Asymmetric Drug-Drug Interaction Prediction Based on Generative Adversarial Networks and Knowledge Graph

Area of Science:

• Pharmacology
• Bioinformatics
• Artificial Intelligence

Background:

• Drug-drug interaction (DDI) prediction is crucial but challenging due to biological complexity.
• Deep learning methods have advanced DDI prediction, but often neglect the asymmetrical nature of interactions.
• This asymmetry can lead to information loss in feature learning.

Purpose of the Study:

• To propose a novel method, KGGAN-DDI, for predicting potential drug-drug interactions.
• To address the challenge of neglected asymmetry in DDI prediction.
• To enhance the accuracy and efficiency of DDI prediction models.

Main Methods:

• Utilized a knowledge graph embedding module to encode asymmetric drug pair associations.
• Employed a dual-generator Generative Adversarial Network (GAN) for realistic sample generation.
• Incorporated a least squares loss function to stabilize the optimization process and mitigate vanishing gradients.

Main Results:

• KGGAN-DDI demonstrated superior performance compared to existing state-of-the-art methods in extensive experiments.
• The knowledge graph embedding effectively captured and encoded asymmetric drug interactions.
• The dual-generator GAN and least squares loss improved prediction accuracy and model stability.

Conclusions:

• KGGAN-DDI effectively predicts drug-drug interactions by addressing the asymmetry challenge.
• The proposed method offers enhanced feature representation and contextual relevance for drug interactions.
• Experimental results and case studies validate the effectiveness and superiority of KGGAN-DDI.