InChIKey-Deduplicated ClassyFire/ChemOnt Label Collection

This dataset is an InChIKey-deduplicated collection of ClassyFire/ChemOnt chemical class labels assembled from multiple independently generated ClassyFire-labelled resources contributed by several laboratories and collaborators. To the best of our knowledge, this is the largest collected and openly distributable set of ClassyFire labels currently available. The released table contains 73,105,281 unique compounds. Every row contains a full InChIKey, one representative SMILES string, and numeric ChemOnt class identifiers. Where available, rows also include external database identifiers: 68,693,298 compounds have a PubChem CID, 30,187,323 compounds have a ZINC identifier, and 25,775,344 compounds have both. Class labels are stored as numeric ChemOnt IDs to keep the dataset compact and practical for large-scale machine-learning workflows. The main hierarchy path follows the ClassyFire hierarchy and is encoded as numeric identifiers for kingdom, superclass, class, subclass, and direct parent. When no label is available for a given hierarchy layer, the corresponding value is encoded as null. TSV format The main release file is a tab-separated table with one row per unique InChIKey. Columns: inchikey: full 27-character InChIKey used as the deduplication key. cid: PubChem CID, when available. Empty otherwise. zinc_id: ZINC identifier, when available. Empty otherwise. smiles: one representative SMILES string for the InChIKey. classification_json: numeric ChemOnt labels. Rows use a five-position hierarchy path corresponding to [kingdom, superclass, class, subclass, direct_parent]; missing layer values are encoded as null. Compact externally contributed single-ID rows were excluded from this release. sources: source provenance flags for the retained row and identifiers, such as pubchem, zinc, or pubchem+zinc. The accompanying ChemOnt dictionary TSV maps each numeric ID to its CHEMONTID and label name. In that dictionary, parent_numeric_id is the parent ChemOnt numeric ID; null means that the term has no parent. The value 0 is not missing, but the real numeric ID for CHEMONTID:0000000, Organic compounds. Dataset summary Unique InChIKeys / compounds: 73,105,281 Rows with representative SMILES: 73,105,281 Rows with PubChem CID: 68,693,298 Rows with ZINC ID: 30,187,323 Rows with both PubChem CID and ZINC ID: 25,775,344 ZINC-only new InChIKeys: 4,411,979 ChemOnt class coverage Coverage remains highly uneven across ChemOnt. Counts below refer to ChemOnt classes with at least the indicated number of InChIKey-deduplicated examples in this release. Total ChemOnt classes considered: 4,825 Classes with no examples: 1,026 Classes with exactly 1 example: 99 Classes with at least 1 example: 3,799 Classes with at least 10 examples: 3,432 Classes with at least 100 examples: 2,751 Classes with at least 1,000 examples: 1,766 Classes with at least 10,000 examples: 854 Classes with at least 100,000 examples: 283 The ontology-wide zero-count total and the hierarchy-layer missing totals answer different questions. There are 1,026 ChemOnt dictionary labels that never occur anywhere in the released classification_json values. Of these, 59 are labels that belong to the retained five-position ClassyFire path but have no examples in this release. The other 967 are valid ChemOnt terms, but they are not part of the supervised target set used here. In practical terms, these 967 terms are present in ChemOnt but outside the retained five-position ClassyFire classification path: kingdom, superclass, class, subclass, and direct_parent. No label from this outside-path set appears in the released dataset; every ChemOnt label with at least one example appears in one of the five retained hierarchy positions. Separately, there are 80 unique labels missing from at least one normalized hierarchy-layer reference set. This layer-specific missing-label set includes the 59 zero-example target labels above, plus 21 labels that do occur elsewhere in the five-position path but are absent from one specific layer. The 967 outside-path ChemOnt terms should therefore be treated as unused or auxiliary ontology terms for this release, not as missing supervised target classes. They are provided in the zero-count ChemOnt class TSV rather than repeated in the layer-coverage section. Hierarchy layer coverage The counts below summarize coverage for the five-position hierarchical ClassyFire path after normalizing