Google DeepMind Uses Gemma-Based Model to Uncover Potential Cancer Therapy Pathway

In a breakthrough that highlights the growing role of artificial intelligence in scientific discovery, Google DeepMind, in collaboration with Yale University, has unveiled Cell2Sentence-Scale 27B (C2S-Scale) — a 27-billion parameter foundation model designed to interpret the “language” of individual cells.

Built on the Gemma family of open models, the system has already led to the discovery of a promising new cancer therapy pathway.

Google researchers tasked their new C2S-Scale 27B model with finding a drug that could act as a conditional amplifier—enhancing immune signals only in “immune-context-positive” environments where low interferon levels exist. Demonstrating advanced reasoning beyond smaller models, C2S-Scale analysed over 4,000 drugs using a dual-context virtual screen.

The model compared real patient samples with active immune interactions to isolated cell data lacking such context, identifying drugs that specifically boosted antigen presentation in the immune-relevant setting. Notably, while 10–30% of these hits were already documented, many were entirely novel discoveries, revealing the model’s potential for context-aware drug discovery.

This type of conditional reasoning, researchers said, was an emergent capability that smaller models failed to capture.

To validate the model’s findings, DeepMind and Yale researchers tested silmitasertib on human neuroendocrine cells. When combined with low-dose interferon, the drug significantly boosted antigen presentation — the process by which tumor cells become visible to immune cells — by roughly 50%. On its own, neither agent produced a comparable effect.

This dual-context discovery suggests a new approach for turning “cold” tumors, which typically evade the immune system, into “hot” ones that can be targeted by immunotherapy.

Beyond its immediate cancer implications, the C2S-Scale model provides a blueprint for AI-driven biological discovery, allowing researchers to conduct large-scale virtual drug screenings and uncover context-dependent biological mechanisms.