Bayer Built a Knowledge Graph That Predicts Heart Failure Drug Targets

Cardiovascular disease is the leading cause of death worldwide. Developing new treatments costs an average of US$2 billion per drug, takes 10 to 15 years, and fails 70% of the time in phase 2 trials. A central bottleneck is identifying the right therapeutic targets: genome-wide association studies have flagged thousands of genetic variants linked to chronic heart disease, but most are too small in effect to translate into actionable drug candidates.

What’s been missing is a way to connect genetic data with the structural and functional reality of what’s happening inside a patient’s heart — the individual-level phenotypes that sit closer to disease mechanisms than statistical associations alone.

A team at Bayer AG’s R&D Pharmaceuticals division across Wuppertal, Berlin and Cambridge, Massachusetts, have been developing an internal knowledge graph platform integrating primary data with prior knowledge, setting out to close that gap for heart failure. Their work has so far led to five programs in experimental validation across the Cardiovascular and Renal therapeutic areas, driven by in silico discovery approaches.

Recently, the Bayer team worked with researchers at Imperial College London and the British Heart Foundation, led by O’Regan and Rajoob, to apply similar methods to cardiac imaging data. They built CardioKG: a knowledge graph in Neo4j that integrates computer vision-derived cardiac imaging data with 18 biological databases, modeling over a million relationships between genes, diseases, drugs, pathways, and anatomical structures.

Their findings, published in Nature Cardiovascular Research in December 2025, identified druggable gene targets across heart failure, atrial fibrillation and myocardial infarction. For heart failure, five of the ten top-ranked genes were druggable — AR, APP, GATA2, EGR1 and EP300 — collectively addressable by 48 existing medications.

Why graph? Connecting what no single database could

Before CardioKG, the standard knowledge graph approach to predicting which genes might cause heart disease relied on information from curated biological databases — genomic annotations, protein interactions, pathway ontologies, drug-target mappings. These were useful for organizing known biology, but they had a structural blind spot: no individual-level data linking genes to what was actually observable in patients’ organs. Primary patient data is crucial to understanding complex multi-faceted diseases like heart failure. Applying machine learning and artificial intelligence methods to these data is critical for identifying disease patient populations and predicting therapeutic hypotheses with a higher chance of success in clinical trials.

“The idea is that the graph is not only biomedical entities but rather it’s a patient-centric graph. What you can use such a graph for is things like patient subtyping, biomarker identification and drug effect prediction,” says Mahmoud Ibrahim, Data Science and AI Mission Lead at Bayer’s R&D Pharmaceuticals division.

The relevant data existed — genomic databases like DisGeNET and BioGRID, drug interaction databases like DrugBank and DGIdb, pathway databases, and 200,000+ cardiac imaging phenotypes from the UK Biobank — but it was scattered across 18 separate repositories with no unified data model. Researchers could query one database at a time but couldn’t traverse the connections between them.

Meanwhile, existing graph embedding techniques like Node2Vec and ComplEx could capture proximity in a graph but failed to preserve the directionality of biological relationships. A gene that causes a pathway to activate is fundamentally different from a gene that is activated by that pathway. Flattening these into undirected similarity scores destroyed the very information needed for drug targeting.

Building CardioKG: 19 data sources in one connected graph

The CardioKG team, led by Rajoob and O’Regan at Imperial College London, built a directed, multimodal knowledge graph in Neo4j, integrating data that had never been connected in a single graph structure before. The final graph contained 33,277 nodes across 12 entity types — genes (18,606), human participants (9,584), medications (2,106), molecular pathways (1,707), diseases (1,036), anatomical regions (160), and six additional categories — connected by 1,195,437 relationships.

The critical architectural decision was representing individual patients directly as nodes in the graph, connected to genetic information and to cardiac MRI-derived imaging traits as nodes in the graph. Each of 21 imaging measurements — including left ventricular ejection fraction, chamber volumes, wall thickness, aortic distensibility, and strain rates — became a node connected to anatomical regions, diseases, and individual participants. Computer vision pipelines extracted over 200,000 image-derived phenotypes that were semantically linked to genes and diseases through the graph structure.

This meant a query could traverse from a specific cardiac measurement (say, reduced ejection fraction) through anatomical structures (left ventricle) to gene expression profiles, to known drug targets, to candidate medications — all within a single connected graph.

To preserve the directionality of biological relationships in their embeddings, the team developed a novel directed variational graph auto-encoder (DVGAE). The architecture processed four inputs — node type, node properties, edge properties, and edge direction — achieving 98.1% edge-wise accuracy in reconstructing the original graph. Three machine learning classifiers were then trained on the embeddings; SVM achieved the best performance with AUC-ROC scores of 0.80 for heart failure, 0.78 for atrial fibrillation, and 0.83 for myocardial infarction.

Including imaging changed what the graph could see. For heart failure, 66 disease-relevant pathways surfaced with imaging data against four without. Methotrexate emerged as a repurposing candidate, with heart failure patients already taking it for rheumatoid arthritis showing higher survival probability than those treated with hydroxychloroquine.

A reusable pattern for graph-powered drug discovery

The architecture of CardioKG — imaging phenotypes as graph nodes, directed embeddings, multi-database fusion — is generalizable to any disease where imaging captures organ-level phenotypes: neurological conditions via brain MRI, liver disease via abdominal imaging, musculoskeletal disorders via joint imaging. The team has made the full codebase and analysis scripts publicly available on GitHub and Zenodo.

Within Bayer’s R&D pipeline, the patient-centric knowledge graph approach aligns with broader investments in agentic AI and graph-powered intelligence, and our overarching strategy to capitalize on differentiated and proprietary primary patient data to enable more accurate predictions of the outcomes of therapeutic interventions. Bayer’s existing Neo4j deployment supports active work in knowledge graph construction, GraphRAG, and semantic layer development — positioning knowledge graphs as a reusable patient data integration layer, coupled with state-of-the-art Machine Learning methodology, rather than a one-off research project.

“This multi-model data approach powered by machine learning has the potential to accelerate the timeline to find treatments for patients with heart disease, the number one killer worldwide.” — Nature Cardiovascular Research, Editorial (January 2026)