A 3-year immunology mystery gets its breakthrough — with GPT-5!

Hi everyone, it's Shii! Today I found a story that gets me really excited: an AI helped a scientist clear up a puzzle that had been nagging his lab for 3 years — and the setting is immunology. Sounds tough, but let's take a peek together!

OpenAI News openai.com

What was announced?

This one comes from OpenAI's News. Immunologist Derya Unutmaz (a researcher at The Jackson Laboratory and the University of Connecticut) used GPT-5 Pro to find a breakthrough on a T cell mystery his lab had been unable to crack since 2022.

Because it's a News-style post, it reads more like "look what we achieved" than a how-to for a product. Still, the content is a great, concrete look at how AI can actually help at the research bench.

Why it matters

T cells are key players in our immune system, fighting cancer, autoimmune disease, and infections. Unutmaz's lab was studying how these T cells grow and branch into different types, and how that connects to glucose.

But the experimental data held a strange result they couldn't explain. Without an answer, the lab shelved the question and moved on to other urgent work — a piece of homework left in a drawer for 3 years.

"We have the data but can't interpret it" is a common moment in research, and the key idea here is that AI might help exactly there.

What changes

Unutmaz handed that dormant data and question to GPT-5 Pro. Drawing on a huge body of prior research, the model came back with a new mechanistic hypothesis to explain the observed T cell behavior.

And in hindsight it made sense — a connection that Unutmaz himself, and no one in his lab, had spotted. It sat just outside his own area of expertise, so humans alone had missed it.

Analysis that would normally take months moved forward in far less time. That is the big point: it hints that the very speed of research could change.

Dive Deep

The important thing is that GPT-5 didn't just hand over "the answer" on its own. The flow looked like this:

• The human researcher supplies the data, the domain expertise, and the judgment of whether a hypothesis is biologically plausible
• The AI proposes hypotheses (the interpretive leap) across a huge span of literature
• The researcher checks and validates whatever the model produces

This "the AI leaps, the human verifies" pairing is what worked. In fact, GPT-5 Pro was reported to have correctly predicted the outcome of a separate, unpublished experiment (a lymphoma-targeting T cell experiment), which suggests it isn't merely reciting memorized data.

At the same time, it's worth being careful: the specific biology (which molecules do what) is not fully laid out in a peer-reviewed paper at the time of this announcement. So this is less "AI has fully proven it!" and more "quickly propose a promising hypothesis, then let humans validate it" — that's the fair way to read it.

Wrap-up

• Immunologist Derya Unutmaz used GPT-5 Pro to find a breakthrough on a T cell puzzle (about glucose and differentiation) his lab had left unsolved since 2022
• GPT-5 Pro proposed a fresh mechanistic hypothesis consistent with prior research, surfacing a connection even the experts had missed
• It reportedly predicted the result of an unpublished experiment too, so it's not just memorization
• Still, the deciding factors were human expertise and validation — AI proposes, humans verify, working as a pair
• It could help push cancer and autoimmune research forward, which is the exciting hope here

If you've ever hit a wall interpreting your own data, or you're curious whether AI can be a partner for generating hypotheses (not just for looking things up), I think this one is especially for you!