Biomedical ontology alignment as typed link prediction
Systems receive a source concept and a fixed set of target candidates, then rank candidate–relation pairs for equivalence or directional subsumption — one familiar interface for graph, text, and neuro-symbolic methods.
Source and target denote the same disease. Under the preferred typed metric this is the only fully relevant prediction — entity and relation both match the reference — so a strong system ranks it first.
Illustrative example (NCIT → DOID). The live site renders from the public release; hidden test data is never shown.
One graph, three tasks, two primary metrics
BioKG-Align runs two tracks — the main track, summarised here, and a complex track on OWL class-expression generation (more in the FAQ).
- 5 track ontologies
- NCIT, DOID, ORDO, SNOMED, FMA — projected into one knowledge graph.
- 3 task pairs
NCIT-DOID,SNOMED-FMA,SNOMED-NCIT.- Primary metrics
- Preferred Typed (Relation-Aware) MRR and Hierarchy-Aware Typed nDCG@10.
- Candidate sets
- 50 target candidates × 3 relations = 150 scored rows per query.
- Three methodological entry points
- Triples → + node properties → + Datalog rules — graph, text, and neuro-symbolic methods.
The unified graph spans multiple projected ontologies with three task pairs.
- Input
- A source entity and a fixed candidate set, per query.
- Output
- A ranked list of (candidate, relation) pairs.
- Correctness
- A hit needs both the target entity and the relation to match the hidden reference.
Current competition status
Status: In preparation
- Starting Kit Release
- Late May 2026
- Dataset Publication
- Early June 2026
- Competition Proposal Acceptance Decision
- Mid June 2026
- Competition Starts, Leaderboard Opens
- Late June 2026 Provisional
- Competition Ends
- Late October 2026 Provisional