RGD Pathway Ontology: GO Mappings, GO-CAM Completeness and Module Seeds

Bottom line: The Pathway Ontology (PW) from the Rat Genome Database (RGD)
has 2,703 live terms. It is curated against human, rat and mouse genes, and RGD
draws 200 of its pathways as interactive diagrams. We are using it three ways.
First, as a source of PW→GO mappings: 635 PW terms now have at least one
GO biological-process mapping used for matching, 19 of them from hand curation. Where PW's own definition cites a GO
term and our lexical match also finds one (397 terms), the two agree on 388.
Separately, 44 GO ids cited in PW definitions are obsolete and one no longer resolves. Second, as a
yardstick for GO-CAM completeness. Across the 144 diagram pathways that
describe normal biology, the median pathway has only 29% of its RGD-curated
human genes in any human GO-CAM. The best single GO-CAM covers a median of 2
genes of a pathway, and only 56 of the 144 have any human GO-CAM model whose
process falls under the mapped GO term. Third, as a source of module seeds. We parsed
all 200 RGD diagram pages into structured records (3,209 interaction edges,
1,715 gene links, 651 chemicals, 580 links to other pathways). These were
triaged against our 714 modules. Ninety already have a matching module, 14 are
ready to model, 9 have only a broad GO process term or none, and 56 are disease
or drug pathways and out of scope. Thirteen modules have now been derived from RGD
diagrams: the pilot srebp_signaling; the linked blood set coagulation_cascade,
fibrinolysis and kallikrein_kinin_cascade; a second set,
acetylcholine_signaling, dopamine_receptor_signaling,
visual_phototransduction and prostanoid_biosynthesis; and a third set,
protein_kinase_c_signaling, dna_cytosine_methylation_cycle,
reelin_signaling, trail_apoptotic_signaling and nmj_achr_clustering. All
thirteen pass schema and module validation.

Why this source

GO-CAM and our modules both need a pathway frame: which gene products work
together, in what order, toward which process. Reactome supplies that frame for
human. KEGG has licensing limits, and the UniPathway vocabulary is archived (see the
UniPathway project). PW is CC-BY, actively maintained (release
2026-05-16 used here), and curated by people at RGD. It also has three
properties the others lack:

• A separate pathway ontology with GO cross-references. Most PW definitions
  cite the GO process they correspond to. That makes PW a natural bridge for
  checking whether GO has a process term for a pathway, and whether the term
  is current.
• Hand-drawn diagrams with clickable regions. Each diagram's image map
  links genes, chemicals and sub-pathways and records interactions as
  A ---> B, A ---| B or A ---- B. That is close to the causal structure
  a module or GO-CAM needs, but the edge signs cannot be taken at face value:
  in the SREBP diagram, cholesterol and 25-hydroxycholesterol are drawn as
  activating SREBP processing, which they inhibit (see the pilot below).
• Curated gene membership with evidence. RGD's curated rows (TAS, IDA, IMP,
  ISO and others, 6,205 human rows, excluding RGD's own 33 IEA rows) are kept separate from bulk imports (KEGG IEA,
  SMPDB and PID, 23,480 human rows), so we can measure against the curated set.

Methods

Everything is regenerated by two scripts in
RGD_PATHWAY_ONTOLOGY/scripts/. No result is
hardcoded; the figures on this page are read from data/summary.yaml.

uv run python projects/RGD_PATHWAY_ONTOLOGY/scripts/fetch_rgd_pathways.py
uv run python projects/RGD_PATHWAY_ONTOLOGY/scripts/analyze_rgd_pathways.py

fetch_rgd_pathways.py downloads pw.obo, go-basic.obo, RGD's human and rat
PW gene annotation files and the HGNC symbol table into the git-ignored
.cache/rgd_pw/. It reads the list of diagram pathways from RGD's own pathway
selector page, fetches each diagram page (up to 3 MB of HTML each, 86 MB in all, cached and
not committed), and writes one compact record per pathway to
RGD_PATHWAY_ONTOLOGY/pathways/. A record
holds the title, RGD's free-text description, any GO and KEGG ids the page cites
as its "associated" terms, the diagram image URL, the image-map regions
(sub-pathway links, RGD gene links, chemicals normalized to ChEBI, PubChem,
KEGG or HMDB ids, interaction edges and unlinked labels) and the page's
references. Gene tables are not scraped from the HTML. The bulk annotation
files carry the same rows with evidence codes.

RGD returns HTTP 403 to the default python-requests user agent, so the
fetcher sends its own identifying user agent.

analyze_rgd_pathways.py writes four files to
RGD_PATHWAY_ONTOLOGY/data/:

Module matching uses three tests. A module that cites the PW id matches. A
module matches when at least two of its human members are in the pathway and
they make up at least half its members (mapped through HGNC). A module also
matches when it uses the mapped GO term or one of its descendants. Seeds are
then put in tiers:

Seeds with more than 100 curated genes are flagged broad: they should be
split along their sub-pathways before modelling.

Findings

PW→GO mappings

• Coverage is good for metabolism and named signaling pathways. 635 of
  2,703 live PW terms have at least one current GO process mapping used for
  matching. 19 of them come from the hand-curated set; the other 9 curated terms
  have only a broad GO term or none. 636 terms
  cite GO in their definition and 417 have a lexical match.
• PW's definition citations are reliable. Where both exist, the definition
  citation and the lexical match agree for 388 of 397 terms. In no case is the
  cited GO term an ancestor of the lexical match. This supports treating PW's
  definition citations as closer to exact than relatedMatch, after spot
  checks. The GO ids typed on RGD's diagram pages are less reliable. 4 of 45
  disagree with the other sources for the same pathway, and three of these look
  wrong: urea cycle is cited as GO:0000076, a DNA replication term whose local id
  repeats the PW id; ERK5 signaling as the JNK cascade; and Toll-like receptor
  signaling as the Drosophila Toll pathway. The fourth, de novo pyrimidine
  synthesis cited as 'de novo' UMP biosynthesis, is defensible, since the
  mammalian route ends at UMP; it is not in an is_a chain with the lexical
  match, so it is flagged as non-equivalent. The script demotes all four to
  relatedMatch and leaves them out of GO-CAM and module matching.
• PW cites retired GO terms. 44 definition citations point at obsolete GO
  terms, and one at an id that no longer resolves. Several are amino-acid "metabolic process" terms that GO has since
  obsoleted, such as cysteine, histidine, tyrosine, phenylalanine and
  tryptophan metabolic process. PW:0000543 (protein kinase A (PKA) signaling
  pathway) cites GO:0010737, now "obsolete protein kinase A signaling", both in its definition
  and on its diagram page. GO names no replacement, but the current term is
  GO:0141156 cAMP/PKA signal transduction. These can go back to RGD as a
  concrete update list. In the SSSOM file the 45 obsolete citation sites (44
  distinct PW→GO pairs) carry pw2go:qa/OBSOLETE_TARGET and the unresolvable
  one pw2go:qa/MISSING_TARGET; 12 of the 45 list GO's replacement in
  see_also.
• Some well-known pathways have no GO process match by label. Examples are
  p53 signaling, AMPK signaling, PKA signaling, hypoxia-inducible factor,
  angiotensin II and vasopressin signaling, renin-angiotensin, and
  mitochondrial and cytosolic iron-sulfur cluster assembly. Some do have a GO
  term under another name, for example GO:0072331 (signal transduction by p53
  class mediator) and GO:0141156 (cAMP/PKA signal transduction). All 28 such
  diagram pathways are now mapped by hand; see Curated GO mappings below.

