ADGRA1 (GPR123) — review notes

UniProt: Q86SQ6 (AGRA1_HUMAN), 560 aa, Swiss-Prot, PE 1: Evidence at protein level,
HGNC:13838, chromosome 10. Accession verified independently against
projects/paint/human-no-IBA-simple.csv (human,Q86SQ6,ADGRA1) and against the UniProt REST
record, which returns uniProtkbId: AGRA1_HUMAN, geneName: ADGRA1, synonym GPR123.

Computed evidence for everything numeric below lives in
ADGRA1-bioinformatics/ (analyze_adgra1.py → RESULTS.md, results.json).
Nothing in that report is hardcoded; it is re-read from UniProt, IntAct and QuickGO on each run.

Row reconciliation (done before reviewing, per CLAUDE.md)

The stub under-seeded by 20. Cause is the known one:
GOAValidator.seed_missing_annotations keys entries on
(GO ID, evidence_type, reference, negated, qualifier) and omits WITH/FROM, so the 22
GO:0005515 rows — which differ only in the partner accession — collapsed to 2 entries
(one per reference). All 22 were restored so every partner carries its own verdict; that
matters here because the partners split 8/13 on whether an affinity was actually measured.

What ADGRA1 is

ADGRA1 is the structural outlier of the 33-member adhesion GPCR family. Two independent
sources say so:

• UniProt's own feature table: TOPO_DOM 1..19 Extracellular, then TRANSMEM 20..40. The
  canonical isoform (Q86SQ6-3) has a 19-residue extracellular N-terminus — there is
  physically no room for the GAIN/GPS module every other aGPCR carries.
• PMID:12565841
• PMID:41961591

The mass of the protein is on the inside: TOPO_DOM 306..560 Cytoplasmic, a 255-residue
tail ending in ...WKNETTV. ETTV matches the class I PDZ-binding consensus X-[ST]-X-[VIL]
exactly, which the analysis script checks against the retrieved sequence rather than asserting.
PMID:41961591
and the motif was already predicted from sequence conservation twenty years ago
PMID:17212699.

UniProt does not annotate this motif. There is no FT MOTIF feature in the entry
(checked programmatically), even though the entry's own CC -!- INTERACTION block lists 21
PDZ-domain partners. See "UniProt corrections" below.

Expression is CNS-restricted
PMID:17212699,
and HPA calls the gene Tissue enriched (brain).

Finding 1 — the protein binding rows are a PDZ interactome, and they say so

All 22 GO:0005515 rows come from two papers, and both are PDZ-specific by construction:

• PMID:24550280 — proteomic peptide-phage display.
  PMID:24550280;
  PMID:24550280
  The ADGRA1 "interactor" here is its C-terminal peptide, selected by a DLG1 PDZ domain.
• PMID:36115835 — the holdup assay.
  PMID:36115835

Measured, not asserted: 21 of 21 GOA partners carry at least one annotated PDZ domain
(RESULTS.md Q2; range 2–13 domains, all reviewed Swiss-Prot entries at canonical length —
no TrEMBL or ORFeome substitutions, unlike ACRV1's Q86WV8). IntAct's own curation agrees
independently: every record annotates the ADGRA1 side as feature PDZ-binding motif
(featureTypes: sufficient to bind) and the partner side as PDZ domain.

Topology check: passes. The tail is cytoplasmic (TOPO_DOM 306..560) and all 21 partners
are cytoplasmic/membrane-associated scaffolds. This is the opposite of the ACRV1 pattern,
where a secretory-lumen protein had only cytosol-facing partners.

So GO:0005515 protein binding is uninformative where a mechanism is in hand:
MODIFY → GO:0030165 PDZ domain binding on all 22 rows.

Do not over-read the motif, though. The one experiment that tested what it does found
nothing: PMID:41961591
The motif binds PDZ domains; it does not target the receptor to synapses. The review says the
first and not the second.

Finding 2 — GOA's 21 partners were chosen by domain count, not by affinity

This started as a numeric discrepancy and turned out to be the bug report, as usual.

