In a world where we ask artificial intelligence to summarize articles and create images, scientists have now tasked an AI with a much bigger challenge: finding a new way to fight cancer. And it just delivered a promising lead.
Google, on October 15, announced a major scientific milestone: an AI model from its Gemma family has helped uncover a potential new way to treat cancer.
In collaboration with Yale University, researchers at Google DeepMind and Google Research developed Cell2Sentence-Scale 27B (C2S-Scale), a massive 27 billion parameter AI model designed to understand the “language” of individual cells.
According to a Google blog post by Shekoofeh Azizi and Bryan Perozzi, “C2S-Scale generated a novel hypothesis about cancer cellular behavior and we have since confirmed its prediction with experimental validation in living cells.” The finding, they said, “reveals a promising new pathway for developing therapies to fight cancer.”
Understanding the model
C2S-Scale is part of Google’s Gemma family of open-source models, and it marks a new chapter in biological research powered by AI. The model was built to analyze data at the single-cell level, which allows scientists to study how individual cells behave in complex biological environments.
One of the biggest challenges in cancer treatment is that many tumors are “cold” — meaning they remain invisible to the body’s immune system. Google’s researchers tasked the AI model with finding a drug that could make these tumors “hot,” or visible to immune cells, by boosting antigen presentation, the process through which cells display signals that alert the immune system.
In simpler terms, the goal was to help the immune system “see” hidden tumors, but only under the right conditions. To do this, the AI ran a dual-context virtual screen across more than 4,000 drug compounds. It compared how each drug behaved in two situations:
- Immune-Context-Positive, where immune signaling was present but weak
- Immune-Context-Neutral, where no immune activity was detected
The model then highlighted drugs that only worked in the first condition, where immune activity was low but present.
The surprising drug candidate
The model identified an existing compound called silmitasertib (CX-4945), known as a kinase CK2 inhibitor, as a promising candidate.
The AI made a striking prediction: silmitasertib could significantly enhance antigen presentation, but only when used in an “immune-context-positive” environment, where low levels of interferon (a key immune protein) were already present.
This prediction was especially notable because silmitasertib had not previously been linked to boosting MHC-I expression or antigen presentation in any published research.
To verify this, researchers tested the model’s prediction in human neuroendocrine cell models, which the AI had never encountered during training. The results aligned with what the AI had predicted:
- Silmitasertib alone had no measurable effect
- Low doses of interferon alone had only a mild impact
- But when combined, the two produced a remarkable 50% increase in antigen presentation, effectively making the cancer cells more visible to the immune system
In other words, the combination helped “light up” the cancer cells, potentially making them easier for the immune system to find and attack.
What it means for the future of AI and medicine
The success of C2S-Scale demonstrates that larger AI models can do more than improve existing research tasks; they can generate entirely new scientific ideas.
As Google noted, the result “provides a blueprint for a new kind of biological discovery,” showing how scaled AI systems can uncover complex biological interactions that traditional methods might miss.
Google says the teams at Yale are now conducting deeper studies into the mechanism discovered by the model and testing other AI-generated hypotheses in various immune settings.
To accelerate progress, Google has made the C2S-Scale 27B model and its resources available to the scientific community. Researchers can access the model on Hugging Face and explore its code on GitHub, potentially building on this discovery to find more treatments in the future.
Arm and Meta have announced a long-term strategic partnership to advance AI efficiency across every level of computing — from software to large-scale data center infrastructure.