A Drug-Drug Interaction Prediction Method Based on Atomic 3D Position Encoding and Elastic Message Passing Graph...

Area of Science:

• Pharmacology
• Computational Chemistry
• Artificial Intelligence

Background:

• Drug-drug interactions (DDIs) are critical in clinical practice.
• Current graph neural network (GNN) methods for DDI prediction often neglect atomic 3D structures and are susceptible to noise, limiting accuracy.

Purpose of the Study:

• To develop a more accurate and robust drug-drug interaction prediction model.
• To address limitations of existing GNN-based DDI prediction methods by incorporating 3D molecular information and enhancing model resilience.

Main Methods:

• Proposed a novel method, A3DPE-EMPGNN (atomic 3D position encoding and elastic message passing graph neural network).
• Constructed an atomic feature network using attention and message passing with 3D position encoding.
• Developed a molecular feature network with multi-head attention for inter-drug interactions.
• Implemented adversarial attack detection and defense with supervised and contrastive loss learning.

Main Results:

• Achieved over 98% accuracy across ACC, AUC, AP, and F1-score on two real-world datasets.
• Demonstrated superior performance compared to state-of-the-art GNN-based DDI prediction models.
• The proposed method shows enhanced robustness against adversarial attacks.

Conclusions:

• The A3DPE-EMPGNN method effectively predicts drug-drug interactions by leveraging atomic 3D structural information.
• The integration of adversarial defense strategies significantly improves model robustness.
• This approach represents a substantial advancement in computational DDI prediction.