AI Reasoning Model Generates Diagnostic Leads for Unresolved Rare Disease Cases

As scientific evidence evolves and clinical data remain fragmented across systems, many cases continue to go unresolved. Although reanalyzing existing records can reveal missed variants or overlooked disease links, the process places a substantial workload on clinical laboratories. A new study now shows how an artificial intelligence (AI) reasoning model can help specialists revisit previously unsolved pediatric cases and generate testable leads for further evaluation.

Boston Children’s Hospital’s Manton Center for Orphan Disease Research, with Harvard University and OpenAI, used the OpenAI o3 Deep Research reasoning model to reanalyze de-identified clinical and genomic data from difficult cases. The workflow asked the model to propose the most plausible molecular explanation and present its supporting evidence across phenotype, inheritance, variant annotations, data-quality patterns, and scientific literature. Human experts then reviewed outputs using ACMG/AMP criteria and proceeded to additional testing only when warranted.

Each case file included standardized Human Phenotype Ontology (HPO) terms, selected clinician notes, demographic metadata, and a filtered variant table capturing rarity, predicted protein effect, ClinVar status, and signal quality across available family members, often trios. A result counted as a diagnosis only after qualified clinicians adjudicated the evidence, a Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory confirmed a pathogenic or likely pathogenic variant, and the care team returned the finding to the family. The model itself did not render diagnoses; it generated evidence-linked hypotheses to focus expert review.

The study, published June 18, 2026, in NEJM AI, applied the workflow to 376 previously unsolved cases spanning neurodevelopmental disorders, rare neuromuscular disease, early psychosis, and sudden unexpected death in pediatrics. Physicians established 18 diagnoses after expert review and clinical confirmation, yielding a 4.8% increase. Cohort-level yields were 10.0% (neurodevelopmental), 6.6% (neuromuscular), 13.3% (early psychosis), and 1.0% (sudden unexpected death in pediatrics).

Before examining unsolved cases, the team refined prompts on previously solved datasets. The approach recovered the correct gene and variant in duplicate runs for 48 of 51 mixed rare-disease cases, returned the correct diagnosis for 45 of 57 neuromuscular cases, and identified the correct gene in all 15 long-read genomes while naming both disease-causing alleles in 12. Model confidence scores tracked with correctness (mean 85.6 for consistent true calls versus 42.1 for incorrect or unknown) but were not used for adjudication.

The workflow demonstrated flexibility by inferring a 22q11.2 deletion from low-quality call patterns in an early-psychosis case, later confirmed by follow-up sequencing. It also surfaced digenic explanations in select cases and highlighted a possible mechanistic link between an S1PR1 deletion and vitiligo that will require experimental validation. Reported limitations include the retrospective design, heterogeneous cohorts, lack of blinding to model confidence, and no measurement of time, cost, clinician effort, or false-positive workload; prospective multi-center comparisons were recommended.

The authors emphasized that the research does not support using general-purpose models for diagnosis and that every result passed through human adjudication and laboratory confirmation. Newer general-purpose or purpose-built systems were not evaluated, and broader deployment would require attention to privacy, security, auditability, and local regulation. The Manton Center will lead the next phase through an OpenAI Foundation grant to develop a platform-agnostic, low-cost genetics AI copilot for rare-disease case analysis.

“The bottleneck is time. An expert can devote only so much of their day to any one particular person,” said Dr. Catherine Brownstein, Boston Children’s Hospital’s Manton Center for Orphan Disease Research.

“Researchers like Catherine and me can’t possibly keep 8,000 different diseases in our heads. That’s the power of AI,” said Alan Beggs, director of the Manton Center for Orphan Disease Research.

Related Links
Boston Children’s Hospital’s Manton Center for Orphan Disease Research