idse-HE: Hybrid embedding graph neural network for drug side effects prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug development faces high failure rates due to unexpected side effects.
• Current in silico methods often rely on single data perspectives (e.g., chemical structure, network topology, or knowledge graphs).
• A unified approach is needed to integrate diverse drug information for improved side effect prediction.

Purpose of the Study:

• To develop a novel hybrid embedding graph neural network model (idse-HE) for predicting potential drug side effects.
• To effectively fuse drug features from both macroscopic biological networks and microscopic molecular structures.
• To enhance the accuracy and reliability of in silico drug side effect prediction.

Main Methods:

• Proposed idse-HE, a hybrid embedding graph neural network integrating graph and node embedding modules.
• Fused drug chemical structure, substructure sequence, and network topology information.
• Utilized drug and side effect representations as implicit factors to reconstruct the original matrix for prediction.

Main Results:

• idse-HE demonstrated superior performance compared to existing advanced methods in predicting drug side effects.
• The model exhibited stable and robust performance across all evaluated indicators.
• Confirmed several previously unrecognised real drug-side effect pairs from the model's predictions.

Conclusions:

• idse-HE effectively integrates multi-modal drug data for accurate side effect prediction.
• The model offers a significant advancement over single-perspective approaches in drug discovery.
• A web server is available for researchers to access data, code, and reproduce results.