Genetic Medicine · canada
Taking Blood’s Epigenetic Clues Across Tissues: Machine Learning Charts a Path for Prenatal Variant Interpretation
Using Down syndrome as a proof of concept, a research team converted DNA methylation signatures from postnatal blood into a cross-tissue classifier, with the aim of adding another layer of evidence for interpreting variants of uncertain significance identified by prenatal sequencing; however, hurdles including expansion to other diseases and prospective validation remain before it can become a routine clinical test.
Prenatal genetic sequencing can sometimes identify a variant without being able to determine whether it actually causes disease. Such “variants of uncertain significance” can leave families and medical teams facing difficult decisions with insufficient information. A research team at The Hospital for Sick Children (SickKids) in Toronto, Canada, has proposed a machine-learning approach that seeks to transform epigenetic data accumulated from postnatal blood into an interpretation tool for prenatal samples such as placental tissue and umbilical cord blood.
The research focuses on DNA methylation “episignatures”: certain genetic disorders leave relatively stable, disease-specific methylation patterns across the genome that can help determine whether genetic variants are pathogenic. However, most existing signatures are derived from postnatal whole blood, and methylation changes with cell type and developmental stage, so directly applying them to prenatal tissues often does not work.
This proof-of-concept study, published in The American Journal of Human Genetics, selected trisomy 21, or Down syndrome, as its model. The researchers first established a blood-derived trisomy 21 episignature using 266 samples, then integrated public data to train a machine-learning model on 850 trisomy 21 and control samples spanning six prenatal and postnatal tissues. According to the paper’s abstract, the model accurately predicted trisomy 21 status across all tested tissues, suggesting that disease-associated signals may contain a shared component that is not restricted by cell type.
A step closer to practical application involved combining an established postnatal episignature with a small number of prenatal samples for training. The results showed that this approach could generate a classifier capable of identifying prenatal samples, meaning researchers may be able to begin building prenatal epigenetic interpretation models without having to collect large quantities of difficult-to-obtain fetal tissue anew for every disease.
However, this result does not constitute a new Down syndrome screening test, nor has it yet been shown to improve the clinical accuracy of current prenatal diagnosis. Trisomy 21 has clear and widespread genomic effects, making it more likely than a single-gene disorder to produce a recognizable signal. Whether the model can be extended to rarer neurodevelopmental disorders with greater phenotypic variability, and whether it can reliably interpret specific variants of uncertain significance, still require validation on a case-by-case basis.
Before the approach can enter clinical workflows, further research will also need to conduct blinded, prospective validation using independent and representative prenatal cases, and clarify how differences in gestational stage, sampled tissue, maternal cell contamination, and experimental platforms affect the results. At this stage, the method is better viewed as a technical starting point for overcoming tissue barriers than as a mature test that can independently determine medical decisions during pregnancy.