PseudoCAP annotation review: P. aeruginosa seeds of PAINT IBAs

Bottom line: PseudoCAP, the community annotation group of the
Pseudomonas Genome Database, made most of the experimental GO annotations on
Pseudomonas aeruginosa PAO1 (1,841 of 2,941 experimental rows in QuickGO).
PAINT curators use some of those rows as seeds for ancestral-node (IBD)
annotations, which reach many other proteins as IBAs, so an error in a seed is
copied to every descendant. Of 471 PAO1 proteins that seed an IBD, 398 have a
PseudoCAP annotation supporting at least one seeding; for 139 node terms (109
proteins) the PAO1 protein is the only seed. We reviewed all 109 of those
proteins: 106 have full reviews, and the review files for pvdE, pqsA and PA2403
are still to be written. For 21 genes the seed is wrong or
is an indirect mutant phenotype, and the IBD it supports should be withdrawn or
repointed. Examples are catalase activity on a heme exporter (ccmC), lipid
biosynthesis on an esterase (estA), DNA endonuclease activity from a mistyped
GO id (pilU), and signal transduction on a signal synthase (lasI). For 23 the
seed is right for PAO1 but the node carries the term to proteins or lineages it
does not fit. For 30 the claim is true but the PseudoCAP row misstates the
evidence code, qualifier or reference, and 32 are sound. The reviews record 116
defects for PseudoCAP and 75 for PAINT, written up in
PSEUDOCAP/defects/.

Tier 2, the 289 proteins whose node terms have other seeds as well, is now
reviewed too: 284 of 289 genes, 342 seedings and 43,061 leaf IBAs. There most
seeds hold and most nodes stand on their other seeds, but the reviews still
record 156 PseudoCAP defects and 41 for PAINT, including eleven seed rows that
describe a different gene or organism and a pair of choline transporters whose
annotations are swapped. Across all PAO1 reviews the reports now hold 272
PseudoCAP, 116 PAINT, 15 UniProt and 12 other defects.

Why start with IBA seeds

An IBA is a projection of an IBD made at a PANTHER tree node (see
IBA_REVIEW). PAINT curators place the IBD after reading the
experimental annotations of the node's descendants, so a seed annotation is
load-bearing in a way an ordinary experimental row is not. Two features of the
PseudoCAP corpus make the seeds worth checking first:

• Mutant phenotypes as functions. Half of the PseudoCAP rows that support a
  seeding are IMP (291 IMP, 300 IDA, 15 EXP, 4 IEP). A mutant phenotype can be
  an indirect effect: the clearest case found so far is ccmC, a cytochrome c
  maturation heme exporter whose IMP catalase activity row (PMID:16824107) is
  the only seed of a catalase IBD propagating to 143 leaves, including back onto
  ccmC itself. Overall, 71 molecular-function node terms are supported only
  by IMP or IGI from PAO1 (8,129 leaf IBAs), and 14 of them are sole-seeded.
  Many of these are probably sound (sigma factor, TonB-dependent siderophore
  receptor activity inferred from mutants in well-understood families), but
  each needs a look.
• Sole seeds. When the PAO1 protein is the only seed, there is no second
  line of evidence for the curator's node call.

Data and method

The worklist is built by
build_iba_seed_worklist.py and
written to iba_seed_worklist.yaml
(generated 2026-10-28). It downloads, into the git-ignored .cache/paint/:

• PANTHER PAINT IBD.gaf (node annotations with their seed lists) and the leaf
  IBA file gene_association.paint_uniprot.gaf.gz;
• the PAO1 proteome (UniProt, taxon 208964, 5,583 entries);
• all PAO1 experimental annotations from QuickGO (ECO:0000269 and ECO:0000006
  descendants);
• go-basic.obo (release 2026-07-26).

A PAO1 protein seeds a node term when it is listed in that IBD row's WITH
column. A seeding is PseudoCAP-supported when one of the protein's
non-negated experimental annotations, assigned by PseudoCAP, is the node term
or an is_a/part_of descendant of it. Every PAO1 seeding found at least one
supporting experimental annotation, which is a useful check on the join.
Leaf IBAs are counted per node and term from the leaf IBA file. That file
covers only PANTHER's reference proteomes, so the counts are lower bounds:
GOA also projects IBAs onto the rest of UniProtKB. For three nodes checked in
QuickGO on 2026-10-28 (all terms from the node together), the GOA rows were
925 for PTN001247822 (ccmC; 143 catalase IBAs in the reference file), 1,770 for
PTN005219546 (estA; 960 per term) and 369 for PTN002388087 (PA4079; 27).

