World model for autonomous scientific discovery in biomedical science

Abstract

Our goal is to establish an AI that can autonomously make scientific discoveries in the field of biology and medicine. In the field of biology and medicine, it is extremely difficult to determine the first principle due to the extreme complexity of the research subject, and a data-driven approach is taken, but the cost of data generation is also high. Therefore, we examined the potential of reinforcement learning based on the world model in the field of biology and medicine as a means of efficiently acquiring models from complex objects. First, we created a simulation environment that takes into account various parameters of the neuronal cell culture process and used it to test whether reinforcement learning based on the world model can autonomously discover from scratch, without prior knowledge, the optimal culture method for efficiently differentiating neuronal cells. As a result, we confirmed that the world model-based reinforcement learning "Dreamer v3," which has high image reconstruction capability, can autonomously discover the culture conditions after experiencing about 10 cell culture experiments.