Astellas Deploys Structure Prediction AI "Boltz-2 NIM" to Democratize and Accelerate Drug Discovery

Astellas Pharma is a global life sciences company focusing on turning innovative science into treatments for patients. Its therapies in oncology, ophthalmology, urology, women’s health, and immunology address conditions ranging from today’s most challenging cancers to helping people with chronic conditions live with fewer limits to reimagining life after organ transplantation.    

To reach more patients, Astellas is working to deliver transformative medicines by pioneering science in breakthrough areas of R&D through strategic use of advanced technologies and AI—with the goal of bringing innovative medicines to patients faster.

As part of this effort, Astellas participated in the Tokyo-1 Project, established by Xeureka Inc. (a subsidiary of Mitsui & Co., Ltd.) to build an innovation hub in the Japanese healthcare industry. Currently, Astellas actively utilizes several NVIDIA resources, including the DGX H100/H200, and developed "astABpLM” in 2024 to improve screening efficiency in antibody research. "astABpLM” is a proprietary language model specialized for predicting antibody physical properties using the NVIDIA BioNeMo™ Framework.

Astellas has also improved efficiency by introducing AI and simulation technologies into research for small- and medium-molecule drugs such as PROTACs. Effective drug discovery research for complex molecules relies on collaboration between computational chemists, who handle protein structure prediction and theoretical prediction of candidate compounds, and medicinal chemists, who are responsible for compound design and synthesis. Traditionally, this research has required computational chemists to provide a high volume of experimental compounds, which can create bottlenecks and slow the drug discovery process.

To address this challenge and streamline collaboration, Astellas is leveraging AI to "democratize" structure-based drug design (SBDD). Kazuya Nagaoka, who promotes AI and simulation technologies in compound-related modality research, explains the reason: "Until now, simulations were performed exclusively by computational chemists. However, with the emergence of Boltz-2 NIM, we believed that medicinal chemists themselves could proceed with compound design and hypothesis verification based on structures without having to request help from computational chemists."

"Until now, simulations using protein-compound complex structures required for SBDD were performed exclusively by computational chemists. However, by utilizing the revolutionary AI model Boltz-2 NIM, we believed that medicinal chemists themselves could proceed with structure-based compound design and hypothesis verification.

Kazuya Nagaoka
Associate Director Astellas Pharma Inc. DigitalX, R&D Digital, Modality Informatics Team Lead of Digital Chemical Modality Discovery

Astellas Pharma adopted NVIDIA’s "Boltz-2 NIM" to promote democratization, utilizing the Tokyo-1 drug discovery support platform for the execution environment. Boltz-2 NIM is a microservice developed and provided by NVIDIA based on Boltz-2, a biomolecular model jointly developed by the Massachusetts Institute of Technology (MIT) and the biotech company Recursion Pharmaceuticals. It has the following features:

• Unlike previous biomolecular models, the original Boltz-2 can predict compound binding affinity in addition to protein structure, making it suitable for use by both computational chemists and medicinal chemists.
• As the name NIM (NVIDIA inference microservices) suggests, it is provided as a microservice, requiring minimal programming effort. It can be operated directly via commands or within a Docker environment.
• Because it is optimized for NVIDIA GPUs using the NVIDIA TensorRT optimization tool, it can perform predictions faster on Tokyo-1 compared to the original Boltz-2 (Python-based) provided through GitHub.

Astellas developed a browser-based front end to promote internal use. Keigo Ide, who is responsible for drug discovery AI research and solution development, explains: "Ease of use is essential for driving democratization across the organization. We had experience with Docker-based development, and the NVIDIA Digital Biology Team supported us in optimizing Boltz-2 NIM for our environment, so implementation went smoothly."

One of the users is medicinal chemist Yosuke Matsumura. "Binding affinity prediction simulations were previously far too complex—an area entirely out of reach for medicinal chemists. Now, with the combination of Boltz-2 NIM and the front end developed by Ide, results can be obtained simply by entering information and clicking an execution button. While we still request complex simulations from Nagaoka, we can execute simple simulations immediately and apply the results to our next ideas. From a medicinal chemist's perspective, this is a truly transformative change."

