Building Explainable Graph Neural Network by Sparse Learning for the Drug-Protein Binding Prediction

Area of Science:

• Computational Chemistry
• Bioinformatics
• Machine Learning

Background:

• Explainable Graph Neural Networks (GNNs) are used for drug-protein binding prediction.
• Current GNN models often identify chemically invalid key structures.
• Manual thresholding is required to pinpoint important substructures.

Purpose of the Study:

• To develop a novel explainable GNN model for drug-protein binding prediction.
• To ensure identified key drug structures are chemically valid.
• To improve the accuracy and interpretability of GNN-based drug-target interaction analysis.

Main Methods:

• Proposed Sparse Learning to Graph Neural Networks (SLGNN).
• Utilized a chemical-substructure-based graph representation for drug molecules.
• Incorporated generalized fused lasso with message-passing algorithms.

Main Results:

• SLGNN successfully identified chemically valid substructures critical for drug-protein binding.
• The identified substructures demonstrated enhanced predictive power compared to state-of-the-art methods.
• SLGNN eliminated the need for manual threshold selection.

Conclusions:

• SLGNN offers a robust and interpretable approach to drug-protein binding prediction.
• The method ensures chemical validity of identified key drug structures.
• SLGNN advances explainable AI in drug discovery and development.