Graph neural network-based drug-drug interaction prediction.

Area of Science:

• Pharmacology
• Computer Science
• Bioinformatics

Background:

• Accurate prediction of drug-drug interactions (DDIs) is vital due to increasing polypharmacy and the need for treatment efficacy and patient safety.
• Adverse DDIs can lead to toxicity, reduced efficacy, or fatal outcomes, highlighting the importance of reliable prediction methods.

Purpose of the Study:

• To extend existing graph neural network (GNN) models for improved drug-drug interaction (DDI) prediction.
• To investigate the performance of various GNN architectures, including basic and advanced models, on DDI datasets.

Main Methods:

• Developed and evaluated advanced GNN models by integrating techniques like skip connections and post-processing layers, building upon existing architectures (SAGE, GAT, GCN).
• Compared the performance of basic GNN models against more complex ones, including adaptive graph diffusion models, across three distinct DDI datasets.

Main Results:

• Experimental results indicated that simpler GNN models occasionally outperformed advanced ones on specific evaluation metrics.
• Graph Convolutional Network (GCN) with skip connections, GCN with NGNN, and SAGE with NGNN demonstrated competitive accuracy compared to other baseline models.

Conclusions:

• The study underscores that simpler GNN architectures can achieve comparable or superior performance in DDI prediction compared to highly complex models.
• The findings suggest a nuanced approach to GNN model selection for DDI prediction, emphasizing empirical evaluation on specific datasets.