Hierarchical Graph Attention Network with Positive and Negative Attentions for Improved Interpretability: ISA-PN

Area of Science:

• Computational chemistry
• Materials science
• Artificial intelligence in chemistry

Background:

• Deep learning (DL) models are advancing chemistry and materials science.
• Interpreting DL models is crucial for understanding molecular structure-property relationships.
• Current attention mechanisms offer limited interpretability for molecular substructures.

Purpose of the Study:

• To develop a versatile segmentation method and an interpretable subgraph attention (ISA) network.
• To enhance the understanding of molecular structure-property relationships using DL.
• To improve the interpretability of DL models in chemical applications.

Main Methods:

• Introduced a versatile segmentation method.
• Developed an interpretable subgraph attention (ISA) network with positive and negative streams (ISA-PN).
• Validated ISA models using datasets for aqueous solubility, lipophilicity, and melting temperature.

Main Results:

• The ISA-PN model quantifies contributions of molecular substructures via positive and negative attention scores.
• ISA-PN demonstrated significantly improved interpretability compared to ISA and GC-Net.
• Maintained similar accuracy levels to existing models while enhancing interpretability.

Conclusions:

• The ISA-PN model provides meaningful insights into substructure contributions to molecular properties.
• This work enhances the interpretability of DL models in chemical applications.
• The ISA-PN model is effective for elucidating quantitative molecular structure-property relationships.