root-encoded rows back to the kingdom, superclass, class, subclass, and direct_parent positions. All rows in the filtered release contain a five-position hierarchy path; compact externally contributed single-label rows were excluded. Kingdom: Reference labels: 2 Observed labels: 2 Labels with no examples: 0 Rows with label: 73,105,281 Percent of hierarchical rows: 100.000000% Missing labels: none Superclass: Reference labels: 32 Observed labels: 32 Labels with no examples: 0 Rows with label: 73,105,228 Percent of hierarchical rows: 99.999928% Missing labels: none Class: Reference labels: 649 Observed labels: 642 Labels with no examples: 7 Rows with label: 72,985,494 Percent of hierarchical rows: 99.836145% Missing labels: Thiolactones; Organo-alkali-metal compounds; Sulfinohydroximic acids; Sulfonohydrazonamides; Carboperoxoselenoic acids; Furo[3,4-d]pyrimidine glycosides; Carbides Subclass: Reference labels: 2,811 Observed labels: 2,766 Labels with no examples: 45 Rows with label: 62,098,531 Percent of hierarchical rows: 84.943974% Missing labels: Organolithium compounds; Organogold compounds; Actinide hyponitrites; Actinide hypophosphites; Actinide superoxides; Alkaline earth metal hyponitrites; Lanthanide hypophosphites; Lanthanide peroxomonosulfates; Lanthanide peroxodisulfates; Metalloid hyposulfites; Alkali metal ferrates; Lanthanide arsenites; Organometalloid compounds; 1-phosphoribosyl-imidazoles; Betalains; Intermetallic post-transition metal compounds; Miscellaneous antimonites; Thiazetidines; Azomethine ylides; Sulfinohydroximic acids; Dalbergiquinols; Pseudoguaianes; Ophiobolane sesterterpenoids; Thieno[3,4-c]pyridazines; 4’-prenylated 2-arybenzofurans; Phenanthro-perylenequinones; Imidoyl carbenes; Phosphonated cyclopropyl nucleosides; (3’->5’)-dinucleotide phosphorodithioates; 2’,3’-cyclic pyrimidine nucleoside phosphorothioates; Cyanide salts; Inorganic lead salts; Plicamine-type amaryllidaceae alkaloids; Sulfur-containing organoarsenic compounds; 6,7-dihydropteridines; Organolead sulfides; 4-Anilinoquinolines; Phosphiranes; Nitrimines; Metal cyanates; Dialkylzincs; Organomagnesium halides; Dialkylmagnesiums; Diazonium perchlorates; Organometallic peroxides Direct parent: Reference labels: 3,680 Observed labels: 3,618 Labels with no examples: 62 Rows with label: 73,105,274 Percent of hierarchical rows: 99.999990% Missing labels: Organolithium compounds; Organogold compounds; Actinide hyponitrites; Actinide hypophosphites; Actinide superoxides; Alkaline earth metal hyponitrites; Lanthanide hypophosphites; Lanthanide peroxomonosulfates; Lanthanide peroxodisulfates; Metalloid hyposulfites; Alkali metal ferrates; Lanthanide arsenites; Epoxyeicosatrienoic acids; Alkyl fluorides; Thiolactones; CDP-monoacylglycerols; Organo-alkali-metal compounds; 1-phosphoribosyl-imidazoles; Miscellaneous ferrates; Purine 3’-deoxyribonucleoside polyphosphates; Azomethine ylides; Telluroethers; Seleniranes; Sulfinohydroximic acids; Dalbergiquinols; Dihydropteroic acids and derivatives; S-acylglutathiones; Aromatic alcohols; Sulfonohydrazonamides; Carboperoxoselenoic acids; 1-alpha,2-epoxy-3-oxo-5-alpha-steroids; 3-beta-hydroxy-6-oxo-5-alpha-steroids; N-nitrosated N-hydroxyguanidines; 2,3-dienoyl fatty acyl thioesters; Thieno[3,4-c]pyridazines; 4’-prenylated 2-arybenzofurans; Phenanthro-perylenequinones; Imidoyl carbenes; 7-prenylated flavans; Phosphonated cyclopropyl purine nucleosides; Phosphonated cyclopropyl pyrimidine nucleosides; Furo[3,4-d]pyrimidine glycosides; (3’->5’)-cyclic dinucleotide bis(phosphoromonothioate) analogues; (3’->5’)-dinucleotide phosphorodithioates; 2’,3’-cyclic pyrimidine nucleoside phosphorothioates; 4-Hydroxymercuribenzoic acids; Triorganotin metal salts; Plicamine-type amaryllidaceae alkaloids; Thiohydroximic acids and derivatives; Very long-chain 3-enoyl CoAs; 6,7-dihydropteridines; Glucosylceramides; Deoxyribo- and ribonucleoside phosphonates; 4-Anilinoquinolines; Substituted imidazole-4-acetic acid and analogues; Carbides; Aziridinylquinones; Dialkylzincs; Organomagnesium halides; Alkylmagnesium halides; Arylmagnesium halides; Organic thionitrous acids Warning: Despite the size of this release, many ChemOnt classes remain orphan classes, meaning that no compound in this collection is currently assigned to them. Other classes remain singleton or near-singleton classes, represented by only one or very few examples. This dataset should therefore be considered a large but incomplete snapshot