A multi-modal vision knowledge graph of cardiovascular disease

Understanding gene-disease associations is important for uncovering pathological mechanisms and identifying potential therapeutic targets. Knowledge graphs offer a powerful solution for representing and integrating data from multiple biomedical sources, but lack individual-level information on target organ structure and function. Here we developed CardioKG, a knowledge graph integrating over 200,000 computer vision-derived cardiovascular phenotypes from biomedical images with data extracted from 18 diverse biological databases modelling over a million relationships. A variational graph auto-encoder was used to generate node embeddings from the knowledge graph, which were used as input features to predict gene-disease associations, assess druggability and propose drug repurposing strategies. The model predicted new genetic associations and therapeutic strategies for leading causes of cardiovascular disease which were also associated with improved survival. Candidate therapies included methotrexate for heart failure and gliptins for atrial fibrillation. Imaging enhanced the ability to leverage biological data for pathway discovery. These capabilities represent an important step toward using biomedical imaging to enhance graph-structured models for identifying treatable disease mechanisms.