Google DeepMind, in collaboration with Yale University, has unveiled a 27-billion-parameter AI model (dubbed C2S‑Scale 27B / “Cell2Sentence-Scale”) that mined thousands of single-cell and tumor-immune datasets for fresh therapeutic insights.

The system screened over 4,000 drug candidates across immune-context positive and neutral cell environments, identifying Silmitasertib (CX-4945) - a known CK2 inhibitor - as a previously unreported amplifier of antigen presentation when combined with low-dose interferon in “cold” tumours. 

Laboratory validation in human neuroendocrine tumour cells showed a ~50 % boost in MHC-I antigen expression under the predicted conditions, making the tumour cells more visible to the immune system. 

While still in the pre-clinical phase, researchers are calling this generative-AI driven hypothesis-generation a milestone in biomedical science.
The model not only reproduced known drug-biology links (10-30 % of its hits) but suggested novel ones, suggesting AI may now generate hypotheses rather than only recognise patterns. 

If further pre-clinical and clinical work validates the finding, this approach could accelerate the repurposing of existing compounds and unlock entirely new therapeutic pathways - reshaping how drug discovery is done.
However, scientists caution the road from AI model to approved treatment is long and complex.