Fin-H2AN: Fingerprint-based Heterogeneous Hypergraph Attention Network for Molecular Property Prediction

ICLR 2026 Conference Submission20609 Authors

Keywords: Molecular Property Prediction, Heterogeneous Hypergraph, Hypergraph Learning, Drug discovery

Abstract: Molecular property prediction plays a crucial role in drug discovery and material design but faces major challenges such as constructing meaningful molecular features and defining effective similarity metrics between molecules. Recently, graph neural networks (GNNs) have shown great success in learning molecular representations from graphs via capturing local atomic interactions. However, they often struggle to capture global chemical motifs and complex long-range dependencies between molecules and substructures. In this paper, we present Fin-H2AN, a novel Fingerprint-based Heterogeneous Hyper- graph Attention Network to address the molecular property prediction problem. In this model, we propose a novel heterogeneous hypergraph structure by defining higher-order relations between molecules by leveraging diverse substructures derived from multiple molecular fingerprints. In this way, all molecules and their different fingerprints are embedded into a unified hypergraph. We employ a heterogeneous hypergraph attention model to learn meaningful molecular representations from the proposed heterogeneous hypergraph. It captures higher-order relations among molecules and integrates unique features of different molecular fingerprints into the molecular embeddings. These embeddings are then used for molecular property prediction. Extensive experiments on eight MoleculeNet benchmark datasets demonstrate that Fin-H2AN outperforms the state-of-the-art molecular property prediction models and show its effectiveness in capturing both local and global molecular information.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 20609