Conserved-function recovery

Bottom line: pick a function that almost every genome in a clade must
carry, establish its presence in each proteome independently of GO (here from
InterPro family membership), and ask how well GO annotation recovers it. Both
kinds of disagreement are informative: a false negative is a proteome where
annotation lags the sequence, a false positive is a protein GO calls by the
function's name that does not have it. The first case, the bacterial F-type ATP
synthase, recovers well (precision 0.967, recall 0.976 over 15,525 bacterial
reference proteomes), and its false positives led to a systematic error: almost
every flagellar and type III secretion export ATPase (FliI/SctN) in UniProt is
annotated as an ATP synthase, through PAINT, TreeGrafter, InterPro2GO and GO's
own logical inference. The experiment is a genome-content cousin of
pathway satisfiability, which asks the same kind of
question of a whole module; here the unit is one conserved complex and the
point is to audit the annotation, not the genome.

Opened 2026-09-27 as a case study inside the
TreeGrafter evaluation; made a project of its own on
2026-10-14.

Case 1: the bacterial F-type ATP synthase

The test

A function that nearly every organism in a clade carries gives a truth set that
does not depend on GO: "does this proteome encode a complete F-type ATP
synthase?", decided from InterPro family membership of all eight subunits. The
test asks how well plain GO annotation recovers it
(completeness_fetch.py,
completeness_analyze.py,
results).

• Proteomes: 15,525 bacterial reference proteomes with BUSCO completeness ≥ 95%.
  Of these, 14,959 encode all eight F-type subunit families.
• Accuracy: "has a GO:0046933 protein" scores precision 0.967 and recall 0.976
  against that truth.
• False negatives: new TrEMBL entries that have InterPro cross-references but no
  GO annotations yet.
• False positives: mostly V/A-type ATPases and FliI/SctN. The FliI/SctN false
  positives are genuine annotation errors, and they are the subject of this page.
• Genuine absences: the complex is missing from phytoplasmas, and V/A-type
  ATPases replace it in many Clostridia, Spirochaetia and Bacteroidota.

What the false positives were

• Almost every UniProtKB entry in the FliI/SctN family (IPR005714) currently
  receives ATP-synthase GO terms. FliI and SctN are the soluble ATPases of the
  flagellar and injectisome (type III) export systems: they couple ATP hydrolysis
  to substrate docking and chaperone release and stimulate export, while the
  proton-motive-force-coupled step happens at the membrane export gate. They are
  homologs of the F1/V1/A1 α/β subunits (FliI resembles both F1 α and β;
  PMID:17202259), and in PTHR15184 the nearest synthase-side clade is F1-β.
  They hydrolyse ATP rather than make it. Of the 11,205 UniProtKB entries in
  their family (InterPro IPR005714), 97% carry GO:0046933 proton-transporting ATP synthase activity,
  rotational mechanism and 97% carry GO:0015986 proton motive force-driven ATP
  synthesis. Among the 29 reviewed Swiss-Prot entries, all 29 carry GO:0015986
  and 20 carry GO:0046933 (counts).
• One misplaced PAINT node explains the synthase terms. Both IBDs (for
  GO:0046933 and GO:0045259 proton-transporting ATP synthase complex) sit on
  PANTHER:PTN008558586 in PTHR15184. In the live PANTHER tree that node is a
  duplication. Its two children are the F1-β clade PTN008558588, which
  holds every experimental seed cited for those IBDs, and the FliI/SctN
  export-ATPase clade PTN000390097 (SF9 SPI-1, SF62 SPI-2, SF81 flagellar) (placement).
• PAINT IBA passes GO:0046933 to 28 proteins with the FliI/SctN signature
  IPR005714 in the reference genomes, plus a Chlamydia FliI (O84722) that lacks
  the signature (targets). They
  include Salmonella FliI, InvC and SsaN and E. coli FliI.
• TreeGrafter then passes GO:0046933 to every grafted FliI/SctN. All five graft
  nodes seen in our reviews descend from PTN000390097.
• The IBD was placed one node too high (too ancient; the family review's
  TOO_DEEP, below, says the same): on the duplication
  immediately above the F1-β / FliI–SctN split rather than on the F1-β clade.
  TreeGrafter faithfully reproduces the error rather than causing it.
• Two more pipelines add errors of their own:
• InterPro2GO maps the SctN-specific entry IPR013380 to GO:0046961 proton-
  transporting ATPase activity, rotational mechanism and to GO:0006754 ATP
  biosynthetic process.
• It maps the shared F1/V1/A1 N-terminal domain IPR004100 to GO:1902600 proton
  transmembrane transport and GO:0046034 ATP metabolic process.
• GO's logical inference (GO_REF:0000108) then turns each wrong molecular
  function into a wrong process through part_of links: GO:0046933 is
  part_of GO:0015986, and GO:0046961 is_a GO:0015078 proton
  transmembrane transporter activity, which is part_of GO:1902600.
• Nine full reviews reject every such row: 35 rows, 31 REMOVE and 4
  MARK_AS_OVER_ANNOTATED, with none accepted
  (per-row table).

