Amid the surging open-source AI ecosystem, Google's open-source edge-side model family has reached a new milestone. On August 20 local time, Google officially announced that the total number of downloads for its open-source large model Gemma has exceeded 1 billion. Since its release, more than 100,000 customized variants have been developed by global developers based on this series, successfully building a vast "Gemmaverse" open-source ecosystem.

From the perspective of practical applications, the Gemma family has already gone beyond the boundaries of simple dialogue generation and penetrated into core scenarios across various vertical industries. Currently, the model is widely applied in on-site image analysis in space, standardization of medical reports, design of innovative cancer treatment plans, and animal language recognition among other cutting-edge fields.

In biomedical research, Gemma has shown disruptive application value. Google has launched a specialized vertical model called MedGemma, tailored for medical scenarios to support structured medical reports and medical imaging assistance. Meanwhile, it has collaborated with top research institutions, leveraging the Gemma foundation to tackle cancer mechanisms. Among them, the Cell2Sentence-Scale 27B (C2S-Scale) model, jointly developed by Google DeepMind, Google Research, and Yale University, has become a focal point in the industry.

The traditional approach to cancer vaccines is to enable the immune system to identify "identity tags" to locate enemies. However, when facing many "cold tumors" lacking unique antigens on their surface and highly expressing PD-L1 proteins, the immune system often suffers from "face blindness," unable to recognize or be suppressed even under massive attack. To solve this treatment challenge, C2S-Scale innovatively sorts single-cell gene expression data by activity level, converting it into "gene sentences" similar to text, allowing large language models to directly "read" cells.

Based on this, the model efficiently completed virtual screening of more than 4,000 drugs, ultimately predicting a drug combination of silmitasertib combined with low-dose interferon. Most of the candidate drugs selected had no known association with the target, and the selected silmitasertib (CX-4945) is a CK2 kinase inhibitor.

This new treatment hypothesis proposed by AI was subsequently validated in human live cell experiments. The experimental results showed that using the drug alone or low-dose interferon alone had little effect, but the combination significantly increased antigen presentation by about 50%, successfully confirming that the combination could significantly enhance antigen presentation levels, potentially transforming immunotherapy-insensitive "cold tumors" into "hot tumors." This is one of the few successful cases where AI has generated a new cancer treatment mechanism and obtained laboratory validation.

Although the relevant findings have been fully open-sourced for further research by global research teams, the discovery is still at the stage of in vitro cell experiments and has not yet entered clinical trials. Nevertheless, the 1 billion downloads not only prove the strong capabilities of Gemma as a low-cost technical foundation but also mark the deep involvement of open-source large models in the mining of original biological data, ushering in a new era for life science research.