HyperMolFusion: A Hypergraph-Enhanced Multi-Modal Fusion Framework for Accurate Molecular Property Prediction

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Cheminformatics

Background:

• Deep learning methods for molecular property prediction often struggle with integrating multi-source, heterogeneous chemical data.
• Existing models fail to holistically represent molecular structures and capture high-order synergistic interactions.

Purpose of the Study:

• To introduce HyperMolFusion, a hypergraph-enhanced multi-modal fusion model for improved molecular property prediction.
• To address limitations in integrating atomic, fingerprint, and motif-level chemical information.

Main Methods:

• Developed HyperMolFusion using a hypergraph framework to model chemical motifs as hyperedges, capturing high-order correlations.
• Incorporated AtomConv for local atomic interactions, HyperConv for motif-level analysis, and a mixed molecular fingerprint module.
• Utilized a chemically guided attention (CGA) mechanism for dynamic fusion of multi-level features into hierarchical representations.

Main Results:

• HyperMolFusion demonstrated promising performance across eight MoleculeNet benchmarks for both regression and classification tasks.
• Achieved notable results including RMSE of 0.611 for lipophilicity and ROC-AUC of 0.935 for ClinTox.
• Effectively alleviated over-smoothing and preserved structural information through hierarchical representation learning.

Conclusions:

• HyperMolFusion offers a systematic and effective solution for molecular property prediction by integrating diverse chemical information via hypergraph modeling.
• The model provides a more reliable computational tool to enhance the efficiency and accuracy of drug development pipelines.
• This work highlights the potential of hypergraph neural networks in advancing cheminformatics and accelerating drug discovery.