IntAct holds 124 ADGRA1 records over 80 partners; 122 are holdup assay from PMID:36115835.
GOA/UniProt keep 21. Asking which 21:

• UniProt's NbExp equals the IntAct record count for 21 of 21 partners — so NbExp here
  is not a count of experiments at all. It counts how many PDZ domains of the same partner
  protein were assayed inside one dataset. PATJ's NbExp=9 is nine PATJ PDZ domains, not
  nine experiments. (Third instance of the NbExp trap this campaign, after ACRV1 and
  ADAMTSL5; the first two were sub-methods of one screen, this one is domains of one protein.)
• The cut is exactly NbExp ≥ 2. 22 IntAct partners have ≥2 records; 21 are in GOA. The one
  exception is NHERF4/PDZD3 — checked rather than rounded away.
• Consequence one: 13 of the 21 GOA partners have no quantified affinity. IntAct carries a
  kd:1(molar) placeholder for them, consistent with the paper's own statement that
  PMID:36115835
• Consequence two: 23 partners with a genuinely measured Kd are excluded. The sharpest
  statement is a single pair, because it isolates the variable: SNX27, at 3.7 µM the tightest
  binder measured anywhere in the dataset, is excluded, while DLG1 at 4.6 µM — the second
  tightest — is retained. Affinity is not what separates them; PDZ-domain count is (SNX27 has
  one PDZ domain, DLG1 has three). Of the four tightest binders overall — SNX27 3.7, DLG1 4.6,
  MAST2 4.9, MAGI3 5.1 µM — three are excluded and only DLG1 is kept.

(Corrected in round 2. The first version ranked the three excluded partners against DLG1 in
the wrong direction. Lower Kd is tighter, so MAST2 (4.9) and MAGI3 (5.1) are in fact weaker
than the retained DLG1 (4.6), and DLG1 is itself second-tightest overall rather than a
ceiling. The reviewer caught it; I recomputed the full ranking from results.json and the
reviewer is right. The corrected form is the stronger argument anyway, because one pair
isolates the variable where three loosely-ranked numbers did not. The four retracted phrasings
are now in WITHDRAWN_PHRASES in audit_adgra1_claims.py, and that check was verified to fire
against the exact blocked revision via git show 3ce38c99e. Note the guard's limitation, which
this very paragraph ran into: a literal-phrase matcher cannot distinguish a claim being
retracted from a claim being asserted, so the retracted wording is described here rather than
reproduced. It also cannot catch a paraphrase — prose surfaces still need re-reading by hand.)

I did not convert this into per-row MARK_AS_OVER_ANNOTATED verdicts on the 13. The
kd:1(molar) placeholder means "not quantified", and IntAct's negative flag is False on
all 124 records, so the pairs were curated as detected positives. Treating an
under-determined placeholder as a refutation would be over-calling in the same shape as the
AADACL2 topology argument that measurement declined to confirm. The affinity split is recorded
per row as a fact and the verdict is the same MODIFY for all 22.

Finding 3 — the adhesion-GPCR classification block, alive in GOA

Campaign calibration says the "structural class became a molecular function" error was swept
out of GOA when Swiss-Prot keyword annotations (GO_REF:0000043, IEA:UniProtKB-KW) were
retired around April 2026. On ADGRA1 that is only half true, and the half that survives is
worth reporting.

Checked at the UniProt layer first, as instructed: the entry carries no EC=, and its
DR GO; block contains no IEA:UniProtKB-KW line at all — the SPKW sweep did reach this
entry. KW G protein-coupled receptor; Receptor; Transducer; remain but no longer generate GO.

But GOA still carries GO:0004930 four times, and the three non-IEA lines cite papers that
contain no functional experiment of any kind:

The TAS row is not a one-gene problem. Querying QuickGO by reference rather than by gene:
PMID:15203201 carries 78 annotations across 27 distinct entities — GO:0016020 on 27,
GO:0007186 on 26, GO:0004930 on 25 — i.e. essentially the entire human adhesion-GPCR
family, all TAS, all from GDB, from a paper that performed no perturbation on any of them.
Applying the two-question projection test from ACTR8/ACTRT3: (1) 27 entities from one
reference; (2) the "functional" term does not stay on a perturbed gene, because no gene was
perturbed — the paper is a catalogue. So this is a block projection of family membership into
a molecular function, and it entered through GDB TAS, a route the SPKW retirement never
touched.

But the term is nonetheless correct, which is the part I did not expect. See Finding 4.

Finding 4 — a 21-year-old unsupported annotation that became true in 2026

PMID:41961591 (Cell Reports, Apr 2026) measured it:
PMID:41961591, validated
by dose–response —
PMID:41961591
— with GαoB as an unresponsive control.

