DeepMind validates cancer hypothesis Cold tumors: make hot
Google DeepMind and Yale University report on an AI model that generated a biological hypothesis for making tumors visible to the immune system. This hypothesis was subsequently confirmed in living cells. Alphabet-CEO Sundar Pichai calls the validation in cells a milestone.
Quelle: Own representation
AI-assisted cancer research
An open Gemma-based AI model named C2S-Scale 27B generated a new hypothesis about cancer biology. It predicted that inhibition of the CK2 kinase by the drug Silmitasertib would markedly increase antigen presentation of tumor cells, but only with weak interferon signaling. This prediction was confirmed in human neuroendocrine cell models. The combination of Silmitasertib and low-dose interferon increased antigen presentation by about 50 percent ( Google Blog). ). This represents a rare case of "AI → hypothesis → wet lab → hit" ( Google Blog).
C2S-Scale 27B is a large language model for single-cell biology. It converts gene expression profiles of cells into "sentences" of genes ("cell sentences"). As a result, it can answer biological questions in natural language, interpret and simulate responses to interventions ( Google Research Blog). ). The work is based on a scientific preprint describing the scaling of this method to 27 billion parameters ( bioRxiv).
Quelle: ausgezeichnet.org
Artificial intelligence plays a key role in analyzing complex biological data and the development of new therapeutic approaches.
Biological Foundations
"Cold" tumors are cancer types in which the immune system hardly penetrates. This is often due to weak antigen presentation, stromal barriers, and immunosuppressive signals. "Hot" tumors, on the other hand, have many immune cells and respond more to immunotherapy ( Molecular Cancer). ). Antigen presentation, often via MHC-I molecules, makes tumor cells recognizable to T cells. Interferons can generally boost this presentation ( PMC NCBI). ). Silmitasertib (CX-4945) is a clinically investigated inhibitor of the CK2 kinase, which regulates diverse cellular processes ( PMC NCBI).
Quelle: spektrum.de
CAR-T cell therapy is a promising approach to activate the immune system in the fight against cancer.
Current status & development
In April 2025, researchers from Yale, Google Research and DeepMind presented the scaled "Cell2Sentence" idea. Single-cell data are turned into text, and LLMs learn to "read" and to "write" biology ( Google Research Blog; bioRxiv).
On October 15, 2025 it was announced that the 27B model generated a clinically potent hypothesis and passed the first real-world laboratory test. The researchers let C2S-Scale think across two contexts: once with weak interferon signaling (immune-context-positive) and once without immune context. Over 4,000 compounds were simulated. The model identified Silmitasertib as a "conditional amplifier" that upregulates antigen presentation only in the appropriate interferon environment ( Google Blog).
Subsequently, human neuroendocrine cell models that were not known to the model were tested. Silmitasertib alone showed no effect. Low-dose interferon alone had a small effect. The combination led to about 50 percent more antigen presentation, which means increased "visibility" for the immune system ( Google Blog).
Preprint, code and models are openly accessible. The Yale partner explains background and points to GitHub/Hugging Face ( vandijklab.org; Hugging Face; GitHub).
Analysis & Assessment
This approach aims to move from "AI in analysis" to "AI as discoverer," by generating hypotheses rather than just prioritizing them ( Google Research Blog). ). Another driver is clinical need, as many tumors remain "cold" for immunotherapies, and ways to targeted "warming" are a key question ( Molecular Cancer). ). The release of preprint, code and models accelerates replication and extension and promotes the reputation of Gemma/C2S ( vandijklab.org; Hugging Face).
Silmitasertib is known, but the interferon-dependent amplification of antigen presentation shown here is new and context-dependent. This potentially reduces side effects, as the "boost" occurs only in the appropriate immunological milieu ( Google Blog; PMC NCBI).
Quelle: YouTube
The C2S-Scale 27B model generated a hypothesis that was confirmed in living cells (in vitro). Silmitasertib plus low-dose interferon increased antigen presentation by about 50 percent ( Google Blog). ). C2S-Scale translates single-cell data into "cell sentences" and follows clear scaling laws; the preprint is public ( Google Research Blog; bioRxiv).
It remains unclear how transferable this is across different tumor types, the dose window, safety, and duration of effect in animals and humans. So far there are no preclinical animal data or clinical data for this combination in this mechanism ( Google Blog). ). Headlines that talk about "AI cures cancer" or suggest clinical efficacy exaggerate the finding. The work shows a lab-validated hypothesis, not a patient study ( Economic Times).
Sundar Pichai summarizes the news as an "exciting milestone" and points to the validation in cells ( X.com). ). The Yale partner lab (van Dijk Lab) positions C2S-Scale as a platform for "virtual cells" and open collaboration ( vandijklab.org). ). Independent reports classify the step as a potential new path for immunotherapies, but emphasize the early state of research ( Decrypt). ). Skepticism remains about clinical translatability, as interferon pathways are complex and not free from side effects ( PMC NCBI).
Quelle: user-added
This diagram shows methods for enriching and detecting circulating tumor cells, circulating tumor DNA, exosomes, and tumor-cell–associated platelets from blood samples.
Outlook & Open Questions
For researchers, the open resources facilitate replication, comparison, and extension. Preprint, code and models are directly available ( bioRxiv; GitHub; Hugging Face). ). Clinicians should note that this is an exciting mechanism, but there are no clinical efficacy data. Literature on antigen presentation and interferon helps with context ( PMC NCBI). ). Interested readers should consult primary sources and verify whether results are confirmed in animals and humans.
Quelle: YouTube
Open questions include which tumor entities truly benefit from interferon-dependent CK2 inhibition and how large the clinically relevant in vivo effect is. In addition, dosing and combinations that are safe and effective, as well as interactions with established therapies such as checkpoint inhibitors, need to be clarified. The robustness of the effect outside the tested neuroendocrine models is also open. According to Google/Yale, mechanistic follow-up work and further tests are underway; peer review and preclinical animal models are planned ( ). The statement "DeepMind and Yale: AI hypothesis for cancer validated" is essentially true. An AI generated a new, context-dependent idea, and the lab confirmed it in cells. This bridge from simulation to experiment has been rarely achieved ( Google Blog; bioRxiv).
). The path to patients is long, but the open, replicable and biologically plausible pattern shows how AI can accelerate discoveries in the future ( Google Blog). ); vandijklab.org; Google Research Blog).