GPT-5 helps immunologist Derya Unutmaz solve a 3-year-old T cell mystery

GPT-5 is now predicting experimental results to crack open immunology puzzles. On June 23, 2026, OpenAI shared how GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old T cell mystery. The model identified N-linked glycosylation interference as the reason the glucose inhibitor 2-DG spiked inflammatory T cells, and it proposed a mannose rescue experiment whose result it predicted exactly. ASAP summarizes the announcement from the primary source.

The 3-year paradox: starve the cells, and they multiply

Immunologist Derya Unutmaz had spent 3 years unable to explain why a glucose inhibitor spiked inflammatory T cell subsets. The glucose inhibitor 2-DG (2-deoxyglucose) was expected to starve T cells of energy, yet it sharply increased inflammatory subsets instead. If it had only blocked energy, cell activity should have fallen, so the opposite result was the 3-year question. Unutmaz's lab had no mechanism that explained the paradox.

On June 23, 2026, OpenAI described how GPT-5 worked through the puzzle. Unutmaz fed the model unpublished flow cytometry data, and the model returned not a single answer but a causal hypothesis paired with a test to confirm it.

The mechanism GPT-5 identified: structure, not metabolism

GPT-5 reasoned that 2-DG triggers N-linked glycosylation interference, disrupting the sugar chains on cell-surface receptors. The explanation is that the inhibitor is not merely starving the cells of energy but physically disturbing the sugar coatings on their receptors. The model moved past the metabolic assumption to a structural cause.

Why this case matters: not an answer, but a testable hypothesis

The weight of this case is not that the model got the answer right, but that it delivered the answer in a testable form. Instead of stopping at the observation of the 2-DG paradox, the model proposed a mannose rescue experiment to bypass the block, and it predicted exactly the result that Unutmaz's lab had already run but never published. That the data was unpublished is the crux: it means the model got right a result that could not have been in its training data. In other words, it did not look up an answer; it reasoned out a mechanism and computed the outcome of the next experiment. In science, value comes not from a "plausible explanation" but from a "prediction that can be proven wrong," and GPT-5's answer belongs to the latter.

How to read the numbers

The figure of "3 years" stands in for the difficulty of the problem. That a single lab held onto it for 3 years without building a mechanism means the paradox was a genuine gap, not a trivial measurement error. Still, it is too early to generalize from one success. The result rests on the premise that an expert like Unutmaz selected clean data to feed in and had the capacity to verify the model's output through experiment. To avoid overstatement, one has to read this alongside the fact that AI did not complete the discovery alone; the expert's question design and verification capability filled half of the outcome.

What it means for research on the ground in Korea

The same mechanism could lead to cancer therapy that boosts CAR-T cells' tumor-killing capacity, and the finding may give leads for autoimmune research as well. For bio and medical research in Korea, the point worth chewing on is that the decisive edge may lie not only in large-scale experimental facilities but in accumulated unpublished data and good questions. If flow cytometry data locked in a drawer becomes the biggest lever precisely when paired with a model, then how you build out data management and experimental-verification pipelines is what separates the competitive from the rest. AI is moving from generating hypotheses to predicting experiments and speeding up research, but without the verification capacity to catch that speed, a prediction remains just a plausible sentence.

Source: ASAP summary of OpenAI's "How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery" (June 23, 2026; GPT-5 solved immunologist Derya Unutmaz's 3-year-old T cell mystery, the glucose inhibitor 2-DG spiked inflammatory T cells, the model proposed N-linked glycosylation interference, it predicted the mannose rescue result, and suggested cancer CAR-T and autoimmune applications).