Google, Yale’s AI model finds potential way to make cancers more visible to immune system
Google and Yale’s 27-billion-parameter C2S-Scale model has produced a lab-confirmed finding that could help make some tumours easier to target.
Google, in collaboration with Yale University, has announced Cell2Sentence-Scale 27B, or C2S-Scale, a large artificial intelligence (AI) model designed to read and predict what individual cells are doing.
The model has already led to a notable discovery, predicting a new way to make some cancers more visible to the immune system, a prediction that laboratory experiments have confirmed.
This suggests that AI could help uncover entirely new pathways for cancer therapy, potentially speeding up how scientists develop and test treatments.
Alphabet and Google chief Sundar Pichai noted that the model “generated a novel hypothesis about cancer cellular behaviour, which scientists experimentally validated in living cells.”
“With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer,” he added in a post on X.
C2S-Scale is built on the open Gemma family of models and is now available to the research community.
It works by analysing single-cell data, which records activity inside individual cells, such as which genes are switched on or which proteins are active. By treating these measurements like a “language,” the model can spot patterns, predict how cells will respond, and even suggest experiments.
According to a blog post, the team focused on a major challenge in cancer therapy. Many tumours are immunologically “cold,” meaning the immune system struggles to detect them. Cells display fragments of internal proteins on their surface to alert immune cells, a process called antigen presentation. Interferon is a natural immune signal that can boost this process, but in many tumours, interferon levels are too low to be effective.
The researchers asked C2S-Scale to find a drug that could act as a conditional amplifier, one that would boost antigen presentation only when a weak interferon signal was present, leaving cells without immune signals unchanged.
To test this, the team ran a two-part virtual screen. One context used patient-derived samples that preserved tumour-immune interactions, while the other used isolated cell lines with no immune context. The model simulated the effects of over 4,000 drugs and highlighted those that increased antigen presentation only in the patient-like setting.
A standout prediction was silmitasertib, also known as CX-4945, which inhibits the enzyme CK2. While CK2 is involved in many cellular functions, the model suggested silmitasertib could specifically boost antigen presentation in the presence of low interferon.
Laboratory tests on human neuroendocrine cells, which the AI had not seen before, confirmed this. Silmitasertib alone had no effect, low-dose interferon alone had a modest effect, and the combination produced a roughly 50% increase in antigen presentation.
The teams are continuing this research, and C2S-Scale is available for other scientists to explore and build upon. The model and resources are available on HuggingFace and GitHub for researchers.
Edited by Jyoti Narayan