"X involved in Y" — Precomposed GO Terms and Their Decomposition

Bottom line: GO contains 784 live terms whose label reads "X involved in
Y", and it has already obsoleted 1,265 more, 883 of them with a comment that
the term "is pre-composed and should be captured as a GO-CAM model". Our gene
reviews use 349 distinct such terms in 1,462 existing_annotations rows
across 735 genes (39 species folders, species_folders in data/summary.yaml);
8 of those rows sit on obsolete terms (14 when the project started). The schema could not say where the Y context goes when a
term is replaced by its X. This project added decomposed_into to MODIFY
(pointing at a module, a GO-CAM or the gene's core functions, or recording that
Y is not carried), worked the 14 obsolete-term rows and a chaperone-mediated
autophagy pilot with a new module, and set a module policy: modules are the
reusable mechanisms (X), not the contexts (Y). A short-lived
annotation-extension slot was removed, because X plus part_of(Y) just
restates the retired class.

What "precomposed" means, and what "replace it with a GO-CAM" means

A precomposed term packs two statements into one class. For example, for
GO:0061740 protein targeting to lysosome involved in chaperone-mediated
autophagy, GO's logical definition is

GO:0061740 ≡ GO:0006622 protein targeting to lysosome
             and part_of some GO:0061684 chaperone-mediated autophagy

The first part, the genus (X), says what the gene product does. The second
part, the differentia (Y, here attached by part_of), says what larger
process that activity serves. Annotating HSPA8 to GO:0061740 asserts both at
once. GO uses the same shape at three levels:

Precomposition does not scale. GO would need a term for every X crossed with
every Y. Each pair is also a little GO-CAM squeezed into a class name, and it
cannot say which Y instance, which upstream or downstream activity, or in
what cell. GO has therefore been retiring these terms. The obsolete stanzas
say so: 501 carry "obsoleted because it is a pre-composed term that should be
captured as a GO-CAM model" (e.g. GO:1902284 neuron projection extension
involved in neuron projection guidance, go-ontology#31330), 200 "represents
a GO-CAM model", and 40 "pre-composed and should be represented as a GO-CAM
model". A 2023 GO annotation call also noted that "BP involved in BP" and "MF
involved in BP" terms are better represented as GO-CAMs
(Annotation Conf. Call 2023-05-16;
only the search-result summary was reachable from here, not the page itself).

"Replace with a GO-CAM" means un-packing the class back into the graph it
abbreviates:

The decomposed form keeps every assertion the class made: HSPA8 is still
inferred involved_in both X and Y. It also gains what the class could not
hold, namely the MF that does the work, the location, and the causal links to
the other participants. Nothing is lost if the Y context is kept somewhere
linked to the annotation. The loss happens when people "decompose" by
splitting the row into two unrelated annotations (involved_in X,
involved_in Y), because that drops the fact that X happens as part of Y.

Where the context should go, in this framework:

1. A module or production GO-CAM, when the gene takes part in a
   multi-participant mechanism. This is the real replacement: it adds the
   activity that does the work, its inputs, and the causal links around it.
2. The gene's core_functions, when Y is central to the gene. These are
   already a gene-local GO-CAM (molecular_function + directly_involved_in +
   locations + in_complex).
3. Nowhere, recorded as context_dropped, when Y is peripheral (typically
   a propagated developmental phenotype) or was itself the over-annotation.

Not an annotation extension. involved_in X with part_of(Y) is exactly
GO's logical definition of the retired class, so writing it is
precomposition in different syntax. GO asked for these terms to be "captured
as a GO-CAM model", not re-encoded. We briefly had an extension slot and
removed it (2026-10-25); see the schema section.

GO's own pointers behave accordingly. replaced_by names a single term and
is applied to existing annotations automatically, so an obsoleted
precomposition necessarily keeps one half. consider lists both halves, but
it is advice for a curator, not an automated mapping.

Inventory (GO release 2026-07-26)

Generated by scripts/inventory.py
into data/. No result is
hardcoded, so re-running it against a newer go.obo updates every number
here.

Figures refreshed 2026-11-07, after merging main; they are a snapshot, and
data/summary.yaml is authoritative. The first snapshot
(2026-10-18) had 1,023 rows over 519 genes. Current review actions on these
rows (after round 6): KEEP_AS_NON_CORE 690, ACCEPT 358, MODIFY 169,
MARK_AS_OVER_ANNOTATED 116, UNDECIDED 77, REMOVE 26, NEW 17, TOO_GENERAL 9. Apart from the rows this
project decomposed, the existing MODIFYs do not decompose: in the first
snapshot, one of 71 proposed the bare genus and the other 70 were lateral moves to a different term (e.g.
DNA strand elongation involved in DNA replication → G1/S transition of mitotic cell cycle for Drosophila chromatin factors). Evidence: IMP 368,
IEA 310, ISS 193, IDA 160, IBA 112, ISO 110, IGI 83, other 115.

