Researchers from Google DeepMind, in collaboration with Yale University, have created the C2S-Scale 27B model, which has 27 billion parameters and is based on the Gemma family of models. This system converts complex gene expression data from individual cells into so-called cell sentences, making it possible to analyze cell behavior similarly to natural language. The model was trained on more than a billion single-cell profiles and can simulate responses to thousands of drugs in various biological environments.
The research focused on the problem of cold tumors, which hide from the immune system. The model was tasked with finding a drug that would amplify the immune signal only in an environment with low levels of interferon, a key protein in immune signaling. Using a dual virtual screening process, the model examined more than 4,000 drugs in two contexts: one with intact tumor-immune interactions from patient samples and another with isolated cell lines lacking an immune context.
Key Discovery: Silmitasertib and Its Effects
The model identified silmitasertib, also known as CX-4945, an inhibitor of the CK2 kinase. According to the prediction, this drug would significantly increase antigen presentation only in an environment with low levels of interferon, where it would act as a conditional amplifier. In a neutral environment without an immune context, it would have no effect. This discovery was novel because silmitasertib had not previously been associated with enhanced MHC-I expression or antigen presentation.
Validation was carried out in Yale University laboratories using human neuroendocrine cell models that the model had not encountered during training. The experiments confirmed the prediction: silmitasertib alone did not increase antigen presentation. A low dose of interferon had only a modest effect. However, combining the two led to a synergistic increase of 50%, making the tumor more visible to the immune system.
Experimental Validation and Additional Details
The researchers conducted tests in living cells, where the combination of silmitasertib and a low dose of interferon caused a significant increase in antigen presentation. This effect was specific to an environment with an immune context, exactly as the model had predicted. In the overall screening, the model identified 10 to 30% of the drugs as known candidates, while the remainder were unexpected candidates with no previous association with immunomodulation.
Sundar Pichai, CEO of Google, described this discovery as an important milestone in the application of AI in science. The model did not merely process data, but generated a new hypothesis that scientists then validated. The research was conducted under the leadership of Shekoofeh Azizi and Bryan Perozzi from Google, in collaboration with a laboratory at Yale.
The C2S-Scale 27B model is now freely available on platforms such as Hugging Face and GitHub, including the code and a scientific preprint on bioRxiv. This approach allows other scientists to test the predictions and advance the research. Teams at Yale are now investigating the mechanism behind this discovery and testing additional model predictions in other immune contexts.
