ARBA00004173 mitochondrion (GO:0005739)

View original ARBA rule on UniProt

Type: ARBA
Status: COMPLETE
Action: MODIFY
Confidence: 0.35

Description

Aggregated ARBA localization rule that asserts the UniProt SUBCELLULAR LOCATION "Mitochondrion" (SL-0173) for eukaryotic proteins matching any of 1,490 OR-ed condition sets built from CATH FunFam, InterPro and PANTHER signatures. Via the UniProtKB-SubCell -> GO mapping (GO_REF:0000044) the assertion is propagated to GO:0005739 (mitochondrion) with evidence IEA / ECO:0007322. The rule currently annotates 536,854 unreviewed (TrEMBL) proteins and 0 reviewed proteins. It is not a single-family rule but an umbrella aggregating hundreds of independent branch-level classifiers of very uneven quality, ranging from canonical mitochondrial families (MIC60, TOM40, TIM/Tim10-like, cytochrome c oxidase subunits, sideroflexins) to bare single unlabeled FunFams and clade-overfitted singletons.

Analysis Summary

Condition-set counts describe the sets recorded in this review, which may omit the full rule.

0
Domain Pairs Analyzed
12
Recorded condition sets
0
Subset Relationships
0
Redundant Annotations

Review Summary

ARBA00004173 is not one rule but an umbrella of 1,490 independent, OR-ed branch-level classifiers that all emit the same UniProt SUBCELLULAR LOCATION value (Mitochondrion, SL-0173), which GO_REF:0000044 converts to GO:0005739 (IEA/ECO:0007322) on 536,854 unreviewed proteins with zero reviewed proteins acting as controls. The rule is genuinely heterogeneous rather than uniformly bad. Many branches are unimpeachable - MIC60 (CS15), TOM40 (CS74), Tim10-like translocase subunits (CS23), mitochondrially encoded cytochrome b (CS1), cytochrome c oxidase subunits, sideroflexins (CS38) - and for these GO:0005739 is correct, if often underspecified relative to available submitochondrial child terms. The GO curator complaint (geneontology/go-annotation#6412, @hattrill, about Drosophila ScpX / Q9VJ43) exposes a specific and structural defect rather than a one-off mistake. The branch whose signature pair matches Q9VJ43's architecture is CS779 (FunFam 3.30.1050.10:FF:000001 AND FunFam 3.40.47.10:FF:000016 AND NOT Bacteria/Archaea/Viruses), i.e. it requires the N-terminal thiolase domain in addition to the C-terminal SCP2 domain. But in the mouse orthologue P32020 the mitochondrion assertion is scoped by UniProt to "Isoform SCP2" (P32020-2), the short alternative-initiation product that LACKS the thiolase domain, while the comment scoped to "Isoform SCPx" (P32020-1, the full-length architecture CS779 actually detects) lists Peroxisome only. ARBA has evidently learned the SUBCELLULAR LOCATION comment while discarding its `molecule:` (isoform) qualifier, so the branch selects exactly the isoform that is not annotated mitochondrial. The curator's diagnosis ("only the short isoform P32020_2 is potentially mitochondrial") is correct and, on the UniProt record, even understated. The same SCPx architecture simultaneously fires ARBA00004496 (Cytoplasm; CS2894 is the identical FunFam pair) and ARBA00004275 (Peroxisome; CS80 is 3.40.47.10:FF:000016 alone), so the three sibling location rules jointly reproduce the union of the two isoforms' recorded locations, of which only Peroxisome is right for SCPx. How far this generalizes is not established here: 139/1,490 (9.3%) of this rule's condition sets have a signature set that also appears verbatim in the Cytoplasm rule and 19/1,490 (1.3%) in the Peroxisome rule (3 in both), but co-annotation to mitochondrion plus cytoplasm or peroxisome is routine and often correct (fumarase, aconitase, several aminoacyl-tRNA synthetases, beta-oxidation enzymes), so these counts are an UPPER BOUND on candidate sets for per-set triage, not by themselves evidence of collapse. Distinguishing genuine dual localization from isoform collapse requires inspecting each set - or a baseline measuring how often reviewed proteins matching those signatures legitimately carry both SL values - and neither was done. The isoform-collapse mechanism itself is verified for exactly one orthologue, mouse P32020; human P22307 and rat P11915 were not checked, and ARBA's actual training provenance for CS779 was not established. Independently of scale, the literature actively argues against mitochondrial residence for this family: Li, Fan & Papadopoulos (PMID:26901662) showed by live-cell confocal imaging that the N-terminal presequence of SCPX/SCP2 is not sufficient to direct the proteins to mitochondria, and that PTS1-mediated peroxisomal targeting dominates. Structurally the rule is far outside any parsimony envelope: 693 sets (46.5%) are a single FunFam plus taxon (240 with no resolved label at all), 1,303 (87%) have no positive taxon condition beyond NOT(Bacteria/Archaea/Viruses), only 422 (28%) contain any signature whose label mentions "mitochondri*" (a weak proxy), and of the 885 distinct InterPro entries used, 493 have some InterPro2GO mapping but only 35 are mapped to GO:0005739 by InterPro curators. That last figure is a weak proxy too and is NOT evidence that InterPro curators rejected a mitochondrial location: InterPro2GO maps a term only where it holds for every match, its CC coverage is deliberately sparse, and - decisively - this rule's condition sets are conjunctive, so for the 131 InterPro+InterPro+PANTHER, 99 InterPro x3, 99 InterPro+PANTHER and 67 InterPro x2 sets no individual entry need imply mitochondrion for the set to. Only the 81 single-InterPro sets permit a like-for-like comparison, and that restricted statistic was not computed. Recommended action is MODIFY rather than DEPRECATE: the biologically sound branches are numerous and valuable, and wholesale retirement would lose a large amount of correct annotation. The deeper fix is SPLIT into per-family rules, mirroring the falcon report's framing that the rule should survive only "as a container for branch-level classifiers". Caveats on this review: `just analyze-rule ARBA00004173` refuses to run on rules with more than 12 condition sets, so no pairwise overlap statistics, heatmap or `entries` view could be generated, and `sync-rule-review-single` / `render-rule` are not runnable. Only one deep-research provider succeeded (falcon); the perplexity run failed with HTTP 401 (quota exhausted). Twelve illustrative condition sets are enumerated above with their true 1-based numbers; the other 1,478 have not been individually adjudicated.

Action Rationale

MODIFY, not DEPRECATE. Unlike a rule that is wrong throughout, ARBA00004173 contains a large core of correct branches keyed on whole-protein mitochondrial families (MIC60, TOM40, Tim chaperones, COX subunits, sideroflexins, MT-encoded OXPHOS subunits) for which GO:0005739 is exactly the right assertion; deleting the rule would discard correct annotation on a large fraction of 536,854 proteins. The defect the GO curators reported is specific, mechanistically identifiable, and fixable: ARBA is learning UniProt SUBCELLULAR LOCATION comments without honouring their `molecule:` (isoform) scoping, so an architecture-specific signature (thiolase + SCP2 FunFam pair, CS779) inherits a location that UniProt records only for a different, domain-deficient isoform (P32020-2). Because the failure is a property of the training procedure rather than of one hand-written branch, it warrants a targeted engineering fix plus branch pruning rather than either blanket acceptance or blanket removal. Its prevalence is unquantified: 139 of this rule's condition sets share a signature set with the Cytoplasm rule ARBA00004496 and 19 with the Peroxisome rule ARBA00004275, but genuine dual localization produces the same signal, so those counts bound the triage list from above rather than counting defects, and the mechanism is confirmed only for mouse P32020. A secondary, independent problem is the very large tail of low-information branches (693 single-FunFam sets, 240 unlabeled) and 140 singleton positive taxon clades that look like training-set residue. If UniProt is unwilling to make ARBA isoform-aware, the fallback should be SPLIT into per-family rules so that each branch can be accepted, audited or retired on its own evidence.

GO Annotations

GO:0005739 - mitochondrion
Aspect: C

Rule Definition

Condition Sets

Condition Set 1

3 condition(s)
Notes:

Sound. Mitochondrially encoded cytochrome b (MT-CYB and orthologues) is a bona fide mitochondrial inner-membrane protein, so GO:0005739 is correct (arguably underspecified relative to mitochondrial respiratory chain complex III / inner membrane). I specifically checked the plant chloroplast cytochrome b6f false-positive risk and it does not materialise: Arabidopsis cytochrome b6 (P56773) carries IPR005797 (N-terminal), not IPR005798, and plant petD (e.g. P56774, P05643, Q332U6) carries IPR005798 but not IPR027387, so the two-InterPro conjunction excludes the plastid subunits. This is an example of the rule's good branches.

Condition Set 2

3 condition(s)
Notes:

Biologically correct target (COX2 is mitochondrially encoded and resides in the mitochondrial inner membrane) but the positive taxon condition is an artefact: COX2 is pan-eukaryotic, yet this branch is confined to Euarchontoglires, and the near-identical CS13 (IPR036257 + IPR045187) is confined to Laurasiatheria. This is a recall-losing, training-set-shaped clade restriction rather than a biological one; the two branches should be merged and generalised.

