TreeGrafter Inference Evaluation

Bottom line: TreeGrafter grafts a protein that is not in a PANTHER reference
tree onto the best-matching node and copies that node's GO terms to it as IEA
annotations (GO_REF:0000118), with no curator in the loop. We took every such
annotation in the review corpus at commit f81b9f300 (1,074 annotations on
577 reviewed proteins) and tallied how reviewers treated them, alongside the
curated PAINT/IBA set as a contrast. Reviewers accepted 43.2% of TreeGrafter
annotations as-is, removed 11.1% outright and down-graded 26.4% in all
(REMOVE, MODIFY or MARK_AS_OVER_ANNOTATED), against 73.1% accepted for
PAINT/IBA; molecular-function terms fared worst, with 41.5% down-graded. When
another pipeline reproduced the same TreeGrafter call (GO_REF:0000120)
acceptance rose to 77%, though reviewers could see that label. In about three
of four down-graded cases (four of five of those assigned a failure mode) the
tree placement was sound and the inherited term was the problem (too coarse, a
sibling term, or a generic or out-of-context term). The errors cluster by
family: 26 of the 77 PANTHER families with at least four reviewed annotations
had half or more of their terms down-graded. Two mechanical checks pick out
down-grades well: a GO term whose EC number contradicts the protein's own
UniProt EC (53 rows, 92% down-graded) and a query organism outside the clade
of the PAINT node the term came from (45 rows, 69%); see
Propagation rules and graft paths.
Much of what TreeGrafter says is right and said nowhere else: 623 rated rows
are the protein's only annotation to that term or a related one, and the
reviews kept 70% of them; where it and another pipeline give related terms and the
review keeps only one, it keeps TreeGrafter more often (see
What TreeGrafter gets right).

We did this because TreeGrafter output is routinely conflated with curated
PAINT/IBA, and knowing where automated grafting over-reaches gives PANTHER and
PAINT curators concrete families to fix. 65% of the rows come from the
Pseudomonas putida KT2440 batch, so the rates are directional, and they are
pinned to a commit because the corpus keeps growing.

Overview

TreeGrafter (Tang et al. 2019,
doi:10.1093/bioinformatics/bty625)
is the algorithm — bundled into InterProScan — that grafts a query protein
onto the most appropriate PANTHER reference phylogenetic tree and then
propagates the GO annotations attached to the grafting node down onto the query.

The critical distinction this project keeps straight is TreeGrafter vs.
PAINT/IBA, which are routinely conflated:

So the TreeGrafter-based inferences are the IEA / GO_REF:0000118 /
assigned-by TreeGrafter annotations — not the IBA ones. (We verified this
directly in the corpus GOA: every GO_REF:0000118 row is IEA / TreeGrafter
/ PANTHER.)

One more reference matters, and it changes what the GO_REF:0000118 set is.
GO_REF:0000120 is UniProt's "Combined Automated Annotation using Multiple
IEA Methods": per its GO reference-collection definition it "integrates
identical annotations from multiple electronic pipelines, including UniRule,
ARBA, InterPro2GO, TreeGrafter2GO (GO_REF:0000118), RHEA2GO, KeyWord2GO,
SubCellular2GO, EC2GO and EnsEMBL Compara", listing the contributing pipelines
pipe-separated in WITH/FROM. In this corpus every GO_REF:0000120 row
with a PANTHER:PTN… in WITH/FROM (979 rows in the cached GOA files at
f81b9f300; 891 at the 2026-09-06 snapshot) also lists at least one other
pipeline (InterPro in most cases), and no reviewed protein has the same term
under both GO_REF:0000118 and GO_REF:0000002 (0 shared pairs in the
summary sidecar). So the three electronic references
partition the predictions by corroboration:

The headline rates below are therefore for the uncorroborated TreeGrafter
output. The corroborated set and the uncorroborated InterPro2GO set are
evaluated alongside it in Corroboration.

This project evaluates how good those automated TreeGrafter inferences are when
each is held to standard GO review criteria, using the AIGR corpus of
agent-written gene reviews as the reference standard (see
Reference standard).

Method

Each genes/*/*/*-ai-review.yaml records, per existing annotation, an
evidence_type, an original_reference_id, and a reviewer action from the
AIGR action enum (ACCEPT, KEEP_AS_NON_CORE, MODIFY,
MARK_AS_OVER_ANNOTATED, REMOVE, UNDECIDED, …). We treat the AIGR
adjudication as the reference judgement (see
Reference standard) and ask how the annotations with
evidence_type: IEA and original_reference_id: GO_REF:0000118 (TreeGrafter)
were treated. The PAINT/IBA set (GO_REF:0000033) is reported alongside purely
as a contrast — it is a different, curator-driven pipeline, on a largely
different set of genes. Existing annotations marked retired: true (rows that
a later GOA refresh no longer contains, kept in the review for provenance) are
not live annotations and are skipped by every script.

Reproduce everything on this page (about three minutes; only graft_check.py
needs the network):

uv run --with pyyaml projects/TREEGRAFTER/analyze_treegrafter.py   # review/contrast/summary
python3 projects/TREEGRAFTER/analyze_placement.py                   # placement + family hotspots
uv run --with pyyaml projects/TREEGRAFTER/classify_failure_modes.py # failure modes
python3 projects/TREEGRAFTER/snapshot_drift.py                      # drift vs the 2026-09-06 snapshot
uv run --with pyyaml projects/TREEGRAFTER/exemplar_independence.py  # exemplar citation check
python3 projects/TREEGRAFTER/graft_check.py                         # live UniProt; or --refresh-actions offline
python3 projects/TREEGRAFTER/propagation_rules.py --go-obo go-basic.obo  # rule scoring (GO from purl.obolibrary.org)
uv run --with pyyaml projects/TREEGRAFTER/tg_versus_other_sources.py --go-obo go-basic.obo  # TreeGrafter vs other sources
uv run --with pyyaml projects/TREEGRAFTER/tg_versus_protnlm.py --go-obo go-basic.obo        # TreeGrafter vs ProtNLM2
python3 projects/TREEGRAFTER/ec_contradiction_scale.py              # live UniProt + QuickGO
python3 projects/TREEGRAFTER/trace_graft_paths.py                  # examples: live PANTHER tree API + QuickGO
python3 projects/TREEGRAFTER/trace_graft_paths.py --all --panther-trees Tree_MSF \
    --ibd-gaf IBD.gaf --paint-total PAINT_TreeGrafter_Annotations_TOTAL.txt.gz  # every row, offline

analyze_treegrafter.py --out-dir DIR writes its three sidecars elsewhere, so
the numbers can be checked against a newer tree without touching the committed
files. It writes:

• TREEGRAFTER/treegrafter_review.tsv — one row per reviewed TreeGrafter annotation (gene, taxon, term, aspect, action).
• TREEGRAFTER/treegrafter_contrast.tsv — one row per reviewed corroborated-PANTHER (GO_REF:0000120) and InterPro2GO (GO_REF:0000002) annotation on the review files that carry TreeGrafter rows.
• TREEGRAFTER/treegrafter_summary.tsv — action counts for all populations; a headline rates block (accept, retained, REMOVE-only, MODIFY, MARK_AS_OVER_ANNOTATED, down-graded) per population and per aspect; per-taxon breakdown; most frequently down-graded terms.

analyze_placement.py joins each annotation to its PANTHER family / subfamily
/ graft node and writes
treegrafter_family_hotspots.tsv;
classify_failure_modes.py assigns every down-graded annotation a failure
mode and writes
treegrafter_failure_modes.tsv
(see the failure-modes page).

Results (computed at commit f81b9f300, 2026-10-03)

Provenance. Every number on this page and its sub-page comes from the
sidecars regenerated on 2026-10-03 at commit f81b9f300 with the
commands above (the genes/ tree was clean at that commit). The corpus keeps
growing; re-run the commands rather than trusting these figures on a later
tree, and re-curate failure_mode_curated.tsv whenever rows cross into or out
of the down-graded set.

5,549 review files were scanned, yielding 1,074 reviewed TreeGrafter
annotations (GO_REF:0000118) across 577 review files (557 distinct gene
symbols). No row is PENDING/NEW.

Relative to the previous refresh (2026-10-01 at 943b98815; 998 annotations /
560 files), 80 annotations were added, 4 dropped and 75 actions changed, and
the down-graded set went from 299 to 283 (66 rows left it, 50 entered,
all 50 of them newly added rows). Three events account for all of it:

• The rejection re-review is now in the
  tables. All 75 changed actions are rows audited in
  rereview-2026-09-24/: 54 MARK_AS_OVER_ANNOTATED → KEEP_AS_NON_CORE,
  7 REMOVE → MARK_AS_OVER_ANNOTATED, 4 MARK_AS_OVER_ANNOTATED →
  ACCEPT, 3 MARK_AS_OVER_ANNOTATED → MODIFY, 3 REMOVE → UNDECIDED,
  2 REMOVE → KEEP_AS_NON_CORE, 1 REMOVE → ACCEPT and 1 REMOVE →
  MODIFY.
• The unicellular holozoan rows are now in the tables. 65 of the 80 added
  rows are the holozoan case study
  proteins (S. rosetta 41, Capsaspora 19, the sponge Oscarella pearsei
  5), 41 of them down-graded; the other 15 are new P. putida pyrimidine
  rows (pyrB–pyrF, pyrC', carB; 9 down-graded).
• Retired and merged rows are gone. The 4 dropped rows are the three
  viral rows that main's GOA refresh marked retired: true (9CAUD/g022
GO:0006302, REMOVE; 9CAUD/dfrP GO:0046452 ACCEPT and GO:0046655
KEEP_AS_NON_CORE; see the 2026-09-28 note) and the duplicate
  PSEPK/aroQ folder merged into aroQ-III (its GO:0019631 row survives
  under aroQ-III). The previous refresh still counted the three retired rows;
  the scripts now skip retired: true.

Drift since the 2026-09-06 snapshot

An earlier version of this page was frozen at a 2026-09-06 snapshot (898
annotations / 510 files, on main at e3350f43d).
snapshot_drift.py diffs the two
(treegrafter_snapshot_drift.tsv):

Most of the drift comes from two audits: the 2026-09-20 full-gene
re-review recorded in TREEGRAFTER/rereview-2026-09-20/
(below), plus the follow-up reviews of its awaiting genes, and the
2026-09-24 rejection re-review. Of the
131 changed actions, 57 are on genes in the 2026-09-20 folder and 71 on rows
audited on 2026-09-24 (the other three are IRE1 GO:0051082, fabF
GO:0005829 and fruA GO:0090563); of the 98 rows that left the
down-graded set, 37 are on 2026-09-20 genes and 61 on 2026-09-24 rows. The
most common transitions are MARK_AS_OVER_ANNOTATED → KEEP_AS_NON_CORE
(53 rows), REMOVE → MARK_AS_OVER_ANNOTATED (14),
MARK_AS_OVER_ANNOTATED → ACCEPT (11), REMOVE → UNDECIDED (10) and
MODIFY → ACCEPT (10). fogD illustrates the 2026-09-20 path: the re-review moved its six
REMOVE rows to UNDECIDED rather than reject on pathway context alone, and
the later follow-up review settled them as MODIFY / MARK_AS_OVER_ANNOTATED
(four rows) and KEEP_AS_NON_CORE (two), so in the snapshot diff those four
stay down-graded.

