A Gemma model helped discover a new potential cancer therapy pathway (blog.google)

AI research team released Cell2Sentence-Scale 27B (C2S-Scale), a 27-billion-parameter foundation model in the Gemma family built to read single-cell biology. Using a dual-context virtual screen (patient-like “immune-context-positive” samples with low interferon signaling vs immune-context-neutral cell lines), C2S-Scale simulated the effects of >4,000 drugs and predicted context-specific hits. Crucially, the model flagged the CK2 inhibitor silmitasertib (CX-4945) as a conditional amplifier: little effect alone or in neutral context, but strong synergy with low-dose interferon in the immune-context-positive condition. Lab tests in human neuroendocrine cells (a cell type unseen during training) confirmed the prediction—silmitasertib plus low-dose interferon produced ~50% increased MHC-I antigen presentation versus controls—turning a “cold” tumor signal substantially “hotter.” This result matters because it demonstrates an emergent capability from scale: larger biology models can perform context-dependent reasoning and generate novel, testable hypotheses rather than merely recalling known links. Methodologically, the dual-context in silico screen biases discovery toward patient-relevant biology and recovered both known hits (~10–30%) and surprising candidates for follow-up. The validated prediction suggests a new combination strategy to sensitize tumors to immunotherapy and provides a blueprint for AI-driven high-throughput hypothesis generation. C2S-Scale and resources are now publicly available for researchers to extend and validate further findings.