TransFOL: A Logical Query Model for Complex Relational Reasoning in Drug-Drug Interaction

Area of Science:

• Pharmacology
• Bioinformatics
• Artificial Intelligence

Background:

• Predicting drug-drug interactions (DDIs) is vital for drug discovery and safety.
• Traditional wet lab methods for DDI identification are costly and time-consuming.
• Existing deep learning models for DDIs lack comprehensive reasoning capabilities and fail to integrate diverse biomedical factors.

Purpose of the Study:

• To develop an advanced model for predicting drug-drug interactions (DDIs) that overcomes the limitations of existing methods.
• To incorporate complex biomedical factors beyond simple drug pairs into DDI prediction.
• To frame DDI prediction as a link prediction problem on knowledge graphs for enhanced reasoning.

Main Methods:

• Proposed TransFOL, a DDI prediction model integrating Cross-Transformer and Graph Convolutional Networks (GCNs) within a first-order logical query framework.
• Constructed a biomedical query graph to learn entity and relation embeddings.
• Employed an enhancement module to aggregate semantic information and utilized Cross-Transformer for node semantic encoding and GCN for neighbor aggregation.

Main Results:

• TransFOL outperformed state-of-the-art methods on traditional DDI prediction tasks across two benchmark datasets.
• The model demonstrated strong performance and generalization capabilities in more realistic settings with diverse biomedical information.
• Experimental results validated the model's ability to handle complex drug reasoning tasks.

Conclusions:

• TransFOL offers a powerful and generalizable approach for drug-drug interaction prediction.
• The model's ability to integrate complex biomedical factors enhances its applicability in real-world drug discovery and recommendation scenarios.
• This knowledge graph-based approach advances the field of computational pharmacology and drug safety.