Cross-dependent graph neural networks for molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning
• Graph theory

Background:

• Molecular property prediction is crucial for drug discovery and materials science.
• Graph neural networks (GNNs) show promise in learning molecular representations by leveraging graph structures.
• Existing GNNs may not fully exploit the interplay between atomic and bonding information.

Purpose of the Study:

• To develop a novel GNN framework, CD-MVGNN, for enhanced molecular representation learning.
• To simultaneously consider atom (node) and bond (edge) information for more accurate predictions.
• To improve the expressive power and interpretability of GNNs in cheminformatics.

Main Methods:

• Proposed a multi-view GNN (MVGNN) framework with parallel atom-central and bond-central views.
• Introduced a cross-dependent message-passing scheme to facilitate information exchange between views.
• Theoretically justified the model's expressiveness in distinguishing non-isomorphic graphs.

Main Results:

• CD-MVGNN demonstrated significantly superior performance compared to state-of-the-art models on various benchmarks.
• Experimental results validate the model's effectiveness in molecular property prediction.
• Visualization of node importance confirmed the model's interpretability and consistency with chemical knowledge.

Conclusions:

• CD-MVGNN offers an effective approach for molecular representation learning by integrating multi-view information.
• The cross-dependent message-passing mechanism enhances GNNs' ability to capture complex molecular features.
• The model provides interpretable insights into molecular properties, aiding scientific discovery.