GO-CAM completeness

GO-CAM coverage of these pathways is scattered rather than pathway-shaped. The
144 in-scope diagram pathways have a median of 29% of their RGD-curated human
genes in some human GO-CAM, and 42 reach half. Those genes are spread over many
unrelated models. The best single model covers a median of 2 genes, and the
median share of genes in GO-CAMs about the mapped process is 0. That last
number undercounts modelled pathways (see the caveat below the table). Examples:

The two columns fail in opposite directions, so neither is a completeness
measure on its own. A gene that appears in a GO-CAM about something else does
not make the pathway modelled, so the "any GO-CAM" column overcounts. The
process-level column only counts models in which some activity is annotated to
the mapped GO process or a descendant. GO-CAMs usually annotate each activity to
the step it performs, not to the whole pathway, so this column undercounts.

The SREBP row shows the undercount. The human GO-CAM 6205c24300000880
contains six of the seven curated genes and models the pathway step by step.
Its activities carry step-level process terms: GO:0036316 (SREBP-SCAP complex
retention in ER), GO:0051604, GO:0045542, GO:0006607, GO:0046889 and
GO:1905552. None of these is under the mapped term GO:0032933 (SREBP
signaling pathway), so the pathway scores 0%. A 0 in the process-level column
therefore does not tell an unmodelled pathway apart from one modelled with
step-level terms. best_gocam_overlap, read together with the model title,
separates the two cases for SREBP. Both columns count human GO-CAMs only. The
Wnt and citric acid cycle zeros have other causes. The only GO-CAM carrying the
Wnt term is a Drosophila model, so canonical Wnt signaling has no human model
at all. The one human model carrying the TCA-cycle term contains none of RGD's
curated genes (below).

The citric acid cycle row also shows a limit of the gene basis. Eleven GO-CAMs
carry the TCA-cycle term: one mouse, one fly, eight fission-yeast and one human
model (SIRT3 regulation, with CS and MDH2). The human model contains none of the
11 RGD-curated human genes. RGD's curated human set lists the succinyl-CoA, succinate dehydrogenase,
fumarase and 2-oxoglutarate dehydrogenase steps but omits CS, ACO2, the IDH3
subunits and MDH2, which reach the human term only through the KEGG and SMPDB imports.
The curated basis avoids KEGG's over-broad lists, but it can also miss core
members. The completeness table therefore reports both counts
(n_genes_rgd_curated, n_genes_all).

Module seeds

Nine of the 10 candidates of the 2026-10-20 triage were modelled in the second
and third sets (below). The tenth, IL-10 signaling, left the queue through a
gene-overlap match and still has no module:

IL-10 signaling
(PW:0000515) now matches il22_signaling, which merged on main meanwhile, by gene
overlap only (4 of its 7 genes are in the pathway). IL-10 got its own module in the sixth set.

Re-running the triage after merging main on 2026-10-18 moved six former candidates. Five now
match only by GO term to Drosophila or plant modules: RNA polymerase I, II and III
transcription, mitochondrial protein import, and the microRNA pathway. The sixth, insulin
secretion, now has a mammalian module. The microRNA pathway was modelled in the fifth set;
the RNA polymerase pathways are still worth modelling for mammals.

Several related_module_only seeds were mammalian pathways whose only module
match was a bacterial one: nucleotide excision repair, mismatch repair,
homologous recombination and non-homologous end joining. All four now have
human modules (fifth set).

Two caveats apply to the has_module tier. A gene-overlap match needs only
half of a module's genes in the pathway, so a broad pathway such as extrinsic
apoptosis also matches the NF-kB and NLR modules. A has-module call means a
related module exists, not that the module covers the whole RGD pathway. A PW id
counts as cited only when a module lists it as an evidence source; before
2026-10-26 any mention counted, so a pathway a module's notes named as out of scope
(kinin signaling, PW:0001480, in kallikrein_kinin_cascade) was miscounted as
covered.

Curated GO mappings

The 28 diagram pathways that had no GO process mapping were mapped by hand in
data/pw_go_curated.sssom.yaml,
an SSSOM set with one comment per row. The script merges it into the generated set
under pw2go:src/CURATED_MAPPING. It stops with an error if a curated GO id is
missing, obsolete, not a biological process or mislabelled. Only exactMatch,
closeMatch and narrowMatch rows are used for GO-CAM and module matching. A
broad term such as GO:0007186 G protein-coupled receptor signaling pathway would
otherwise match every module under it.

The 14 new candidates, all modelled in the fourth set below:

For hormones without a receptor signaling pathway term, GO uses cellular response to <ligand> stimulus. That branch covers vasopressin, CRH, PTH and
glucagon, and it is also the process of the mouse glucagon GO-CAM
636d9ce800000147. In GO a receptor signaling pathway sits under the cellular
response to its ligand, so the response term is strictly the broader one. These
rows still use closeMatch, because a PW "signaling pathway" includes the
downstream response. The V2 vasopressin branch maps to the same GO:1904117
with broadMatch, because it is one receptor's share of that response.

The ATM row's GO:0000077 DNA damage checkpoint signaling also covers
ATR-dependent checkpoints, so one of the four GO-CAMs it matches
(65039e8700000123, ATR activation) is an ATR model, not an ATM one.

PKA signaling and beta-adrenergic signaling should extend
gpcr_camp_pka_signaling rather than start new modules. That module is a compact
five-gene module (ADRB2, GNAS, ADCY5, PRKACA, CREB1) grounded in GO:0007188, so
the triage cannot link it to either pathway.

The five related_module_only pathways are spliceosome, mRNA nuclear export and
p53 signaling, whose only matches are Drosophila or C. elegans modules, and
calcium-mediated signaling (matched by t_cell_receptor_signaling). The fifth,
renin-angiotensin, matched kallikrein_kinin_cascade only because that module's
notes discuss GO:0002003 angiotensin maturation. It is now has_module by gene
overlap with the angiotensin II module, which does not model the cascade (see the
fourth set).

The 9 no_specific_go_term pathways show where GO is coarser than PW. For the
receptor pathways both GO branches were checked: ... receptor signaling pathway and cellular response to <ligand> stimulus.

