Machine Learning May Make Prenatal Genetic Testing More Reliable

Advances in genome sequencing are opening a new window onto fetal health, allowing clinicians to examine an unborn baby’s DNA for changes associated with genetic and neurodevelopmental conditions. Yet the same technology that can reveal potentially important mutations can also generate an unsettling problem: genetic variants that science cannot confidently classify. These findings, known as variants of uncertain significance, or VUS, can leave families and physicians without a clear answer about whether a DNA change is harmless, disease-causing or somewhere in between.

Researchers at The Hospital for Sick Children (SickKids) in Toronto have developed a machine-learning approach that could help resolve some of that uncertainty. The method converts blood-based epigenetic patterns into “tissue-agnostic” episignatures—molecular signals that can identify a genetic condition regardless of whether the DNA came from blood, amniotic fluid, placental tissue or another biological source. The advance could eventually make epigenetic testing more useful in prenatal medicine, where access to fetal tissues is limited and clinical decisions often must be made before birth.

The work focuses on epigenetics, the system of chemical modifications that regulates how genes operate without changing the underlying DNA sequence. One of the most important epigenetic mechanisms is DNA methylation, in which chemical groups attach to specific DNA bases and influence whether nearby genes are active or silent. Certain genetic disorders produce distinctive, disease-associated methylation patterns across the genome. These patterns, called episignatures, can act like molecular fingerprints, helping clinicians determine whether a genetic variant is likely to disrupt normal development.

The SickKids team, led by Clinical Geneticist and Senior Associate Scientist Rosanna Weksberg and Senior Research Associate Sanaa Choufani, has helped establish more than 60 episignatures. Many have already been clinically validated for diagnostic use. Traditionally, however, these signatures have been considered tissue-specific. A methylation pattern identified in blood might not be directly applicable to DNA extracted from prenatal samples, because different tissues undergo distinct developmental and regulatory processes. That limitation has prevented many families undergoing prenatal testing from benefiting from episignature-based analysis.

To test whether this barrier could be overcome, the researchers first generated a blood-derived episignature for Down syndrome, a condition caused in most cases by an extra copy of chromosome 21. The signature was built using samples from 266 people with Down syndrome. The team then used publicly available DNA methylation data from 850 individuals with and without the condition to train a machine-learning model. The dataset included six prenatal and postnatal tissue types, allowing the researchers to compare disease-associated methylation patterns across tissues with very different biological functions.

Rather than searching for a single methylation site, the model analyzed coordinated changes across numerous regions of the genome. Machine-learning algorithms can identify combinations of features that are difficult to recognize through manual inspection, including patterns that remain biologically meaningful even when their strength varies between tissues. In this case, the model was designed to preserve the core information of the blood-derived Down syndrome signature while adapting it to the molecular characteristics of other tissues.

The results demonstrated that the transformed signature accurately recognized the Down syndrome pattern in every tissue type tested. This finding suggests that a blood-derived episignature can be computationally converted into a broader diagnostic signal without losing its ability to distinguish affected and unaffected samples. The result does not mean that every genetic condition will produce a universal signature, but it provides proof that tissue-specificity—one of the major challenges in epigenetic diagnostics—may be reduced with carefully trained models and sufficiently diverse reference data.

The potential clinical impact is particularly important for prenatal diagnosis. Amniotic fluid and placental tissue may be available during pregnancy, but they are not equivalent to blood and can contain limited amounts of DNA. A tissue-agnostic episignature could allow clinicians to compare prenatal methylation data with established diagnostic patterns even when the original signature was discovered in postnatal blood. In practical terms, the approach could help interpret uncertain variants and provide families with more precise information at a time when uncertainty can have profound emotional and medical consequences.

The researchers also believe the strategy could eventually expand testing beyond conventional blood samples. Saliva and oral swabs, which are easier and less invasive to collect, may become useful sources of DNA if their molecular signals can be reliably connected to disease-associated episignatures. Such applications would require extensive validation across larger and more diverse populations, as well as careful assessment of false-positive and false-negative results. Machine-learning systems must also be tested in real clinical settings to ensure that their predictions remain reliable when samples vary in quality, ancestry, developmental stage and medical history.

The study, published in The American Journal of Human Genetics, represents an early but significant step toward more flexible epigenetic diagnostics. By combining genome-scale methylation analysis with machine learning, the SickKids team has shown that information discovered in one tissue can potentially be translated to others. The researchers say the long-term goal is to shorten the diagnostic odyssey experienced by many children and families affected by rare genetic disorders, while giving clinicians clearer evidence for interpreting uncertain variants. Supported by the Canadian Institutes of Health Research, the work could help advance a more individualized approach to prenatal and pediatric care, in which molecular data are used not only to diagnose disease but also to guide families through some of medicine’s most difficult decisions.

Web References: The American Journal of Human Genetics study; EpigenCentral; Rosanna Weksberg; SickKids Genetics & Genome Biology.

References: American Journal of Human Genetics; The Hospital for Sick Children; Canadian Institutes of Health Research.

Keywords

Prenatal genetic testing, variants of uncertain significance, episignatures, DNA methylation, epigenetics, tissue-agnostic diagnostics, machine learning, Down syndrome, prenatal medicine, genetic disorders, medical genetics, SickKids