Researchers from the MRC Laboratory of Medical Sciences have taken an important step toward more data-driven and personalised cardiology by integrating cardiac imaging into a biomedical knowledge graph. The work, led by Dr Khaled Rjoob and Professor Declan O’Regan from the Computational Cardiac Imaging Group, demonstrates how combining detailed heart images with genetic and pharmacological data can significantly improve the identification of disease mechanisms and potential treatments.
Knowledge graphs are widely used to connect information on genes, diseases, molecular pathways and therapies across large biological databases. Until now, however, these networks largely lacked individual-level data on how an affected organ actually looks and functions. With the development of CardioKG, imaging data has been added to a knowledge graph for the first time, providing a more complete and physiologically grounded representation of cardiovascular disease. The study is published in the journal Nature Cardiovascular Research.
Training CardioKG
CardioKG was trained using cardiac imaging data from more than 9,500 participants in the UK Biobank, including 4,280 individuals with atrial fibrillation, heart failure or previous myocardial infarction, alongside 5,304 healthy controls. From these scans, the researchers extracted over 200,000 image-based traits that capture variation in cardiac structure and function. These traits were then integrated with information from 18 biological databases, allowing artificial intelligence to learn relationships between imaging features, genes and drug targets.
According to Professor O’Regan, the inclusion of imaging data fundamentally changed the predictive power of the model. “One of the advantages of knowledge graphs is that they integrate information about genes, drugs and diseases,” he explains. “By adding detailed heart imaging, we transformed how effectively new disease genes and therapeutic opportunities could be identified.”
Gene-disease associations
The enriched knowledge graph enabled the team to predict previously unrecognised gene–disease associations and to identify opportunities for drug repurposing. Among the most notable findings was the prediction that methotrexate, a drug commonly used in rheumatoid arthritis, may have beneficial effects in heart failure. The model also suggested that gliptins, widely prescribed for diabetes, could be effective in patients with atrial fibrillation. In addition, the analysis pointed to a potentially protective effect of caffeine in a subgroup of patients with atrial fibrillation and a rapid, irregular heart rhythm, a finding that aligns with emerging evidence from other studies.
Beyond cardiology, CardioKG serves as a proof of concept for a broader class of imaging-enriched knowledge graphs. The same approach could be applied wherever high-quality organ imaging is available, including brain imaging for neurological diseases or body composition imaging for metabolic disorders. By linking organ-level structure and function directly to molecular and pharmacological data, such models could accelerate the discovery of new therapeutic targets and support more efficient drug development.
Evolving CardioKG
From an industry perspective, the ability to rapidly generate ranked lists of high-priority genes and candidate drugs offers pharmaceutical developers a powerful starting point for validation and clinical translation. As Dr Rjoob notes, the next step is to evolve CardioKG into a dynamic, patient-centred framework. “By capturing disease trajectories over time, we can move towards personalised predictions of when disease is likely to develop and which interventions are most likely to be effective.”
The study highlights how the convergence of imaging, artificial intelligence and biomedical knowledge graphs is opening new pathways toward precision medicine, shifting discovery from isolated datasets to integrated, clinically meaningful models.
