Relation extraction from DailyMed structured product labels by optimally combining crowd, experts and machines

Area of Science:

• Computational biology
• Pharmacology
• Data science

Background:

• Machine learning in drug discovery relies on high-quality data, but open-source drug indication data often lacks consistency, accuracy, and context.
• Existing databases fail to differentiate between treatments targeting underlying pathology and those offering only symptomatic relief, limiting the utility of computational drug repositioning.
• The lack of proper provenance and overlap in drug indication sources results in poor-quality predictions and limited new insights.

Purpose of the Study:

• To develop a higher-quality drug-disease relation dataset suitable for drug discovery and repositioning.
• To improve the accuracy and reliability of machine learning models used in drug repurposing.
• To demonstrate the effectiveness of combining weak supervision (programmatic labeling and crowdsourcing) with deep learning for relation extraction.

Main Methods:

• Utilized weak supervision techniques, including programmatic labeling and crowdsourcing, to extract drug-disease relations from DailyMed text.
• Employed deep learning methods for relation extraction to create a comprehensive drug-disease dataset.
• Constructed a machine learning model to classify drug-disease relations into four categories: treatment, symptomatic relief, contradiction, and effect.

Main Results:

• Generated a high-quality drug-disease relation dataset with significant overlap with the manually curated DrugCentral dataset.
• The developed machine learning model achieved a 15.5% improvement in F1 score (71.8%) using Bi-LSTM compared to the best discrete methods.
• The model successfully classified drug-disease relations into treatment, symptomatic relief, contradiction, and effect categories.

Conclusions:

• High-quality data is essential for building accurate and reliable drug repurposing prediction models.
• The integration of crowdsourcing, expert knowledge, and machine learning methods effectively enhances datasets and predictive model performance.
• This approach offers a scalable solution for improving drug-disease relation databases, thereby advancing drug discovery and repositioning efforts.