Single-Organism IBA Phylogenetic Distance Analysis

Single-Organism IBA Phylogenetic Distance Analysis

Overview

A recurrent question in IBA review is how to evaluate IBA annotations whose WITH/FROM column cites experimental descendants from only a single organism (excluding self-citations where the target gene's species is also a donor).

As established in IBA_REVIEW.md, a single-organism donor list is not intrinsically weak when the donor and target are close orthologs and the PAINT curator placed the ancestral IBD node at an appropriate clade (for example, mouse $\to$ human within Euarchontoglires or Euteleostomi). However, when a single organism's experimental annotation is used to seed an ancestral IBD deep in the tree—such as at Opisthokonta, Eukaryota, or LUCA—and then propagated down across kingdoms, domains, or multiple gene-duplication boundaries, the risk of over-propagation rises substantially.

This sub-project defines five quantitative distance and donor-topology metrics to measure "how far down (and up) the tree" a single-organism or single-clade IBA travels, benchmarks them against existing review actions across all 28,945 positive GOA IBA annotations in the repository, and uses the resulting high-distance filter to identify candidate genes for targeted re-review.

Corpus-Wide Prevalence

Across genes/*/*/*-goa.tsv and interpro/panther/*/*-paint.tsv (reproducible via uv run python projects/IBA_REVIEW/single_organism_distance.py --summary):

Action Classification: Propagation Failure vs Term Granularity Refinement

To avoid conflating true homology over-propagation with routine ontology-granularity adjustments, all tables below separate non-retained review actions into two distinct categories:

Five Quantitative Distance & Donor-Topology Metrics

Implemented in single_organism_distance.py:

1. Taxonomic LCA Clade Depth ($d_{\text{LCA}}$) & Branch Steps ($s_{\text{up}} + s_{\text{down}}$)

For each donor–target pair $(O_{\text{donor}}, O_{\text{target}})$, we find their Lowest Common Ancestor (LCA) along the UniProt/NCBI species OC phylogeny:
- $d_{\text{LCA}}$: Clade depth of $\text{LCA}(O_{\text{donor}}, O_{\text{target}})$ from cellular_organisms ($d_{\text{LCA}} = 0$ at cellular_organisms/LUCA, $1$ at Eukaryota/Bacteria, $2$ at Metazoa/Fungi/Viridiplantae, $3\text{–}4$ at Bilateria/Protostomia/Deuterostomia/Ascomycota, $6+$ at Euteleostomi/Tetrapoda/Euarchontoglires).
- $s_{\text{up}} + s_{\text{down}}$: Total taxonomic branch steps up from $O_{\text{donor}}$ to $\text{LCA}$ plus down from $\text{LCA}$ to $O_{\text{target}}$.

Across the 3,883 rated non-self single-organism IBA rows with resolved donor lineages (plus 7 rows in Tier 0 with unresolved donor lineage), Propagation Failure (% Fail) and % REMOVE climb substantially as the donor–target LCA moves from close clades to cross-kingdom and cross-domain (LUCA) splits:

Phylogenetic Tier Representative LCA Clades Rated Keep Fail % Fail Refn % Refn % Down % REMOVE
Tier 1: Close Clade Euarchontoglires, Eutheria, Murinae, Caenorhabditis 1,519 1,302 121 8.0% 96 6.3% 14.3% 2.5%
Tier 2: Intermediate / Unenumerated Clade Saccharomycotina, Pezizomycotina, Magnoliopsida, plus unenumerated residual clades (Mammalia, Fungi, Bilateria, Xenopus, etc.) 101 76 7 6.9% 18 17.8% 24.8% 2.0%
Tier 3: Phylum / Subphylum Split Euteleostomi (Teleost $\leftrightarrow$ Mammal), Ascomycota, Ecdysozoa 657 512 90 13.7% 55 8.4% 22.1% 5.6%
Tier 4: Deep Metazoa / Bilateria Split Metazoa / Bilateria (Protostomia [Worm/Fly] $\leftrightarrow$ Deuterostomia [Vertebrate]) 542 426 84 15.5% 32 5.9% 21.4% 4.4%
Tier 5: Cross-Kingdom Opisthokonta / Amorphea / Bacteria Opisthokonta (Fungi $\leftrightarrow$ Metazoa), Amorphea, cross-phylum Bacteria 808 617 122 15.1% 69 8.5% 23.6% 5.6%
Tier 6: Cross-Supergroup (Eukaryota root) Eukaryota (Viridiplantae/Algae $\leftrightarrow$ Opisthokonta/Protist) 95 64 27 28.4% 4 4.2% 32.6% 12.6%
Tier 7: Cross-Domain (LUCA) cellular_organisms (Bacteria/Archaea $\leftrightarrow$ Eukaryota) 161 105 40 24.8% 16 9.9% 34.8% 16.1%
Excluded: Tier 0 (Unknown donor lineage) Unresolved external donor taxon 7 6 1 14.3% 0 0.0% 14.3% 14.3%

Key observations from Table 1:
- Clade-to-Tier mapping (classify_evolutionary_span): single_organism_distance.py maps each donor–target LCA clade to Tiers 0–7: Unknown $\to$ Tier 0; Cellular organisms (LUCA) $\to$ Tier 7; Eukaryota $\to$ Tier 6 when either species is in Viridiplantae/Algae (Arabidopsis, Oryza, Triticum, Chlamydomonas), else Tier 5 (Amorphea or Opisthokonta); Bacteria/Archaea $\to$ Tier 5 vs intra-phylum bacterial clades (Bacillati, Gammaproteobacteria, Enterobacteriaceae, Actinomycetes) $\to$ Tier 3; Metazoa $\to$ Tier 4; phylum/subphylum clades (Euteleostomi, Dikarya, Ascomycota, Viridiplantae, Embryophyta, Ecdysozoa, Arthropoda) $\to$ Tier 3; order/subphylum fungal and plant clades (Saccharomycotina, Pezizomycotina, Magnoliopsida, Pentapetalae, rosids) $\to$ Tier 2; and close mammalian/model genera and superorders (Eutheria, Boreoeutheria, Euarchontoglires, Primates, Hominidae, Glires, Rodentia, Murinae, Caenorhabditis, Aspergillus, etc.) $\to$ Tier 1, with an explicit fallback (Tier 2: Other ({lca})) routing any unenumerated LCA clade to Tier 2 (see §Limitations item 4).
- Separating Propagation Failure (% Fail) from Term Refinement (% Refn) confirms that the distance effect is driven by genuine over-propagation (REMOVE + MARK_AS_OVER_ANNOTATED), which more than triples from 8.0% in Tier 1 to 28.4% in Tier 6 and 24.8% in Tier 7, while outright % REMOVE increases more than 6-fold (2.5% $\to$ 12.6% $\to$ 16.1%).
- By contrast, Term Refinement (% Refn) shows no systematic trend with distance (4.2%–9.9% across Tiers 1 and 3–7), and separating it explains the high total downgrade rate in Tier 2 ($n = 101$): 18 of the 25 downgraded rows (17.8%) are MODIFY/TOO_GENERAL granularity refinements, whereas only 7 rows (6.9%, 2.0% REMOVE) are propagation failures.
- Note that sample sizes in Tier 2 ($n = 101$), Tier 6 ($n = 95$), and Tier 7 ($n = 161$) are much smaller than in Tiers 1, 3, 4, and 5 ($n = 542\text{–}1{,}519$), so point estimates in those smaller tiers carry wider sampling uncertainty.

