Problem
Research questions and friction points this paper is trying to address.
Molecular Machine Learning
Self-supervised Learning
Molecular Structure Similarity
Multi-channel learning for integrating structural hierarchies into context-dependent molecular representation
🤖 AI Summary
Existing self-supervised methods struggle to jointly capture multi-level dynamic correlations—structural similarity, scaffold composition, and local chemical environments—due to molecular data scarcity and the inherent difficulty of modeling property–environment coupling. To address this, we propose the Multi-channel Hierarchical Graph Neural Network (MH-GNN), the first model to explicitly encode atoms, bonds, functional groups, and topological substructures as parallel, hierarchical channels. We design a hierarchical gated attention mechanism for cross-scale adaptive fusion and introduce conformation-aware message passing to overcome representational limitations of single-embedding paradigms. Evaluated on 12 molecular property prediction and reactivity tasks, MH-GNN achieves state-of-the-art performance, reduces cross-dataset generalization error by 19.7%, and significantly enhances modeling capability for stereoelectronic effects and long-range substituent influences.
Innovation
Methods, ideas, or system contributions that make the work stand out.
structural hierarchy
multi-channel integration
enhanced property prediction
🔎 Similar Papers
No similar papers found.
