October 2025 – The Yale University, is proud to announce the launch of Cell2Sentence-Scale 27B (C2S-Scale), a groundbreaking 27-billion parameter foundation model engineered to interpret the molecular “language” of individual cells.
Built upon the Gemma family of open models, C2S-Scale represents a significant advancement in single-cell analysis and introduces a transformative approach to biological discovery powered by large-scale artificial intelligence.
A Milestone for AI in Biological Research
The introduction of C2S-Scale marks a pivotal moment in the intersection of AI and life sciences. During its testing phase, the model generated an entirely novel hypothesis on cancer cell behavior, which was later experimentally validated in living cells. This successful prediction uncovers a promising new biological pathway with potential implications for future cancer therapies.
This achievement builds upon earlier research demonstrating that biological AI models, like language models, follow predictable scaling laws—where increasing model size leads to improved performance. Yet, this work also posed a deeper question: Can larger biological models go beyond existing tasks to uncover new biological mechanisms? C2S-Scale 27B provides a compelling answer—yes, they can.
How C2S-Scale 27B Works
A persistent challenge in cancer immunotherapy is that many tumors remain “cold,” meaning they are invisible to the immune system. A critical strategy for overcoming this involves enhancing antigen presentation, enabling tumors to display immune-activating signals.
C2S-Scale 27B was tasked with identifying a conditional amplifier—a drug that would boost immune signaling only within an “immune-context-positive” environment (where low levels of interferon are present) but remain inactive in neutral conditions. This form of context-dependent reasoning emerged as a capability unique to large-scale models.
To identify the right candidate, researchers designed a dual-context virtual screen involving two experimental conditions:
- Immune-Context-Positive: Real patient samples with intact immune interactions and low interferon signaling.
- Immune-Context-Neutral: Isolated tumor cell lines lacking immune context.
Simulating over 4,000 drug interactions, the model highlighted several promising compounds—about 10–30% of which were previously known, with the remainder representing novel, unreported candidates.
From AI Prediction to Laboratory Validation
Among these predictions, the model identified a striking context-specific effect for the kinase CK2 inhibitor silmitasertib (CX-4945). C2S-Scale predicted that silmitasertib would significantly enhance antigen presentation only in immune-context-positive conditions—a novel hypothesis not yet reported in scientific literature.
Subsequent laboratory experiments on human neuroendocrine cells confirmed the AI’s predictions:
- Silmitasertib alone did not increase antigen presentation (MHC-I).
- Low-dose interferon alone produced a modest effect.
- The combination of silmitasertib and low-dose interferon yielded a synergistic 50% increase in antigen presentation—effectively making “cold” tumors “hot.”
This outcome marks a major step forward in using AI to discover condition-specific biological mechanisms and opens new directions for combination cancer therapies.
A Blueprint for AI-Driven Biological Discovery
C2S-Scale 27B demonstrates that scaling AI models in biology not only improves prediction accuracy but also enables the generation of novel, testable hypotheses. It sets a precedent for using foundation models to conduct virtual high-throughput drug screens and uncover context-conditioned biological phenomena.
Collaborating teams at Yale are now extending this research to explore additional AI-generated predictions across different immune environments. With continued preclinical and clinical validation, these insights could help accelerate the discovery of next-generation immunotherapies.
Access and Availability
The C2S-Scale 27B model and its associated resources are now available to the research community. Researchers and institutions are invited to explore, collaborate, and build upon this work to further advance our shared mission—translating the language of life through artificial intelligence.