By organism folder (by_organism in data/summary.yaml): human 737 rows,
mouse 150, rat 140, DROME 120, worm 88, DICDI 75, yeast 35, other 106.

Rows on already-obsolete terms (fix first)

The GATA3 consider pairs show the information loss directly. GO offers
both halves as separate terms; only a model can keep "negative regulation of
proliferation as part of mesonephros development", and the question is
whether that context is worth a model at all.

Where GO-CAMs already realize the decomposition

The check looks for a cached GO-CAM activity whose BP is Y or a descendant of
Y. Examples:

• HSPA8 (human; mouse and rat by ISO/ISS) — model 63c0ac2b00001252:
  protein-macromolecule adaptor activity part_of GO:0061740. The production
  model still uses the precomposed BP itself, so obsoleting the term would
  mean nesting protein targeting to lysosome inside chaperone-mediated autophagy in the model as well. A GO-CAM is only "the replacement" once its
  own process nodes are decomposed.
• SYK (mouse rows) — human model 66c7d41500000715: protein tyrosine kinase activity part_of neutrophil activation involved in immune response. This is an ortholog hit through the same-symbol proxy.
• RAD51 (human; SCHPO rad51 by ortholog) — 69a0c46f00002668:
  single-stranded DNA binding part_of double-strand break repair involved in meiotic recombination, which realizes Y = meiotic cell cycle for the
  row chromosome organization involved in meiotic cell cycle.
• ATG7 — 6446bfcb00001774: Atg12 activating enzyme activity part_of
  autophagosome assembly, which realizes protein lipidation involved in autophagosome assembly with a better MF than the precomposed BP carries.

Production GO-CAMs are themselves full of precomposed BP nodes. "Replace with
a GO-CAM" is therefore an instruction to the GO-CAM curators too, not only to
the GAF.

Recurring Y contexts

data/module_candidates.yaml
groups rows by Y. There are 195 contexts; 93 have three or more genes, and
70 of those are named by no module (crude test: the GO id does not appear in
any module YAML). They are not a module to-do list. Read them through the
module policy below: some Y's are really mechanisms (module material), most
are tissue or developmental contexts (not).

Broad contexts such as apoptotic process (26 genes), immune response, or
anatomical structure development mostly mark grouping terms (class D).

Module policy: modules are the composable parts

A precomposed term is usually mechanism X used in context Y:
cell migration in sprouting angiogenesis, negative regulation of FGFR
signaling in ureteric bud formation, protein targeting to lysosome in
chaperone-mediated autophagy. Modules should model the X side.

• Build a module when X (or Y) is a reusable, multi-participant mechanism
  with its own steps: chaperone-mediated autophagy, ESCRT-I sorting, the
  hair-cell mechanotransduction complex, caspase activation and inhibition.
  Such a module is used across many contexts.
• Do not build a module per developmental or tissue context. Ureteric bud
  branching, sprouting angiogenesis, heart valve morphogenesis and mesonephros
  development are not mechanisms. They are programmes that reuse mechanisms we
  already model (FGFR, GDNF/RET, Notch, VEGFR, BMP, Hedgehog signalling) in a
  particular place and order. A module per context would duplicate the
  signalling modules with a tissue label attached, and it would recreate the
  X × Y explosion that GO is removing, one level up.
• Contexts are expressed as context, not as new mechanisms. When a
  developmental programme is genuinely worth curating, model it as a concrete
  composition. That can be a production GO-CAM, which is instance-level and
  causal and is GO's intended home. It can also be a concrete module whose
  context carries the anatomical_locations, developmental_stages and
  cell_types, and whose parts conforms_to the existing signalling
  templates. Both already exist in the schema. Do this only when the programme
  is the object of curation, never just to give one annotation somewhere to
  point.
• So a row points at its mechanism. A VEGFA row on "positive regulation of
  cell migration involved in sprouting angiogenesis" decomposes to positive
  regulation of cell migration, with decomposed_into naming the
  vegfr_signaling element VEGFA occupies. The angiogenesis context is either
  in that module's context (if it is the module's scope) or dropped.

Triage scheme for every row

Each row gets one decomposition outcome, recorded alongside the ordinary
action. Every outcome that keeps X is a MODIFY to X with decomposed_into
saying where Y went:

Outcomes:

• DECOMPOSE_MODEL — MODIFY to the genus; Y lives in a named GO-CAM activity or module element (decomposed_into.gocam_associations / module_associations), or in the gene's core functions (in_core_functions).
• DROP_CONTEXT — MODIFY to the genus with context_dropped: true, because Y is peripheral context that does not warrant a model, or was itself the over-annotation (e.g. Drosophila acetyltransferase complex members on DNA strand elongation involved in DNA replication). The reason says which.
• KEEP_PRECOMPOSED — the term is the right grain (D-i/D-ii) or GO is not retiring it. Record the reason.
• Ordinary REMOVE / MARK_AS_OVER_ANNOTATED when neither X nor Y holds. Decomposing does not rescue a wrong annotation.

