Transcription Factor Annotation Review Project

Bottom line: having a DNA-binding domain does not make a protein a
transcription factor, and GO's dbTF annotations (GO:0003700 and descendants)
are the place where that confusion shows up. February 2026 local snapshots
were used to compare InterPro domains, UniProt KW-0805 proteins and the
GREEKC consortium's 1,449-protein curated dbTF target set. In the 1,447-protein
GOA comparison set, GOA and GREEKC agree on 1,385 proteins, 95.6%; 62 proteins
are GOA-only. Seven of those 62 were reviewed in detail: the netrins NTN1 and
NTN3 had dbTF activity and cis-regulatory DNA binding REMOVEd. HDAC4, RPS3,
RFXAP and RFXANK had dbTF activator or DNA-binding rows either MODIFYed to
GO:0003713 transcription coactivator activity, MODIFYed to GO:0000976
transcription cis-regulatory region binding, or REMOVEd. NME2 instead proved
to be a non-core, noncanonical promoter/G-quadruplex regulator. The netrin
case traced to POU-seeded rows at PAINT node PTN000180816 in the cached
PTHR10574 export; a family review records the provenance discrepancy that
still needs upstream tracing.

The false-negative direction came out clean. Homeobox-fold ceramide synthases,
the bHLH inhibitors ID1-4 that lack a basic region, and the NCOA coactivators
are all correctly excluded from dbTF. The planned ML second pass has not been
run: 945 proteins in the 1,448-protein February 2026 raw QuickGO extraction
have IEA-only dbTF support and are still waiting for DeepTFactor validation,
and 50 GREEKC-only proteins are unreviewed.

Objective

Use AI/ML tools as a second-pass review to validate, challenge, and enhance existing GO annotations for transcription factors and transcriptional regulators. This addresses the gap between traditional homology-based annotation and modern deep learning approaches.

Background

The Problem

Current GO annotations for TFs rely heavily on:

• Pfam/InterPro domain detection (HMMs)
• Sequence homology (BLAST)
• Manual curation from literature

These methods have limitations:

• Miss novel TF families with no homology to known TFs
• Cannot distinguish functional subtypes (activator vs repressor)
• Don't capture proteins that lack canonical DNA-binding domains but function as cofactors
• Propagate annotation errors through homology transfer

Recent Advances (per Palsson et al. 2026)

Workflow

Phase 1: Sequence-based prediction

For each protein annotated with TF-related GO terms:

Input: UniProt sequence
    │
    ├─► DeepTFactor
    │   └─► Binary TF prediction + saliency maps (DBD regions)
    │
    ├─► InterPro/Pfam
    │   └─► Known DBD domains (zinc finger, HTH, etc.)
    │
    └─► ESM-2 embeddings
        └─► Cluster with known TF families

Discrepancy flags:

• GOA says TF but DeepTFactor says non-TF (or vice versa)
• DeepTFactor highlights non-canonical regions (no Pfam DBD match)
• Protein clusters with TFs in embedding space but lacks TF annotation

Phase 2: Functional subtype analysis

Current GO doesn't cleanly distinguish:

• Sequence-specific DNA-binding TFs vs general TFs
• Transcriptional activators vs repressors
• Pioneer factors vs non-pioneer factors
• Direct DNA binders vs cofactors

Approach:

1. Extract proteins with GO:0003700 (DNA-binding TF activity) and children
2. Cross-reference with:
   • GO:0001228 (activator) vs GO:0001227 (repressor)
   • Presence/absence of transactivation domains
   • Literature evidence codes (IDA, IMP vs IEA, ISS)
3. Flag proteins with:
   • Only IEA evidence for TF function
   • Conflicting activator/repressor annotations
   • TF annotation but no detectable DBD

Phase 3: Structure-based validation

For proteins where sequence methods are ambiguous:

Input: UniProt ID
    │
    ├─► AlphaFold structure
    │
    ├─► Foldseek search
    │   └─► Find structural neighbors (even with low sequence identity)
    │
    └─► Compare GO annotations of structural neighbors