• Receptor pathways.
• Bradykinin has neither kind of term. GO:0140162 (venom-mediated vasodilation
  by activation of bradykinin receptor signaling pathway) names a bradykinin
  pathway term that does not exist, and the kinin cascade terms are kinin
  generation, not signalling. Kinin signaling therefore maps only to
  GO:0007186.
• The calcium-sensing receptor has only GO:0071277 cellular response to
  calcium ion, which covers every calcium response.
• The V2 vasopressin branch and angiotensin II via AT2 are receptor subtypes of
  pathways that do have terms (GO:1904117, GO:0038166). GO does not split
  those terms by receptor.
• CaMKII signaling. GO obsoleted GO:0099004 calmodulin dependent kinase
  signaling pathway. The term it suggests instead, GO:0061762, is the
  CAMKK-AMPK cascade, so CaMKII signaling maps only to GO:0019722. (ATM
  signaling maps broadly to GO:0042770, and narrowly to GO:0000077 DNA damage
  checkpoint signaling, so it is a candidate.)
• Iron-sulfur assembly. Mitochondrial and cytosolic iron-sulfur cluster
  assembly both map only to GO:0016226, because GO does not separate the two
  machineries.
• Others. Titanium dioxide nanoparticle response maps to GO:1990267.
  Histone modification has no process term at all (sssom:NoTermFound): GO
  obsoleted GO:0016570 histone modification and its children in favour of
  molecular functions under GO:0140993 histone modifying activity.

These are candidates for GO term requests, or for modules grounded on molecular
functions rather than on a process term. The V2 vasopressin and AT2 branches have
since left the tier, because the fourth-set modules cite them as evidence; 7 remain.

Pilot: SREBP signaling

modules/srebp_signaling.yaml was
derived from the RGD SREBP diagram,
PW:0000753. The diagram's
seven gene products and edges map directly onto five module steps:

1. sterol sensing and ER retention (SCAP, INSIG1, INSIG2)
2. SCAP-escorted COPII transport to the Golgi
3. site-1 cleavage (MBTPS1)
4. site-2 intramembrane cleavage (MBTPS2)
5. nuclear transcription (SREBF1, SREBF2)

A feedback edge runs from step 5 back to INSIG1. The RGD diagram supplied the
boundary and topology. The molecular functions were then chosen from each
protein's experimental GO annotations, not from the diagram. The edge types
needed curator judgment. RGD's ----> is "positively affects", which covers
causation, provision of a substrate and regulation alike, and the sign can be
wrong. The diagram draws cholesterol and 25-hydroxycholesterol as positively
affecting "Srebf at Golgi", but sterols block SREBP transport and
processing. The module models that link as NEGATIVELY_REGULATES. Anyone
reusing diagram_edges from the other 199 records should check every sign
against the literature.

After the first pass, each step was given quote-checked support from the primary papers that defined it (13 PubMed abstracts). It was also cross-checked against an independent human GO-CAM, 6205c24300000880, which models six of the seven gene products with matching molecular functions (it uses the parent GO:0000981 for SREBF1/2, where the module uses GO:0001228). The model also contains AKT1 and PCK1, which the module leaves out: their reported effect on INSIG opposes retention and rests on one cancer-cell study (corrected separately in ai4curation/ai-gene-review#5055). That model omits SCAP, and it includes KPNB1 for SREBP nuclear import, a step the RGD description mentions but the module does not yet model. A curator-built GO-CAM for the same pathway is the strongest check available on a diagram-derived module, so the seeds' best_gocam field is worth reading before modelling the next candidates. It is a lead, not a ranking. For most pathways no model carries the mapped process, and ties between equally overlapping models then go to the lowest model id.

The pilot shows a limit of the route: the diagram has no step-level GO terms,
so every activity needs separate grounding. SCAP's escort role has only the
generic cargo receptor activity (GO:0038024), which the module now uses along
with sterol sensor activity (GO:0032935), matching SCAP's gene review. Reviews of
all seven member genes merged in ai4curation/ai-gene-review#5215.

Blood set: coagulation, fibrinolysis, kallikrein-kinin

The next three candidates were modelled together, because their RGD diagrams
link to each other
(PW:0000474,
PW:0000481,
PW:0000526). Members shared
between diagrams were given one home each:

• modules/kallikrein_kinin_cascade.yaml
  (GO:0002254 kinin cascade) owns contact activation. Factor XII and plasma
  kallikrein activate each other, and kallikrein releases bradykinin from
  high-molecular-weight kininogen. Tissue kallikrein releases kallidin. The
  module also covers the B1 and B2 receptors, the kininases and the inhibitors
  C1-inhibitor, alpha-2-macroglobulin and kallistatin.
• modules/coagulation_cascade.yaml
  (GO:0072378 blood coagulation, fibrin clot formation) starts where contact
  activation hands off, at factor XI activation. Its parts are extrinsic
  initiation, intrinsic amplification, thrombin feedback, prothrombinase,
  fibrin formation, factor XIIIa cross-linking, and the natural anticoagulants:
  antithrombin, TFPI, the protein C/protein S/thrombomodulin system and protein
  Z/ZPI.
• modules/fibrinolysis.yaml (GO:0042730)
  takes the cross-linked clot as its input. It covers plasminogen activation by
  tPA and uPA, plasmin degradation of fibrin, and inhibition by PAI-1 and the
  other serpins, alpha-2-antiplasmin and TAFI.

The triage re-run counts four seeds as has_module that were not before: the
three pathways and the protein C anticoagulant pathway (PW:0000480), which
coagulation_cascade declares as an evidence source because its protein C part
holds five of that pathway's six genes (all but PROCR). Kinin signaling
(PW:0001480), which an earlier count included, is outside kallikrein_kinin_cascade
and went back to needs_go_mapping; it is now curated to a broad GO term (see
Curated GO mappings).

Most of the curation work was in the non-enzymatic members, because the RGD
diagrams draw them as plain nodes:

• Substrates. Kininogen is the substrate the kallikreins cut. It gets only
  the hormone activity of the released kinin, which its own experimental
  annotations and the human GO-CAM 6246724f00000632 both use. It gets no
  kallikrein-kinin process term. A comparator check in QuickGO found no
  kininogen among the annotations to either kallikrein-kinin cascade term.
• Fibrinogen. In coagulation fibrinogen is thrombin's substrate, but it
  polymerises itself, so it gets structural molecule activity and fibrin
  polymerisation. In fibrinolysis, fibrin is the template that speeds up its own
  lysis (the cofactor case in CLAUDE.md). GO has no molecular-function term for
  that template role, so the participant is given processes and no molecular
  function.
• Protease cofactors. Tissue factor and factors Va and VIIIa get
  endopeptidase activator activity, and protein S peptidase activator activity.
  These are curator judgments, because these proteins have no experimental
  molecular-function annotation that fits. Protein Z and the factor XIII B
  subunit are left without a molecular function rather than given one GO cannot
  support.

GO-CAM support is thin. Curated models cover protein C activation (THBD, F2,
PROC: 666b894f00000232), bradykinin degradation and receptor binding (ACE,
KNG1, BDKRB2: 6246724f00000632) and plasminogen's role in matrix breakdown
(PLAU, PLAUR, PLG: 68fac5ed00001527). The procoagulant steps, tissue
kallikrein and every inhibitor step have no curated model. Of the roughly 50
member gene products, only VWF, A2M and APOH have gene reviews.

