How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery

Immunologist Derya Unutmaz and his team spent three years trying to understand T cell behavior under specific conditions, but traditional experimental and data analysis methods failed to provide clear answers. T cells are central to the immune system, and their abnormal behavior is closely linked to cancer and autoimmune diseases.

The core of the puzzle was the response mechanism of T cells in a certain microenvironment—why some T cells enter an 'exhausted' state while others remain active. This question had long puzzled the immunology community because existing models could not explain the observed heterogeneity.

Unutmaz fed GPT-5 Pro a large amount of single-cell RNA sequencing data, protein interaction networks, and literature summaries accumulated by his team, asking it to find potential patterns and mechanisms. Within hours, GPT-5 Pro proposed a previously unconsidered hypothesis of cross-regulation between signaling pathways, involving the synergistic action of transcription factors FOXO1 and mTOR.

The hypothesis predicted that under low-nutrient conditions, nuclear localization of FOXO1 inhibits mTOR activity, preventing T cells from entering an exhausted state; the opposite occurs in high-nutrient environments. This mechanism explained why T cell fates differ across microenvironments.

Unutmaz's team then validated the hypothesis through CRISPR gene editing and in vitro experiments, confirming a novel regulatory relationship between FOXO1 and mTOR. The experimental results were highly consistent with GPT-5's predictions, thus solving the three-year puzzle.

This case demonstrates that large language models can not only retrieve and summarize knowledge but also perform cross-domain reasoning to generate testable hypotheses. GPT-5 Pro integrated disparate information from immunology, cell biology, and biochemistry, uncovering connections that human researchers might have missed.

However, the breakthrough relied on high-quality, structured input data (single-cell sequencing and protein interaction data) and the researchers' critical evaluation of the output. GPT-5 itself lacks experimental validation capabilities, and its reasoning process remains a 'black box'—we do not know why it selected these genes and pathways.

Moreover, the generalizability of the result remains to be tested. Can other laboratories replicate the finding in independent datasets? Could GPT-5's hypothesis have blind spots due to biases in its training data? These questions remain unanswered.

If the mechanism holds under broader physiological and pathological conditions, it could provide new targets for cancer immunotherapy: by modulating the balance between FOXO1 and mTOR, it may prevent T cell exhaustion in the tumor microenvironment, thereby enhancing the durability of immunotherapies.

In autoimmune diseases, the finding might explain why certain T cells become hyperactive at inflammatory sites, offering clues for designing more precise immunosuppressants. Unutmaz's team has already begun discussions with pharmaceutical companies about potential collaborations.