AI-Assisted Reanalysis of Rare Genetic Diseases Using Reasoning Models

openai · Healthcare · 2026 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Rare genetic disease reanalysis

Researchers wanted to revisit unsolved rare-disease cases where clues are scattered across variants, records, family data, and evolving literature.

Constraints:

Architecture

The rare-disease workflow places o3 Deep Research as an explanation-first reasoning layer over already-analyzed clinical/genomic case packets. Each packet combines standardized phenotype terms, clinician notes, metadata, and filtered variant evidence. The model proposes plausible molecular explanations with evidence, then human experts review candidates using ACMG/AMP clinical-lab standards, resolve disagreements, order additional testing when appropriate, and count diagnoses only after clinical confirmation.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  de-identified-case-packets["de-identified case packets"]
  variant-evidence-table["variant evidence table"]
  o3-deep-research["o3 Deep Research"]
  specialist-review["specialist review"]
  additional-testing["additional testing"]
  clinical-confirmation["clinical confirmation"]
  de-identified-case-packets --> variant-evidence-table
  de-identified-case-packets --> o3-deep-research
  variant-evidence-table --> o3-deep-research
  o3-deep-research --> specialist-review
  specialist-review --> additional-testing
  additional-testing --> clinical-confirmation
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
de-identified case packets Clinical and genomic packets containing HPO terms, clinician notes, metadata, and filtered variant tables. o3 Deep Research, variant evidence table
variant evidence table Structured variant evidence including rarity, predicted protein effect, ClinVar classification, family signal quality, and inheritance context. de-identified case packets, o3 Deep Research
o3 Deep Research Reasoning model that proposes evidence-linked molecular explanations and shows its work. de-identified case packets, variant evidence table, specialist review
specialist review Human ACMG/AMP-style review by at least two team members, with consensus resolution. o3 Deep Research, additional testing
additional testing Follow-up testing and clinical-laboratory confirmation for candidate diagnoses. specialist review, clinical confirmation
clinical confirmation Final physician/lab confirmation stage; model output alone is never diagnosis. additional testing

Data flow

The team assembles a de-identified packet for each unsolved case, including HPO terms, notes, metadata, and a filtered variant table. o3 Deep Research reads that packet and proposes the most plausible molecular explanation with its reasoning. Specialists review the candidate, resolve disagreements by consensus, and only then pursue additional testing and clinical confirmation.

4 supporting quotes
“For each case, the team assembled a de-identified packet containing standardized Human Phenotype Ontology terms to describe the patient’s clinical presentation, occasional clinician notes and any descriptive clinical diagnosis, metadata such as age and gender, and a filtered variant table.”article, stated
“The table captured each variant’s rarity, its predicted effect on the encoded protein, ClinVar classification, and signal quality across available family members.”article, implied
“The team asked the model to propose the most plausible molecular explanation and to show its work.”article, stated
“At least two team members reviewed each candidate, disagreements were resolved by consensus, and a model output was never treated as a diagnosis.”article, stated

Why this architecture

The system is intentionally assistive rather than autonomous because rare-disease diagnosis is high-stakes and evidence evolves over time. The model helps specialists connect phenotype, inheritance, variant evidence, and literature, but it does not diagnose patients or make clinical decisions; human review and lab confirmation are the safety boundary.

4 supporting quotes
“Even with genomic sequencing, many people with rare diseases never receive a clear genetic diagnosis. Roughly half remain undiagnosed after extensive testing and specialist review.”article, implied
“As new gene-disease relationships, case reports, and classification evidence accumulate, unsolved cases can become newly interpretable.”article, stated
“Instead of returning only a ranked gene, it was asked to connect the clinical features, inheritance pattern, variant evidence, and scientific literature into a justification that a human reviewer could interrogate.”article, stated
“The model did not diagnose any patient or make any clinical decision. It produced evidence-linked hypotheses for specialists to review and, where appropriate, investigate through additional testing and confirm in a clinical laboratory.”article, stated

Models used: OpenAI o3 Deep Research

Scale

Alternatives considered (roads not taken)

Ranked-gene output without explanatory justification

Rejected because: The workflow needed interrogable reasoning connecting clinical features, inheritance, variant evidence, and literature.

“Instead of returning only a ranked gene, it was asked to connect the clinical features, inheritance pattern, variant evidence, and scientific literature into a justification that a human reviewer could interrogate.”article, stated

Let the model make or replace diagnoses

Rejected because: Outputs were only candidate hypotheses for expert review, testing, and clinical-lab confirmation.

“The model did not diagnose any patient or make any clinical decision. It produced evidence-linked hypotheses for specialists to review and, where appropriate, investigate through additional testing and confirm in a clinical laboratory.”article, stated

Patterns

human-in-the-loop-gate (clinical-safety)

The model generated hypotheses, but physicians and clinical labs performed review, testing, and confirmation.

“The model did not diagnose any patient or make any clinical decision.”article, implied

confidence-thresholding (quality)

Self-reported confidence scores were analyzed against known-case correctness.

“The model’s self-reported confidence scores tracked with correct diagnoses in these previously solved cases: the mean minimum score was 85.6 for consistently correct calls and 42.1 for incorrect or unknown calls.”article, stated

Tradeoffs

How to deploy reasoning models in medicine.

Chose expert-reviewed hypothesis generation over direct automated diagnosis.

Rationale: The workflow gains lead generation while avoiding unsupported clinical decision-making by the model.

“It produced evidence-linked hypotheses for specialists to review and, where appropriate, investigate through additional testing and confirm in a clinical laboratory.”article, stated

Gotchas (what bit them)

Old unsolved cases can become solvable as scientific knowledge changes.

Root cause: Variants, phenotypes, and literature evolve after initial sequencing and may be split across incompatible records.

Resolution: Periodically reanalyze with an explanation-first model layer and expert adjudication.

“Rare-disease reanalysis is both a scientific and a maintenance problem.”article, stated

Evaluation

The researchers refined on solved cases, measured duplicate-run recovery, confidence calibration, and additional yield on unsolved cases.

Metrics: 48/51 solved-case recovery · 45/57 neuromuscular cases · 18 diagnoses · 4.8% additional yield

“Following expert review, additional testing, and clinical confirmation, physicians established diagnoses in 18 cases—an additional diagnostic yield of 4.8% after earlier analysis by specialists.”article, stated

Outcomes

Research workflow, not production clinical deployment; marked non-agentic because the model is a reasoning assistant layer.

“This research is not evidence that patients, clinicians, or customers should use OpenAI models to diagnose disease or make medical decisions.”article, implied