GPT-5 Pro helps Derya Unutmaz crack a three-year T-cell mystery

Basic science research moves slowly partly because datasets outlast the hypotheses that generated them. A flow cytometry dataset collected by immunologist Derya Unutmaz at The Jackson Laboratory for Genomic Medicine had been sitting largely unresolved since 2022, its central question about how glucose metabolism shapes T-cell fate stubbornly unanswered. OpenAI published a case study describing what happened when Unutmaz fed that data to GPT-5 Pro.

The lab had been treating human CD4+ T cells with 2-deoxyglucose (2DG), a compound that blocks glucose metabolism. After removing 2DG and priming the cells with IL-2, they observed a lasting shift toward a proinflammatory Th17-like state, a finding they could document but not explain. GPT-5 Pro reportedly proposed that disrupted N-linked glycosylation during priming was the mechanistic driver, and that memory T cells rather than naive T cells were the key population behind the effect. Unutmaz described the model's output as a "remarkable insight" that was outside his and his lab's immediate expertise.

The more striking part of the account is a prediction test Unutmaz ran separately. He asked GPT-5 Pro to forecast the outcome of an experiment he had already conducted involving CD8+ T cells targeting a type of lymphoma, and the model reportedly got it right, correctly predicting the boost in those cells' ability to kill the lymphoma cells. A correct prediction on a held-out experiment is a harder test than generating a plausible hypothesis on the same dataset.

The honest caveat is that OpenAI is the source telling this story, and the reported mechanistic explanation came paired with follow-up experiments, not a validated conclusion. The account also surfaces a constraint the retrieved material phrases directly: the approach worked because a domain expert had the right dataset, knew which question had remained unanswered, understood why earlier analysis had stalled, and could recognize when the model's answer was biologically plausible rather than merely fluent. That combination is not universally available.

If the mechanism holds up, the implications reportedly extend to cancer and autoimmune disease research. For labs sitting on stalled datasets, the more immediate question is whether their unanswered questions are in the same position Unutmaz's were in 2022: answerable in principle, waiting for the right question to be put to the right model.