Progress Report
Co-evolution of Human and AI-Robots to Expand Science Frontiers[1] Smart and Skilled AI-robot Scientists
Progress until FY2024
1. Outline of the project
Great scientists are intuitive, efficient, and dexterous—able to successfully carry out even highly delicate experiments. This topic focuses on developing AI-robot scientists who embody these qualities and help achieve efficient scientific exploration.
- The AI generates promising hypotheses by integrating knowledge across multiple fields, then further refines those hypotheses through its own interpretation of its experimental results.
- The robot performs precise experiments that even skilled human scientists may find difficult.
- Together, the AI and robot enable efficient discovery by focusing on essential trial and error.
2. Outcome so far
In the life sciences, sample sizes are often limited, making it difficult to formulate strong hypotheses. As a result, scientists sometimes have to rely on guesswork. Furthermore, detailed observations and microscopic manipulations of model organisms (e.g. Xenopus, Arabidopsis, mice) require high skill levels, and even experienced scientists can fail.
To address these challenges, we have developed a system in which: Science AI proposes promising hypotheses; Robots conduct precise experiments; The AI analyzes the data and automatically generates experimental reports; These reports are used to further refine the hypotheses.
We have already developed: Science AI capable of proposing hypotheses based on diverse datasets; Robots that perform precision experiments difficult even for experts; Science AI that analyzes fine structures and generates automated reports.
As a concrete example, we have applied these technologies to develop alternatives to pesticides— “medicines” for plants. In this study, we aim to create compounds that control the opening and closing of stomata to enhance plants’ resistance to environmental stress (Uozumi PI).
Here’s how it works: Science AI analyzes chemical structure data and academic literature to propose candidate compounds (Takeuchi PI, Sasano PI); These compounds are then synthesized (Arisawa PI); Robots automatically evaluate their performance (Arai PI); The results, such as the degree of stomatal opening, are automatically measured and compiled into reports by the Science AI (Mori PI).
We have shown that this system can: Generate hypotheses more efficiently than human scientists; Analyze data at much higher speeds; Achieve experiment quality and speed that surpass both novice and expert scientists.
3. Future plans
At the cellular level, plants and animals share many cellular similarities. Therefore, we are working to apply the same technologies to drug discovery for humans. We have already discovered compounds with better pharmaceutical properties than those previously identified by human scientists (Arisawa PI and others).
We are also developing Science AI capable of generating hypotheses about gene regulatory networks (Kamiya PI, Takebe PI, Maruyama PI, Tanimura PI, and others).
In addition, Science AI is used to explore nutrient combinations in plant culture media. We have discovered new media formulations that promote root growth more effectively than those developed by human scientists (Takeuchi PI, Sato PI, and others).
Through the development of core enabling technologies, we aim not only to advance science but also to promote real-world applications across diverse fields.