Ono has introduced advanced methods that combine AI and knowledge graphs aimed at forecasting and hypothesis generation in drug discovery. By structuring the relationship between genes, disease, and pharmaceuticals, we thereby extract useful knowledge. We use Tokyo-1 to handle extensive computation, achieving sophisticated analysis in a realistic time frame. Based on the results derived from collaboration with Healx and the use of open source, we are advancing construction of an independent model. In addition, we have introduced explainable artificial intelligence (XAI) and large-scale language models, working to integrate in-house data, with the aim of improving both the quality and speed of drug discovery.

In the Tokyo-1 sub-working group, multiple pharmaceutical manufacturers are engaged in ultra-high-speed structure-based virtual screening (SBVS) using GPU. Through the development of small-molecule compound preliminary processing programs and high-speed tools, we are achieving processing speeds over 100 times faster than conventional methods. Each company has reported results, such as the complete analysis of seven million compounds in just a few days. At present, we are aiming to construct processes to enable screening at a challenging scale of 100 billion compounds in just days, in collaboration with AI.

We are working to improve antibody parameter forecasting and activity using the various publicly available protein and antibody language models. By conducting extensive calculations using Tokyo-1, we identified antibodies with up to 30 times improved binding activity from among the tens of antibodies planned for testing. Currently, we are training foundational protein language models on antibody data with the goal of constructing a model capable of making more precise predictions.

As part of our structure-based approach to drug discovery, we obtain three-dimensional protein structures using X-ray crystallography, cryo-electron microscopy, and AI technology. We conduct molecular dynamics calculations based on the structural information obtained, allowing for detailed analysis of the proteins and small-molecule compounds binding mode and interaction. In addition, we calculate the free energies of binding through free-energy perturbation plus (FEP+) for precision forecasting of compound binding strength. We use these results for the design of new compounds and optimization of candidates, striving for greater efficiency in drug discovery research.

We are driving “social drug discovery” through the collaboration of medicinal chemists and AI. By fusing the intelligence of medicinal chemists with that of AI, we aim to resolve issues in molecular design. Using the Tokyo-1 supercomputer, we are conducting reinforced learning of AI with feedback from medicinal chemists. Presently, we are proceeding with the development of a medicinal chemist AI agent and interaction tests between medicinal chemists and AI agents, accelerating implementation of a collaborative system whereby humans and AI advance together.

We evaluate compound activity and physicochemical properties using visualized maps of compound design and synthesis. By utilizing AI forecasting to strengthen the design section of our Design - Make - Test - Analyze (DMTA) cycle in particular, we eliminate wasteful compositions in an effort to shorten the time frame. We use key evaluation data in the next stage of design and have introduced an automated synthesis machine for efficient synthesis. Through integration of computational science with experimental work, we aim to optimize the DMTA cycle and further accelerate drug discovery.