Condition Set 3

4 condition(s)
Notes:

Sound single-signature branch. IPR019133 is a whole-protein family (MICOS/mitofilin MIC60) that is definitionally mitochondrial inner-membrane, so a single InterPro family plus the prokaryote/virus exclusion is acceptable here. This is exactly the case the falcon report says is permissible for single-InterPro branches, and it contrasts sharply with the bare single-FunFam branches (CS608, CS812) below.

Condition Set 4

6 condition(s)
Notes:

Sound but internally redundant. The small Tim chaperones are intermembrane space proteins, so GO:0005739 is correct. However IPR004217 (domain), IPR035427 (homologous superfamily of that same domain) and IPR050673 (family) all describe the same sequence region; the conjunction adds no discriminating power over IPR050673 alone. This shape (nested domain/superfamily/family triples) recurs throughout the rule and is the main driver of the SIGNIFICANT condition-overlap assessment.

Condition Set 5

4 condition(s)
Notes:

The falcon deep-research report's flagship positive example: IPR004686 is the sideroflexin (SFXN) family, whose members are multi-pass mitochondrial inner-membrane transporters. GO:0005739 is well supported; the report argues mitochondrial inner membrane would be more informative for this branch. Retain, and consider propagating a submitochondrial child term.

Condition Set 6

6 condition(s)
Notes:

Sound. TOM40 is the mitochondrial outer-membrane import channel. Here the nesting is defensible because IPR023614 is a promiscuous beta-barrel superfamily and IPR037930 supplies the specificity; the conjunction usefully excludes VDACs from the generic porin superfamily hit. Retain.

Condition Set 7

6 condition(s)
Notes:

Included for contrast with the issue #6412 complaint. This is the OTHER SCP2-domain-containing branch in the rule - it targets HSDL2 (SDR + SCP2 domain), not the SCPx/thiolase architecture. Q9VJ43 does NOT match it (no SDR signature). It should therefore not be confused with the branch that actually fires on the Drosophila protein; it is CS779, not this set, that produces the reported false positive.

Condition Set 8

6 condition(s)
Notes:

A thiolase branch pinned to a specific PANTHER subfamily (PTHR18919:SF153). Q9VJ43 is assigned to PTHR42870:SF1 (NON-SPECIFIC LIPID-TRANSFER PROTEIN-LIKE 2), so this set does not match it. Included as a positive illustration of the falcon report's point that a narrow PANTHER subfamily condition is a much safer localization proxy than a bare structural superfamily/FunFam, because paralogues within a thiolase-like fold differ in targeting (peroxisomal vs mitochondrial vs cytosolic).

Condition Set 9

4 condition(s)
Notes:

Worst-case shape, included as a representative of the unlabeled single-FunFam class. CATH 2.130.10.10 is a beta-propeller superfamily - one of the most promiscuous folds in the proteome, occurring in nuclear, cytosolic, membrane-trafficking and mitochondrial proteins alike. A single unlabeled FunFam from it, conjoined only with NOT(Bacteria/Archaea/Viruses), provides no localization signal whatsoever. Branches of this shape should be dropped unless individually benchmarked against reviewed entries.

Condition Set 10

5 condition(s)
Notes:

THE DEFECTIVE BRANCH underlying geneontology/go-annotation#6412. This is the condition set whose signature pair matches the architecture of Drosophila ScpX (Q9VJ43, 544 aa): an N-terminal thiolase domain (CATH 3.40.47.10, IPR020616/IPR055140/IPR016039) plus a C-terminal SCP2 sterol-binding domain (CATH 3.30.1050.10, IPR003033/IPR036527). Exhaustive scanning of all 1,490 condition sets shows only CS779 and CS812 could match Q9VJ43's signature complement; the InterPro/PANTHER SCP2 sets (CS249, CS303) cannot. (I could not query CATH FunFam assignments for Q9VJ43 directly, so this is an architecture-level match, not a verified FunFam hit.) The defect: the mitochondrial evidence for this family comes exclusively from an ISOFORM-SCOPED UniProt annotation. In mouse SCP2_MOUSE (P32020), ALTERNATIVE PRODUCTS records alternative initiation with two isoforms; the SUBCELLULAR LOCATION comment scoped to "Isoform SCPx" (P32020-1, the full-length thiolase+SCP2 protein that this FunFam pair detects) lists Peroxisome ONLY (ECO:0000269, PMID:26901662), whereas the comment scoped to "Isoform SCP2" (P32020-2, the short alternative-initiation product that LACKS the thiolase domain) lists Cytoplasm, Peroxisome, Endoplasmic reticulum and Mitochondrion (ECO:0000269, PMID:11003606 and PMID:26901662). ARBA has learned the SUBCELLULAR LOCATION comment while discarding its `molecule:` (isoform) qualifier, so it requires the thiolase FunFam - i.e. selects precisely the isoform that is NOT annotated mitochondrial. The same collapse is visible across sibling rules: the identical FunFam pair is CS2894 of ARBA00004496 (Cytoplasm), and 3.40.47.10:FF:000016 alone is CS80 of ARBA00004275 (Peroxisome), so this one architecture triggers three location rules and reproduces the union of the two isoforms' locations. Only the peroxisome call is consistent with what UniProt records for isoform SCPx. RETIRE or repair this branch (and CS2894 of ARBA00004496).

Condition Set 11

4 condition(s)
Notes:

The second (and only other) set that could match Q9VJ43. A single, entirely UNLABELED FunFam from the thiolase CATH superfamily 3.40.47.10 plus the prokaryote/virus exclusion. Thiolase-fold proteins are distributed across mitochondria, peroxisomes and cytosol, so a bare FunFam from this superfamily cannot discriminate compartment. Representative of the 693 single-FunFam branches (46.5% of the rule), 240 of which carry no resolved label at all.

Condition Set 12

2 condition(s)
Notes:

Representative clade-overfitting artefact. The FunFam label identifies ACAT1, the mitochondrial matrix acetoacetyl-CoA thiolase - a genuinely pan-metazoan (indeed pan-eukaryotic) mitochondrial enzyme - yet the branch is restricted to Anura (frogs). 187 of 1,490 sets carry a positive taxon condition and they use 140 distinct clades, each appearing only once or twice (Anura, Pongo, Bos, Homo, Mus, Kluyveromyces, Oryza, Camelineae, "Aspergillus subgen. Circumdati", ...). These read as training-set clade artefacts rather than biology: they neither add precision nor reflect the taxonomic distribution of the family, and they silently sacrifice recall.

Assessments

OVERLY_COMPLEX

1,490 OR-ed condition sets, i.e. ~124x the 12-condition-set threshold above which the project's own tooling refuses to analyse a rule. The complexity is not merely large but poorly structured: 693 sets (46.5%) are a single FunFam plus taxon, 240 of those FunFams have no resolved label at all; 180 sets are two FunFams, 131 are InterPro+InterPro+PANTHER, 102 are three FunFams, 81 a single InterPro, 19 a single PANTHER. Condition-type usage is taxon 1,490, FunFam 1,359, InterPro 885, PANTHER 287. All 1,490 signature sets are distinct (no exact duplicates), so the complexity is not literal duplication - it is the absence of any shared mechanistic basis: the sets have nothing in common except emitting the same location string. Nested InterPro triples (domain + homologous superfamily + family for the same region, e.g. CS23) add conditions without adding discrimination. A rule of this shape cannot be reviewed, benchmarked or maintained as a unit, which is why the deeper recommendation is to split it into per-family rules.

MODERATE

Literature support is bimodal, and a single enum cannot express that; MODERATE is chosen as the rule-level summary. For a substantial minority of branches the support is effectively STRONG: MIC60, TOM40, the small Tim chaperones, mitochondrially encoded cytochrome b and cytochrome c oxidase subunits, and sideroflexins are textbook mitochondrial proteins with abundant experimental localization evidence, and the falcon report singles out the sideroflexin branch (CS38/IPR004686) as well supported. For the branch the GO curator complained about, support is CONTRADICTED: Li, Fan & Papadopoulos (PMID:26901662) tested the putative N-terminal mitochondrial presequence of SCPX/SCP2 directly by live-cell confocal imaging of fluorescent fusions and concluded it is not sufficient to localize either protein to mitochondria, while C-terminal PTS1-mediated peroxisomal targeting is. The only primary evidence for mitochondrial SCP2 is Starodub et al. (PMID:11003606), an immunofluorescence study of the SHORT SCP-2 protein in L-cell fibroblasts that ranks mitochondria third behind peroxisomes and ER - and UniProt correctly scopes that evidence to isoform P32020-2, not to the SCPx architecture that CS779 detects. For the very large tail of unlabeled single-FunFam branches there is no identifiable literature at all, because the branches cannot even be resolved to a named family. The falcon report additionally warns that rules trained on noisy source annotations reinforce systematic errors, which is precisely what the isoform-collapse defect demonstrates.

