Google’s new AI model points the way to making hidden tumors visible

An artificial intelligence system has discovered a potential method for making hidden tumors visible to the immune system, researchers from Google and Yale University reported.

The 27-billion-parameter model, called Cell2Sentence-Scale 27B, predicted that combining an existing drug called silmitasertib with low doses of interferon could enhance the immune system’s ability to detect certain cancers. Laboratory experiments subsequently confirmed the prediction, showing a 50 percent increase in antigen presentation, the process by which cancer cells become visible to immune cells.

The discovery, published as a preprint on bioRxiv and announced on Google’s research blog, addresses a fundamental challenge in cancer immunotherapy; many tumors remain “cold,” or invisible to the body’s immune defenses, rendering them resistant to otherwise promising treatments.

“This announcement marks a milestone for AI in science,” the research team wrote in their blog post, noting that the discovery “reveals a promising new pathway for developing therapies to fight cancer.”

Reading our genetic information like a book

Rather than training the AI on medical images or clinical data, researchers at Google Research and Yale’s van Dijk Lab taught it to interpret genetic information as a form of language, converting cellular RNA sequences into “cell sentences” that could be processed like text.

The system was built on Google’s Gemma-2 architecture, originally designed for natural language processing. By training progressively larger versions of the model, from 410 million to 27 billion parameters, on more than one billion tokens of biological data, the researchers discovered that only the largest models could perform the kind of conditional reasoning necessary for drug discovery.

When tasked with finding drugs that would amplify immune responses only in specific biological contexts, the 27-billion-parameter model identified an unexpected candidate: silmitasertib, a kinase inhibitor that had never been associated with enhancing tumor visibility to the immune system.

“This required a level of conditional reasoning that appeared to be an emergent capability of scale,” the researchers noted in their paper. “Our smaller models could not resolve this context-dependent effect.”

A hidden synergy comes to light

The true test came in the laboratory at Yale, where scientists evaluated the AI’s hypothesis using human neuroendocrine cells, a cell type the model had never encountered during training.

The experiments followed a precise protocol to test the predicted synergy. Cells treated with silmitasertib alone showed no change in antigen presentation, and low-dose interferon produced modest improvements. But when combined, as the AI had predicted, the drugs produced a marked amplification effect, increasing antigen presentation by approximately 50 percent.

This validation is particularly significant because silmitasertib’s maker, Senhwa Biosciences, had been developing the drug for other cancer applications. While CK2, the enzyme targeted by silmitasertib, had been studied in various immune contexts, its potential role in enhancing MHC-I expression (the mechanism for antigen presentation) had not been previously identified.

Maybe all we need is scale

The study adds weight to a contentious idea in AI that simply making models larger can unlock qualitatively new capabilities, a claim critics argue distracts from other important factors like data quality and architecture design. The team tested five model sizes and observed consistent improvements in performance as parameters increased, with the ability to identify context-dependent drug interactions emerging only at the largest scale.

“These scaling trends are observed in both full fine-tuning and parameter-efficient regimes,” the researchers wrote, suggesting that the approach could be adapted for institutions with varying computational resources.

The model family ranges from 410 million to 27 billion parameters and was trained using a combination of Google’s TPU infrastructure and open-source tools. The researchers employed modern reinforcement learning techniques during fine-tuning to optimize the models for biological reasoning tasks.

When saving time means saving lives

The discovery could accelerate the traditionally lengthy process of drug development, particularly for combination therapies. Instead of screening thousands of compounds through trial and error, researchers could use AI models to generate targeted hypotheses about drug interactions.

The approach is especially relevant for immunotherapy, where combination treatments often prove more effective than single agents. Current immunotherapy drugs like pembrolizumab (Keytruda) and nivolumab (Opdivo) have transformed treatment for some cancers but remain ineffective for many patients whose tumors evade immune detection.

Dr. David van Dijk, whose Yale laboratory collaborated on the research, indicated that his team is now investigating the mechanisms underlying the discovered pathway and testing additional AI-generated predictions in other immune contexts.

Giving everyone the chance to benefit

In keeping with recent trends in AI research, Google and Yale have made the C2S-Scale models publicly available through Hugging Face and GitHub, letting other teams test the findings and explore new uses.

The release includes models of varying sizes to accommodate different computational budgets, from the full 27-billion-parameter version to smaller variants suitable for academic laboratories with limited resources.

This open approach contrasts with the proprietary strategies of some pharmaceutical companies and could accelerate validation and expansion of the technique. Researchers can fine-tune the models for specific biological questions or integrate them into existing drug discovery pipelines.

An AI-generated light at the end of the tunnel

Significant challenges remain before the discovery could translate to clinical applications. The findings require extensive preclinical validation in animal models, followed by human clinical trials that typically span years.

The research also raises questions about the interpretability of AI-generated hypotheses. While the model successfully predicted the drug combination’s effect, understanding why silmitasertib works in this specific context remains an active area of investigation. The mechanism by which CK2 inhibition amplifies interferon signaling to enhance antigen presentation requires further study, as well. Then there are the safety concerns, as both silmitasertib and interferon have known side effects and their combination in whole organisms may produce interactions not apparent in cell culture.

Additionally, the model’s predictions were validated in one specific cell type. Whether the same synergistic effect occurs across different cancer types and in the complex tumor microenvironment of living patients remains unknown. Cancer’s notorious heterogeneity means that therapies effective in laboratory settings frequently fail to translate to clinical success.

Despite the hurdles, the work between Google and Yale represents an important proof of concept. By showing that large language models can generate hypotheses worth testing, and that some hold up in the lab, the collaboration hints at a new role for AI in discovery science: helping researchers see patterns they might have missed.