MM-HNN: A MoleculeNet-Benchmarked Molecular Hypergraph Neural Network for Molecular Property Prediction
Bingsai Wang, Guifei Zhou, Chunyan LiIntroduction:
Accurately predicting molecular properties is a key task in computational chemistry and drug discovery. Traditional molecular graph representations often struggle to capture higher-order connections, limiting their ability to model complex molecular structures.
Methods:
The Molecular Hypergraph Neural Network (MHNN) leverages hypergraphs to model higher-order relationships between atoms, such as conjugated structures. Inspired by the MHNN framework, we propose a MoleculeNet-benchmarked Molecular Hypergraph Neural Network (MMHNN), which extends the hypergraph-based molecular representation paradigm to large-scale benchmark molecular property prediction tasks.
Results:
MM-HNN is systematically evaluated on ESOL, FreeSolv, Lipophilicity (regression), and HIV, BACE, BBBP, and Tox21 (classification). MM-HNN achieves the best performance on Free- Solv and Tox21, while demonstrating highly competitive results across the remaining benchmarks.
Discussion:
The stronger performance on FreeSolv and Tox21 suggests that hypergraph representations excel at molecular property prediction. Ablation analysis shows that incorporating Higher-Order Connections (HOCs), i.e., multi-atom hyperedges representing conjugated structures, yields numerical performance improvements across all seven datasets, with statistically significant improvements observed on Lipophilicity, BACE, and BBBP, suggesting task-dependent benefits of higher-order topology.
Conclusion:
This study demonstrates that hypergraph-based architectures are a powerful tool for molecular property prediction, but are not universally superior to pretrained graph-based models. The findings underscore the need to align representation choices with target property characteristics.