Transformer-based graphs for drug-drug interaction with chemical knowledge embedding

Area of Science:

• Pharmacology
• Computational Chemistry
• Artificial Intelligence

Background:

• Drug-drug interactions (DDIs) pose significant medical risks, necessitating accurate identification methods.
• Deep learning enhances DDI prediction but often lacks chemical information integration and interpretability.
• Current models struggle to elucidate the mechanisms behind DDIs.

Purpose of the Study:

• To introduce TRACE, a novel transformer-based graph representation learning framework for DDI prediction.
• To integrate chemical knowledge and enhance the interpretability of DDI prediction models.
• To improve the accuracy and generalization ability of DDI prediction.

Main Methods:

• Developed TRACE, a transformer-based graph representation learning framework.
• Integrated chemical knowledge into the model architecture.
• Utilized an attention mechanism for interpretability and substructure identification.

Main Results:

• TRACE demonstrated superior performance compared to state-of-the-art models in both in-distribution and out-of-distribution settings.
• The framework successfully identified high-risk chemical substructures associated with DDIs.
• Achieved strong predictive performance and generalization ability.

Conclusions:

• TRACE offers a robust and interpretable approach for DDI prediction.
• The model enhances understanding of DDI mechanisms through substructure analysis.
• TRACE supports safer drug development and informs combination therapy strategies.