Second set: acetylcholine, dopamine, phototransduction, prostanoids

Four more candidates were modelled, one module each:

• modules/acetylcholine_signaling.yaml
  (PW:0000841, with its
  nicotinic and muscarinic sub-pathways PW:0000842 and PW:0000843) runs from
  choline uptake (SLC5A7) and ACh synthesis (CHAT) to reception and hydrolysis
  (ACHE). Vesicular loading and release are references to
  synaptic_vesicle_neurotransmitter_loading and synaptic_vesicle_exocytosis.
  Reception is a choice between five nicotinic receptor subtypes and the two
  muscarinic G-protein arms, M1/M3/M5 to Gq/11 and M2/M4 to Gi/o. GRK2/GRK5 and
  RGS2/RGS4/RGS6 terminate the signal. RGD's agrin, LRP4, MuSK, DOK7 and rapsyn
  genes are left out: they build the neuromuscular junction rather than transduce
  the signal, and GO keeps them under AChR clustering (GO:0071340), not
  acetylcholine signaling. They are modelled separately in nmj_achr_clustering
  (third set).
• modules/dopamine_receptor_signaling.yaml
  (PW:0000394) is the striatal
  instance of gpcr_camp_pka_signaling. D1-like receptors act through Golf and
  ADCY5, then PKA and DARPP-32, which inhibits PP1; D2-like receptors inhibit
  ADCY5 through Gi/o. Regulatory parts cover the DARPP-32 phospho-switch (CDK5,
  PP2A, calcineurin, CK1), GRK/arrestin/RGS9-2 termination, STEP and the
  beta-arrestin 2/PP2A/AKT arm. Dopamine synthesis stays in
  catecholamine_biosynthesis. SRC is left out as peripheral, and the D1-D2
  heteromer Gq arm is recorded as contested rather than asserted.
• modules/visual_phototransduction.yaml
  (PW:0000962) gives each step,
  from opsin through transducin, PDE6, the CNG channel and the NCKX exchanger,
  a rod and a cone variant. Recovery covers GRK1/GRK7, the arrestins, recoverin,
  the RGS9 GAP complex and cGMP resynthesis by retGC and the GCAPs, and the retinoid
  cycle is a reference to retinoid_visual_cycle. RGD's human list has two
  changes. CIMIP3 is dropped: it is a sperm-axoneme protein whose alias GUCA1ANB
  ("GUCA1A neighbor") probably put it there. GUCA1A is added, because RGD's own
  diagram draws Guca1a acting on the retinal guanylate cyclase.
• modules/prostanoid_biosynthesis.yaml
  (PW:0001243) is a reaction
  chain grounded in Rhea and ChEBI. PTGS1 or PTGS2 makes PGH2, which branches to
  PGE2, PGD2, PGI2, TXA2 and PGF2alpha through the terminal synthases. AKR1A1 is
  left out: nothing but RGD's own text supports a prostaglandin reaction for it,
  so it is recorded as a knowledge gap. Arachidonate release by cPLA2, the
  receptor sub-pathways and prostanoid inactivation are out of scope.

These modules also turned up three possible upstream errors:

• In the human cone-opsin GO-CAMs for OPN1SW and the OPN1MW genes, the PDE6C
  activity is part of GO:0070966 (no-go mRNA decay). The OPN1LW model places the
  same activity in GO:0141190 (transducin-mediated opsin signaling pathway).
• PRXL2B's IBA/ISS term GO:0047017 is defined by PGD2 reduction, but the
  primary study of mouse and pig PRXL2B (PMID:18006499) found that PGD2 is not a
  substrate. Its own reaction (RHEA:28214) has no GO molecular-function term.
• RCVRN's TAS row for guanylate cyclase activator activity (GO:0008048) cites the
  1991 paper behind a role recoverin is now thought not to have. That paper is not
  cached, so this is a question rather than a finding.

GO-CAM coverage is uneven. Curated human models cover the full rod and cone
activation cascades (6918f23700000035, 68fac5ed00003872 and the other cone
opsin models), each dopamine receptor's G-protein step and several nicotinic and
muscarinic receptors. No model covers phototransduction recovery, the DARPP-32
switch or any prostanoid synthase.

Third set: PKC, DNA methylation, reelin, TRAIL, NMJ clustering

The last four candidates were modelled, plus one module that the acetylcholine
work split off:

• modules/protein_kinase_c_signaling.yaml
  (PW:0000840) models the PKC
  life cycle. The steps are HSP90/CDC37 chaperoning, priming by PDPK1 (with
  mTORC2, which the RGD text names), activation by isoform class, anchoring by
  RACK1 and PICK1, and dephosphorylation by PHLPP1/2. The classes are a variant
  set: conventional PKCs need DAG and calcium, novel PKCs need DAG only, and
  atypical PKCs sit on the PAR3/PAR6 scaffold. The second-messenger step is a
  reference to gpcr_plc_calcium_pkc_signaling. RGD's substrate and cross-talk
  genes (FSCN1, BTK, LCK, FYN, RASA1) are left out, and LDB3, PCNT and COPB1 are
  described in prose only.
• modules/dna_cytosine_methylation_cycle.yaml
  (PW:0001340) covers de novo
  methylation (DNMT3A/B, with DNMT3L), UHRF1-directed maintenance by DNMT1,
  active demethylation by TET1-3 and TDG, and the methyl-CpG readers. TDG and
  DNMT3L are not on RGD's list; RGD's own text and the GO definition of active
  demethylation support adding them. Left out: the transcription factors on
  RGD's list, the CXXC-domain histone enzymes (they belong to the
  histone-modification sub-pathway), MBD4 (a mismatch glycosylase) and UHRF2.
  By gene overlap alone it also matches the broad chromatin
  modification/remodeling pathway (PW:0001337), the loose-match case described
  above.
• modules/reelin_signaling.yaml
  (PW:0000388) runs from RELN
  binding VLDLR and LRP8 to DAB1 phosphorylation by FYN and SRC, then the PI3K
  and LIS1 branches. DAB1 degradation by the SOCS7/CUL5 ligase ends the signal.
  The NMDA and AMPA receptor subunits are downstream targets, not participants.
  The integrins and the PAF-AH catalytic subunits are context.
• modules/trail_apoptotic_signaling.yaml
  (PW:0000682) is the
  concrete TRAIL instance of death_receptor_apoptotic_signaling. TRAIL binds
  DR4 or DR5, which assemble a DISC of FADD and caspase-8, held back by c-FLIP.
  Caspase-10 and the decoy receptors are optional parts. The direct caspase
  route and the BID-to-mitochondria route are a variant set that refers to the
  intrinsic and execution modules. TRADD is left out, because studies of the
  endogenous DISC find it is not recruited.
• modules/nmj_achr_clustering.yaml
  covers the clustering arm of PW:0000842: neural agrin, LRP4, MuSK, DOK7 with
  CRK/CRKL, and rapsyn. The muscle receptor is the thing clustered, not a
  participant. No receptor subunit carries GO:0071340, which agrin, LRP4, MuSK
  and rapsyn do.

