Chemical Identity Lost in Regulation: code and data
Geert Van Haute, Stijn Goedertier, Pieter Fannes, Maxim Van de Wynckel
Code and data for "Chemical Identity Lost in Regulation: A Study of Semantic Interoperability in European Chemical Substance Data" (Van Haute, Goedertier, Fannes, Van de Wynckel), Poster & Demo track, SEMANTiCS 2026, Ghent. European regulatory datasets represent chemical substances heterogeneously — as single molecules, mixtures, UVCBs, substance groups, or analytical sum parameters — identified only by CAS numbers and free-text names rather than structure-based identifiers. This repository integrates 18 ECHA-based datasets (41,813 records) under a seven-tier linkability taxonomy: only 62.5% of 33,452 unique entries link to a defined molecular structure via InChIKey, and CAS numbers prove unreliable as cross-dataset identifiers (70% of structure-resolved InChIKeys occur in only one source). Sentence-embedding models fail to substitute for structure in ChemOnt classification (Hit@1 ≤ 5.75%); Claude Sonnet 4.6 reaches 39.2% but still misclassifies the majority. Structure-defined substances are integrated into an RDF knowledge graph (SKOS, XKOS, ChEBI, ~48,400 cross-domain annotation links), with a separate LLM-based scope-mapping step yielding 304 validated SKOS crosswalk triples between regulatory groups and ChemOnt classes. Contents: the R/Python analysis pipeline, the SPARQL/ROBOT-based RDF integration scripts, the resulting knowledge graph (data/processed/rdf/), and the poster/paper sources. Licensed CC BY 4.0.