Headline

Three rates, reported side by side because they answer different questions
(all from the headline rates block of the summary sidecar):

MARK_AS_OVER_ANNOTATED sits in both retained and down-graded on
purpose: the term is not wrong, but it is flagged as over-reaching. Read
together: roughly one TreeGrafter inference in ten is judged wrong outright,
about three in four survive in some form, and fewer than half are accepted
as-is. The accept rate was 41.0% at an earlier 415-annotation snapshot,
41.3% at the 2026-09-06 snapshot, 45.1% at the 2026-09-27 refresh and 44.6%
at the 2026-10-01 refresh; the rise since the snapshot is mostly the
2026-09-20 re-review above, not new data. The fall to 43.2% since 2026-10-01
is the 80 added rows (15 accepted, 50 down-graded), mostly holozoan. The
rejection re-review barely moves the accept rate (5 rows to ACCEPT); what it
moves is the split between MARK_AS_OVER_ANNOTATED (11.5% → 8.0%) and
KEEP_AS_NON_CORE (21.0% → 25.8%), so the down-graded rate falls (30.0% →
26.4%) while the retained rate is almost unchanged (77.2% → 77.0%).

The PAINT/IBA column is not a like-for-like comparison. It is every IBA row
in the corpus (3,394 review files, dominated by human and model-organism
genes), while TreeGrafter by construction annotates sequences that are not
in the PANTHER reference tree — so the two populations barely share genes.
Restricted to the same 577 review files, there are only 20 IBA rows on 9
files (16 ACCEPT, 3 KEEP_AS_NON_CORE, 1 UNDECIDED): too few to say
anything. The contrast shows that curated IBA on well-studied genes fares
better than automated grafting on non-model genes; it does not isolate the
effect of the pipeline from the effect of the gene set. The
corroboration contrast
below, which is on the same files, is the better-controlled comparison.

By GO aspect

Molecular-function propagations are the worst category by a wide margin (one
in five MF rows is removed outright); CC terms are rarely wrong but are
often parked as non-core.

Corroboration: TreeGrafter-only vs multi-method vs InterPro2GO-only

Each population below is restricted to the 577 review files that carry
TreeGrafter rows, but covers only the subset of those files that carry it
(412 and 407 files respectively), so the three are close to — not exactly —
the same proteins:

Three things follow:

1. Corroboration is the strongest single predictor of a good TreeGrafter
   call. The same algorithm's output is accepted 77% of the time when another
   pipeline reproduces it and 43% when none does. UniProt's GO_REF:0000120
   merge is, in effect, already a QC gate, and the GO_REF:0000118 residue is
   the part that failed it.

   Caveat — the reviewers were not blind to the label.
   original_reference_id is visible in the review YAML while the annotation
   is being adjudicated, and "combined multiple IEA methods"
   (GO_REF:0000120) reads as visibly stronger provenance than a lone
   GO_REF:0000118. So the 77%-vs-43% gap may be partly caused by the label
   rather than only predicted by it. The direction of the effect is very
   likely real — corroboration by an independent pipeline is genuine evidence
   — but the magnitude should not be taken at face value from this corpus;
   treat it as an upper bound until the provenance-blinded test in Next steps
   is run.

2. Uncorroborated InterPro2GO is down-graded more often than uncorroborated
   TreeGrafter — 51% (54% for MF) — but mostly by MARK_AS_OVER_ANNOTATED
   (32%); its REMOVE rate (6.5%) is lower than TreeGrafter's (11.1%), and
   its retained rate is similar (79% vs 77%). Its down-graded terms are
   dominated by coarse ancestors — catalytic activity (45), membrane (24),
   oxidoreductase activity (21), nucleotide binding (15). So InterPro2GO's
   uncorroborated residue is mostly uninformative, while TreeGrafter's is
   more often wrong. This qualifies the graft-check hypothesis on the
   failure-modes page: InterPro does out-resolve PANTHER for particular
   proteins (AprA, S-crystallin, Mcr1), but the informative InterPro2GO calls
   have mostly already been merged into GO_REF:0000120.

3. Taxon composition. TreeGrafter's down-grade rate is now lower on the
   P. putida KT2440 rows (21.9%, 154/703) than elsewhere (34.8%, 129/371);
   it was 29.3% vs 31.4% at the 2026-10-01 refresh and 31% vs 42% at the
   2026-09-06 snapshot. The gap reopened for two reasons: 56 of the 66 rows
   that the rejection re-review moved out of the down-graded set since
   2026-10-01 are P. putida rows, and the holozoan rows added outside
   P. putida are mostly down-graded (41 of 65).
   InterPro2GO is down-graded 56.6% on P. putida and 35.3% elsewhere (the
   P. putida reviews were strict about generic MF terms). Computed from the
   PSEPK review files in the review and contrast sidecars.

Where TreeGrafter inferences fail

Deep dive: Failure Modes & Tree Placement
joins every down-graded annotation to the PANTHER family/subfamily it was
grafted onto (and the ancestral PTN graft node), and assigns each one a
failure mode. Short answer: in about three cases out of four (four of five
of those assigned a mode) the placement is fine and the inherited term is the
problem — too coarse or a sibling term from the family node (47%), or a
generic / out-of-context localization or process (25%). About one in seven
(42 annotations on 29 proteins, e.g. aprA, fcs, mdh, mqo1–3, dapE,
and Capsaspora coWts) is a mis-placement, genuine pseudo-enzymes are
rare (5 annotations, 3 proteins), and 12% are left unclassified (about half
of them the choanoflagellate cadherin over-annotations).

Mode 3 is not a grafting error. It records a reviewer informativeness
policy — generic CC terms (cytosol, cytoplasm, membrane), uninformative
binding terms and out-of-context process terms are down-graded because the
review standard prefers specific, core terms, not because the graft landed in
the wrong place or the term is false. 33 of the 70 are cellular-component rows
assigned by aspect alone (34 CC rows in all); the rejection re-review already
moved most redundant cytosol/cytoplasm rows to KEEP_AS_NON_CORE, which
is why mode 3 fell from 93 rows to 70. Mode 3 also now holds host-context
calls on the holozoan proteins (adherens-junction and junction-assembly
processes on choanoflagellate cadherins, regulation of organ growth on
S. rosetta warts, cell fate specification on Capsaspora CoBra,
mannan biosynthesis on couscous). A different review policy would move many
of the CC rows into retained; read the mode-1/2/4 counts (180 annotations,
64% of the down-grades) as the grafting-attributable part.

The five rotary-ATPase rows (fliI ×3, sctN ×2, GO:0046933) are filed as
mode 1 by the operational rule on the sub-page (correct export-ATPase
subfamily, wrong term from a node above it); the
case study traces that term to a PAINT
IBD placed on a duplication node.

Protein counts are distinct review files, not gene symbols — 577 files
carry only 557 symbols (mdh, ALB, dapF and others span more than one
file), so a symbol-keyed count reads mode 4 as 28 (the two mdh proteins
merge) and mode 1 as 104.

The TreeGrafter terms most often down-graded cluster in two patterns:

1. Over-specific catalytic activity propagated to the wrong paralog/subfamily.
   The grafting node carries a precise enzymatic MF that the query has diverged
   away from: proton-transporting ATP synthase activity, rotational mechanism
   (GO:0046933, 5; the rotary-ATPase case),
   fatty acid synthase activity (GO:0004312, 4),
   phosphotransferase activity, phosphate group as acceptor (GO:0016776, 4),
   carotenoid dioxygenase activity (GO:0010436, 4),
   NADH dehydrogenase activity (GO:0003954, 3; it was 6 before the
   rejection re-review kept three of the complex I rows as non-core),
   spermidine synthase activity (GO:0004766, 3),
   (S)-2-hydroxyglutarate dehydrogenase activity (GO:0047545, 3).
   TreeGrafter places a sequence on a tree node but cannot tell that the
   catalytic residues, or the whole substrate specificity, have changed.
   (triacylglycerol lipase activity on fogD was re-reviewed to
   UNDECIDED on 2026-09-20 — the re-review found that the family source Ayr1
   does have lipase activity — and the later follow-up review marked it
   MARK_AS_OVER_ANNOTATED: the source activity is real, but FogD sits in a
   distinct subfamily with no hydrolase assay.)

2. Generic / uninformative localization. Before the rejection re-review,
   cytosol (GO:0005829, 18 rows) and cytoplasm (GO:0005737, 9) were the
   two most frequently down-graded TreeGrafter terms. The re-review judged
   most of them true but redundant and moved them to KEEP_AS_NON_CORE, so
   only 2 of each remain down-graded; the generic CC terms now most often
   down-graded are membrane (GO:0016020, 4), plasma membrane (3) and
   nucleus (3). The MF analogue, identical protein binding (GO:0042802),
   is down to 1 row.

3. Animal-only terms on unicellular holozoans. Each of the three
   S. rosetta cadherins carries the same ten down-graded rows from one
   graft node (PTN000616280, PTHR24027:SF422), so adherens junction,
   catenin complex, beta-catenin binding, cell-cell junction assembly,
   cell morphogenesis, calcium-dependent cell-cell adhesion and others
   each appear three times (see the
   case study).

There are also process-level mis-propagations of two kinds: a mechanistically
wrong sibling process (lipopolysaccharide core region biosynthetic process
on the mcr phosphoethanolamine transferases, which modify lipid A, not the
core oligosaccharide), and pathway terms propagated onto hosts that lack the
pathway (sucrose biosynthetic process on P. putida fbp).

Family hotspots (upstream targets)

treegrafter_family_hotspots.tsv
aggregates the reviewer outcome over all 1,074 TreeGrafter annotations per
PANTHER family, subfamily and graft node. Of the 77 families with at least
four reviewed annotations, 26 have half or more of their propagated terms
down-graded and 27 have none — the errors are concentrated, not diffuse.
Families with ≥60% down-graded (modes counted from
treegrafter_failure_modes.tsv):

Relative to the 2026-09-06 snapshot, three families are no longer
hotspots: PTHR48078 (ilvA-I/ilvA-II, 6/6 → 0/6 down-graded) after the
2026-09-20 re-review; PTHR11556 (FBPase, 7/11 → 1/11) and PTHR11632 (SDH
flavoprotein, 71% → 3/7 = 43%, aprA electron-transfer and plasma-membrane
rows now ACCEPT). PTHR44169 (fogD) went 6/6 → 0/6 at the 2026-09-20
re-review and is back at 4/6 after the follow-up review. Since the
2026-10-01 refresh, the rejection re-review took PTHR11271 (guaD/hutF,
3/5 → 1/5) and PTHR30435 (flgE/flgG, 3/4 → 1/4) off the table and moved
PTHR42995 from 4/4 to 3/4 (accD GO:0009329 restored); PTHR21272
(aroQ-III, aroQ1, aroQ2, 3/3) drops below the four-annotation
threshold now that the duplicate aroQ folder is merged. PTHR24027 and
PTHR24356 (holozoan rows) and PTHR43137 (pyrC) are new to the table.

