Symmetry-Sensitive Analysis of Molecular Graph Neural Network Models.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Molecular interpretability

Background:

• Graph convolutional neural networks (GCNNs) are increasingly used for predicting molecular properties.
• The 'black-box' nature of GCNNs limits their interpretability and adoption in chemistry.
• Understanding model predictions is crucial for scientific discovery and validation.

Purpose of the Study:

• To develop a symmetry-sensitive interpretation method for GCNNs in molecular chemistry.
• To enhance the interpretability of GCNN models by aligning explanations with chemical intuition.
• To provide a computationally efficient tool for understanding GCNN predictions.

Main Methods:

• Introduction of the MolgraphX explainer, a novel method for interpreting GCNNs.
• Focus on highlighting the importance of specific molecular substructures in predictions.
• Validation using diverse datasets of small organic molecules with varying properties.

Main Results:

• MolgraphX effectively highlights key molecular substructures influencing GCNN predictions.
• The method provides explanations consistent with chemical intuition.
• Demonstrated computational efficiency and efficacy across multiple datasets.

Conclusions:

• The proposed MolgraphX method bridges the gap between GCNN accuracy and chemical understanding.
• Offers chemists a valuable tool for interpreting GCNN predictions in molecular chemistry.
• Facilitates deeper insights into chemical mechanisms underlying molecular properties.