These modules turned up more possible upstream errors:

• RGD. The PW:0000840 text says PDPK1 phosphorylates "the activation loop of
  PLC"; it should say PKC.
• GOA.
• PRKCZ, an atypical PKC, has an EXP annotation to calcium,diacylglycerol-
  dependent kinase activity (GO:0004698). PRKCD, a calcium-independent novel
  PKC, has an IDA one.
• TET1-3 lack GO:0141167 active DNA demethylation, which only TDG carries,
  although the term's definition starts with TET oxidation.
• GO-CAM, c-FLIP. 685de18700006529 types c-FLIP with cysteine-type
  endopeptidase activity, but its caspase-like domain lacks the catalytic
  cysteine.
• GO-CAM, TRAIL models. Both use generic receptor activity for DR4/DR5
  instead of GO:0036463 TRAIL receptor activity, and add a TRADD step that
  rests on an overexpression study.
• GO-CAM, reelin.
• Neither human reelin model includes VLDLR.
• In 66187e4700002704, FYN's evidence is a paper on Nogo-A
  (PMID:16979591), and CUL5's is a CUL5-ARIH2 structure paper (PMID:34518685).
  Neither is about reelin or DAB1.
• GO-CAM, NMJ. The mouse model 64c3044100000089 gives rapsyn the inner-ear
  alpha9 receptor (Chrna9) as input, not the muscle receptor.
• GO-CAM, DNA methylation. 6882d2b800000042 makes HLTF's H3K23
  ubiquitination part of maintenance methylation. Its source, PMID:40680746,
  raises a role in DNMT1 recruitment only as a possibility.

Fourth set: the 14 curated-mapping candidates

The 14 candidates from the curated mappings became 13 modules. Angiotensin II
signaling and its AT1 and AT2 branches share one module, and so do vasopressin
signaling and its V2 branch. The triage re-run moves all 14 to has_module, and
the AT2 and V2 branches leave no_specific_go_term, because the modules cite
them as evidence.

• Hormone receptors.
• modules/angiotensin_ii_receptor_signaling.yaml
  (PW:0000244,
  PW:0000527,
  PW:0000528) treats the
  two receptors as a variant set off one ligand. AT1 signals through Gq/11-PLC
  and JAK2/TYK2, with GRK/beta-arrestin desensitisation and AGTRAP/ARAP1
  trafficking. AT2 signals through Gi-linked SHP-1/MKP-1 phosphatases and
  MTUS1/ZBTB16. AGT is only the ligand source: the proteases of the cascade
  belong to a separate renin-angiotensin pathway, PW:0000492, which this module
  does not model.
• modules/vasopressin_receptor_signaling.yaml
  (PW:0000491,
  PW:0000713) has a V1a/V1b
  Gq/11-PLC arm and a renal V2 arm. The V2 arm runs through Gs, ADCY6, cAMP and
  AKAP-anchored PKA to AQP2 exocytosis with STX3/STX4. AQP2 is a participant,
  because it performs the water transport.
• Four Gs-coupled modules share one structure: ligand, receptor, Gs, then
  adenylyl cyclase and PKA by reference to gpcr_camp_pka_signaling, then
  GRK/beta-arrestin termination.
  modules/crh_receptor_signaling.yaml
  (PW:0000493) covers CRH
  and the urocortins on CRHR1/CRHR2.
  modules/beta_adrenergic_signaling.yaml
  (PW:0000794) covers
  ADRB1/ADRB2, including the PKA-dependent switch to Gi.
  modules/pth_receptor_signaling.yaml
  (PW:0001135) adds the PTH1R
  Gq arm, by reference to gpcr_plc_calcium_pkc_signaling, and NHERF/ezrin
  control of coupling.
  modules/glucagon_receptor_signaling.yaml
  (PW:0000676) covers GCGR
  with ADCY5/ADCY6.
• Signalling cassettes and sensors.
• modules/camp_pka_signal_transduction.yaml
  (PW:0000543) is the shared
  cAMP/PKA core: ADCY1-10, the PKA holoenzymes, AKAP anchoring (including EZR)
  and PDE termination. The receptor step refers to gpcr_camp_pka_signaling.
  EPAC, CNG channels and the PKA substrates are out of scope.
• modules/ampk_energy_sensing.yaml
  (PW:0000542) has the AMPK
  heterotrimer with family subunit choices, LKB1 and CAMKK upstream kinases
  and the Thr172 phosphatases as variant sets, and the TSC1-TSC2 GAP relay to
  mTORC1. The phosphorylated substrates (raptor, ACC1/2, TBC1D1, HMGCR) are
  outputs, and CPT1A is a reference to carnitine_shuttle.
• modules/hif_oxygen_sensing.yaml
  (PW:0000360) models all 20
  RGD genes: the EGLN and HIF1AN dioxygenases as the oxygen sensors,
  VHL-mediated degradation, and HIF-alpha/ARNT transcription with
  CREBBP/EP300.
• modules/hair_cell_mechanotransduction.yaml
  (PW:0001503) runs from the
  CDH23/PCDH15 tip link and the USH1 upper density to the TMC1/TMC2 channel,
  MYO1C adaptation and the VGLUT3/otoferlin ribbon synapse. The channel
  subunits TMC1, TMC2, LHFPL5, TMIE and CIB2 are not on RGD's list; without
  them the module has no transducer. Twelve bundle-building genes are left
  for a stereocilium-organization module, and POTEI is a probable orthology
  artefact from rat actin.
• DNA damage, NAD and glucose uptake.
• modules/atm_dsb_signaling.yaml
  (PW:0001361) covers MRN
  sensing, KAT5-dependent ATM activation, the H2AX-MDC1-RNF8-RNF168 chromatin
  cascade and the BRCA1-A complex. The BRCA1-A subunits UIMC1, ABRAXAS1,
  BABAM1, BABAM2 and BARD1 are added. CHK2 and p53 are outputs.
• modules/nad_consumption.yaml
  (PW:0002581) groups the
  three NAD+-cleaving enzyme classes, sirtuins, PARPs and CD38/BST1/SARM1, and
  cites the three sub-pathways PW:0002582-PW:0002584. ZC3HAV1, a catalytically
  inactive PARP, is left out.
• modules/glut4_insulin_stimulated_glucose_uptake.yaml
  (PW:0000639) takes AKT2
  from insulin_receptor_signaling. It then models the TBC1D4/TBC1D1 RabGAPs
  and DENND4C, a RAB10 (adipocyte) or RAB8A/RAB13 (muscle) choice, vesicle
  motors, an optional RAC1-PAK1-ARP2/3 actin step, STX4/SNAP23/VAMP2 fusion
  and SLC2A4. Of RGD's 54 genes, AKT1, IRAP (LNPEP, which is cargo) and the
  contested CAP-CBL-TC10 arm are left out; the last is a knowledge gap.
  GO:0044381 is placed only on SLC2A4, since no trafficking protein carries
  it in human, mouse or rat GOA.