Supporting Evidence:

  • PMID:26901662: the N-terminal mitochondrial targeting sequence is not powerful enough to localize the SCPX and SCP2 proteins to the mitochondria.
  • PMID:26901662: The Scp2 gene contains two transcription initiation sites, giving rise to the 58 kDa sterol carrier protein-x (SCPX) and 15 kDa pro-SCP2 proteins
  • PMID:11003606: detected SCP-2 in peroxisomes > endoplasmic reticulum > mitochondria >
  • file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md: Sideroflexins are mitochondrial inner-membrane proteins, although their transported substrates and physiological roles differ among paralogues. That branch is well supported for GO:0005739.
  • file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md: Thus, family rules trained on noisy source annotations can reinforce systematic errors.
  • https://github.com/geneontology/go-annotation/issues/6412: From orthology to P32020, only the sort isoform P32020_2 is potentially mitochondrial.
SIGNIFICANT

Quantitative pairwise overlap could not be computed because `just analyze-rule` refuses to run on rules with >12 condition sets, so `pairwise_overlap` and `entries` are empty. Qualitative and cross-rule evidence nonetheless makes SIGNIFICANT the right call, on three distinct axes. (1) Within-set redundancy: the recurrent InterPro triple shape (domain + its homologous superfamily + the corresponding family, 99 sets of shape InterPro x3, e.g. CS23 = IPR004217 + IPR035427 + IPR050673) conjoins nested annotations of the same sequence region, which cannot increase specificity. (2) Between-set overlap within the rule: 693 single-FunFam sets drawn from broad CATH superfamilies will match overlapping protein sets, and multi-FunFam sets built from the same superfamily (3.40.47.10 appears in CS779, CS812, CS1124 among others) partially subsume one another. (3) Cross-rule overlap, which is the most consequential: 139/1,490 (9.3%) of this rule's condition sets have a signature set that appears identically in the Cytoplasm rule ARBA00004496, 19/1,490 (1.3%) in the Peroxisome rule ARBA00004275, and 3 in both. These compartments are not mutually exclusive - mitochondrion/cytoplasm and mitochondrion/peroxisome co-annotation is routine and frequently correct (fumarase, aconitase, several aminoacyl-tRNA synthetases, beta-oxidation enzymes), and the falcon report makes the same point: "For dual-localized families, GO:0005739 should coexist with the other supported component rather than suppress it." The 139 + 19 counts are therefore an upper bound on the sets that warrant per-set triage, not a count of collapsed antecedents; separating true dual localization from isoform collapse requires inspecting each set, which this review did not do. What is demonstrated is the single worked example CS779, where the shared signature requires the thiolase FunFam that isoform P32020-2 lacks: the same FunFam pair is CS2894 of ARBA00004496, and its thiolase half alone is CS80 of ARBA00004275. Separately, of the 885 distinct InterPro entries used by this rule, 493 already carry some InterPro2GO mapping but only 35 are mapped to GO:0005739 by InterPro curators. This is an argument from silence and should not be read as InterPro curators declining a mitochondrial claim: InterPro2GO maps a term only where it holds for all matches, CC coverage there is sparse by design, and this rule's condition sets are conjunctive, so in the 131 InterPro+InterPro+PANTHER, 99 InterPro x3, 99 InterPro+PANTHER and 67 InterPro x2 sets no single entry need imply mitochondrion for the set to. The comparison is like-for-like only in the 81 single-InterPro sets, and that restricted count was not computed.

APPROPRIATE

GO:0005739 (mitochondrion) is the right granularity for a rule whose source assertion is the UniProt subcellular-location keyword Mitochondrion (SL-0173): for every true positive it is a correct, conservative cellular-component term, and it is the term GO_REF:0000044 legitimately produces from that keyword. The aspect and branch are right (cellular component, not molecular function or biological process), so it is not MISMATCHED, and it is certainly not TOO_NARROW. The honest qualifier is that it is frequently UNDERSPECIFIED rather than wrong: many branches identify families whose submitochondrial compartment is well established (inner membrane for sideroflexins and OXPHOS subunits, outer membrane for TOM40, intermembrane space for the Tim chaperones) and could support a child term in addition to the parent. I have deliberately not scored this TOO_BROAD, because the false positives this rule generates - Q9VJ43 being the reported case - are not failures of term granularity at all: no mitochondrial child or parent term would be correct there. Fixing them requires fixing the antecedent (isoform scoping, branch pruning), not the consequent. Scoring TOO_BROAD would misdirect the fix.

TOO_BROAD

The dominant taxonomic condition is NOT(Bacteria/Archaea/Viruses), present as the ONLY taxon constraint in 1,303/1,490 sets (87%). This is a coarse "eukaryote-ish" filter and carries almost no organelle information: it does not distinguish mitochondria from cytosol, nucleus, ER, Golgi, peroxisome or - in plants and algae - plastid, and it is largely redundant with signatures that are already eukaryote-restricted. Combined with the 693 bare single-FunFam branches, this is the principal false-positive generator, and it is exactly how the Drosophila ScpX case arises: CS779's taxonomic condition does nothing to prevent a peroxisomal/cytosolic lipid-transfer protein from being called mitochondrial. A separate, opposite artefact coexists with this over-breadth and is worth recording even though the enum can only carry one value: the 187 sets that do have a positive taxon condition use 140 distinct clades, each occurring only once or twice (Anura, Pongo, Bos, Homo, Mus, Kluyveromyces, Oryza, Camelineae, "Aspergillus subgen. Circumdati", ...). These are TOO_NARROW in effect and read as training-set residue rather than biology - CS1124 confines the FunFam labelled "acetyl-CoA acetyltransferase, mitochondrial" (ACAT1), a pan-metazoan mitochondrial matrix thiolase, to frogs; CS9 and CS13 split cytochrome c oxidase subunit II between Euarchontoglires and Laurasiatheria. They sacrifice recall without buying precision. The rule therefore manages to be simultaneously too broad where it matters (the 87% with no positive constraint) and arbitrarily too narrow where it does constrain.

References (7)