Re-review of 2026-09-20

TREEGRAFTER/rereview-2026-09-20/
holds the audit records of a full-gene re-review of 29 genes carrying
TreeGrafter rows (28 of them in the current TreeGrafter set), done by agent
reviewers (three of the seven records name Codex; the others name no
reviewer), which re-examined every annotation on each gene (not only the
GO_REF:0000118 rows) and wrote the new actions into the gene review YAMLs:

In total 19 genes are reviewed and 10 are awaiting_adjudication in
these records. The awaiting genes' new actions are already in the YAMLs and
therefore in every table above: of the 37 rows on these genes that left the
down-graded set since the 2026-09-06 snapshot, 13 (5 now UNDECIDED, 5
ACCEPT, 3 KEEP_AS_NON_CORE) are on awaiting genes, so those rows are
provisional until a curator adjudicates. Five of the awaiting genes
(fogD, K9IJK6, K9IMD0, K9IFT7, Cgas), and IRE1, have since had
follow-up reviews of focused OpenScientist reports, which changed 16
TreeGrafter actions (see the 2026-10-01 note); the re-review records
themselves were not updated. The
cache-reports/ subfolder preserves six OpenScientist reports reused as
evidence in that pass (see its README; they are evidence inputs, not accepted
conclusions).

Reference standard

• What the reference is. Every rate on this page measures agreement with
  the AIGR agent review of each gene, written under the repository's curation
  guidelines and fixed at the pinned commit (f81b9f300) before scoring.
  There is no separately human-rated subset, so agreement with expert
  curators is not quoted.
• Label visibility. Reviewers saw original_reference_id (and hence
  whether a row was TreeGrafter, corroborated or InterPro2GO) while deciding.
  This affects the corroboration contrast (see the caveat there) and possibly
  the TreeGrafter-vs-IBA contrast.
• Agent reports among the cited evidence. 200 of the 1,074 reviewed
  TreeGrafter annotations (61 of the 283 down-graded) cite an OpenScientist or
  Falcon file (blinded hypothesis report or Falcon deep research) in their
  review.supported_by, alongside the literature
  (exemplar_independence.py).
• The ten blinded exemplars. The ten OpenScientist/Falcon exemplars on the
  failure-modes page
  were drawn only from rows that were already down-graded, with no
  accepted-annotation controls, and n = 10, so no claim about the check's
  reliability (sensitivity, specificity or false-positive rate) can be made.
  Their blinded verdicts can be compared fairly with the review actions
  recorded when each exemplar was selected (aprA REMOVE, fcs MODIFY,
  OCTS1 MARK_AS_OVER_ANNOTATED, eryAIII MODIFY, mcr-1 MODIFY,
  NaPMT3 REMOVE, ADAR2 REMOVE, NaUGT1 REMOVE, aceK MODIFY,
  ahpC MODIFY), but not with the current actions: some of those reviews
  were later revised with the reports in hand (aceK → ACCEPT, NaUGT1 →
  UNDECIDED), and 9 of the 10 exemplar annotations now cite the report
  (exemplar_independence.tsv).
• Drift. The reference is revised over time (131 TreeGrafter actions
  changed since the 2026-09-06 snapshot, 75 of them since the 2026-10-01
  refresh), which is why results are pinned to
  a commit.

Caveats

• Corpus composition. 65% of the TreeGrafter rows (703 of 1,074, under two
  taxon labels for KT2440) come from Pseudomonas putida KT2440, so the rates
  are largely a P. putida result. The 95% Wald interval on the 43.2% accept
  rate is roughly ±3 pp, but the taxon skew and reference drift matter more
  than the sampling error. Directional, not a frozen benchmark.
• The reference standard is the AIGR agent review corpus at the pinned
  commit, which is revised over time (see
  Reference standard).
• KEEP_AS_NON_CORE is not an error — the inference is correct but
  peripheral. Accept + non-core is 69.0% for TreeGrafter vs 88.9% for
  corpus-wide IBA; adding MARK_AS_OVER_ANNOTATED gives the retained rates
  in the headline table.
• 50 TreeGrafter rows (4.7%) are UNDECIDED, 8 of them on genes still
  awaiting_adjudication in the 2026-09-20 re-review records.

Rejection re-review (2026-09-24)

The rejection rate is only meaningful if the rejections themselves hold up, so
every TreeGrafter row then marked REMOVE or MARK_AS_OVER_ANNOTATED was
re-examined against a fixed rule set: REMOVE requires positive contrary
evidence, broad-but-true terms are not over-annotations, a too-coarse ancestor
is MODIFY rather than REMOVE, and every term definition is checked in
QuickGO. A term that is true but is a redundant ancestor of a carried term, or
a redundant generic compartment, is KEEP_AS_NON_CORE rather than ACCEPT:
the enum reserves ACCEPT for terms "representing the core function of the
gene", so restoring a bare ancestor as core would assert something the evidence
does not. Rules, per-batch records and a generated summary are in
TREEGRAFTER/rereview-2026-09-24/.
The nine genes already re-audited on 2026-09-20 were left as recorded there.

192 rejected rows across 166 gene entries — 165 proteins, since one protein
had two review folders that this audit merged — of which 117 (61%) stand and
75 (39%) were relaxed —
56 to KEEP_AS_NON_CORE, 7 REMOVE → MARK_AS_OVER_ANNOTATED, 5 to ACCEPT,
4 to MODIFY, 3 REMOVE → UNDECIDED. Only five rows are restored as core
functions, each one a term the protein itself performs: the MurJ flippase
reaction (GO:0034204), the AccA/AccD carboxyltransferase complex
(GO:0009329), bis-MGD biosynthesis on MobA (GO:1902758), and the two
lymphotoxin-alpha signalling processes on K9IWR0.

This is the largest change to the down-graded population since the
2026-09-06 snapshot, and since the 2026-10-03 refresh it is in every table on
this page and in the failure-mode classification. Every audited row that is
still live has the same action in treegrafter_review.tsv as in its batch
record (190 of 192; the other two are the retired g022 row and the row on
the duplicate aroQ folder that was merged into aroQ-III), and 126 of them
remain down-graded. Against the 2026-10-01 tables it moved 66 rows out of the
down-graded set and none into it. The relaxations say more about the
first-pass reviews than about TreeGrafter, and they cut unevenly across the
four failure modes:

• Redundant locations (34 rows). cytosol/cytoplasm on soluble bacterial
  enzymes had been down-graded purely because a sibling location row existed. A
  broad true term is not an error, but a redundant compartment is not a core
  function either, so these are KEEP_AS_NON_CORE. Mode 3 (generic/context)
  is therefore the mode that shrinks most: 40 of the 66 rows that left the
  down-graded set at the 2026-10-03 refresh were mode 3, and it went from 93
  rows (31%) to 70 (25%) even after the holozoan host-context rows were
  added. It is a real failure
  only where the location is incompatible with the protein — secreted
  cystatin cpi-2, exported flagellar hook flgE, periplasmic alr — and
  those rejections stand.
• True ancestors of a carried specific term (~20 rows). oxidoreductase activity on betA, glycosyltransferase activity on murG, protein transport on secD/secF, deaminase activity on guaD; each verified by
  QuickGO ancestry and kept as non-core. The six complex I subunits carrying
  NADH dehydrogenase activity are a granularity case the classification
  does not separate out — a whole-complex activity on a single subunit that
  lacks the NADH site; the three still down-graded (nuoE/nuoH/nuoI) are
  curated as mode 3. On nuoE/nuoH/nuoI, where GOA carries
  no GO:0008137 row, that becomes MODIFY → GO:0008137 with
  contributes_to; on nuoG/nuoL/nuoM, which already carry a separate
  GO:0008137 row taking the same qualifier correction, proposing it again
  would be redundancy, so the ancestor is kept as non-core instead.
• Rejections resting only on "no target-specific assay" (~12 rows). Absence
  of a target experiment does not refute a supported phylogenetic inference;
  these became MARK_AS_OVER_ANNOTATED, or UNDECIDED in the three cases where
  the substrate itself is genuinely unknown (ptxD ×2, retS).
• Label read instead of definition (1 row, but instructive). GO:0009329 "acetate CoA-transferase complex" on PSEPK accD was removed as a different
  enzyme; the term's definition is the AccA/AccD carboxyltransferase
  component of acetyl-CoA carboxylase, so the propagation was more specific
  than the term the gene already carried. The term's label and its capable_of
  axiom to GO:0008775 contradict its own definition — worth raising with GO.
  Confirmed against the live QuickGO ontology API on 2026-10-02 rather than only
  against the cached label: the definition returns verbatim (xrefs
  PMID:2719476, PMID:8423010), GO:0009317 is a current ancestry relation,
  GO:0032283 (the plastid accD subcomplex) is still the sole child, and the
  capable_of GO:0008775 relation is present with a 2015-06-18 addition date.
  The term is not obsolete and its aspect is cellular_component, so the
  mismatch really is label-and-axiom versus definition.

What survives scrutiny almost intact is mode 4, the paralog /
wrong-subfamily catalytic transfer (only the two ptxD rows left it, as
UNDECIDED): spermidine synthase on the PMT methyltransferases,
carotenoid dioxygenase on the lignostilbene dioxygenases, LDH on malate
dehydrogenases, cysteine synthase vs. O-acetylhomoserine sulfhydrylase. The
hotspot list is built on exactly these, so it is the part of the analysis the
re-review leaves standing — and the right input to the upstream tickets.

About a third of the 117 retained rejections had their reason strengthened
with the specific evidence (EC numbers, PANTHER subfamily vs. graft node,
cached substrate-panel quotes) rather than left on family-level doubt.

Deeper analyses

• Failure Modes & Tree Placement — joins all
  283 down-graded annotations to their PANTHER graft point, plus a lightweight
  PANTHER-vs-InterPro graft check on ten
  exemplars. Key result: the placement is usually sound; the error is the GO term
  attached to the graft node, and InterPro sometimes resolves the protein better
  than PANTHER (e.g. IPR011803 AprA, IPR003083 S-crystallin).
• ATP-synthase terms on flagellar/T3SS export ATPases — now its own
  project, Conserved-function recovery, where
  it is the first case. A case of inherited PAINT over-placement. The GO:0046933 /
  GO:0045259 IBD sits on the duplication node PTN008558586 in PTHR15184, above
  both the F1-β and the FliI/SctN clades. PAINT IBA and TreeGrafter then label
  nearly every FliI/SctN protein an ATP synthase, and InterPro2GO (IPR013380,
  IPR004100) and GO_REF:0000108 add further wrong rows. Nine full reviews
  (FliI in Caulobacter, H. pylori, P. putida, E. coli and Salmonella;
  SctN in Salmonella ×2, Yersinia and Shigella) remove 31 of the 35 affected
  rows and mark the other 4 as over-annotations. The five TreeGrafter rows of
  this case (fliI ×3, sctN ×2) are in the tables, all REMOVE.
• Unicellular holozoans: Hippo pathway case study
  — TreeGrafter on choanoflagellate, Capsaspora and sponge proteins, which
  have no IBA rows because none of them is a PANTHER reference genome. The
  commonest failure is grafting onto animal-only nodes:
• choanoflagellate cadherins onto a Bilateria node;
• Capsaspora integrin betas onto the vertebrate ITGBL1 node;
• Capsaspora T-box factors onto an all-animal node carrying "cell fate
  specification".

