Graph Neural Network-Based Molecular Property Prediction with Patch Aggregation.

Area of Science:

• Computational chemistry
• Machine learning
• Quantum mechanics

Background:

• Graph neural networks (GNNs) are effective for molecular property prediction.
• Current GNN aggregation methods increase parameters and computational cost without guaranteed accuracy improvements.

Purpose of the Study:

• To introduce a novel, parameter-efficient edge-to-node aggregation mechanism for GNNs.
• To enhance the accuracy and efficiency of GNNs in predicting molecular properties.

Main Methods:

• Developed "patch aggregation," inspired by Multi-Head Attention and Mixture of Experts.
• Integrated patch aggregation into state-of-the-art GNN models (SchNet, DimeNet++, SphereNet, TensorNet, VisNet).
• Compared patch aggregation against existing methods (sum, MLP, softmax, set transformer).

Main Results:

• Patch aggregation consistently outperformed existing aggregation techniques in predicting QM9 thermodynamic properties and MD17 energies/forces.
• The method demonstrated improved prediction accuracy and parameter efficiency.
• Patch aggregation proved applicable across diverse GNN architectures.

Conclusions:

• Patch aggregation is a superior edge-to-node mechanism for GNNs in molecular property prediction.
• It offers enhanced accuracy and computational efficiency, suitable for resource-limited applications.
• This method represents a significant advancement for GNNs in computational chemistry.