GPT-5 Pro Cracks a 3-Year T-Cell Puzzle, Lab Confirms

OpenAI published an account of GPT-5 Pro helping immunologist Derya Unutmaz untangle a result that had sat unexplained on his bench since 2022. Unutmaz, a researcher at The Jackson Laboratory for Genomic Medicine, had cultured human T cells with 2-deoxy-D-glucose — a compound that interferes with how cells use sugar — and watched them behave in ways his lab could not pin to a clean mechanism. Years of expertise had not closed the gap. The model did, in the space of a conversation.

The key move was a distinction Unutmaz's team had not fully separated. Blocking glucose metabolism can act in more than one way, and GPT-5 Pro argued that the data pointed not to blunted glycolysis — the obvious culprit — but to disrupted N-linked glycosylation, a sugar-tagging process cells use to build and display proteins. From there the model went further: it fingered the IL-2 receptor pathway as the driver and predicted that the effect would be carried by memory T cells rather than naïve ones. Each of those was a specific, falsifiable claim, not a hand-wave.

The most striking part was a prediction the model could not have copied from the literature, because the experiment had never been published. Asked to simulate engineering CD8+ T cells with an anti-CD19 CAR — the same architecture behind approved blood-cancer therapies — to attack CD19-positive lymphoma, GPT-5 Pro forecast that a brief dose of 2-DG during priming would lower the exhaustion markers PD-1 and LAG-3, preserve the cells' killing power, and improve their ability to destroy tumor cells over repeated encounters. When Unutmaz's lab ran it, the prediction held.

OpenAI framed the episode carefully, and so should anyone reading it. This was an expert-led workflow, not an autonomous discovery: a domain scientist with decades of immunology behind him chose the questions, interrogated the model's reasoning, and confirmed every claim at the bench. The model functioned as an accelerant for hypothesis generation — compressing what might have been months of literature-cross-referencing and dead-end experiments into a sharply prioritized shortlist. The account appears alongside other case studies in OpenAI's broader "early science acceleration" write-up, part of a push to show frontier models contributing to real research rather than benchmark scores.

It lands amid a wave of AI-for-science claims that range from rigorous to overhyped, and this one earns attention precisely because the bar was concrete: a novel, mechanistic, experimentally verified prediction in cancer immunology. It echoes a pattern BitsMinds has tracked elsewhere — from OpenAI's dedicated science model GPT-Rosalind to Mayo Clinic's cancer-detection work — in which the most credible results come not from AI working alone, but from a specialist using it as an instrument. The honest caveat, which OpenAI's own life-science benchmarks underline, is that today's models still fail most hard biology tasks; the wins are real, but they are not yet routine.