DGX: Uncovering General Behavior of Deep Graph Models With Model-Level Explanation

Area of Science:

• Artificial Intelligence
• Bioinformatics
• Computational Biology

Background:

• Deep graph learning models are increasingly used for complex systems, especially in bioinformatics.
• Existing explanation methods struggle to uncover general graph patterns guiding these models.
• A versatile explanation technique for black-box deep graph models is needed.

Purpose of the Study:

• To propose DGX, a novel deep graph model explainer.
• To generate explanatory graphs that reveal underlying graph patterns.
• To provide diverse and customizable explanations for trained deep graph models.

Main Methods:

• DGX generates multiple, distinguishable explanatory graphs.
• It encodes structural knowledge captured by graph neural networks.
• Explanations can be customized using prior knowledge or constraints.

Main Results:

• DGX effectively explains deep graph models on synthetic and real-world graph data.
• It identified groups of mutagenic compounds in a mutagenicity prediction task.
• DGX revealed structural patterns differentiating autism spectrum disorder from controls in brain networks.

Conclusions:

• DGX offers an effective, diverse, and customizable approach to explaining deep graph models.
• It enhances understanding of models trained on complex graph data.
• DGX has significant applications in biomedicine and bioinformatics for mechanism discovery.