Google DeepMind Identifies New Potential for Cancer Immunotherapy Using Gemma-Based Model

C2S-Scale generated new hypotheses regarding the behavior of cancer cells. Specifically, it identified drugs that act as conditional amplifiers to make tumors "hot" relative to the immune system. The research team confirmed these predictions through validation using live cells in a laboratory setting.

The drug in question is Silmitaceltob, a CK2 kinase inhibitor. The model predicted a "contextual split," where antigen presentation is strongly promoted only in specific immune environments while having a smaller effect in others. The enhancement of MHC-I expression via CK2 inhibition had not been reported in previous literature.

The research team raised the question of whether larger models can acquire new capabilities rather than simply improving accuracy on existing tasks. They stated that this discovery serves as a milestone demonstrating that the true value of scaling lies in unknown discoveries.