AI Can Detect Genetic Risk Using Blood Tests

New artificial intelligence models can yield much more nuanced and detailed assessments of genetic risks for 10 inherited diseases, researchers reported last 28 August.

This kind of machine learning has the potential to be a powerful new tool for helping clinical geneticists more accurately screen for inherited diseases and can greatly improve on test results that are often murky or uncertain, according to a study of the AI models published in the journal Science.

Tapping more than 1.3 million electronic health records generated by routine lab tests, researchers at the Icahn School of Medicine at Mount Sinai in New York used their models to focus on 1,648 rare variants in 31 genes corresponding to 10 "autosomal dominant" diseases, meaning diseases in which risk can be inherited with only one copy of a mutated gene from one parent.

A machine learning, or ML, model was constructed for each of 10 diseases: arrhythmogenic right ventricular cardiomyopathy, familial breast cancer, familial hypercholesterolemia, hypertrophic cardiomyopathy, adult hypophosphatasia, long QT syndrome, Lynch syndrome, monogenic diabetes, polycystic kidney disease and von Willebrand disease.

The authors reported the models succeeded in generating scores for the "penetrance" of each of the hundreds of genetic variants -- or how likely a variance is to ultimately result in a disease. The "ML penetrance" scores range from 0 to 1, with a higher score closer to 1 suggesting a variant may be more likely to contribute to disease, while a lower score indicates minimal or no risk.

Senior study author Ron Do, the Charles Bronfman Professor in Personalized Medicine at the Icahn School, said such AI-generated penetrance scores represent a vast improvement over existing testing which can yield only simple "yes/no" answers for diseases such as high blood pressure, diabetes and cancer, whose genetic risks don't actually fit into such neat, binary categories.

"Our study shows that machine learning-based penetrance is valuable not only for classic hereditary conditions such as familial breast cancer, familial hypercholesterolemia, or long QT syndrome, but also for diseases with murkier boundaries like monogenic diabetes, cardiomyopathies and kidney disease," he told UPI in emailed statements.

"These conditions exist on a spectrum, and our approach quantifies risk in a way that reflects that spectrum. By combining genetic information with real-world health data such as lab values and vital signs, we can provide more nuanced and clinically relevant risk estimates," he said.

One of the problems with current genetic testing methods is that for many patients, "receiving a genetic test result that is labeled 'uncertain' can be frustrating and anxiety-provoking, because it leaves them and their families without clear guidance," Do said. "Our work shows that ML-based penetrance has the potential to help reduce that uncertainty."

The likelihood of an inherited disease risk manifesting itself can be refined by drawing on millions of routine health records, the authors state. For example, it showed that patients with high ML penetrance scores had measurable differences in cholesterol, heart rhythms or kidney function.

"For patients, this could mean more personalized risk assessments and earlier interventions if warranted," Do added. "While this approach does not replace conventional penetrance metrics, it adds an additional layer of evidence that can help patients and their clinicians make more informed decisions."

The AI-generated data produced some surprises, as well: some genetic variants which had been considered "uncertain" showed clear signs of producing disease, while others previously thought to be likely culprits manifested few real-world effects.

The next step is to expand the model to include more diseases and to address a wider range of genetic changes and more diverse populations, the authors say.