1B. Total Taxonomic Branch Steps ($s_{\text{up}} + s_{\text{down}}$)

Total Taxonomic Branch Steps ($s_{\text{up}} + s_{\text{down}}$) Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
02–06 steps (Close: Genus/Family) 160 22 13.8% 13 8.1% 21.9% 11 6.9%
07–12 steps (Order/Superorder, e.g. Rodent–Primate) 796 66 8.3% 59 7.4% 15.7% 33 4.1%
13–18 steps (Phylum/Subphylum, e.g. Fish–Mammal) 1,340 114 8.5% 97 7.2% 15.7% 38 2.8%
19–24 steps (Kingdom/Supergroup, e.g. Yeast–Human) 991 180 18.2% 79 8.0% 26.1% 76 7.7%
25+ steps (Deep Bilateria / Cross-Kingdom / Cross-Domain) 596 109 18.3% 42 7.0% 25.3% 26 4.4%
Excluded (unknown donor lineage, total_tax_steps < 0) 7 1 14.3% 0 0.0% 14.3% 1 14.3%

Note on non-monotonicity and UniProt OC lineage density: Raw taxonomic step counts are non-monotonic at the extremes. The 02–06 steps bin ($n = 160$, 13.8% Fail, 6.9% REMOVE) is enriched for intragenus/intraphylum bacterial and fungal paralog transfers, and the 25+ steps bin (4.4% REMOVE, 18.3% Fail) has a lower REMOVE rate than 19–24 steps (7.7% REMOVE, 18.2% Fail) because metazoan UniProt OC lineages contain many more intermediate taxonomic ranks (~14–16 ranks in vertebrates/insects) than bacterial or fungal lineages (~6–9 ranks). Consequently, a cross-domain Bacteria $\to$ Bacteria or Yeast $\to$ Human pair often falls into 19–24 steps, whereas a Worm $\to$ Human (Bilateria) pair falls into 25+ steps. Named phylogenetic tiers (Table 1) are therefore a better-calibrated metric than raw OC rank counts.

2. Excess PAINT Node Depth ($\Delta_{\text{depth}} = d_{\text{LCA}} - d_{\text{PTN}}$)

Even when the single experimental donor and target are close relatives (e.g., mouse $\to$ human), a PAINT curator may have placed the ancestral IBD assertion at LUCA or Eukaryota rather than at Mammalia or Vertebrata.

Because UniProt OC rank depth varies across kingdoms, the lookup table PAINT_TAXON_DEPTH in single_organism_distance.py maps all 63 explicit positive-IBD taxon:<id> identifiers across interpro/panther/*/*-paint.tsv to an approximate OC clade depth ($d_{\text{PTN}} = 0$ for taxon:131567 cellular_organisms; $1$ for Eukaryota/Opisthokonta/Bacteria/Archaea; $2\text{–}5$ for intermediate Metazoa/Bilateria/Chordata/kingdom nodes; $6+$ for Euteleostomi/Tetrapoda/Amniota/Mammalia/Euarchontoglires and close genus/species nodes such as Caenorhabditis taxon:6237). Rows in *-paint.tsv whose taxon column is a bare taxon: string (meaning no taxon constraint was recorded in the PAINT GAF; see src/ai_gene_review/etl/panther_paint.py) are reported separately rather than assigned a clade depth. Restricting to all 1,751 rated rows on close donor–target pairs ($d_{\text{LCA}} \ge 6$: 917 in locally cached PAINT families with an explicit node taxid, 246 in cached families with an unrecorded blank taxon:, and 588 in uncached families) holds donor–target evolutionary distance constant while isolating the effect of PAINT ancestral node placement ($d_{\text{PTN}}$):

PAINT Ancestral Node Placement ($d_{\text{PTN}}$) on Close Taxonomic Pairs ($d_{\text{LCA}} \ge 6$, total $n = 1{,}751$) Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
Scoped within clade / Vertebrata ($d_{\text{PTN}} \ge 6$, e.g. Euteleostomi, Tetrapoda, Mammalia, Caenorhabditis) 486 31 6.4% 36 7.4% 13.8% 2 0.4%
Intermediate Metazoa / Bilateria / Chordata ($2 \le d_{\text{PTN}} \le 5$) 298 23 7.7% 24 8.1% 15.8% 10 3.4%
Over-deeply placed at LUCA / Eukaryota / Opisthokonta ($d_{\text{PTN}} \le 1$, $\Delta_{\text{depth}} \ge 5$) 133 10 7.5% 8 6.0% 13.5% 5 3.8%
Unrecorded / unmapped PAINT taxon (blank taxon: in *-paint.tsv) 246 20 8.1% 11 4.5% 12.6% 12 4.9%
Uncached PAINT family (no local interpro/panther/*/*-paint.tsv record) 588 60 10.2% 42 7.1% 17.3% 18 3.1%

Within cached PAINT families with an explicit node taxon on close donor–target pairs ($n = 917$), placing a single-organism IBD at LUCA/Eukaryota/Opisthokonta ($d_{\text{PTN}} \le 1$, $n = 133$) or at intermediate pre-vertebrate nodes ($2 \le d_{\text{PTN}} \le 5$, $n = 298$) raises the outright REMOVE rate 8- to 9-fold (3.4%–3.8% vs 0.4% when scoped within the clade, $d_{\text{PTN}} \ge 6$), and increases propagation failure from 6.4% to 7.5%–7.7%. (An additional 246 rows in cached families have a blank taxon: column in *-paint.tsv—where the PAINT GAF did not record a node taxon constraint, spanning both shallow mouse-only nodes and deeper nodes—and show an 8.1% Fail / 4.9% REMOVE rate.)

3. PANTHER Subfamily Dilution (SAME_SF vs CROSS_SF)

Joining the target protein and single-organism donor protein against cached interpro/panther/*/*-entries.csv tests whether they reside in the same PANTHER subfamily (SAME_SF, typically direct 1:1 orthologs) or different subfamilies (CROSS_SF, separated by at least one gene-duplication/subfamily-divergence event). Across the 3,883 rated rows with known donor lineage (1,379 with resolved subfamily for both donor and target; 2,504 in UNKNOWN_SF where the PANTHER family or external donor seed is not in the local *-entries.csv cache):

Taxonomic Regime Subfamily Relationship Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
Close / moderate taxa ($d_{\text{LCA}} \ge 4$) SAME_SF (same PANTHER subfamily) 509 43 8.4% 37 7.3% 15.7% 12 2.4%
Close / moderate taxa ($d_{\text{LCA}} \ge 4$) CROSS_SF (cross-subfamily paralog) 242 26 10.7% 31 12.8% 23.6% 11 4.5%
Distant taxa ($0 \le d_{\text{LCA}} \le 3$) SAME_SF (same PANTHER subfamily) 313 40 12.8% 31 9.9% 22.7% 17 5.4%
Distant taxa ($0 \le d_{\text{LCA}} \le 3$) CROSS_SF (cross-subfamily paralog) 315 77 24.4% 18 5.7% 30.2% 29 9.2%

Crossing a PANTHER subfamily boundary (CROSS_SF) increases the Propagation Failure (% Fail) rate from 8.4% $\to$ 10.7% in close/moderate taxa and nearly doubles it from 12.8% $\to$ 24.4% in distant taxa, while nearly doubling the % REMOVE rate (2.4% $\to$ 4.5% and 5.4% $\to$ 9.2%).

