OpenAI has shared how its reasoning model GPT-5 Pro helped an immunologist solve a puzzle that had stumped his lab for three years[1]. In an experiment involving T cells, the immune cells closely tied to cancer and autoimmune disease, the model offered a coherent mechanism for a phenomenon that experts could not explain. Notably, the AI did not deliver a diagnosis or a final verdict; it complemented human expertise and lit a path forward for the research.

An Experiment Shelved for Three Years

The mystery dates back to 2022[1]. Derya Unutmaz, an immunologist and professor at The Jackson Laboratory and the University of Connecticut, was studying how glucose, a type of sugar, affects the way T cells develop and take on their roles. T cells help the body fight viruses and cancer cells, respond to some bacteria and parasites, and distinguish healthy cells from threats, branching into different roles as they mature.

In the experiment, T cells early in their development were placed in either a low-glucose environment or one containing deoxyglucose, a glucose-like molecule[1]. Deoxyglucose interferes with a cell's ability to use glucose, disrupting energy production and protein construction. Because both conditions limit energy, the team expected similar results.

The outcome, however, was different[1]. The T cells exposed to deoxyglucose overwhelmingly produced cells involved in the inflammatory response. In the low-glucose environment, some cells did specialize as inflammatory-response cells, but not in the numbers seen with deoxyglucose. What is more, the effect of early exposure to deoxyglucose persisted even after the molecule was removed. A lack of energy alone could not explain it, suggesting something else was at work. Unable to pin down the reason at the time, the team shelved the experiment and moved on to other tasks.

The Answer GPT-5 Pro Pointed To: Glycosylation

The turning point came in late 2025[1]. With the release of GPT-5 Pro, Unutmaz fed the model an unpublished figure of flow-cytometry scatterplots showing the different T-cell subsets and asked what might explain the data and which experiments to run next.

GPT-5 Pro suggested that the driver was not a simple lack of energy but impaired N-linked glycosylation—the process by which cells attach sugar chains to proteins—during priming[1]. It further predicted that the responding cells were memory T cells rather than naive ones. The model then proposed concrete follow-up experiments, including an elegant mannose rescue experiment that restores glycoprotein maturation without restoring glycolysis. The lab had already run that very experiment, and the results matched the model's predictions exactly.

"GPT-5 came up with this really remarkable insight that retrospectively makes perfect sense," Unutmaz said[1]. The connection sat just outside his own area of expertise, which is why neither he nor anyone in his lab had spotted it.

Predicting an Unpublished Result

Unutmaz then tested whether GPT-5 Pro could predict the outcome of an experiment[1]. He chose one he had already carried out himself, involving T cells that target a type of lymphoma. His experiment had shown that these CD8+ T cells had an enhanced ability to kill lymphoma cells.

When he asked GPT-5 Pro to simulate the same experiment, the model correctly predicted the boost in the CD8+ cells' killing ability[1]. Since he had not yet published the results, the model could not have gleaned the answer from the internet. "That was the moment I felt like, okay, these models have now come to a point where they really, truly understand," he recalled.

Unutmaz says models like GPT-5 Pro now function more like collaborators[1]. They can process hundreds of new papers published every week to streamline literature reviews, help surface questions that remain unanswered, and assist in honing hypotheses. "The number of things you can do to address your hypothesis is vast," he said. "You have countless approaches, and you don't know which one will be the best strategy." So he uses the model to simulate experiments and predict outcomes, narrowing down which experiments are worth running in the lab. That can save weeks, months, or even years of work, potentially accelerating biology in a major way.

Expertise Still Required, and the Need for Caution

Even so, expertise does not become unnecessary[1]. AI may generate an insight, but it is still people who must judge its significance and plausibility. Without Unutmaz's level of expertise, OpenAI notes, one could not have told whether the mechanistic insight the model flagged in his immune cell experiments was important.

The power to accelerate research also demands responsible handling[1]. While it can advance biology and medicine, it could also lower barriers for bad actors seeking to design biological or chemical weapons. OpenAI describes its approach to tracking such risks and building safeguards against capabilities that could cause severe harm in its Preparedness Framework.

Unutmaz himself is optimistic about where AI is headed[1]. Lately he has also used Codex and GPT-5.2 Deep Research to compile large-scale cancer mutation datasets and to create research materials, including a substantial draft textbook focused on T cells, all aimed at accelerating precision immunotherapy. "To not only be able to witness it historically but participate a little bit, I feel truly lucky and privileged to do that," he said.

Summary

OpenAI's GPT-5 Pro offered immunologist Derya Unutmaz a coherent hypothesis—centered on impaired N-linked glycosylation and memory T cells—for a T-cell mystery he had shelved for three years, and even correctly predicted the outcome of an unpublished experiment[1]. Rather than handing over the answer itself, the model acts as a collaborator that proposes hypotheses experts can verify. The final judgment still rests on human expertise, and guarding against misuse remains essential, but this stands out as a concrete example of AI accelerating scientific discovery.

Source: https://openai.com/index/accelerating-science-gpt-5/