Three qualifications that the review records rather than smooths over:

1. The construct is mouse, not human.
   PMID:41961591
   So for the human gene this supports an ISS row with supporting_entities: [UniProtKB:Q8C4G9], not IDA.
2. The activation is agonist-independent. No ligand is known, and the receptor's entire
   19-residue ectodomain is dispensable for it:
   PMID:41961591.
3. GO:0004930's definition is "Combining with an extracellular signal and transmitting the
   signal across the membrane by activating an associated G-protein…" (QuickGO). The second
   conjunct is now demonstrated; the first is unestablished, and experiment 2 above argues
   it is not what is happening in the assay. Same for GO:0007186, whose definition begins
   "The series of molecular signals initiated by a ligand binding to its receptor."

Verdict: ACCEPT GO:0004930 and GO:0007186 — GO routinely annotates orphan and
constitutively active GPCRs here and there is no ligand-free sibling term (checked: no such
term found; GO:0004888 also requires "combining with a signal"). The ligand half is raised as
an ontology question, not converted into a GO action, and the three classification papers are
flagged MISCITED in reference_review — which is precisely what that field is for. A NEW
GO:0004930 ISS row gives the term its first real evidence line.

So the fold/class-propagation lead came back partly confirmed and partly inverted: the
annotation provenance is as bad as predicted (a 27-gene block from a catalogue paper), and the
claim turns out to be right anyway. REMOVE would have been wrong.

Finding 5 — PAINT node placement here is good, and that is the answer to the paralog question

Both halves of the node question, from RESULTS.md Q5:

The paralog-transfer hypothesis is NOT confirmed, and the denominator makes that a
measurement rather than a hedge. The ADGRA2 donors carry a large body of their own
experimental annotation — mouse Adgra2 alone has 15 experimental rows (angiogenesis, Wnt
signalling, CNS development), human ADGRA2 has 9 including GO:1990909 Wnt signalosome IDA,
zebrafish adgra2 has 10. Of all of that, exactly two of the most generic terms reached
ADGRA1, and every ADGRA2-specific Wnt/BBB term stayed on the ADGRA2-only node
PTN002914520. PAINT walled the characterised paralog's biology off correctly.

What is true is narrower and is recorded as provenance, not as a verdict: 1 of 6 IBA
donor tokens is an ADGRA1 ortholog. GO:0005886 and GO:0007166 reach ADGRA1 from a donor set
that is entirely ADGRA2/ADGRA3. Following the ACTA2 discipline of separating the two
claims: that is provenance, and on its own supports no verdict. The circularity question
resolves the other way — the terms are independently corroborated for ADGRA1 by its own
ortholog node (postsynaptic density is plasma membrane) and by surface-labelled HA-ADGRA1
puncta in the 2026 paper. So both rows are ACCEPT with the paralog-only donor set noted.

Precision check (the ACRV1 test): negative. Mouse Adgra1 holds GO:0014069 and
GO:0098978 by IDA/EXP from PMID:28935861, i.e. the same terms that propagated — not more
specific ones. So no downward MODIFY is warranted, unlike ACRV1 where the IBA landed three
levels above its donor's IDA. Reported because a null result from a check is a finding.
The donor evidence itself is sound:
PMID:28935861

The reciprocal half: what PTN002914505 did not give

PTN002914505's human reach is exactly ADGRA1, and it gives two CC terms and nothing else.
ADGRA1 therefore receives no molecular function and no functional biological process from
PAINT at all, and its only two IBA terms trace to a 2017 sub-cellular proteomics survey.

The 2026 Cell Reports work is ortholog-node evidence of exactly the kind that node exists to
propagate — G-protein coupling and a PV-interneuron-specific conditional-knockout phenotype in
mouse Adgra1, the very gene behind MGI:MGI:1277167. Recommending it to PAINT is the single
highest-value action available on this gene, and it is filed in suggested_questions.

Finding 6 — the compartment and the process did not line up (round-2 fix)

GOA places ADGRA1 at the glutamatergic synapse and postsynaptic density, both by IBA from
the 2017 proteomics survey. The 2026 study localises it at the inhibitory synapse:
PMID:41961591,
from HA-ADGRA1 delivered by AAV into the dentate gyrus of PV-Cre and SST-Cre mice.

The first version of this review filed the process half of that experiment (GO:0032230) and
not the compartment half, so core_functions asserted a molecular function at the
glutamatergic postsynapse while asserting positive regulation of GABAergic transmission as the
process. GO:0098982 GABA-ergic synapse is now filed as a second NEW ISS row on the same
reference, same organism, same evidence strength, and added to the locations of both core
functions. The existing GO:0098978 ACCEPT is untouched — deferring to the SynGO curator is
right, and the two terms are not exclusive for a receptor at a subset of synapses of both kinds.