4. Exact PANTHER Tree Edge Steps (steps_from_ibd) & Duplications (duplication_on_path)

The TreeGrafter path audit (projects/TREEGRAFTER/graft_paths_all.tsv) records exact PANTHER tree edge steps (steps_from_ibd) and whether a gene-duplication node lies between the IBD assertion node and the target (duplication_on_path). Because graft_paths_all.tsv audits TreeGrafter (IEA) graft paths on non-PANTHER reference-proteome species (such as 9AVES, 9POAL, AEDAE) rather than the 3,890 GOA IBA population on PANTHER reference species, its direct row-level overlap with GOA IBA annotations is only $n = 2$ rows (Table 4B). Evaluating all 386 rated single-taxon IBD paths in the TreeGrafter (IEA) audit (Table 4A) provides a complementary measurement of how tree-step length and intervening duplications affect single-donor IBD propagation failure:

Table 4A: All Single-Taxon IBD Paths in projects/TREEGRAFTER/graft_paths_all.tsv ($n = 386$ TreeGrafter IEA paths)

Tree Path Regime (steps_from_ibd, duplication_on_path) Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
< 10 tree steps, duplication_on_path = False 210 28 13.3% 11 5.2% 18.6% 11 5.2%
< 10 tree steps, duplication_on_path = True 164 28 17.1% 12 7.3% 24.4% 13 7.9%
>= 10 tree steps, duplication_on_path = False 2 0 0.0% 0 0.0% 0.0% 0 0.0%
>= 10 tree steps, duplication_on_path = True 10 2 20.0% 1 10.0% 30.0% 1 10.0%

Table 4B: Matched Non-Self GOA IBA Subset Overlapping graft_paths_all.tsv ($n = 2$)

Tree Path Regime (steps_from_ibd, duplication_on_path) Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
< 10 tree steps, duplication_on_path = False 2 0 0.0% 0 0.0% 0.0% 0 0.0%
< 10 tree steps, duplication_on_path = True 0 0 0.0% 0 0.0% 0.0% 0 0.0%
>= 10 tree steps, duplication_on_path = False 0 0 0.0% 0 0.0% 0.0% 0 0.0%
>= 10 tree steps, duplication_on_path = True 0 0 0.0% 0 0.0% 0.0% 0 0.0%

In Table 4A ($n = 386$ TreeGrafter IEA paths), crossing at least one duplication node (duplication_on_path = True) on paths of < 10 steps increases the Propagation Failure (% Fail) rate from 13.3% $\to$ 17.1% and the % REMOVE rate from 5.2% $\to$ 7.9% ($n = 210$ vs $n = 164$), while paths of $\ge 10$ tree steps are rare in graft_paths_all.tsv ($n = 12$ total).

5. Donor-Support Topology Regimes: Single-Gene vs Multi-Paralog vs Single-Clade Jump vs Bracketed Multi-Species ($d_{\text{MRCA}}$ vs $d_{\text{LCA}}$)

A natural follow-up question is whether the elevated propagation failure rate of single-organism IBAs is caused by having only one experimental seed gene, or by lack of phylogenetic bracketing around the target organism. To test this across all 9,410 resolved non-self GOA IBA rows (9,246 rated with a completed review action: 9,090 with resolved donor lineages and 156 in Tier 0), single_organism_distance.py --regimes computes both the donor-set MRCA clade depth ($d_{\text{MRCA}}$) across all multi-species donor taxa in WITH/FROM and the closest donor–target LCA depth ($d_{\text{LCA}}$), partitioning every non-self IBA into four topological regimes:

  1. Regime 1 (R1_SINGLE_GENE: 1 taxon, 1 distinct seed): Strictly one experimental donor gene in a single non-self species ($n = 3{,}139$ rated with resolved donor lineage; deduplicated on resolved accession via canonical_seed_gene, with all_non_self = 0 and rated_resolved_lineage = 0 rows reclassified from R2 to R1 as reported in Section 5E, mrca_clade = "single_taxon", and mrca_depth = -1).
  2. Regime 2 (R2_MULTI_PARALOG_SINGLE_TAXON: 1 taxon, $\ge 2$ distinct seeds): Multiple experimental paralog/co-annotated seed genes, all from the same single non-self species ($n = 744$ rated; note that $3{,}139 + 744 = 3{,}883$, matching the Table 1 resolved-lineage single-organism total).
  3. Regime 3 (R3_SINGLE_CLADE_JUMP: $\ge 2$ donor taxa, $d_{\text{MRCA}} > d_{\text{LCA}}$): Multiple donor species are cited in WITH/FROM, but all donor species coalesce within a single restricted clade ($d_{\text{MRCA}} > d_{\text{LCA}}$) before jumping out to the target organism—for example, mouse + rat (Rodentia / Glires) jumping to worm or yeast, worm + fly (Ecdysozoa) jumping to human, or M. tuberculosis + P. aeruginosa (Bacteria) jumping to human ($n = 1{,}262$ rated).
  4. Regime 4 (R4_BRACKETED_MULTI_SPECIES: $\ge 2$ donor taxa, $d_{\text{MRCA}} \le d_{\text{LCA}}$): Multiple donor species span across or phylogenetically bracket the target organism's LCA—for example, yeast + mouse bracketing human, or fly + zebrafish bracketing human ($n = 3{,}945$ rated).

Table 5A: Propagation Failure & Refinement by Donor-Support Topology Regime ($n = 9{,}090$ resolved-lineage rated rows; $n = 9{,}246$ incl. Tier 0)

Donor-Support Topology Regime Rated Keep Fail % Fail Refn % Refn % Down % REMOVE
R1: Single donor gene (1 taxon, 1 distinct seed) 3,139 2,503 398 12.7% 238 7.6% 20.3% 4.7%
R2: Multi-paralog single taxon (1 taxon, >= 2 distinct seeds) 744 599 93 12.5% 52 7.0% 19.5% 4.7%
R3: Multi-species single-clade jump (>= 2 taxa, $d_{\text{MRCA}} > d_{\text{LCA}}$) 1,262 1,033 141 11.2% 88 7.0% 18.1% 4.0%
R4: Bracketed multi-species (>= 2 taxa, $d_{\text{MRCA}} \le d_{\text{LCA}}$) 3,945 3,420 227 5.8% 298 7.6% 13.3% 2.1%
Excluded: Tier 0 / Unknown donor lineage (7 single-taxon + 149 multi-taxon) 156 131 10 6.4% 15 9.6% 16.0% 3.2%

Table 5B: Donor-Support Regimes Stratified by Phylogenetic Distance Band ($d_{\text{LCA}}$ Tiers)

