A Domain Adaptive Interpretable Substructure-Aware Graph Attention Network for Drug-Drug Interaction Prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Accurate drug-drug interaction (DDI) prediction is vital for clinical efficacy and drug safety.
• Existing computational methods for DDI prediction often overlook critical substructural features and struggle with cross-domain generalization.

Purpose of the Study:

• To introduce SAGAN, a domain-adaptive, interpretable, substructure-aware graph attention network for enhanced DDI prediction.
• To address limitations in current DDI prediction models by incorporating substructural insights and improving generalization capabilities.

Main Methods:

• Developed a novel substructure segmentation method using attention mechanisms and unsupervised clustering to identify key interaction regions.
• Applied a conditional domain adversarial network to achieve cross-domain generalization by optimizing source domain and adversarial losses.
• Evaluated SAGAN on four real-world datasets in both in-domain and cross-domain scenarios.

Main Results:

• SAGAN demonstrated superior performance compared to state-of-the-art DDI prediction models across all evaluated datasets and scenarios.
• The model successfully extracted pharmacologically significant substructures, aiding in the identification of novel interaction sites.
• Visualization results confirmed the model's ability to pinpoint critical regions involved in drug interactions.

Conclusions:

• SAGAN offers a significant advancement in DDI prediction by integrating substructure awareness and domain adaptation.
• The model's interpretability and generalization capabilities provide valuable tools for drug developers in screening and optimization.
• The developed substructure-aware approach can aid in discovering new local interaction sites and refining drug structures.