Ranking: tier 1 genes are the only seed of at least one node term;
tier 2 genes share every node with other seeds. Within a tier, genes are
ordered by the leaf IBAs that hang on their sole-seeded nodes, then by total
leaf IBAs.

Per-gene review protocol

Each gene gets a full review (all GOA rows, core functions, notes). On top of
that, the notes file carries a ## PseudoCAP IBA-seed assessment table with,
for every seeding node: the PTN node, node term, number of seeds, leaf IBAs,
the seed annotation, and a verdict:

• SOUND: the paper directly shows the node term (or a descendant).
• OVER-SPECIFIC: the seed is right but the node term is more specific than
  the clade can carry, or vice versa.
• INDIRECT-PHENOTYPE: the seed is a downstream mutant phenotype, not the
  gene's own activity or process.
• WRONG: the paper does not support the row.
• UNVERIFIABLE: the paper cannot be read (abstract-only, no full text).

Errors in a PseudoCAP row are recorded as review.source_defects with
resource: PSEUDOCAP (added to SourceResourceEnum for this project). When a
wrong seed leaves the IBD unsupported, a second defect with
resource: GO_CENTRAL names the PTN node, so just export-source-defects
produces one report for PseudoCAP and one for PAINT.

Results: tier 1 (sole seeds)

All 109 tier 1 proteins have been assessed, and 106 have full reviews. The
review files for pvdE, pqsA and PA2403 still need writing (see below), so they
are tallied separately; pvdE's notes and seed assessment (seed sound, node too
broad) are committed. Each gene is classified by the outcome for
its sole-seeded node terms. Leaf IBA counts are the worklist's
reference-proteome counts for those terms, so they are lower bounds.

• Withdraw or repoint (21): estA, ccmC, pilU, pilM, pilN, wbpD, PA0920,
  liuC, PA5377, lasI, PA3697, cysB, ptrB, PA5208, PA1340, lhpL, znuC, phnB,
  pslC, motB, nrdJa.
• Node does not fit the clade (23): PA5458, PA3760, mgtE, aruC, gbuA,
  mucB, pchB, nirJ, fpr, fur, nmoA, tgpA, dctQ, PA4503, PA4504, pqsC, wzm,
  phzC1, thiI, pelA, xcpV, braF, PA5359.
• Fix the seed row, keep the IBD (30): fleN, hutI, hutF, hom, thrB, PA3891,
  the six Sha subunits PA1054-PA1059, PA2280, phzS, lpxA, xcpX, xcpS, xcpW,
  xcpQ, xcpR, PA0692, dgcB, PA4521, pilA, oprB, PA4505, gltA, arcD, PA2580,
  fliD.
• Sound (32): lap, pqsH, PA1024, pilB, phzM, hutH, braD, waaL, wbpY, cif,
  pilD, pscC, relA, hfq, dksA, cifR, PA4079, PA4624, oprG, pilT, fimU, arcA,
  lasB, PA1486, PA5085, PA0906, dctP, PA1341, mucD, pfeA, pvdH, eddB.

The classification is by the worst outcome among a gene's sole-seeded nodes;
several "sound" genes still have defects on other rows (for example lap and
lasB, which are given secretion-system terms for being secretion cargo).

Upstream reports

export_pseae_defects.py collects the
source_defects recorded in all PAO1 reviews into one report per resource in
PSEUDOCAP/defects/. After tier 2, and after
the merge from main that brought in the AST-module and eut reviews, they stand
at 272 for PseudoCAP (123 wrong term, 84 wrong evidence code, 41 wrong
reference, 14 wrong qualifier, 8 wrong gene product, 2 other), 116 for PAINT/GO Central,
23 for UniProt, 7 for InterPro (which now has its own report, since the
InterPro2GO mappings outgrew the catch-all) and 5 for other resources: 423 in
all, matching the 423 source_defects entries recorded under genes/PSEAE. They
include defects from PAO1 reviews made before this project. The tier-1 counts
quoted below are the totals as they stood at the end of tier 1 (116, 75, 5 and
11).

