Hokkaido University develops system to integrate chemists' intuition into reaction pathway search algorithms

Computational investigation of chemical reactions is highly complex and has traditionally required trial-and-error calculations for enormous populations of candidates. A research group at Hokkaido University, which includes Specially Appointed Assistant Professor Pinku Nath from the Institute for Chemical Reaction Design and Discovery (ICReDD), Postdoctoral Fellow Yuriko Ono from the Faculty of Science, and Specially Appointed Professor Yu Harabuchi, Specially Appointed Professor Yasunori Yamamoto, Professor Satoshi Maeda, Professor Tetsuya Taketsugu, and Professor Masaharu Yoshioka from ICReDD, has developed a knowledge system called "ChemOntology" which systematizes chemists' experience and judgment criteria to efficiently control the Artificial Force Induced Reaction (AFIR) method. AFIR is a computational chemistry-based reaction pathway search technique. This enables calculations to focus on the most promising reaction candidates, successfully reducing the time and effort required for computation. The system is anticipated to accelerate the discovery and optimization of chemical reactions in various fields, including drug discovery, battery materials, and catalyst development. The research was published online in ACS Catalysis.

Understanding how chemical reactions proceed is difficult through conducting experiments alone. Accordingly, computational chemistry is utilized to predict reaction pathways using computers. The AFIR method developed by Maeda can computationally determine all possible chemical reactions. However, it comprehensively searches through candidates including known reactions and structures that chemists' intuition would deem impossible. In other words, it considers candidates that experimental chemists would judge as "meaningless", resulting in enormous computational time.

In this study, ChemOntology, a knowledge system that organizes chemists' experience and judgment criteria, was combined with the AFIR method.

In ChemOntology, knowledge about elementary reactions such as oxidative addition, used by chemists to understand reactions, is expressed as knowledge for controlling reactions in the AFIR method. Additionally, multiple knowledge engines were developed to perform processes including "structural unit identification," "chemical unit identification," "reaction center identification," and "functional group identification," which experimental chemists use to interpret reaction progression. These engines cooperate to appropriately navigate reaction pathway searches containing suitable combinations of elementary reactions. This works by classifying molecular structures and judging important parts much like chemists analyzing reactions through experiments. As a result, AFIR can preferentially search from "meaningful reaction candidates."

To convert human knowledge into a form digestible by computers, specialized knowledge about chemical reactions was broken down into elements such as "concepts," "classifications," "judgment criteria," and "exceptions," and was organized as a hierarchical rule system. For example, intuitive judgments that chemists naturally make, such as "this bond is easily broken" or "this structure is impossible," were defined as explicit conditional expressions and classification dictionaries; it was implemented as reusable knowledge. This mechanism enables computers to "narrow down candidates while searching," just like chemists do. The system's machines perform the "thinking that chemists do in their minds," with its key feature being that human knowledge is directly integrated within the search algorithm itself, rather than in a separate search engine.

Data-driven research methods using AI technology conventionally require training with large datasets. However, the knowledge-driven ChemOntology can apply the knowledge that chemists use to analyze chemical reactions to the AFIR method, thus eliminating the need for such training. Additionally, the basis for reaction pathway search decisions is clear and based on chemical knowledge, rather than judgments based on pattern recognition by machines. This enables it to explore more complete reaction pathways while reducing computational costs. Although conventional approaches centered on "trial-and-error searching" considering energy conditions, the introduction of ChemOntology enables "rational progression of calculations using chemists' knowledge."

The newly developed system is expected to accelerate various chemical research areas, including the discovery of new drug candidates, exploration of next-generation battery materials, and design of catalytic reactions. In particular, advances in research which integrates experiments, computational science, and information science are anticipated to enable human knowledge and AI to cooperatively contribute to the discovery and optimization of new chemical reactions.

Journal Information
Publication: ACS Catalysis
Title: ChemOntology: A Reusable Explicit Chemical Ontology-Based Method to Expedite Reaction Path Searches
DOI: 10.1021/acscatal.5c06298