Researchers at the Computational Cardiac Imaging Group at the MRC Laboratory of Medical Sciences, London, have developed a new method that could help accelerate the identification of drugs for heart disease by combining cardiac imaging with artificial intelligence (AI)–driven knowledge graphs. The new method, published in Nature Cardiovascular Research, integrates detailed heart structure and function from medical images directly into a biological knowledge graph. Called CardioKG, this approach aids the prediction of gene–disease links which can identify existing drugs that could be repurposed to treat cardiovascular conditions better than currently available treatments.
Knowledge graphs are a tool for collecting information from biological databases and linking what is already known about genes, diseases, treatments, molecular pathways and symptoms in a structured network. But they have lacked the kind of detailed, individual-level information about how the affected organ actually looks and functions.
CardioKG closes that gap by adding imaging data to a knowledge graph, the first time images have been integrated in this way. While knowledge graphs can connect a broad array of information including genomics, molecular pathways, diseases, and drugs, they have relied on abstracted or population-level data. By incorporating imaging-derived phenotypes, the researchers can now capture patient-level variation in how disease affects the heart itself.
To build the model, the team gathered cardiac imaging data from 4,280 patients with atrial fibrillation, heart attack, or heart failure from the UK Biobank along with 5,304 healthy participants. From these imaging data, the researchers generated more than 200,000 image-based traits defining heart structure and function. These traits were then integrated with information from 18 biological databases, creating a network of over a million relationships linking genes, diseases, pathways, and drugs.
The use of AI was key to deriving knowledge from the imaging-knowledge graph integration. The researchers used a variational graph auto-encoder to learn embeddings from the knowledge graph. This provided predictions of gene–disease associations and drug repurposing possibilities. The addition of imaging data improved the model’s performance and predictive capabilities, by creating imaging-derived “endophenotypes” that are closer to disease mechanisms than many observable traits.
Using CardioKG, the team identified new disease-associated genes and predicted several potential drug repurposing opportunities. Among these were methotrexate, a drug commonly used to treat rheumatoid arthritis, as a candidate for heart failure, and gliptins, used in diabetes, as potential treatments for atrial fibrillation. The model also suggested a protective association between caffeine and atrial fibrillation in patients with irregular and fast heart rhythms.
“What’s exciting is there are other recent studies in the field which support our preliminary findings,” said senior author Declan O’Regan, PhD, a principal investigator at MRC. “[This] highlights the huge potential of knowledge graphs in uncovering existing drugs that might be repurposed as new treatments.”
The researchers noted that this new method adds power to previous genome-wide association studies (GWAS) that have identified many disease-linked variants but aren’t able to specifically identify actionable treatment targets. “An important bottleneck has been the limited availability of individual-level phenotypes that can be linked to other semantic information in the network,” the researchers wrote. Earlier studies have suggested the potential of knowledge graphs, but the development of CardioKG shows that adding imaging data enriches these GWAS, improve pathway discovery, and drug target prioritization.
The implications for clinical care point to earlier and more precise identification of therapeutic targets on a patient-by-patient basis. By quickly identifying high-priority genes and candidate drugs, imaging-enhanced knowledge graphs could also guide pharmaceutical development and allow for more targeted clinical trials. Data from this study also showed that predicted drug repurposing for heart failure could improve patient survival.
The MRC team is now planning to further refine their approach using CardioKG. “Building on this work, we will extend the knowledge graph into a dynamic, patient-centered framework that captures real disease trajectories,” said first author Khaled Rjoob, PhD, a postdoctoral researcher at Imperial College London. The researchers will also seek to include more diverse imaging datasets and noted there is an opportunity to apply the same approach to include the brain and body fat, where imaging is already widely used.
