OpenAI's immunology story shows a realistic version of scientific acceleration: an expert revisits a hard problem, the model surfaces a plausible mechanism, and the lab gets a better path for follow-up.
The case centers on immunologist Derya Unutmaz and a three-year-old question about how glucose affects T cell development. His lab had shelved the experiment after the results did not fit the expected explanation.
The model helped reconnect the mechanism
OpenAI says GPT-5 Pro suggested that deoxyglucose interfered with construction of IL-2, a protein involved in whether T cells become Th17 inflammatory-response cells. That gave the lab a mechanism that fit the confusing experimental pattern.
The model also predicted the outcome of another experiment involving CD8+ T cells and lymphoma cells, according to the OpenAI post. The result had not yet been published, so the claim sits inside OpenAI's reported case rather than public independent validation.
More shelved problems become reviewable
This is the realistic near-term impact: not science becoming automatic, but more abandoned or stuck problems becoming easier to inspect again.
Models can help researchers scan literature, connect adjacent mechanisms, propose explanations, and narrow which experiments deserve time. Human expertise still decides what to test, what the evidence supports, and how to handle uncertainty.
If this trajectory holds, we will hear more stories like this. The surprising part will become less surprising: a model catching a connection just outside the team's working frame, then handing the expert a better question.
Related services: Quality, Knowledge, OpenAI expertise