Research employs AI to diagnose unsolved genetic syndromes
Using a deep learning tool, new research from the Division of Medical Genetics illuminates the importance of in-depth genetic testing, especially for hard-to-diagnose patients.
“Non-canonical splice variants in thoracic aortic dissection cases and Marfan syndrome with negative genetic testing” recently was published in Nature’s NPJ Genomic Medicine, revealing how an advanced AI tool, SpliceAI, detected an enrichment of splice-site genetic variants in aortic dissection cases compared to healthy controls.
“People with heritable thoracic aortic disease, like Marfan syndrome, are at risk of the aorta tearing, which can be deadly,” explained David Murdock, MD, assistant professor of medical genetics and first author of the paper. “Doctors try to find genetic causes early, but current tests often miss some patients, even those with classic symptoms. This study looked at non-canonical splice variants – genetic changes that don’t occur in the usual ‘hotspots’ but might still affect how a gene is used to make proteins.”
Researchers used SpliceAI, an AI splice prediction tool developed by Illumina and trained on large datasets of human gene sequences and known splicing outcomes, to review all of their “unsolved cases” – including those with early-onset aortic dissections or signs of Marfan syndrome with negative genetic tests. They also employed the technology on thoracic aortic disease cases in the large biobanks of the UK and Penn Medicine and showed in lab experiments that these variants could alter genes implicated in heritable thoracic aortic disease.
“What we were looking for was to see if individuals who have thoracic aortic disease show an enrichment of these kinds of splice-site variants,” he said. “In the past, these kinds of variants would have been ignored, or overlooked because we couldn’t really understand what they do. But with the use of a tool like SpliceAI, we can have a much better idea of whether those variants are disease-causing or not.”
Researchers found that these overlooked variants, discoverable by SpliceAI, were more common in people with aortic dissections than in healthy controls. Some of these variants were in genes that are already known to cause syndromes like Marfan or Loeys-Dietz, even when the patients didn’t exhibit the usual physical hallmarks of these conditions.
“SpliceAI can detect ‘cryptic’ splice sites, non-standard changes that still alter the splicing process. It has proven to outperform older tools in both sensitivity, finding real issues, and specificity, avoiding false positives,” he added. “We suspect some of these splice variants lead to milder, atypical presentations of known syndromes.”
These findings suggest that non-canonical splice variants are an underrecognized contributor to thoracic aortic disease, particularly in sporadic dissection and unsolved Marfan syndrome cases, highlighting the potential of advanced AI-based splice prediction tools in genetic diagnostics.
“We think these variants may contribute to a broader spectrum of disease than previously recognized,” Murdock said, adding that this technology could be expanded to other genetic diseases.
SpliceAI, he added, is already being used in clinical settings due to its robust capabilities.