Protocol to explain graph neural network predictions using an edge-centric Shapley value-based approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cheminformatics

Background:

• Graph neural networks (GNNs) are powerful tools for analyzing complex relational data, particularly in chemistry.
• Explaining the decision-making process of GNNs is crucial for trust and interpretability, especially in scientific applications.
• Existing explanation methods may not fully capture the nuances of GNNs operating on graph-structured data.

Purpose of the Study:

• To introduce EdgeSHAPer, a novel workflow for interpreting GNNs.
• To approximate Shapley values for GNN explanations using Monte Carlo sampling.
• To provide a protocol applicable to chemical datasets and adaptable for user-defined data.

Main Methods:

• Implementation of a workflow using Python scripts for GNN explanation.
• Approximation of Shapley values via Monte Carlo sampling for feature attribution.
• Detailed steps for neural network training, explanation, and feature mapping analysis.

Main Results:

• EdgeSHAPer provides a systematic approach to explain GNN behavior.
• The workflow successfully applies to a chemical dataset, demonstrating its utility.
• The protocol is generalizable to various datasets beyond the initial chemical application.

Conclusions:

• EdgeSHAPer offers a robust method for enhancing the interpretability of graph neural networks.
• The workflow facilitates a deeper understanding of feature importance in GNN predictions.
• This protocol serves as a valuable resource for researchers working with GNNs in cheminformatics and beyond.