Raw YAML

View Source YAML
id: ARBA00004173
description: >-
  Aggregated ARBA localization rule that asserts the UniProt SUBCELLULAR LOCATION
  "Mitochondrion" (SL-0173) for eukaryotic proteins matching any of 1,490 OR-ed
  condition sets built from CATH FunFam, InterPro and PANTHER signatures. Via the
  UniProtKB-SubCell -> GO mapping (GO_REF:0000044) the assertion is propagated to
  GO:0005739 (mitochondrion) with evidence IEA / ECO:0007322. The rule currently
  annotates 536,854 unreviewed (TrEMBL) proteins and 0 reviewed proteins. It is not
  a single-family rule but an umbrella aggregating hundreds of independent
  branch-level classifiers of very uneven quality, ranging from canonical
  mitochondrial families (MIC60, TOM40, TIM/Tim10-like, cytochrome c oxidase
  subunits, sideroflexins) to bare single unlabeled FunFams and clade-overfitted
  singletons.
status: COMPLETE
rule_type: ARBA
rule:
  rule_id: ARBA00004173
  # 1490 condition sets total; 12 illustrative sets enumerated below with their
  # true 1-based `number` from rules/arba/ARBA00004173/ARBA00004173.enriched.json.
  # The full set cannot be enumerated here and `just analyze-rule` refuses to run
  # on rules with >12 condition sets, so pairwise_overlap and entries stay empty.
  condition_sets:
  - number: 1
    conditions:
    - condition_type: INTERPRO
      value: IPR005798
      curie: InterPro:IPR005798
      label: Cytochrome b/b6, C-terminal
      interpro_type: DOMAIN
    - condition_type: INTERPRO
      value: IPR027387
      curie: InterPro:IPR027387
      label: Cytochrome b/b6-like domain superfamily
      interpro_type: HOMOLOGOUS_SUPERFAMILY
    - condition_type: TAXON
      value: Eukaryota
      curie: NCBITaxon:2759
      label: Eukaryota
    notes: >-
      Sound. Mitochondrially encoded cytochrome b (MT-CYB and orthologues) is a
      bona fide mitochondrial inner-membrane protein, so GO:0005739 is correct
      (arguably underspecified relative to mitochondrial respiratory chain complex
      III / inner membrane). I specifically checked the plant chloroplast
      cytochrome b6f false-positive risk and it does not materialise: Arabidopsis
      cytochrome b6 (P56773) carries IPR005797 (N-terminal), not IPR005798, and
      plant petD (e.g. P56774, P05643, Q332U6) carries IPR005798 but not
      IPR027387, so the two-InterPro conjunction excludes the plastid subunits.
      This is an example of the rule's good branches.
  - number: 9
    conditions:
    - condition_type: INTERPRO
      value: IPR001505
      curie: InterPro:IPR001505
      label: Copper centre Cu(A)
      interpro_type: BINDING_SITE
    - condition_type: INTERPRO
      value: IPR002429
      curie: InterPro:IPR002429
      label: Cytochrome c oxidase subunit II-like C-terminal
      interpro_type: DOMAIN
    - condition_type: TAXON
      value: Euarchontoglires
      curie: NCBITaxon:314146
      label: Euarchontoglires
    notes: >-
      Biologically correct target (COX2 is mitochondrially encoded and resides in
      the mitochondrial inner membrane) but the positive taxon condition is an
      artefact: COX2 is pan-eukaryotic, yet this branch is confined to
      Euarchontoglires, and the near-identical CS13 (IPR036257 + IPR045187) is
      confined to Laurasiatheria. This is a recall-losing, training-set-shaped
      clade restriction rather than a biological one; the two branches should be
      merged and generalised.
  - number: 15
    conditions:
    - condition_type: INTERPRO
      value: IPR019133
      curie: InterPro:IPR019133
      label: MICOS complex subunit MIC60
      interpro_type: FAMILY
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      Sound single-signature branch. IPR019133 is a whole-protein family
      (MICOS/mitofilin MIC60) that is definitionally mitochondrial inner-membrane,
      so a single InterPro family plus the prokaryote/virus exclusion is
      acceptable here. This is exactly the case the falcon report says is
      permissible for single-InterPro branches, and it contrasts sharply with the
      bare single-FunFam branches (CS608, CS812) below.
  - number: 23
    conditions:
    - condition_type: INTERPRO
      value: IPR004217
      curie: InterPro:IPR004217
      label: Tim10-like
      interpro_type: DOMAIN
    - condition_type: INTERPRO
      value: IPR035427
      curie: InterPro:IPR035427
      label: Tim10-like domain superfamily
      interpro_type: HOMOLOGOUS_SUPERFAMILY
    - condition_type: INTERPRO
      value: IPR050673
      curie: InterPro:IPR050673
      label: Mitochondrial import inner membrane translocase subunit
      interpro_type: FAMILY
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      Sound but internally redundant. The small Tim chaperones are intermembrane
      space proteins, so GO:0005739 is correct. However IPR004217 (domain),
      IPR035427 (homologous superfamily of that same domain) and IPR050673
      (family) all describe the same sequence region; the conjunction adds no
      discriminating power over IPR050673 alone. This shape (nested
      domain/superfamily/family triples) recurs throughout the rule and is the
      main driver of the SIGNIFICANT condition-overlap assessment.
  - number: 38
    conditions:
    - condition_type: INTERPRO
      value: IPR004686
      curie: InterPro:IPR004686
      label: Tricarboxylate/iron carrier
      interpro_type: FAMILY
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      The falcon deep-research report's flagship positive example: IPR004686 is
      the sideroflexin (SFXN) family, whose members are multi-pass mitochondrial
      inner-membrane transporters. GO:0005739 is well supported; the report argues
      mitochondrial inner membrane would be more informative for this branch.
      Retain, and consider propagating a submitochondrial child term.
  - number: 74
    conditions:
    - condition_type: INTERPRO
      value: IPR023614
      curie: InterPro:IPR023614
      label: Porin domain superfamily
      interpro_type: HOMOLOGOUS_SUPERFAMILY
    - condition_type: INTERPRO
      value: IPR027246
      curie: InterPro:IPR027246
      label: Eukaryotic porin/Tom40
      interpro_type: FAMILY
    - condition_type: INTERPRO
      value: IPR037930
      curie: InterPro:IPR037930
      label: Tom40
      interpro_type: FAMILY
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      Sound. TOM40 is the mitochondrial outer-membrane import channel. Here the
      nesting is defensible because IPR023614 is a promiscuous beta-barrel
      superfamily and IPR037930 supplies the specificity; the conjunction usefully
      excludes VDACs from the generic porin superfamily hit. Retain.
  - number: 249
    conditions:
    - condition_type: INTERPRO
      value: IPR002347
      curie: InterPro:IPR002347
      label: Short-chain dehydrogenase/reductase SDR
      interpro_type: FAMILY
    - condition_type: INTERPRO
      value: IPR003033
      curie: InterPro:IPR003033
      label: SCP2 sterol-binding domain
      interpro_type: DOMAIN
    - condition_type: INTERPRO
      value: IPR051935
      curie: InterPro:IPR051935
      label: Hydroxysteroid dehydrogenase-like protein 2
      interpro_type: FAMILY
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      Included for contrast with the issue #6412 complaint. This is the OTHER
      SCP2-domain-containing branch in the rule - it targets HSDL2
      (SDR + SCP2 domain), not the SCPx/thiolase architecture. Q9VJ43 does NOT
      match it (no SDR signature). It should therefore not be confused with the
      branch that actually fires on the Drosophila protein; it is CS779, not this
      set, that produces the reported false positive.
  - number: 303
    conditions:
    - condition_type: INTERPRO
      value: IPR016039
      curie: InterPro:IPR016039
      label: Thiolase-like
      interpro_type: HOMOLOGOUS_SUPERFAMILY
    - condition_type: INTERPRO
      value: IPR020613
      curie: InterPro:IPR020613
      label: Thiolase, conserved site
      interpro_type: CONSERVED_SITE
    - condition_type: PANTHER
      value: PTHR18919:SF153
      curie: PANTHER:PTHR18919:SF153
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      A thiolase branch pinned to a specific PANTHER subfamily (PTHR18919:SF153).
      Q9VJ43 is assigned to PTHR42870:SF1 (NON-SPECIFIC LIPID-TRANSFER
      PROTEIN-LIKE 2), so this set does not match it. Included as a positive
      illustration of the falcon report's point that a narrow PANTHER subfamily
      condition is a much safer localization proxy than a bare structural
      superfamily/FunFam, because paralogues within a thiolase-like fold differ in
      targeting (peroxisomal vs mitochondrial vs cytosolic).
  - number: 608
    conditions:
    - condition_type: FUNFAM
      value: 2.130.10.10:FF:000404
      curie: CATH.FunFam:2.130.10.10:FF:000404
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      Worst-case shape, included as a representative of the unlabeled
      single-FunFam class. CATH 2.130.10.10 is a beta-propeller superfamily -
      one of the most promiscuous folds in the proteome, occurring in nuclear,
      cytosolic, membrane-trafficking and mitochondrial proteins alike. A single
      unlabeled FunFam from it, conjoined only with NOT(Bacteria/Archaea/Viruses),
      provides no localization signal whatsoever. Branches of this shape should be
      dropped unless individually benchmarked against reviewed entries.
  - number: 779
    conditions:
    - condition_type: FUNFAM
      value: 3.30.1050.10:FF:000001
      curie: CATH.FunFam:3.30.1050.10:FF:000001
      label: Putative Non-specific lipid-transfer protein
    - condition_type: FUNFAM
      value: 3.40.47.10:FF:000016
      curie: CATH.FunFam:3.40.47.10:FF:000016
      label: Non-specific lipid-transfer protein
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      THE DEFECTIVE BRANCH underlying geneontology/go-annotation#6412. This is the
      condition set whose signature pair matches the architecture of Drosophila
      ScpX (Q9VJ43, 544 aa): an N-terminal thiolase domain (CATH 3.40.47.10,
      IPR020616/IPR055140/IPR016039) plus a C-terminal SCP2 sterol-binding domain
      (CATH 3.30.1050.10, IPR003033/IPR036527). Exhaustive scanning of all 1,490
      condition sets shows only CS779 and CS812 could match Q9VJ43's signature
      complement; the InterPro/PANTHER SCP2 sets (CS249, CS303) cannot. (I could
      not query CATH FunFam assignments for Q9VJ43 directly, so this is an
      architecture-level match, not a verified FunFam hit.) The defect: the
      mitochondrial evidence for this family comes exclusively from an
      ISOFORM-SCOPED UniProt annotation. In mouse SCP2_MOUSE (P32020), ALTERNATIVE
      PRODUCTS records alternative initiation with two isoforms; the SUBCELLULAR
      LOCATION comment scoped to "Isoform SCPx" (P32020-1, the full-length
      thiolase+SCP2 protein that this FunFam pair detects) lists Peroxisome ONLY
      (ECO:0000269, PMID:26901662), whereas the comment scoped to "Isoform SCP2"
      (P32020-2, the short alternative-initiation product that LACKS the thiolase
      domain) lists Cytoplasm, Peroxisome, Endoplasmic reticulum and Mitochondrion
      (ECO:0000269, PMID:11003606 and PMID:26901662). ARBA has learned the
      SUBCELLULAR LOCATION comment while discarding its `molecule:` (isoform)
      qualifier, so it requires the thiolase FunFam - i.e. selects precisely the
      isoform that is NOT annotated mitochondrial. The same collapse is visible
      across sibling rules: the identical FunFam pair is CS2894 of ARBA00004496
      (Cytoplasm), and 3.40.47.10:FF:000016 alone is CS80 of ARBA00004275
      (Peroxisome), so this one architecture triggers three location rules and
      reproduces the union of the two isoforms' locations. Only the peroxisome
      call is consistent with what UniProt records for isoform SCPx. RETIRE or
      repair this branch (and CS2894 of ARBA00004496).
  - number: 812
    conditions:
    - condition_type: FUNFAM
      value: 3.40.47.10:FF:000020
      curie: CATH.FunFam:3.40.47.10:FF:000020
    - condition_type: TAXON
      value: Archaea
      curie: NCBITaxon:2157
      label: Archaea
      negated: true
    - condition_type: TAXON
      value: Bacteria
      curie: NCBITaxon:2
      label: Bacteria
      negated: true
    - condition_type: TAXON
      value: Viruses
      curie: NCBITaxon:10239
      label: Viruses
      negated: true
    notes: >-
      The second (and only other) set that could match Q9VJ43. A single, entirely
      UNLABELED FunFam from the thiolase CATH superfamily 3.40.47.10 plus the
      prokaryote/virus exclusion. Thiolase-fold proteins are distributed across
      mitochondria, peroxisomes and cytosol, so a bare FunFam from this
      superfamily cannot discriminate compartment. Representative of the 693
      single-FunFam branches (46.5% of the rule), 240 of which carry no resolved
      label at all.
  - number: 1124
    conditions:
    - condition_type: FUNFAM
      value: 3.40.47.10:FF:000007
      curie: CATH.FunFam:3.40.47.10:FF:000007
      label: acetyl-CoA acetyltransferase, mitochondrial
    - condition_type: TAXON
      value: Anura
      curie: NCBITaxon:8342
      label: Anura
    notes: >-
      Representative clade-overfitting artefact. The FunFam label identifies
      ACAT1, the mitochondrial matrix acetoacetyl-CoA thiolase - a genuinely
      pan-metazoan (indeed pan-eukaryotic) mitochondrial enzyme - yet the branch
      is restricted to Anura (frogs). 187 of 1,490 sets carry a positive taxon
      condition and they use 140 distinct clades, each appearing only once or
      twice (Anura, Pongo, Bos, Homo, Mus, Kluyveromyces, Oryza, Camelineae,
      "Aspergillus subgen. Circumdati", ...). These read as training-set clade
      artefacts rather than biology: they neither add precision nor reflect the
      taxonomic distribution of the family, and they silently sacrifice recall.
  go_annotations:
  - go_id: GO:0005739
    go_label: mitochondrion
    aspect: C
  entries: []
  reviewed_protein_count: 0
  unreviewed_protein_count: 536854
  created_date: '2020-05-12'
  modified_date: '2025-12-14'
review_summary: >-
  ARBA00004173 is not one rule but an umbrella of 1,490 independent, OR-ed
  branch-level classifiers that all emit the same UniProt SUBCELLULAR LOCATION
  value (Mitochondrion, SL-0173), which GO_REF:0000044 converts to GO:0005739
  (IEA/ECO:0007322) on 536,854 unreviewed proteins with zero reviewed proteins
  acting as controls. The rule is genuinely heterogeneous rather than uniformly
  bad. Many branches are unimpeachable - MIC60 (CS15), TOM40 (CS74), Tim10-like
  translocase subunits (CS23), mitochondrially encoded cytochrome b (CS1),
  cytochrome c oxidase subunits, sideroflexins (CS38) - and for these GO:0005739
  is correct, if often underspecified relative to available submitochondrial
  child terms.

