Google DeepMind’s Gemma-Based Model Uncovers a New Cancer Therapy Pathway
Google DeepMind and Yale University, researchers have unveiled Cell2Sentence-Scale 27B (C2S-Scale), a 27 billion parameter foundation model designed to interpret the “language” of individual cells.
Created on October 16|Last edited on October 16
Comment
In a collaboration between Google DeepMind and Yale University, researchers have unveiled Cell2Sentence-Scale 27B (C2S-Scale), a 27 billion parameter foundation model designed to interpret the “language” of individual cells. Built upon Google’s Gemma family of open models, C2S-Scale represents one of the largest and most advanced models for single-cell analysis to date. The model’s most notable achievement so far is the identification of a new potential pathway for cancer treatment, a discovery confirmed through laboratory experiments.
Building the C2S-Scale 27B Model
C2S-Scale 27B builds on earlier research showing that biological AI models follow the same scaling laws as language models: the larger the model, the better it performs. But the key question was whether scale enables not just better predictions, but new insights entirely. By training on massive single-cell datasets, the model was designed to understand how different cell types communicate through molecular signals—effectively translating cellular activity into a kind of “sentence structure” that AI could interpret and reason about.
Discovering a New Immune Pathway
Cancer immunotherapy often struggles with “cold” tumors, those that evade detection by the immune system. C2S-Scale was tasked with finding a drug that could selectively boost immune activity in cells that already show weak immune signaling, but not in those without any immune context. To test this, researchers designed a virtual drug screen that simulated thousands of compounds in two environments: one with low-level immune signaling and one without. The goal was to identify drugs that would amplify immune response only when the right context was present.
From AI Prediction to Lab Validation
Among thousands of compounds, C2S-Scale highlighted silmitasertib (CX-4945), a kinase CK2 inhibitor, as a top candidate. The model predicted that silmitasertib would significantly enhance antigen presentation—the process that makes tumor cells visible to immune cells—when combined with low doses of interferon, a natural immune-signaling molecule. Laboratory testing confirmed this prediction. Silmitasertib alone showed no significant effect, and interferon alone produced a modest one, but the combination led to a striking 50 percent increase in antigen presentation. This confirmed that the AI model’s prediction reflected a real, biological interaction rather than a statistical coincidence.
Implications for Cancer Research and Drug Discovery
This finding could mark a major shift in how researchers approach cancer immunotherapy. By revealing that CK2 inhibition can act as a context-dependent amplifier of immune signaling, the model opens the door to new combination treatments that “heat up” previously resistant tumors. Beyond cancer, the success of C2S-Scale demonstrates that large biological foundation models can generate novel, testable hypotheses about cellular behavior—effectively turning AI into a discovery engine for biology. Researchers at Yale are now extending this work, testing additional model-generated hypotheses across different immune environments.
Access and Collaboration
Google DeepMind has made the C2S-Scale 27B model and its associated resources available to the research community. The model, code, and datasets can be accessed through Hugging Face, GitHub, and the preprint publication on bioRxiv. The open release aims to accelerate further discoveries by allowing other scientists to refine, validate, and expand upon the pathways identified by the model.
Add a comment