Confidence follows the rule in REVIEW_PROTOCOL:
CONFIRMED needs the full text, primary data we checked, or an abstract that
states the decisive fact outright; a call that rests on what an abstract does
not say is SUSPECTED. On 2026-10-29 we audited the 66 PseudoCAP defects at
PROBABLE or CONFIRMED whose papers we hold only as abstracts, after the xcp
review-as-IDA call turned out to have been made from an abstract (the full
text, supplied by the user, later confirmed it). Of the 66, 18 dropped to
SUSPECTED, 2 dropped from CONFIRMED to PROBABLE, 7 rose to CONFIRMED on a
full-text read (lhpL, pilA, waaL), and 39 were unchanged; none was removed and
no annotation action changed. Six more were then raised to CONFIRMED because
the abstract states the decisive fact outright (PA1340 and PA2280: the paper
is about another organism; cysB: CysB is a transcription factor whose mutant is
an auxotroph; estA: an overexpression result coded IDA). The user then
supplied four full texts, which settled 16 more calls on 2026-11-05:
PMID:16030218 (the only sha mutant is the PA1054 insertion, so the ten IMP
evidence-code defects on PA1055-PA1059 are CONFIRMED), PMID:4290562 (step III
is never assayed and there is no hutI or hutF mutant; histidase is assayed
directly, so hutH's IMP should be IDA), PMID:7768797 (oprB's evidence is
whole-cell uptake in mutant and overexpressing strains, so IMP) and
PMID:9224878 (its Fig. 1 is a pilA mutant twitching assay, so pilA's IMP
stands and that defect was removed). The PseudoCAP report stood at 36
CONFIRMED, 56 PROBABLE and 24 SUSPECTED at the end of tier 1, and the PAINT
report at 6, 56 and 13. After tier 2 and the tree check (below) the totals are
96, 126 and 50 for PseudoCAP and 32, 83 and 1 for PAINT.

Checked against the tree

Node-level calls were first made from the PAINT leaf list, which says what a node
reaches but not how the clade is shaped. panther_tree.py
now pulls the family's real topology from the PANTHER treeinfo service and reports a
node's event type, its subfamily composition, every leaf, and the smallest clade holding
a given set of accessions. The trees for 103 of the 105 nodes we filed against were
fetched this way; the other two (hutH, pfeS) were resolved by hand, and pfeS turned out
not to sit in the family UniProt assigns it.

Every defect that had been left SUSPECTED pending the tree was re-examined on 2026-11-05.
The PAINT report moved from 17 CONFIRMED, 82 PROBABLE and 17 SUSPECTED to 37 CONFIRMED,
77 PROBABLE and 2 SUSPECTED (116 entries in all). The 32/83/1 split quoted here earlier was
exact when first written, but the leaf-evidence audit in 1f556e0 regenerated the reports and
changed it; the totals still summed to 116, which is why the stale figures survived several
reading passes. The check does three different things:

• It names the target. A request to move an IBD can now give the node to move it to:
  astA's dihydrolase term to PTN000764522 (the node is a duplication whose two children
  are exactly the AstB and AstA/AruG subfamilies), xcpV's two T2SS terms to PTN002215342,
  pfeA's colicin receptor to PTN002012591, thrB's homoserine kinase to PTN002326066,
  hutF's two terms to PTN000138520, PA1257's cystine transport to PTN002024099.
• It shows when there is no target. For fur the family is homogeneous but siderophore
  production does not track the topology, so the fix is to retype the node term rather
  than move it. For PA4088 the gene sits in a Pseudomonas duplication whose third member
  is PAO1's own HemL, so no HemL-only subclade exists and the remedy became a leaf-level
  divergence note. For ligD and aruC the clade is uniform, so the term is either right or
  wrong family-wide: ligD's became a withdrawal request, and aruC's "move it" half was
  dropped.
• It refutes. hutI's 52 leaves really are all imidazolonepropionases, so that breadth
  objection was withdrawn and the defect narrowed to its citation. PA4521's clade is a
  clean one-ortholog-per-organism group, so the breadth objection was withdrawn entirely
  and only the unverified seed remains — SUSPECTED, waiting on the full text of
  PMID:9835532. It is one of the two entries still SUSPECTED; the other is mucB's
  PTN002215378 GO:0032885 entry, which states its own counter-consideration. arcA's flagged leaves mostly survived scrutiny: three of four
  are arginine deiminases whose names lag, leaving one archaeal duplication pair.
  xcpV is the sharpest case: our own leaf-count analysis had read the node as spanning the
  GspI/PilV split, but no PilV leaf is in the tree at all — the extra proteins were
  non-reference members grafted by sequence. That conclusion is corrected in
  genes/PSEAE/xcpV/xcpV-bioinformatics/RESULTS.md.

