Coverage and quality

Every number here is computed from the published atlas by npm test and the build, not quoted from memory. The unflattering ones are included on purpose: an atlas that only reports what it does well is not telling you how much to trust it.

Coverage

163 families across 54 DNA-binding-domain classes. Total written text: 141,111 words in the long entries, median 872 per family.

How cohesive the families are

Median similarity of a family's members to its representative motif. A family below 0.5 is a bucket of unrelated matrices, not one motif — and its entry says so.

Evidence behind the matrices

HOCOMOCO records how each matrix was measured. In vitro evidence supports "this protein binds this sequence"; a ChIP-only matrix records where the protein was crosslinked, which need not be a sequence it binds itself.

Where the data and the labels disagree

All of these are listed individually on the label checks page.

Citations

2,090 references across 163 entries, 1,780 of them distinct; median 13 per family, minimum 8. Citation text is regenerated from PubMed's own record, so it cannot drift from the identifier beside it. What is not checked is whether each sentence is a fair reading of the paper it cites — see methods.

Anchoring and fact-check

2,656 inline citations, 1,235 history steps and 521 open questions.

Reading each anchored sentence against its paper found 106 problems: 66 contradictions, 15 overstatements, 17 anchors to the wrong paper and 6 unsourced claims. Every one is listed.

Entries that decline to answer

30 of 163 families state that they have no single consensus rather than being given one. That is a result, not a gap: a KRAB zinc-finger hit identifies a kind of protein, not a site.