Rare pediatric diseases often linger without a genetic answer despite whole‑genome sequencing. About half of affected children stay undiagnosed after exhaustive testing, because clinicians must wade through millions of variants and fragmented records. Researchers from Boston Children’s Hospital’s Manton Center, Harvard and OpenAI applied the o3 Deep Research reasoning model to 376 previously unsolved cases, hoping to surface evidence‑linked hypotheses for review.

Each case packet combined standardized Human Phenotype Ontology terms, parental genotype data and a filtered variant table. The model was prompted to propose the most plausible molecular explanation and to display its reasoning. In validation runs on 51 known diagnoses, the system recovered the correct gene in 48 cases; confidence scores above 80 correlated with accurate calls, guiding reviewers toward the most promising leads.

After expert review, additional testing and CLIA‑certified confirmation, physicians established diagnoses in 18 of the 376 cases, yielding a 4.8% increase over prior analyses. Seven of those were rediscoveries already listed in public databases, underscoring the difficulty of synthesizing dispersed evidence. The study demonstrates that an explanation‑first AI layer can generate testable hypotheses, making periodic reanalysis of legacy genomes more scalable.