Distance Band ($d_{\text{LCA}}$ Tier of Closest Donor) Donor-Support Regime Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
Close / Intermediate Clade (Tiers 1–2) R1: Single donor gene 1,266 106 8.4% 89 7.0% 15.4% 35 2.8%
Close / Intermediate Clade (Tiers 1–2) R2: Multi-paralog single taxon 354 22 6.2% 25 7.1% 13.3% 5 1.4%
Close / Intermediate Clade (Tiers 1–2) R3: Multi-species single-clade jump 323 13 4.0% 24 7.4% 11.5% 7 2.2%
Close / Intermediate Clade (Tiers 1–2) R4: Bracketed multi-species 1,338 71 5.3% 92 6.9% 12.2% 22 1.6%
Phylum / Bilateria Split (Tiers 3–4) R1: Single donor gene 938 124 13.2% 68 7.2% 20.5% 43 4.6%
Phylum / Bilateria Split (Tiers 3–4) R2: Multi-paralog single taxon 261 50 19.2% 19 7.3% 26.4% 18 6.9%
Phylum / Bilateria Split (Tiers 3–4) R3: Multi-species single-clade jump 458 55 12.0% 28 6.1% 18.1% 19 4.1%
Phylum / Bilateria Split (Tiers 3–4) R4: Bracketed multi-species 1,978 106 5.4% 169 8.5% 13.9% 41 2.1%
Cross-Kingdom / Supergroup / LUCA (Tiers 5–7) R1: Single donor gene 935 168 18.0% 81 8.7% 26.6% 71 7.6%
Cross-Kingdom / Supergroup / LUCA (Tiers 5–7) R2: Multi-paralog single taxon 129 21 16.3% 8 6.2% 22.5% 12 9.3%
Cross-Kingdom / Supergroup / LUCA (Tiers 5–7) R3: Multi-species single-clade jump 481 73 15.2% 36 7.5% 22.7% 25 5.2%
Cross-Kingdom / Supergroup / LUCA (Tiers 5–7) R4: Bracketed multi-species 629 50 7.9% 37 5.9% 13.8% 21 3.3%

Table 5C: Donor-Support Regimes by Individual Phylogenetic Tier (Tiers 1–7)

Phylogenetic Tier ($d_{\text{LCA}}$ of Closest Donor) Donor-Support Regime Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
Tier 1: Close Clade R1 / R2 / R3 / R4 1,208 / 311 / 321 / 1,187 106 / 15 / 13 / 62 8.8% / 4.8% / 4.0% / 5.2% 83 / 13 / 22 / 75 6.9% / 4.2% / 6.9% / 6.3% 15.6% / 9.0% / 10.9% / 11.5% 35 / 3 / 7 / 18 2.9% / 1.0% / 2.2% / 1.5%
Tier 2: Intermediate Clade R1 / R2 / R3 / R4 58 / 43 / 2 / 151 0 / 7 / 0 / 9 0.0% / 16.3% / 0.0% / 6.0% 6 / 12 / 2 / 17 10.3% / 27.9% / 100.0% / 11.3% 10.3% / 44.2% / 100.0% / 17.2% 0 / 2 / 0 / 4 0.0% / 4.7% / 0.0% / 2.6%
Tier 3: Phylum / Subphylum Split R1 / R2 / R3 / R4 512 / 145 / 140 / 1,414 65 / 25 / 6 / 81 12.7% / 17.2% / 4.3% / 5.7% 41 / 14 / 8 / 115 8.0% / 9.7% / 5.7% / 8.1% 20.7% / 26.9% / 10.0% / 13.9% 26 / 11 / 1 / 33 5.1% / 7.6% / 0.7% / 2.3%
Tier 4: Deep Metazoa / Bilateria Split R1 / R2 / R3 / R4 426 / 116 / 318 / 564 59 / 25 / 49 / 25 13.8% / 21.6% / 15.4% / 4.4% 27 / 5 / 20 / 54 6.3% / 4.3% / 6.3% / 9.6% 20.2% / 25.9% / 21.7% / 14.0% 17 / 7 / 18 / 8 4.0% / 6.0% / 5.7% / 1.4%
Tier 5: Cross-Kingdom Opisthokonta / Amorphea R1 / R2 / R3 / R4 720 / 88 / 372 / 492 114 / 8 / 57 / 43 15.8% / 9.1% / 15.3% / 8.7% 68 / 1 / 20 / 34 9.4% / 1.1% / 5.4% / 6.9% 25.3% / 10.2% / 20.7% / 15.7% 40 / 5 / 16 / 16 5.6% / 5.7% / 4.3% / 3.3%
Tier 6: Cross-Supergroup (Eukaryota root) R1 / R2 / R3 / R4 70 / 25 / 39 / 135 15 / 12 / 4 / 7 21.4% / 48.0% / 10.3% / 5.2% 1 / 3 / 2 / 3 1.4% / 12.0% / 5.1% / 2.2% 22.9% / 60.0% / 15.4% / 7.4% 6 / 6 / 4 / 5 8.6% / 24.0% / 10.3% / 3.7%
Tier 7: Cross-Domain (LUCA) R1 / R2 / R3 / R4 145 / 16 / 70 / 2 39 / 1 / 12 / 0 26.9% / 6.2% / 17.1% / 0.0% 12 / 4 / 14 / 0 8.3% / 25.0% / 20.0% / 0.0% 35.2% / 31.2% / 37.1% / 0.0% 25 / 1 / 5 / 0 17.2% / 6.2% / 7.1% / 0.0%

Table 5D: Donor-Support Regimes Cross-Tabulated with PANTHER Subfamily Status (SAME_SF vs CROSS_SF)

Because $d_{\text{MRCA}}$ vs $d_{\text{LCA}}$ measures species topology only, a phylogenetically bracketed (R4) row can still cross a paralogy boundary when the target and donor(s) sit in different PANTHER subfamilies (see human/PARD3B GO:0000226 below). Crossing Metric 5 (R1–R4) with Metric 3 separates the species-bracketing effect from the subfamily-divergence effect. On multi-donor rows (R2–R4), SAME_SF means target_sf in donor_sfs_list (i.e. at least one resolved donor shares the target's PANTHER subfamily, with the sub-rows below splitting SAME_SF by the number of distinct resolved donor subfamilies in donor_sfs_list: all resolved-SF donors in target_sf where donor_sfs_list == [target_sf], noting that additional donors on the row may lack a cached SF, vs mixed donor SFs where len(donor_sfs_list) > 1 so $\ge 1$ resolved donor matches target_sf and $\ge 1$ resolved donor sits in another subfamily), CROSS_SF means $\ge 1$ donor has a resolved subfamily and none matches target_sf, and UNKNOWN_SF means either the target or all donors lack a cached PANTHER subfamily:

Donor-Support Regime + Subfamily Status (and Resolved-SF Coverage) Rated ($n$) Fail % Fail Refn % Refn % Down REMOVE % REMOVE
R1: Single donor gene + SAME_SF (1,064 / 3,139 = 33.9% resolved) 697 71 10.2% 59 8.5% 18.7% 25 3.6%
R1: Single donor gene + CROSS_SF 367 66 18.0% 24 6.5% 24.5% 28 7.6%
R1: Single donor gene + UNKNOWN_SF 2,075 261 12.6% 155 7.5% 20.0% 96 4.6%
R2: Multi-paralog single taxon + SAME_SF (any donor matches) (315 / 744 = 42.3% resolved) 125 12 9.6% 9 7.2% 16.8% 4 3.2%
$\quad\hookrightarrow$ R2 SAME_SF: all resolved donors in target_sf 46 5 10.9% 6 13.0% 23.9% 3 6.5%
$\quad\hookrightarrow$ R2 SAME_SF: mixed donor SFs ($\ge 1$ in target_sf) 79 7 8.9% 3 3.8% 12.7% 1 1.3%
R2: Multi-paralog single taxon + CROSS_SF 190 37 19.5% 25 13.2% 32.6% 12 6.3%
R2: Multi-paralog single taxon + UNKNOWN_SF 429 44 10.3% 18 4.2% 14.5% 19 4.4%
R3: Multi-species single-clade jump + SAME_SF (any donor matches) (763 / 1,262 = 60.5% resolved) 252 24 9.5% 25 9.9% 19.4% 7 2.8%
$\quad\hookrightarrow$ R3 SAME_SF: all resolved donors in target_sf 218 17 7.8% 22 10.1% 17.9% 6 2.8%
$\quad\hookrightarrow$ R3 SAME_SF: mixed donor SFs ($\ge 1$ in target_sf) 34 7 20.6% 3 8.8% 29.4% 1 2.9%
R3: Multi-species single-clade jump + CROSS_SF 511 67 13.1% 39 7.6% 20.7% 26 5.1%
R3: Multi-species single-clade jump + UNKNOWN_SF 499 50 10.0% 24 4.8% 14.8% 18 3.6%
R4: Bracketed multi-species + SAME_SF (any donor matches) (2,535 / 3,945 = 64.3% resolved) 1,353 41 3.0% 105 7.8% 10.8% 19 1.4%
$\quad\hookrightarrow$ R4 SAME_SF: all resolved donors in target_sf 795 25 3.1% 55 6.9% 10.1% 11 1.4%
$\quad\hookrightarrow$ R4 SAME_SF: mixed donor SFs ($\ge 1$ in target_sf) 558 16 2.9% 50 9.0% 11.8% 8 1.4%
R4: Bracketed multi-species + CROSS_SF 1,182 117 9.9% 106 9.0% 18.9% 44 3.7%
R4: Bracketed multi-species + UNKNOWN_SF 1,410 69 4.9% 87 6.2% 11.1% 21 1.5%

Key takeaways from Tables 5A–5D:
- Multi-species single-clade jumps (R3: $d_{\text{MRCA}} > d_{\text{LCA}}$) behave like single-organism jumps (R1/R2), not like bracketed multi-species IBAs (R4): Citing two or more donor species from the same donor clade (e.g. mouse + rat, worm + fly, or two bacteria) yields an overall failure rate of 11.2% (4.0% REMOVE, $n = 1{,}262$)—nearly double that of phylogenetically bracketed multi-species IBAs (5.8% Fail, 2.1% REMOVE, $n = 3{,}945$). Holding closest-donor distance constant, unbracketed single-clade jumps fail at 2.2-fold higher rates in Tiers 3–4 (12.0% Fail in R3, $n = 458$, vs 5.4% Fail in R4, $n = 1{,}978$; and 15.4% Fail, 5.7% REMOVE in Tier 4 Bilateria jumps, $n = 318$, vs 4.4% Fail, 1.4% REMOVE in R4, $n = 564$) and 1.9-fold higher rates in Tiers 5–7 (15.2% Fail in R3, $n = 481$, vs 7.9% Fail in R4, $n = 629$).
- Band-stratified comparison of multi-paralog (R2) vs single-gene (R1) single-taxon IBAs: Although the unstratified failure rates of R2 (12.5%) and R1 (12.7%) look identical, R2 has a more proximal tier composition than R1 (47.6% of R2 rows are in Tiers 1–2 vs 40.3% of R1 rows). Within Tiers 1–2, R2 has a slightly lower failure rate than R1 (6.2% on $n = 354$ vs 8.4% on $n = 1{,}266$), whereas in Tiers 3–4 (Phylum/Bilateria Split) R2 reverses above R1 (19.2% Fail, 95% Wilson CI $14.8\text{–}24.4\%$, $n = 261$, vs 13.2% Fail, 95% CI $11.2\text{–}15.5\%$, $n = 938$), and in small-$n$ Tier 6 (Viridiplantae $\leftrightarrow$ Opisthokonta) R2 reaches 48.0% Fail ($12/25$, 95% CI $30.0\text{–}66.5\%$, hypothesis-generating given small $n$) because multiple co-annotated paralogs in one distant donor species often reflect a lineage-specific gene-family expansion (CROSS_SF: 19.5% Fail in Table 5D).
- Subfamily divergence (CROSS_SF) triples the failure rate even inside bracketed multi-species (R4) transfers: In Table 5D, R4 rows where at least one resolved donor shares the target's PANTHER subfamily (R4 + SAME_SF, $n = 1{,}353$) have only a 3.0% Fail (1.4% REMOVE) rate—and restricting to rows where all resolved-SF donors sit in the target's subfamily gives an identical 3.1% Fail (1.4% REMOVE, $n = 795$) (with 2.9% Fail, 1.4% REMOVE on the $n = 558$ mixed-subfamily rows where the target's subfamily still has direct donor support; the smaller R3 + SAME_SF mixed-donor cell at 20.6% Fail, $7/34$, 95% Wilson CI $10.3\text{–}36.8\%$, vs 7.8% Fail, 95% CI $4.9\text{–}12.1\%$ on $n = 218$ all-donor rows has wide overlapping confidence intervals and is hypothesis-generating given small $n$)—whereas R4 rows where the target sits in a different subfamily from all resolved donors (R4 + CROSS_SF, $n = 1{,}182$) jump to 9.9% Fail (3.7% REMOVE). Species bracketing (R4) therefore does not protect against paralog divergence when no experimental donor resides in the target's subfamily.
- Term refinement (% Refn) is invariant across all four regimes (7.0%–7.6%), confirming that the separation between unbracketed (R1, R2, R3: 11.2%–12.7% Fail) and bracketed (R4: 5.8% Fail) transfers is driven specifically by homology over-propagation rather than ontology granularity preferences.