Per-gene verdicts

Verdicts on the PseudoCAP seed and on the IBD it feeds. "Defects" are the
source_defects recorded in the review (P = PseudoCAP, G = GO_CENTRAL/PAINT).

Not completed

The review files for pvdE, pqsA and PA2403 were not written: the agent's write
of each review YAML was blocked by a safety classifier (pvdE, pqsA), and the
agent stopped before PA2403. pvdE's notes, seed assessment and node analysis
are committed. The research findings for the other two:

• pqsA: the cited paper (PMID:12426334) attributes the PA0996 result to a
  different study; PMID:18728009 shows a pqsA mutant makes no quinolines. The
  seeding node PTN002458530 is the family root, and the child node
  PTN001186225 swaps terms between PqsA and FadD22.
• PA2403: the gene is FpvG, which releases iron from ferripyoverdine
  (PMID:28192658, PMID:31924850), so its pyoverdine biosynthesis seed
  (PMID:12207696, an expression-profiling study) looks wrong or indirect.

Patterns

• Indirect phenotypes as functions (ccmC, estA): a mutant lacking catalase
  activity because it cannot mature heme proteins was annotated as having
  catalase activity, and PAINT used it as the only seed of a catalase node.
  An esterase whose deletion lowers rhamnolipid output was annotated to
  glycolipid and lipid biosynthesis, and seeds a 960-leaf lipid biosynthesis IBD.
• Inflated evidence codes (fleN): IDA on a paper that reports a sequence
  motif.
• Wrong GO id (pilU, fimT): PseudoCAP entered GO:0016888 DNA endonuclease
  activity, the id after GO:0016887 ATP hydrolysis activity, which it gave PilB
  and PilT from the same paper (PMID:18174131) on the same day. PAINT then used
  the pilU row as the only seed of a node, so a pilus retraction ATPase clade
  inherits endonuclease activity. The same GO:0016888 row also sits on fimT, a
  minor pilin that the paper's abstract does not mention, so it looks copied
  onto the wrong gene product as well; the fimT review asks for it to be
  deleted. fimT seeds no node and receives no IBA.
• Wrong references (ccmC, phzS): the cited PMID is a different paper
  (E. coli CcmABC biochemistry for ccmC; a crystallization note for phzS).
  The pilB/pilU paper looked like a mis-keyed neighbour of the ccmC paper but
  is correct.
• Evidence codes that overstate the assay (fleN, xcpX, xcpS, xcpW, xcpQ,
  xcpR, PA0692, PA5458, PA3891, PA4503, PA4504, dgcB): IDA on a sequence motif,
  on cross-species complementation, on an operon deletion or on a review; IMP
  for complex membership. All six Xcp type II secretion process rows cite one
  review (PMID:9818381) as IDA. We first flagged this from the abstract alone,
  which was not enough: a review from the laboratory that did the work can
  report original data. The full text (supplied by the user, read 2026-10-28)
  settles it. Its only new observations are unpublished notes on XcpT and XcpR
  multimerization, XcpR membrane association and an XcpP-XcpQ interaction,
  with no original secretion data, so the rows should be TAS.
• Substrate terms on clades where the binding protein sets the substrate
  (PA4503, PA4504, braF, dctQ): dipeptide transport seeded on permease nodes
  that also hold NikB/NikC (nickel), SapB (putrescine) and GsiD (glutathione);
  C4-dicarboxylate transport on a YiaM (diketogulonate) clade.
• Over-specific activity (nrdJa): a thioredoxin-dependent diphosphate
  reductase term on an enzyme that prefers triphosphates and was never assayed
  with thioredoxin.
• Subunit and partner conflation (PA5377, PA1055, PA1057): carnitine
  transport, which needs the separate binding protein CaiX, seeded on the
  permease node; Sha antiporter activity, measured for the whole PA1054-PA1059
  cluster after deleting only PA1054, given as IMP enables to each subunit.
• Consequences as processes (PA0920): an alanyl-phosphatidylglycerol
  synthase annotated to phospholipid homeostasis.
• Pathway context that does not travel (liuC): terpene catabolism is true
  for PAO1 only because its terpene pathway feeds the leucine pathway, but it
  was seeded on a hydratase node whose other members lack it.
• Phenotypes as locations (pilM, pilN): IMP "mutant has no pili" recorded
  as located_in type IV pilus, and given identically to pilN, pilO and pilP;
  confirmed on pilN.
• Gene-specific roles on family-root nodes (pchB, nirJ, fpr, fur): the
  seed is right for the PAO1 gene, but the node is the whole family. Salicylate
  synthesis on the chorismate mutase root, heme d1 synthesis on an archaeal
  radical SAM clade, heme catabolism on ferredoxin reductases that include
  E. coli Fpr, and siderophore repression on Fur proteins of bacteria that
  make no siderophores.
• One citation for a whole operon (hom): the ThrH/ThrB kinase paper cited
  for hom, thrB and thrH alike.
• Another organism's paper coded IDA (thiI): a Salmonella genetics study
  given to PA5118 as IDA.
• Old papers that predate the gene (hutI, lpxA): a 1967 pathway paper cannot tie
  an activity to PA5092. Here the term is right and the node reaches human
  AMDHD1, but its only anchor needs replacing.
• Wrong pathway (wbpD): an acetyltransferase of the UDP-ManNAc3NAcA
  pathway annotated to UDP-glucuronate biosynthesis, seeding an 8-leaf IBD.
• Cargo annotated as machinery (lap): a secreted enzyme given secretion
  system process terms.
• Sound seeds on nodes that are too broad (pvdE, pqsH, PA5458, PA3760): the
  PseudoCAP row is right, but the node carries the term to proteins that are a
  different kind of thing (soluble ATPase subunits given a transmembrane
  transporter activity), to lineages without the process (LPS biosynthesis on
  Gram-positive peptidoglycan O-acetyltransferases), or rests a large clade on a
  single seed. PA3760's N-acetylglucosamine transport seed sits on a general
  PTS Enzyme I node that also holds PtsI and the nitrogen-PTS PtsP.

