Self-explainable Molecular Property Prediction via Multi-view Hypergraph Learning

Xudong Chen, Yixin Liu, Xiaoxiao Ma, Kaize Ding

Keywords: Hypergraph, Explainable AI, Molecular Properties Prediction

Abstract: Explainable molecular property prediction plays a critical role in drug discovery and materials science. Existing graph neural network (GNN)-based approaches usually rely on pairwise atomic interactions for molecular modeling and interpretation. However, such atom-centered modeling often neglects the cooperative effects of atomic groups and lacks the guidance of chemical rules, thereby limiting both prediction accuracy and interpretability. To address these challenges, we propose a multi-view **Hyper**graph learning method for **S**elf-**E**xplainable **M**olecular property prediction(HyperSEM). Our method introduces a hyperedge-driven explanation paradigm, where atomic groups are explicitly modeled as hyperedges to capture high-order cooperative effects, and multi-view hypergraphs are constructed to jointly integrate chemical rules and data-driven signals. Furthermore, we design a molecular structure-informed hypergraph convolution to preserve both high-order atomic-group interactions and low-order structural features, and an information-bottleneck-guided self explanation to jointly generate predictions and explanations. Extensive experimental results show that HyperSEM outperforms existing state-of-the-art methods on seven benchmark datasets, demonstrating dual advantages in prediction accuracy and interpretability.

Primary Area: interpretability and explainable AI

Submission Number: 10470