The CLAUDE.md rules still apply. Rows sharing a term must get consistent
actions. A true precomposed term is never REMOVEd to deduplicate. Experimental
rows whose full text we have not read are not overruled. The "do not add
what curators declined" test also applies to Y: being necessary for Y (a
knockout phenotype) is not the same as doing part of Y, and many class C
rows fail that test.

Schema: decomposed_into on MODIFY (implemented)

Before this project, Review had only action, reason and
proposed_replacement_terms (id and label), so a MODIFY to X could not say
where Y went except in prose.

Decisions (project owner): keep MODIFY, with no new action (2026-10-18);
record Y only by reference to a model, not as an annotation extension
(2026-10-25). For one week the schema also had proposed_replacement_extensions
(X plus part_of(Y)). It was removed because it restated the retired class,
added none of the structure a model adds, duplicated the module link where both
existed, and nothing in the pipeline consumed it. Its 15 rows were converted:
the 8 HSPA8 rows already linked to the GO-CAM and module; the 7 GATA3, NOTCH1
and DNMT1/Dnmt1 rows became context_dropped.

review:
  action: MODIFY
  proposed_replacement_terms:
    - id: GO:0006622
      label: protein targeting to lysosome
  decomposed_into:                        # Decomposition: where Y now lives
    gocam_associations:                   # GoCamAssociation (same class modules use)
      - model: gomodel:63c0ac2b00001252
        activity: gomodel:63c0ac2b00001252/63c0ac2b00001260
    module_associations:                  # ModuleAssociation: modules/<module>.yaml + element id
      - module: chaperone_mediated_autophagy
        element: hspa8_substrate_adaptor
    in_core_functions: true               # Y recorded in core_functions[].directly_involved_in
    context_dropped: false                # true = Y deliberately not carried (peripheral, or the over-annotation)
    notes: ...

Validation (src/ai_gene_review/validation/decomposition.py, run by just validate, tested in tests/test_decomposition_validation.py):

"To X" is judged against GO when just validate runs ontology checks: the
replacement counts if it is an is_a ancestor of the combined term, or if
its label or a synonym equals the label-derived genus. That catches drift
such as cell proliferation → cell population proliferation. Without the
ontology, only an exact label match counts. Module element ids must be
node, part or annoton ids, not cited CURIEs.

Lateral MODIFYs, where the replacement is not the genus, are not flagged.
Known asymmetry: decomposed_into records where Y went, but nothing
records a dropped X. When the genus fails and only the context survives, as
for GATA3's ureteric-bud rows, the outcome is an ordinary action
(MARK_AS_OVER_ANNOTATED, or a lateral MODIFY) with the reason saying so.
The gene-review HTML template shows decomposed_into as badges.

Worked cases (round 1)

Each case is recorded in the gene review with decomposed_into and validates
cleanly. After round 1, 15 rows carried a recorded decomposition (16 before
human DNMT1 was retired upstream). After round 3, rows_decomposition_recorded
in data/summary.yaml is 47; besides this project's rows, that includes
decompositions recorded by other reviews merged to main.

The obsolete-term rows

Pilot: chaperone-mediated autophagy (GO:0061740)

• New module
  modules/chaperone_mediated_autophagy.yaml.
  It has two steps: (1) HSPA8 recognizes the KFERQ-motif substrate and
  delivers it to the lysosome (protein-macromolecule adaptor activity, part of
  protein targeting to lysosome); (2) LAMP-2A multimerizes into the
  translocation complex, and HSPA8 then disassembles it. It is grounded in
  PMID:2799391, PMID:11559757, PMID:18644871 and PMID:26212789, and linked to
  production GO-CAM 63c0ac2b00001252. The LAMP-2A function carries no GO
  id. The GO-CAM's signaling adaptor activity describes a relay rather
  than substrate binding. Transmembrane protein transporter activity would
  claim LAMP-2A is the conduit, which is not established. Cargo receptor
  activity is defined for delivery to transport vesicles.
• HSPA8 (human IDA ×2, IMP, TAS; mouse ISO, ISS; rat ISS, ISO; 8 rows):
  DECOMPOSE_MODEL. MODIFY → GO:0006622 protein targeting to lysosome, with
  decomposed_into pointing at the GO-CAM activity (the human ortholog's, for mouse and rat), the module annoton
  hspa8_substrate_adaptor, and in_core_functions. Each core function's
  directly_involved_in swaps GO:0061740 for GO:0006622 and keeps CMA.
• CLU (IDA, PMID:9228033): DROP_CONTEXT. Decomposition exposed a wrong
  Y. The evidence is LRP2/megalin receptor-mediated endocytosis of
  extracellular CLU–Aβ complexes, followed by lysosomal degradation. That is
  not CMA, which translocates cytosolic KFERQ proteins through LAMP-2A.
  MODIFY → GO:0006622 with context_dropped: true. The CLU core function now
  lists protein targeting to lysosome instead of the CMA-specific term.

