudinal follow-up across subsequent pregnancies showed that some genetically confirmed homozygotes initially did not show the aberrant fragmentome, then developed it in later pregnancies. That finding points to variability in phenotypic expression over time. Host physiology, pregnancy-related factors, or other modifying influences may alter how much DNase1L3 dysfunction is visible in plasma at a given moment.
This is where the biology becomes clinically consequential. A cfDNA assay that assumes fragmentomic changes are caused only by tissue injury or fetal fraction will miss another layer of variation: inherited or acquired nuclease impairment. The result is not merely theoretical noise. It can alter interpretation in prenatal screening and may affect any liquid biopsy built on fragment size or motif assumptions, including oncology and transplant applications.
A biomarker concept with broader implications for cfDNA assays
The authors argue that the R206C-associated fragmentomic phenotype could serve as a biomarker for future DNase1L3-targeted therapeutic studies. That is a reasonable extension of the data because the model is detecting an enzymatic consequence rather than a downstream disease label. If DNase1L3 function can be restored therapeutically, fragmentomics could offer a direct molecular readout of response.
The study also reinforces a larger principle in cfDNA medicine: host biology can shape sequencing signals as much as the tissue of origin does. As liquid biopsy platforms expand, the assumption that low-coverage plasma sequencing only reports on fetal, tumor, or transplant-derived DNA becomes harder to defend. Nuclease biology belongs in the interpretive framework.
The work still leaves room for caution. The training and validation sets were enriched for boundary cases and homozygous variant samples, so the reported performance should not be read as a population-wide prevalence estimate. The findings nevertheless establish a clear proof of concept: machine learning can detect a DNase1L3 impairment signature directly from cfDNA fragmentomics, even in sparse sequencing data, and even when the genotype itself is not directly measurable from the assay.
What remains unresolved is how often similar fragmentomic states arise in routine clinical practice from non-genetic causes, and how best to integrate this signal into existing cfDNA pipelines without overcalling genotype-phenotype relationships that are not one-to-one.