These are catalogued for PAINT as pattern 17 of IBA_REVIEW, next to the closely related pattern 16 (a phenotype-derived molecular function as an IBD seed, worked through on the same ccmC node from the Shewanella side).

Results: tier 2 (nodes with other seeds)

Tier 2 is the 289 PAO1 proteins whose PseudoCAP-supported node terms also have
other seeds, so a wrong seed does not by itself sink the IBD. 284 are reviewed;
the five Xcp/Hxc secretion genes below are not. Together they cover 342
seedings and 43,061 leaf IBAs (reference-proteome counts, so lower bounds).

Most seeds hold up, and most nodes stand on their other seeds, which is the
expected result for this tier: the question here is less "does the IBD
survive" than "is the PseudoCAP row itself right". The reviews record 156
PseudoCAP defects (60 CONFIRMED, 70 PROBABLE, 26 SUSPECTED) on 130 of the 284
genes, 41 for PAINT/GO Central (11, 26, 4), 10 for UniProt and 1 elsewhere.
By kind, the PseudoCAP defects are 77 wrong term, 43 wrong evidence code, 22
wrong reference, 6 wrong gene product and 6 wrong qualifier.

Seeds that are wrong

Eleven seed rows are about something other than the gene they sit on, and all
are CONFIRMED from the text:

In each case the term itself is usually still true on other evidence, so the
row is kept and the reference or evidence code is corrected rather than the
claim withdrawn.

Node terms that do not fit the clade

41 PAINT defects, the 11 CONFIRMED ones being:

• argA (PTN004744913): the node term is the ArgJ-type N-acetyltransferase
  reaction, but every seed is an acetyl-CoA N-acetylglutamate synthase.
• ccoP1 (PTN002435498): cytochrome-c oxidase activity on a node of
  cytochromes c4/c552 and bc1 subunits; the electron-transfer parent fits.
• waaC (PTN001575314): heptosyltransferase II, from one E. coli WaaF
  seed, on a node spanning heptosyltransferases I, II and III.
• wzz (PTN000794252): tyrosine kinase activity, from Wzc and Etk, on a
  clade whose Wzz-type members end a few residues after the second TM helix.
• astA (PTN000764516): N-succinylarginine dihydrolase at the root of a
  family that also holds the unrelated AstA/AruG succinyltransferases, so
  seven succinyltransferase leaves inherit the dihydrolase.
• galU (PTN000501197): LPS core biosynthesis on a UTP-glucose-1-phosphate
  uridylyltransferase clade including Bacillus, Staphylococcus and
  Mycobacterium, which make no LPS.
• wapR (PTN000531351), PA5455 (PTN002288302, both seeds from one paper
  with no enzyme assay), speC and spuE (missing NOT annotations:
  AZIN1 on the ornithine decarboxylase node, putrescine binding on SpuE),
  tpx (PTN002454730, the node term is obsolete).