Worked cases (round 2): caspase regulation

The MF terms GO:0043027 cysteine-type endopeptidase inhibitor activity
involved in apoptotic process, GO:0008656 … activator activity involved in
apoptotic process and GO:0043028 … regulator activity involved in apoptotic
process, across 22 genes. Each one's X is a genuine molecular function:
GO:0004869 cysteine-type endopeptidase inhibitor activity or GO:0140608
cysteine-type endopeptidase activator activity. Y (apoptotic process) is
modelled by the apoptosis modules.

Ontology note. GO places neither combined term under its genus.
GO:0043027 is_a only GO:0043028, and GO:0008656 is_a GO:0016505 and GO:0043028.
Neither is_a GO:0004869 or GO:0140608, so annotations to the combined term do
not reach the genus by reasoning. Ticket 3 in
upstream-tickets.md asks
for this to be fixed.

Non-core rows. MODIFY carries no core/non-core flag. When a KEEP_AS_NON_CORE
row on a combined term is decomposed, the non-core judgment survives only in
reason, or in whether a core function names the genus. Where the genus itself
is doubtful, prefer MARK_AS_OVER_ANNOTATED (CASP1, BCL2L13) to a MODIFY that
would silently promote it.

The inventory now also counts reviewer-authored combined terms: replacement
terms, core-function terms and isoform-specific terms that a review
introduced, rather than GOA rows it reviewed, plus combined terms used
anywhere in a module annoton (function, processes, targets). These are in
data/authored_precomposed.yaml.
After round 6 there are 302 uses: 271 across 166 genes (164 in
directly_involved_in, 99 as MODIFY replacements, 6 as core-function MFs
and 2 as isoform-specific terms) and 31 in 12 modules (30 as annoton
processes and 1 as an annoton function, GO:0099626 in
synaptic_vesicle_exocytosis). 56 of them are on the label-only MOMP terms
(round 3) and are flagged label_only; they need no decomposition. 71 more
are on the five terms kept in rounds 5 and 6 and are flagged kept. That
leaves 175 to work through. Label-only and kept rows are also left out of
module_candidates.yaml: a label-only filler is the regulated process, and a
kept term is not being decomposed, so neither is a context to model. This is a second work stream:
reviews have been actively moving annotations toward these terms.

Worked cases (round 3): label-only MOMP terms and ATG8 lipidation

Round 3 started with the largest group of reviewer-authored combined terms
and found it was not precomposed at all.

GO:1901030 / GO:1901029 (positive / negative) regulation of MOMP involved in apoptotic signaling pathway — no decomposition

56 authored uses (BID, PMAIP1, BBC3, BCL2L11, MCL1, BCL2, and 10 annoton
processes in bcl2_family_momp_control and trail_apoptotic_signaling) plus
10 GOA rows (BAK1, GSK3B, LRRK2, rat Gsk3b, ACAA2, HSPA1A, SLC25A5). GO's
logical definition of GO:1901030 is biological regulation and positively_regulates GO:0097345 mitochondrial outer membrane
permeabilization, with no part_of differentia. The "involved in apoptotic
signaling pathway" is only in the label. It is entailed rather than added,
because GO:0097345 is itself defined as MOMP "as part of the apoptotic
signaling pathway" (part_of GO:0097190). Removing the context leaves the
same term; there is no broader regulation of MOMP class to move to, and
GO:1901028 lists "regulation of MOMP" as an EXACT synonym. Outcome: keep
every row; no decomposition. The BH3-only, BCL-2 and BAK rows are judged
on their own evidence as before.

This is a new triage class, E. label-only: a regulation term whose
"involved in Y" is entailed by its own definition. The rule tests that
entailment directly: the genus is the bare biological regulation, there is no
part_of differentia and no nested filler, and Y (the label text after
"involved in") is an is_a/part_of ancestor of the regulated process. The
inventory flags these terms, their rows (rows_label_only) and their authored
uses (label_only: true, uses_label_only), so they no longer count as
decomposition work. Four live GO terms qualify: the three MOMP regulation terms
(GO:1901028/29/30) and GO:0061081 positive regulation of myeloid leukocyte
cytokine production involved in immune response, whose regulated process
GO:0061082 is_a GO:0002367 cytokine production involved in immune response.
GO:0061081 is the weakest of the four: its entailment runs through GO:0002367,
which is itself precomposed and in this project's queue. If GO retires
GO:0002367 and re-parents GO:0061082, the "immune response" ancestor may go and
the term will drop out of class E on the next inventory run. The MOMP terms
have no such dependency: GO:0097345's part_of GO:0097190 is a plain context
axiom.