  The GO curator complaint (geneontology/go-annotation#6412, @hattrill, about
  Drosophila ScpX / Q9VJ43) exposes a specific and structural defect rather than a
  one-off mistake. The branch whose signature pair matches Q9VJ43's architecture
  is CS779 (FunFam 3.30.1050.10:FF:000001 AND FunFam 3.40.47.10:FF:000016 AND
  NOT Bacteria/Archaea/Viruses), i.e. it requires the N-terminal thiolase domain
  in addition to the C-terminal SCP2 domain. But in the mouse orthologue P32020
  the mitochondrion assertion is scoped by UniProt to "Isoform SCP2" (P32020-2),
  the short alternative-initiation product that LACKS the thiolase domain, while
  the comment scoped to "Isoform SCPx" (P32020-1, the full-length architecture
  CS779 actually detects) lists Peroxisome only. ARBA has evidently learned the
  SUBCELLULAR LOCATION comment while discarding its `molecule:` (isoform)
  qualifier, so the branch selects exactly the isoform that is not annotated
  mitochondrial. The curator's diagnosis ("only the short isoform P32020_2 is
  potentially mitochondrial") is correct and, on the UniProt record, even
  understated.

  The same SCPx architecture simultaneously fires ARBA00004496 (Cytoplasm; CS2894
  is the identical FunFam pair) and ARBA00004275 (Peroxisome; CS80 is
  3.40.47.10:FF:000016 alone), so the three sibling location rules jointly
  reproduce the union of the two isoforms' recorded locations, of which only
  Peroxisome is right for SCPx. How far this generalizes is not established here:
  139/1,490 (9.3%) of this rule's condition sets have a signature set that also
  appears verbatim in the Cytoplasm rule and 19/1,490 (1.3%) in the Peroxisome
  rule (3 in both), but co-annotation to mitochondrion plus cytoplasm or
  peroxisome is routine and often correct (fumarase, aconitase, several
  aminoacyl-tRNA synthetases, beta-oxidation enzymes), so these counts are an
  UPPER BOUND on candidate sets for per-set triage, not by themselves evidence of
  collapse. Distinguishing genuine dual localization from isoform collapse
  requires inspecting each set - or a baseline measuring how often reviewed
  proteins matching those signatures legitimately carry both SL values - and
  neither was done. The isoform-collapse mechanism itself is verified for exactly
  one orthologue, mouse P32020; human P22307 and rat P11915 were not checked, and
  ARBA's actual training provenance for CS779 was not established.
  Independently of scale, the literature actively argues against mitochondrial residence for this
  family: Li, Fan & Papadopoulos (PMID:26901662) showed by live-cell confocal
  imaging that the N-terminal presequence of SCPX/SCP2 is not sufficient to direct
  the proteins to mitochondria, and that PTS1-mediated peroxisomal targeting
  dominates.

  Structurally the rule is far outside any parsimony envelope: 693 sets (46.5%)
  are a single FunFam plus taxon (240 with no resolved label at all), 1,303
  (87%) have no positive taxon condition beyond NOT(Bacteria/Archaea/Viruses),
  only 422 (28%) contain any signature whose label mentions "mitochondri*" (a
  weak proxy), and of the 885 distinct InterPro entries used, 493 have some
  InterPro2GO mapping but only 35 are mapped to GO:0005739 by InterPro curators.
  That last figure is a weak proxy too and is NOT evidence that InterPro curators
  rejected a mitochondrial location: InterPro2GO maps a term only where it holds
  for every match, its CC coverage is deliberately sparse, and - decisively - this
  rule's condition sets are conjunctive, so for the 131 InterPro+InterPro+PANTHER,
  99 InterPro x3, 99 InterPro+PANTHER and 67 InterPro x2 sets no individual entry
  need imply mitochondrion for the set to. Only the 81 single-InterPro sets permit
  a like-for-like comparison, and that restricted statistic was not computed.

  Recommended action is MODIFY rather than DEPRECATE: the biologically sound
  branches are numerous and valuable, and wholesale retirement would lose a large
  amount of correct annotation. The deeper fix is SPLIT into per-family rules,
  mirroring the falcon report's framing that the rule should survive only "as a
  container for branch-level classifiers".

