Cell2Sentence / C2S-Scale (Yale, Google)

Cell2Sentence, and its scaled successor C2S-Scale from Yale and Google, converts a cell into text by writing its gene names in rank order of expression, so a standard language model can be trained on cells and prompted about them in natural language. The 2025 bioRxiv preprint scaled Gemma-based models to 27B parameters on these cell sentences. Its headline result is a prediction that silmitasertib, a CK2 inhibitor, boosts antigen presentation under low interferon, which was validated in vitro, an early example of an AI-generated hypothesis confirmed in the wet lab and relevant to cancer immunotherapy. Text tokenisation is lossy because expression magnitudes are discarded, and only one hypothesis has been validated so far. For a newcomer: it turns a cell into a sentence so a chatbot can reason about biology, and one of its ideas has already checked out in the lab.