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

Researchers developed an explainable graph neural network framework that uses group lasso regularization to predict compound-protein affinity and identify critical molecular substructures in drug discovery. The approach leverages activity-cliff molecule pairs to improve predictions for tyrosine-protein kinases and other targets, demonstrating enhanced interpretability and accuracy in molecular property prediction.

This research addresses a fundamental challenge in computational drug discovery: developing AI models that not only predict molecular properties accurately but also explain their predictions in chemically meaningful ways. Traditional machine learning approaches often function as black boxes, making it difficult for medicinal chemists to understand which molecular features drive drug efficacy or toxicity. By integrating graph neural networks with sparse group lasso regularization, the researchers created a framework that simultaneously improves prediction accuracy while highlighting the specific molecular subgraphs responsible for compound-protein interactions.

The methodology builds on established principles in structural biology and machine learning. Activity-cliff molecules—compounds with similar structures but dramatically different biological activities—provide ideal training data for understanding structure-activity relationships. The sparse group lasso regularization prunes irrelevant molecular connections while preserving those critical for binding affinity, effectively creating a sparse, interpretable model. Testing across six tyrosine-protein kinases demonstrates the framework's applicability to clinically relevant targets.

For the pharmaceutical and biotech industries, this work streamlines lead optimization by providing computational guidance on which molecular modifications most impact target binding. Chemists can focus synthesis efforts on validated structural modifications rather than conducting exhaustive combinatorial screening. The improved interpretability reduces the risk of developing compounds with unintended properties, accelerating development timelines and reducing costs. The framework's success on kinase targets suggests applicability across protein families.

Future development should validate predictions against new experimental data and extend the approach to larger, more diverse compound libraries. Integration with other explainable AI techniques could further enhance mechanistic insights into protein-ligand interactions, potentially transforming how computational and medicinal chemists collaborate in early-stage drug discovery.

• →Graph neural networks with sparse group lasso regularization improve both prediction accuracy and explainability in compound-protein affinity modeling.
• →The framework identifies critical molecular substructures driving drug efficacy, enabling more focused lead optimization strategies.
• →Activity-cliff molecule pairs provide valuable training data for understanding how small structural changes produce large biological activity differences.
• →Improved model interpretability reduces synthesis failures and accelerates drug discovery pipelines by validating structural modifications computationally.
• →Results demonstrate applicability across multiple tyrosine-protein kinase targets, suggesting broader utility across protein families in drug discovery.