A New York-based immunologist says GPT-5 Pro helped him make sense of a three-year-old experiment that had stumped his lab, offering a possible explanation for how a nutrient affects the development of immune cells tied to cancer, autoimmune disease and infection.
Derya Unutmaz, a professor at The Jackson Laboratory and the University of Connecticut, said the model helped him revisit research from 2022 that examined how glucose influences the way T cells specialize. T cells are central to the immune system’s response to viruses, bacteria, parasites and cancer, and scientists have long been interested in how their development can push them toward different roles in the body.
Unutmaz’s original experiment compared two conditions in early T cell development. In one, the cells were placed in a low-glucose environment. In the other, the cells were exposed to deoxyglucose, a glucose-like molecule that interferes with how cells use sugar. The team expected similar results, since both conditions limit the cells’ access to glucose-related energy. Instead, the deoxyglucose group produced far more inflammatory-response cells, while the low-glucose group did not show the same effect. The result persisted even after the molecule was removed.
At the time, Unutmaz and his team could not explain the difference and set the question aside. That changed when GPT-5 Pro became available in late 2025. After uploading the data, Unutmaz asked the model to analyze it.
According to OpenAI’s account, GPT-5 Pro suggested that deoxyglucose was interfering with the production of IL-2, a protein that can keep T cells from becoming a Th17 cell, a type associated with inflammatory responses. In that interpretation, deoxyglucose would effectively remove a brake on Th17 development, helping explain why the cells in that condition shifted so strongly toward the inflammatory pathway.
Unutmaz said the explanation was one he had not considered, even though it made sense in hindsight. He described the insight as being just outside his own expertise and outside the immediate assumptions of his lab.
The immunologist then tested the model in another way. He asked GPT-5 Pro to simulate a separate experiment involving CD8+ T cells that target lymphoma. In Unutmaz’s own study, those cells showed an improved ability to kill lymphoma cells. The model reportedly predicted the same result before the findings had been published, which Unutmaz took as evidence that the system was doing more than simply retrieving public information.
For Unutmaz, the experience suggests that large language models are becoming research collaborators rather than just search tools. He said they can help scientists review the flood of papers published each week, narrow down unanswered questions and simulate possible experimental outcomes before lab work begins.
That, in turn, could save researchers weeks, months or even years by helping them choose the most promising hypotheses to test. But Unutmaz also stressed that AI output still needs expert scrutiny. A model may surface a useful idea, but specialists must decide whether it is biologically sound and worth pursuing.
The story also points to broader questions about the pace and risks of AI-assisted science. OpenAI said the same capabilities that may accelerate biology and medicine could also lower barriers for misuse, including by people seeking to design harmful biological or chemical agents. The company said its Preparedness Framework is meant to track and mitigate those risks.
Still, Unutmaz is optimistic. He has also used advanced AI tools for tasks including large-scale cancer mutation datasets and T cell-focused research materials. In his view, the current wave of AI differs from past technological shifts because it can actively contribute to scientific reasoning.
For now, he said, the opportunity to witness and participate in that shift feels like a rare privilege.