  Caveats on this review: `just analyze-rule ARBA00004173` refuses to run on rules
  with more than 12 condition sets, so no pairwise overlap statistics, heatmap or
  `entries` view could be generated, and `sync-rule-review-single` / `render-rule`
  are not runnable. Only one deep-research provider succeeded (falcon); the
  perplexity run failed with HTTP 401 (quota exhausted). Twelve illustrative
  condition sets are enumerated above with their true 1-based numbers; the other
  1,478 have not been individually adjudicated.
action: MODIFY
action_rationale: >-
  MODIFY, not DEPRECATE. Unlike a rule that is wrong throughout, ARBA00004173
  contains a large core of correct branches keyed on whole-protein mitochondrial
  families (MIC60, TOM40, Tim chaperones, COX subunits, sideroflexins, MT-encoded
  OXPHOS subunits) for which GO:0005739 is exactly the right assertion; deleting
  the rule would discard correct annotation on a large fraction of 536,854
  proteins. The defect the GO curators reported is specific, mechanistically
  identifiable, and fixable: ARBA is learning UniProt SUBCELLULAR LOCATION
  comments without honouring their `molecule:` (isoform) scoping, so an
  architecture-specific signature (thiolase + SCP2 FunFam pair, CS779) inherits a
  location that UniProt records only for a different, domain-deficient isoform
  (P32020-2). Because the failure is a property of the training procedure rather
  than of one hand-written branch, it warrants a targeted engineering fix plus
  branch pruning rather than either blanket acceptance or blanket removal. Its
  prevalence is unquantified: 139 of this rule's condition sets share a signature
  set with the Cytoplasm rule ARBA00004496 and 19 with the Peroxisome rule
  ARBA00004275, but genuine dual localization produces the same signal, so those
  counts bound the triage list from above rather than counting defects, and the
  mechanism is confirmed only for mouse P32020. A
  secondary, independent problem is the very large tail of low-information
  branches (693 single-FunFam sets, 240 unlabeled) and 140 singleton positive
  taxon clades that look like training-set residue. If UniProt is unwilling to
  make ARBA isoform-aware, the fallback should be SPLIT into per-family rules so
  that each branch can be accepted, audited or retired on its own evidence.
suggested_modifications:
- >-
  Retire or repair CS779 (FunFam 3.30.1050.10:FF:000001 AND FunFam
  3.40.47.10:FF:000016 AND NOT Bacteria/Archaea/Viruses) and its sibling CS2894 in
  the Cytoplasm rule ARBA00004496 (the identical FunFam pair). The mitochondrion
  and cytoplasm calls for the SCPx thiolase+SCP2 architecture derive from a
  UniProt SUBCELLULAR LOCATION comment scoped to isoform P32020-2, which lacks the
  thiolase domain that these branches require. For this architecture only the
  peroxisome call (ARBA00004275 CS80) should be retained. This directly resolves
  the false GO:0005739 IEA on Drosophila ScpX (Q9VJ43) reported in
  geneontology/go-annotation#6412.
- >-
  Make ARBA isoform-aware at the training stage: exclude, or explicitly scope,
  SUBCELLULAR LOCATION comments that carry a `molecule:` (isoform) qualifier
  whenever the learned signature is architecture-specific and the named isoform
  differs in domain content from the displayed sequence. At minimum, do not let a
  location asserted for an alternative-initiation or alternative-splice product be
  transferred by a signature that requires a domain absent from that product.
- >-
  Audit the 139 condition sets whose signature set also appears in ARBA00004496
  (Cytoplasm) and the 19 that also appear in ARBA00004275 (Peroxisome) - 155
  distinct sets, since 3 appear in both. They are an upper bound on candidates for
  collapsed isoform- or
  condition-specific localization, not a defect count: many families genuinely
  occupy more than one of these compartments, so each set needs individual
  inspection - ideally against a baseline of how often reviewed proteins matching
  the same signature legitimately carry both SUBCELLULAR LOCATION values. Sets
  where the shared signature requires a domain absent from the isoform bearing the
  location comment (the CS779 pattern) are the ones to triage first.
- >-
  Drop, or gate behind benchmarking, condition sets consisting of a single FunFam
  plus only NOT(Bacteria/Archaea/Viruses) where the FunFam has no
  mitochondria-specific label (693 single-FunFam sets, of which 240 have no
  resolved label at all; e.g. CS608 = 2.130.10.10:FF:000404, CS812 =
  3.40.47.10:FF:000020). Require each retained branch to have at least one
  reviewed Swiss-Prot positive exemplar with experimental mitochondrial evidence.
- >-
  Replace the 140 singleton positive taxon clades (187 sets; e.g. CS1124 =
  mitochondrial acetyl-CoA acetyltransferase FunFam restricted to Anura, CS9 =
  COX2 restricted to Euarchontoglires, CS13 = COX2 restricted to Laurasiatheria)
  with evidence-based taxonomic scope derived from the phylogenetic distribution
  of experimentally localized orthologues. Merge branches that differ only in
  clade.
- >-
  Collapse redundant nested InterPro conjunctions where a domain, its homologous
  superfamily and the corresponding family all describe the same region and the
  family alone is diagnostic (e.g. CS23: IPR004217 + IPR035427 + IPR050673).
  Retain nesting only where the broad signature is promiscuous and the narrow one
  supplies genuine discrimination (e.g. CS74 Tom40 within the porin superfamily).
- >-
  Where branch-level evidence supports it, propagate a submitochondrial child term
  in addition to GO:0005739 (mitochondrial inner membrane for sideroflexins/CS38
  and the OXPHOS branches, mitochondrial outer membrane for TOM40/CS74,
  mitochondrial intermembrane space for the Tim chaperones/CS23) rather than
  leaving every prediction at the organelle level.
- >-
  Strategic option (deeper fix): SPLIT ARBA00004173 into per-family rules so that
  each branch carries its own provenance, taxonomic scope, GO granularity and
  benchmark, instead of 1,490 classifiers of wildly differing quality sharing one
  rule identifier and one blanket ECO:0007322 evidence code. This also makes
  individual branches citable and retractable in response to GO curator reports.
parsimony:
  assessment: OVERLY_COMPLEX
  notes: >-
    1,490 OR-ed condition sets, i.e. ~124x the 12-condition-set threshold above
    which the project's own tooling refuses to analyse a rule. The complexity is
    not merely large but poorly structured: 693 sets (46.5%) are a single FunFam
    plus taxon, 240 of those FunFams have no resolved label at all; 180 sets are
    two FunFams, 131 are InterPro+InterPro+PANTHER, 102 are three FunFams, 81 a
    single InterPro, 19 a single PANTHER. Condition-type usage is taxon 1,490,
    FunFam 1,359, InterPro 885, PANTHER 287. All 1,490 signature sets are distinct
    (no exact duplicates), so the complexity is not literal duplication - it is
    the absence of any shared mechanistic basis: the sets have nothing in common
    except emitting the same location string. Nested InterPro triples
    (domain + homologous superfamily + family for the same region, e.g. CS23) add
    conditions without adding discrimination. A rule of this shape cannot be
    reviewed, benchmarked or maintained as a unit, which is why the deeper
    recommendation is to split it into per-family rules.
  supported_by:
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-analysis-notes.md
    supporting_text: Rule ARBA00004173 has 1490 condition sets, which exceeds the
      maximum of 12. Analysis is skipped for rules with too many condition sets as
      they would require excessive UniProt API queries and take too long.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: Treat each of the 1,490 condition sets as an independently
      testable classifier.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: The rule should be retained only as a container for
      branch-level classifiers, with systematic removal or refinement of
      generic-domain, fragment-prone, paralogue-ambiguous, and organelle-confounded
      branches.
literature_support:
  assessment: MODERATE
  notes: >-
    Literature support is bimodal, and a single enum cannot express that; MODERATE
    is chosen as the rule-level summary. For a substantial minority of branches the
    support is effectively STRONG: MIC60, TOM40, the small Tim chaperones,
    mitochondrially encoded cytochrome b and cytochrome c oxidase subunits, and
    sideroflexins are textbook mitochondrial proteins with abundant experimental
