The paper titled "CHEM: Causally and Hierarchically Explaining Molecules" by the joint research team of Professor Jong-June Jeon's laboratory (Kyungdong Woo, Soyong Cho, ChangHyun Kim) and Professor Jinhee Choi's laboratory (Donghyeon Kim, and Kimoon Na) at the University of Seoul has been accepted to the 2025 Conference on Information and Knowledge Management (CIKM, 2025), one of the top international conferences in computer science.

CIKM is a prestigious international conference organized by the Association for Computing Machinery (ACM) and is highly competitive, having an acceptance rate of less than 20-30%. In computer science, acceptance at a top conference is more difficult than publication in a top journal.e

▶ Process of incorporating prior molecular information into the neural network model

In this study, the joint research team developed a Graph Neural Network (GNN) model that can identify and predict causal subgraphs that determine the properties of molecules. Although existing GNN models rely solely on data information to identify the important molecular substructures, this study employs prior knowledge of functional groups in molecules to improve both the prediction performance and explainability of the neural network. This is a major achievement where actual chemical structure information in the training phase is reflected so that complex neural network results are presented in a form that can be intuitively understood by chemical experts. This achievement is expected to have broad applications in areas such as chemical hazard assessment, drug development, and chemical reactivity analysis. 

This research outcome is the result of interdisciplinary research conducted as part of the "Development of Environmental Disease Prediction Model Based on Molecular Toxicity Network" project funded by the Ministry of Environment. Further, it is significant as it was a collaboration between environmental toxicity experts and data science and artificial intelligence scholars. The research team stated, "This study is significant in that it simultaneously addresses the issues of explainability and generalization performance by learning causal substructures within molecular structures," and added, "It is important to note that the causal subgraphs were validated on real molecular data using chemical domain knowledge."