Google's AI Cracks a New Cancer Code (decrypt.co)

Google DeepMind, working with Yale, announced that a 27‑billion‑parameter single‑cell foundation model built on the open‑source Gemma family—Cell2Sentence‑Scale 27B (C2S‑Scale)—generated a novel, experimentally validated cancer‑therapy hypothesis. After simulating effects of more than 4,000 drug candidates under two immune states, the model predicted that silmitasertib (CX‑4945), a CK2 kinase inhibitor, would act as a conditional amplifier: when combined with low‑dose interferon (an immune‑active context) it raised MHC‑I antigen presentation in human neuroendocrine cells by roughly 50%, effectively making otherwise “cold” tumor cells more visible to the immune system. Lab experiments confirmed the prediction, and Yale teams are now probing the underlying mechanism. The result is significant because it suggests large biological AI models can invent new, testable scientific hypotheses—not merely fit known patterns—and run high‑throughput virtual drug screens to prioritize experiments. Key technical points: C2S‑Scale is a 27B parameter model for single‑cell analysis, evaluated across immune‑active vs immune‑inactive simulated conditions, and produced a previously unreported link between CK2 inhibition and enhanced antigen presentation. Caveats remain: the work is preclinical, not yet peer‑reviewed, and clinical translation will require substantial follow‑up. Still, the study offers a blueprint for AI‑driven discovery pipelines that generate mechanistic hypotheses for lab validation.