Two triage observations came out of this set. PW:0000492 renin-angiotensin
cascade is now has_module only because all 21 angiotensin II module genes are
in its gene list, although the module leaves its proteases out. This is the
loose gene-overlap case described above. The cascade was modelled in the fifth set.
The best_gocam leads for PW:0000639 (63d320cd00001536, RAC1 degradation in
migration) and PW:0000244 (67f85f2b00004426, a C5a receptor model) are
gene-overlap ties with no relation to the pathway.

Possible upstream errors found in this set:

• GOA, catalytic terms on non-catalytic proteins.
• STRADA has GO:0004672 protein kinase activity by IDA from PMID:12805220,
  which calls STRAD a pseudokinase.
• PRKAB2, a non-catalytic AMPK subunit, has GO:0004679 by IDA.
• CAMKK2 has GO:0004713 protein tyrosine kinase activity by TAS.
• Most mono-ADP-ribosyltransferase PARPs (PARP3, 4, 6, 8, 10-12, 14-16,
  TIPARP) carry GO:0003950 poly-ADP-ribosyltransferase activity by IBA and
  sometimes IDA. This looks like a PAINT node at the family root and is worth
  an IBA review.
• GOA, conflicts and odd rows.
• SAT2 has both enables and NOT|enables GO:0004145 by IDA.
• SIRT7 has protein methyltransferase activity (GO:0008276, IDA). The
  module does not adopt it.
• PARP14 has an IEA row for GO:0019677 that looks over-propagated.
• GO:1904117 cellular response to vasopressin has a CDC6 IEA row.
• GOA, missing terms. In each case below the term is missing in human GOA.
• SNX13 has no GTPase activator activity, though its Gs-alpha GAP activity is
  annotated as a process.
• ATM has GO:0035979 only by ISS/IEA.
• RIF1, TBC1D4, TBC1D1, STX4, VAMP2 and MYO1C have no experimental molecular
  function.
• AQP2 has no GO:0070295 renal water absorption.
• No oxygen sensor (EGLN1-3, VHL, HIF1AN) carries GO:0097411. This may be
  the convention of annotating sensors to regulation of the pathway.
• CD38/BST1 have no GO:0019677.
• GO-CAM.
• 68d5ebd600000330, a GCGR-to-Gi model, gives the GCGR activity
  adenylate cyclase-activating GO:0007189 while its GNAI1 activity uses the
  inhibiting GO:0007193.
• In 617b481400000460, the KAT5 node is typed as bare protein binding.
• In 65a1f4f800000487, H2AX regulates NBN directly and skips MDC1.
• In 696022cd00000603, a SIRT3 activity is part of mitochondrial fusion,
  which does not fit the model's TCA-cycle title.
• PTH1R (67c10cc400004508) and CRHR2 (69729a3800007299) are typed with
  generic receptor terms although specific ones carry IDA rows.
• GO term placement. GO:0141156 cAMP/PKA signal transduction is carried in
  human GOA by effectors and three AKAPs, but by no cyclase, PKA subunit or PDE.
  The module places it on the cyclase, PKA and AKAP steps, reading the absence as
  the term being new.
• Coverage gaps. No GO-CAM models the hair-cell channel, the oxygen
  sensors, the AT1 JAK/STAT arm, the AT2 branch or GLUT4 translocation. Most
  members of these modules have no gene review; the modules' notes list them.

Fifth set: repair, RNA, p53, retinoids, the RAS cascade, Fe-S and PCP

The fifth set modelled 11 mammalian pathways whose only match was a GO term
on a bacterial, Drosophila or plant module, plus the renin-angiotensin cascade.
That cascade had counted as covered only by gene overlap. Each module cites one
PW id as evidence; mRNA nuclear export also cites its NXF1-NXT1 child,
PW:0001625. The triage re-run moves all 11 to has_module. PW:0002635,
mitochondrial iron-sulfur cluster biogenesis, moved there too, through
mitochondrial_iron_sulfur_cluster_assembly, which merged on main
meanwhile.

• DNA repair.
• modules/nucleotide_excision_repair.yaml
  (PW:0000130) has two
  recognition routes as a variant set. Global-genome repair goes through
  UV-DDB and XPC-RAD23B-CETN2. Transcription-coupled repair goes through
  stalled RNA polymerase II, CSB, CRL4-CSA, UVSSA and ELOF1. The next steps
  are TFIIH opening, XPA/RPA verification, dual incision by XPF-ERCC1 and
  XPG, and repair synthesis and ligation. It models 36 of 39 RGD genes and
  adds 9 that the steps need. The CDK-activating kinase is left out, because
  it leaves TFIIH during repair.
• modules/eukaryotic_mismatch_repair.yaml
  (PW:0000662) follows the
  reaction reconstituted from purified human proteins. MutSalpha or MutSbeta
  recognises the mismatch, RFC loads PCNA, MutLalpha nicks, EXO1 excises,
  RPA protects the gap, and Pol delta and LIG1 fill and seal it. MutLbeta and
  MutLgamma are context.
• modules/homologous_recombination_dsb_repair.yaml
  (PW:0000202) refers to
  atm_dsb_signaling for the BRCA1-A arm. It then runs from short- and
  long-range resection through RPA coating, BRCA1-PALB2-BRCA2 RAD51 loading
  and strand invasion to junction dissolution or resolution, the last as
  variant sets.
• modules/nonhomologous_end_joining.yaml
  (PW:0000203) is classical
  NHEJ: Ku, DNA-PKcs, optional end processing and gap filling, and
  XRCC4-LIG4 with XLF and PAXX. It takes 53BP1-RIF1 from the ATM module.
  Alternative end joining has its own GO term and is context. CtIP, which RGD
  curates here, belongs to the resection step of the HR module.
• Signalling.
• modules/p53_signaling.yaml
  (PW:0000718) is the p53
  regulatory circuit: stress-kinase phosphorylation, MDM2/MDM4 feedback, ARF
  inhibition of MDM2, acetylation and SIRT1 deacetylation, and transactivation.
  The p53 target genes are outputs under the substrate-vs-participant rule;
  MDM2 is the exception, because it does work in the loop. Nine of the 40 RGD
  genes are participants. The apoptosis and G1-arrest outputs are references
  to existing modules.
• modules/wnt_planar_cell_polarity.yaml
  (PW:0000200) takes ligand
  secretion and Frizzled reception from wnt_signaling. It then models the
  ROR2/PTK7 co-receptors, the core PCP proteins, and the DAAM1-RHOA-ROCK and
  RAC1-JNK arms. It avoids the Wnt/PCP terms GO plans to obsolete
  (ai4curation/ai-gene-review#5781).
• modules/renin_angiotensin_cascade.yaml
  (PW:0000492) is the
  proteolytic cascade: renin with ATP6AP2, ACE or chymase or cathepsin G, the
  ACE2/neprilysin/PRCP arm to angiotensin-(1-7) on MAS1, and ENPEP and ANPEP.
  The receptor step is a reference to angiotensin_ii_receptor_signaling.
  Angiotensinogen is the proteases' substrate and the hormone source, and it
  carries no GO:0002003, following the AGT worked example in CLAUDE.md.
• RNA and metabolism.
• modules/mirna_biogenesis_and_silencing.yaml
  (PW:0000808) runs from the
  Microprocessor, XPO5 export and DICER1 with TARBP2/PRKRA to Argonaute
  loading, TNRC6, CCR4-NOT and PAN2-PAN3 deadenylation, decapping and XRN1.
  The miRNA editing and uridylation enzymes are context.
• modules/mrna_nuclear_export.yaml
  (PW:0001587) has TREX
  loading, NXF1-NXT1, optional TREX-2, the pore, DDX19-GLE1 release and an
  optional CRM1 route. It models the pore inline, because no human pore
  module exists.
• modules/retinoic_acid_metabolism.yaml
  (PW:0001004) runs from
  RBP4-STRA6 delivery through the RDH10-DHRS3 pair and ALDH1A1-3 to CRABP
  binding and CYP26 catabolism. Ester storage is optional and refers to
  retinoid_visual_cycle. It does not cite PW:0000141 retinol metabolism:
  that pathway is intestinal absorption and shares only 2 of its 11 genes
  with this module.
• modules/cytosolic_iron_sulfur_assembly.yaml
  (PW:0002636) covers all
  nine RGD genes. The NDOR1-CIAPIN1 electrons feed the NUBP1-NUBP2 scaffold
  and the CIAO1-CIAO2B-MMS19 targeting complex, with the mitochondrial module
  upstream. GO has no process term for the cytosolic machinery, so the
  module uses GO:0044572 [4Fe-4S] cluster assembly.

