AI doesn't replace geneticists—it makes them faster at rare disease diagnosis

A collaboration between Boston Children's Hospital and OpenAI used an LLM to reanalyze 376 unsolved rare disease cases, leading to 18 new diagnoses. The model drastically reduces the search space from thousands of genetic variants to a handful of hypotheses, saving human analysts hours of work per case.

The key bottleneck is not the model's capability but rather the time-consuming process of filtering variants, checking databases, and reading literature. The LLM excels at data integration and literature synthesis, surfacing connections that human analysts miss due to time constraints. The model requires careful prompting and expert validation to avoid errors.

Teams should watch for this workflow scaling beyond rare disease to other diagnostic challenges. The true value is in enabling periodic reanalysis of old cases as new knowledge emerges—a task that is impractical manually. This creates an operational shift: the genome stays static, but the interpretation layer becomes dynamic.

The findings are compelling but preliminary. The 5% diagnosis rate on previously unsolved cases is meaningful but modest. Hype around "AI diagnosing diseases" should be tempered—the model augments expert analysts, not replaces them. The tradeoff is between AI speed and the risk of false positives that waste expert attention.