HYG-mol: An Interpretable Multimodal Hypergraph Framework for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Accurate molecular property prediction is crucial for drug discovery.
• Existing methods like graph neural networks and language models have limitations in capturing complex molecular structures and semantics.
• There's a need for methods that integrate both higher-order structural information and chemically grounded semantic understanding.

Purpose of the Study:

• To develop an interpretable molecular property prediction framework, HYG-mol.
• To address limitations of existing methods by integrating hypergraph-based structural modeling with multimodal chemical semantics.
• To improve the accuracy and interpretability of molecular property predictions.

Main Methods:

• Representing molecules as hypergraphs, encoding substructures (functional groups, rings) as hyperedges.
• Fusing chemical semantic information from pretrained language models with atomic-level physicochemical descriptors.
• Employing a hypergraph attention network to model cross-scale interactions and identify relevant substructures.

Main Results:

• HYG-mol consistently outperforms state-of-the-art methods on MoleculeNet benchmark datasets for both classification and regression tasks.
• Ablation studies confirm the effectiveness of the proposed hypergraph representation and semantic integration.
• Interpretability analysis shows strong correspondence between model-identified substructures and known chemically meaningful motifs.

Conclusions:

• HYG-mol offers a unified and interpretable framework for molecular property prediction.
• Explicitly grounding chemical semantics in higher-order structural representations enhances predictive performance.
• The method advances the field by providing a more comprehensive approach to structure-property relationships in drug discovery.