    localization evidence, and the falcon report singles out the sideroflexin
    branch (CS38/IPR004686) as well supported. For the branch the GO curator
    complained about, support is CONTRADICTED: Li, Fan & Papadopoulos
    (PMID:26901662) tested the putative N-terminal mitochondrial presequence of
    SCPX/SCP2 directly by live-cell confocal imaging of fluorescent fusions and
    concluded it is not sufficient to localize either protein to mitochondria,
    while C-terminal PTS1-mediated peroxisomal targeting is. The only primary
    evidence for mitochondrial SCP2 is Starodub et al. (PMID:11003606), an
    immunofluorescence study of the SHORT SCP-2 protein in L-cell fibroblasts that
    ranks mitochondria third behind peroxisomes and ER - and UniProt correctly
    scopes that evidence to isoform P32020-2, not to the SCPx architecture that
    CS779 detects. For the very large tail of unlabeled single-FunFam branches
    there is no identifiable literature at all, because the branches cannot even be
    resolved to a named family. The falcon report additionally warns that rules
    trained on noisy source annotations reinforce systematic errors, which is
    precisely what the isoform-collapse defect demonstrates.
  supported_by:
  - reference_id: PMID:26901662
    supporting_text: the N-terminal mitochondrial targeting sequence is not powerful
      enough to localize the SCPX and SCP2 proteins to the mitochondria.
  - reference_id: PMID:26901662
    supporting_text: The Scp2 gene contains two transcription initiation sites,
      giving rise to the 58 kDa sterol carrier protein-x (SCPX) and 15 kDa pro-SCP2
      proteins
  - reference_id: PMID:11003606
    supporting_text: detected SCP-2 in peroxisomes > endoplasmic reticulum >
      mitochondria >
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: Sideroflexins are mitochondrial inner-membrane proteins,
      although their transported substrates and physiological roles differ among
      paralogues. That branch is well supported for GO:0005739.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: Thus, family rules trained on noisy source annotations can
      reinforce systematic errors.
  - reference_id: https://github.com/geneontology/go-annotation/issues/6412
    supporting_text: From orthology to P32020, only the sort isoform P32020_2 is
      potentially mitochondrial.
condition_overlap:
  assessment: SIGNIFICANT
  notes: >-
    Quantitative pairwise overlap could not be computed because `just analyze-rule`
    refuses to run on rules with >12 condition sets, so `pairwise_overlap` and
    `entries` are empty. Qualitative and cross-rule evidence nonetheless makes
    SIGNIFICANT the right call, on three distinct axes. (1) Within-set redundancy:
    the recurrent InterPro triple shape (domain + its homologous superfamily + the
    corresponding family, 99 sets of shape InterPro x3, e.g. CS23 = IPR004217 +
    IPR035427 + IPR050673) conjoins nested annotations of the same sequence region,
    which cannot increase specificity. (2) Between-set overlap within the rule:
    693 single-FunFam sets drawn from broad CATH superfamilies will match
    overlapping protein sets, and multi-FunFam sets built from the same superfamily
    (3.40.47.10 appears in CS779, CS812, CS1124 among others) partially subsume one
    another. (3) Cross-rule overlap, which is the most consequential: 139/1,490
    (9.3%) of this rule's condition sets have a signature set that appears
    identically in the Cytoplasm rule ARBA00004496, 19/1,490 (1.3%) in the
    Peroxisome rule ARBA00004275, and 3 in both. These compartments are not
    mutually exclusive - mitochondrion/cytoplasm and mitochondrion/peroxisome
    co-annotation is routine and frequently correct (fumarase, aconitase, several
    aminoacyl-tRNA synthetases, beta-oxidation enzymes), and the falcon report
    makes the same point: "For dual-localized families, GO:0005739 should coexist
    with the other supported component rather than suppress it." The 139 + 19
    counts are therefore an upper bound on the sets that warrant per-set triage,
    not a count of collapsed antecedents; separating true dual localization from
    isoform collapse requires inspecting each set, which this review did not do.
    What is demonstrated is the single worked example CS779, where the shared
    signature requires the thiolase FunFam that isoform P32020-2 lacks: the same
    FunFam pair is CS2894 of ARBA00004496, and its thiolase half alone is CS80 of
    ARBA00004275. Separately, of the 885 distinct InterPro entries used by this
    rule, 493 already carry some InterPro2GO mapping but only 35 are mapped to
    GO:0005739 by InterPro curators. This is an argument from silence and should
    not be read as InterPro curators declining a mitochondrial claim: InterPro2GO
    maps a term only where it holds for all matches, CC coverage there is sparse by
    design, and this rule's condition sets are conjunctive, so in the 131
    InterPro+InterPro+PANTHER, 99 InterPro x3, 99 InterPro+PANTHER and 67
    InterPro x2 sets no single entry need imply mitochondrion for the set to. The
    comparison is like-for-like only in the 81 single-InterPro sets, and that
    restricted count was not computed.
  supported_by:
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-analysis-notes.md
    supporting_text: Rule ARBA00004173 has 1490 condition sets, which exceeds the
      maximum of 12. Analysis is skipped for rules with too many condition sets as
      they would require excessive UniProt API queries and take too long.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: Multiple signatures may either add specificity or simply be
      nested/redundant annotations of the same sequence region.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: Where a PANTHER subfamily is sufficient and more specific, a
      parent-family condition may be redundant.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: For dual-localized families, GO:0005739 should coexist with the
      other supported component rather than suppress it.
go_specificity:
  assessment: APPROPRIATE
  notes: >-
    GO:0005739 (mitochondrion) is the right granularity for a rule whose source
    assertion is the UniProt subcellular-location keyword Mitochondrion (SL-0173):
    for every true positive it is a correct, conservative cellular-component term,
    and it is the term GO_REF:0000044 legitimately produces from that keyword. The
    aspect and branch are right (cellular component, not molecular function or
    biological process), so it is not MISMATCHED, and it is certainly not
    TOO_NARROW. The honest qualifier is that it is frequently UNDERSPECIFIED rather
    than wrong: many branches identify families whose submitochondrial compartment
    is well established (inner membrane for sideroflexins and OXPHOS subunits,
    outer membrane for TOM40, intermembrane space for the Tim chaperones) and could
    support a child term in addition to the parent. I have deliberately not scored
    this TOO_BROAD, because the false positives this rule generates - Q9VJ43 being
    the reported case - are not failures of term granularity at all: no
    mitochondrial child or parent term would be correct there. Fixing them requires
    fixing the antecedent (isoform scoping, branch pruning), not the consequent.
    Scoring TOO_BROAD would misdirect the fix.
  supported_by:
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: GO:0005739 denotes the mitochondrion as a cellular component.
      It is appropriate if the protein resides in any mitochondrial compartment
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: For branches supported by submitochondrial evidence, more
      precise cellular-component terms should be propagated in addition to the
      parent
taxonomic_scope:
  assessment: TOO_BROAD
  notes: >-
    The dominant taxonomic condition is NOT(Bacteria/Archaea/Viruses), present as
    the ONLY taxon constraint in 1,303/1,490 sets (87%). This is a coarse
    "eukaryote-ish" filter and carries almost no organelle information: it does not
    distinguish mitochondria from cytosol, nucleus, ER, Golgi, peroxisome or - in
    plants and algae - plastid, and it is largely redundant with signatures that
    are already eukaryote-restricted. Combined with the 693 bare single-FunFam
    branches, this is the principal false-positive generator, and it is exactly how
    the Drosophila ScpX case arises: CS779's taxonomic condition does nothing to
    prevent a peroxisomal/cytosolic lipid-transfer protein from being called
    mitochondrial.

