Comparing methods for drug-gene interaction prediction on the biomedical literature knowledge graph: performance...

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Identifying novel drug-target interactions is critical for drug discovery and repurposing.
• Knowledge graphs constructed from biomedical literature offer a powerful approach to predict these interactions.
• Link prediction methods can uncover missing connections between drug and gene entities.

Purpose of the Study:

• To compare the performance of state-of-the-art graph embedding and contextual path analysis methods for predicting drug-gene interactions.
• To evaluate the explainability of these link prediction methods.
• To assess the utility of predicted interactions for drug repurposing.

Main Methods:

• Construction of a biomedical knowledge graph from literature using text mining.
• Application and comparison of graph embedding techniques (e.g., TransE, DistMult) and contextual path analysis.
• Training a decision tree on model predictions to enhance explainability.
• Validation of predicted drug-gene interactions against external databases.

Main Results:

• A trade-off was observed between the predictive accuracy and the explainability of the link prediction methods.
• Graph embedding methods generally achieved higher predictive accuracy.
• Contextual path analysis offered more interpretable predictions.
• The decision tree approach successfully aided in understanding the prediction rationale.
• Predicted interactions showed promising results when tested for drug repurposing applications.

Conclusions:

• Link prediction on biomedical knowledge graphs is a viable strategy for identifying novel drug-gene interactions.
• The choice of method depends on the balance required between predictive performance and the need for explainable artificial intelligence (XAI).
• The findings support the use of these computational approaches to accelerate drug discovery and repurposing efforts.