Limitations and Methodological Confounds

  1. Circularity between phylogenetic distance and reviewer policy (and how --candidates works): The review actions (REMOVE, MARK_AS_OVER_ANNOTATED, MODIFY, ACCEPT, KEEP_AS_NON_CORE) in genes/*/*/*-ai-review.yaml are not an independent experimental ground truth. They were produced by AI and human reviewers following IBA_REVIEW.md guidelines that explicitly instruct reviewers to check phylogenetic node placement, subfamily divergence, and cross-kingdom pathway differences. Consequently, higher failure rates in Tiers 5–7, CROSS_SF, and unbracketed regimes (R1–R3) reflect both genuine biological divergence and heightened reviewer scrutiny of deep-node propagations. Conversely, some high-distance single-organism and single-clade IBAs in the baseline corpus remained at ACCEPT or KEEP_AS_NON_CORE even when the reviewer's own reason text noted a biological caveat. Specifically, --candidates in single_organism_distance.py filters on tier >= 4, action in {"ACCEPT", "KEEP_AS_NON_CORE"}, and reason matching 13 skeptical keywords ("do not", "does not", "lacks", "absent", "plant", "yeast-specific", "only in", "artifact", "over-annotat", "mislead", "cytosolic", "not directly", "unlike"); it does not filter on CROSS_SF, and passing --all-regimes extends --candidates, --tier, --tsv, and --json across all four regimes (R1–R4). Thus, --candidates serves as an internal reason-vs-action inconsistency detector, which we supplemented with PANTHER family/ortholog consistency sweeps to assemble the re-review candidates below. Note that harmonizing candidate proposals with sibling reviews on the same PAINT node is itself subject to the same-pipeline circularity caveat, since those sibling reviews were generated under the same curation guidelines.
  2. Sample size variation across strata: While Tiers 1, 3, 4, and 5 have large denominators ($n = 542\text{–}1{,}519$), Tier 2 ($n = 101$), Tier 6 ($n = 95$, including $n = 25$ in R2), Tier 7 ($n = 161$, including $n = 16$ in R2 and $n = 2$ in R4), the 02–06 steps bin ($n = 160$), and the >= 10 TreeGrafter steps bins ($n = 2$ and $n = 10$) have smaller sample sizes where a handful of rows shift percentages by several points.
  3. Partial local cache coverage, differential unresolved-seed exclusion, and Table 5D subfamily-resolution differences across regimes: Roughly 34% of close-pair single-organism IBAs (588 / 1,751 in Metric 2, plus 246 / 1,751 with an unrecorded blank taxon: in *-paint.tsv) and 64% of all rated single-organism rows (2,504 / 3,883 in Metric 3) belong to PANTHER families or external seed accessions not yet cached or lacking taxon metadata under interpro/panther/. Furthermore, because a GOA IBA row enters the regime analysis only when every WITH/FROM seed token resolves (not has_unresolved), multi-seed rows have a higher chance of containing at least one uncached UniProtKB: token: as reported by single_organism_distance.py --regimes (Section 5E, using n_distinct_donors via canonical_seed_gene), 303 single-distinct-seed vs 1,558 multi-distinct-seed non-self GOA IBA rows (1,861 total) are excluded due to $\ge 1$ unresolved seed token (this 303 / 1,558 split is provably invariant between raw len(seed_toks) and deduplicated n_distinct_donors, because any token present in mod_to_acc resolves by construction so no row with len(seed_toks) > 1 and n_distinct_donors == 1 can satisfy has_unresolved). Because multi-donor rows (R3/R4) are retained only when all donor tokens come from locally cached PANTHER families or MOD namespaces—and require only $\ge 1$ donor to have a cached PANTHER subfamily to enter SAME_SF/CROSS_SF—resolved subfamily coverage in Table 5D rises from 33.9% in R1 (1,064 / 3,139) and 42.3% in R2 (315 / 744) to 60.5% in R3 (763 / 1,262) and 64.3% in R4 (2,535 / 3,945). Within-regime SAME_SF vs CROSS_SF contrasts are unaffected by this coverage gradient, but cross-regime Table 5D cell comparisons (e.g. R1 + CROSS_SF vs R4 + CROSS_SF) compare a 34% subset of R1 against a 64% subset of R4. Among resolved rows, Metrics 1 and 1B cover 99.8% (3,883 / 3,890) of rated single-organism rows (7 excluded in Tier 0 from Sus scrofa 9823 lacking a local *-uniprot.txt file), and Metric 5 covers 98.3% (9,090 / 9,246) of rated non-self resolved IBA rows across Regimes 1–4 (156 excluded in Tier 0: 7 single-taxon and 149 multi-taxon rows citing at least one donor species without a local *-uniprot.txt lineage).
  4. Coarse Xenbase taxon mapping and Tier 2 residual clade fallback: In MOD_TO_TAXON, Xenbase is mapped at the prefix level to ("8355", "Xenopus laevis") when the Xenbase: token is not resolved via a local genes/XENTR/*/*-uniprot.txt cross-reference (build_mod_to_acc). For multi-donor rows that cite both an uncached Xenbase: token (mapped to 8355 X. laevis) and a Xenopus tropicalis UniProtKB: token (8364), this prefix fallback treats them as two congeneric Xenopus species (n_donor_taxa = 2) rather than a single donor species, shifting such rows from R1/R2 into R3/R4. Separately, because classify_evolutionary_span routes any unenumerated LCA clade string to Tier 2: Other ({lca}), Tier 2 ($n = 101$ rated single-organism rows) combines 99 explicitly mapped fungal/plant order/subphylum rows (Magnoliopsida: 51, Saccharomycotina: 39, rosids: 6, Pentapetalae: 2, Pezizomycotina: 1) with 2 fallback rows (Poaceae: 1, Pseudomonadati: 1).
  5. Exact vs term-fallback review-action matching (rev_by_exact vs rev_by_term): Many older *-ai-review.yaml files omit supporting_entities on IBA entries. When matching a GOA IBA row to its review action, single_organism_distance.py first looks up (go_id, frozenset(supporting_entities)) (EXACT) and falls back to rev_by_term[go_id] (TERM_FALLBACK, which retains the last IBA entry for that go_id). Across the 9,090 resolved-lineage rated rows (Section 5E), EXACT matching covers 68.1% of R1 (2,138 / 3,139), 71.0% of R2 (528 / 744), 73.1% of R3 (923 / 1,262), and 76.1% of R4 (3,001 / 3,945). When a gene review omits supporting_entities and has multiple GOA IBA rows for the same go_id in different regimes, TERM_FALLBACK assigns the same review action to each row.

Candidate Failure Patterns Identified for Re-Review (and Valid Deep Conservation)

Combining uv run python projects/IBA_REVIEW/single_organism_distance.py --candidates (and --candidates --all-regimes) with PANTHER family/ortholog consistency sweeps across single-organism (R1/R2), single-clade jump (R3), and subfamily-divergent (CROSS_SF) reviews surfaces six recurring biological failure mechanisms—comprising 15 initial single-organism candidate genes (PR #5703 for Tiers 6–7 and PR #5704 for Tiers 4–5), 17 sibling-harmonization genes (PR #5857), 2 cross-subfamily high-distance audit genes (PR #5861), and 5 multi-paralog (R2), single-clade jump (R3), and subfamily-divergent (R4 + CROSS_SF) candidate genes (PR #6157)—alongside valid deep conservation:

  1. Cross-Domain Organelle / Pathway Projection (Bacteria $\leftrightarrow$ Eukaryota, Tier 7)
  2. Mechanism: A bacterial enzyme or plastid-targeted nuclear factor is used to seed a LUCA or bacterial node for a compartment-specific or pathway-specific claim that differs between bacteria, plastids, and eukaryotic cytosol/mitochondria, or a MOD over-annotation on a plastid subunit is propagated to bacteria.
  3. Single-organism (R1) candidates (PR #5703):
    • human/MDH1 (GO:0006099 tricarboxylic acid cycle from M. tuberculosis mdh UniProtKB:P9WK13 via PTN000600336 at LUCA; KEEP_AS_NON_CORE $\to$ REMOVE in PR #5703): M. tuberculosis has a single TCA-cycle MDH, whereas eukaryotic MDH1 is the cytosolic malate–aspartate shuttle isoform (mitochondrial MDH2 runs the TCA cycle).
    • BACSU/sigF (GO:0000976 transcription cis-regulatory region binding from an Arabidopsis thaliana plastid sigma factor TAIR:locus:2084420 via PTN000779307; ACCEPT $\to$ REMOVE in PR #5703): Bacterial sigma factors recognize promoter DNA only as part of the RNA polymerase holoenzyme (GO:0016987 sigma factor activity; GO:0003677 DNA binding), aligning sigF with paralog BACSU/sigG.

