InChIKey-Deduplicated ClassyFire/ChemOnt Label Collection

The largest labelled map of chemical space this vertical has. InChIKey-Deduplicated ClassyFire/ChemOnt Label Collection attaches ClassyFire functional classifications, expressed against the ChemOnt 2.1 ontology, to 73,356,229 structurally unique compounds — version v4, published June 2, 2026. Three upstream libraries feed it: PubChem contributes 68,693,295 rows carrying a CID, ZINC20 contributes 30,187,323 rows carrying a ZINC ID, and Enamine REAL adds 250,948 screening-compound rows, with 25,775,344 compounds overlapping the two major sources. Because deduplication runs on the InChIKey hash rather than on database entries, the collection holds one row per unique molecule — the same structure cannot inflate a training set three times over.

Every compound carries the full five-tier classification path — kingdom, superclass, class, subclass and direct parent — plus four label columns reaching beyond the strict path. Coverage spans 4,120 of ChemOnt's 4,824 classes, or 85.4%. Get a sample of this dataset scoped to the classes you actually model.

No payload tour needed — the schema is the story. Every row in the Parquet frames presents the same twelve-column silhouette, and the label columns arrive natively as uint16 arrays rather than delimited strings, so reading the identity columns never touches hundreds of millions of label entries:

CLASSYFIRE/CHEMONT LABEL COLLECTION v4 — ROW SHAPE AS TYPED IN THE PARQUET FRAMES
cid       : int64            (present on PubChem-derived rows)
zinc_id   : uint32           (present on ZINC-derived rows)
smiles    : string           (structure string)
inchikey  : string           (deduplication key, one row per unique compound)

kingdom_ids              : list<uint16>  -- decodes via vocabulary.json
superclass_ids           : list<uint16>
class_ids                : list<uint16>
subclass_ids             : list<uint16>
direct_parent_ids        : list<uint16>
intermediate_nodes_ids   : list<uint16>
alternative_parents_ids  : list<uint16>
substituents_ids         : list<uint16>
mapped_features_ids      : list<uint16>

PARTITION PLAN (80/10/10, STRATIFIED BY DEEPEST RESOLVED CHEMONT SLOT)
train.parquet       58.7 M rows
validation.parquet   7.3 M rows
test.parquet         7.3 M rows

What the rows establish: identity travels on the InChIKey hash — the same key the Kaggle LogP compilation ships — so enrichment joins never depend on spelled-out names. Each label column decodes through vocabulary.json into ChemOnt terms, and the ChemOnt dictionary resolves numeric ids to ids, names and parents; a decoded path climbs from kingdom through superclass, class, subclass and direct parent, with alternative placements and structural features riding alongside. Request a sample cut to named classes and the same columns arrive filtered, hierarchy intact.

Thirteen fields define every row, all marked verified against the release documentation during the August 2026 research pass. Four carry identity: cid (int64) on PubChem-derived rows, zinc_id (uint32) on ZINC-derived rows, the SMILES structure string and the InChIKey hash doing the deduplication. Nine carry labels, packed as list<uint16> index arrays: the five tiers of the strict classification path — kingdom, superclass, class, subclass, direct parent — then intermediate nodes along the route, alternative parents where a structure legitimately belongs to more than one branch, and substituent and mapped-feature vocabularies that work as auxiliary multi-label targets.

The reference files that make those indices readable travel alongside the frames and fold under additional fields on request: vocabulary.json mapping indices to ChemOnt terms, the dictionary resolving numeric ids to names and parents, the ChemOnt 2.1 ontology for full ancestor closure, and the non-hierarchical label column. Their handling is confirmed with you before anything ships.