of currently recoverable ClassyFire-labelled chemical space, not a complete coverage of the ChemOnt ontology. Provenance and limitations We attempted to preserve and trace label provenance to the best of our ability. However, because these labels reached us through an aggregative chain, with some laboratories sharing labels that had themselves been assembled from labels collected by other collaborators, the exact origin of every individual label is unfortunately not always recoverable. It remains unclear whether all contributed labels were generated through the public ClassyFire APIs, through other released ClassyFire software components, through database snapshots, through the original ClassyFire authors’ large-scale classifications, or through other local aggregation procedures. This release certainly includes the labels published in 2016 by the ClassyFire authors on their website, as well as the CANOPUS ClassyFire labels, but the precise lineage of all additional labels cannot be guaranteed. Warning: The current public ClassyFire API is difficult to use for large-scale verification. In our experience it is flaky, changes behavior over time, and is practically limited to approximately 12 requests per minute. Because the labels aggregated here were generated across different laboratories and at different times, we could not exhaustively revalidate every assignment against the current public service. Users should therefore treat the labels as a large community-aggregated resource, with the possibility of historical version differences or occasional classification errors. Where different laboratories provided overlapping labels, often themselves assembled from labels originally collected by additional collaborators, we did not observe disagreements after matching by the same normalized InChIKey. We did, however, encounter cases where structure or InChIKey normalization issues could make chemically related or tautomeric molecules appear under different InChIKeys, and these cases could be associated with substantially different ChemOnt assignments. For this reason, the dataset should be interpreted as InChIKey-resolved rather than as a fully tautomer-normalized chemical identity resource. The large PubChem coverage of this release may also provide a clue about label origin. The original ClassyFire publication reports PubChem-scale classification, including more than 60 million PubChem molecules and more than 77 million classified compounds overall. Since this release contains 68,693,298 PubChem-linked compounds, our hypothesis is that a substantial fraction of the PubChem-labelled portion may overlap with, or derive from, the PubChem-scale ClassyFire submission described in that work, which to our knowledge is not clearly downloadable from PubChem, but we suppose might have been downloadable at some point in the past. Note: We do not have proof of this PubChem-scale provenance. This interpretation is based only on the compatible order of magnitude and PubChem coverage. Broader motivation We hope this dataset will facilitate future open-source model development and make large-scale ClassyFire/ChemOnt prediction more accessible. More broadly, we hope that collecting and releasing these labels will support more scientific, transparent, and reproducible research in chemical structural classification, an area where large reusable reference datasets and auditable modelling efforts have historically been limited. This release is intended as a twin resource to our full PubChem NPC classification release, available at https://zenodo.org/records/19701309. We are actively working on both SMARTS-based classifiers, including genetic-algorithm approaches developed in earth-metabolome-initiative/smarts-evolution, and neural-network models for chemical structural classification. We invite researchers who have additional ClassyFire labels, especially labels for underrepresented, orphan, or singleton ChemOnt classes, to contact us so that future versions of this collection can be expanded and improved. We also welcome additional labelled SMILES or structure-to-ClassyFire mappings from trustworthy sources. If new reliable labels become available, especially for currently underrepresented ChemOnt classes, we intend to incorporate them and publish updated releases of this dataset. We thank all contributors, their collaborators, and the groups who generated and shared the underlying labels that made this aggregation possible.