  4. Single-clade jump (R3) candidate (PR #6157):

    • human/ASAH2 (GO:0042759 long-chain fatty acid biosynthetic process from two bacterial neutral ceramidases in M. tuberculosis and P. aeruginosa UniProtKB:O06769|UniProtKB:Q9I596 via PTN004377481 at LUCA; baseline action ACCEPT $\to$ proposed MARK_AS_OVER_ANNOTATED): ASAH2 is a catabolic plasma-membrane neutral ceramidase (GO:0006672 ceramide metabolic process) that cleaves the amide bond of ceramide to form sphingosine (the physiological salvage-pathway product, GO:0046512) while releasing the fatty acyl chain as a hydrolytic leaving group (PMID:10781606, PMID:26190575) rather than performing long-chain fatty acid biosynthesis; re-reviewing to MARK_AS_OVER_ANNOTATED harmonizes human/ASAH2 with 1:1 vertebrate ortholog DANRE/asah2 on the same PTN004377481 node.

  5. Plant Secondary Metabolism / Subfamily Expansion Projected to Fungi & Animals (Viridiplantae $\leftrightarrow$ Opisthokonta, Tier 6)

  6. Mechanism: Land plants expanded enzyme families for specialized pathways (agmatine-route putrescine synthesis, purine/cytokinin nucleoside hydrolysis, plastidial MEP-pathway isoprenoid synthesis) that are absent, non-physiological, or reversed in fungi and metazoans.

  7. Single-organism (R1/R2) candidates (PR #5703 & PR #5857):

    • SCHPO/SPAC11D3.09 & SCHPO/SPAPB24D3.03 (GO:0009446 putrescine biosynthetic process from Arabidopsis agmatinases AT4G08870/AT4G08900 via PTN000144486 at Eukaryota; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5703): Fungi synthesize putrescine from ornithine via ornithine decarboxylase and lack an endogenous arginine decarboxylase pathway to generate agmatine.
    • SCHPO/SPAC17G8.02 & SCHPO/SPBC1683.06c (GO:0006152 purine nucleoside catabolic process, a Regime 2 Tier 6 IBA from Arabidopsis NRH1/NRH2 AGI_LocusCode:AT1G05620|AGI_LocusCode:AT2G36310 via PTN000266727 at Eukaryota, KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED alongside GO:0008477 purine nucleosidase activity in PR #5703): Fungal Urh1-family enzymes are pyrimidine-preferring uridine hydrolases (GO:0006218 / GO:0045437) with negligible purine nucleosidase activity.
    • human/IDI1 & human/IDI2 (GO:0009240 isopentenyl diphosphate biosynthetic process from Arabidopsis IPP1 AGI_LocusCode:AT5G16440 via PTN000853999 at Eukaryota; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5703, plus sibling yeast/IDI1 harmonized to MARK_AS_OVER_ANNOTATED in PR #5857): In the eukaryotic mevalonate pathway, MVD synthesizes IPP first and IDI1/IDI2 catalyze its isomerization to DMAPP (GO:0004452 / GO:0050992).

  8. Yeast-Specific Organelle Topologies & Subfamily/Pathway Divergence Projected to Metazoa & Amoebozoa (Fungi $\leftrightarrow$ Metazoa/Amoebozoa, Tier 5–6)

  9. Mechanism: S. cerevisiae closed-mitosis / nucleus–vacuole junction (NVJ) processes, yeast-specific pathways, or fermentation ester synthases are seeded at Eukaryota or Opisthokonta from SGD annotations and propagated to organisms that lack a yeast-type vacuole, NVJ, or fermentation pathway, or across a subfamily duplication boundary.
  10. Single-organism (R1) candidates (PR #5704, PR #5857, & PR #5861):
    • human/ATG4D (GO:0034727 piecemeal microautophagy of the nucleus from S. cerevisiae ATG4 SGD:S000005167 via PTN002503342 at Eukaryota; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5704, matching human/ATG4C and worm/atg-4.1), human/ATG2B (SGD:S000005186 via PTN000324023) and DICDI/atg1 (SGD:S000003148 via PTN000681273; KEEP_AS_NON_CORE $\to$ UNDECIDED in PR #5704, matching human/ATG2A and human/ULK1), plus core autophagy machinery siblings human/ATG5, human/ATG7, worm/atg-5, worm/atg-7, SCHPO/atg2, SCHPO/atg5, and SCHPO/atg13 harmonized to MARK_AS_OVER_ANNOTATED on GO:0034727 in PR #5857.
    • human/CALR3 (GO:0036503 ERAD pathway from S. cerevisiae CNE1 SGD:S000000039 via PTN000125949 at Eukaryota, CROSS_SF: PTHR11073:SF3 CALRETICULIN-3 vs PTHR11073:SF1 CALNEXIN 14D-RELATED; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5861) and human/IDO1 (GO:0034354 'de novo' NAD biosynthetic process from L-tryptophan from S. cerevisiae BNA2 SGD:S000003923 via PTN000127988 at Opisthokonta, CROSS_SF: PTHR28657:SF2 INDOLEAMINE 2,3-DIOXYGENASE 1 vs PTHR28657:SF5 INDOLEAMINE 2,3-DIOXYGENASE; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5861).

  11. Multi-paralog single-taxon (R2) candidate (PR #6157):

    • human/ABHD2 (GO:0051792 medium-chain fatty acid biosynthetic process from three S. cerevisiae fermentation medium-chain fatty acid ethyl ester synthases/esterases EHT1|EEB1|YMR210W SGD:S000000381|SGD:S000004823|SGD:S000006016 and GO:0051793 medium-chain fatty acid catabolic process from SGD:S000000381|SGD:S000006016 via PTN000082647 at Opisthokonta; baseline actions KEEP_AS_NON_CORE $\to$ proposed REMOVE on GO:0051792 and MARK_AS_OVER_ANNOTATED on GO:0051793): Human ABHD2 (PTHR10794:SF45 MONOACYLGLYCEROL LIPASE ABHD2) is a progesterone-activated plasma-membrane 2-arachidonoylglycerol (long-chain monoacylglycerol) hydrolase (GO:0047372 acylglycerol lipase activity, PMID:26989199) with no fatty-acid biosynthetic activity (REMOVE on GO:0051792) and only generic in vitro short/medium-chain $p$-nitrophenyl esterase activity (MARK_AS_OVER_ANNOTATED on GO:0051793).

  12. Metazoan Endomembrane / Tissue / Developmental Specialization Projected to Fungi or Invertebrates (Metazoa $\to$ Fungi / Protostomia, Tier 4–5)

  13. Mechanism: Metazoan- or vertebrate-specific organelle localization, cofactor-binding, or developmental symmetry traits from one or two animal donors are placed at Bilateria, Opisthokonta, or LUCA and inherited by fungi or invertebrates lacking the donor pathway or organelle specialization.
  14. Single-organism (R1) candidates (PR #5704):
    • SCHPO/lsb6 (GO:0005802 trans-Golgi network, GO:0007030 Golgi organization, GO:0007032 endosome organization from Drosophila Pi4KIIalpha FB:FBgn0037339 via PTN000305230; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5704).
    • SCHPO/oar2 (GO:0048038 quinone binding from human mitochondrial CBR4 UniProtKB:Q8N4T8 via PTN002449501 at LUCA; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5704).
    • SCHPO/cit1 (GO:0005975 carbohydrate metabolic process from human CS UniProtKB:O75390 via PTN000201223 at LUCA; KEEP_AS_NON_CORE $\to$ MARK_AS_OVER_ANNOTATED in PR #5704).

