Human dbTF annotations

A DNA-binding domain is not a transcription factor

AI Gene Review · projects/TRANSCRIPTION_FACTORS/

Bottom line

  • A February 2026 comparison used 1,447 GOA dbTF proteins (GO:0003700 and descendants) and 1,449 GREEKC dbTFs; the two agree on 1,385, or 95.6%
  • Seven of the 62 GOA-only proteins were reviewed in detail: six carried dbTF/cis-regulatory over-annotations and NME2 was retained as a non-core G-quadruplex regulator
  • The false-negative direction is clean: ID1-4, the NCOA coactivators and homeobox-fold ceramide synthases are all correctly excluded

The distinction that gets lost

A dbTF binds a specific DNA sequence and regulates transcription from it.

Two other things are routinely confused with it:

  • coTF — a coactivator or corepressor. Acts on transcription, does not bind the site. GO:0003713 / GO:0003714.
  • GTF — a general factor or Mediator subunit. Part of the machinery, not a sequence-specific regulator.

And a DBD alone proves nothing: ceramide synthases carry a homeobox-like fold, the ID proteins carry a bHLH without its basic region.

Three lists, three scopes

What the seven reviews found

The 62 proteins GOA calls dbTF and GREEKC does not include enzymes (HDAC4, ABHD2), a noncanonical G-quadruplex regulator (NME2), a ribosomal protein (RPS3), RFX complex scaffolding subunits (RFXAP, RFXANK), and two secreted guidance molecules (NTN1, NTN3).

Reviewed outcomes:

Gene Row Action
NTN1, NTN3 dbTF activity, cis-regulatory DNA binding (IBA) REMOVE
HDAC4 activator; cis-regulatory DNA binding (IDA) MODIFY → GO:0003713; MODIFY → GO:0000976
RPS3 DNA-binding transcription activator activity (IMP) MODIFY → GO:0003713
RFXAP activator and regulatory-region DNA binding (IDA) MODIFY → GO:0003713
RFXANK activator; regulatory-region DNA binding (IDA) MODIFY → GO:0003713; REMOVE
NME2 DNA-binding transcription activator activity (IDA) KEEP_AS_NON_CORE

One node, two families

What the netrin case shows

The IBA WITH/FROM list is entirely POU domain proteins: POU2F1, POU1F1, POU5F1, POU4F1-3 and their fly, mouse, rat and worm orthologs.

  • Netrins sit in PTHR10574; the verified POU5F1 seed sits in PTHR11636
  • Netrins are secreted axon-guidance cues with laminin N, EGF and NTR domains and no DNA-binding domain
  • The POU-seeded rows are present in the cached PTHR10574 export; whether the mismatch came from tree placement, identifier mapping or export assembly still needs upstream tracing

Recorded in the family review interpro/panther/PTHR10574/PTHR10574-review.md.

Evidence, and what is left to check

Of the 1,448 dbTF-annotated proteins in the raw QuickGO extraction, 945 have IEA-only support. That is the pool where a machine-learning second pass could add something.

Planned but not run:

  • DeepTFactor over the 945 IEA-only proteins, with saliency maps checked against Pfam DBD positions
  • The remaining 55 of the 62 GOA-only proteins
  • The 50 GREEKC-only proteins, mostly zinc fingers, as candidate false negatives
  • MEIOSIN (C9JSJ3), the one bHLH-domain protein whose exclusion is unresolved

Status

Done: three-way set comparison (GOA, UniProt KW-0805, GREEKC); seven gene reviews; the PTHR10574 family review; a synthesized guideline document.

Files: projects/TRANSCRIPTION_FACTORS.md (objective, workflow, tooling plan) and, in projects/TRANSCRIPTION_FACTORS/:

  • dbTF-discrepancy-analysis.md — the set comparisons and evidence-code breakdown
  • greekc-goa-comparison.md · PANTHER-IBA-error-report.md
  • tf-synthesized-guidelines.md · human-dbTF-list.tsv and the diff files

Open: the ML pass, 105 unreviewed discrepant proteins, and a cross-check against TFCheckpoint 2.0.