Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Networks with Group Lasso Regularization

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Cheminformatics

Background:

• Explainable AI (XAI) accelerates drug discovery by enhancing molecular representation and property prediction.
• Developing end-to-end explainable models for structure-activity relationships (SAR) is challenging due to limited data and sensitivity of molecular properties to small structural changes.
• Identifying key molecular moieties linked to compound-protein affinity for specific targets is crucial for efficient drug design.

Purpose of the Study:

• To propose a framework using graph neural networks (GNNs) for predicting compound-protein affinity and explaining property differences in target-specific drug discovery.
• To enhance model explainability through structure-aware loss functions and regularization techniques (group lasso, sparse group lasso).
• To identify critical molecular substructures influencing compound-protein affinity and guide lead optimization.

Main Methods:

• Implemented a GNN framework leveraging property and structure information from molecule pairs exhibiting activity cliffs.
• Employed group lasso and sparse group lasso regularizations to prune and highlight relevant molecular subgraphs for activity differences.
• Applied the framework to activity cliff data for tyrosine-protein kinases (Src, Abl, Tec families) and anaplastic lymphoma kinase.

Main Results:

• Integrating common- and uncommon-node information with sparse group lasso improved molecular property prediction accuracy, indicated by lower RMSE and higher Pearson's correlation coefficients.
• Regularization techniques enhanced GNN feature attribution, boosting graph-level global direction scores and atom-level coloring accuracy.
• The framework demonstrated improved model interpretability in identifying critical molecular substructures for specific protein targets.

Conclusions:

• The proposed GNN framework effectively predicts compound-protein affinity and explains property differences in target-specific drug discovery.
• Structure-aware regularization significantly enhances the explainability and accuracy of AI models in SAR studies.
• This approach strengthens AI interpretability in drug discovery pipelines, aiding in the identification of key moieties for lead optimization.