There are also two cross-family mis-placements: Capsaspora Warts with the
citron/ROCK kinases, and the S. rosetta yorkie candidate with the MAGI
family. One correct graft still inherits animal-tissue IBDs from LATS node
PTN002390470, including regulation of organ growth on a unicellular
organism. Since the 2026-10-03 refresh its 65 TreeGrafter rows
(S. rosetta 41, Capsaspora 19, Oscarella 5; 41 down-graded) are in
every table on this page. The row-level classifier files the down-graded
ones as mode 3 (16, host context and CC), mode 0 (17: the cadherin
over-annotations the reviews call untested rather than impossible, and two
warts rows), mode 1 (5) and mode 4 (3: coWts ×2, coITGB2 cell-cell
adhesion). The case study reads the cadherin set as a whole as a
taxon-blind graft ("4-like"); row by row, only host-context reasons are
stated explicitly enough to classify.
- Corpus-wide provenance (side query, 2026-10-11). A GO-DB analysis of
PomBase's contribution to TreeGrafter
(brief; the build
scripts are not in this repo) walks every TreeGrafter graft node in
goa_uniprot_all up the PANTHER 19 trees to the nearest same-term IBD. Of
166.8M TreeGrafter annotations, 90.1% trace to a same-term IBA on an
ancestor, usually several steps above the graft: only 4.9M come from an IBD
on the graft node itself. The other 9.9% have no same-term IBA on their
path. For 3.0M of these an ancestor now carries an IBA to a more specific
term, consistent with PAINT refining the IBD after the TreeGrafter run
(dated 2026-07-27) while TreeGrafter still emits the coarser parent. Example:
PTHR28283 root PTN002001416 now has GO:0110034, but 2,350 TreeGrafter
rows still carry the parent GO:1902660 (negative regulation of glucose
mediated signaling pathway), including on bacterial proteins. This
run-lag is a second source of the "too coarse" and out-of-context inherited
terms counted above, separate from IBD placement. It is not measured in
this corpus.
- OpenScientist blinded verification uses a dedicated TreeGrafter prompt
template,
templates/treegrafter_function_hypothesis.md,
which frames each propagated term as a blinded "GENE has " hypothesis
and instructs the agent to actively test the failure modes
(granularity, pseudo-enzyme, mis-placement). Run via
gene-hypothesis-research openscientist <org> <gene> --annotation-term-id <GO> --as-function-hypothesis --template templates/treegrafter_function_hypothesis.md.
See Reference standard for how the ten existing
runs can be compared with the reviews (against the actions recorded at
selection, not the current ones).

What TreeGrafter gets right

Added 2026-10-14.
tg_versus_other_sources.py sets
each reviewed TreeGrafter row beside the other annotations on the same
protein, related by GO is_a/part_of (same term, ancestor, descendant, or a
sibling sharing a direct parent), and records which side the review kept
(tg_versus_other_sources.tsv,
counts in
tg_versus_other_sources_summary.tsv).
"Kept" is ACCEPT or KEEP_AS_NON_CORE; "down-graded" is REMOVE,
MODIFY, MARK_AS_OVER_ANNOTATED or TOO_GENERAL. Recomputed 2026-10-16
from the live reviews, after TOO_GENERAL was first counted as down-graded.

The only source, and kept. 623 rated TreeGrafter rows have no other
annotation on the protein, from any source, to the same term, an ancestor or
a descendant. The reviews kept 439 (70.5%): 232 of 262 CC (89%), 70 of 118 MF
(59%), 137 of 243 BP (56%). TreeGrafter overall was kept 72% of the time in
the pinned f81b9f300 snapshot (this sole-source figure is from the live
reviews, so the two are close rather than a measured gap): the statements it
alone makes are not worse than the rest. Several carry the
protein's defining activity: squid ADAR2 (double-stranded RNA adenosine
deaminase activity), naked mole-rat Cgas (2′,3′-cyclic GMP-AMP synthase
activity), P. putida relA (GTP diphosphokinase activity), hummingbird
TAS1R3 (sweet taste receptor activity), Capsaspora coYki
(transcription coactivator activity), P. putida tig (protein folding
chaperone).

More specific than another pipeline. 159 TreeGrafter rows are
descendants of an InterPro2GO term on the same protein (79 of them kept), 50
of an ARBA term (27 kept), 14 of a UniRule term (11 kept) and 141 of a
combined-IEA term (123 kept). Where the review kept one side of a related
pair and down-graded the other (distinct TreeGrafter rows):

Examples where TreeGrafter is the side kept:

• Deeper than InterPro2GO. Vampire bat K9J287 deoxyribonuclease I
  activity (InterPro2GO: catalytic activity, DNA nuclease activity; both
  MODIFY); naked mole-rat Scn9a sodium ion transmembrane transport
  (InterPro2GO: monoatomic ion transport, sodium ion transport, transmembrane
  transport); P. putida PP_1084 thioredoxin peroxidase activity
  (antioxidant activity, oxidoreductase activity), hpd/hmgB L-tyrosine
  catabolic process (aromatic amino acid metabolic process), glpF glycerol
  transmembrane transport.
• A sibling InterPro2GO term the review rejected. ADAR2
  double-stranded RNA adenosine deaminase activity (InterPro2GO adenosine
  deaminase activity, MODIFY); pdxH pyridoxal 5′-phosphate biosynthetic
  process (InterPro2GO pyridoxine biosynthetic process, REMOVE); paaJ
  phenylacetate catabolic process (InterPro2GO 3,4-dihydroxybenzoate
  catabolic process, over-annotated).
• Against ARBA. PP_2320 peptidoglycan L,D-transpeptidase activity
  ACCEPT against ARBA carboxypeptidase activity REMOVE (a GO ancestor, but
  the wrong reaction: "The closely matched ErfK/LdtPae2 subfamily resolves the
  YkuD reaction as lipoprotein-acceptor L,D-transpeptidation"); squid AP180
  vesicle budding from membrane kept against ARBA receptor-mediated
  endocytosis over-annotated; dnaQ 3′-5′ exonuclease, visC
  2-polyprenylphenol 6-hydroxylase and dinB error-prone translesion
  synthesis where ARBA has only exonuclease, monooxygenase or DNA metabolic
  process. True TreeGrafter-versus-ARBA sibling conflicts are rare here:
  three pairs across the 47 proteins that carry both.

The reverse cases exist and are listed in the same file: 35 TreeGrafter rows
down-graded where InterPro2GO's related term was kept. Most are a specific
TreeGrafter term that is wrong under a coarse InterPro2GO parent that is
still true (mdh, nspC, eryAI–III, the stilbene dioxygenases, the
PMTs, the choanoflagellate cadherins). The combined-IEA call won 13
sibling conflicts against TreeGrafter.

Against ProtNLM2

The ProtNLM2 evaluation and this project barely
overlap, for a structural reason. TreeGrafter annotates only proteins outside
PANTHER's reference genomes; reference-genome proteins get PAINT IBA rows
instead. Most ProtNLM cohorts are reference genomes: the 44 reviewed horse
genes carry 39 IBA rows and no TreeGrafter row, and the same holds for the
fly, fission yeast and Neurospora cohorts. Of the 162 ProtNLM
prediction-review files at the time, 79 were on proteins with IBA rows and 10
on proteins with reviewed TreeGrafter rows (one has both). There are now 202,
the 40 DONKEY40 files carrying TreeGrafter rather than IBA rows; the same 10
are in the TreeGrafter snapshot.
tg_versus_protnlm.py prints these counts
from each protein's GOA file and pairs the two sources on those 10 (tg_versus_protnlm.tsv). The
scales differ: TreeGrafter rows carry a gene-review action, ProtNLM
predictions a VDCL assessment from a prediction review.

• TreeGrafter says more, and most of it is kept. 29 TreeGrafter rows
  (22 kept, 4 down-graded, 3 undecided) against 12 ProtNLM GO predictions
  (5 COR, 3 CNN, 3 LSP, 1 NPI). Two of the ten proteins (CDK16 in blue
  whale, a MarR regulator in Chromobacterium) have no ProtNLM GO output at
  all.
• Where both speak, TreeGrafter is as specific or more. Four pairs relate:
  two share cytoplasm; on dog CNNM4 TreeGrafter's magnesium ion transport
  (ACCEPT) sits under ProtNLM's monoatomic ion transport (LSP); on pufferfish
  gas7a TreeGrafter adds six more (clathrin-coated pit and vesicle, plasma
  membrane, clathrin coat assembly, clathrin-dependent endocytosis, neuron
  projection morphogenesis; five accepted, one kept as non-core) where ProtNLM
  gives only cytoplasm.
• Both fall for the same pseudophosphatase. On mussel MTMR9, ProtNLM's
  phosphatidylinositol-3-phosphate phosphatase activity is NPI and
  TreeGrafter's phosphatidylinositol dephosphorylation is MODIFY;
  TreeGrafter's protein phosphatase binding, the regulatory-partner role, is
  accepted.
• ProtNLM's correct new terms are complementary. Its five COR calls are
  terms TreeGrafter does not give at all: extracellular matrix and its
  organization on primate OLFML2A (where TreeGrafter's only term, signal
  transduction, is over-annotated), L-ascorbic acid binding on quail PAM,
  autophagosome on cod tbc1d14, and determination of heart left/right
  asymmetry on macaque TTC39C.
• The LSP and CNN judgements are made against the protein's GOA record, which
  includes the TreeGrafter rows, so ProtNLM is in part being scored against
  TreeGrafter here.

Ten proteins is too few for rates. Running ProtNLM2 on the 577 reviewed
proteins is not an option: its output exists only for its published accession
list, which includes 15 of them. So we ran a larger test on a cohort chosen
from that list: the DONKEY40 cohort. It has 40 donkey proteins
with both ProtNLM2 GO output and TreeGrafter rows, each paired with a horse
ortholog carrying PAINT IBA rows, reviewed 2026-10-14 and re-reviewed
2026-10-23.

