C2S-Scale 27B: Google created a neural network for cancer immunotherapy
ШІ
Google, together with Yale University, presented a new basic artificial intelligence model, C2S-Scale 27B, equipped with 27 billion parameters, designed to analyze the “language” of individual cells and find potential ways to treat cancer.
According to the developers , the system put forward a hypothesis about the behavior of cancer cells, which was later confirmed experimentally on living biological samples.
” This discovery opens up new prospects for creating cancer therapies ,” Google said.
The model builds on previous research that showed that biological and linguistic systems follow similar scaling laws: increasing the size of a system increases its efficiency.
How C2S-Scale 27B works
One of the key challenges with immunotherapy is that many tumors remain “cold,” meaning they are not visible to the immune system. To activate them, cells must present signals more intensely through a process known as “antigen presentation.”
The task of C2S-Scale 27B was to find a drug that would act as a conditional enhancer: stimulating an immune response only in a specific “immunopositive” environment with low levels of interferon, insufficient for independent antigen activation. Smaller models could not solve this problem due to the complexity of conditional reasoning.
To do this, we created a virtual dual-context screening that included two stages:
- Immunopositive context: The model received real patient samples where tumors and immune cells interacted, with low levels of interferon.
- Immune-neutral context: data from isolated cell lines without an immune environment.
More than 4,000 drugs were simulated in both contexts, and the model determined which ones enhanced antigen presentation only in the first case, focusing on clinically important scenarios.
Approximately 10–30% of these drugs were already mentioned in the scientific literature, the rest were unexpected discoveries.
Experimental verification
C2S-Scale 27B revealed a significant “context difference” for the CK2 kinase inhibitor silmitasertib (CX-4945). The model predicted a significant increase in antigen presentation in the “immune-positive” context and almost no effect in the “immune-neutral” context.
The researchers then tested the hypothesis on human neuroendocrine cells, which were not used in training the model. The results showed:
- silmitasertib alone did not alter antigen presentation;
- low dose interferon had a moderate effect;
- Combining silmitasertib with low-dose interferon created a significant synergistic enhancement, increasing antigen presentation by approximately 50% and making the tumor more visible to the immune system.
These numerical predictions were validated in the laboratory, demonstrating the potential of C2S-Scale in discovering new conditional interferon enhancers to convert “cold” tumors into “hot” ones, more sensitive to immunotherapy.
” This is only the first step, but an experimentally validated basis for combination therapies, where several drugs work synergistically, has already been created ,” the authors note.
The Yale teams are continuing their research by testing other AI predictions in different immune contexts. If preclinical and clinical trials confirm the results, this approach could significantly accelerate the development of new therapeutic strategies against cancer.