AOP-Wiki Key Event Components and assays: LLM extraction, cross-model consensus

Cross-model consensus of LLM-extracted Key Event Components (KEC) and assay descriptions for AOP-Wiki. Supplementary data for "From AOP-Wiki to AOPMAPPER: trust-aware AI-assisted enrichment of Adverse Outcome Pathway knowledge". Archived data: doi:10.5281/zenodo.23554426. The curated results are used to enrich AOPMapper. Code that generates all of it: github.com/ideaconsult/aopmapper-curation.

What you are looking at

Large language models read the free text of AOP-Wiki Key Events (KEs) and extracted structured knowledge from it. The outputs of several models were compared and grouped, so that what the models extracted independently ends up in one cluster. The support of a cluster is the number of models that produced it - the more models agree, the more it can be trusted.

• Key Event Components (KEC) are triples of a biological object, a process and an action (for example a protein, a process it takes part in, and how it changes). AOP-Wiki already holds curated triples for many KEs, so each cluster is compared with them: confirmed (the curated triple was also found by the models), missed (a curated triple no model found) or novel (found by the models, absent from AOP-Wiki).
• Assays describe how a KE is measured. They are extracted from the free-text measurement methodology of AOP-Wiki Key Events: assay name, method type, measured endpoint and its direction, throughput, ontology terms and the sentence of the source text that supports it. AOP-Wiki has no field for structured assay information and no curated assays, so there is no confirmed or missed status; clusters are read by support and by the quoted evidence. The records are exported in a proposed XML schema based on the Methods2AOP recommendation for per-assay metadata.

The pilot (116 Key Events of OECD-endorsed AOPs, AOP-Wiki release 2026-04-01) compared 72 models and was used to choose the consensus ensemble. The full corpus (AOP-Wiki release 2026-07-01) was then processed with the ten selected models, which form the consensus.

Key Event Components, full corpus (AOP-Wiki 2026-07-01)

• KEC consensus view
  Interactive page for all Key Events. For each KE: the curated AOP-Wiki triples (green when confirmed by a cluster of model triples, amber when missed) and the novel clusters (blue), each with its support bar and the models behind it. A panel summary and a support histogram are at the top.

Assays, full corpus (AOP-Wiki 2026-07-01)

• Assay consensus view large page
  One collapsible block per Key Event, with the strongest cross-model agreement first. Each assay cluster has a support bar and a table with one row per model: assay name, method type, endpoint, direction (with its confidence margin), throughput and the quoted evidence from the source text.

Key Event Components, pilot (AOP-Wiki 2026-04-01)

• KEC116 pilot consensus view
  The same view for the 116-Key-Event pilot, where 72 models were compared. The models included are listed in the page's panel summary.

XML exports (download)

• kec_consensus.xml (open rendered) about 12 MB
  The KEC consensus (full corpus) as schema-valid XML.
• assays.xml (open rendered) about 9 MB
  The curated assays as schema-valid XML, in the proposed assay schema.

Consensus data (download)

• viz_assay_consensus.json about 37 MB
  The assay consensus behind the assay view: clusters, per-model members and evidence. Token log-probabilities are reduced to the endpoint-direction margin.

Model panel (download)

• model_panel.tsv
  The 72 models compared in the pilot, with the ranking criteria; the ten selected for the consensus are marked in the Selected column.

Schemas and stylesheets (download)

• methods2aop_schema_proposal.xsd
  XSD 1.1 schema for the curated assays: a proposed extension of the AOP-Wiki XML, based on the Methods2AOP recommendation.
• kec_consensus_schema.xsd
  XSD 1.1 schema for the KEC consensus.
• assays.xsl and kec_consensus.xsl
  XSLT stylesheets that render the two XML files as HTML.

Validate with an XSD 1.1 validator, for example Python xmlschema.

Licence: CC BY-SA 4.0 (see LICENSE). Some pages are large and may take a moment to load.