Machine learning could help classify genetic variants in prenatal testing

An approach based on machine learning may reveal signatures of disease in prenatal tissues based on information obtained from postnatal blood samples.

Prenatal testing – where the genetic information of a fetus can be obtained from samples of the maternal blood, the placenta or the amniotic fluid – typically relies on DNA sequencing to identify genetic variants that are associated with disease. However, in many instances, whether a specific variant is harmful or benign remains unclear, leaving future parents without answers. In a proof-of-concept study, researchers from the Hospital for Sick Children in Toronto, Canada, developed a machine learning model that could help clinicians evaluate these 'variants of uncertain significance' using blood-based epigenetic patterns.

'We're thrilled that our model can bring a new level of precision to prenatal testing, where so many questions remain to be answered,' said Dr Rosanna Weksberg, clinical geneticist and joint leader of the study published in the American Journal of Human Genetics.

The team focused on episignatures, which are patterns of chemical tags attached to DNA that act as unique markers for specific genetic conditions. Their study built on previous work in which they used episignatures from blood samples to distinguish between disease-causing and benign genetic variants. However, this approach remained tissue-specific: episignatures established from postnatal whole blood could not be applied to samples from other tissues, such as the ones obtained during testing of the amniotic fluid or placenta.

'If we could take these blood-derived signatures and make them tissue-agnostic, we could overcome one of the biggest limitations in prenatal diagnostics,' said Dr Sanaa Choufani, a senior research associate and joint leader of the study.

In response, the team developed a cross-tissue classification strategy using a machine learning model. To test the new approach, they first generated an episignature for Down's syndrome that was derived from postnatal blood samples from individuals with the condition. The model was then trained on genomic data extracted from six pre- and postnatal tissue types obtained from people with and without Down's syndrome. It was able to accurately identify the generated episignature in all sample types, suggesting that it is possible to establish episignatures using blood samples that remain relevant for other tissues and in the context of prenatal diagnosis.

Going forward, the researchers propose that this approach could be used to convert vast amounts of blood-derived episignatures into tissue-agnostic ones, opening the door to more extensive and effective diagnoses of rare genetic conditions. They also suggest that their work could enable a wider range of sample types to be used, including saliva and oral swabs that are easier to collect. However, further research and rigorous validation in the clinic are needed before the approach can be used for diagnosis.

The impact of AI on genomics will be explored at this year's PET Annual Conference, AI and Automation in Fertility and Genomics: Pipeline? Or Pipe Dream?, taking place in central London on Wednesday 9 December 2026.

Find out more and register here.