Google DeepMind and Yale University announced on October 15 the development of C2S-Scale 27B, a 27-billion-parameter artificial intelligence foundation model that has successfully generated and experimentally validated a novel hypothesis for cancer treatment. Built on Google's Gemma family of models, this represents one of the largest AI systems ever developed to study cellular behavior and marks what the company calls "a milestone for AI in science."
The breakthrough centers on addressing one of cancer immunotherapy's most challenging problems: converting "cold" tumors that evade immune detection into "hot" tumors that can be recognized and attacked by the immune system. These cold tumors typically resist existing immunotherapy treatments, representing a significant unmet medical need in oncology.
AI-Driven Drug Discovery Through Virtual Screening
The C2S-Scale 27B model was designed to identify drugs that could boost immune signaling through antigen presentation, but only under specific biological conditions. Using what researchers termed a dual-context virtual screen, the AI analyzed more than 4,000 drugs across patient tumor samples and isolated cell data to identify compounds that would selectively enhance immune activation in relevant biological settings.
The model's analysis revealed surprising results. While some predictions involved known drugs, approximately 10 to 30 percent were entirely new candidates with no previous connection to cancer immunotherapy or immune modulation. Among the most significant findings was the identification of silmitasertib (CX-4945), a kinase CK2 inhibitor, as a potential conditional amplifier drug.
The AI predicted that silmitasertib would dramatically increase antigen presentation only when used in an "immune-context-positive" environment where low levels of interferon were already present. Crucially, neither the drug alone nor interferon alone would produce significant effects, but their combination could potentially trigger a substantial immune response against tumors.
Experimental Validation Confirms AI Predictions
Yale scientists tested the AI's hypothesis using human neuroendocrine cell models that were not part of the model's training data. The experimental validation confirmed the AI's prediction with striking accuracy. Treatment with silmitasertib alone produced no change in antigen presentation, while low-dose interferon alone had only modest effects.
However, combining both treatments resulted in approximately a 50 percent increase in antigen presentation, effectively making tumor cells more visible to the immune system. This finding demonstrates that the AI model didn't merely process biological data but reasoned through complex cellular contexts to uncover how specific conditions determine treatment success.
"What made this prediction so exciting was that it was a novel idea," Google researchers noted. "Although CK2 has been implicated in many cellular functions, including as a modulator of the immune system, inhibiting CK2 via silmitasertib has not been reported in the literature to explicitly enhance MHC-I expression or antigen presentation."
Implications for Cancer Treatment Development
The discovery suggests a potential new pathway for developing combination therapies targeting immunotherapy-resistant tumors. Google CEO Sundar Pichai emphasized the clinical potential, stating, "With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer."
Research teams at Yale are now investigating the exact mechanism underlying this immune system effect and testing additional predictions generated by the AI system. The collaboration demonstrates how large-scale AI can function as a virtual laboratory, conducting thousands of simulated experiments to reveal previously unknown relationships between drugs, cells, and immune signals.
Advancing AI-Driven Biological Discovery
DeepMind researchers described the findings as evidence that scaling up biological AI models produces more than improved accuracy—it can generate entirely new hypotheses. "The true promise of scaling lies in the creation of new ideas, and the discovery of the unknown," the research team stated.
The C2S-Scale 27B model and accompanying tools have been made publicly available on Hugging Face and GitHub, with a scientific preprint posted on bioRxiv. This approach provides what researchers call "a blueprint for a new kind of biological discovery," utilizing large-scale AI systems to run virtual drug screens and propose biologically grounded hypotheses for laboratory testing.
However, experts note that these findings represent only the initial step in a lengthy development process. The results have yet to undergo peer review or clinical validation, and any therapeutic application would require years of additional research and clinical trials before reaching patients.
