DWPL-GCNMF: Structure-Aware Dynamic Weighted Pseudo-Label Learning for Adverse Drug Reaction Prediction.

Area of Science:

• Pharmacogenomics
• Computational Biology
• Drug Safety

Background:

• Adverse drug reaction (ADR) prediction is crucial for drug safety.
• Current methods struggle with extremely sparse data, limiting generalization and introducing bias.
• Existing positive-unlabeled learning strategies are prone to confirmation bias.

Purpose of the Study:

• To develop a robust framework for ADR prediction that overcomes data sparsity and confirmation bias.
• To improve the accuracy and stability of predicting potential ADRs.
• To enhance the prioritization of potential ADRs for further investigation.

Main Methods:

• Proposed DWPL-GCNMF, a structure-aware semisupervised framework.
• Combined graph convolutional network (GCN)-based structural embeddings from a drug-protein knowledge graph with matrix factorization.
• Implemented a dynamic weighted pseudolabel learning strategy to mitigate confirmation bias and improve training data.
• Fused multiple base models to enhance prediction stability.

Main Results:

• DWPL-GCNMF demonstrated superior performance on benchmark datasets (DrugBank and SIDER) with high sparsity (∼2.6% and ∼2.3% observed associations).
• Achieved improved Area Under Precision-Recall (AUPR) and F1 scores compared to baseline methods.
• Significantly enhanced top-K prioritization of potential ADRs, with Recall@15 reaching up to 0.8078 and 0.8146.

Conclusions:

• Structure-aware representation learning combined with dynamic weighted pseudolabeling offers a robust approach for ADR prediction.
• The proposed framework effectively addresses data sparsity challenges in ADR prediction.
• DWPL-GCNMF provides a valuable tool for prioritizing potential ADRs in drug development and safety monitoring.