Associate Professor Masahito Ohue, Graduate Student Apakorn Kengkanna, and Specially Appointed Junior Associate Professor Yuta Kikuchi of the Department of Computer Science, School of Computing at the Institute of Science Tokyo and Professor Takashi Niwa of the Graduate School of Pharmaceutical Sciences at Kyushu University have developed a new generative AI method that enables efficient catalyst design. This pre-trained generative model, called CatDRX, extracts and learns features individually from catalyst structures and proposes catalysts that can accommodate specified reaction conditions. It makes it possible to propose catalyst structures that promote arbitrary chemical reactions while simultaneously predicting their catalytic activity. The work was published online in Communications Chemistry. CatDRX is available for download from the program-sharing site GitHub as open-source software (https://github.com/ohuelab/CatDRX).
CatDRX consists of three modules: a catalyst embedding module that extracts feature vectors from matrices representing catalyst structures using neural networks; a condition embedding module that represents chemical structures as molecular graphs and extracts feature vectors using graph neural networks, while treating other reaction conditions as vectors combining direct numerical features and one-hot vectors; and an autoencoder module that can propose catalysts suitable for specified reaction conditions by learning these feature vectors in an integrated manner.
When the performance of CatDRX was evaluated in catalyst proposal and catalytic activity prediction tasks, it showed stable and good performance across a wide range of reaction conditions. When four different reaction conditions were input to propose catalysts, it was shown that the system could propose diverse structures while retaining the general shapes of catalysts commonly seen in each reaction. Additionally, by choosing whether to perform sampling from the latent space from distributions close to existing catalysts or completely randomly, it became possible to adjust the balance between the validity of proposed catalysts and the scope of exploration.
Furthermore, in comparisons of catalytic activity prediction across eight different chemical reactions, CatDRX demonstrated performance equal to or better than existing methods. On the other hand, challenges were observed in cases such as C-C cross-coupling, where the chemical reactions and catalysts to be predicted deviated from the training data, resulting in insufficient prediction performance and reduced accuracy. This is a common issue in AI prediction, and it is said that expanding the data used for learning and the features to be learned would increase the applicability to a wider range of reaction systems.
The research group aims to build a more general-purpose model capable of handling more diverse reactions and catalysts by expanding the dataset used for pre-training. They are also considering further improvements, as new challenges have been observed in some areas, such as catalytic activity prediction and the feasibility of proposed catalysts.
This achievement is a cross-disciplinary research outcome accomplished through the collaborative efforts of Ohue, who specializes in information science, and Niwa, who specializes in organic synthetic chemistry. It is an excellent example demonstrating the importance of integrating different areas of expertise in today's research fields, which are becoming increasingly complex and diverse.
Journal Information
Publication: Communications Chemistry
Title: Reaction-conditioned generative model for catalyst design and optimization with CatDRX
DOI: 10.1038/s42004-025-01732-7
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.