Enhancing drug-drug interaction prediction by three-way decision and knowledge graph embedding

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Drug-drug interaction (DDI) prediction is crucial for pharmaceutical research and clinical practice.
• Current computational methods often struggle with data incompleteness and uncertainty, leading to ambiguous classification of DDIs.
• Biomedical knowledge graphs offer rich supplementary information valuable for enhancing DDI prediction.

Purpose of the Study:

• To develop an advanced computational method for more accurate DDI prediction.
• To address the limitations of binary classification in existing DDI prediction models.
• To leverage biomedical knowledge graphs and three-way decision for improved DDI identification.

Main Methods:

• A three-way decision-based method (3WDDI) was proposed, integrating knowledge graph embeddings as supplementary features.
• Convolutional Neural Networks (CNNs) were used to classify drug pairs into positive, negative, and boundary regions based on chemical structure features.
• Delay decisions for boundary-region samples were made by incorporating knowledge graph embedding features.

Main Results:

• The 3WDDI method achieved high performance metrics: Accuracy (0.8922), AUPR (0.9614), AUC (0.9582), and F1-score (0.8930).
• The integration of knowledge graph embeddings significantly enhanced the accuracy of DDI prediction.
• 3WDDI demonstrated superior performance compared to several baseline DDI prediction models.

Conclusions:

• The proposed 3WDDI method effectively enhances drug-drug interaction prediction accuracy by utilizing three-way decision and knowledge graph embeddings.
• This approach successfully addresses the uncertainty and incompleteness issues inherent in traditional DDI prediction methods.
• 3WDDI represents a significant advancement in computational DDI prediction, offering improved reliability for pharmaceutical research and clinical applications.