PAINT Human No-IBA Gene Review Project

Bottom line: PAINT curators annotate ancestral nodes in PANTHER trees, and those calls reach human genes as IBA annotations; human genes with no IBA at all may be poorly characterized, divergent, or lack orthologs with experimental evidence. This project works through a list of 7,593 such genes (7,524 distinct symbols), giving each two deep research reports and a full AI-assisted review of its existing GO annotations. We did this to find where phylogenetic annotation has a gap or a wrong node, and to see what the literature supports for genes that inherit nothing. As of 2026-09-26, 715 of the listed genes have a completed review (not all written under this project), covering 24,409 annotation decisions: 11,926 ACCEPT, 5,786 KEEP_AS_NON_CORE, 2,778 MARK_AS_OVER_ANNOTATED, 1,475 MODIFY, 1,338 REMOVE, 902 NEW and 204 UNDECIDED. The status section below (635 genes, 2026-09-04) predates this count.

The headline lessons are that the no-IBA list is stale (most genes in the 2026-09-04 batch now receive IBAs), that the recurring real gap is families with no molecular-function IBD at all, and that gene names mislead: PLD3 and PLD4, named as phospholipases D, are 5'-3' exonucleases, and PLD5 is catalytically inactive. The 2026-09-04 batch added structured PANTHER FamilyReviews for all 19 of its families; one of them (PTHR48178, PEX2) found an IBD placed on a PEX2/PAF1 name confusion.

Overview

Review human genes that lack IBA (Inferred from Biological Ancestor) annotations. These genes are candidates for PAINT (Phylogenetic Annotation INference Tool) but currently have no phylogenetically-inferred annotations, meaning they may be:
- Poorly characterized
- Have unique/divergent functions
- Lack clear orthologs with experimental evidence

For each gene, the workflow generates at least 2 deep research reports (from different providers) and completes AI-assisted annotation review.

Source: ai4curation/ai-gene-review

Data Source

• Spreadsheet: https://docs.google.com/spreadsheets/d/12bR3FZ7XrUXL86IKJc__K6QbFBEsSlQeSdiVXwFsjEI/
• Local: projects/paint/human-no-IBA-simple.csv (format: species,uniprot_id,gene_symbol)
• Total genes: 7,593

Model Species

Primary: Homo sapiens (human)
- UniProt species code: human
- Genes without IBA annotations are priority targets

Completed Reviews

165 genes with COMPLETE status (as of 2026-01-25)

Highlighted Reviews (Notable Findings)

Notable Findings

PLD3/PLD4/PLD5 Misannotation

These proteins are named "phospholipase D" but are actually:
- PLD3/PLD4: 5'-3' exonucleases with immune regulatory functions
- PLD5: Catalytically inactive pseudoenzyme

This is a prime example of misleading gene nomenclature that AI review can flag.

Batch Processing Infrastructure

The project has scaled to industrial batch processing:
- 50 batches prepared (2,500 gene capacity)
- Parallel architecture for continuous pipeline
- Can process 100+ genes per hour with deep research automation

Reproducibility

Per-gene workflow:

just fetch-gene human GENE
just deep-research human GENE --provider falcon
just deep-research human GENE --provider cyberian
# Review and complete ai-review.yaml
just validate human GENE

List all completed PAINT genes:

comm -12 <(cut -d',' -f3 projects/paint/human-no-IBA-simple.csv | sort) \
         <(grep -l "status: COMPLETE" genes/human/*/*.yaml | xargs dirname | xargs -I{} basename {} | sort)

Supplementary files in projects/PAINT/:
- human-no-IBA-simple.csv - Gene list (species, uniprot_id, gene_symbol)
- human-no-IBA.tsv - Full annotation data

STATUS

Project Statistics (2026-09-04):
- Total genes in project: 7,593
- PAINT genes completed: 635 (8.4%)
- Ready for review (have deep research but not complete): 6
(ERVMER34-1, PEX11A, SUMF2, TAX1BP1, TMEM67, TMF1)
- Structured PANTHER FamilyReviews written for reviewed genes' families: 19 of 19
(interpro/panther/<PTHR>/<PTHR>-review.yaml)

Progress

• [x] Infrastructure setup (batch processing, parallel deep research)
• [x] 635 PAINT gene reviews completed
• [x] Family-level review dimension added: paired FamilyReview per reviewed gene
• [ ] Complete reviews for remaining 6 genes with deep research
• [x] FamilyReviews for all 19 families of the 2026-09-04 batch (the three
  InterPro-outage stragglers — PTHR11494, PTHR15414, PTHR48482 — recovered
  and completed the same day)
• [ ] Re-fetch IL10 GOA: the cached snapshot still carries now-obsolete
  GO:0005615 rows that PAINT has already migrated to GO:0005576
• [ ] Re-derive the no-IBA source list against current GOA (see 2026-09-04 note)
• [ ] Scale deep research to all genes
• [ ] Full project completion (7,593 genes)

