For many families affected by rare disease, genomic sequencing does not immediately deliver an answer. Even after extensive testing and specialist review, roughly half of patients still remain without a clear genetic diagnosis.
The difficulty is not only the amount of data. A single case can involve thousands or millions of genetic variants, incomplete clinical records, changing variant classifications and a scientific literature base that keeps expanding. What once looked inconclusive can become meaningful years later.
Researchers tested an AI-assisted workflow
A team from Boston Children’s Hospital’s Manton Center for Orphan Disease Research, Harvard University and OpenAI evaluated whether OpenAI o3 Deep Research could help experts revisit rare disease cases that had already resisted diagnosis.
The study, published on June 18, 2026, in NEJM AI, used de-identified clinical and genomic information from 376 previously reviewed but unsolved cases. The model was used to generate evidence-linked candidate explanations for expert review.
After clinical specialists reviewed the candidates, ordered additional testing where needed and confirmed findings through established clinical processes, physicians were able to diagnose 18 cases. That added a diagnostic yield of 4.8% beyond prior specialist analysis.
The model generated leads, not diagnoses
The researchers emphasized that the AI system did not diagnose patients, make treatment decisions or determine what information should be returned to families. Its role was to identify possible explanations that human experts could investigate.
A case was counted as diagnosed only when qualified experts reviewed the evidence, classified the variant as pathogenic or likely pathogenic, confirmed it in a CLIA-certified laboratory and returned the result through the clinical team.
Why reanalysis can change the outcome
Rare disease genetics is constantly changing. New gene-disease associations are reported, public databases are updated, old variants are reclassified and new case reports add evidence that did not exist when a patient was first tested.
At the same time, patient information can be scattered across systems that use different formats and vocabularies. Symptoms, family history, sequencing results and clinician notes may not be easy to connect. Periodic reanalysis can help bring those pieces together against a newer scientific background.
How each case was prepared
For every case, the team created a de-identified packet for analysis. These packets included standardized Human Phenotype Ontology terms describing the patient’s features, occasional clinician notes, any descriptive clinical diagnosis, age and gender information and a filtered genetic variant table.
The variant table summarized rarity, predicted effect on the protein, ClinVar classification and signal quality across available family members. Most cases included genomic data from the child and both biological parents.
An explanation-first design
The AI workflow was not built to simply rank genes. Researchers asked the model to connect symptoms, inheritance pattern, variant evidence and scientific literature into a coherent explanation that reviewers could examine.
Specialists then assessed the model’s candidates using the ACMG/AMP framework used by clinical laboratories to classify genetic variants. At least two team members reviewed each candidate, and disagreements were settled by consensus.
Workflow tested before unsolved cases
Before applying the process to unresolved cases, the team evaluated it on cases with known answers. In 51 rare disease cases, the workflow recovered the correct gene and variant in duplicate runs for 48 cases.
In 57 neuromuscular cases, it returned the correct diagnosis in duplicate runs for 45 cases. In a 15-case long-read genome set, it named the correct gene in all 15 cases and identified both disease-causing alleles in 12 cases.
The model’s self-reported confidence scores also helped reviewers prioritize work. Consistently correct calls had a mean minimum score of 85.6, while incorrect or unknown calls had a mean minimum score of 42.1. The scores were not treated as calibrated probabilities or as a replacement for expert judgment.
Diagnostic yield by cohort
The researchers then applied the workflow to four groups of previously unsolved cases: neurodevelopmental conditions, rare neuromuscular disease, early psychosis in children and adolescents and sudden unexpected death in pediatrics.
| 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% |
The early psychosis group was small, so its percentage carries a wide confidence interval. The yield also depended on whether a given cohort was likely to involve a single-gene cause.
Some answers were already in fragmented records
Among the 18 diagnoses, seven were rediscoveries. They had been established outside the local research workflow but were missing from the record the team reviewed.
In several cases, the relevant variants had already been listed as pathogenic or likely pathogenic in public databases. That finding highlighted a practical challenge in rare disease work: important information can exist, but still fail to reach the right clinical record at the right time.
Complex cases produced unexpected leads
One early psychosis case showed how the model could point beyond the variants explicitly listed in the input. It noticed a pattern of low-quality calls on chromosome 22 and connected that signal with cardiac, immune, neurodevelopmental and psychiatric features.
From that pattern, it proposed a 22q11.2 deletion associated with DiGeorge syndrome. Follow-up genome sequencing later confirmed the deletion.
