Adverse Drug Event Prediction Using Noisy Literature-Derived Knowledge Graphs: Algorithm Development and Validation.

Area of Science:

• Computational linguistics
• Bioinformatics
• Machine learning

Background:

• Adverse drug events (ADEs) pose significant global health and economic burdens.
• Post-marketing surveillance (pharmacovigilance) is crucial for identifying ADEs not detected during clinical trials.
• Biomedical literature is a rich source for ADE discovery, often processed using knowledge graphs (KGs) and natural language processing (NLP).

Purpose of the Study:

• To address noise (false positives/negatives) in literature-derived KGs caused by NLP inference.
• To leverage NLP-derived confidence scores in representation learning for improved ADE prediction.
• To develop and evaluate weighted graph representation learning methods that account for inference inaccuracies.

Main Methods:

• Developed Weighted DeepWalk and Weighted TransE by incorporating confidence scores into existing representation learning algorithms.
• Learned node representations from the large-scale Semantic MEDLINE Database KG (over 93 million relations).
• Compared weighted methods against unweighted counterparts and the state-of-the-art Embedding of Semantic Predications using benchmark datasets for ADE prediction.

Main Results:

• Weighted representation learning methods significantly outperformed their unweighted versions and Embedding of Semantic Predications.
• Achieved performance improvements of up to 5.75% in F1-score and 8.4% in area under the ROC curve.
• Demonstrated enhanced state-of-the-art ADE prediction accuracy from literature-derived KGs.

Conclusions:

• The developed classification models can assist pharmacovigilance teams in discovering novel ADEs.
• Highlighting the critical importance of modeling inaccuracies within inferred KGs for effective representation learning.
• These findings advance computational approaches to drug safety monitoring.