In brief
The Gemma-based model generated and validated a new cancer-therapy hypothesis in human cell experiments.
It identified silmitasertib as a conditional amplifier that boosts immune visibility of tumors.
The result hints that scaled biological AI models can invent—not just optimize—new scientific ideas.

Google DeepMind said Wednesday that its latest biological artificial-intelligence system has generated and experimentally confirmed a new hypothesis for cancer treatment, a result the company calls “a milestone for AI in science.”

“With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer,” Google CEO Sundar Pichai tweeted.

An exciting milestone for AI in science: Our C2S-Scale 27B foundation model, built with @Yale and based on Gemma, generated a novel hypothesis about cancer cellular behavior, which scientists experimentally validated in living cells. 

With more preclinical and clinical tests,…

— Sundar Pichai (@sundarpichai) October 15, 2025

In collaboration with Yale University, DeepMind researchers released a 27-billion-parameter foundation model for single-cell analysis called Cell2Sentence-Scale 27B (C2S-Scale), built on Google’s open-source Gemma family of models. The model was able to generate “a novel hypothesis about cancer cellular behavior and we have since confirmed its prediction with experimental validation in living cells. This discovery reveals a promising new pathway for developing therapies to fight cancer,” the company wrote in a blog post today.

The finding centers on one of the hardest problems in cancer immunotherapy: how to make so-called cold tumors, which are invisible to the immune system, more hot and thus more responsive to treatment. DeepMind said its model successfully identified a conditional amplifier drug that could boost immune visibility only in certain biological contexts.



To test the idea, C2S-Scale analyzed patient tumor data and simulated the effects of more than 4,000 drug candidates under two conditions: one where immune signaling was active and one where it was not. The model predicted that silmitasertib (CX-4945), a kinase CK2 inhibitor, would dramatically increase antigen presentation—a key immune trigger—but only in the immune-active setting.

“What made this prediction so exciting was that it was a novel idea,” Google wrote. “Although CK2 has been implicated in many cellular functions, including as a modulator of the immune system, inhibiting CK2 via silmitasertib has not been reported in the literature to explicitly enhance MHC-I expression or antigen presentation. This highlights that the model was generating a new, testable hypothesis, and not just repeating known facts.”

Laboratory experiments confirmed the prediction. When human neuroendocrine cells were treated with both silmitasertib and low-dose interferon, antigen presentation rose by roughly 50 percent, effectively making the tumor cells more visible to the immune system.

DeepMind researchers described the discovery as evidence that scaling up biological AI models doesn’t just improve accuracy—it can produce entirely new hypotheses. “The true promise of scaling lies in the creation of new ideas, and the discovery of the unknown,” the post said.

Teams at Yale are now probing the mechanism that underlies this immune-system effect and testing other AI-generated predictions. DeepMind said the work “provides a blueprint for a new kind of biological discovery,” one that uses large-scale AI systems to run virtual drug screens and propose biologically grounded hypotheses for lab testing.

The model and accompanying tools are publicly available on Hugging Face and GitHub, with a scientific preprint posted on bioRxiv.

Still, experts caution that such findings represent only the first step in a long process. The results have yet to undergo peer review or clinical validation, and any therapeutic application would require years of additional research and trials.

Generally Intelligent Newsletter

A weekly AI journey narrated by Gen, a generative AI model.