Predicting drug characteristics using biomedical text embedding.

Area of Science:

• Pharmacology
• Bioinformatics
• Computational Biology

Background:

• Drug-drug interactions (DDIs) are a significant cause of preventable medical injuries.
• Existing methods for predicting DDIs are ineffective for new drugs with limited interaction data.
• Traditional matrix completion approaches require prior knowledge of drug interactions.

Purpose of the Study:

• To develop a novel approach for predicting drug-drug interactions (DDIs) applicable to both new and established drugs.
• To overcome the limitations of existing DDI prediction methods when dealing with drugs lacking interaction data.
• To leverage biomedical text embeddings for enhanced DDI prediction.

Main Methods:

• Proposed Adjacency Biomedical Text Embedding (ABTE), a hybrid approach combining known drug interactions with biomedical text embeddings.
• Utilized concept embedding, incorporating biomedical information, for improved text representation.
• Evaluated ABTE's performance against existing DDI prediction models and matrix factorization techniques.

Main Results:

• ABTE demonstrated superior performance compared to current DDI prediction models and matrix factorization methods.
• The concept embedding approach within ABTE yielded the highest prediction accuracy.
• Biomedical text embeddings proved effective for other drug-related prediction tasks, such as pregnancy drug safety classification.

Conclusions:

• Text and concept embeddings derived from biomedical corpora can accurately predict drug-drug interactions and drug safety.
• Embedding-based models achieve performance comparable to models using hand-crafted features.
• Text embeddings offer a data-driven approach, relying solely on published literature without manual feature engineering.