Gemma AI identifies a potential path against cancer

An AI model based on Gemma identified a possible way to make some tumors more visible to the immune system. Google and Yale University tested the prediction in human cells and observed an approximately 50% increase in antigen presentation, a process that helps the body's defenses recognize tumor cells.

The result is not yet a treatment. It is an AI-generated hypothesis validated in the laboratory, an early step that will need to clear preclinical studies and clinical trials before it can have a medical application.

A model that analyzes the language of cells

The model is called Cell2Sentence-Scale 27B, or C2S-Scale. It has 27 billion parameters and is built on the open Gemma model family. Its goal is to interpret single-cell data as if it were a language, detecting relationships between biological signals, drugs and cellular behavior.

Google and Yale used it to search for a drug that could act as a conditional amplifier: it had to increase an immune signal only when a small amount of interferon was already present, a protein that coordinates part of the body's defensive response. The signal had to be insufficient on its own, but useful when combined with the drug.

To find it, the team simulated the effect of more than 4,000 drugs in two scenarios:

• Patient samples showing interaction between the tumor and the immune system, along with low levels of interferon.
• Isolated cells without that immune context.

The goal was to find drugs that worked in the first scenario, which more closely resembles a real tumor, but not in the second. The candidates included drugs that were already known, as well as others with no documented connection to this effect.

The prediction that made it to the lab

The model highlighted silmitasertib, also known as CX-4945, an inhibitor of the protein kinase CK2. It predicted that the drug would increase antigen presentation when interferon levels were low, but would have little or no effect without that context.

The prediction mattered because it did not simply repeat a known connection. Although CK2 is involved in several cellular functions and had already been linked to the immune system, there was no specific description of silmitasertib increasing the expression of MHC-I, a molecule that helps the body's defenses detect abnormal cells.

The researchers tested the hypothesis in human neuroendocrine cell models that the model had not seen during training. The results were as follows:

• Silmitasertib on its own did not change antigen presentation.
• A low dose of interferon produced a moderate effect.
• The combination of both produced a marked, synergistic increase of approximately 50%.

The tests were repeated in the laboratory and confirmed the model's prediction. In simple terms, the combination could help turn a tumor considered “cold”, meaning poorly visible to the immune system, into one that is easier to recognize and more responsive to immunotherapy.

What changes and what remains to be proven

For you, the news does not mean that a new cancer therapy already exists. It means AI can help find drug combinations and biological effects that were not obvious when each piece of data was analyzed separately.

The Yale team is still studying how this interaction works and testing other predictions from the model. Before involving patients, it will need to establish the combination's safety, efficacy and appropriate dose in more preclinical models and then in clinical trials.

Google has made C2S-Scale 27B and its resources available to the research community. The important question will be whether other predictions from the model are validated outside this experiment. That will show whether it can become a regular tool for designing therapies or whether this result remains a promising proof of concept.