Multi-GraphDDI: Multi-Feature Fusion and Interaction for Graph-Based Drug-Drug Interaction Prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) pose significant risks to therapeutic efficacy and patient safety.
• Accurate computational prediction of DDIs is crucial for drug discovery and clinical decision support.

Purpose of the Study:

• To develop a novel structure-only framework, Multi-GraphDDI, for predicting drug-drug interactions (DDIs).
• To evaluate the performance of Multi-GraphDDI against existing methods on benchmark datasets.

Main Methods:

• Multi-GraphDDI encodes molecular fingerprints (ECFP4, PubChem, pharmacophore) as image channels and uses a graph isomorphism network (GIN) for graph representation.
• A bidirectional feature-interaction module with cross-attention aligns image-based and graph-based molecular representations.
• The fused features are utilized to predict drug-drug interaction scores.

Main Results:

• Multi-GraphDDI achieved high performance across ChCh-Miner, ZhangDDI, and DeepDDI datasets, with AUC/AUPR/F1 scores reaching up to 0.9986/0.9998/0.9730.
• The proposed framework outperformed competing DDI prediction methods.
• Integration of heterogeneous structural cues through coarse- and fine-grained feature interaction proved effective.

Conclusions:

• Multi-GraphDDI offers an effective and scalable solution for DDI prediction by leveraging complementary structural information.
• The structure-only approach eliminates the need for external biological networks, simplifying the prediction process.
• This method holds promise for advancing drug discovery and clinical decision support systems.