Google DeepMind has introduced an innovative artificial intelligence biomodel, Cell2Sentence-Scale (C2S-Scale), with 27 billion parameters, capable of both formulating and testing scientific hypotheses in the field of biomedicine. This solution is marked as a new milestone in the application of AI for cancer treatment research.
This is reported by Business • Media
Capabilities of the Cell2Sentence-Scale Model
In collaboration with Yale University, the DeepMind team utilized a model built on the Gemma architecture to analyze data at the single-cell level. C2S-Scale not only formulates biological hypotheses but is also capable of testing them in laboratory conditions. During the research, the model discovered that the drug silmitasertib (CX-4945) can significantly enhance the immune response, making tumor cells more vulnerable to the immune system. This finding is considered a potential breakthrough in the development of new strategies for treating cancer.
Research Results and Prospects
To test the hypothesis, C2S-Scale modeled the effects of over 4000 different substances under conditions of active immune signaling. The model predicted that silmitasertib significantly increases antigen presentation—a crucial mechanism for activating the immune response—but only when the immune system is active. Experiments on human cells confirmed these predictions: in combination with low doses of interferon, the level of antigen presentation increased by 50%.
“AI has for the first time proposed a new, previously unreported combination with clinical potential.”
This success demonstrates the scalability of biomodels not only for more accurate predictions but also for generating fundamentally new ideas in basic biology. Researchers at Yale University have already begun studying the mechanism behind this effect and are testing other predictions of the system.
As Google CEO Sundar Pichai notes, this discovery will serve as a foundation for new clinical trials and underscores the significant role of artificial intelligence in biomedical science. The code, model, and working tools are already available on Hugging Face and GitHub platforms, and the research preprint has been published on bioRxiv.
At the same time, experts emphasize that the results obtained have not yet undergone proper peer review and require extensive validation before potential application in clinical practice.