DisenKGE-DDI: A Knowledge Graph Embedding Framework Based on Disentangled Graph Attention Networks for Drug-Drug...

Area of Science:

• Pharmacology and Bioinformatics
• Artificial Intelligence in Medicine
• Computational Drug Discovery

Background:

• Drug-drug interactions (DDIs) pose risks to therapeutic efficacy and patient safety.
• Accurate DDI prediction is crucial for clinical safety and rational drug use.
• Existing deep learning models for DDI prediction have limitations in capturing both interaction directionality and diversity.

Purpose of the Study:

• To develop a novel framework, DisenKGE-DDI, for enhanced DDI prediction.
• To address the limitations of existing methods by incorporating both micro- and macro-disentanglement mechanisms.
• To improve the comprehensive modeling of pharmacological relationships for more accurate DDI prediction.

Main Methods:

• Introduced DisenKGE-DDI, a framework based on a disentangled graph attention network.
• Implemented micro-disentanglement using a factor-aware relation-based message aggregation and dual-layer attention.
• Applied macro-disentanglement with mutual information regularization to ensure independence of semantic components.

Main Results:

• DisenKGE-DDI demonstrated superior efficacy compared to state-of-the-art methods on public benchmark datasets.
• The framework effectively captures intricate local semantic features and diverse interaction characteristics.
• The proposed disentanglement mechanisms enhance the adaptiveness and comprehensiveness of drug embeddings.

Conclusions:

• DisenKGE-DDI offers a significant advancement in DDI prediction accuracy and reliability.
• The framework's ability to model both directional and diverse interaction aspects is key to its success.
• This approach holds promise for improving drug safety and guiding clinical decision-making.