FORSMILE
JA
AI2026/06/07

DeepMind's "Co-Scientist" Emerges! The Era of AI as a Partner in Scientific Research

Google DeepMind has announced "Co-Scientist," a new AI designed to accelerate scientific research. Built on Gemini, this AI autonomously generates and tests hypotheses, and has already achieved groundbreaking results in biology. The process of scientific discovery may be fundamentally transformed.

Back to Blog

The Birth of an "AI Partner" to Accelerate Scientific Discovery

Google DeepMind has announced "Co-Scientist," a new multi-agent AI system with the potential to fundamentally transform the process of scientific research. This AI is not just a data analysis tool; it is designed to collaborate with researchers as a "Co-Scientist," autonomously generating, discussing, and evolving new hypotheses for complex scientific problems. It has already achieved remarkable results in the life sciences, such as identifying new drug candidates for acute myeloid leukemia and accelerating infectious disease research, demonstrating its potential to dramatically shorten the cycle of scientific discovery.

How Co-Scientist Works: Collaboration by Multiple AI Agents

At the core of Co-Scientist is a "multi-agent architecture" where multiple AI agents, each with a specialized role, work together. The system uses Google's Gemini as its foundational model and is designed to mimic scientific thought processes. Specifically, a "Generating Agent" creates hypotheses based on literature, a "Reflecting Agent" critically evaluates them as a virtual peer reviewer, a "Ranking Agent" prioritizes promising hypotheses in a tournament format, and an "Evolving Agent" merges and refines promising hypotheses. This cycle is autonomously repeated, allowing the AI to continuously improve the quality of hypotheses from multiple angles without fixating on a single idea.

Technical Details and Specific Achievements

Based on research goals set by researchers in natural language, Co-Scientist autonomously utilizes tools such as web search, literature database referencing, and Python code execution to build and test hypotheses. For example, when a Cambridge University researcher input a research proposal on influenza, Co-Scientist presented promising hypotheses that the researcher had not yet considered. Through subsequent dialogue and the addition of unpublished data, it reportedly led to the identification of specific targets (amino acids) that could potentially shorten an experimental process, which would typically take 2-3 years, to just a few months. This "Human-in-the-loop" approach, where humans and AI collaborate in research, balances both the quality and speed of discovery.

Impact on Engineers and Future Prospects

The emergence of Co-Scientist is a symbolic event, demonstrating AI's advancement into the realm of creative intellectual work. For Japanese engineers, how to utilize or develop such advanced AI agents will be a crucial future theme. Specifically, skills in developing specialized tools and APIs for AI, designing workflows to coordinate multiple agents, and advanced prompt engineering to evaluate the validity of AI-generated hypotheses and direct subsequent actions will become essential. The future where "AI co-scientists" play an active role in all research and development fields, not just life sciences but also material development, software engineering, and even social sciences, is rapidly approaching.

📦
Amazon で関連書籍・ツールを検索
artificial intelligence machine learning LLM book
Amazonで探す →(アソシエイトリンク)
Related articles