Each year, millions of children worldwide are born with severe genetic disorders caused by variants in a single gene. Pinpointing the culprit variant from the tens of thousands typical in any individual’s exome is an immense challenge, requiring extensive domain knowledge. As a result, the diagnostic rate for these rare Mendelian diseases hovers between just 30-40%.
However, a powerful new artificial intelligence (AI) system called AI-MARRVEL (AIM) could significantly boost diagnoses for impacted children and their families as per study published in NEJM. Developed by an international team of researchers, AIM was trained on high-quality genetic samples clinically diagnosed by certified experts, along with features encoding genetic principles and clinical knowledge.
When evaluated across three independent real-world patient datasets, AIM doubled the number of accurate diagnoses compared to current standard methods and tools. It achieved a 98% precision rate in identifying diagnosable cases from a pool of 871 previously unsolved samples, successfully pinpointing the causative variants in 57% of those cases.
“Existing bioinformatics tools often struggle with limited accuracy, coding variants, and simulated rather than real patient data,” explains lead researcher Dr. Ana Zlatanova. “AIM overcomes these limitations through an AI model trained on expertly curated clinical cases and datasets encompassing the full complexity of human genotypic and phenotypic information.”
In addition to improving diagnostic rates for new patients, the researchers found AIM could aid clinicians in periodically reanalyzing older, unsolved cases as novel disease genes are discovered over time. The AI system’s performance further improved when fine-tuned for specific use cases like recessive disorders and trio analyses of parents and children.
Remarkably, AIM even demonstrated potential for identifying entirely new disease genes. When tested on cases from the Undiagnosed Diseases Network, it correctly predicted two newly reported disease-causing genes that had not yet been incorporated into its training data.
“The successes we’ve seen with AIM across primary diagnoses, reanalysis, and novel gene discovery underscore the immense potential of machine learning to transform the clinical genomics landscape,” says Dr. Zlatanova. “With further optimization and broader deployment, this AI system could help ensure far more children receive life-changing genetic diagnoses moving forward.”
The international team included experts from Stanford University, the Hospices Civils de Lyon, Harvard Medical School, and the Montreal Neurological Institute among other leading institutions. They plan to continue iterating on AIM while making it available for clinical use worldwide.