Not class E:

• Grouping intersections. GO:0003006 developmental process involved in
  reproduction is defined as GO:0022414 reproductive process and GO:0032502 developmental process, and GO:0022412 likewise. Y is in the axiom as a
  second genus, not as a part_of, so these stay class D-ii. The inventory
  records only the first genus of such two-genus definitions.
• Regulation terms whose Y is not entailed. GO:0051946/GO:0051951
  (positive) regulation of glutamate uptake involved in transmission of nerve
  impulse regulate GO:0051935 glutamate reuptake, which has no ancestor
  labelled "transmission of nerve impulse".
• A specific genus that regulates Y. GO:0071631 mating pheromone
  secretion involved in positive regulation of conjugation with cellular
  fusion has a real Y.

GO's synonyms on GO:1901030 and GO:1901029 call the context-free phrase BROAD,
which disagrees with the logical definition; that is upstream ticket 5.

GO:0061739 protein lipidation involved in autophagosome assembly → GO:0006501 C-terminal protein lipidation

Why the decomposition is safe: every ATG3/ATG7 review already accepts
GO:0000045 and lists it in its core functions, so the context is carried
without the precomposed term, and the MFs name the Atg8-family substrate.
Human ATG16L1 already accepts GO:0006501 (IDA), so the decomposed form was
established in the corpus. Participation check: the E1 and E2 catalyse steps of
the conjugation cascade, so a lipidation process term is theirs. They are not
substrates. ATG8 lipidation also occurs outside autophagosome assembly (the
yeast Cvt pathway; LC3-associated phagocytosis and conjugation to single
membranes in animals), which is a second reason to prefer the genus.

Lessons from round 3

1. Read the logical definition before decomposing. The label pattern
   "X involved in Y" overcounts: when the definition has no part_of, the Y
   is either entailed by X's own definition (MOMP) or absent (grouping
   classes).
2. The genus can be the better term, not just the shorter one. For
   yeast ATG7, the precomposed child excluded a context in which the gene
   acts (Cvt).
3. A node concept is not an annotation. Module annotons take the genus;
   the node that models the step may keep the exact precomposed class as its
   concept.

Worked cases (round 4): cadherin binding

GO:0098641 cadherin binding involved in cell-cell adhesion (class A, MF
part_of BP) is defined as GO:0045296 cadherin binding and part_of GO:0098609 cell-cell adhesion. It is also is_a GO:0098632 cell-cell
adhesion mediator activity. That looked like a possible lateral target, but
GO:0098632 is itself defined as GO:0005515 protein binding and part_of GO:0098609. Its label doesn't say "involved in", but its definition is the same
X-plus-context shape, so moving to it would not decompose anything. The
decomposition is to the genus GO:0045296, with cell-cell adhesion carried by
the core functions. Every cadherin review already lists a cell-cell adhesion
process there: GO:0007156 homophilic cell-cell adhesion (CDH1), GO:0044331
cell-cell adhesion mediated by cadherin (CDH2, CDH3, CDH11, FAT4), and
GO:0098609 (DSG2, DSC2; DSG2 also GO:0002159 desmosome assembly). No module
covers adherens junctions or desmosomes yet; under the module policy that
would be a mechanism worth modelling, but it is not needed to place the
context here.

Worked cases (round 5): two terms that stay

The plan's next two authored terms turned out to be the right grain. Neither
is decomposed, and no review changes. Both are now listed, with their
reasons, in
keep_precomposed.yaml,
which the inventory reads. Together they account for 49 authored uses and 41
GOA rows.

GO:0106074 aminoacyl-tRNA metabolism involved in translational fidelity — kept

The term has no logical definition (class D-iii). Its definition, "Any process
which detects an amino-acid acetylated tRNA is charged with the correct amino
acid, or removes incorrect amino acids from a charged tRNA", is aminoacyl-tRNA
editing, and its RELATED synonyms say so ("aminoacyl-tRNA proofreading",
"aminoacyl-tRNA correction"). GO has no other term for that process.

• The X half has no term of its own. "Aminoacyl-tRNA metabolism" is not a
  GO class. The nearest genus, GO:0006399 tRNA metabolic process, covers all
  of tRNA metabolism (modification, processing, aminoacylation, degradation),
  so decomposing to it would discard what the term says.
• The Y half is already carried. The term is is_a GO:0006450
  regulation of translational fidelity. A decomposition has nowhere better
  to put "translational fidelity" than that second parent.
• GO anchors the editing activity here. GO:0002161 aminoacyl-tRNA
  deacylase activity is part_of GO:0106074. Every review that uses the term
  in core_functions (15 genes: the class I and II synthetases' editing
  domains, AARSD1/AlaX, DTD1, ANKRD16 and others) pairs it with GO:0002161 or
  one of its children (Ser-tRNA(Ala), Gly-tRNA(Ala), Ala-tRNA(Pro),
  D-aminoacyl-tRNA deacylase, L-seryl-tRNA(Thr) hydrolase activity). The
  exception is ANKRD16, whose activity has no GO term yet: it binds AlaRS
  and captures serine that AlaRS has misactivated, on its own lysine side
  chains, before the serine can reach tRNA(Ala). Its review records that as
  proposed_molecular_function: misactivated amino acid acceptor activity.

