GPT-5 Pro cracked a 3-year-old immunology mystery in one afternoon

I’ve been watching the “AI for science” hype cycle for a while now, and most of it is just that — hype. But every so often a story comes along that makes me sit up and pay attention. This is one of them.

Derya Unutmaz is an immunologist at the Jackson Laboratory. He’s been studying T cell behavior for years, specifically how these cells decide whether to attack or stand down. Three years ago, he hit a wall. A particular pattern of T cell activation didn’t fit any known model. He had data, he had hypotheses, but nothing clicked. The mystery sat there, unresolved, for three years.

Then he tried GPT-5 Pro.

Not as a search engine or a glorified autocomplete. He fed it the raw experimental data, the contradictory results, the dead ends. He described the problem in plain language, the way you’d explain it to a colleague over coffee. The model didn’t just spit out a textbook answer. It connected the dots between signaling pathways that Unutmaz hadn’t considered together. It suggested a mechanism involving a specific cytokine feedback loop that, when tested, turned out to be exactly what was happening.

Three years of frustration, resolved in a single afternoon.

What strikes me about this isn’t just the speed. It’s the nature of the insight. This wasn’t a lookup. GPT-5 Pro didn’t have a cached answer for “weird T cell activation pattern in autoimmune context.” It synthesized information from immunology, systems biology, and even some oncology literature that Unutmaz wasn’t deeply familiar with. The model effectively played the role of a brilliant, hyper-literate collaborator who has read every paper in the field and can spot patterns across disciplines.

Now, I’m not naive. I know GPT-5 Pro has limitations. It can hallucinate, it can be confidently wrong, and it doesn’t truly “understand” immunology. But in this case, the output was testable, and it passed. That’s the bar that matters in science.

The implications for cancer and autoimmune research are obvious but worth stating. If an AI can help untangle T cell behavior — one of the most complex and context-dependent systems in the human body — then it can probably help with a lot of other stubborn problems. Drug target identification, biomarker discovery, patient stratification. The usual bottlenecks in translational research.

Unutmaz himself said something that stuck with me: “I no longer think of GPT-5 as a tool. It’s more like a junior colleague who has read everything and thinks differently.” That’s a shift in framing that matters. We’ve spent years treating AI as a calculator or a search engine. Maybe we should start treating it as a thinking partner — flawed, yes, but capable of seeing things we miss.

I’ll be watching to see if this holds up in other labs. One success story doesn’t prove a paradigm shift. But if GPT-5 Pro can do this for immunology, I’m genuinely curious what it might do for fields I know less about. Materials science. Neuroscience. Climate modeling.

For now, I’m just glad a good scientist got his answer. And I’m a little jealous I didn’t think of trying this myself.