Last updated: 2026-09-04

NOTES

2026-09-04

20-gene batch with paired PANTHER family reviews

Completed finishing reviews (all validate with zero warnings, status COMPLETE)
for 20 genes: BCKDHA, BCKDHB, CD28, CTLA4, NDUFS2, NDUFV1, PEX2, PEX10,
PEX11B, PEX13, PEX16, ORMDL3, MBL2, MTCH2, IL10, ERLEC1, CFAP61, LOXHD1,
GPATCH11, NAALADL2 — and, new for this project, wrote structured FamilyReviews
(node-level PAINT/IBD adjudication) for all 19 of their PANTHER families
(CD28/CTLA4 share PTHR11494; the three families stalled by a multi-hour
InterPro API outage were recovered and completed the same day). Across the
19: residue validator 506 checks pass / 0 fail, family-gene crosscheck 0
conflicts.

Key findings:
- The no-IBA source list is stale. Most of the 20 "no-IBA" genes now
receive IBAs (PAINT IBDs dated 2022–2026): PEX11B, ORMDL3, CFAP61, LOXHD1,
BCKDHA/B, PEX13, PEX16, MTCH2, MBL2 among them. The recurring real gap is
narrower and invisible to a has-IBA test: families lacking any
molecular-function IBD (BCKDH E1, PEX13, PEX16, NDUFV1's eukaryotic node),
leaving human genes with only uninformative protein-binding IPI rows as MFs.
- PTHR48178 (PEX2): the Cdc73/Paf1-complex IBD is a homonym confusion
(PEX2 synonym PAF1 vs the PAF1 elongation factor), seeded by the target's
own miscited IDA — WRONG_NODE, retraction recommended.
- IL10: four GO_REF:0000024 ISS rows trace to mouse TNF (P06804), not
mouse IL-10 — a wrong-accession entry set argued from fold non-homology.
- PTHR10404 (NAALADL2): human NAALADL2 descends from the carboxypeptidase
IBD node yet receives no IBA (silent pruning); recommended an explicit IRD,
with a machine-checked residue site proving loss of the catalytic Glu pair.
- GPATCH11: PAINT's 2026-02-25 snapshot already withdrew the kinetochore
IBD (go-annotation#6450); the 2017 GOA kinetochore IBA is stale propagation.
- BCKDHA: draft carried a systematic 45-residue precursor-vs-mature
numbering error, now corrected against the UniProt SQ block.

Provenance: history records under history/genes/human/<GENE>/ and
history/other/<PTHR>/ (one per gene and per family review).

2026-02-04

Major batch annotation review session
- Reviewed 81 genes using annotation-reviewer agent
- PAINT-specific completions: 165 → 207 (+42)
- Total completions: 246 → 328 (+82)
- Remaining with deep research: 104 → 22 (-82)

Notable genes reviewed include:
- Fe-S cluster assembly pathway: HSCB, HSPA9, IBA57, ISCA1, ISCA2, ISCU, NFS1, NFU1, MMS19, CIAO1, BOLA3
- Apoptosis/autophagy: BCL2, BCL2L1, BECN1, CASP9, ATG4D, ATG5, ATG7, DRAM1, DRAM2
- Transcription factors: GATA3, FOXO1, IRF8, OLIG2, ASCL1
- Signaling: NOTCH1, LRRK2, AXIN1, FAS
- Disease-relevant: HTT (Huntington), FXN (Friedreich ataxia), CBS (homocystinuria)

Note: Subagent status updates weren't persisting; fixed manually.

2026-01-25

Project reorganization and stats update
- Created top-level PAINT.md project file
- Renamed folder from paint/ to PAINT/ for consistency
- Updated stats: 165 COMPLETE reviews (was showing 14 - massively out of date)
- Started cyberian server for deep research
- Identified 285 genes needing cyberian deep research
- Created batch processing script
- First gene (ABCB7) completed in 17 minutes

2025-12-18(Massive Scaling Session)

Major Achievement: Transformed from manual workflow to industrialized batch processing
- Expanded from 297 to 1,806 gene folders (+508% growth)
- Submitted 1,500+ cyberian deep research jobs
- Created 50 batches (2,500 gene capacity)
- Established fully parallel, async pipeline

2025-12-18 (Review Session)

Completed reviews for:
- ICA1, ICA1L - BAR domain proteins
- SOCS4 - E3 ligase adaptor
- PLD3, PLD4, PLD5 - Discovered misannotation (NOT phospholipases)
- RASA3 - Bifunctional RasGAP

Key finding: PLD3/PLD4/PLD5 nomenclature is misleading - they are exonucleases, not phospholipases.

Slides

• Slides (Marp source: PAINT-slides.md) — AI generated