Subtrack 1 — Global Equivalence Alignment

The traditional, whole-ontology alignment setting of Bio-ML 2026 (Track 1 — Equivalence).

For each of the three ontology pairs you submit one complete equivalence alignment between the source and target ontologies, in the OAEI Alignment RDF format, using full OWL IRIs. Submissions are scored organiser-side against a hidden test reference; the headline number is the repaired, coherence-aware F1, with the standard F1 and a reasoner-based Global Coherence figure reported alongside.

The three ontology pairs

Pair	Source → Target	Reference used for scoring
`NCIT-DOID`	NCIt → DOID	repaired (headline) + standard
`SNOMED-FMA`	SNOMED CT → FMA	repaired (headline) + standard
`SNOMED-NCIT`	SNOMED CT → NCIt	repaired (headline) + standard

The source ontologies are not re-hosted in the OAEI-ML/bio-ml dataset — obtain each from its original publisher (see the ontologies page and the dataset’s own ontologies.md).

A semi-supervised setting

The equivalence reference is split source-stratified 60 / 10 / 30 into train / validation / test. The public training slice is released as refs_equiv/train.tsv (one SrcEntity, TgtEntity, Score correspondence per line, full IRIs); the validation and test slices are held back and scored organiser-side. You may train, tune, and threshold on the public correspondences however you like; you then submit a full alignment over the two ontologies, and the organisers score the portion that falls on the hidden test entities.

Because the test reference is hidden, Subtrack 1 is scored organiser-side. Participants validate their submission’s format locally (see submission format) and submit; the leaderboard reports the scores.

Two references, two semantics

The gold equivalence reference is derived from UMLS/Mondo and is therefore, as published, potentially logically incoherent. Every submission is scored against two views of it:

Standard reference — traditional P/R/F1. The complete (possibly-incoherent) reference, scored with ordinary set-based precision, recall and F1 over equivalence correspondences.

Repaired reference — coherence-aware, relation-agnostic P/R/F1 (headline). The reference after mapping repair. Correspondences that repair fully removed are annotated ? and are ignored from both the precision and the recall denominators (they neither help nor hurt). Correspondences that repair weakened rather than removed survive as subsumption (< / >), and are credited relation-agnostically — a reference < or > is satisfied by a predicted correspondence of any relation between the same entities. This is the headline metric.

\mathrm{Precision} = \frac{\vert A \cap R^{+} \vert}{\vert A \setminus R^{?} \vert}, \qquad \mathrm{Recall} = \frac{\vert A \cap R^{+} \vert}{\vert R^{+} \vert}, \qquad \mathrm{F1} = \frac{2\,\mathrm{P}\,\mathrm{R}}{\mathrm{P} + \mathrm{R}},

where $R^{+}$ are the surviving (credited) reference correspondences and $R^{?}$ the ?-flagged ones removed from both denominators. Both P/R/F1 sets are macro-averaged across the three pairs.

How the repaired reference was constructed

Repair follows the LargeBio “remove-if-any” convention, computed as the union of removed across three established repair tools — ALCOMO (run under the ELK reasoner), LogMap-repair, and AML-repair. A correspondence is dropped (annotated ?) if any of the three removes it; where LogMap merely weakens an equivalence to a subsumption, the weakened < / > is kept (subsumption takes priority). owl:deprecated correspondences are out-of-task and dropped from both the reference and the predictions before scoring. Full construction detail is deferred to the supplementary materials (available at track launch).

Global Coherence

Alongside P/R/F1, each submitted alignment is checked for the logical incoherence it induces when merged with the two ontologies. A reasoner computes global_coherence — the degree of incoherence in $[0, 1]$ (lower is better) — together with the number of unsatisfiable classes and the size of the merged class set. Because it requires a reasoner, Global Coherence is computed organiser-side only; it is never part of the participant-side kit.

Scoring and leaderboard columns

Subtrack 1 is scored organiser-side and published to the Track 1 — Global Alignment CodaBench leaderboard. Its columns are macro_f1_repaired (headline), macro_precision_repaired, macro_recall_repaired, macro_f1_standard, global_coherence, and the per-pair f1_repaired_<pair>. Organiser baseline numbers are on the baselines page.

For the exact file shape, a worked example, a copy-paste template, and local validation, see the submission format.