Of the 14 rows on the term, 13 are electronic and one is a reviewer-proposed
NEW row (ISS, P. putida thrS). The electronic rows come from GO_REF:0000108
(logical inference from inter-ontology links, 8 rows) and GO_REF:0000120
(combined IEA, 5 rows). Both infer the process from the deacylase activity
through the same part_of GO:0106074 link the keep decision rests on: the
human DTD1 row lists GO:0002161, GO:0051499 and GO:0106026 as its
supporting_entities. Their reviews judged each row on its merits: 12 are
ACCEPT, one is the NEW row, and one is MARK_AS_OVER_ANNOTATED (the truncated
human DTD1 product A0A2R8YCT7, which lacks the editing site and so has none of
those activities; the term stays on the canonical Q8TEA8). None is a MODIFY
away from the term.

GO:0006271 DNA strand elongation involved in DNA replication — kept

The term is class B with a logical definition (GO:0022616 DNA strand elongation and part_of GO:0006260 DNA replication), so it could be
decomposed mechanically. It should not be:

• It names a stage, not a mechanism used in a context. Its EXACT synonym
  is "DNA replication elongation", and Reactome has the same unit
  (R-HSA-69190 "DNA strand elongation"). It stands to replication as
  GO:0006414 translational elongation stands to translation, a term whose
  label happens not to say "involved in".
• Its subclasses are stages too. GO:0006272 leading strand elongation
  and GO:0006273 lagging strand elongation are is_a GO:0006271. Retiring
  the parent would leave them under the bare genus, which is wrong for both.
• The genus is broader in a way that matters. GO:0022616 also covers
  strand extension outside replication. The genes that carry GO:0006271 are
  replicative polymerases (POLD2 of polymerase delta, and the B-family
  polymerase of phage BPPH2), the sliding clamp (ECOLI/dnaN, BACSU/dnaN, PSEPK/dnaN), the
  clamp loader (RFC subunits, CG8142) and the MCM helicase (MCM4, mcm-4).
  For them, "which strand-extension process" is the information.

Of the 27 GOA rows, 21 are ACCEPT on these core replication proteins. The
others were already MODIFYed for reasons unrelated to the label: the four
Drosophila Enok acetyltransferase complex members (Br140, Eaf6, Ing5,
enok; IDA, PMID:27198229) move to GO:0000082 G1/S transition of mitotic
cell cycle, elg1 to GO:0006261 DNA-templated DNA replication, and
ECOLI/dnaC to GO:0031297 replication fork processing. Those
reviews correct over-annotations; they are not decompositions and are left
as they are.

Keeping the class does not settle each row's granularity. DROME/RfC3,
DROME/RfC38, DROME/RfC4, yeast/RFC3 and human/POLD2 also carry GO:0006272
leading strand elongation or GO:0006273 lagging strand elongation
beside their accepted GO:0006271 rows. Whether a given row should be refined
to one of those children is an ordinary review question, separate from
whether GO:0006271 should exist.

Lessons from round 5

1. "Involved in" in a label is not a reason to decompose. Some terms name
   a process that GO has no other word for (GO:0106074), and some name a
   stage of the Y process that has its own subclasses (GO:0006271). Ask
   whether the X half, standing alone, would still say what the gene does.
   If no X term exists, or the bare X would lose which process the gene
   serves, keep the term and record why.
2. There is no mechanical test for this. An EXACT synonym without
   "involved in", or is_a children, looked like a usable signal. Across the
   inventory it flags more than a hundred terms, and many of those EXACT
   synonyms just reword the X-in-Y pair ("coronary angiogenesis" for angiogenesis involved in
   coronary vascular morphogenesis), and grouping classes have children too.
   Keep decisions are made term by term and listed in
   keep_precomposed.yaml.
3. Upstream, these are renames, not obsoletions. Renaming GO:0006271 to
   its EXACT synonym, and GO:0106074 to the editing process it defines, would
   take both out of the "involved in" set without moving an annotation
   (ticket 6).

Worked cases (round 6): hearing detection kept, developmental apoptosis decomposed

GO:0050910 detection of mechanical stimulus involved in sensory perception of sound — kept, with GO:0050973 and GO:0050974

The plan had this term down for a module. The module now exists:
hair_cell_mechanotransduction, built from RGD PW:0001503, models the
CDH23/PCDH15 tip link, the TMC1/TMC2 channel with LHFPL5, TMIE and CIB2, the
upper tip-link density and the inner-hair-cell synapse. Its annotons already
use GO:0050910, and its notes record that the same apparatus serves balance
in vestibular hair cells, where it uses the sibling GO:0050973 (TMC2 carries
both; the MYO1C adaptation motor carries GO:0050973).