• Geography: none, deliberately. This is chemical space, not economic space — a compound carries a structure, not a country. Geography enters downstream, when class labels feed regional portfolios, sourcing screens or jurisdiction-specific compliance work layered on records such as ECHA CHEM.
• Temporal: a pinned snapshot. Version v4 was published June 2, 2026 — a fixed edition rather than a moving feed, which is exactly what reproducible modelling wants: pin the version string and every result stays traceable to the exact labelled bytes you trained on.
• Granularity: one row per unique InChIKey compound — 73,356,229 of them — pre-partitioned into an 80/10/10 train/validation/test split stratified by the deepest resolved ChemOnt slot (58.7M / 7.3M / 7.3M rows), so compounds cannot leak across partitions the way they do when teams informally re-split a shared corpus. One subtlety belongs in your data card: deduplication skips tautomer normalization, so tautomers may surface as separate entries — collapse them before splitting if tautomer state affects your task.

API, files, or your warehouse. Daily, weekly, or hourly.

You pick the channel and the cadence; the thirteen-field dictionary above travels unchanged across all three. Teams training models tend toward the Parquet frames loaded once beside their feature store; product surfaces embedding class lookups take a scoped feed; SQL-first shops take warehouse load with the label indices resolved into readable ChemOnt terms. Name the classes, tiers or upstream libraries when you request the sample — the sample ships first either way, and changing cadence afterward is a settings conversation, not a re-integration project.

A labelled corpus at this scale answers one question millions of times over — what chemical class is this structure? — and that single answer composes into unrelated workflows:

• Property and activity modelling — class labels serve as categorical features or multi-label auxiliary targets beside property endpoints; joined to Kaggle's 426k PubChem compilation on the shared InChIKey spine, they turn a lipophilicity regression into a chemistry-aware one.
• Leakage-resistant benchmarking — the stratified 80/10/10 partition arrives pre-built, so train/test contamination stops being the finding reviewers catch.
• Portfolio and supplier screening — classify a procurement catalog or a competitor's substance list into ChemOnt classes and see which branches of chemistry the company actually lives in, not which ones its marketing claims.
• Impact-factor proxies — life-cycle assessment figures grouped by the same ClassyFire/ChemOnt taxonomy join on class labels directly, composing hazard, property and impact layers into one feature store.
• Regulatory cross-referencing — attach ontology-backed classes to substance inventories such as ECHA CHEM, adding a structural dimension to compliance lists that stop at substance identity.
• Search and similarity products — the substituent and mapped-feature vocabularies give structured facets for browsing and filtering chemical catalogs.

The modelling workflows live in the ML model training use case and the citation-grade research use case.

Data scientists and ML engineers get the rarest artifact in applied chemistry ML: a corpus whose split, labels and identity key were designed together, so no derivation leaks silently into the pipeline. Developers and data-product builders get class lookups at catalogue scale without operating a classifier themselves. Competitive-intel and product teams get portfolio maps that place any substance list onto ChemOnt's branch structure. Journalists, academics and students get coverage statistics — 73.4 million compounds, 85.4% of classes observed — attributed to a citable versioned release. All delivered daily, weekly, or hourly.

Provenance note — the collection compiles PubChem, ZINC20 and Enamine REAL structures classified by ClassyFire against the ChemOnt 2.1 ontology, distributed as the versioned Zenodo release v4 (June 2, 2026). It is the only Parquet-native distribution among the twelve primary specialty-chemicals datasets Datadory catalogs.

Label-trust note — the compiler states plainly that a complete re-check of every label against the live classification service is impractical at this scale. Audit a random sample, document the caveat, and report benchmarks as trained on the v4 snapshot rather than on freshly regenerated labels.

Completeness note — 4,120 of 4,824 ChemOnt classes carry examples; 704 hold none, concentrated in exotic inorganic, lanthanide and actinide chemistry the upstream libraries barely sample. Treat classes outside the observed set as out-of-scope for training and reporting rather than as defects to fix.

Where to go next — for property targets on the same identity key, Kaggle — 426k PubChem Molecules for LogP Prediction; for impact factors grouped by this taxonomy, Dryad - LCA Impact Data for Chemicals (proxy classification); for the regulatory mirror, ECHA CHEM - EU Chemicals Regulatory Database (REACH & C&L); and for evaluated physical constants behind the structures, NIST Chemistry WebBook. Background reading: SMILES notation, PubChem-derived dataset, the ClassyFire/ChemOnt labels deep dive, the specialty chemicals data guide and the ranked shortlist of the best specialty chemicals datasets.