KGMAEDDI: Knowledge Graph and Molecular-Graph Masked Autoencoder for Drug-Drug Interaction Prediction.

Area of Science:

• Pharmacology
• Bioinformatics
• Computational Chemistry

Background:

• Drug-drug interaction (DDI) prediction is crucial for safe drug development.
• Previous methods often ignored drug molecular structures or biomedical knowledge graph (KG) relationships.
• Integrating both structural and semantic information is needed for accurate DDI prediction.

Purpose of the Study:

• To propose KGMAEDDI, a novel framework for DDI prediction.
• To effectively integrate molecular structures and KG semantic knowledge.
• To improve the accuracy and robustness of DDI prediction models.

Main Methods:

• Utilized a message-passing neural network for drug molecular structure feature extraction.
• Employed a knowledge-aware attention network for KG semantic representation learning.
• Integrated modalities using a masked autoencoder and bi-directional cross-attention for mutual reconstruction and alignment.

Main Results:

• KGMAEDDI demonstrated superior performance over state-of-the-art baselines on the DrugBank dataset.
• The framework achieved high accuracy in both binary and multi-class DDI prediction settings.
• The results validate the effectiveness of integrating structural and semantic information for DDI prediction.

Conclusions:

• KGMAEDDI successfully integrates molecular structures and KG semantic knowledge for enhanced DDI prediction.
• The proposed fusion strategy effectively aligns drug representations in a shared latent space.
• This approach offers a promising direction for advancing computational DDI prediction.