In other cases, the model suggested that two genes together better explained the patient’s condition. Variants in LAMA2 and FOXP1 helped explain combined muscle and neurodevelopmental findings in one case. Another involved a previously unrecognized digenic explanation connected to TTN and SRPK3.
A new biological hypothesis for vitiligo
The workflow also generated a possible research lead involving vitiligo. In one neurodevelopmental case, it identified an 11-amino-acid deletion in S1PR1 in a person with vitiligo.
S1PR1 encodes a cell-surface receptor involved in signaling, immune-cell movement and tissue biology. The model connected evidence suggesting that the deletion could affect receptor structure and signaling, potentially reducing pigment production while helping immune cells remain in the skin.
This proposed S1PR1-vitiligo relationship has not been proven and requires experimental validation. Still, it showed how AI-assisted analysis may help translate scattered evidence from immunology, structural biology and clinical genetics into testable ideas.
Broader disease patterns may emerge
The neuromuscular cohort also raised the possibility that some genes may have wider clinical effects than currently recognized. Damaging variants in HSPB8 and CDK13 did not fully match the best-known disorders linked to those genes.
The researchers said more cases and laboratory work will be needed to determine whether those findings represent true phenotype expansion.
Kyra receives an answer after years
One of the diagnoses involved Kyra, whose symptoms first became clear when she was 9 years old. Her mother noticed during karate class that Kyra was no longer getting as low in her stances. She also slowed down during soccer practice and began walking and running on her toes.
Her pediatrician could not identify the cause of her muscle weakness, beginning a nearly 20-year search involving tests, treatments and specialist consultations.
The reanalysis linked Kyra’s condition to a frameshift variant in HSPB8. She was diagnosed with a form of myofibrillar myopathy, a disease in which abnormal protein structures accumulate in muscle fibers and contribute to weakness.
A genetic counselor from the Manton Center called Kyra about a week before her 28th birthday. By then, she had used a wheelchair and depended on a ventilator since age 13, although her condition later plateaued. Because her form of myofibrillar myopathy is extremely rare, much remains unknown about its long-term course, but the diagnosis offered long-awaited clarity.
Clear limits on medical use
The study does not show that patients, clinicians or customers should use OpenAI models to diagnose disease or make medical decisions. It also does not endorse OpenAI o3 Deep Research, ChatGPT or any other OpenAI product as a diagnostic tool.
All diagnoses in the study were made by physicians and qualified clinical experts through standard clinical review, additional testing and laboratory confirmation.
Study limitations remain important
The study was retrospective and included heterogeneous cohorts. Reviewers were not blinded to the model’s confidence scores.
The researchers did not measure time saved, cost, clinician workload, false-positive burden or changes in patient care. They also did not systematically evaluate every type of genetic variation, including structural variants, repeat expansions, deep-intronic changes and mosaicism.
Large language models can misread context or produce explanations that sound plausible but do not hold up under expert review. In this study, the model broadened the search and helped focus human analysis, but it did not determine the final answer.
Privacy and safeguards
The research used de-identified information. No protected health information was used or transmitted outside approved environments.
Any wider clinical deployment would require strong controls for privacy, security, auditability and local regulation. AI access would not replace sequencing infrastructure, genetic counseling, confirmatory testing or specialist expertise.
Next stage of research
Future prospective, multi-center studies will need to compare AI-assisted reanalysis with standard practice. Researchers said key measures should include diagnostic yield, time to candidate, clinician effort, false-positive workload, cost and effects on patient care.
Reproducibility and safety will also require versioned prompts, reference checks, audit logs and better-calibrated uncertainty estimates.
From research model to genetics copilot
This study used OpenAI o3 Deep Research. Newer general-purpose models can search and synthesize more scientific material, while purpose-built systems such as GPT-Rosalind are designed for deeper life-sciences work, including analysis of variant effects on protein structure and function.
Those newer capabilities were not tested in this study and will need their own evaluations and access controls.
OpenAI supported the initial research, while the next phase will be led by the Manton Center through a grant from the OpenAI Foundation. The grant will support work on a platform-agnostic, low-cost genetics AI copilot for clinical teams analyzing rare disease cases.
Keeping pace with discovery
The central idea is not that AI replaces medical diagnosis. It is that carefully reviewed AI-generated leads may help specialists revisit unresolved cases as science advances.
For families still waiting, that could mean an old inconclusive result may one day become a path to an answer.
