When Do Quantum Mechanical Descriptors Help Graph Neural Networks to Predict Chemical Properties?

Area of Science:

• Computational chemistry
• Machine learning
• Materials science

Background:

• Deep graph neural networks (GNNs) predict chemical properties but struggle with extrapolation.
• Quantum mechanical (QM) descriptors can improve GNN generalizability.
• QM calculations are computationally intensive.

Purpose of the Study:

• Investigate the impact of QM descriptors on GNN performance for chemical property prediction.
• Determine when QM descriptors benefit GNNs for molecular property prediction.
• Provide guidelines for integrating QM descriptors into GNN workflows.

Main Methods:

• Systematic analysis of atom, bond, and molecular QM descriptors.
• Evaluation of directed message passing neural networks (D-MPNNs).
• Prediction of 16 molecular properties across various tasks and dataset sizes.

Main Results:

• QM descriptors primarily benefit D-MPNNs on small datasets with high target correlation and accurate computation.
• Using QM descriptors can be costly without benefit or introduce noise, degrading performance.
• Strategic integration offers physics-informed, data-efficient modeling.

Conclusions:

• QM descriptors enhance GNNs when used judiciously, particularly for small datasets.
• Guidelines and tools are provided for effective QM descriptor integration.
• This approach streamlines drug and material design through improved chemical property prediction.