Google DeepMind and Yale University developed C2S-Scale, a 27 billion parameter foundation model built on the Gemma family of open models, to advance single-cell analysis. This AI system identified a potential cancer therapy pairing: the kinase inhibitor silmitasertib (CX-4945) and low-dose interferon, which together amplify antigen presentation in tumor cells. The model’s prediction was validated through lab experiments showing a 50% increase in MHC-I expression when both drugs were applied, making tumors more visible to immune cells.

The breakthrough stems from C2S-Scale’s ability to simulate drug effects in two distinct biological contexts. Researchers used a dual-context virtual screen, testing over 4,000 drugs across immune-active and neutral environments. The model highlighted silmitasertib as a candidate that only boosted antigen presentation when combined with interferon in patient-derived tumor tissue, suggesting a context-dependent mechanism. This approach leveraged scaling laws from natural language processing to uncover novel biological insights.

Experiments confirmed C2S-Scale’s hypothesis: silmitasertib alone had no effect, while interferon alone showed modest results. The combination triggered MHC-I upregulation in neuroendocrine cells, a cell type the model had never encountered during training. This synergy demonstrates how AI can predict previously unexplored drug mechanisms, bypassing traditional trial-and-error methods.

C2S-Scale’s code, preprints, and datasets are now publicly available on Hugging Face and GitHub, enabling researchers to replicate the study or explore new hypotheses. The model’s success underscores the potential of large-scale AI in biomedical discovery, particularly for designing combination therapies targeting cold tumors resistant to immunotherapy. As Yale teams investigate the underlying biology, this work provides a roadmap for AI-driven drug repurposing in oncology.