How this atlas was made

What is generated and what is computed

Two different things sit on every family page, and they have different reliability.

Computed from the matrices — the member list, the data consensus, the representative motif, the cohesion value, the evidence tags, the cross-species transfer flags, and every entry on the label checks page. These are arithmetic over the 4,232 motifs in JASPAR 2024, HOCOMOCO v14, CIS-BP and the Vierstra archetypes. They are reproducible from the repository and do not involve a language model.

Written by a model — the short summary, the long entry, the caveats, and the choice of which papers to cite. Each family was written by an agent that read its computed data file first, then searched the literature and wrote from what it retrieved.

What is checked

• Every citation. All 1,780 distinct PMIDs behind the 2,090 references resolve against PubMed, and each one names an author the citation claims. The printed citation text is then regenerated from PubMed's own record, so what you read cannot drift from the identifier beside it. A wrong identifier surfaces as a mismatch rather than sitting there looking plausible.
• Structure. Every entry carries the five required sections in order, within set word budgets, with at least six references.
• Claims about the motifs. Where an entry states a consensus, it is compared with the consensus computed from the member matrices. 30 families state that they have no single consensus rather than being given one, and the pages and the map show that instead of substituting the medoid. Statements about a named model's evidence type, a family's cohesion or a motif's information content are linted against the computed values.
• Retractions. Every cited PMID is checked against PubMed's retraction flag. Citing a retraction notice is allowed — an entry may need to say a study was withdrawn — but citing the withdrawn work is not.
• Every anchored sentence, against its source. A second pass tied each substantive claim to the paper that supports it, marked inline, and read that paper to do so. It flagged 106 problems in 70 entries: 66 where the prose contradicted its source or the atlas's own data, 15 overstatements, 17 anchors to a paper that did not actually state the claim, and 6 claims nobody could source. 101 were corrected, and 5 flags turned out to be wrong themselves and were rejected. One retracted paper was found cited and removed. The full list is published, problem by problem.

What is not checked

• Sentences without an anchor. Anchored sentences have been read against their source, but framing sentences and some textbook background carry no anchor, and those rest on the model's training rather than on a retrieved paper.
• Abstracts, mostly. Most anchors were checked against the abstract, because many papers are not open access. A claim drawn from a paper's main text but absent from its abstract can still be mis-stated.
• Claims with no citation. Several entries state textbook background without a reference, because the reference list is capped at twelve. Those sentences rest on the model's training, not on a retrieved source.
• Completeness. No claim is made that an entry covers the important literature for its family, only that what it cites is real.

Known weak points

• 18 families are buckets, not motifs. Their members do not resemble one another (cohesion below 0.5), and the KRAB zinc-finger, unnamed-archetype and "other" entries say so in their first sentence. A hit to one of these identifies a kind of thing, not a site.
• 631 of 1595 HOCOMOCO matrices are ChIP-only. A ChIP-derived matrix records where a protein was crosslinked, which need not be a sequence it binds itself. Those are flagged in amber in every member table.
• 83 CIS-BP entries are inherited from a non-human protein, down to 55% DNA-binding-domain identity — the human RFX8 motif comes from a louse. Each is labelled with its species and identity.
• 53 motifs sit under a label their own matrix contradicts, and 68 uncharacterised zinc-finger motifs are essentially a known family's motif. Both are listed on the label checks page rather than quietly corrected.
• HOCOMOCO v14 has no dedicated paper. The motif ids used throughout are v14 (H14CORE), but the citable description is the 2024 paper for v12, which used ChIP-seq and HT-SELEX only. The GHT-SELEX, SMiLE-seq and PBM evidence codes postdate it.
• The process made one systematic error of its own. The first version of the writing brief mis-decoded three HOCOMOCO evidence letters (G, I and B, which are all in vitro assays) as "integrative" or "other collections". Twelve sentences in ten entries repeated it, understating the evidence behind those models, before it was caught and corrected.
• Family assignment is from the label, not the matrix. A motif's family comes from the protein name it carries, routed through a curated gene-symbol table. Only 27 of the 4,232 were assigned from similarity instead, and a further 256 carry a label so generic ("ZNF") that they are grouped as unnamed archetypes rather than placed. That independence from the matrices is what makes the label checks possible — but it also means a mislabelled matrix lands in the wrong family until someone notices.

Reproducing it

Everything is in the repository: the family ontology, the gene-symbol assignment tables, the written entries as JSON, and the scripts. npm run families rebuilds the assignment and the merged atlas, npm test runs the integrity checks, and verify-refs.mjs re-checks every citation against PubMed.

Corrections

If something here is wrong, it probably is somewhere. Corrections are the fastest way to fix it, and they are public.