• TreeGrafter on a mammal reproduces IBA. 199 of the 225 TreeGrafter rows
  (GO_REF:0000118 plus merged GO_REF:0000120 rows with a PTN node) carry
  the same term as the horse ortholog's IBA, and no review attributes a
  down-grade to graft placement.
• A correct graft does not make the rows much more reliable. The donkey
  reviews down-graded 17.3% of the rows by this page's definition (39 rows,
  REMOVE + MODIFY + MARK_AS_OVER_ANNOTATED) and 22.7% counting the 12
  TOO_GENERAL rows, against 26.9% for this corpus, whose snapshot has no
  TOO_GENERAL rows. The gap is modest, rests on 40 genes (PARD3B and PIKFYVE
  contribute six down-grades each), and mixes taxon with review vintage.
• Where donkey rows were down-graded or left UNDECIDED (64 rows), the cause
  was not graft placement.
• The input sequence (16 rows on three proteins): a PIKFYVE fragment with
  no kinase domain, a TNFRSF1A model missing part of its ectodomain, and a
  truncated SNAP23. TreeGrafter does not check that the protein keeps the
  domain the function needs.
• PAINT terms that over-reach (13 rows, the same terms horse receives as
  IBA), such as the SNAP25 synaptic terms on SNAP23.
• PAINT terms that are right but imprecise (14 MODIFY rows, again the
  same terms as horse IBA), twelve of them refined to a descendant.
• True but too general (12 TOO_GENERAL rows).
• Six rows whose term horse IBA does not carry. One, ATG3 GO:0019776,
  is a release lag: horse IBA already carries its replacement,
  GO:0141046.
• Three rows that no available evidence decides (FLCN, PARD3B, RBM41).
• ProtNLM2 on the same proteins (69 GO predictions) is judged 68% correct
  and 13% wrong. The caveat above applies with more force: 36 of the 47
  correct calls are CNN or LSP, judged against a GOA record that contains the
  TreeGrafter rows, and the same agentic process wrote both the prediction
  reviews and the gene reviews.
• Where it overlaps the tree-based terms, it is mostly an ancestor of them:
  28 of 38 related predictions, 24 of them correct.
• All 11 of its COR calls are terms neither TreeGrafter nor IBA gives.

Propagation rules and graft paths

Added 2026-10-14. Three questions about the 1,074 reviewed rows: can a
mechanical check, run without reading the literature, pick out the rows
reviewers went on to down-grade; which PAINT node does each reviewed term come
from; and how far do the clearest errors reach beyond the corpus?

Every TreeGrafter row is a TreeGrafter-only call. GOA merges identical
calls from several electronic pipelines into GO_REF:0000120, so a
GO_REF:0000118 row is one no other pipeline made. None of the 1,074 shares
its GO term with another electronic annotation on the same protein (9 share
it with an IBA, 10 with an experimental or ISS row). This is why the
corroboration contrast above exists at all, and it means the EC-match control
below is empty by construction.

Rules scored against the reviews

propagation_rules.py scores nine
rules from UniProt records, GOA rows and GO (go-basic.obo, release
2026-07-26) against the actions in
treegrafter_review.tsv. The base
rate is 27.6% down-graded over 1,024 rated rows (UNDECIDED excluded).
Results in propagation_rules.tsv,
every hit in propagation_rule_hits.tsv.

Two more rules flag too few rows to read: a family or subfamily name
containing INACTIVE (3 rows, all eryA, from the PTHR43775:SF51 name rather than
a pseudoenzyme) and animal tissue vocabulary on a non-metazoan protein (1 row,
S. rosetta warts regulation of organ growth); the eukaryote-only-CC on a
prokaryote rule flags nothing.

The rules are scored against actions written by the same agentic review
process, so a high down-grade rate partly measures agreement with the reviews'
own reasoning rather than independent ground truth. The strongest hits do not
simply restate the rule: purM GO:0004637, for example, was argued from
domain architecture (PF00586/PF02769, IPR004733), not from EC numbers. The
independent test is running the EC rule over all of GOA (Next steps) and
checking a sample with a human curator.

• EC contradiction is the one strong signal. The rule compares the EC
  cross-references of the GO term with the EC numbers in the protein's
  UniProt DE lines, and fires only when every pair differs at a level both
  numbers specify, so a coarse term (oxidoreductase activity, 1.-.-.-) is
  compatible with a specific EC rather than contradicted. A spot check of the
  hits that were not down-graded found three causes. One was a rule bug, now
  fixed: GO's is_a can cross EC branches (GO:0008834, EC 2.5.1.31, is_a
  GO:0045547, EC 2.5.1.87), so a term that subsumes another GOA term whose
  EC does match is now exempt (uppS). The other two are genuine false
  positives with a stated reason: fadB (an EC split by chain length that the
  reviewer accepted) and metZ (a lyase versus transferase EC class for the
  same PLP chemistry); ilvA-I/ilvA-II (dual threonine/serine substrate) are
  kept as non-core. Not one reviewed TreeGrafter MF term with an EC
  cross-reference agrees with its protein's UniProt EC (0 of 82 across all
  repo GOA files), which fits GOA's merging: an agreeing call would also have
  been made by EC2GO and so would sit under GO_REF:0000120.
• Broad compartments and true-but-coarse parents are kept. The two rules
  that might have been expected to flag over-annotation do the opposite:
  reviewers retain such rows, usually as KEEP_AS_NON_CORE.
• The rest are weak. Sibling-term and name-overlap rules roughly double
  the base rate on small numbers; the paralog rule mostly reflects the
  paralog-rich P. putida batch.

How far the EC contradictions reach

ec_contradiction_scale.py takes
each (UniProt EC, GO term) pair behind a corpus EC contradiction, counts the
UniProtKB entries that have that EC and that GO term but not the term's own
EC (which removes fused or multifunctional proteins for which the term may be
right), and samples 40 of them in QuickGO to see where the term comes from
(ec_contradiction_scale.tsv,
44 pairs). Restricted to the 36 pairs from single-EC proteins whose corpus
rows were all down-graded, they add up to about 150,000 protein–term pairs,
and in almost every sample all or nearly all of the annotations are
TreeGrafter's:

The samples are drawn from the first 500 hits UniProt returns, not uniformly;
the counts use UniProt's own go: search, which covers every evidence code.

Graft paths: which PAINT node the term came from

trace_graft_paths.py fetches the live
PANTHER 19 tree for a family, walks up from the graft node to the nearest
ancestor carrying an IBA for the same GO term, and summarises the clades under
that IBD (graft_paths.yaml for seven
examples). It reproduces the rotary-ATPase case independently: all five
fliI/sctN grafts trace to the duplication node PTN008558586, 3 to 7
steps up. The other examples:

• Fusion-seeded root IBDs (purM, moaA, moeB). The IBD sits on the
  family root (LUCA) and its only experimental seeds are multi-domain
  proteins: trifunctional GART from fly, mouse, rat and S. pombe for the
  PurD activity in PTHR10520; human and mouse MOCS1 (MoaA fused to MoaC) for
  cPMP synthase in PTHR22960; and yeast Uba4, which carries a rhodanese
  domain, for thiosulfate sulfurtransferase in PTHR10953. The activity
  belongs to a domain the standalone bacterial proteins lack. In PTHR10953
  the root spans the whole E1 superfamily, 1,090 reference leaves including
  ubiquitin, SUMO and NEDD8 activating enzymes and ATG7. trpC fits by name
  (fungal TRP1 fuses IGPS and PRAI) but was not traced.
• A root IBD across a duplication (eryAIII). The fatty acid synthase
  IBD is on the root of PTHR43775, seeded by animal FASN and a Dictyostelium
  protein; a bacterial duplication clade of type I PKS (36 leaves) inherits it.
• A mixed-function node (dapE). The acetylornithine deacetylase IBD sits
  on a bacterial speciation node in PTHR43808:SF31, seeded by E. coli ArgE
  and a Xanthomonas protein, whose other members include a DapE and several
  dipeptidases.
• No same-term IBD left (mdh). No ancestor of the graft node carries an
  IBA for L-lactate dehydrogenase activity in the current tree, consistent
  with a PAINT change since the TreeGrafter run (the same lag the PomBase
  side query above found at scale).

All 1,074 rows, offline. trace_graft_paths.py --all repeats the trace
for every reviewed row against the PANTHER 19 trees (Tree_MSF/*.tree from
PANTHER's TreeGrafter data download), PAINT's current IBD.gaf (PANTHER
v19.0, created 2026-08-28) and PAINT_TreeGrafter_Annotations_TOTAL.txt.gz
(the AN-to-PTN map, and the GO set TreeGrafter assigns at each node today);
the PANTHER API times out on too many families for a full run. It reproduces
the API results for the examples above. Each row's query organism is checked
against the IBD node's clade using the full NCBI lineage from the UniProt
taxonomy API (UniProt's OC lines omit ranks such as Bilateria and
Opisthokonta). Output in
graft_paths_all.tsv, scores in
graft_path_rules.tsv:

• 939 rows trace to a same-term IBD, a median of 4 steps above the graft node
  (maximum 22); 543 of those IBDs are on the family root and 494 paths cross a
  duplication. 129 rows (12%) have no same-term IBD left on the path: the term
  is no longer in TreeGrafter's current set for that node, so the GOA row
  predates a PAINT change. Reviewers down-graded 31% of these, close to the
  base rate. Two pyrC' rows (allantoinase activity, purine nucleobase
  catabolic process) now meet a NOT IBD on the path; both were removed. Four
  rows have no graft node in GOA.

• Taxon scope is the structural signal. The outside-clade hits are the
  choanoflagellate cadherins on a Bilateria node, the Capsaspora coITGB2
  integrin and S. rosetta yorkie on Eumetazoa nodes, the N. attenuata
  putrescine N-methyltransferases under a bacterial spermidine-synthase IBD,
  P. putida pyrE under a eukaryotic IBD and davD/gabD-II under a
  Terrabacteria IBD. Accepted exceptions include organelle-derived proteins
  that sit in bacterial clades by descent (octopus DDO). The 0-step rule
  overlaps it heavily (18 of its 28 rows are the cadherins) and is not an
  independent signal.
• The inside/outside split covers 840 of the 903 rated rows that reach a
  same-term IBD. The other 63 are undecidable: 62 have no resolvable clade
  name anywhere up the tree and one has no taxon id (11 of the 63 were
  down-graded). And 406 of the 795 "inside" rows sit under a LUCA IBD, which
  every organism is inside by definition. Those 406 and the other 389 are
  down-graded at the same rate (25.4% each), so the inside bucket is a
  baseline, not a test; the rule's content is the 45 outside rows.
• Depth and root placement are not. Root and non-root IBDs, and short
  and long paths, are down-graded at about the base rate.
• Combined. EC contradiction, outside-clade and a NOT IBD on the path
  together flag 96 rated rows, 81% down-graded, catching 28% of all
  down-grades; only 3 rows trip both of the first two.

NOTES

Controlled OpenScientist batch (2026-10-15)

Sixteen blinded hypotheses ("GENE has TERM") were pre-registered in
openscientist-controls/selection.yaml.
The review's action on each row was frozen at 1b7f8f330, before any report
was read. All 16 reports are written. Five runs timed out at OpenScientist's
7,200 s ceiling with 3 iterations and were retried once with 2 (PP_2482, relA,
PP_4594, metZ, pdxH), which departs from the pre-registered protocol.
Verdicts and their scoring:
openscientist-controls/results.yaml.

