o3-assisted reanalysis yields 11 new diagnoses and seven rediscoveries

OpenAI o3 Deep Research, used as the reasoning layer of an LLM-assisted genome reanalysis workflow (Human Phenotype Ontology terms, clinician notes and a filtered variant table in; explanation-linked candidate diagnoses out)

Not assessed

Exact release not established.

OpenAI

The model proposed molecular explanations from clinical information and genetic variants.

Researchers used o3 Deep Research to analyze de-identified clinical information and genetic variants.

The model received clinical notes, phenotype terms, and filtered genetic variants. It proposed a molecular diagnosis with supporting evidence.

Not assessed

Researchers gave the model de-identified symptoms, clinical notes, and genetic variants that earlier analyses had not resolved. They refined the prompt on known diagnoses, then asked for explanations of unresolved cases. At least two experts reviewed each candidate, and a certified laboratory confirmed diagnoses before clinicians returned results to families.

Not assessed

o3 Deep Research read each de-identified case packet and proposed the most plausible molecular explanation with its supporting evidence. It diagnosed no one. Humans did the rest. At least two reviewers adjudicated each candidate under ACMG/AMP criteria, a CLIA-certified laboratory confirmed each variant, and the clinical team returned results to families. Its distinct contribution was surfacing candidates in cases that commercial and institutional analyses and multidisciplinary teams had already reviewed. In one early-psychosis case it inferred a 22q11.2 deletion absent from the input data, which genome sequencing then confirmed. Seven of the 18 were rediscoveries already established externally. The study was retrospective, and reviewers were not blinded to model confidence. The authors call for prospective evaluation with comparator baselines, so the gain over renewed human reanalysis is unmeasured.

o3 Deep Research read each de-identified case packet and proposed the most plausible molecular explanation with its supporting evidence. It diagnosed no one. Humans did the rest. At least two reviewers adjudicated each candidate under ACMG/AMP criteria, a CLIA-certified laboratory confirmed each variant, and the clinical team returned results to families. Its distinct contribution was surfacing candidates in cases that commercial and institutional analyses and multidisciplinary teams had already reviewed. In one early-psychosis case it inferred a 22q11.2 deletion absent from the input data, which genome sequencing then confirmed. Seven of the 18 were rediscoveries already established externally. The study was retrospective, and reviewers were not blinded to model confidence. The authors call for prospective evaluation with comparator baselines, so the gain over renewed human reanalysis is unmeasured.

Not assessed

Researchers confirmed 18 diagnoses in local records: 11 new diagnoses and seven already established elsewhere.

Improved Physical Health

Not assessed

People. Organizations.

Families whose unresolved local records were reexamined at Boston Children’s Hospital.

Eighteen families received confirmed local diagnoses, including 11 newly diagnosed families.

Not assessed

Not assessed

Yes

Not assessed

Not assessed

Not assessed

Experts reviewed the proposed diagnoses under ACMG/AMP criteria and arranged laboratory confirmation. Researchers counted confirmed diagnoses against 376 locally unresolved cases.

A retrospective reanalysis of 376 rare-disease records at Boston Children’s Hospital, mostly involving a child and both parents.

Confirmed diagnoses, distinguishing new diagnoses from those already established elsewhere.

prior local research records; no contemporaneous expert-reanalysis control

Reference: 376 cases unresolved in the local research record, including some diagnosed elsewhere. Result: 18 confirmed local diagnoses: 11 new and seven rediscoveries; 4.8% overall.

Not assessed

There was no simultaneous expert reanalysis using the same updated evidence. Some findings relied on later gene discoveries or revised variant classifications. The study did not measure treatment improvements, savings, or clinician time, and involved OpenAI employees and funding.

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Developer: OpenAI. Operator: Boston Children's Hospital, The Manton Center for Orphan Disease Research (with Harvard Medical School and OpenAI co-authors).

Benefit: Boston, United States. Developer: United States. Operator: United States.

Not assessed

Not assessed

Not assessed

Not assessed

Geneticists must review candidates and arrange confirmatory tests. The data need approved, protected handling. Expected yield depends on the prior work and the newer evidence available for each cohort.

Not assessed

The paper calls for prospective multicenter evaluation, but no later development after the June 2026 publication was found.

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

AIBD-0211-source-0001–0002

Numbered references below. [1] [2]

Not assessed

Developer, Independent [1] [2]

The study’s supplements and disclosure document were inaccessible. [1] [2]

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

People. Organizations.

People: 2. Organizations: 1.

Families received 18 confirmed local diagnoses, including 11 new findings. A patient interview describes the value of receiving an explanation. The study measured no changes in health or treatment, and no comparison isolates AI’s added value over an expert reanalysis.

Not assessed

Not assessed

Not assessed

Not assessed

Independently reported

Not assessed

Not assessed

Clinical genomics / rare disease diagnosis. Hospital rare-disease research center; retrospective reanalysis study.

OpenAI o3 Deep Research is a general-purpose reasoning and research model, not a diagnostic device. The Manton Center team built an explanation-first reanalysis workflow around it to propose candidate molecular diagnoses for unsolved cases. These diagnoses were the workflow's intended output. OpenAI states that the study does not describe or endorse a customer use of o3 for diagnosis. Brownstein said that she was shocked that a commercial system could find new diagnoses in genomes analyzed many times.

Date basis unassessed: 2026 (precision unassessed). Updated: 25 October 2026.

Not assessed

SHA-256: 62db4c890827…

oecd_aim: 2026-06-20-3da6.

Not assessed

Not assessed

8 revisions. Latest: 25 October 2026.