Our Take
The win here is not that AI invented a new drug, but that it connected existing biomedical data across domains fast enough to matter in a neonatal crisis, and the connection held up in practice.
Why it matters
Rare disease diagnosis still relies on slow manual literature review and educated guessing. This case shows that open-source knowledge graphs can compress diagnosis-to-treatment timelines from years to weeks, with direct clinical impact. For rare disease families, speed is survival.
Do this week
Healthcare AI teams: audit your knowledge graph's coverage of rare genetic variants and existing drug-indication pairs this month so you can identify gaps before your next pediatric case.
How a Knowledge Graph Found a Hidden Treatment
Jorie Kraus was born with a rare deletion on chromosome 10, causing severe motor dysfunction. Her muscles did not work properly. Breathing was labored. Her heart, legs, and larynx were all compromised. In the first 73 days of her life, spent in a neonatal intensive care unit, her parents faced repeated warnings that her condition was likely terminal or permanently disabling.
At Mayo Clinic, physician Whitney Thompson and colleague Laura Lambert rapidly sequenced Jorie's genome and fed the genetic data into the Biomedical Data Translator, an open-source knowledge graph funded by the National Institutes of Health. The tool is designed to harmonize and reason across disparate biomedical datasets—research papers, clinical databases, drug properties, genetic annotations—to identify unexpected drug-disease connections.
The tool identified Klonopin, a widely available muscle relaxer commonly used for seizure disorders and panic attacks, as a compound with characteristics that could counteract many of Jorie's neurodevelopmental and motor symptoms. Doctors began treatment. The response was immediate and dramatic. "It was almost like a light switch," her mother Joanie Kraus said at STAT's Breakthrough Summit West in San Francisco on Tuesday. Within weeks, a child who had spent two years with developmental plateaus began moving freely, maneuvering around obstacles, holding objects, and speaking. She is now almost 3 years old and developing normally, though she will require lifelong neurodevelopmental care.
Thompson emphasized the role of the AI system: "I don't think we would have gotten there without the AI tool. It's able to make inferences across all the biomedical literature, things that we wouldn't have been able to connect otherwise. So the AI portion of this was absolutely critical."
Speed and Connectivity Are the Bottlenecks
The typical path from symptom to diagnosis for ultra-rare genetic conditions takes years. Families are shuttled between specialists, undergo multiple rounds of testing, and often never receive a name for their condition, let alone a treatment option. When a treatment does exist, it is buried in literature that no single clinician can reasonably consume.
This case demonstrates that the bottleneck is not biological—Klonopin already existed—but informational. The knowledge graph compressed months or years of manual literature review into hours by making probabilistic inferences across datasets that humans could not reasonably cross-reference. The tool is not inventing new biology; it is operating at human-scale speed for clinical decision-making.
Jorie's case has already replicated. Her parents shared a second case of a 5-year-old with a different chromosome 10 mutation but the same diagnosis. That child's neurologist followed the same steps and prescribed Klonopin. The child went from non-verbal to speaking sentences.
However, barriers remain. Genomic sequencing is not yet widely available. Rapid laboratory testing to confirm a treatment hypothesis is difficult across most health systems. The Biomedical Data Translator itself is a work in progress when it comes to consistent, reliable deployment across diverse geographies and hospital networks.
What Clinicians and Health Systems Should Do Now
For rare disease centers and pediatric hospital systems, the signal is clear: genomic sequencing and AI-assisted drug-disease matching should be part of the diagnostic protocol for unexplained neonatal presentation, not a last resort after months of failed conventional workup. The timeline advantage is real.
This does not mean deploying a proprietary black-box tool. The Biomedical Data Translator is open-source and NIH-funded. Health systems should evaluate whether integrating it into their diagnostic workflow for rare pediatric cases is feasible now, rather than waiting for a commercial vendor to repackage the same capability.
For researchers building knowledge graphs: Jorie's case shows that coverage of rare genetic variants and existing drug-indication pairs is incomplete. Auditing your data for gaps in these domains and prioritizing them is not an academic exercise—it directly affects which families get diagnosed and which treatment options are visible when they matter most.