• Accepted rows mostly hold up. Of the 9 accepted rows, the run supports
• ADAR2, Cgas and relA, controls whose defining activity comes only from
  TreeGrafter, are all supported.
• TAS1R3 is partially supported, and the disagreement is about the GO
  term, not the protein. The run argues that sweet taste is a property of
  the receptor pair, not of the T1R3 subunit, and that hummingbirds have no
  TAS1R2. But it also cites the finding that the hummingbird's T1R1–T1R3
  pair senses sugar (Baldwin et al. 2014, PMID:25146290). So the question
  is how to annotate a subunit with its complex's function.
• The EC-contradiction rule is 1 for 2.
• The fadB row, accepted although it trips the rule, is supported.
• The metZ row is refuted: carbon-sulfur lyase activity on an
  O-succinylhomoserine sulfhydrylase. The protein is 82% identical to the
  characterised P. aeruginosa MetZ, an EC 2.5.1.- transferase. GO has no
  term for that exact activity; the nearest, GO:0051009 O-acetylhomoserine
  sulfhydrylase activity, is a transferase too. So the review's reasoning
  that the lyase term is "a reasonable available parent" does not hold. The
  re-review below changed the row to REMOVE and added a term request.
• The outside-clade rule's test case holds. DDO, whose query lies outside
  the IBD node's clade, is supported.
• Both "TreeGrafter over the other pipeline" calls are supported: PP_2320
  over ARBA, and pdxH over InterPro2GO.
• UNDECIDED rows: the term is right in kind but wrong in depth.
• Both MoaA paralogs (PP_1969, PP_2482) are supported as GTP
  3′,8′-cyclases (a verdict the re-review below did not adopt).
• The other five fault the depth of the term, not the family placement:
  • eryCIV: too general; a dTDP-sugar aminotransferase.
  • PP_3571: too general. PANTHER's own subfamily call is SF31,
    N-acetyl-L-citrulline deacetylase; the genome's ArgE ortholog is
    PP_5186.
  • gabD-II: the NAD+ in the term is too specific; the cofactor
    pocket allows NAD(P)+.
  • ptxD: keep the term and add hydroxypyruvate reductase. The run
    also rejects the phosphite dehydrogenase reading of the gene symbol.
  • PP_4594: the sibling activity within its subfamily cannot be
    separated by sequence alone.
• None of the 16 runs found a pseudo-enzyme.

Re-review (2026-10-15). The eight rows in question were re-reviewed with
the reports as evidence, checking each decisive claim against repo data first:

• PP_1969, PP_2482: first moved to ACCEPT because all 17 MoaA ligand
  positions are conserved, then reverted after PR review asked why the two
  function-level MoaA models do not match them. They score 269 and 268 bits
  against NCBIfam TIGR02666 (trusted cutoff 284; every characterised MoaA
  checked scores 370 or more), and reach only level -1 of HAMAP MF_01225. The
  score is lost across both ends of the protein, not at ligand positions or
  through a missing segment, and PANTHER subfamily SF0 is treated as
  non-discriminative here as it is for PP_3571, gabD-II and PP_4594
  (PP_1969-bioinformatics/RESULTS.md). So the run's SUPPORTED verdict, which
  did not consider these models, is not adopted. PP_2482 sits next to a
  molybdoenzyme gene cluster, the main lead for a test.
• eryCIV: it is a DesI-type sugar 4-aminotransferase, and GO already has
  the specific term.
• PP_3571: the KT2440 ArgE ortholog is PP_5186, so a generic amidase
  term replaces the specific one. The run's N-acetyl-citrulline suggestion
  was not taken up, because 380 members of PANTHER subfamily SF31 are named
  succinyl-diaminopimelate desuccinylase.
• metZ: the reaction has no GO term, so it now has a structured term
  request: "O-succinylhomoserine sulfhydrylase activity", parent
  GO:0016765, skos:exactMatch RHEA:27826. The core function points to
  that request.
• PP_4594, ptxD, gabD-II: these stay open, because evidence the runs did
  not consider points elsewhere.
• PP_4594 is equidistant from the gamma-synthases and gamma-lyases.
• ptxD sits in a kgu gene cluster, which suggests a 2-keto-6-phosphogluconate
  reductase.
• gabD-II is as close to the ectoine-pathway DoeC as to GabD, and sits
  between doeD- and doeB-type genes. That last finding also adds a term
  request: "L-aspartate-4-semialdehyde dehydrogenase (non-phosphorylating)
  activity" (RHEA:45764).

The re-review also corrected several of the runs' claims:
- PtxD has no Trp314.
- Human CTH is in PANTHER subfamily SF15, not SF80.
- The PP_2482 UniProt record has no binding-site features.

Sixteen runs give no rates. What they show is how the faults split: where
TreeGrafter errs here, the protein is usually in the right family and the
term is the wrong depth. A wrong reaction class (metZ) is rare.

2026-10-14

• Brief rewritten (neutral title "Agentic review of the TreeGrafter
  pipeline"; numbers pinned to f81b9f300, matching this page) and linked
  from the manifest; it adds four tree case studies and the rule results.
• New: Propagation rules and graft paths
  with propagation_rules.py, ec_contradiction_scale.py and
  trace_graft_paths.py and their sidecars. GO release 2026-07-26 for the
  rules; PANTHER 19 trees with the 2026-08-28 PAINT IBDs for the trace.
• Checked the corpus against main at 75e426908: the same 1,074
  TreeGrafter rows on 577 proteins (6,994 review files). Since f81b9f300,
  six rows then UNDECIDED were adjudicated, all REMOVE (PP_3394 ×3,
  STABO mdr ×3), the IRE1 unfolded protein binding row was retired and
  a Pks1 secondary metabolite biosynthetic process row added, which moves
  the down-graded share from 26.4% to 26.9%. The headline tables stay pinned;
  the rule scripts read the committed treegrafter_review.tsv.
• The FliI/SctN case study (formerly a sub-page here, with scripts in
  TREEGRAFTER/rotary_atpase/) moved to a project of its own,
  Conserved-function recovery, since it came
  from an F-type ATP synthase completeness test rather than from the
  TreeGrafter tally.
• New: What TreeGrafter gets right, from
  tg_versus_other_sources.py (sole-source rows, and head-to-head outcomes
  against InterPro2GO, ARBA, UniRule and the combined IEA calls).
• New: "Against ProtNLM2" (tg_versus_protnlm.py): horse and the other
  reference-genome ProtNLM cohorts get IBA, not TreeGrafter; 10 proteins
  overlap.
• "Feed the hotspot list upstream" reworded: no PAINT/PANTHER tickets are to
  be filed yet; the fix points are collected under Next steps.

2026-10-03

• Refreshed at f81b9f300, the merge of main into this branch, which
  brought in the 2026-09-24 rejection re-review (PR #3165), main's GOA refresh
  that marked the viral TreeGrafter rows retired: true (PR #3503) and the
  holozoan case study. This is the refresh the 2026-09-28 note deferred.
  Re-ran analyze_treegrafter.py, analyze_placement.py,
  classify_failure_modes.py, snapshot_drift.py, exemplar_independence.py
  and graft_check.py --refresh-actions (offline) and rewrote every number on
  both pages from the sidecars: 998 → 1,074 annotations, 560 → 577 files,
  299 → 283 down-grades; accept 44.6% → 43.2%, retained 77.2% → 77.0%,
  REMOVE 11.2% → 11.1%, MARK_AS_OVER_ANNOTATED 11.5% → 8.0%,
  KEEP_AS_NON_CORE 21.0% → 25.8%, down-graded 30.0% → 26.4%, UNDECIDED
  44 → 50; MF down-graded 49.3% → 41.5%; corroborated accept 78% → 77%;
  hotspot families 27/72 → 26/77; mode shares 1/3/4/0/2 went 50/31/12/5/1% →
  47/25/15/12/2%.
• Retired rows. None of the scripts knew about retired: true, so the
  2026-10-01 sidecars still counted the three retired viral rows (g022
GO:0006302, dfrP GO:0046452/GO:0046655). analyze_treegrafter.py
  (annotations()), classify_failure_modes.py and
  exemplar_independence.py now skip retired: true existing annotations;
  analyze_placement.py and graft_check.py read the review sidecar and
  inherit the filter.
• Hotspot labels. analyze_placement.py labels a symbol shared by several
  review files as species/review-folder rather than species/symbol, so
  same-species paralogs (PSEPK dapF__Q88CF3 / dapF__Q88GD4) no longer
  collapse to one label. Its docstring no longer quotes stale corpus counts.
• failure_mode_curated.tsv: 19 current overrides whose rows are no longer
  down-graded (most of them relaxed by the rejection re-review, e.g. nuoG/
  nuoL/nuoM GO:0003954, K9IWR0 ×2, mobA, ptxD ×2, retS) are now
  superseded, each note carrying the current action (38 superseded rows in
  all). 22 rows were curated: 3 rows to mode 4 from the explicit
  substrate-mismatch rationales in the rejection re-review records (davA
  and acoA, previously mode 0; hutF, previously heuristic mode 1), 18 newly
  down-graded rows whose reviewer reason states the mode (14 holozoan rows:
  9 host-context mode 3, 3 mis-placement mode 4 for coWts ×2 and coITGB2
GO:0098609, 2 sibling or too-narrow mode 1 for TLN and coITGB2
GO:0007160; 4 pyrimidine-pathway granularity rows, broad parents of a
  retained de novo UMP term, on pyrC ×2, pyrD and pyrE), and
  flgE cytosol, whose strengthened reason now trips the pseudo-enzyme
  keywords (kept at mode 3). The remaining new down-grades are left to the
  heuristic or mode 0 (33 mode-0 rows now, 17 of them holozoan).
• Retitled "Rejection re-review (2026-09-24) — post-snapshot, not in the
  tables above" to "Rejection re-review (2026-09-24)": all 190 live audited
  rows are in the tables with the actions their batch records give. The
  holozoan rows (65, 41 down-graded) are also in the tables; the case study's
  callout says so. Marked the "Refresh the snapshot" next step done.

2026-10-01

• Refreshed at 943b98815. The 2026-09-27 numbers were computed at
  fff7793a6, which no longer exists on main after a rebase; main has
  since added reviews (including the rotary-ATPase case study) and follow-up
  reviews of OpenScientist reports on fogD, K9IJK6, K9IMD0, K9IFT7,
  Cgas and IRE1. Re-ran analyze_treegrafter.py, analyze_placement.py,
  classify_failure_modes.py, snapshot_drift.py, exemplar_independence.py
  and graft_check.py --refresh-actions (offline) and rewrote every number on
  both pages from the sidecars: 983 → 998 annotations, 550 → 560 files,
  279 → 299 down-grades, accept 45.1% → 44.6%, retained 76.2% → 77.2%,
  REMOVE 10.8% → 11.2%, down-graded 28.4% → 30.0%, UNDECIDED 59 → 44;
  MF down-graded 47.7% → 49.3%; hotspot families 24/71 → 27/72. The 15 added
  rows include the five TreeGrafter GO:0046933 rows of the rotary-ATPase case
  (four new, plus PSEPK fliI), now in every table.
• failure_mode_curated.tsv: no current override left the down-graded set.
  Six superseded rows re-entered it (fogD ×4, IRE1 GO:0070059,
  K9IJK6 GO:0014909); their new reasons were re-read and they are current
  again with the same mode. Seven newly down-graded rows were curated from
  the reviewer's reason: the four new rotary-ATPase rows and Cgas
  GO:0071360 (node/sibling term, mode 1), PP_2608 GO:0050661 (mode 1),
  K9IJK6 GO:0048008 (mode 3). The rest are left to the heuristic, and two
  PP_2608 rows stay mode 0 (16 mode-0 rows now). Superseded-row notes now
  carry the current action. Mode shares barely moved: 1 50%, 3 31%, 4 12%.