"With the introduction of Boltz-2 NIM, simple simulations can be executed immediately, and the results can be applied to the next idea. This is a game-changing development for medicinal chemists."

Yosuke Matsumura
Senior Researcher, Ph.D. in Pharmaceutical Sciences Astellas Pharma Inc. Innovation Lab Innovation Core Lab/Core Discovery Unit

Kazuya explains the effects of introducing Boltz-2 NIM: "We applied it to a drug discovery project targeting membrane proteins and believe we succeeded in taking the quality and speed of therapeutically relevant compound design proposals to the next level."

Furthermore, medicinal chemist Dr. Matsumura states: "Based on hypotheses derived from Boltz-2 NIM, we selected 10 compounds from a library and conducted assays. The results showed that all 10 compounds exhibited the expected medicinal effects. Initially, we anticipated needing many more assays, but by narrowing them down with Boltz-2 NIM, we achieved a reduction of more than 90% in evaluated compounds and more than 70% in the research period."

Regarding performance, Dr. Ide provides the following evaluation: "When applied to complex modeling and activity prediction for a certain target, the open source Boltz-2 written in Python took over 20 seconds. However, using Boltz-2 NIM optimized with TensorRT, results were obtained in just 7 seconds. We consider this a significant advantage of using NIM."

Toshiyuki Ohfusa, team lead of lab automation at Astellas, is providing applications that are easier for medicinal chemists to use. "Medicinal chemists prefer handling visually intuitive structural formula data, but Boltz-2 NIM does not support that. Therefore, we developed an application with functions to draw compounds as structural formulas and convert them into SMILES notation. We also implemented features such as the automatic sequential execution of multiple compound combinations for specific proteins to shorten the DMTA (Design-Make-Test-Analyze) cycle." This application will be rolled out across the company.  

The browser-based application developed by Dr. Ide is already provided to various departments globally, and its use is spreading. "As of the end of February 2026, the application has more than 50 regular users, and the total number of users has exceeded 2,000."

Astellas Pharma is expanding its use of NIM. In compound modality drug discovery, they plan to deploy NIM such as DiffDock, GenMol, and MolMIM on Tokyo-1 and combine them with Boltz-2 NIM to further promote democratization. Expansion to other modalities has already occurred. To enable consistent execution from structure and sequence design to binding prediction for proteins and peptides, they implemented a "De Novo Protein/Peptide Design Workflow" on Tokyo-1 using "RFdiffusion NIM" for 3D structures, "ProteinMPNN NIM" for sequence design, and Boltz-2 NIM.

Kazuya evaluates the benefits of NIM: "NIM, which is performance-optimized and ready to use as a container, is a groundbreaking mechanism. With the success of Boltz-2 NIM, we want to actively utilize other NIM in the future and look forward to NVIDIA's further expansion of the lineup."

Against the backdrop of diversifying drug discovery modalities, the ability to utilize AI and high-performance computing resources determines the competitiveness of pharmaceutical companies. With Tokyo-1 as its computational foundation, Astellas Pharma's efforts to democratize drug discovery are well worth watching.

"A process that takes over 20 seconds with open -source Boltz-2 takes only 7 seconds with Boltz-2 NIM optimized by TensorRT. This is a major benefit of using NIM."

Keigo Ide
Ph.D. in Engineering Astellas Pharma Inc. DigitalX, R&D Digital, Modality Informatics Team Lead of Digital Biological Modality Discovery

"To make Boltz-2 NIM easier to use, we developed an application with structural formula drawing functions and automatic execution for multiple protein-compound combinations."

Toshiyuki Ohfusa
Ph.D. in Pharmaceutical Sciences Astellas Pharma Inc. DigitalX, R&D Digital, Modality Informatics Team Lead of Lab-Automation

“We are expanding NIM deployment on Tokyo-1 to democratize drug discovery across multiple modalities. For compound discovery, they combine DiffDock, GenMol, and MolMIM with Boltz-2 NIM. For proteins and peptides, they utilize an end-to-end 'De Novo Protein/Peptide Design Workflow' integrating RFdiffusion, ProteinMPNN, and Boltz-2 NIMs”