The NER and mismatch repair modules both model PCNA, RFC, Pol delta and LIG1,
because no human replication module exists to refer to. A shared human
gap-filling module could replace both copies.

Possible upstream errors found in this set:

• GOA.
• PMS1 has a contributes_to DNA endonuclease IDA row whose paper assays human
  MLH1-PMS2 and names PMS1 only as the yeast subunit.
• PCNA has a purine mismatch glycosylase IDA row whose abstract reports only
  binding to MutS complexes.
• TNRC6A's only adaptor-activity row is HDA from an RNA-interactome screen.
• RDH10's NAD+-dependent activity cites a paper reporting a preference for
  NADP.
• MAS1 has an angiotensin II receptor TAS row from 1988, before
  angiotensin-(1-7) was known as its ligand.
• ACE2 has a peptidyl-dipeptidase IDA row whose paper says it acts only as a
  carboxypeptidase on the angiotensins.
• NUP214 and NUP42 carry a karyopherin term, nuclear export signal receptor
  activity.
• GO-CAM.
• The TC-NER model 6796b94c00002938 types CSB and UVSSA with activities
  other than their reviewed and experimental ones.
• The resection models 65a1f4f800000855 and 65a1f4f800001380 use
  replication-fork resection (GO:0110025) where DSB processing
  (GO:0000729) looks intended.
• 613aae0000001005 places RAD51 single-stranded DNA binding in the
  presynaptic intermediate filament cytoskeleton.
• p53 is typed with four different transcription factor terms across models.
• Coverage. No human GO-CAM models canonical miRNA biogenesis, mismatch
  repair, NHEJ, CIA, PCP or retinoic acid synthesis.

Sixth set: transcription, mTOR, cytokines, lipids and calcium

The sixth set modelled the remaining mid-sized mammalian pathways and IL-10
signaling in nine modules. Each cites one PW id as evidence. The re-run moves
eight seeds to has_module; IL-10 was already there by IL-22 gene overlap and
now has its own module. Two broad seeds also moved, but only by gene overlap:
calcium-mediated signaling (PW:0001140) and eukaryotic RNA polymerase
transcription (PW:0001275). Both remain unmodelled as a whole.

• Transcription.
• modules/rna_polymerase_i_transcription.yaml
  (PW:0000126) has UBF and SL1
  promoter recognition, RRN3 recruitment, the Pol I enzyme as one complex with
  13 units, and TTF1/CAVIN1 termination. It models 23 of RGD's 37 genes; the
  general elongation factors and the signalling regulators are left out.
• modules/rna_polymerase_iii_transcription.yaml
  (PW:0000128) has the three
  promoter types as a variant set, TFIIIB with BRF1 or BRF2, the Pol III enzyme
  with POLR3G/POLR3GL as a paralog choice, and optional MAF1 repression. It
  models 36 of 58 genes; tRNA processing (PW:0001573) is left out. The
  subunits shared with Pol I are written identically in both modules.
• mTOR. modules/mtor_complex_signaling.yaml
  (PW:0000180) models mTORC1 and
  mTORC2 as complexes, RHEB, and amino-acid sensing through Ragulator, the Rag
  GTPases, GATOR1/2, FLCN-FNIP and the sensors. The TSC complex, AKT and AMPK
  inputs refer to ampk_energy_sensing and pi3k_akt_mtor. S6K, 4E-BP1, ULK1,
  TFEB and the mTORC2 substrates are kinase targets, not participants.
• Cytokine receptors.
• modules/leptin_receptor_signaling.yaml
  (PW:0000363) is the LepRb
  JAK2-STAT3 cascade with SH2B1, the SHP2-ERK and IRS-PI3K branches, and
  SOCS3/PTP1B feedback. It fills in the receptor cascade that
  leptin_melanocortin_signaling compresses, and takes the ligand step from it.
• modules/il10_signaling.yaml
  (PW:0000515) has IL10RA/IL10RB
  assembly, JAK1 and TYK2, and STAT3 docking. SOCS3 is a transcriptional output,
  because it blocks IL-6 rather than IL-10 signalling.
• Both take STAT dimerisation and transcription from jak_stat_signaling.
• Lipids.
• modules/cardiolipin_metabolism.yaml
  (PW:0001344) runs from
  TAMM41, PGS1, PTPMT1 and CRLS1 to remodelling by a deacylase variant set,
  tafazzin and HADHA. Its chemicals are grounded to ChEBI from the Rhea
  reactions. LCLAT1 is left out: two studies of the human protein found no
  lysocardiolipin activity. The seed's GO-CAM lead is this pathway's own human
  model.
• modules/intestinal_retinoid_absorption.yaml
  (PW:0000141) covers dietary
  vitamin A uptake, the part retinoic_acid_metabolism left out. Its steps are
  retinyl ester hydrolysis, SCARB1/CD36 carotenoid uptake, BCO1 cleavage,
  retinal reduction (the enzyme is unknown, so AKR1B10 and RDH11 are
  candidates), CRBP2, LRAT and chylomicron packaging by reference, ISX feedback
  and ABCA1 efflux.
• Calcium. GO has no specific process term for either pathway.
• modules/camkii_signaling.yaml
  (PW:0001185) has calmodulin
  activation of the CaMKII holoenzyme, Thr286 autophosphorylation, GluN2B
  docking, the CaMKIIN inhibitors and phosphatase termination. The substrates
  are kinase targets. It does not use the obsolete GO:0099004.
• modules/calcium_sensing_receptor_signaling.yaml
  (PW:0001141) has CASR with
  Gq/11 and Gi/o routes, PLC and cAMP steps by reference, AP2S1-dependent
  endocytosis and GRK/arrestin termination. GNA11 and AP2S1 are the familial
  hypocalciuric hypercalcaemia genes.

