CatDRX is a generative AI framework developed at Institute of Science Tokyo, which enables the design of new chemical catalysts based on the specific chemical reactions in which they are used. The model learns from large reaction datasets and predicts how well a catalyst will perform, while also proposing new catalyst structures. Validated across various reaction types, CatDRX offers a promising strategy to accelerate catalyst discovery for a wide range of chemical and industrial processes.
Catalysts are chemical compounds that speed up chemical reactions. These are used everywhere, from pharmaceutical production to material science and green chemistry. However, designing a new catalyst is a slow process and often requires long cycles of experimental trial and error. While computational simulations are being used, current methods are limited to only specific reaction types, restricting their applications.
To overcome these limitations, a research team from Institute of Science Tokyo (Science Tokyo), Japan, developed Catalyst Discovery framework based on a ReaXion-conditioned variational autoencoder (CatDRX), a new generative artificial intelligence (AI) framework that designs catalysts under real reaction conditions. The study was led by Associate Professor Masahito Ohue, along with graduate student Apakorn Kengkanna and Specially Appointed Assistant Professor Yuta Kikuchi from the Department of Computer Science, School of Computing at Science Tokyo, and Professor Takashi Niwa from the Graduate School of Pharmaceutical Sciences, Kyushu University, Japan. Their findings were published in Volume 8 of the journal Communications Chemistry on October 23, 2025.
"What makes CatDRX different is how it learns. Instead of only focusing on the catalyst structure, it also considers the entire reaction environment," explains lead author, Ohue.
The model leverages a conditional variational autoencoder architecture that learns from both catalyst structures and reaction conditions, including reactants, product, reagents, and reaction time. This rich context allows the model to predict catalyst performance, such as percent yield or selectivity, as well as propose entirely new catalyst structures that are tailored to a given chemical reaction.
To begin with, the researchers trained CatDRX on a large public reaction database and fine-tuned it using downstream catalytic datasets. This enabled the model to learn general chemical patterns while adapting to specific reaction types. To analyze the performance of the model, CatDRX was evaluated across several real-world case studies, including reactions like the Suzuki-Miyaura cross-coupling and the asymmetric Pictet-Spengler reactions—two widely used chemical reactions in organic synthesis.
The results were impressive. CatDRX showed strong performance in each case and accurately predicted catalytic activity, generating catalyst candidates that resembled known successful examples. It also suggested new catalyst structures worth testing in laboratory experiments. Using standard computational chemistry tools such as the density functional theory, the team confirmed some of these candidates, demonstrating that the model's predictions aligned with known mechanistic behavior.
"One key strength of CatDRX is its flexibility," comments Ohue. "You can either ask the model to search for entirely new catalysts or explore small structural variations around known catalysts for improved performance."
What makes the model special is its intrinsic ability to capture the complex trade-offs in catalyst behavior. Designing catalysts often requires balancing activity, selectivity, and practical constraints. CatDRX brings these factors together by linking reaction conditions directly to performance. Additionally, the ability of the model to navigate chemical space both globally and locally with optimization technique significantly shortens the time taken for the identification of potential catalysts.
Looking ahead, the researchers aim to further improve the model by including data such as adding temperature conditions and reaction concentrations. They also plan to expand the catalyst libraries and add further refinements in the generative approach so that it suggests structures that are easier to synthesize. While these improvements hold promise, the current model of CatDRX is a powerful example of how AI can support sustainable chemistry, helping researchers design catalysts in a more efficient and eco-friendly manner.
