Explainable artificial intelligence reveals divergent learning in pharmacophore-based hierarchical pooling graph...

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Graph neural networks (GNNs) are increasingly used for molecular property prediction.
• Hierarchical pooling offers multi-scale representation learning but its impact on GNN interpretability is underexplored.
• Rationalizing GNN predictions is crucial for understanding molecular behavior.

Purpose of the Study:

• To investigate the effect of hierarchical pooling on GNNs for molecular property prediction.
• To compare the interpretability of GNNs with and without hierarchical pooling using explainable AI (XAI).
• To integrate pharmacophore features into hierarchical GNN architectures.

Main Methods:

• Designed GNN architectural variants incorporating pharmacophore features with hierarchical pooling at different levels.
• Applied XAI methods to analyze feature importance and substructure attribution across different GNN models.
• Evaluated compound classification performance and compared explanation characteristics qualitatively and quantitatively.

Main Results:

• GNN models utilizing pharmacophore-based graph reduction or hierarchical pooling achieved performance comparable to models using complete graph representations.
• XAI analyses revealed distinct internal learning characteristics between the different GNN variants.
• The study demonstrated that reduced graph GNNs can match prediction accuracy while potentially offering different interpretability insights.

Conclusions:

• Hierarchical pooling in GNNs is a viable strategy for molecular property prediction, yielding performance comparable to traditional methods.
• Integrating pharmacophore features and employing XAI provides valuable insights into the decision-making processes of these GNN variants.
• The choice of hierarchical pooling strategy impacts GNN learning characteristics and interpretability, offering flexibility in model design.