MRGCDDI: Multi-Relation Graph Contrastive Learning Without Data Augmentation for Drug-Drug Interaction Events...

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence

Background:

• Predicting drug-drug interactions (DDIs) is crucial for therapeutic safety and reducing adverse drug events.
• Graph neural network (GNN) models show promise for DDI prediction but often neglect essential drug structure and interaction data.
• Existing GNN models may require data augmentation, introducing noise and complexity.

Purpose of the Study:

• To develop an advanced deep learning method, MRGCDDI, for accurate DDI prediction.
• To effectively integrate drug molecular structure information and multi-relational DDI network data.
• To improve DDI prediction by leveraging contrastive learning without data augmentation.

Main Methods:

• MRGCDDI utilizes a contrastive learning approach without data augmentation to preserve graph data semantics.
• The method integrates drug features derived from molecular graphs.
• It incorporates information from multi-relational drug-drug interaction networks.

Main Results:

• MRGCDDI demonstrated superior performance compared to state-of-the-art methods on two benchmark datasets (Deng's and Ryu's).
• Significant improvements were observed in accuracy, Macro-F1, Macro-Recall, and Macro-Precision.
• For instance, Deng's dataset saw accuracy increase by 4.33% and Macro-F1 by 11.57%.

Conclusions:

• MRGCDDI offers an effective and robust approach for DDI prediction.
• The method enhances therapeutic safety by accurately identifying potential drug interactions.
• The integration of molecular structure and network information via contrastive learning is key to its success.