A multimodal vision knowledge graph of cardiovascular disease

Area of Science:

• Biomedical Informatics
• Cardiovascular Research
• Computational Biology

Background:

• Gene-disease associations are crucial for understanding disease mechanisms and finding therapeutic targets.
• Existing knowledge graphs lack individual-level data on organ structure and function.
• Biomedical imaging offers rich phenotypic information not fully integrated into current models.

Purpose of the Study:

• To develop CardioKG, a knowledge graph integrating cardiovascular imaging phenotypes with biological data.
• To utilize graph embeddings for predicting gene-disease associations and therapeutic opportunities.
• To explore drug repurposing strategies for cardiovascular diseases using integrated data.

Main Methods:

• Integrated over 200,000 computer vision-derived cardiovascular phenotypes from images.
• Combined imaging data with information from 18 biological databases.
• Employed a variational graph auto-encoder for node embedding and prediction tasks.

Main Results:

• Modeled over a million relationships within the CardioKG knowledge graph.
• Successfully predicted gene-disease associations and identified potential drug candidates.
• Identified methotrexate for heart failure and gliptins for atrial fibrillation as candidate therapies.
• Demonstrated that imaging data integration enhances pathway discovery and therapeutic opportunity identification.

Conclusions:

• CardioKG effectively integrates diverse biomedical data, including imaging phenotypes.
• The developed graph-based model aids in discovering treatable cardiovascular disease mechanisms.
• Biomedical imaging data significantly enhances graph-structured models for therapeutic target identification.