Last month, Google officially released a new artificial intelligence system calledAI co-scientist. This system is developed based on the latest Gemini 2.0 model and aims to assist researchers in generating innovative hypotheses and research plans.
Researchers only need to describe their research goals in natural language, such as "a deeper understanding of the transmission mechanisms of pathogenic microorganisms,"AI co-scientistand it will propose testable hypotheses, summarize existing literature in relevant fields, and provide possible experimental design plans.
AI co-scientistIt is not intended for automating scientific research but serves as a collaborative tool to help scientists collect information more efficiently, optimize research ideas, and ultimately promote scientific progress.
The AI co-scientist system includes a series of specialized agents: Generation, Reflection, Ranking, Evolution, Proximity, and Meta-review. Researchers can naturally describe their research goals, and the system automatically generates scientific hypotheses, reviews, and experimental designs. These agents iterate through automatic feedback, continuously generating, evaluating, and optimizing hypotheses, forming a self-improving loop.
Designed specifically for collaboration, researchers can interact with the system by providing seed ideas or feedback. The AI co-scientist also uses tools like web searches and dedicated AI models to improve the quality of generated hypotheses.
To validate the practical utility of the AI co-scientist, Google conducted real-world experimental validations in multiple biomedical fields, including drug repurposing, discovery of new therapeutic targets, and elucidation of antimicrobial resistance mechanisms.
Under expert guidance, the AI co-scientist explored a previously discovered but unpublished phenomenon by their team: how "capsid-forming phage-inducible chromosomal islands" (cf-PICIs) exist across multiple bacterial species. Without prior access to experimental data, the AI co-scientist independently proposed that cf-PICIs could interact with various phage tail structures, thereby expanding their host range. This computational inference had already been experimentally validated in the laboratory.
In drug repurposing research for acute myeloid leukemia (AML), the drugs suggested by the AI co-scientist were experimentally verified to effectively inhibit tumor cell activity at clinically relevant concentrations.
Additionally, the system successfully proposed novel epigenetic targets for hepatic fibrosis and provided new insights into the study of antimicrobial resistance mechanisms.
Despite the limited sample size, experts rated the novelty and impact of the results generated by the AI co-scientist highly, surpassing other similar models. Moreover, through the Elo self-assessment metric, the output quality of the system continues to improve with increased computational resources.
