Drug discovery traditionally takes over 10 years and costs hundreds of billions of yen, with only about one in tens of thousands of candidate compounds ultimately approved as a new drug. Now, a new approach combining generative AI with computer simulation is poised to transform this process by automating candidate screening and enabling precise molecular predictions in a fraction of the time.

Why Is This Hybrid AI-Simulation Approach Different from Traditional Drug Discovery?

The breakthrough lies in deploying each technology where it performs best. Generative AI rapidly screens thousands of candidate compounds by learning patterns from historical chemical data, while computer simulation then validates the most promising candidates through detailed molecular analysis based on physical chemistry principles. This two-stage process dramatically reduces trial-and-error experiments that would otherwise require years of laboratory work.

Associate Professor Masahito Ohue of the School of Computing explains the strategic advantage of this hybrid approach. "AI, which learns patterns from large volumes of data and automates prediction and decision-making, is effective for discovering and screening candidate compounds. Computer simulation, which reproduces and predicts phenomena on the basis of theory and mathematical formulas, is then essential for precise evaluation after the candidates have been narrowed down," he stated.

What Are the Key Technologies Enabling This Shift?

• Generative AI Models: These systems learn patterns from large volumes of chemical and biological data, automating the discovery and initial screening of candidate compounds without requiring researchers to manually test each one.
• Molecular Simulation Techniques: Methods like free energy perturbation provide theoretical predictions based on physical chemistry principles, enabling precise evaluation of how drug candidates will behave at the molecular level.
• Protein Language Models: Specialized AI models trained on protein sequences and structures, applying the same transformer-based reasoning used in large language models to biological data instead of natural language.
• Compound Language Models: Similar to protein models but focused on small molecules and drug candidates, enabling AI to reason about chemical properties and drug-like characteristics.
• Genome Language Models: AI systems trained on genetic data to identify disease targets and predict how compounds will interact with specific biological pathways.

How Do Recent Breakthroughs Like PairMap and CatDRX Work?

One concrete example is PairMap, a method released in January 2025 that uses a technique called free energy perturbation (FEP) to predict how strongly drug compounds bind to disease-causing proteins. The innovation addresses a critical limitation: conventional FEP struggles when comparing compounds with very different molecular structures. PairMap solves this by automatically generating intermediate compounds between two candidates, enabling step-by-step calculations that maintain accuracy even across structurally diverse options.

Another breakthrough came in October 2025 with CatDRX, a generative AI model that proposes catalysts suited to specific chemical reactions. By training on past experimental data, CatDRX helps researchers identify the right catalysts for organic synthesis, a process that previously relied heavily on trial and error. The model was developed through joint research with organic synthesis specialists who helped Ohue understand the practical challenges researchers face.

"Information science has countless applications. This joint research led me to think constantly about what challenges exist in which fields, and how information science can be used to solve them," said Masahito Ohue.

Masahito Ohue, Associate Professor at the School of Computing

What Impact Could This Have on Drug Development?

The practical implications are substantial. By reducing the number of compounds requiring expensive laboratory testing, this approach shortens development timelines and lowers costs. For diseases like influenza, where researchers need to predict how strongly a drug binds to viral proteins, the accuracy gains translate directly into faster identification of viable treatments. The binding strength between a protein and a compound is expressed as a value called binding free energy, and predicting this accurately is essential for narrowing down candidate compounds efficiently.

The vision extends beyond incremental improvements. Ohue envisions a future where generative AI serves as a capable research assistant that can respond instantly to emerging threats. "For example, the goal is generative AI for drug discovery that, when a new, unknown virus emerges, could respond instantly if asked, 'I want to quickly develop a drug effective against this virus,'" he explained.

How Do Recent AI Breakthroughs Support This Vision?

The 2024 Nobel Prize in Chemistry recognized computational protein design and structure prediction, with AlphaFold 2 playing a central role. The subsequent AlphaFold 3 expanded capabilities to predict structures of proteins, nucleic acids like DNA and RNA, and small molecules that could serve as drug candidates. However, Ohue noted that AlphaFold 3 currently achieves only 60 to 70 percent accuracy, indicating substantial room for improvement before it becomes a practical tool for drug discovery.

Large language models (LLMs) have also advanced remarkably, and researchers are now applying the same principles used by LLMs trained on natural language to biological and chemical data. By changing what the models are trained on, it is possible to build protein language models, compound language models, genome language models, and others. Results that could lead to generative AI for drug discovery are gradually emerging, with these technologies expected to prove valuable in the coming years.

Steps to Implement AI-Assisted Drug Discovery in Research Programs

• Identify Target Proteins: Begin by determining which disease-causing proteins or pathogens your research should focus on, then use AI models to predict which compounds might bind effectively to those targets.
• Screen Candidates with Generative AI: Deploy generative AI models trained on historical chemical reaction data to rapidly propose and rank thousands of candidate compounds, dramatically reducing the initial pool for laboratory testing.
• Validate with Molecular Simulation: Use computer simulation techniques like free energy perturbation to predict binding strength and molecular behavior for the most promising candidates before committing resources to expensive experiments.
• Collaborate Across Disciplines: Partner with domain experts in organic synthesis, biochemistry, and clinical research to ensure AI-generated predictions align with practical constraints and regulatory requirements.
• Iterate and Refine: Feed experimental results back into the AI models to continuously improve prediction accuracy and expand the scope of compounds and reactions the system can handle.

What Challenges Remain Before This Becomes Standard Practice?

Despite the progress, significant hurdles persist. Current generative AI models for drug discovery work best with organic reactions and compounds where sufficient experimental data has accumulated. Expanding these approaches to inorganic materials and less-studied reaction types requires more training data. Additionally, the integration of AI predictions with regulatory requirements and clinical trial protocols remains an ongoing challenge.

The convergence of generative AI, molecular simulation, and specialized language models represents a fundamental shift in how researchers approach drug discovery. Rather than replacing human expertise, these tools amplify researcher capabilities, enabling faster identification of promising compounds and more efficient allocation of laboratory resources. As these technologies mature, the timeline from disease identification to viable drug candidate could shrink from over a decade to months, potentially transforming how medicine responds to emerging health threats.