The PROBABLE ones repeat tier-1 shapes: a term that suits one subfamily placed
at a family root (PA4088, PA1611 and pfeS for histidine phosphotransfer
activity on kinases with no HPt domain, liuE, speE2, ligD, arcC, PA2077,
PA2127, cupA3), a location that does not hold across the clade (wspR, azu),
and a catalytic term given to a non-catalytic subunit (parC's ATP binding,
rpoS's RNA polymerase activity).

Patterns new in tier 2

• Paralog mix-ups (PA3933/PA5291, PA5134, dadA2, folP, pvdJ): tier 1 found
  single mis-keyed references; tier 2 found whole sets of rows on the wrong
  member of a family, in both directions between two genes.
• Subunit-level coding of a complex result (ccoP1, ccoP2): one operon
  mutant phenotype given to all eight cbb3 oxidase subunits as enables. The
  formate dehydrogenase subunits were checked for the same pattern and are
  clean: their rows come from biochemistry on the purified complex.
• Location rows that the assay cannot support (pmpM, wzx, PA0938, ampD,
  ampDh2, ampDh3, ampG, PA4218): periplasmic side of plasma membrane on
  12-TM transporters, and cell wall on cytoplasmic or inner-membrane
  proteins.
• Regulators annotated with their targets' activities (gbdR, ptxR, mexZ,
  nfxB, nalC, nalD, atuR, glmR): phospholipase and phosphatase activity,
  exotoxin A's ADP-ribosyltransferase activity, negative regulation of transporter activity for repressing a pump's genes, and terpene catabolism
  for repressing the atu operon.
• Gene model problems found on the way (PA4596, PA2057, betT1): annotated
  start sites that truncate a DNA-binding domain, a signal peptide, or the 18
  residues an active BetT1 needs.

Not completed (tier 2)

xcpT, xcpU, PA0681 (HxcT), PA0685 (HxcQ) and xqhA are fetched but unreviewed:
the agent's draft was stopped by a safety classifier. Its analysis, which is
not in any review file, found the tier-1 Xcp pattern again (the 1998 review
PMID:9818381 cited as IDA, which should be TAS) and judged both nodes sound.
The algC audit in the same batch also went unwritten; it found the
PMID:22513190 IMP seed to be an indirect phenotype, since that paper measures
polysaccharide output and never assays phosphomannomutase.

Plan

1. ~~Review every tier 1 gene.~~ Done except for writing the pqsA, PA2403 and
   pvdE review files.
2. Reconcile small inconsistencies between reviews: xcpX's complex row
   (PMID:9466253) has no reference defect, while the similar xcpW and xcpV rows
   (PMID:8331069) do.
3. Send the PseudoCAP and PAINT reports upstream, starting with the 21
   withdraw-or-repoint cases. The PAINT report now names a target node for each
   relocatable IBD (see "Checked against the tree").
4. ~~Tier 2 (289 proteins that share each node with other seeds).~~ Done for
   284; the five Xcp/Hxc genes above still need review files.
5. Reconcile tier-2 inconsistencies between parallel batches: MODIFY versus
   TOO_GENERAL for the same general term (rpoS against algU/pvdS; wbpE against
   wbpD/wbpY; nrdA against nrdJa), whether an IDA code resting on a mutant
   phenotype is worth a defect (PA1014 filed one, migA and wapR did not), and
   the braC/braG disagreement over D-alanine transport, where only the braC
   review used the ITC paper. The np20 review also judges its pyocyanin seed an
   indirect phenotype, where the earlier fur review called the same seed
   reasonable.
6. The rest of the PseudoCAP corpus (2,069 rows on PAO1 in QuickGO),
   prioritising IMP rows on molecular-function terms.

Re-run the builder when PAINT or QuickGO data change; it only re-downloads
files missing from .cache/paint/, so delete that folder to refresh.