KG-Predict: A knowledge graph computational framework for drug repurposing.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Large-scale biological data offers opportunities for drug discovery and repurposing.
• Knowledge graph methods integrate diverse data but struggle with complex entity relationships.
• Existing approaches have limitations in modeling context-sensitive inter-relationships.

Purpose of the Study:

• To develop KG-Predict, a knowledge graph computational framework for drug repurposing.
• To overcome limitations in modeling complex biomedical entity relationships.
• To infer novel drug-disease interactions for therapeutic applications.

Main Methods:

• Constructed a knowledge graph (GP-KG) integrating genotypic and phenotypic data (1,246,726 associations, 61,146 entities).
• Developed KG-Predict to learn low-dimensional representations of entities and relations from GP-KG.
• Utilized learned representations to infer new drug-disease interactions.

Main Results:

• KG-Predict achieved high cross-validation performance (AUROC=0.981, AUPR=0.409, MRR=0.261), outperforming state-of-the-art methods.
• Applied to Alzheimer's disease, KG-Predict successfully prioritized FDA-approved and clinical trial anti-AD drugs.
• Demonstrated strong predictive accuracy for drug repurposing candidates (AD: AUROC=0.868, AUPR=0.364).

Conclusions:

• KG-Predict is an effective computational framework for drug repurposing.
• The framework accurately models complex biomedical relationships for predicting new drug-disease links.
• KG-Predict holds significant potential for accelerating drug discovery and identifying novel therapeutics.