Google has released a new AI model called Cell2Sentence-Scale 27B (C2S-Scale) in collaboration with Yale that can potentially offer a new pathway for developing therapies to fight cancer. Built on the Gemma models, the 27 billion parameter model was able to generate a novel hypothesis around how cancer cells behave, which the team later confirmed as correct. 

CEO Sundar Pichai hailed the achievement on X calling it “an exciting milestone for AI in science.”

The AI model was designed to understand the “language” of individual cells. 

One of the biggest challenges with cancer immunotherapy is that most tumours are “cold,” or invisible to the body’s immune system. To make them discoverable, tumours have to be forced to display immune-triggering signals through a process called antigen presentation. 

The C2S-Scale 27B model was asked to find a drug that can act as a ‘conditional amplifier,’ so the immune signal is boosted in a specific “immune-context-positive” environment where some key immune-signalling protein already existed. 

Google said that while the smaller AI model didn’t have the reasoning capabilities to resolve this problem, when scaled up the model eventually could.

The team then simulated 4,000 drugs across patient samples for the model to predict which ones would only boost antigen presentation in the patient-relevant setting. The AI model was able to highlight between 10-30% drugs which were already known while the others were entirely novel.

Google has made the C2S-Scale 27B model available on GitHub and Hugging Face.