Sources and licenses

Used for

Every bioactivity measurement on this site: Kᵢ, Kd, IC₅₀, EC₅₀, AC₅₀, potency and the other recorded readouts, together with assay metadata, target names and organism. Each measurement keeps its ChEMBL assay identifier, which is what makes the row traceable.

Used for

Compound identity: structure (SMILES), formula, molecular weight, IUPAC name, synonyms and the 2D depictions shown on each dossier. PubChem is the authority the identity audit checks every record against, and the structure descriptors are computed from the SMILES it returns rather than from the local copy. PubChem also aggregates the prose this site quotes: the descriptive record on each dossier’s Summary — what a compound is, what it is recognized for, what a body does with it and how it is scheduled — alongside the pharmacology and toxicology passages, each attributed to the depositor database that wrote it (DrugBank, LiverTox, HSDB, NCI Thesaurus, ChEBI, FDA, DEA) rather than to PubChem itself. Every passage is quoted whole; none is summarized.

Used for

Pharmacological classification on the Families page: which compounds are hallucinogens, dissociative anesthetics, muscarinic antagonists and so on. Each class is a MeSH descriptor carrying NLM's own definition, assigned to the compound by PubChem rather than inferred here. The previous build derived this axis from regular expressions over compound names, which is a taxonomy with no author and no citation; it is not used.

Used for

1,399 equilibrium binding constants determined under a single screening program, with the radioligand and tissue preparation each was measured against. 1,066 of them carry a PubMed identifier, so they resolve to the paper that reported the measurement rather than to a database record — until these were ingested, every reference_pmid in the archive was empty. Compounds are matched to PDSP's free-text ligand names only where the name appears verbatim in that compound's verified PubChem synonym list; a name two compounds both claim is refused rather than assigned.

Used for

Duration of acute subjective effects and plasma elimination half-life, on each compound's behavioral summary. Every value is a sentence quoted verbatim from the abstract of a named controlled trial or pharmacokinetic study, and each citation is re-resolved against PubMed on every audit run — an entry whose PMID stops resolving to the recorded paper is withheld rather than shown. Candidates are harvested automatically but promoted by hand, because a pattern match cannot tell a compound's half-life from one mentioned beside it. PubMed is also the only source here that is not a re-publisher: ChEMBL and PubChem carry a measurement a year or more after the laboratory reported it, and sometimes never, so where the databases lag, a paper is read and transcribed instead of showing nothing. That reaches two things. Measurements: 64 rows on this site are transcribed by hand from a named paper, each carrying the paper's own wording for the number and where in the paper it sits (tools/literature_measurements.json). And compounds: ten records exist here because a paper characterized them and ChEMBL had not curated the result — three halogenated tryptamines from a 2026 Molecular Psychiatry paper, and seven ring-deleted analogues of LSD from a 2026 PNAS paper — identity still from PubChem and audited weekly against it, admission and every value from the paper, said on the face of the dossier (tools/literature_compounds.json). What the paper cannot supply is a structure anyone can re-check, so the name the paper prints is parsed by OPSIN and has to return the structure PubChem holds for the cited record; two further analogues from the PNAS paper are declined on that rule, not because the evidence is thin — the paper names and characterizes them as fully as the seven — but because PubChem holds no record of either structure, and a structure held only here would sit permanently unaudited while looking like every audited record. PubMed also supplies the bibliography on each dossier's Literature section: every PMID this archive cites anywhere — on an assay row, beside a quoted passage of pharmacology or toxicology, or against a curated duration — is resolved once to its title, journal, year and first authors. 330 identifiers resolve, and 72 of the 224 compounds carry at least one; the rest cite no publication and their dossiers say so. What a paper contributed is computed from this archive's own rows, and no summary of a study's content is written anywhere on this site.

Used for

The Trials plate on each dossier: how many registered studies name the compound as an intervention, their phases and statuses, and the most recent of them with enrollment, sponsor and conditions. Every row resolves to its own NCT record. Nothing is aggregated across studies and no outcome is reported — a registered trial is a protocol somebody filed, not a result, so withdrawn and terminated studies are counted alongside the rest rather than dropped to make a compound look better studied than it is.

Used for

The bibliography in each dossier's History section, and nothing else. No database this site reads publishes a history heading, but the history of these compounds is written at length under peer review, and reporting an absence in our sources as though it were an absence in the literature was misleading. So the works are cited rather than paraphrased: this site does not summarize them, quote them, or state what they conclude. A book is attached to a specific compound only where the work's own title names it — Mike Jay's Mescaline appears on mescaline's dossier for that reason and on no other — because deciding what a book covers without opening it is the same invention the rest of this build refuses. Every author, imprint, year, ISBN and DOI was checked against the publisher, a stable library record or a review in an academic journal; the MIT Press volume was confirmed independently through CrossRef. Curated by hand in tools/history_sources.json.

Used for

The DRIFT grid on the Compound–target analysis page, and nothing else. DRIFT ranks targets by FP2 fingerprint similarity between a compound and the known ligands of each target, which is a statement about chemical structure rather than about binding — so it is shaded on its own 0–1 axis, never on a pX axis, and never merged with a measurement. DRIFT also returns an estimated pK for some rows; the service reports that these estimates do not affect its ranking, and this site shades none of them. Where an estimate exists it appears in the cell readout marked as an estimate. Just over half of what DRIFT returns names something that is not a molecular target — cell lines, organisms, tissue preparations, and the placeholders “No relevant target”, “Unchecked” and “ADMET” — and all of it is excluded by the same rule that keeps such rows off the measured grid.

Used for

What each target is, on the target pages: protein name, gene, entry name, and the FUNCTION annotation — quoted verbatim, with UniProt’s own inline PubMed markers left in the sentence rather than stripped out of it. Curated subcellular locations come from the same record. 159 of the archive’s 964 target names resolve to an accession; the rest are cell lines, organisms and assay descriptions, or names no gene lists as a synonym, and they show nothing rather than a guess.

Used for

Where a target is expressed, in three separate readouts that are never merged: consensus tissue RNA (nTPM), which shades the anatomy plates on each compound dossier; single cell type RNA (nCPM), which gives the cell-type bars on a target page; and subcellular location by antibody immunofluorescence, which is a different assay from both and is labeled as such. Transcript is not protein, and a target’s gene being transcribed in a region is not evidence that a compound reaches it.

Used for

The highest-confidence human functional partners of each resolved target, on its page. STRING’s combined score mixes evidence channels, so experiment, database and text mining are shown as separate columns beside it — 5-HT2A’s top partner GRM2 scores 0.998 combined but 0.34 experimental against 0.997 text mining, and collapsing that into one number would hide which kind of claim it is.

Used for

Descriptor calculation (molecular weight, exact mass, cLogP, TPSA, hydrogen-bond counts, rotatable bonds, sp³ fraction) and Bemis–Murcko scaffold derivation. Everything RDKit produces is labeled as computed, never as measured.

Used for

Reads a systematic chemical name into a structure, as a gate rather than as a value: it produces nothing this site displays. Where a compound is held because a paper characterized it, the name the paper prints is handed to OPSIN and the compound is refused unless the structure that comes back carries the InChIKey PubChem holds for the cited record. That check replaced an earlier one that required the paper's name to appear verbatim in PubChem's synonym list, which refused compounds for having been published too recently to be indexed and, when it passed, established only that a string occurred in a list. It can only ever refuse: admission needs PubChem to return a structure and OPSIN to return the same one from the paper's own words.