Informative relational learning for adverse reaction prediction with enhanced generalization to novel drugs

Area of Science:

• Pharmacovigilance
• Machine Learning
• Biomedical Informatics

Background:

• Accurate prediction of adverse drug reactions (ADRs) is crucial for drug safety.
• Current machine learning methods struggle with inadequate ADR representations and poor generalization to novel drugs.

Purpose of the Study:

• To develop an advanced machine learning model for predicting ADRs.
• To improve ADR representation learning and enhance generalization to new drugs.

Main Methods:

• Constructed a multi-source, multi-relational ADR graph integrating hierarchical structure and co-occurrence data.
• Applied a relational graph convolutional network (R-GCN) for relation-aware ADR representation learning.
• Utilized a Conditional Domain Adversarial Network (CDAN) and a dual-branch mixture-of-experts (Dual-MoE) module to improve generalization and model ADR patterns.

Main Results:

• The proposed method consistently outperformed seven baseline methods, achieving significant F1 score improvements.
• Demonstrated enhanced performance on uncommon ADRs and improved robustness under data sparsity.
• Showcased more balanced precision-recall trade-offs compared to existing approaches.

Conclusions:

• The developed knowledge-guided approach effectively addresses limitations in current ADR prediction models.
• The method offers a promising advancement in drug safety surveillance through improved ADR prediction.
• The findings highlight the potential of integrating complex biomedical information for robust drug safety analysis.