AI-Driven Prediction of Gene-Disease Associations Using Computer Vision-Derived Cardiovascular Phenotypes

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Nature Cardiovascular Research https://doi.org/10.1038/s44161-025-00757-4 
Dr. Khaled Rjoob  et al.

Points

The study developed CardioKG which is a multimodal knowledge graph that integrates over two hundred thousand computer vision derived phenotypes with data from eighteen diverse biological databases to model relationships.
Researchers utilized imaging data from over nine thousand participants to capture variations in heart structure and function across patients with atrial fibrillation heart failure and prior history of heart attacks.
A variational graph auto encoder was used to generate node embeddings from the knowledge graph to predict gene disease associations and assess the druggability of various cardiovascular therapeutic targets.
The integration of individual level heart imaging data significantly improved the accuracy of the model in identifying potential drug repurposing strategies such as using methotrexate for the treatment of heart failure.
These findings support the implementation of biomedical imaging to enhance graph structured models which help clinicians and researchers identify treatable disease mechanisms and improve survival rates in high risk populations.

Summary

This study introduced CardioKG, a multimodal knowledge graph designed to integrate high-dimensional imaging data with biological databases to enhance the prediction of gene-disease associations and therapeutic targets. Researchers utilized cardiovascular MRI data from 4,280 UK Biobank participants with atrial fibrillation, heart failure, or myocardial infarction, alongside 5,304 healthy controls. By extracting over 200,000 computer vision-derived cardiovascular phenotypes, the model captured fine-grained structural and functional variations to supplement traditional genomic and proteomic data.

The computational framework integrated these imaging traits with data from 18 diverse biological databases, modeling more than one million unique relationships. A variational graph auto-encoder was employed to generate node embeddings, allowing for the systematic assessment of druggability and the identification of drug repurposing strategies. The integration of individual-level phenotypic data was found to transform the model’s predictive power, enabling more accurate mapping of genetic influences on cardiac morphology and function compared to graphs lacking imaging inputs.

Clinical applications of CardioKG successfully identified novel therapeutic opportunities for leading causes of cardiovascular mortality, which were associated with improved patient survival. Specific findings included the potential repurposing of methotrexate for heart failure management and gliptins for the treatment of atrial fibrillation. By enhancing pathway discovery through the addition of imaging-based traits, this approach demonstrates the utility of graph-structured AI models in uncovering treatable disease mechanisms and accelerating the development of precision cardiovascular therapies.

Link to the article: https://www.nature.com/articles/s44161-025-00757-4

References

Rjoob, K., McGurk, K. A., Zheng, S. L., Curran, L., Ibrahim, M., Zeng, L., Kim, V., Tahasildar, S., Kalaie, S., Senevirathne, D. S., Gifani, P., Losev, V., Zheng, J., Bai, W., de Marvao, A., Ware, J. S., Bender, C., & O’Regan, D. P. (2025). A multimodal vision knowledge graph of cardiovascular disease. Nature Cardiovascular Research, 1–16. https://doi.org/10.1038/s44161-025-00757-4