CD-MVGNN

CD-MVGNN predicts molecular properties by applying a cross-dependent multi-view graph neural network that integrates atom-centric and bond-centric graph representations.

Key Features:

• Graph Neural Networks (GNNs): Employs graph neural networks to capture and process complex relationships in molecular graphs.
• MVGNN multi-view framework: Implements the MVGNN framework to model multiple graph perspectives for molecules.
• Atom-centric and bond-centric views: Represents both atoms (nodes) and bonds (edges) as distinct but complementary views to incorporate atomic and bonding information.
• Cross-dependent message-passing: Uses a cross-dependent message-passing scheme to enable information exchange between atom and bond views.
• Expressiveness for non-isomorphic graphs: Provides a theoretical foundation that enables distinction of non-isomorphic molecular graphs.
• Interpretability via node importance visualization: Produces node importance metrics and visualizations that align with established scientific knowledge.
• Benchmark performance: Validated with extensive experiments on challenging benchmark datasets and reported to outperform state-of-the-art methods.

Scientific Applications:

• Molecular property prediction: Predicts physicochemical or biological properties of molecules from graph representations.
• Discrimination of structurally similar molecules: Distinguishes molecules with similar structures but distinct properties by identifying non-isomorphic graph differences.
• Interpretation of predictive features: Identifies important atoms or bonds via node importance metrics to support mechanistic interpretation.
• Computational chemistry and bioinformatics tasks: Applies to tasks in computational chemistry and bioinformatics that require accurate molecular representation and prediction.

Methodology:

Implements the MVGNN multi-view GNN framework with atom-centric and bond-centric perspectives, employs a cross-dependent message-passing scheme to learn molecular representations via GNNs, visualizes node importance metrics, and evaluates performance on benchmark datasets.