Biomedical knowledge graphs fuse ontology‑derived hierarchies with transversal links among heterogeneous entities such as phenotypes, diseases, genes, proteins, and patients, yielding a hybrid topology. This raises the question of whether hyperbolic embeddings—naturally suited for tree‑like structures—remain beneficial when the graph also contains many non‑hierarchical relations.
We conduct a preliminary study of hyperbolic graph representation learning for Mendelian‑disease differential diagnosis on a patient‑integrated biomedical graph. First, on isolated ontology subgraphs, hyperbolic models achieve strong predictive performance while using far fewer dimensions than Euclidean baselines, confirming their efficiency in encoding hierarchical information. Next, we evaluate the models on a link‑prediction task that ranks candidate diseases for each patient. Results indicate that hyperbolic embeddings can both exploit the biomedical hierarchy and support diagnostic reasoning over heterogeneous patient‑level graphs.
These findings suggest that hyperbolic space is a powerful tool for biomedical knowledge graphs with pronounced hierarchical structure, especially in low‑dimensional, differential‑diagnosis scenarios.
Review