ogbl-biokg

Predict missing biomedical triples by ranking the true head or tail above sampled replacements of the same entity type. The official metric is filtered mean reciprocal rank.

The aggregate graph has 93,773 diseases, proteins, drugs, side effects, and protein functions joined by 51 directed relation types. These include 38 drug-drug and eight protein-protein relations plus drug-protein, drug-side-effect, and function relations.

The standard task provides entity types, relation types, and graph structure but no unified molecular, sequence, or textual node feature matrix. Same-type relations are represented symmetrically with both directions.

Triples are randomly split because reliable timestamps for the underlying experiments and observations could not be recovered. Negative heads and tails are sampled only from the correct entity type and filtered against known triples.

The source repositories mix experimental measurements, curated annotations, and automatically extracted metadata with different confidence and possible contradictions. Random splitting can place closely related evidence across splits, and missing facts are not reliable negatives. External biomedical features must be disclosed.