Use cases:

• Validate DBD predictions via structural fold
• Identify potential DNA-binding surfaces
• Link "dark" proteins to annotated TF families via structure

Phase 4: Transcriptomic validation (where available)

For bacterial TFs, leverage iModulon data:

Input: Gene symbol
    │
    ├─► iModulonDB lookup
    │   └─► Is this gene associated with a known iModulon?
    │
    └─► Compare predicted regulon with GO annotations

Validation criteria:

• TF annotation supported by iModulon membership
• Predicted target genes consistent with GO BP annotations

Implementation

Tools to integrate

Data sources

1. GOA files: QuickGO API for current annotations
2. UniProt: Sequence and reviewed annotation status
3. Publications: PMID cache in publications/ folder

Output format

For each reviewed protein, generate entries in the standard ai-review.yaml format:

existing_annotations:
  - term:
      id: GO:0003700
      label: DNA-binding transcription factor activity
    evidence_type: IEA
    original_reference_id: InterPro:IPR000123
    review:
      action: MODIFY  # or ACCEPT, REMOVE, MARK_AS_OVER_ANNOTATED
      summary: |
        DeepTFactor predicts TF with high confidence (0.95).
        Saliency highlights residues 45-120 (zinc finger region).
        However, no experimental evidence for DNA binding.
      proposed_replacement_terms:
        - id: GO:0003712
          label: transcription coregulator activity
      ml_validation:
        deeptfactor_score: 0.95
        deeptfactor_regions: "45-120 (C2H2 zinc finger)"
        interpro_domains: ["IPR007087"]
        structural_neighbors: ["P12345 (0.85 TM-score)"]

Prioritization

High priority targets

1. Human TFs with IEA-only evidence: 945 proteins
2. Proteins with conflicting annotations: activator AND repressor
3. "TF" annotations lacking any DBD domain: potential cofactors
4. Bacterial y-ome candidates: uncharacterized E. coli proteins predicted as TFs

Species focus

Evaluation metrics

Per-protein metrics

• Agreement rate: DeepTFactor vs existing GOA
• DBD detection rate: Pfam vs DeepTFactor saliency
• Evidence upgrade potential: IEA → computational support

Cohort metrics

• False positive rate in IEA annotations (estimated by ML disagreement)
• Novel TF candidates identified (no current annotation)
• Annotation refinement suggestions (too general → more specific)

References

Key methodological papers

1. Kim GB et al. (2021) DeepTFactor. PNAS 118:e2021171118
2. Kim GB et al. (2023) DeepECtransformer. Nat Commun 14:7370
3. Yu T et al. (2023) CLEAN. Science 379:1358-1363
4. Barrio-Hernandez I et al. (2023) Foldseek Cluster. Nature 622:637-645
5. Palsson BO, Lee SY, Kim GB (2026) AI for gene function discovery. Nat Microbiol 11:350-358

Deep research output

• TRANSCRIPTION_FACTORS/ai-methods-query.txt - Query used for literature scan
• TRANSCRIPTION_FACTORS/ai-methods-tf-prediction-perplexity.md - Literature scan results

GO Consortium TF annotation guidelines

Two related documents exist:

The PDF states it is "replaced by" the journal paper, but this is misleading — both documents are needed:

Journal paper (2021) provides:

• Formal rationale for GTF/dbTF/coTF classification
• Peer-reviewed, citable reference
• Conceptual framework for ontology structure

PDF guidelines (2023) provides content NOT in the paper:

• 4-question checklist for annotation decisions
• Worked examples with specific PMIDs:
• GTF: GTF2H2 (PMID:10924514)
• dbTF: NKX6.3 (PMID:26314965)
• coTF: SIN3A (PMID:22783022) — includes removal of incorrect annotations
• High-throughput data interpretation rules:
• HT-SELEX: supports DNA binding, NOT dbTF without functional evidence
• ChIP-seq: inconclusive alone; needs DBD + functional data
• Bacterial 1-hybrid (B1H) guidance
• Pitfalls section with common annotation errors
• Annotation extensions syntax (has_input, is_active_in, occurs_in)
• Evidence tables showing which experiments support which terms

Synthesized guidelines: TRANSCRIPTION_FACTORS/tf-synthesized-guidelines.md — operational distillation of both sources.