Why the terms are wrong

The reviews give the evidence with verbatim quotes. In brief:

• FliI/SctN are ATP-hydrolysing protein-export ATPases. They are soluble,
  peripheral ATPases that dock export chaperone–substrate complexes, release the
  chaperones, and unfold and hand on the substrates. Their enzyme class is
  EC 7.4.2.8, protein-secreting ATPase (a translocase class; formerly EC 3.6.3.50).
• FliI's ATPase is not an F-type activity. It is insensitive to inhibitors of
  F-, V- and P-type ATPases (PMID:8943245).
• Protons move through the export gate, not the ATPase. In the flagellar
  system the proton motive force is used at the membrane export gate, a
  proton–protein antiporter in which Salmonella FlhA is implicated in the
  proton path (PMID:21934659, PMID:29946050). Injectisome export is also
  PMF-dependent, as the CCCP sensitivity of Ysc secretion shows (PMID:15213145).
  In neither system is the soluble FliI/SctN ATPase a rotary proton pump.
• The resemblance to F1 is real but does not make them synthases. FliJ/SctO is
  γ-stalk-like, and the FliI₆–FliJ ring resembles F1 (PMID:21278755, PMID:17202259).
  But the ring still hydrolyses ATP and has no Fo sector to couple to.

Reviews

Actions by pipeline:

• The four over-annotations are all GO:0046034 ATP metabolic process. It is
  literally true of an ATPase but adds nothing beyond GO:0016887.
• What replaces the wrong terms. The correct molecular function, GO:0008564
  protein-exporting ATPase activity, is kept where it was already in GOA and
  added (NEW) where it was missing.

Other errors the reviews found along the way:

• UniProt naming. The flagellar FliI entries (P26465, P52612, P0CAT8, Q9ZJJ3)
  are named "Flagellum-specific ATP synthase" with EC 7.1.2.2, the
  H⁺-transporting synthase EC, and keep a 1991 "proton translocase" hypothesis in
  their FUNCTION text (PMID:1646201). The injectisome SctN entries already use
  EC 7.4.2.8.
• SsaN host-cell locations. Two AgBase IMP rows place SsaN in host cell
  cytoplasm and host cell membrane. The paper fractionates bacterial cells, and
  the effector, not the ATPase, is what reaches the host (see
  SALTY/sctN2).

Where this sits among the TreeGrafter failure modes

• Unlike most of the failure modes, this is not a problem with the term attached
  to the graft node, nor a mis-graft. The TreeGrafter placements are correct: every
  graft node is inside the FliI/SctN clade.
• The error comes in from PAINT. The ancestral IBD sits on the duplication node,
  so it covers both paralog clades. This is a fifth pattern: inherited PAINT
  over-placement. Fixing the IBD fixes the TreeGrafter output with no change to
  TreeGrafter.
• Earlier reviews did not catch it. The PTHR15184 family review
  (interpro/panther/PTHR15184/PTHR15184-review.yaml) had rated the PTN008558586
  assertions SOUND, having checked the node against its F1-β descendants only. It
  now records both as TOO_DEEP, lists SF9, SF62 and SF81 as diverged
  subfamilies, and scopes GO:0046933 to the F1-β subfamilies and GO:0008564 to
  the export-ATPase subfamilies (updated 2026-09-27).
• In the headline figures since 2026-10-01. All five TreeGrafter rows (fliI
  in CAUVC, HELPJ and PSEPK; sctN in SHIFL and YEREN) are in the main page's
  tables, pinned at f81b9f300 (and in the earlier 943b98815 refresh), all
  REMOVE. In its four-mode classification
  they are filed as mode 1 by the operational rule there (correct subfamily,
  wrong term from a node above it); this page identifies that node term as an
  inherited PAINT IBD.

Candidate upstream fixes (not filed)

1. PANTHER/PAINT: move the GO:0046933 and GO:0045259 IBDs from
   PTN008558586 to PTN008558588, or add an IRD/NOT at PTN000390097. This one
   change fixes both the IBA rows and every TreeGrafter row.
2. InterPro2GO:
3. Drop GO:0046961 and GO:0006754 from IPR013380. That entry is SctN-specific,
   so the mapping is wrong for every member.
4. Remove the IPR004100 → GO:1902600 / GO:0046034 mappings, or move them to
   narrower entries that exclude FliI/SctN. Ideally they would be suppressed only
   when IPR005714 also matches, but a per-entry InterPro2GO mapping probably
   cannot express that condition.
5. Drop the stale IPR005714 → GO:0009058 biosynthetic process mapping.
6. All five are recorded in the InterPro mapping review
   (mapping set).
7. UniProt: rename the flagellar FliI entries to a flagellar export ATPase, use
   EC 7.4.2.8, and drop the "proton translocase" clause.

Further cases

The same design applies to any function with near-universal presence in a
clade and an annotation-independent marker set, for example bacterial RNA
polymerase core subunits, the SecYEG translocon or the bacterial ribosomal
proteins. None has been run yet. For each, the false positives are the part
worth reading.

Reproduce

uv run --with requests projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/export_atpase_scan.py # UniProt counts + PTN008558586 IBA targets
uv run --with requests projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/node_placement.py     # PANTHER tree placement of IBD and graft nodes
uv run --with pyyaml   projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/review_actions.py     # per-row reviewer actions from the gene reviews
uv run --with requests projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/completeness_fetch.py # bacterial F-type completeness (large; to ./cache)
python3 projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/completeness_analyze.py > projects/CONSERVED_FUNCTION_RECOVERY/f_type_atp_synthase/completeness_results.txt

The UniProt, QuickGO and PANTHER queries run against live services, so the counts
will drift between releases. The committed TSVs were generated on 2026-09-27.