A single GPT-5 Pro session did what a top immunology lab couldn't do in three years: explain why a glucose analog called deoxyglucose forces T cells to become inflammatory Th17 cells while low glucose alone does not.
Derya Unutmaz, a professor at The Jackson Laboratory and the University of Connecticut, ran the original experiment in 2022. He exposed developing T cells to low glucose or to deoxyglucose, a molecule that blocks glucose utilization. Both conditions should starve the cells of energy. Yet deoxyglucose triggered a massive spike in Th17 inflammatory cells, while low glucose produced only a modest increase. The effect persisted even after removing the deoxyglucose. The lab shelved the puzzle.
How deoxyglucose tricked T cells
GPT-5 Pro, released in late 2025, made the connection Unutmaz and his team missed. When Unutmaz uploaded the data, the model zeroed in on a protein called IL-2. IL-2 normally acts as a brake that prevents T cells from becoming Th17 cells. Deoxyglucose, GPT-5 proposed, disrupts construction of IL-2. Remove that brake, and the cells freely differentiate into Th17.
"GPT-5 came up with this really remarkable insight that retrospectively, makes perfect sense," Unutmaz told OpenAI. The insight lay just outside his own expertise, which is exactly where LLMs can complement human researchers.
GPT-5 predicts an unpublished lymphoma experiment
Unutmaz then tested GPT-5's ability to simulate biology. He fed it the design of an experiment he had already run on CD8+ T cells targeting a specific lymphoma. The model correctly predicted that those cells would show enhanced killing capability. The results were not on the internet; Unutmaz had not published them yet.
"That was the moment that I felt like, okay, these models have now come to a point where they really, truly understand," he said. Subject matter expertise remains critical - someone without Unutmaz's background could not evaluate the plausibility of the model's output.
What this means for biology research
GPT-5 Pro's ability to process hundreds of new papers weekly and suggest experimental directions is reshaping how labs prioritize experiments. Unutmaz now uses it to simulate outcomes before stepping into the wet lab, cutting weeks to months - even years - from the hypothesis-testing cycle.
The model doesn't replace the scientist; it accelerates the loop between asking and answering. Unutmaz put it bluntly: doing science without AI now would be "like taking both of your hands away, or half of your brain away." The next step is using these simulations to design therapies for the very Th17-driven inflammatory diseases that deoxyglucose hinted at three years ago.
Source: How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
Domain: openai.com
Comments load interactively on the live page.