That looks like "mechanism X used in context Y", but the round-5 test says
keep:

• The bare genus loses the sense. GO:0050982 detection of mechanical
  stimulus also covers non-perceptual detection: GO:0001981 baroreceptor
  detection of arterial stretch, GO:0002000 detection of renal blood flow,
  GO:0035995 detection of muscle stretch. For TMC1 or LHFPL5, "which sense
  this detection serves" is the information.
• The terms are a per-sense family, not one-off combinations. Under
  GO:0050974 detection of mechanical stimulus involved in sensory
  perception sit hearing (GO:0050910), touch, pain and equilibrioception
  (GO:0050973), each the transduction stage of its sense, with EXACT
  synonyms such as "detection of sound" and "sensory transduction of
  sound".
• The shared apparatus already has a term. Where one apparatus serves
  several senses, GO:0050974 states it without either context, and the
  zebrafish lhfpl5a/lhfpl5b, tmc2a/tmc2b and tomt reviews already use it.

So the three terms go on the keep list (30 rows, 22 authored uses), and the
module and reviews are unchanged. The alternative, MODIFY to GO:0050982 with
the sense carried by the module, was considered and rejected because the
genus is a grouping of interoceptive and sensory detection that no review
would choose on its own. The rows on these terms are judged on their own
merits as before (most of the 30 are phenotype or orthology rows on
transcription factors, efferent receptors and bundle proteins, and are
MARK_AS_OVER_ANNOTATED or KEEP_AS_NON_CORE).

GO:1904747 positive regulation of apoptotic process involved in development → GO:0043065 positive regulation of apoptotic process

GO:1904747 is class C and nested: biological regulation and positively_regulates GO:1902742, where GO:1902742 is apoptotic process and part_of GO:0048856 anatomical structure development. Its 12 authored uses
were all MODIFY replacements in six C. elegans engulfment-gene reviews.
They come from the APOPTOSIS project's
assay-readout page. That page
found that the CED-8 paper (PMID:24225442) gave GO:1902742 to engulfment
mutants it used only as a corpse-preserving background. It moved those rows
to positive regulation, because engulfment genes promote death from the
engulfing cell (Reddien 2001; Hoeppner 2001) rather than taking part in the
dying cell's program.

That sign correction stands. What round 6 removes is the developmental
context the term inherited from GO:1902742: it is where the pro-death role
was measured, not a mechanism, and every review calls the role non-core.
Under the class C rule the term unfolds to regulation of X, GO:0043065.

The second line is needed for consistency. Had only the replacements moved,
ced-1, ced-6 and ced-7 would propose GO:0043065 while keeping its child
GO:1904747 on other rows of the same claim. The assay-readout table, the
six notes files, a ced-2 reason and a ced-12 reference note are updated to
match.

Left as they are: the ACCEPT rows on GO:1904747 for worm egl-1 and
Drosophila hid, which need their own reading in the full pass, and the
ced-3, ced-4 and Drosophila Dark rows, which already MODIFY GO:1904747
laterally (to execution phase, apoptotic signaling pathway, or positive
regulation of apoptotic signaling pathway) for reasons unrelated to the
context.

Lessons from round 6

1. A module for the mechanism does not by itself retire the term.
   hair_cell_mechanotransduction is the mechanism, and it still needs a
   process term to say which sense each part serves; GO's per-sense
   detection terms do that, and their common parent covers the shared case.
2. Decompose the gene's own GOA rows when they make the same claim.
   Moving only the replacements would have left three genes proposing a
   term and keeping its child. The non-core judgment survives in the
   reason, as the non-core rule above allows.
3. Another project's choice can be kept and refined. The assay-readout
   page chose the term for its sign; this round keeps the sign and drops
   only the context, and records that on the page itself.

Lessons from round 4

1. A term without "involved in" in its label can still be precomposed in
   its definition. GO:0098632 is protein binding and part_of cell-cell adhesion; the inventory finds only "involved in" labels, so such terms
   are invisible to it. They are not worth chasing as lateral targets.

Lessons from round 1

1. replaced_by is automated by design, so a wrong one is a GO bug, not a
   curator choice. For GO:0000750, the automated migration moves an upstream
   GEF (Arr4/Get3) onto the MAPK cascade itself. A precomposed term's
   replaced_by is only safe when one term is true of every annotated
   participant. Otherwise the term should carry consider only.
   consider is a different thing: it is curator advice and is never applied
   automatically. The GATA3 case (its consider genus is unsupported for this
   gene) is that advice working as intended.
2. Decomposing is a review, not a reformat. Splitting the term into X and
   Y forces each half to be justified separately. In 2 of 7 genes (GATA3
   ureteric bud, CLU), one half failed. The precomposed class had hidden that.
3. Same-term rows move together. The IEA, ISS and ISO copies of a claim get
   the same decomposition, as the action-consistency rules require.
4. Production GO-CAMs still use precomposed terms (the CMA model's HSPA8
   node is GO:0061740), so a module is often the cleaner decomposition target.
5. Most Y contexts are not worth carrying. In round 1 only HSPA8's Y
   (chaperone-mediated autophagy) was a mechanism with a model. GET3's context
   survived inside a lateral replacement term. The other Y's were
   developmental contexts on propagated rows (GATA3, NOTCH1, DNMT1) or wrong
   (CLU).