• Reframed the "Reference independence" subsection as Reference standard:
  the reference is the AIGR agent review fixed at the pinned commit, and the
  scores measure agreement with it. The exemplar caveat now says what can and
  cannot be compared: blinded verdicts against the actions recorded when the
  exemplars were selected (all ten down-graded), not against the current
  actions; with no accepted controls and n = 10, no reliability claim.

• Added the unicellular holozoan case study
  from the ORIGINS_OF_MULTICELLULARITY reviews. Across the 16 literature-based
  reviews, 11 of 53 propagated rows were down-graded, all GO_REF:0000118.

• Cross-family mis-placements. Warts went into PTHR22988, and the
  Yorkie candidate into PTHR10316.
• Grafts onto animal-only nodes, three times. These are choanoflagellate
  cadherins on a node PAINT records at Bilateria, Capsaspora integrin betas
  on the Euteleostomi ITGBL1 node, and Capsaspora T-box factors on an
  all-animal node carrying "cell fate specification".
• Working hypothesis. Reference proteomes escape this because their tree
  position, not the HMM call, sets their IBAs. Fly wts is in PTHR22988 by
  UniProt's classification but takes its IBAs from the LATS node.
• None of these rows are in the frozen 2026-09-06 tables.

• This complements the 2026-09-28 note below. Viral sequences outside the
  trees' taxonomic scope have stopped receiving TreeGrafter terms, but
  unicellular eukaryotes outside a node's PAINT taxon still receive them.

• Second review round on PR #3165. Withdrew two of the previous round's
  relaxations: PSEPK benB GO:0019380 (3-phenylpropionate catabolic
  process) and prpC GO:0005975 (carbohydrate metabolic process) had been
  moved REMOVE → MODIFY on the premise that the term was merely too coarse.
  It is not: GO:0019380 names a different substrate, and propanoate is an
  organic acid rather than a carbohydrate, so both are wrong-substrate or
  wrong-branch propagations — the class this audit retains elsewhere (quiA
GO:0008876). Both are REMOVE again, which also removes a duplication the
  review caught: each MODIFY proposed a term the same review already asserts
  on its own IC/NEW row (GO:0043639, GO:0019543). Split is now 117 stand
  / 75 relaxed. One further row, the g022 double-strand break repair call, was
  relaxed by this audit and then overtaken by main's finding that it is absent
  from GOA entirely; the batch record keeps the reasoning and records that there
  is no live annotation left to relax.

• Corrected a claim this branch had invented and then propagated. The
  2026-09-24 pass wrote that benB "still carries an IC annotation to the
  obsolete GO:0043640". It does not, and never did: GO:0043640 appears
  nowhere in benB-ai-review.yaml on this branch or on main, whose
  proposed_replacement_terms was already GO:0043639. GO:0043640 is
  obsolete (replaced_by GO:0043639, GO release 2026-07-26), which is how it
  reached the sibling benA/benC/benD reviews legitimately and this one by
  mistake. Removed from the review prose, the batch-03 record, this page's
  next-steps (where it would have sent a curator after a row that does not
  exist) and, per docs/history.md, corrected in place in the benB history
  record since the statement was never true.
• accD's core_functions.in_complex now names GO:0009329, the
  carboxyltransferase subcomplex the GO:0009329 row argues for, rather than
  the coarser GO:0009317 it previously kept alongside that argument.
• Smaller fixes: murB/ubiK prose no longer calls the broader cytoplasm row
  "accepted" while the narrower cytosol row is non-core (the underlying
  inversion is now named in the next-steps); a vestigial "Re-review." opener
  removed from pdxJ; summarize.py records how many terms its capped table
  leaves out, so summary.tsv is self-describing.

2026-09-29

• Review round on the rejection re-review (PR #3165). Corrected three
  mis-transcribed UniProt accessions in shipped reason text (guaD cited
  gshB's Q88D35 instead of Q88F18; tyrB Q88NM1 for Q88LG1; davA
  Q88RC2 for Q88QV2) and the same errors in three batch records — six sites,
  each re-checked against the gene's own AC line.
• Settled the ACCEPT / KEEP_AS_NON_CORE convention, which the first pass
  had applied inconsistently (some cytosol rows went each way). ACCEPT is
  defined as retaining a term as the gene's core function, so a redundant
  ancestor of a carried term, or a redundant generic compartment, is
  KEEP_AS_NON_CORE. 26 rows moved, leaving five ACCEPTs, each a term the
  protein itself performs. The prose in every moved row was rewritten to match.
• nuoG/nuoL/nuoM had been given MODIFY → GO:0008137 while already
  carrying a separate GO:0008137 row, which is the redundancy CLAUDE.md
  rules out; they are now KEEP_AS_NON_CORE, with the qualifier correction left
  on the one row that needs it. nuoE/nuoH/nuoI keep the MODIFY because
  GOA carries no such row for them.
• aroQ-III's core_functions cited a synthesized summary as supporting_text
  against GO_REF:0000120, which has no cached text, so the verbatim validator
  could not catch it. Re-pointed at the UniProt file with four verbatim quotes,
  and PMID:41029715 now carries a reference_review recording that it
  establishes the aroQ step in KT2440 but does not say which of the three
  paralogs it used.
• Also removed the audit-bookkeeping sentences that had crept into
  review.reason across 21 files (batch-file pointers belong in the notes and
  audit records, per #3100), and dropped a duplicate GO:0008812 row this
  branch had added to cache/go/terms.csv.

2026-09-28

• Re-reviewed every TreeGrafter row marked REMOVE or
  MARK_AS_OVER_ANNOTATED (192 rows / 165 genes, excluding the nine genes
  re-audited on 2026-09-20): 114 stand, 78 relaxed. Added the section above;
  records in TREEGRAFTER/rereview-2026-09-24/. The frozen tables were
  deliberately left alone — the re-review is a post-snapshot event, and
  refreshing the sidecars without the matching failure_mode_curated.tsv pass
  would leave the failure-mode analysis describing a population that no longer
  exists. For whoever does the refresh: re-running analyze_treegrafter.py
  alone gave 968 annotations / 540 proteins with REMOVE 88 and
  MARK_AS_OVER_ANNOTATED 47 (against 120 and 113 frozen), i.e. the
  down-graded set shrinks by roughly a third and ~24 of the removed rows are
  mode-3 cellular-component rows that were assigned by construction. The
  PTHR21272 hotspot row's member list is also now three proteins, not four:
  genes/PSEPK/aroQ and genes/PSEPK/aroQ-III were duplicate folders for the
  same protein (Q88IJ6 / PP_3003) and have been merged into aroQ-III, the
  name UniProt gives the locus.
• Nine of the 306 down-grades that classify_failure_modes.py declines to
  guess (mode 0) were in scope here: acoA, benB, davA, groES, nuoM,
  PP_0094, I7J3R9 and NCGR_LOCUS1270 ×2. Their re-review rationales in
  the batch records are written from the evidence and should make the
  hand-classification of that queue easier.
• Independent corroboration of one retained row: the re-review kept the
  REMOVE on PSEPK fliI GO:0046933 (proton-transporting ATP synthase
  activity, rotational mechanism) on the grounds that phylogenetic transfer
  across homologous ATPase families produced a contradicted molecular
  function. The rotary-ATPase case study
  added the following day reaches the same conclusion from the opposite
  direction — a proteome-wide completeness test whose false positives were
  FliI/SctN export ATPases — and locates the error upstream, in a PAINT IBD on
  a duplication node rather than in the graft.
• Viral proteins appear to have dropped out of TreeGrafter. Prompted by
  the GO:0006302 double-strand break repair row on the phiR8-01 family-A
  DNA polymerase (9CAUD/g022, I7J3R9). Of all the reviewed proteins in the
  corpus with a phage or virus taxon, only two had GO_REF:0000118 rows: g022
  (one row, graft node PTN000015309, PTHR10133:SF27) and phiNIT1
  9CAUD/dfrP (D0VXF2, two rows from PTN000167324, PTHR48069).
  In the 2026-07-27 GOA release (QuickGO) neither protein has any
  TreeGrafter annotation, and PANTHER:PTN… is also gone from the with/from
  of their GO_REF:0000120 rows. Their UniProt entries still carry the
  DR PANTHER family lines, so the protein is still classified in the family;
  only the GO propagation stopped. We have not found a release note that says
  so, and two proteins are too few to call it policy. It fits PANTHER trees
  being built from cellular-organism reference proteomes, which would make a
  graft of a viral sequence an extrapolation outside the tree's taxonomic scope.
• The change cuts both ways. The g022 row was one of our down-grades
  (REMOVE; it is the I7J3R9 mode-0 row in the heuristic queue under Next
  steps). But the two dfrP rows had been reviewed as correct:
  GO:0046452 dihydrofolate metabolic process (ACCEPT) and GO:0046655
  folic acid metabolic process (KEEP_AS_NON_CORE). So excluding viruses
  removes true positives as well as the false one. The DHFR keeps its MF and
  GO:0046654 THF-biosynthesis terms through InterPro2GO and UniRule. But
  GO:0046654 is a sibling of GO:0046452 under GO:0006760, not an
  ancestor, so the dihydrofolate statement had no surviving replacement. It is
  re-asserted in the review as a NEW ISS row, grounded in the PAINT IBD on
  PTN000167322 (PTHR48069), which dfrP shares with E. coli folA (SF3).
• Refreshed both genes' GOA and marked the vanished rows retired: true, which
  keeps their reviews for provenance. The frozen 2026-09-06 tables still count
  the three rows; they will drop out at the next snapshot refresh.
• Follow-up: confirm with the PANTHER/GOA side whether viral sequences are
  now deliberately excluded from TreeGrafter.

2026-09-27

• Acted on the TreeGrafter section of
  REVIEW-2026-09-26.
  Retired the frozen 2026-09-06 snapshot: re-ran analyze_treegrafter.py,
  analyze_placement.py, classify_failure_modes.py and graft_check.py at
  fff7793a6 (a pre-rebase commit no longer on main; superseded by the
  2026-10-01 refresh) and rewrote every number on both pages from the sidecars
  (898 → 983 annotations, 510 → 550 files, 306 → 279 down-grades, accept
  41.3% → 45.1%). New snapshot_drift.py records the drift (85 added, 63
  changed actions, 48 rows out of / 21 into the down-graded set; 56 of the
  63 changes are the 2026-09-20 re-review), and the page now links and
  summarises rereview-2026-09-20/ (19 genes reviewed, 10
  awaiting_adjudication).
• failure_mode_curated.tsv gained a status column: 25 overrides whose
  annotation is no longer down-graded are kept but marked superseded (the
  classifier ignores them); ADAR2 GO:0008251 re-filed 4 → 1 (its graft node
  PTN000098697 also supplied the accepted ADAR terms); 9 newly down-graded
  PSEPK rows curated to mode 1 from the reviewer's reason. The other 12 new
  down-grades are left to the heuristic or mode 0 (14 mode-0 rows now).
• Reporting: added retained and REMOVE-only rates next to down-graded
  (summary sidecar headline rates block), relabelled mode 3 as an
  informativeness policy, said explicitly that the IBA contrast is
  corpus-wide on different genes and reported the same-file IBA set (20 rows
  on 9 files — too small to use). Added a Reference-independence section; new
  exemplar_independence.py shows 9/10 exemplar annotations cite the
  OpenScientist/Falcon report they are scored against, so the exemplar
  agreement claims on the failure-modes page were withdrawn.
• Script hygiene: analyze_treegrafter.py docstring no longer contradicts
  itself on GO_REF:0000120, and gains --out-dir for dry runs;
  graft_check.py reads review_action from treegrafter_review.tsv
  instead of hard-coding it (--refresh-actions updates it offline).
• Added the rotary-ATPase case study.
  It began as a first-principles completeness test (F-type ATP synthase across
  15,525 bacterial reference proteomes), whose false positives turned out to be
  FliI/SctN export ATPases annotated as ATP synthases. The TreeGrafter graft points
  are correct here; the error is inherited from a PAINT IBD on a duplication node.
  None of these rows are in the frozen 2026-09-06 tables.

