EquiHGNN: Scalable Rotationally Equivariant Hypergraph Neural Networks

Molecular interactions often involve high-order relationships beyond pairwise connections. Hypergraphs enable multi-way interactions, making them well-suited for complex molecular systems. We introduce EquiHGNN, an Equivariant HyperGraph Neural Network that integrates symmetry-aware representations for molecular modeling. By enforcing the equivariance under transformation groups, EquiHGNN preserves geometric and topological properties, yielding more robust and physically meaningful features. Experiments on small and large molecules show that while high-order interactions add little for small molecules, they consistently outperform 2D graphs on larger ones. Incorporating geometric features further boosts accuracy, underscoring the importance of spatial information in molecular learning. Code: https://github.com/HySonLab/EquiHGNN/.