Researchers from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University, and OpenAI reported using the OpenAI o3 Deep Research reasoning model to reanalyze 376 previously unsolved rare-disease cases. The model generated evidence-linked candidate explanations; after expert review, additional testing, and clinical confirmation, physicians established diagnoses in 18 cases, an additional diagnostic yield of 4.8% over prior specialist review. The study was published June 18, 2026 in NEJM AI.
How the workflow operated
For each case the team assembled a de-identified packet containing standardized Human Phenotype Ontology terms, clinician notes when available, demographic metadata, and a filtered variant table with rarity, predicted protein effect, ClinVar classification, and signal quality across family members. Most cases included data from the child and both biological parents.
The model was asked to propose the most plausible molecular explanation and to show supporting evidence. Outputs were reviewed by at least two experts using the ACMG/AMP framework. No model output was treated as a diagnosis; a finding counted only after a variant was classified as pathogenic or likely pathogenic, confirmed by a CLIA-certified laboratory, and returned to the family through established clinical processes.
Findings, examples and limitations
The reanalysis produced diagnoses across four cohorts: neurodevelopmental (100 cases, 10 diagnoses, 10.0%), neuromuscular disease (61 cases, 4 diagnoses, 6.6%), sudden unexpected death in pediatrics (200 cases, 2 diagnoses, 1.0%), and early psychosis (15 cases, 2 diagnoses, 13.3%). Overall yield was 4.8%. The early-psychosis cohort was small and its percentage has a wide confidence interval. Seven of the 18 findings were rediscoveries already documented outside the reviewed record.
Case examples include a confirmed 22q11.2 deletion consistent with DiGeorge syndrome identified after the model linked low-quality calls with multisystem features; a frameshift variant in HSPB8 that led to a diagnosis of myofibrillar myopathy after nearly two decades of investigation; and hypotheses of digenic explanations involving TTN and SRPK3, and combined variants in LAMA2 and FOXP1. The model also proposed an 11-amino-acid deletion in S1PR1 as a testable mechanistic hypothesis related to vitiligo; this requires additional laboratory validation.
The authors note limitations: the study was retrospective, cohorts were heterogeneous, reviewers were not blinded to model confidence, and the team did not measure time saved, cost, false-positive workload, or systematically evaluate certain variant types such as repeat expansions or mosaicism. The report stresses that the model did not make diagnoses and that qualified clinicians performed all clinical decisions and confirmations.
The Manton Center will lead the next research stage through a grant from the OpenAI Foundation and calls for prospective, multi-center studies to compare LLM-assisted reanalysis with standard practice on diagnostic yield, time to candidate, clinician effort, and other outcomes. Dr. Catherine Brownstein observed, “The bottleneck is time.” Alan Beggs added, “Researchers like Catherine and me can’t possibly keep 8,000 different diseases in our heads.”
Original source: OpenAI News