2026-09-25

• The snapshot callout's drift note was itself quoting exact merged-tree
  figures that a second main merge (3246edc2) had already outdated.
  Rewrote it as a dated floor (≥63 annotations / 28 files across HETGA,
  9AVES, PSEPK, XENLA; four changed actions, naming the two pgm rows that
  cross into the down-graded set) so it no longer re-stales on every merge.
  Added a snapshot marker to the failure-modes sub-page and a caveat there
  that K9IMD0 (draculin) and mdr are contested mode-4 calls awaiting a
  curator. "genes" → "proteins" in the Corroboration and Results text. The
  Caveats composition bullet, which still quoted the earlier 43-row delta as
  "mostly HETGA", was rewritten as the same floor (≥63, HETGA and P. putida
  roughly equal; 656/961 = 68%). Doc-only.

2026-09-19

• Froze the sidecars at the 2026-09-06 snapshot (branch 49d8cc0b, main
b62182cc) and said so on the page: the merged main adds 15 review files
  / 43 GO_REF:0000118 annotations (HETGA 29, PSEPK 12, XENLA 2) and one
  action change (PSEPK fruA GO:0090563 ACCEPT → KEEP_AS_NON_CORE) that
  are not in the tables. The completeness clause is now past tense and
  scoped to the snapshot; the P. putida composition caveat is marked as a
  snapshot property (~70% then, ~68% on the merged tree). No scripts re-run.

2026-09-07

• Worked the four open next steps. (1) GO_REF:0000120 is UniProt's
  multi-pipeline merge (definition confirmed from the GO reference collection),
  so GO_REF:0000118 rows are uncorroborated TreeGrafter calls; evaluated the
  corroborated PANTHER rows (637, 77% accepted) and uncorroborated InterPro2GO
  rows (698, 26% accepted) on the same genes and added the Corroboration
  section. (2) Classified all 306 down-grades into the four failure modes
  (classify_failure_modes.py + failure_mode_curated.tsv, 163 rows curated
  by hand, 134 by keyword heuristic, 9 left unclassified): granularity 47%,
  generic/context 36%, mis-placement 13%, pseudo-enzyme 1%. (3) Added per-family / subfamily /
  graft-node hotspot aggregation (treegrafter_family_hotspots.tsv) and the
  hotspot table. (4) Moved TFP and IRE1 out of the pseudo-enzyme table on the
  failure-modes page (they are node-term and generic-binding cases) and added
  the newly recognised mis-placements (mdh, mqo1–3, dapE/pepV, ptxD, mupP,
  quiA, kdsC, acoA, stilbene dioxygenases).

2026-09-06

• Review pass over the project. The committed sidecars predated the
  P. putida KT2440 batch: regenerated them (corpus 2,775 → 4,540 review
  files; TreeGrafter set 415 → 898 annotations, now covering every
  GO_REF:0000118 row in the corpus GOA). Accept rate unchanged at ~41%,
  MODIFY share fell (12.5% → 8%) and KEEP_AS_NON_CORE rose (17% → 21%).
• Corrected the GO_REF:0000120 paragraph (it is not a "handful" of PANTHER
  rows — ~890, as many as the labelled TreeGrafter set) and the description of
  the mcr LPS-core error (sibling-process error, not host-lacks-pathway).
• Fixed the templates/… links, which resolved above the site root when
  rendered; removed the hard-coded scratch path from run_falcon.sh.
• analyze_treegrafter.py now uses the C YAML loader (minutes → ~45 s) and
  emits the action-by-aspect and per-taxon breakdowns added above.
• Verified the exemplar provenance: all ten OpenScientist and Falcon blinded
  reports and the ten Falcon family reports exist under the gene
  *-hypotheses/ and interpro/panther/PTHR*/ directories, and their verdict
  lines match the tables on the failure-modes page.

Next steps

• Adjudicate the ten awaiting_adjudication genes of the 2026-09-20
  re-review (fogD, PP_3394, K9IJK6, mdr, Cgas, K9IFT7, K9IMD0,
  pvdD, NCGR_LOCUS10166, TOLL9); 13 of the 37 rows on the 2026-09-20
  genes that left the down-graded set since the snapshot are on them, so the current rates are
  provisional there. Five of them have had follow-up reviews since; update
  the records' status when a curator signs off.
• A controlled exemplar sample. Re-run the blinded OpenScientist/Falcon
  check on a fresh sample that includes accepted TreeGrafter rows as
  controls, with the review actions recorded before the reports are read;
  only then quote a sensitivity/specificity.
• Settle the five rows still open after the OpenScientist batch (PP_1969,
  PP_2482, PP_4594, ptxD, gabD-II) with substrate-panel assays (for PP_2482, a
  knockout against its neighbouring molybdoenzyme cluster) (see
  Controlled OpenScientist batch);
  the two term requests (metZ, gabD-II) are collected with the other upstream
  fix points, not filed.
• Blind the corroboration test. Re-adjudicate a sample of GO_REF:0000118
  and GO_REF:0000120 rows with original_reference_id withheld, to separate
  the corroboration effect from the reviewer's visibility of the label.
• Second-pass the heuristic failure-mode assignments. 87 of the 283
  down-grades carry a heuristic mode; 34 of those are decided by construction
  (33 CC rows, 1 low-information binding term), leaving 53 keyword-placed
  MF/BP rows, plus the 33 mode-0 rows: 17 holozoan rows (the five
  untested cadherin over-annotations on each of the three S. rosetta
  cadherins, warts GO:0000082 and GO:0043065) and 16 others (ahpC,
  carB, cbcW ×2, Cgas, groES, mfd, nuoM, PP_0094, PP_0301,
  PP_0312, PP_2608 ×2, pyrC' GO:0006145, pyrE GO:0046132,
  scpC). The keyword
  heuristic is sensitive to reason rewrites: the rejection re-review's
  strengthened reasons moved PP_0312 and scpC from mode 1 to mode 0 and
  benB from mode 0 to mode 4 without any curated call.
• Collect the upstream fix points; do not file tickets yet. Candidate
  PAINT / PANTHER fixes stay listed here until the rule and graft-path
  analyses are complete: a stilbene-dioxygenase subfamily in PTHR10543, the
  GO:0046933 IBD on the PTHR15184 duplication node (see the rotary-ATPase
  case), the EptA node term in PTHR30443, the PMT subfamily in PTHR11558,
  PKS-vs-FAS in PTHR43775 (an IBD on the family root), the M20A subfamilies in
  PTHR43808, the MDH/LDH node in PTHR43128, and the fusion-seeded root IBDs in
  PTHR10520 (purM), PTHR22960 (moaA) and PTHR10953 (moeB) (see
  Propagation rules and graft paths).
• Run the EC-contradiction rule over all of GOA. The UniProt-wide counts
  in ec_contradiction_scale.tsv
  cover only the 36 pairs seen in the corpus; a GOA × UniProt EC join would
  give the full list of TreeGrafter MF rows whose EC contradicts the
  protein's own.
• Test the fusion-seeded-root shape across PANTHER. For every IBD on a
  family root, check whether all its seeds are multi-domain proteins whose
  extra domain the rest of the family lacks.
• Test a domain-coverage rule. In DONKEY40, 16 down-graded or UNDECIDED
  TreeGrafter rows are on fragmentary or truncated donkey gene models. A rule
  that flags a TreeGrafter row when the protein lacks the InterPro domain its
  term depends on (for example a kinase term with no kinase domain) could be
  scored against the reviews in the same way as the other rules.
• Human-rated anchor. Have a curator rate a small shared subset so that
  agreement between the agent reviews and expert judgement can be quoted.
• Broaden the taxon base. 65% of the rows are P. putida KT2440; the 29
  HETGA rows (76% accepted) are a first mammalian batch and the 65 holozoan
  rows (63% down-graded) a first unicellular-eukaryote batch, both too small
  and too hand-picked to generalise from. DONKEY40 adds 225 mammalian rows,
  selected by rule rather than by hand (17.3% down-graded, 39/225).
• File the GO:0009329 label/axiom defect with GO. This is the one item the
  2026-09-24 re-review turned from a suspicion into a verified finding (see the
  label-vs-definition bullet above): the term's definition is the AccA/AccD
  carboxyltransferase component of acetyl-CoA carboxylase, but its label says
  "acetate CoA-transferase complex" and it carries a capable_of axiom to
  GO:0008775 acetate CoA-transferase activity — EC 2.8.3.8, different
  chemistry. Both contradict the definition. interpro/panther/PTHR42995/ PTHR42995-paint.tsv shows GO curators themselves use the term for bacterial
  AccD (an IBD seeded by E. coli P0A9Q5), and QuickGO's own blacklist for the
  term is ~80 NOT-qualified annotations on eukaryotic proteins (QuickGO,
  2026-10-02, the same query as the verification above) — which is what a
  misleading label looks like downstream. Proposed fix: rename to name the
  carboxyltransferase component and drop or re-point the capable_of axiom.
• ~~Refresh the snapshot~~ — done on 2026-10-03 at f81b9f300 (see the
  2026-10-03 note): the rejection re-review and the holozoan rows are in
  every table, the failure-mode table was re-curated, and the hotspot member
  labels are de-collided (analyze_placement.py now labels same-symbol
  proteins by species/review-folder, so the two PSEPK dapF paralogs
  would read PSEPK/dapF__Q88CF3 and PSEPK/dapF__Q88GD4; neither dapF
  row is down-graded any more, so PTHR31689 has no example genes at this
  refresh). Still open from that item: re-pin again when the next batch
  lands.
• Harmonize the rows the rejection re-review left out of scope. Some sibling
  rows now sit beside a relaxed parent (PSEPK ubiA GO:0004659 /
  GO:0016765, zwf GO:0006098). PSEPK murB and ubiK carry a sharper
  version: their GO:0005737 cytoplasm row is ACCEPT (a GO_REF:0000120 row,
  outside this audit's scope) while the more specific GO:0005829 cytosol is
  KEEP_AS_NON_CORE, so the less specific term currently reads as the core one.
  pdxB and pdxJ have the consistent pairing.