OnSIDES database: Extracting adverse drug events from drug labels using natural language processing models.

Area of Science:

• Biomedical Informatics
• Pharmacovigilance
• Natural Language Processing

Background:

• Adverse drug events (ADEs) are a significant cause of death and increased healthcare costs in the US.
• Limited machine-readable databases hinder systematic drug safety studies.
• Advances in natural language processing (NLP) offer opportunities for extracting drug safety information from unstructured text.

Purpose of the Study:

• To develop a machine-readable database of drug-adverse drug event (ADE) pairs.
• To leverage NLP for accurate extraction of ADEs from structured product labels.
• To create a comprehensive resource for enhancing drug safety research.

Main Methods:

• Fine-tuning a PubMedBERT model for ADE term extraction from FDA Structured Product Labels.
• Compiling the OnSIDES (on-label side effects resource) database of drug-ADE pairs.
• Developing the OnSIDES-INTL database by extracting pediatric, serious, and international ADEs.

Main Results:

• Achieved high performance in ADE extraction with an F1 score of 0.90, AUROC of 0.92, and AUPR of 0.95.
• The OnSIDES database contains over 3.6 million drug-ADE pairs for 3,233 drug combinations from 47,211 labels.
• Demonstrated potential applications including novel drug target prediction and ADE enrichment analysis.

Conclusions:

• OnSIDES serves as a valuable and comprehensive resource for advancing drug safety research.
• The database facilitates systematic analysis of drug-ADE relationships.
• Enables prediction of novel drug targets, indications, and ADEs.