    A separate, opposite artefact coexists with this over-breadth and is worth
    recording even though the enum can only carry one value: the 187 sets that do
    have a positive taxon condition use 140 distinct clades, each occurring only
    once or twice (Anura, Pongo, Bos, Homo, Mus, Kluyveromyces, Oryza, Camelineae,
    "Aspergillus subgen. Circumdati", ...). These are TOO_NARROW in effect and
    read as training-set residue rather than biology - CS1124 confines the
    FunFam labelled "acetyl-CoA acetyltransferase, mitochondrial" (ACAT1), a
    pan-metazoan mitochondrial matrix thiolase, to frogs; CS9 and CS13 split
    cytochrome c oxidase subunit II between Euarchontoglires and Laurasiatheria.
    They sacrifice recall without buying precision. The rule therefore manages to
    be simultaneously too broad where it matters (the 87% with no positive
    constraint) and arbitrarily too narrow where it does constrain.
  supported_by:
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: It cannot distinguish mitochondrial from plastid targeting in
      plants and algae, or mitochondrial from peroxisomal, ER, and Golgi
      localization in other eukaryotes.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: They can also be overfitted if derived from sparse reviewed
      examples. Taxonomic exclusions should therefore be supported by phylogenetic
      inspection of experimentally localized orthologues, not merely by absence of
      training examples.
  - reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
    supporting_text: a domain shared by proteins in several compartments is not
      intrinsically a localization signal
confidence: 0.35
references:
- id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
  title: Deep research analysis via Falcon - UniProt rule ARBA00004173
  findings:
  - statement: The rule is an aggregated umbrella of 1,490 independent branch-level
      classifiers and should be audited branch-by-branch rather than accepted or
      rejected wholesale.
    supporting_text: Treat each of the 1,490 condition sets as an independently
      testable classifier.
  - statement: The report's final verdict is that the rule should survive only as a
      container for branch-level classifiers, with generic-domain and
      organelle-confounded branches removed.
    supporting_text: The rule should be retained only as a container for
      branch-level classifiers, with systematic removal or refinement of
      generic-domain, fragment-prone, paralogue-ambiguous, and organelle-confounded
      branches.
  - statement: Isoform choice is explicitly listed as a mechanism by which a
      structural-domain signature can fail to predict localization - the exact
      failure mode reported in go-annotation issue #6412 for CS779.
    supporting_text: alternative initiation, splicing, and competing targeting
      sequences can alter localization.
  - statement: Localization is a property of the complete protein, not of one
      structural domain, so a CATH FunFam match is a weak localization proxy.
    supporting_text: localization is a property of the complete protein—including
      terminal targeting information, transmembrane segments, isoform choice, and
      cellular context—not necessarily of one structural domain
  - statement: Single-signature branches are acceptable only when the signature is a
      mitochondria-specific whole-protein family, and unsafe when it is a generic
      domain - the basis for pruning the 693 bare single-FunFam sets.
    supporting_text: Conditions such as CS7, CS15, CS31, CS34, CS38, and CS39 rely
      on one InterPro entry plus a broad taxonomic exclusion. These are acceptable
      only when the InterPro entry is a mitochondria-specific whole-protein family;
      they are unsafe if it is a generic domain.
  - statement: The sideroflexin branch (CS38 / IPR004686) is a well-supported
      positive example for GO:0005739.
    supporting_text: Sideroflexins are mitochondrial inner-membrane proteins,
      although their transported substrates and physiological roles differ among
      paralogues. That branch is well supported for GO:0005739.
  - statement: GO:0005739 is an appropriate cellular-component term when
      mitochondrial residence holds, but is often underspecified relative to
      submitochondrial child terms.
    supporting_text: For branches supported by submitochondrial evidence, more
      precise cellular-component terms should be propagated in addition to the
      parent
  - statement: The NOT(Bacteria/Archaea/Viruses) filter provides essentially no
      organelle specificity within eukaryotes.
    supporting_text: It cannot distinguish mitochondrial from plastid targeting in
      plants and algae, or mitochondrial from peroxisomal, ER, and Golgi
      localization in other eukaryotes.
  - statement: Positive lineage restrictions can be overfitted to sparse reviewed
      training examples - consistent with the 140 singleton clades observed here.
    supporting_text: They can also be overfitted if derived from sparse reviewed
      examples. Taxonomic exclusions should therefore be supported by phylogenetic
      inspection of experimentally localized orthologues, not merely by absence of
      training examples.
  - statement: Rules trained on noisy source annotations reinforce systematic
      errors - the general form of the isoform-collapse defect documented here.
    supporting_text: Thus, family rules trained on noisy source annotations can
      reinforce systematic errors.
  - statement: Branches without reviewed positive exemplars should be flagged as
      speculative rather than merged invisibly into a high-volume rule.
    supporting_text: Branches with no reviewed positive exemplars should be flagged
      as speculative rather than merged invisibly into the same high-volume rule.
  - statement: Isoform-specific localization should be documented rather than
      collapsed, where the annotation model permits it.
    supporting_text: Conditional or isoform-specific localization should be
      documented where the annotation model permits it.
  reference_review:
    relevance: HIGH
    correctness: VERIFIED
    review_notes: >-
      Falcon deep-research report generated for this rule. Its rule-level
      statistics (1,490 condition sets; 536,854 unreviewed proteins; condition-type
      usage taxon 1,490 / FunFam 1,359 / InterPro 885 / PANTHER 287) were
      independently re-derived from ARBA00004173.enriched.json and match exactly.
      Its identification of IPR004686 as the sideroflexin family and its CS38
      numbering were verified against the rule JSON. The report is an
      LLM-generated synthesis, so its literature citations were treated as leads
      rather than as verified facts, and no claim in this review rests on a
      citation appearing only in that report. The companion perplexity run failed
      (HTTP 401, quota exhausted), so this is the only deep-research provider
      available.
- id: PMID:26901662
  title: Sterol Carrier Protein-2, a Nonspecific Lipid-Transfer Protein, in
    Intracellular Cholesterol Trafficking in Testicular Leydig Cells.
  findings:
  - statement: Live-cell confocal imaging of fluorescent SCPX and SCP2 fusions
      showed that the C-terminal PTS1 signal drives peroxisomal targeting, while
      the putative N-terminal mitochondrial presequence is not sufficient to
      localize either protein to mitochondria. This is direct evidence against the
      mitochondrial assertion that ARBA00004173 CS779 makes for the SCPx
      thiolase+SCP2 architecture.
    supporting_text: the N-terminal mitochondrial targeting sequence is not powerful
      enough to localize the SCPX and SCP2 proteins to the mitochondria.
  - statement: The Scp2 gene uses two transcription initiation sites to produce the
      58 kDa SCPX (thiolase + SCP2 domain) and the 15 kDa pro-SCP2 (SCP2 domain
      only); the two products differ in domain content, which is why an
      isoform-scoped location cannot be transferred by a thiolase-requiring
      signature.
    supporting_text: The Scp2 gene contains two transcription initiation sites,
      giving rise to the 58 kDa sterol carrier protein-x (SCPX) and 15 kDa pro-SCP2
      proteins
  reference_review:
    relevance: HIGH
    correctness: VERIFIED
    review_notes: >-
      PMID verified against the cached publication record (Li NC, Fan J,
      Papadopoulos V, PLoS One 2016; full text available). This is the ECO:0000269
      source that UniProt cites for BOTH the "Isoform SCPx" peroxisome-only
      location and part of the "Isoform SCP2" multi-compartment location on P32020,
      confirmed by fetching the P32020 record from the UniProt REST API. Its
      experimental conclusion directly contradicts mitochondrial targeting of the
      full-length SCPx architecture.
- id: PMID:11003606
  title: Sterol carrier protein-2 localization in endoplasmic reticulum and role in
    phospholipid formation.
  findings:
  - statement: Indirect immunofluorescence and confocal microscopy of L-cell
      fibroblasts overexpressing the short SCP-2 protein ranked its distribution
      peroxisomes > endoplasmic reticulum > mitochondria > lysosomes. This is the
      primary source of the mitochondrial claim, and it concerns the short
      SCP2 product (UniProt isoform P32020-2), not the full-length SCPx
      architecture that ARBA00004173 CS779 requires.
    supporting_text: detected SCP-2 in peroxisomes > endoplasmic reticulum >
      mitochondria >
    full_text_unavailable: true
  reference_review:
    relevance: HIGH
    correctness: VERIFIED
    review_notes: >-
      PMID verified against the cached publication record (Starodub et al., Am J
      Physiol Cell Physiol 2000); the cache is abstract-only
      (full_text_available: false), so only the abstract was assessed. Confirmed
      via the UniProt REST API that P32020 cites this PMID exclusively under the
      SUBCELLULAR LOCATION comment scoped to "Isoform SCP2" (P32020-2) - it is not
      cited for the "Isoform SCPx" comment, which lists Peroxisome only. Cited here
      to establish the isoform scoping of the mitochondrial evidence, not to
      dispute the curator's original annotation.
- id: https://github.com/geneontology/go-annotation/issues/6412
  title: 'go-annotation issue #6412: ARBA00004173 mitochondrion annotation on
    UniProt:Q9VJ43 (Drosophila melanogaster ScpX)'
  findings:
  - statement: A GO curator reported that ARBA00004173 assigns a mitochondrial
      location to Drosophila melanogaster ScpX (Q9VJ43), and correctly diagnosed
      that in the mouse orthologue P32020 only the short isoform is potentially
      mitochondrial, with no evidence that D. melanogaster has an equivalent short
      isoform.
    supporting_text: From orthology to P32020, only the sort isoform P32020_2 is
      potentially mitochondrial.
  reference_review:
    relevance: HIGH
    correctness: VERIFIED
    review_notes: >-
      The complaint was independently verified. QuickGO confirms Q9VJ43 carries
      GO:0005739 IEA via GO_REF:0000044 with with/from UniProtKB-SubCell:SL-0173,
      assigned by UniProt; the UniProt REST record for Q9VJ43 shows three ARBA
      location calls (Cytoplasm/ARBA00004496, Mitochondrion/ARBA00004173,
      Peroxisome/ARBA00004275). The curator's reading of P32020 is correct and, if
      anything, understated - UniProt scopes the mitochondrion location strictly to
      Isoform SCP2 (P32020-2) while Isoform SCPx (P32020-1) is annotated Peroxisome
      only.
- id: file:rules/arba/ARBA00004173/ARBA00004173-analysis-notes.md
  title: ARBA00004173 reproducible rule statistics and `analyze-rule` refusal output
  findings:
  - statement: The rule has 1,490 OR-ed condition sets and exceeds the project's
      analysis threshold of 12, so `just analyze-rule ARBA00004173` refuses to run
      and no pairwise overlap statistics, heatmap or entries view are available.
    supporting_text: Rule ARBA00004173 has 1490 condition sets, which exceeds the
      maximum of 12. Analysis is skipped for rules with too many condition sets as
      they would require excessive UniProt API queries and take too long.
  - statement: All structural counts quoted in this review (condition-set shapes,
      taxon-condition counts, InterPro2GO redundancy, and the signature-set overlap
      with the sibling location rules ARBA00004496 and ARBA00004275) are regenerated
      by `uv run python rules/arba/ARBA00004173/ARBA00004173-stats.py`.
- id: file:rules/arba/ARBA00004173/ARBA00004173.json
  title: ARBA rule ARBA00004173 raw rule definition (1,490 condition sets)
  findings:
  - statement: The rule's single consequent is the UniProt SUBCELLULAR LOCATION
      comment with value "Mitochondrion" (SL-0173), which GO_REF:0000044 maps to
      GO:0005739 with evidence IEA / ECO:0007322. Statistics are 0 reviewed and
      536,854 unreviewed proteins; created 2020-05-12, modified 2025-12-14.
  - statement: Condition-type usage across the 1,490 sets is taxon 1,490, FunFam
      1,359, InterPro 885, PANTHER 287. Set shapes are 693 single FunFam (46.5%),
      180 two FunFams, 131 InterPro+InterPro+PANTHER, 102 three FunFams, 99
      InterPro+PANTHER, 99 three InterPro, 81 single InterPro, 67 two InterPro, 19
      single PANTHER. All 1,490 signature sets are distinct.
  - statement: 1,303/1,490 sets (87%) have no positive taxon restriction, only
      NOT(Bacteria/Archaea/Viruses). The remaining 187 use 140 distinct positive
      clades, each appearing once or twice. Only 422/1,490 (28%) contain at least
      one signature whose resolved label mentions "mitochondri*".
  - statement: Of the 885 distinct InterPro entries used by the rule, 493 have some
      InterPro2GO mapping but only 35 are mapped to GO:0005739 by InterPro2GO, so
      ~96% of the rule's InterPro conditions assert a mitochondrial location that
      InterPro curators did not assert.
- id: file:rules/arba/_interpro2go.txt
  title: InterPro2GO mapping file (GO Consortium)
  findings:
  - statement: Only 35 of the 885 InterPro entries used as conditions in
      ARBA00004173 are mapped to GO:0005739 by InterPro2GO, despite 493 of them
      having at least one InterPro2GO mapping.
supported_by:
- reference_id: file:rules/arba/ARBA00004173/ARBA00004173-analysis-notes.md
  supporting_text: Rule ARBA00004173 has 1490 condition sets, which exceeds the
    maximum of 12. Analysis is skipped for rules with too many condition sets as
    they would require excessive UniProt API queries and take too long.
- reference_id: https://github.com/geneontology/go-annotation/issues/6412
  supporting_text: From orthology to P32020, only the sort isoform P32020_2 is
    potentially mitochondrial.
- reference_id: PMID:26901662
  supporting_text: the N-terminal mitochondrial targeting sequence is not powerful
    enough to localize the SCPX and SCP2 proteins to the mitochondria.
- reference_id: PMID:11003606
  supporting_text: detected SCP-2 in peroxisomes > endoplasmic reticulum >
    mitochondria >
- reference_id: file:rules/arba/ARBA00004173/ARBA00004173-deep-research-falcon.md
  supporting_text: The rule should be retained only as a container for branch-level
    classifiers, with systematic removal or refinement of generic-domain,
    fragment-prone, paralogue-ambiguous, and organelle-confounded branches.