GPT-5 Pro and a shelved T-cell experiment: the chat window as research surface

A three-year-old dataset, uploaded into a text box

The concrete mechanic in OpenAI's story is unglamorous and revealing: Derya Unutmaz, a professor at The Jackson Laboratory and the University of Connecticut, took results from an experiment his lab had shelved in 2022 and uploaded them into GPT-5 Pro. He asked the model to analyze the data. That is the entire frontend of the discovery — a file, a prompt, an answer.

The underlying puzzle was specific. His team had exposed developing T cells to either a low-glucose environment or to deoxyglucose, a glucose-like molecule that disrupts a cell's ability to use glucose. They expected similar results, since both conditions limit the energy T cells need. Instead, the deoxyglucose-exposed cells overwhelmingly became inflammatory-response cells, and the effect persisted even after the molecule was removed. The lab couldn't explain it, so they moved on.

What matters for anyone building AI-assisted tools is how low the barrier to reopening this was. No pipeline, no fine-tuning, no schema. The knowledge sat dormant for three years because there was no cheap way to interrogate it again. The interface removed that cost.

The IL-2 insight came out of a single query

According to OpenAI, GPT-5 Pro proposed that deoxyglucose interfered with construction of a protein called IL-2, which can prevent T cells from becoming an inflammatory-response cell known as Th17. By blocking IL-2, deoxyglucose effectively removed a barrier to Th17 formation — explaining why low-glucose conditions didn't produce Th17 cells at the same numbers.

GPT-5 came up with this really remarkable insight that retrospectively, makes perfect sense.Montana Labs

Unutmaz notes the connection sat just outside his own expertise, which is why neither he nor anyone in his lab saw it. That framing is precise and useful: the model's value here wasn't superhuman reasoning but adjacency — pulling in a mechanism from a neighboring subfield. The output still had to land in front of someone who could recognize it as plausible immunology rather than confident nonsense.

Simulating an unpublished experiment as a validation test

The more interesting move, from an interface-design standpoint, was Unutmaz's own check. He asked GPT-5 Pro to simulate an experiment he had already run but not published — CD8+ T cells with an enhanced ability to kill lymphoma cells. The model predicted the boost correctly, and because the results weren't online, it couldn't have retrieved the answer.

This is the pattern applied teams should copy: before trusting a model on the unknown, feed it a known-but-invisible case and see if it reconstructs the result. Unutmaz built his own confidence test into the workflow rather than taking the first striking answer at face value. That habit is what separates a collaborator from an oracle.

Expertise is the part of the interface OpenAI can't ship

OpenAI is candid that this workflow depends on the human. AI may generate an insight, but someone still has to judge its significance and plausibility — and the piece states plainly that a non-expert wouldn't have known whether the IL-2 mechanism mattered. The same document ties the acceleration to biosecurity risk, pointing to OpenAI's Preparedness Framework and the possibility of lowering barriers for bad actors.

The specific implication for anyone designing research tooling on top of these models: the chat surface is doing real work, but it is only as good as the expert sitting on the other side of it. The productive unit here isn't GPT-5 Pro alone — it's an immunologist who can upload three years of buried data, recognize a plausible mechanism when the model surfaces it, and run a private validation before betting a lab week on the answer. Build for that person, not around them.