OpenAI o3 helps find 18 rare disease diagnoses

This study addresses a persistent problem in rare disease care: many patients remain undiagnosed even after genomic sequencing and expert review. The researchers aimed to show that an AI reasoning model could help specialists revisit old cases as medical knowledge expands, making the reanalysis process more efficient and scalable. By combining clinical features, genetic variant data, and scientific literature into reviewable hypotheses, the model acted as a "reasoning layer" on top of existing genomic pipelines, not as a replacement for clinical judgment.

The study's results are modest but meaningful in this context—a 4.8% diagnostic yield in a heavily pre-analyzed population aligns with single-digit gains seen in similar reanalysis studies. Notably, 7 of the 18 diagnoses were rediscoveries, meaning they were found outside the local workflow but absent from the records reviewed, underscoring the challenge of synthesizing information across fragmented data sources. The model also demonstrated flexibility by suggesting structural variants and digenic explanations that were not explicitly requested, leading to confirmed diagnoses in some cases.

Looking ahead, the Manton Center will lead the next stage through a grant from the OpenAI Foundation, aiming to develop a platform-agnostic, low-cost genetics AI copilot. The success of this approach will hinge on whether future prospective studies can show tangible benefits in diagnostic yield, time, and cost while maintaining rigorous privacy and safety standards. For families with unanswered genetic questions, this research suggests that today's inconclusive results may not be permanent.