Google has unveiled a groundbreaking advancement in AI medical research with the help of a new model called Cell2Sentence Scale 27B (C2S-Scale), which is part of its open Gemma family of models. The model helped researchers to identify a new, experimentally validated pathway that could make resistant cancer tumors more treatable with immunotherapy.
The model is developed by Google along with Yale University researchers, and features a 27 billion parameter foundation model that deciphers the “language of individual cells,” offering a transformative way to analyze cellular behavior and uncover potential therapy avenues for diseases like cancer. The core innovation of C2S-Scale lies in its ability to reason contextually, which has been a challenge for smaller biological models.
Google and Yale tasked the model with identifying a “conditional amplifier”, a drug that strengthens immune signaling only in the right conditions. Google explained that tumors often evade immune detection by becoming “cold,” meaning they don’t trigger immune responses. C2S-Scale sought compounds that could “heat” them up, but only where certain immune signals, like interferon, were already weakly present.
The model was able to analyze over 4,000 drugs under two conditions using a dual context virtual screening technique, one with active immune signaling and one neutral, to determine which had the desired effect exclusively in an immune-active context. The majority of the drugs it flagged were already known to researchers, but several of them were novel.
Among the top results was silmitasertib (CX-4945), which is a CK2 kinase inhibitor. Google explained:
The model predicted a strong increase in antigen presentation when silmitasertib was applied in the “immune-context-positive” setting, but little to no effect in the “immune-context-neutral” one. What made this prediction so exciting was that it was a novel idea.
Although CK2 has been implicated in many cellular functions, including as a modulator of the immune system, inhibiting CK2 via silmitasertib has not been reported in the literature to explicitly enhance MHC-I expression or antigen presentation. This highlights that the model was generating a new, testable hypothesis, and not just repeating known facts.
Google then took this prediction to a lab and actually tested the results. Lab experiments confirmed that silmitasertib alone had no effect, and interferon alone had a modest one, but both together resulted in a 50% increase in antigen presentation. This suggests that silmitasertib could help immunotherapy drugs better recognize and target tumors, especially where immune activation is insufficient.
The research successfully demonstrated that large-scale biological foundation models can go beyond data analysis and can actually provide discoveries.
Yale University researchers are now working on how the discovery would operate in different immune contexts and researching additional drug predictions generated by the model.
The Gemma model is open source and is already publicly available on Hugging Face and GitHub. Google has invited other scientists to build on the model"s capabilities.