Google, in collaboration with Yale University, has introduced a new foundation AI model called C2S-Scale 27B, designed to understand the “language” of individual cells and to unlock insights into cancer behavior.
This model, built upon Google’s Gemma family of open models, was used to predict new drug candidates that could combat tumors by simulating cellular behavior at microscopic scales.
After computational predictions, Google tested the results in human cell models. In an announcement, Google explained that C2S-Scale 27B’s findings were lab-tested on human model cells to verify whether its predictions held true under real biological conditions.
As Google’s CEO Sundar Pichai put it,
“An exciting milestone for AI in science: Our C2S-Scale 27B foundation model, built with Yale (University) and based on Gemma, generated a novel hypothesis about cancer cellular behavior, which scientists experimentally validated in living cells.”
A major significance of the model is its accessibility: C2S-Scale 27B, with its 27 billion parameters, is made available on platforms such as GitHub and Hugging Face. This openness allows researchers worldwide to apply it, extend it, or explore its predictions for new cancer interventions.
As Pichai added,
“With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer.”
Large AI models in language and vision have shown that scale often unlocks emergent behaviors. Google’s team applied a similar principle to biology: by scaling up a biological foundational model, they hoped C2S-Scale 27B would not just learn existing patterns but uncover new ones.
In particular, the model was tasked with identifying a drug that acts as a conditional amplifier—a compound that would strengthen immune signaling only in the specific “immune-context-positive” environment (where low levels of interferon are present), without triggering effects in neutral environments.
The model ran virtual screenings across over 4,000 drugs, comparing their predicted effects across two contexts:
Immune-context-positive (with existing interferon signaling)
Immune-context-neutral (isolated cell lines without immune context)
Among the predicted drug hits, some aligned with known literature, but others were novel and surprising—candidates not previously linked to the screen context.
C2S-Scale 27B specifically singled out the kinase CK2 inhibitor silmitasertib (CX-4945). It predicted a strong increase in antigen presentation when silmitasertib is used in the immune-context-positive setting, but minimal effect in a neutral context.
What is noteworthy is that silmitasertib had not previously been reported to enhance MHC-I expression or antigen presentation, meaning the model proposed a genuinely novel mechanistic hypothesis.
To test its predictions, the team conducted laboratory experiments in human neuroendocrine cell models, which were not part of the model’s training data.
The results confirmed the AI’s hypothesis:
Silmitasertib alone had no effect on antigen presentation.
Low-dose interferon alone produced a modest effect.
Combination of silmitasertib + low-dose interferon led to a synergistic amplification, boosting antigen presentation by roughly 50%.
This finding effectively suggests a way to convert otherwise “cold” tumors—those invisible to the immune system—into “hot” tumors that can be targeted more effectively by immunotherapies.
The research emphasizes that scaling biological AI models can yield not just better predictions, but entirely new scientific hypotheses. It outlines a blueprint for merging AI-driven virtual screening with experimental validation to accelerate biomedical discovery.
The Yale teams are continuing to explore the mechanistic basis of the discovered effect and applying C2S-Scale 27B to other immune contexts.
However, it is early days: more preclinical and clinical testing will be essential before any real therapeutic application emerges.
Google and Yale’s collaboration on C2S-Scale 27B marks a significant stride in merging AI and cancer research. By treating cells like languages to be decoded, the model has not only predicted but experimentally validated a new immunotherapy pathway. Its openness via GitHub and Hugging Face allows the scientific community to build on the foundation. Though much work remains before patient benefit, this development is a milestone in AI’s role in unveiling novel biology.