OpenAI o3 Deep Research Uncovers 18 New Diagnoses in 376 Unsolved Rare Disease Cases

376 unsolved rare disease cases, each already pored over by specialists. OpenAI o3 Deep Research found 18 new diagnoses that everyone else missed. That is a 4.8% additional diagnostic yield after years of expert analysis, and it comes from an AI that does not diagnose anything itself - it just surfaces hypotheses with evidence chains that clinicians can verify.

I have seen too many AI-in-medicine press releases that promise miracles and deliver noise. This one is different because the study, published June 18 in NEJM AI, forced the model to show its reasoning. Researchers from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard, and OpenAI assembled de-identified clinical and genomic packets for each of the 376 cases. The model had to connect phenotype terms (Human Phenotype Ontology), inheritance patterns, variant rarity, protein effects, ClinVar classifications, and the latest literature into a justification that a human could interrogate. No black box allowed.

Why 18 Diagnoses Matters More Than the Number

Roughly half of rare disease patients never get a genetic diagnosis after sequencing. Their genomes sit in a backlog, inert, while knowledge moves on. New gene-disease links, variant reclassifications, and fresh case reports accumulate daily. The problem is not that the data does not contain the answer - it is that nobody has the time to re-analyze thousands of unsolved genomes against a moving target. o3 Deep Research does not replace experts; it acts as an explanation-first reasoning layer that makes periodic reanalysis tractable.

The team validated the workflow first. On 51 previously solved cases with known rare conditions, it recovered the correct gene and variant in duplicate runs for 48. On 57 neuromuscular cases, 45 correct diagnoses. On a 15-case long-read genome set, it named the correct gene in every case and both disease-causing alleles in 12. Those numbers tell me the model understands when it is confident: its self-reported confidence score averaged 85.6 for correct calls versus 42.1 for incorrect ones. Not calibrated probabilities, but a useful signal for triaging expert attention.

How o3 Deep Research Actually Worked

Each case packet included standardized Human Phenotype Ontology terms, occasional clinician notes, age and gender, and a filtered variant table from the child and both biological parents. The model was asked to propose the most plausible molecular explanation and to show its work. Researchers then applied the ACMG/AMP framework - the same one clinical labs use - to classify variants. At least two team members reviewed each candidate; disagreements resolved by consensus. A finding counted as a diagnosis only after CLIA-certified lab confirmation and return to the family.

The model did not make clinical decisions. It produced evidence-linked hypotheses for specialists to review. That distinction is critical. The 18 new diagnoses came from cases that had evaded years of expert scrutiny. One key insight: many of these cases had data split across databases with different identifiers, formats, and vocabularies. The model is good at stitching fragmented records together into a coherent narrative.

What This Means for the Diagnostic Backlog

Every genetics clinic has a growing pile of unsolved genomes. The rate of new knowledge exceeds the rate of manual reanalysis. o3 Deep Research will not clear that backlog alone, but it can make expert-led periodic reanalysis scalable. The study suggests that institutions could run this workflow quarterly or yearly, flagging the most promising leads for human review. That is not a revolution - it is an engineered process improvement with measurable yield.

What I want to see next is a deployment with a real clinical workflow, not a retrospective study. The researchers already showed the model works on varied conditions: neurodevelopmental disorders, neuromuscular diseases, other pediatric syndromes. The next step is to integrate this into a hospital's genomic pipeline and measure the diagnostic yield over time as knowledge updates. If the 4.8% number holds consistently, that is tens of thousands of new diagnoses globally each year.

Source: Using AI to help physicians diagnose rare genetic diseases affecting children
Domain: openai.com