For the families of children with rare genetic diseases, the path to a diagnosis can stretch across years, dozens of specialists, and still end without an answer. A study published June 18, 2026, in NEJM AI suggests that AI-assisted genomic reanalysis may open a narrow but meaningful door for some of those families — provided human geneticists remain firmly in the final seat.
Researchers at Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University, and OpenAI applied the o3 Deep Research reasoning model to 376 de-identified pediatric cases that had previously undergone genetic testing and expert review without reaching a diagnosis. After the AI surfaced candidate gene-phenotype links and human clinical geneticists independently evaluated each output, 18 cases resulted in confirmed diagnoses — an incremental diagnostic yield of 4.8%.
The breakdown by disease area: 10 patients with rare neurodevelopmental disorders, four with neuromuscular disease, two cases of sudden unexpected death in pediatrics, and two patients with early-onset psychosis.
“It got almost 5% new diagnoses, which doesn’t sound like a lot, but considering how many times these had already been analyzed, that’s a huge number, and each one means an answer for a family,” said Catherine Brownstein, scientific director of the genetic investigations arm of the Manton Center and one of the study’s lead researchers.
How the pipeline worked — and what it did not do
The research team fed the o3 model a structured dossier for each case: clinicians’ notes, a description of the patient’s phenotype, and a filtered list of candidate genes. The model was asked to propose the most plausible molecular explanation and to show its reasoning — effectively generating evidence-linked hypotheses, not diagnoses.
From there, every model output required independent review by at least two board-certified clinical geneticists applying the ACMG/AMP variant classification framework. A finding advanced to diagnosis status only after four conditions were met: expert review, pathogenic or likely-pathogenic variant classification, confirmation in a CLIA-certified laboratory, and clinical return of the result to the family. The o3 model made no clinical decisions and issued no diagnoses.
Study type and key limitations
This was a retrospective reanalysis study. The 376 cases represent a selected cohort of patients who had already received prior negative workups — meaning the population was, by design, the hardest to diagnose. There is no prospective trial data yet demonstrating that the pipeline improves outcomes under real-world clinical deployment, nor a controlled comparison arm. The 4.8% yield figure should be understood in that context: it is an incremental gain on top of prior specialist evaluation, not a baseline diagnostic rate.
Conflict of interest
The collaboration is part of a broader initiative: in March 2025, OpenAI committed $50 million to its NextGenAI consortium, a 15-institution research partnership that includes Boston Children’s Hospital alongside Harvard, MIT, Caltech, Oxford, and ten other institutions. That financial relationship between the AI company whose model was evaluated and a member institution whose cases were studied is a material conflict of interest. Readers should weigh that connection when assessing the findings.
Correction, 2026-06-22: An earlier version of this article stated that OpenAI committed $50 million specifically to Boston Children’s Hospital AI initiatives, announced in May 2025. The commitment was made in March 2025 to OpenAI’s NextGenAI consortium — a 15-institution research partnership, of which Boston Children’s is one member — not a bilateral grant to Boston Children’s alone. The conflict-of-interest disclosure has been updated accordingly.
The study represents a peer-reviewed signal that AI-assisted genomic reanalysis can surface actionable hypotheses in cases that have exhausted standard pipelines. Whether that signal holds in a prospective, multicenter setting — and whether the 4.8% yield is reproducible across different institutions — remains to be established.