EdgeSHAPer: Bond-centric Shapley value-based explanation method for graph neural networks

Area of Science:

• Artificial Intelligence
• Cheminformatics
• Computational Chemistry

Background:

• Graph neural networks (GNNs) are powerful deep learning models for graph-structured data.
• GNNs exhibit a "black box" nature, limiting interpretability and trust.
• Existing methods for explaining GNN decisions are limited, especially for edge-centric analysis.

Purpose of the Study:

• To introduce EdgeSHAPer, a novel and generally applicable method for explaining GNN-based models.
• To assess the importance of edges in GNN predictions, particularly in molecular graphs.
• To provide intuitive explanations for GNN predictions in drug discovery tasks.

Main Methods:

• EdgeSHAPer utilizes the Shapley value concept from game theory to quantify edge importance.
• The method is applied to compound activity prediction, a key task in drug discovery.
• Feature mapping is combined with edge importance to generate interpretable explanations.

Main Results:

• EdgeSHAPer provides a higher resolution in differentiating features that determine predictions compared to existing methods.
• The approach successfully identifies minimal pertinent positive feature sets for predictions.
• EdgeSHAPer demonstrates effectiveness in explaining molecular graph-based predictions.

Conclusions:

• EdgeSHAPer offers a valuable tool for understanding and rationalizing GNN decisions.
• The edge-centric approach is particularly relevant for molecular graphs in drug discovery.
• This method enhances the interpretability and trustworthiness of GNN models in scientific applications.