Google, in collaboration with Yale University, has introduced a new foundational model with 27 billion parameters, designed to understand the “language” of individual cells.
The C2S-Scale 27B model proposed a hypothesis about cancer cell behavior, which was later experimentally confirmed using live organic samples.
“This discovery has revealed a promising path for developing new cancer therapies,” the company emphasized.
The model builds on previous research showing that biological and linguistic systems follow similar scaling laws—efficiency increases with size.
How C2S-Scale 27B Works
A major challenge in cancer immunotherapy is that many tumors remain “cold”—invisible to the immune system. One way to “heat” them is to induce signal presentation through a process called “antigen presentation.”
Google tasked C2S-Scale 27B with finding a drug that acts as a conditional amplifier: enhancing immune response only in a specific “immune-positive” environment where a low level of interferon is present but insufficient to independently activate antigen presentation.
This task required conditional reasoning, which smaller models could not handle.
To achieve this, scientists developed a virtual dual-context screening capable of identifying this synergistic effect. It involved two stages:
- Immune-positive context: Experts provided the model with real patient samples where tumor and immune cell interactions were preserved, along with a low interferon signal level;
- Immune-neutral context: The model received data from isolated cell lines devoid of an immune environment.
Google then simulated over 4,000 drugs in both contexts and tasked the model with identifying which enhanced antigen presentation only in the first scenario. This focused the search on clinically significant scenarios.
Among the many options, 10–30% were already mentioned in scientific literature, while the rest were unexpected discoveries.
Experimental Confirmation
The model identified a “striking contextual gap” for the kinase CK2 inhibitor called silmitasertib (CX-4945). The neural network predicted a significant enhancement of antigen presentation when the drug was used in an “immune-positive” context, but almost no effect in an “immune-neutral” one.
Notably, this is a completely new idea that had not been mentioned before.
Researchers then tested the hypothesis in the lab. They used human neuroendocrine cells—samples the model had not “seen” during training. The results showed that:
- Treating cells with silmitasertib alone did not alter antigen presentation;
- Treating with a low dose of interferon alone had a moderate effect;
- The combination of silmitasertib and a low dose of interferon caused a pronounced synergistic enhancement of antigen presentation.
In laboratory experiments, this combination increased antigen presentation by about 50%, making the tumor more visible to the immune system.
The digital prediction was repeatedly confirmed.
C2S-Scale discovered a new conditional interferon amplifier that could help turn “cold” tumors into “hot” ones—more susceptible to immunotherapy.
“While this is only a first step, it already provides an experimentally validated basis for developing new combination therapies, where multiple drugs work together to achieve a stronger effect,” the blog states.
Yale University teams are already exploring the identified mechanism and testing other AI predictions in various immune contexts. With further preclinical and clinical validation, such hypotheses could accelerate the development of new treatments.
Previously, the biotechnology company SpotitEarly began developing a home cancer test based on human breath analysis. The technology combines canine olfaction and artificial intelligence algorithms.
Earlier, in September, scientists developed an AI tool for predicting over 1,000 diseases and forecasting health changes up to 10 years in advance.