OpenAI o3 Deep Research for Rare Genetic Disease Diagnosis

Researchers from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University, and OpenAI have demonstrated that the OpenAI o3 Deep Research reasoning model can help physicians identify candidate explanations for rare genetic diseases that had previously evaded expert analysis. In a study published June 18, 2026, in NEJM AI, the AI-assisted workflow established diagnoses in 18 previously unsolved cases, representing an additional diagnostic yield of 4.8%.

AI-Assisted Reanalysis Yields New Diagnoses

OpenAI o3 Deep Research acted as an explanation-first reasoning layer to surface evidence-linked hypotheses for clinician review. The model did not make clinical decisions or diagnoses; instead, it synthesized clinical features, inheritance patterns, variant evidence, and scientific literature into justifications that specialists could interrogate.

Out of 376 previously analyzed unsolved cases, the model helped establish 18 new diagnoses. The results across specific cohorts are as follows:

Cohort	Cases	Diagnoses surfaced	Yield
Neurodevelopmental	100	10	10.0%
Neuromuscular disease	61	4	6.6%
Sudden unexpected death in pediatrics	200	2	1.0%
Early psychosis	15	2	13.3%
Total	376	18	4.8%

Of these 18 diagnoses, seven were rediscoveries of diagnoses established outside the local research workflow but missing from the reviewed records, highlighting the difficulty of synthesizing information across fragmented data sources.

Technical Workflow and Validation

The research team utilized a structured data packet for each case, consisting of standardized Human Phenotype Ontology terms, clinician notes, patient metadata (age and gender), and a filtered variant table containing rarity, predicted protein effect, ClinVar classification, and signal quality.

Validation on Solved Cases

Before analyzing unsolved cases, the team validated the workflow on known diagnoses:

• General rare conditions: Recovered the correct gene and variant in duplicate runs for 48 of 51 cases.
• Neuromuscular cases: Returned the correct diagnosis in duplicate runs for 45 of 57 cases.
• Long-read genome set: Named the correct gene in all 15 cases and both disease-causing alleles in 12 cases.

Confidence Scoring

The model provided self-reported confidence scores that correlated with accuracy. The mean minimum score for consistently correct calls was 85.6, compared to 42.1 for incorrect or unknown calls. These scores were used to guide expert reviewers toward the most promising candidates rather than as a substitute for clinical adjudication.

Advanced Genomic Insights and Hypotheses

The reasoning model demonstrated flexibility in identifying complex genetic events and proposing novel biological mechanisms:

• Structural Variant Inference: In one early-psychosis case, the model inferred a 22q11.2 deletion associated with DiGeorge syndrome by connecting low-quality calls on chromosome 22 with the patient's clinical features, despite the deletion not being explicitly listed in the input data.
• Digenic Explanations: The model identified cases where two genes better explained a presentation, such as combined variants in LAMA2 and FOXP1, and another involving TTN and SRPK3.
• Novel Mechanistic Hypotheses: The model proposed a possible novel link between an 11-amino-acid deletion in S1PR1 and vitiligo, suggesting the deletion alters receptor structure and signaling to reduce pigment production. This hypothesis requires further experimental validation.
• Phenotype Expansion: The model suggested a broader clinical spectrum for damaging variants in HSPB8 and CDK13 within the neuromuscular cohort.

Limitations and Clinical Guardrails

This study was retrospective and used de-identified information. The researchers emphasized that the model is a research tool and not a diagnostic device. Every AI-generated lead was subject to a rigorous human-led process: review by at least two team members using the ACMG/AMP framework, classification of the variant as pathogenic or likely pathogenic, confirmation by a CLIA-certified laboratory, and final physician review.

Key limitations include:

• The study did not measure time saved, cost, or changes in patient care.
• It did not systematically evaluate structural variants, repeat expansions, or mosaicism.
• LLMs can produce plausible but incorrect explanations, necessitating human adjudication.

Future Directions

The Manton Center, supported by a grant from the OpenAI Foundation, will develop a platform-agnostic, low-cost genetics AI copilot. Future research will focus on prospective multi-center studies to compare LLM-assisted reanalysis with standard practice regarding diagnostic yield, clinician effort, and the burden of false positives.