Overview
Co-Scientist is a multi-agent artificial intelligence (AI) system designed for structured scientific thinking and hypothesis generation. Built upon Gemini, its primary objective is to assist scientists in discovering new original knowledge. The system formulates research hypotheses for experimental verification, conditioned by research objectives and existing scientific evidence. Its architecture facilitates continuous generation, critique, and refinement of hypotheses, leveraging scalable test-time compute.
Research Context
Scientific discovery conventionally relies on scientists generating novel hypotheses for complex problems, followed by rigorous experimental validation. Co-Scientist was developed to augment this established process by providing computational support for hypothesis formulation.
Approach
The design of Co-Scientist incorporates a multi-agent architecture. This architecture utilizes an asynchronous task execution framework, which allows for flexible scaling of computational resources during testing. A key component of the system is its tournament evolution process, engineered for the self-improvement of hypothesis generation capabilities. This process enables agents to continuously generate, critique, and refine hypotheses.
Findings
Automated evaluations of Co-Scientist indicated a sustained benefit from scaling test-time compute, demonstrating an improvement in hypothesis quality over time. While the system is general-purpose, its validation focused on three specific biomedical applications:
- Drug repurposing
- Novel target discovery
- Explaining mechanisms of anti-microbial resistance
In the context of acute myeloid leukemia, Co-Scientist specifically identified new drug repurposing candidates. It also identified synergistic combination therapies for this condition. These identified candidates and therapies underwent validation through in vitro experiments.
Why This Matters
The real-world validations observed with Co-Scientist suggest its potential to accelerate scientific discovery. This acceleration could contribute to an era where AI systems empower scientists in their research efforts, particularly in the process of generating and refining novel hypotheses for experimental verification.