  15. Single-clade jump (R3) candidates (PR #6157):

    • ANOGA/DNAAF1 (GO:0007368 determination of left/right symmetry from two vertebrate Euteleostomi donors Danio rerio|Homo sapiens UniProtKB:Q8NEP3|ZFIN:ZDB-GENE-031114-4 via PTN002915666 at Bilateria; baseline action KEEP_AS_NON_CORE $\to$ proposed REMOVE): Left–right body-axis determination via motile nodal cilia at the embryonic left–right organizer (PMID:29228333) is a deuterostome/vertebrate developmental mechanism absent in dipteran insects such as Anopheles gambiae, whose motile axonemes are restricted to sperm flagella and mechanosensory chordotonal neurons.
    • worm/pdr-1 (GO:0005783 endoplasmic reticulum and GO:0005794 Golgi apparatus from two mammalian Euarchontoglires donors Homo sapiens|Rattus norvegicus RGD:61797|UniProtKB:O60260 via PTN000188399 at Bilateria; baseline action UNDECIDED $\to$ proposed MARK_AS_OVER_ANNOTATED): C. elegans PDR-1 functions in the cytosol and at mitochondria for ubiquitylation and mitophagy (PMID:15961382, PMID:25896323), with no evidence of ER or Golgi residence in worms; re-reviewing to MARK_AS_OVER_ANNOTATED harmonizes worm/pdr-1 with invertebrate ortholog DROME/park on the same PTN000188399 node.

  16. Holdase vs ATP-Dependent Foldase Conflation & Paralog Subfunctionalization Across Vertebrate Duplications (Protostomia $\to$ Deuterostomia, Tier 4, and R4 + CROSS_SF)

  17. Mechanism: Invertebrate Ecdysozoa donors (C. elegans, D. melanogaster) have a single ancestral gene or lineage-specific paralog cluster before vertebrate whole-genome duplication, and their ancestral function set is placed at Bilateria/Eumetazoa and propagated into vertebrate paralogs that act via a different mechanism or lost part of the ancestral toolkit (FUNCTIONAL_DIVERGENCE).
  18. Single-organism (R2) sHSP candidates (PR #5704 & PR #5857): Six or seven Drosophila small heat shock proteins in D. melanogaster (Hsp23 FB:FBgn0001224, Hsp26 FB:FBgn0001225, Hsp27 FB:FBgn0001226, l(2)efl FB:FBgn0011296, CG14207 FB:FBgn0031037, and CG7409 FB:FBgn0035817 on the 20220316 GOA rows of human/CRYAA, human/CRYAB, human/HSPB2, and human/HSPB6, plus Hsp22 FB:FBgn0001223 as a seventh seed on the 20260530 GOA rows of human/HSPB1, DANRE/cryaa, DANRE/cryaba, and DANRE/cryabb) seed GO:0042026 protein refolding on PTN000897708 (cached taxon taxon:6072 Eumetazoa; donor–target LCA Metazoa), propagating to vertebrate and nematode sHSPs. In the pre-#5857 state (after PR #5704 moved human/HSPB2 from KEEP_AS_NON_CORE to MARK_AS_OVER_ANNOTATED), the eight vertebrate sHSP reviews on this node spanned four distinct actions: ACCEPT (human/HSPB1), MARK_AS_OVER_ANNOTATED (DANRE/cryaa, DANRE/cryaba, human/HSPB2), MODIFY (DANRE/cryabb to proposed NTR holdase chaperone activity; human/CRYAB to GO:0050821 protein stabilization; human/HSPB6 to GO:0006457 protein folding), and REMOVE (human/CRYAA). PR #5857 harmonized the eight remaining non-MARK_AS_OVER_ANNOTATED vertebrate and nematode sibling reviews (DANRE/cryabb, human/CRYAB, human/HSPB1, human/HSPB6, worm/hsp-16.1, worm/hsp-16.2 from REMOVE, worm/hsp-16.41, and worm/hsp-16.48) to MARK_AS_OVER_ANNOTATED and confirmed MARK_AS_OVER_ANNOTATED on human/HSPB2 (updating its propagation_review source_status to SUPPORTS_SOURCE_BUT_NOT_TARGET and syncing supporting_entities) because sHSPs act as ATP-independent holdase chaperones that prevent aggregation rather than catalyzing the active refolding step of GO:0042026.

  19. Single-clade jump (R3) and bracketed cross-subfamily (R4 + CROSS_SF) candidate (PR #6157):

    • human/PARD3B (GO:0007155 cell adhesion, a Regime 3 Tier 4 single-clade jump from Ecdysozoa donors C. elegans par-3|D. melanogaster baz FB:FBgn0000163|WB:WBGene00003918 via PTN001046835 where $d_{\text{MRCA}} = 3 > d_{\text{LCA}} = 2$; and GO:0000226 microtubule cytoskeleton organization, a Regime 4 Tier 1 bracketed row from FB:FBgn0000163|MGI:MGI:2135608|WB:WBGene00003918 on the same node where $d_{\text{MRCA}} = 2 \le d_{\text{LCA}} = 9$; baseline actions KEEP_AS_NON_CORE $\to$ proposed MARK_AS_OVER_ANNOTATED): In interpro/panther/PTHR16484/PTHR16484-entries.csv, vertebrate canonical PARD3/Pard3 (Q8TEW0, Q99NH2, Q9Z340) is in PTHR16484:SF10 and C. elegans par-3 (Q17353) is in PTHR16484:SF17, whereas vertebrate PARD3B/Pard3b (Q8TEW8, Q9CSB4) is in the diverged subfamily PTHR16484:SF4 (and MGI:2135608 resolves at the taxon level to Mus musculus via MOD_TO_TAXON["MGI"]). Unlike canonical PARD3, vertebrate PARD3B (Par3L) mutated the FXR/KR motifs and does not bind aPKC (PMID:12459187, PMID:33565714), so propagating family-root PTN001046835 cell-adhesion and microtubule-organization annotations across the vertebrate duplication boundary over-annotates PARD3B (harmonizing human/PARD3B with 1:1 mammalian ortholog HORSE/PARD3B on PTN001046835). Notably, GO:0000226 illustrates why Metric 5 must be paired with Metric 3 (Table 5D): even when a row is phylogenetically bracketed in species space (R4, Tier 1 closest donor), propagation fails when the PAINT node sits above a vertebrate paralogy boundary (CROSS_SF).

  20. Valid Deep Conservation (When Single-Organism and Single-Clade High-Distance IBAs Succeed)

  21. Importantly, 65.2% of Tier 7 (LUCA) and 76.4% of Tier 5 (Opisthokonta) single-organism IBAs—and 81.9% of Regime 3 single-clade jumps overall—are retained (ACCEPT / KEEP_AS_NON_CORE) in the corpus. These succeed when:
    1. The term describes a core catalytic molecular function (F aspect, e.g., malate dehydrogenase activity, isopentenyl-diphosphate delta-isomerase activity, uridine nucleosidase activity, 3-oxoacyl-[acyl-carrier-protein] reductase (NADPH) activity).
    2. Active-site residues and domain architecture are conserved across the SAME_SF orthogroup without neofunctionalizing duplications.
    3. The term does not bake in lineage-specific anatomical, organelle-contact, or pathway-direction assumptions.