OpenAI's GPT-5 Pro Cracks Long-Stalled Mystery in T Cell Research
A three-year-old immunology puzzle finally has an answer, and it came from an AI model rather than a lab bench.
Dr. Derya Unutmaz, a professor at The Jackson Laboratory and the University of Connecticut, shelved a 2022 T cell experiment after his team couldn't explain its results. What cracked it is a clear illustration of where GPT-5 medical research now stands: he uploaded the data to GPT-5 Pro in late 2025 and the model surfaced a protein interference mechanism the entire lab had missed.
The Experiment That Went Unsolved for Three Years
T cells are immune cells that fight viruses, kill cancerous cells, and distinguish healthy tissue from threats. As they mature, they specialize into distinct roles that can shape cancer outcomes, autoimmune conditions, and vulnerability to infection. Unutmaz's 2022 experiment set out to understand how glucose affects which specialization path a T cell takes.
His team exposed T cells to two conditions. One group was placed in a low-glucose environment. The other was treated with deoxyglucose, a glucose-like molecule that blocks a cell's ability to metabolize sugar and interferes with energy production and protein synthesis.
The two conditions should have produced similar results. Instead, the deoxyglucose group overwhelmingly developed into Th17 cells, the T cells associated with inflammatory response.
Far fewer Th17 cells appeared in the low-glucose group. The gap couldn't be explained by energy limitations alone. Unutmaz and his lab couldn't crack it, so the experiment went back in the drawer.
What GPT-5 Pro Found in the Data
After uploading the shelved results to GPT-5 Pro, Unutmaz received a specific mechanistic suggestion. The model proposed that deoxyglucose disrupts the construction of a protein called IL-2. That matters because IL-2 normally acts as a biological brake, preventing T cells from differentiating into Th17 cells. By disrupting IL-2 production, deoxyglucose removed that brake and pushed cells toward the inflammatory pathway at much higher rates.
"GPT-5 came up with this really remarkable insight that retrospectively, makes perfect sense," Unutmaz said. The explanation sat just far enough outside his specific area of expertise that neither he nor anyone in his lab had made the connection.
He then ran a second test. Unutmaz asked GPT-5 Pro to simulate an experiment he had already completed: one examining CD8+ T cells and their ability to target a type of lymphoma. The model correctly predicted the outcome, including the enhanced killing capacity of those cells.
Unutmaz hadn't yet published those results, so the prediction came from scientific reasoning rather than cached internet data.
"That was the moment that I felt like, okay, these models have now come to a point where they really, truly understand," he said.
Faster Science, Still Human-Led
Unutmaz now describes GPT-5 Pro as a collaborator. The model can process hundreds of newly published academic papers each week, help narrow which experiments are worth running, and simulate outcomes before researchers commit lab time. He estimates that AI-assisted filtering can cut weeks, months, or even years from research timelines.
Since solving the T cell puzzle, he has used Codex and GPT-5.2 Deep Research to compile large-scale cancer mutation datasets and generate materials for precision immunotherapy work, including a draft T-cell textbook.
Subject matter expertise remains essential. Unutmaz is clear that someone without his immunology background couldn't have judged whether GPT-5 Pro's mechanistic finding was significant or even plausible. In GPT-5 medical research more broadly, the model surfaces possibilities. Researchers still determine which ones are worth pursuing.
OpenAI also notes that faster biological research capabilities carry dual-use risks. The company's Preparedness Framework outlines how it monitors risks from AI capabilities that could be misused in biology or chemistry.
The full account of Unutmaz's case, including further context on T cell specialization and cancer research applications, is available at the OpenAI GPT-5 immunology research page.