Human dbTF catalogue

From Lovering et al. (2021) "A GO catalogue of human DNA-binding transcription factors" (PMID:34673265) and the February 2026 local QuickGO export:

Source: QuickGO API query for GO:0003700 descendants, taxon 9606, Swiss-Prot reviewed proteins.

Related resources:

• QuickGO dbTF target set
• TFCheckpoint 2.0 — cross-references 13 TF collections

Project History

2026-02-04: Initial setup

1. Literature scan on AI/ML methods for TF prediction from sequence

   • Created query in TRANSCRIPTION_FACTORS/ai-methods-query.txt
   • Ran deep research via perplexity → TRANSCRIPTION_FACTORS/ai-methods-tf-prediction-perplexity.md
   • Key finding: DeepTFactor (2021) is primary tool; transformers underexplored for TF classification

2. Reviewed Palsson et al. 2026 (Nature Microbiology)

   • Comprehensive review of AI for gene function discovery
   • Covers DeepTFactor, DeepECtransformer, CLEAN, Foldseek, iModulons
   • 43 references vs 6 from deep research scan
   • Added to project references

3. Gathered GO annotation guidelines

   • Downloaded TRANSCRIPTION_FACTORS/TF-annotation-guidelines.pdf (GO Consortium)
   • Compared with journal paper (PMID:34461313)
   • Created TRANSCRIPTION_FACTORS/tf-synthesized-guidelines.md — concise operational version

4. Defined ML-based review workflow (this page)

   • 4-phase approach: sequence → subtype → structure → transcriptomics
   • Output format compatible with ai-review.yaml schema

Status

Phase 0: Dataset Comparison ✅

• Done: compare GREEKC curated dbTF set with GOA annotations (95.6% agreement)
• Done: identify 62 GOA-only proteins (potential over-annotations)
• Done: identify 50 GREEKC-only proteins (potential missing annotations)

Phase 1: Manual Review (In Progress)

• Done: review 7 high-priority GOA-only proteins
  • NTN1, NTN3: REMOVE dbTF and cis-regulatory DNA-binding IBAs
  • HDAC4: MODIFY activator row → GO:0003713; MODIFY DNA-binding row → GO:0000976
  • RPS3: MODIFY activator row → GO:0003713
  • RFXAP: MODIFY activator and DNA-binding rows → GO:0003713
  • RFXANK: MODIFY activator row → GO:0003713; REMOVE DNA-binding row
  • NME2: KEEP_AS_NON_CORE (noncanonical promoter/G-quadruplex regulation)
• Done: create PTHR10574 (Netrin/Laminin) family review
• Done: document the POU-seeded PAINT provenance discrepancy at PTN000180816
• Todo: review remaining ~55 GOA-only proteins
• Todo: review 50 GREEKC-only proteins for potential annotation
• Todo: trace whether the netrin PTN000180816 mismatch originated in tree placement, identifier mapping, or export assembly

Tracked in ai-gene-review#4021.

Phase 2: ML Validation (Planned)

• Todo: extract 945 IEA-only proteins (dbTF annotation without experimental evidence)
• Todo: fetch sequences from UniProt (deeptfactor-validation/iea-only-sequences.fasta)
• Todo: create deeptfactor-validation/ analysis pipeline
• Todo: install conda/DeepTFactor environment (requires Python 3.6 + PyTorch 1.2)
• Todo: run DeepTFactor predictions
• Todo: analyze discrepancies

Phase 3-4: Structure & Transcriptomics (Not Started)

• Todo: Foldseek validation for ambiguous cases
• Todo: iModulon integration for bacterial TFs