The big picture: An AI model jointly developed by Google and Yale University has produced a groundbreaking hypothesis about how cancer cells interact with the human immune system. Researchers believe the discovery could represent one of the most significant breakthroughs in cancer therapy to date and has the potential to transform how the disease is treated in the future.

The hypothesis was generated by a 27-billion-parameter foundation model called Cell2Sentence-Scale 27B (C2S-Scale), developed by researchers at Google DeepMind and Yale University. Built on Google's open-source Gemma AI model, C2S-Scale is designed for single-cell analysis, enabling researchers to predict the behavior of cancer cells within living organisms.

Clinical validation has confirmed the model's predictions, potentially paving the way for more effective cancer therapies. The discovery builds on Google's earlier research, which showed that biological AI models follow clear scaling laws: larger models exhibit higher levels of conditional reasoning, similar to the behavior of natural language AI systems.

According to Google, the C2S-Scale 27B model can interpret the "language" of individual living cells, allowing it to transform hard-to-detect "cold" tumors into "hot" tumors. This process makes malignant cells more visible to the immune system and more responsive to therapy.

The new AI model also successfully identified a conditional amplifier drug capable of boosting the body's immune signal in specific contexts – for example, when the immune-signaling protein interferon fails to induce antigen presentation on its own due to insufficient levels. This capability was not observed in smaller AI models tasked with similar challenges.

// Related Stories

Explaining how C2S-Scale was trained to reason through complex biological conditions, Google stated that its researchers designed a so-called "dual-context virtual screen," simulating the effects of more than 4,000 drugs across real-world patient tumor samples and isolated cell line data without any immune context.

When asked to identify drugs that could selectively enhance antigen presentation in the first context, the model highlighted several candidates – only 10 – 30 percent of which were previously known to be effective in cancer treatment. The remaining predictions had no prior known link to the screen or to cancer immunotherapy. These predictions were subsequently validated in clinical applications.

Both Gemma and C2S-Scale 27B are publicly available on Hugging Face and GitHub. Google has also posted a scientific preprint on bioRxiv to assist researchers in running virtual drug screens capable of uncovering potentially life-saving hypotheses. Researchers caution, however, that all predictions will require peer review and clinical validation before being adopted for therapeutic use.