Plan

1. ~~Schema~~ — done: decomposed_into, a validator, tests, and template
   rendering (the extension slot was added and then removed).
2. ~~Obsolete-term rows~~ — done (see above).
3. ~~Pilot~~ — done: HSPA8 in human, mouse and rat, CLU, and the CMA module.
4. Next cases — the clean, high-yield clusters:
5. ~~Caspase regulation~~: done in round 2, above.
6. ESCRT-I (protein transport to vacuole involved in ubiquitin-dependent protein catabolic process via the multivesicular body sorting pathway,
   10 rows, TSG101/VPS28/VPS37A–D/MVB12A/UBAP1). First check whether
   dmel_escrt_machinery can be generalized into a species-neutral ESCRT
   module rather than adding a parallel one.
7. ~~hair-cell mechanotransduction~~: the module
   hair_cell_mechanotransduction now exists (built from RGD PW:0001503),
   and round 6 kept its detection terms.
8. Authored combined terms — work through
   data/authored_precomposed.yaml, skipping label_only entries.
9. ~~GO:1901030/GO:1901029 (MOMP)~~: label-only, no decomposition (round 3).
10. ~~GO:0061739 (ATG8 lipidation)~~: done in round 3.
11. ~~GO:0098641 (cadherin binding)~~: done in round 4. The lateral
    GO:0098632 turned out to be precomposed by definition.
12. Parked: definitionally precomposed binding-plus-adhesion terms
    (GO:0098631 cell adhesion mediator activity, GO:0098632 cell-cell
    adhesion mediator activity and their children). The inventory's label
    match cannot find them, so they need a pass driven by logical
    definitions. CDH1's GO:0098631 IDA row is parked on GO:0098632 until
    then.
13. ~~GO:0106074 (aminoacyl-tRNA editing) and GO:0006271 (replication
    elongation)~~: kept as the right grain in round 5.
14. ~~GO:0050910 / GO:0050973 / GO:0050974 (mechanical detection in
    sensory perception)~~: kept in round 6.
15. ~~GO:1904747 (positive regulation of developmental apoptosis)~~:
    decomposed in round 6 for the C. elegans engulfment genes. Its
    remaining uses (egl-1 and Drosophila hid ACCEPT rows; ced-3, ced-4 and
    Drosophila Dark rows already moved laterally) wait for the full pass.
16. Next: GO:0035735 intraciliary transport involved in cilium assembly
    (10 uses, 8 genes) and GO:0003351 epithelial cilium movement involved
    in extracellular fluid movement (10 uses, 4 genes). The other sensory
    detection terms (chemical, light, temperature) are keep candidates, to
    be decided term by term by the round-6 test.
17. Full pass — walk data/annotations.yaml gene by gene, from class A to
    B, C and D. Check gocam_same_gene and gocam_same_symbol_other_species
    first, and confirm orthology before citing the symbol-matched hits.
18. Modules — only for mechanisms, per the module policy. Developmental
    contexts (sprouting angiogenesis, ureteric bud branching, endocardial
    cushion, cardiac conduction) get no new modules; their rows point at the
    existing signalling modules or drop the context.
19. Upstream — drafts are in
    upstream-tickets.md.
    Collect class D-iii terms and GO-CAM process nodes that
    still use precomposed BPs for go-ontology tickets. From round 1, file a
    ticket on GO:0000750: its automated replaced_by GO:0071507 moves
    non-kinase participants such as the Gpa1 GEF Arr4/Get3 onto the MAPK
    cascade. Ask for those annotations to go to GO:0062038 positive regulation
    of pheromone response MAPK cascade, or for the replaced_by to become
    consider. More generally, ask GO to check that a precomposed term's
    replaced_by holds for all annotated participants before obsoletion.

Reproducing

curl -sSL -o go.obo http://purl.obolibrary.org/obo/go.obo   # full edition: needs intersection_of
uv run python projects/INVOLVED_IN_PRECOMPOSITION/scripts/inventory.py \
    --go go.obo --out projects/INVOLVED_IN_PRECOMPOSITION/data

Outputs: go_terms.yaml (every "involved in" term, live or obsolete, with
its decomposition and obsoletion pointers), annotations.yaml (every review
row, with action, decomposition, GO-CAM and module hits),
authored_precomposed.yaml (combined terms that reviews introduced, as MODIFY
replacements, in core functions or as isoform-specific terms, and combined
terms used in module annotons; small, so committed),
module_candidates.yaml, and summary.yaml. The two large files,
go_terms.yaml and annotations.yaml (about 1 MB each), are git-ignored;
regenerate them locally before a full pass. Decompositions marked source: label_split come from splitting the label into two exact GO labels when GO
asserts no logical definition. They are a heuristic, not a GO assertion.
The script also reads the curated
keep_precomposed.yaml
(--keep), which lists terms judged the right grain; their rows and uses
are flagged kept.