Possible upstream errors found in this set:

• GOA.
• TAFAZZIN's GO:0047184 IDA row uses an acyl-CoA-dependent term, but tafazzin
  is CoA-independent.
• PPM1F has IDA GO:0033192 calmodulin-dependent protein phosphatase activity,
  a calcineurin-type term, for a PP2C phosphatase.
• SAMTOR is annotated to positive regulation of TORC1 signaling by a paper
  showing it inhibits mTORC1.
• MAF1's promoter DNA-binding IDA rows look like ChIP occupancy read as a
  molecular function.
• POLR3H and CRCP carry the whole Pol III complex's polymerase activity.
• GO-CAM.
• The cardiolipin model 67b1629100002092 chains tafazzin and HADHA in
  sequence, though they are alternative routes.
• The IL-10 model 626b20c100000681 has no TYK2-to-STAT3 edge and omits STAT1.
• Both CASR models use GNAQ rather than the genetically validated GNA11.
• RGD. The PW:0001344 text gives phosphatidylglycerol as TAMM41's substrate;
  it is phosphatidic acid, as RGD's own diagram shows.

Next steps

• Spot-check about 50 pw_def_xref mappings and, if they hold, promote PW's
  definition citations from relatedMatch to closeMatch/exactMatch.
• Send RGD the 45 obsolete GO citation sites (44 in definitions, 1 on a
  diagram page; the page one repeats PW:0000543's definition citation, so 44
  distinct pairs), the unresolvable one and the 3 wrong page citations, with
  GO's replacement where one exists, and GO:0141156 for PW:0000543.
  Filtering curation_rule on pw2go:qa/ returns all 50 rows: qa/OBSOLETE_TARGET 45, qa/MISSING_TARGET 1,
  qa/PAGE_XREF_DISAGREES 4 (the three wrong citations and the defensible
  PW:0000862) and qa/ID_ECHO 1 (PW:0000076, which also disagrees).
• Reconcile gpcr_camp_pka_signaling with the fuller
  camp_pka_signal_transduction: the four Gs hormone modules still refer to the
  compact module's cyclase and PKA steps.
• Raise the missing bradykinin (kinin) receptor signalling and CaMKII signalling
  terms with GO, or record them as proposed_new_terms in the modules that need
  them.
• Review the blood-set member genes, starting with the proteases and cofactors
  that carry curator-judgment molecular functions (F3, F5, F8, PROS1).
• Model the remaining pathways. Chromatin remodeling and calcium transport mix
  unrelated complex or channel families and need splitting by family. The broad
  ones (RNA polymerase II, translation, ribosome biogenesis, the spliceosome and
  histone modification, plus RNA polymerase transcription and calcium-mediated
  signaling, which match only by gene overlap) need splitting first. Kinin
  signaling (PW:0001480) needs a decision on kallikrein_kinin_cascade's
  scope.
• Make pi3k_akt_mtor's mTORC1 step a reference to
  mtor_complex_signaling#mtorc1, and move TBC1D7 into the TSC complex in
  ampk_energy_sensing.
• Revisit IL10RB's function in il22_signaling (GO:0004896, PAINT-derived):
  il10_signaling uses its experimental coreceptor activity GO:0015026, which
  fits a subunit that cannot bind ligand alone.
• Factor a human DNA gap-filling module (PCNA, RFC, Pol delta, LIG1) out of the
  NER and mismatch repair modules, and point
  retinoic_acid_receptor_signaling's production step at
  retinoic_acid_metabolism.
• Fix the triage's prose scraping of GO and UniProt ids
  (ai4curation/ai-gene-review#6006).
• Candidate modules split out during the sets: Reelin modulation of NMDA/AMPA
  receptors in the adult synapse, and a human PAR/Crumbs/Scribble polarity module
  (the PKC module mentions the PAR complex only through the atypical PKCs), and a
  stereocilium-organization module for the bundle genes left out of hair-cell
  mechanotransduction.
• Repeat the completeness measure with rat and mouse GO-CAMs. The index holds
  59 rat and 2,209 mouse activity rows.

Status

• [x] Fetch PW, RGD gene annotations and all 200 RGD diagram pages
• [x] Parse diagram pages into structured records (pathways/)
• [x] Candidate PW→GO mappings (data/pw_go_mappings.sssom.yaml; just validate-rgd-mappings)
• [x] GO-CAM completeness table (data/pw_gocam_completeness.tsv)
• [x] Module seeds and triage (data/module_seeds.yaml)
• [x] Pilot module from an RGD diagram (modules/srebp_signaling.yaml)
• [ ] Spot-check and promote definition-citation mappings
• [ ] Report obsolete GO citations to RGD
• [x] Map needs_go_mapping seeds (all 28, in data/pw_go_curated.sssom.yaml)
• [x] Gene reviews for SREBP pathway members
• [x] Model the candidate seeds (9 of the 10 from the 2026-10-20 triage, in the second and third sets; IL-10 left the queue by gene overlap)
• [x] Model the 14 candidates from the curated mappings (13 modules, fourth set)
• [x] Model the mammalian pathways whose only match was another taxon's module, and the RAS cascade (11 modules, fifth set)
• [x] Model the remaining mid-sized mammalian pathways and IL-10 (9 modules, sixth set)

Notes

2026-10-16

Project created. Pipeline, parsed diagram records, mappings, completeness
table, seeds and the SREBP pilot module were added. Inputs: PW release
2026-05-16 and GO release 2026-07-26.

2026-10-26

Second set of four modules (acetylcholine, dopamine receptor, visual
phototransduction, prostanoid biosynthesis). The has_module test now counts
a PW id only when a module cites it as an evidence source.

Third set of five modules (PKC signaling, DNA cytosine methylation cycle,
reelin, TRAIL apoptotic signaling, NMJ AChR clustering). The candidate queue
is empty.

2026-10-26 (curated mappings)

Hand-curated GO mappings for the 28 needs_go_mapping seeds
(data/pw_go_curated.sssom.yaml). Fourteen became candidates, five match only
other-taxon or GO-term modules, and 9 have only a broad GO term or none
(new tier no_specific_go_term).

2026-10-28

Fourth set of 13 modules for the 14 curated-mapping candidates. The re-run
leaves no candidates: 107 has_module, 23 related_module_only and 7
no_specific_go_term.

2026-11-05

Fifth set of 11 modules: nucleotide excision repair, mismatch repair,
homologous recombination, NHEJ, p53 signaling, microRNA biogenesis, mRNA
nuclear export, retinoic acid metabolism, the renin-angiotensin cascade,
cytosolic Fe-S assembly and Wnt/PCP. The re-run gives 118 has_module, 14
related_module_only and 5 no_specific_go_term.

2026-11-10

Sixth set of 9 modules: RNA polymerase I and III transcription, mTOR complex
signaling, leptin receptor and IL-10 signaling, cardiolipin metabolism,
intestinal retinoid absorption, CaMKII and calcium-sensing receptor signaling.
The re-run gives 128 has_module, 6 related_module_only and 3
no_specific_go_term.