Explainable graph neural networks for organic cages.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Porous organic cages are versatile materials with applications in separations, catalysis, and encapsulation.
• Shape persistence, a permanent internal cavity, is crucial for most porous organic cage applications.
• Predicting shape persistence in complex organic cage structures is challenging.

Purpose of the Study:

• To develop accurate and interpretable machine learning models for predicting the shape persistence of porous organic cages.
• To leverage Graph Neural Networks (GNNs) for property prediction in topologically complex molecular systems.
• To provide structural insights for designing novel porous organic cages with desired properties.

Main Methods:

• Development and application of Graph Neural Networks (GNNs) to represent organic cages as graphs.
• Training and validation of GNN models using a computational database of porous organic cages.
• Utilizing integrated gradients to analyze the contribution of molecular components to shape persistence.

Main Results:

• GNN models demonstrated improved prediction accuracy and transferability compared to random forest models.
• Integrated gradients enabled quantitative evaluation of monomer and fragment contributions to shape persistence.
• The explicability of GNNs allowed for interpretation of predictions and identification of key structural features.

Conclusions:

• Graph Neural Networks offer a powerful and interpretable approach for predicting the shape persistence of porous organic cages.
• The developed GNNs provide valuable insights into structure-property relationships, guiding the design of new materials.
• This work advances the application of explainable AI in materials science for accelerated discovery.