We challenged the new C2S scale 27B model. The challenge is to find drugs that act as conditional amplifiers, that is, drugs that enhance immune signals only in certain “immune-positive” settings, where low levels of interferon (an important immune signaling protein) are already present but alone are insufficient to induce antigen presentation. This required a level of conditional reasoning that appeared to be a new ability to scale. Our small model could not resolve this context-dependent effect.
To achieve that, we designed a dual-context virtual screen to find this unique synergy. The virtual screen includes two stages.
Immune context positivity: We provided the model with real patient samples with intact tumor-immune interactions and low levels of interferon signaling. Immune context neutral: Isolated cell line data without immune context was provided to the model.
They then simulated the effects of more than 4,000 drugs in both situations and asked the model to predict which drugs would promote antigen presentation only in the first situation, biasing the screen towards patient-relevant settings. Of the many drug candidates highlighted by the model, some (10-30%) of the hit drugs are already known in the prior literature, while the remaining drugs are surprising hits with no prior known relevance to the screen.
From prediction to experimental verification
The model’s predictions were clear. It identified a significant “context split” in a kinase CK2 inhibitor called cilmitasertib (CX-4945). This model predicted a significant increase in antigen presentation when cilmitasertib was applied in an “immune status positive” setting, but had little effect in an “immune status neutral” setting. What makes this prediction so interesting is that it was a novel idea. Although CK2 is thought to be involved in many cellular functions, including its role as a regulator of the immune system, there are no reports in the literature that inhibition of CK2 by Silmitasertib explicitly enhances MHC-I expression or antigen presentation. This emphasizes that models are not simply repeating known facts, but are generating new testable hypotheses.
However, predictions are only valuable if they can be validated in clinical applications. The real test is done first in a lab and eventually in a doctor’s office.
The next stage of the project was to take this hypothesis to the bench and test it in a human neuroendocrine cell model. This cell type was not recognized at all by the model during training. The experiment demonstrated the following:
Treatment of cells with Silmitasertib alone had no effect on antigen presentation (MHC-I). Treatment of cells with low doses of interferon alone had a moderate effect. Treatment of cells with both silmitasertib and low doses of interferon resulted in a significant synergistic amplification of antigen presentation.
Of note, in our clinical tests, the combination of silmitasertib and low doses of interferon increases antigen presentation by approximately 50%, making tumors more recognizable to the immune system.
The in silico predictions of the model were confirmed multiple times in vitro. The C2S scale successfully identified novel interferon conditional amplifiers, revealing new potential pathways that could make “cold” tumors “hot” and increase their responsiveness to immunotherapy. Although this is an early first step, it provides a powerful, experimentally validated lead for developing new combination therapies that combine multiple drugs to achieve more potent effects.
The results also provide a blueprint for new kinds of biological discoveries. This shows that by following scaling laws and building large-scale models such as C2S scale 27B, we can create predictive models of cell behavior that are powerful enough to perform high-throughput virtual screens, discover context-conditioned biology, and generate biologically grounded hypotheses.
The Yale team is currently investigating the mechanisms uncovered here and testing additional AI-generated predictions in other immune contexts. With further preclinical and clinical validation, such hypotheses could ultimately accelerate the path to new treatments.
Overview of C2S Scale 27B
The new C2S scale 27B model and its resources are now available to the research community. Please use these tools to grow our work and help us continue translating the languages of life.
