Google has introduced a new suite of artificial intelligence (AI) tools that successfully suggested a novel drug combination for cancer treatment, demonstrating promising results in preliminary laboratory tests. This development signals a crucial moment for the integration of AI into scientific research. The AI model, known as Cell2Sentence-Scale 27B (C2S-Scale), features 27 billion parameters and is built upon the Gemma family of open models, designed specifically to interpret cellular language.
According to Shekoofeh Azizi and Brian Perozzi, researchers at Google DeepMind and Google Research, respectively, this announcement represents a significant milestone for AI in the scientific realm. They stated, “C2S-Scale generated a novel hypothesis about cancer cellular behavior, and we have since confirmed its prediction with experimental validation in living cells. This discovery reveals a promising new pathway for developing therapies to fight cancer.”
The C2S-Scale model was trained using a comprehensive dataset that included real-world patient and cell-line data. It posited that the drug silmitasertib could enhance the immune system”s ability to detect early-stage cancerous tumors. Currently, silmitasertib (CX-4945) is undergoing several clinical trials aimed at treating conditions such as multiple myeloma, kidney cancer, medulloblastoma, and advanced solid tumors. In January 2017, the U.S. Food and Drug Administration awarded the drug orphan status for advanced cholangiocarcinoma.
What sets Google”s initiative apart is not merely the rediscovery of silmitasertib but its utilization of extensive cancer biology literature to propose this new application. Typically, pharmaceutical companies invest billions and rely on specialized teams to unearth such insights. Sunil Laxman, a systems biologist at the Institute for Stem Cell Science and Regenerative Medicine in Bengaluru, commented, “It”s a nice result and was a well-chosen problem to test the capabilities of a large language model (LLM). This would have ordinarily taken a focused team of researchers several months to arrive at such a suggestion.”
However, Dr. Laxman emphasized that the model did not provide insights that a trained biologist would not have considered. He described the outcome as “very good, not great,” pointing out that many laboratories may lack access to the extensive chemical compound libraries necessary for testing. While this AI approach has streamlined the process of potential discovery, it does not represent a groundbreaking advancement in cancer biology.
LLMs are fundamental to AI and are developed using human-annotated data to comprehend language and solve complex problems. Drs. Azizi and Perozzi argue that their findings indicate the possibility of creating LLMs that can derive biological rules by rewarding successes and penalizing failures, rather than being explicitly trained on those rules. This methodology mirrors how some of the most advanced chess-playing LLMs were developed.
Siddhartha Gadgil, a mathematics professor at the Indian Institute of Science in Bengaluru, regarded the findings as significant. He noted that the leading AI models in mathematics currently match the capabilities of skilled mathematicians but have not yet reached the genius level. “We cannot predict when an AI will solve the Riemann Hypothesis, but there is no reason to believe it will never happen,” he stated. Gadgil cited the International Mathematical Olympiad 2025, where a model from OpenAI demonstrated the ability to solve problems at a level comparable to a gold medal-winning human participant.
Despite these advancements, the Olympiad”s problems differ from those faced by professional mathematicians and are designed for high school students. The Riemann Hypothesis, a longstanding challenge concerning the nature of prime numbers, remains unsolved and is a focus for many mathematicians. The Clay Mathematical Institute has offered a reward of $1 million for a solution.
Given the extensive literature available to LLMs, they are expected to propose innovative approaches to mathematical problems that may not be immediately apparent to human experts. Prof. Gadgil, who actively utilizes LLM tools in his research, acknowledged that the academic community remains divided on their incorporation, but he believes that models with untapped potential should be embraced by researchers.