GO:0098793 presynapse was considered and declined, and the reasons matter more than the
verdict because they run against the obvious reading. vGAT is a presynaptic marker and the
rescue works when the receptor is restored in the presynaptic PV cell
PMID:41961591 — but
that is a statement about which cell, not about which side of the synapse the protein sits on.
Against it, and keeping the two kinds of argument apart — a distinction the reviewer was right to
press: the localisation case is weak on its own terms, because the sentence is hedged
("suggesting subcellular localization") and rests entirely on overexpressed tagged protein
PMID:41961591.
Those are the grounds for declining GO:0098793. Separately, on function, the paper's own
controls show the presynaptic release apparatus is intact —
PMID:41961591
and PMID:41961591.
That is a mechanistic gap in its own right and says nothing about where the protein is; the first
round of this review recruited it as if it did. Both filed as a knowledge gap rather than a term.
Restraint argued per term from the measured numbers, not applied uniformly: GO:0098982 in,
GO:0098793 out.

Cross-review check against the concurrent ADGRA2 and ADGRA3 reviews

Run per the AADACL2/3/4 lesson: when three paralogs are reviewed independently, compare the
rows they share rather than diverging silently. Sibling PRs #2314 (ADGRA2) and #2315 (ADGRA3).

Eleven rows are shared by all three genes (identical term + evidence code + reference; the two
IBA rows also have byte-identical WITH/FROM). Eight of the eleven are resolved differently.

All three independently reached MODIFY → GO:0030165 on the PDZ rows. Three separate
reviews converging on the same replacement for the same 21-partner holdup dataset is the
strongest single result in this comparison.

All three independently found the PMID:15203201 GDB block. ADGRA2's review derived the
same 78-annotations/27-entities/25-with-GO:0004930 figures I did, and went one step further:
it checked by name which two entities miss GO:0004930, so that the pseudogene recipients
are confirmed rather than inferred from the count. I reproduced that claim from QuickGO and it
is correct — the two are ADGRG3 and ADGRV1, and the pseudogenes ADGRE4P and
ADGRF2P do receive the molecular function. Two reviews deriving the same block independently
is much stronger than either alone; cited in agreement.

Where I disagree, and why it is mostly not a disagreement

The GO:0004930/GO:0007186 split (ADGRA1 ACCEPT, ADGRA3 ACCEPT, ADGRA2 REMOVE) is
evidence-driven, not method-driven, and by the ACTR1A/ACTR10 rule that needs no
reconciliation: all three reviews apply the same standard — annotate what has been measured on
this gene — and the measurements differ.

• ADGRA1: PMID:41961591 (Cell Rep, Apr 2026), TRUPATH 14-sensor panel, Gα13/Gα11/Gα15, dose-response.
• ADGRA3: PMID:40127866 (J Biol Chem, May 2025), "Complex G-protein signaling of the adhesion
  GPCR, ADGRA3" — Gi and Gs, stachel-dependent, DVL-independent. Verified at PubMed; the paper
  is real and is what the sibling says it is.
• ADGRA2: no coupling assay, and gene-specific evidence pointing away from one (IUPHAR lists no
  transducer; UniProt states the characterised tethering activity does not rely on its GPCR
  structure; the one reported heterotrimeric contact runs the other way, Gβγ binding the C-tail).

So three different verdicts on one row can all be right here, unlike the AADACL case where one
row on one node had to have one answer.

One correction I owe the ADGRA2 review

Its REMOVE reason states: "No G-protein coupling assay for ADGRA1, ADGRA2 or ADGRA3 exists in
the systematic adhesion-GPCR coupling literature ... and none for the ADGRA clade."

That clade-wide negative is false, and it is refuted twice over by work the two sibling
reviews found independently: PMID:41961591 assays ADGRA1 and PMID:40127866 assays ADGRA3.
Neither is cited in the ADGRA2 review. The miss is instructive rather than careless for the
first — a hippocampal-circuits paper will not surface in a pharmacology-literature sweep — but
PMID:40127866 is squarely inside the literature that was swept, so the sweep was incomplete.

This does not flip ADGRA2's verdict. Its gene-specific arguments are untouched by either
paper, and coupling in ADGRA1 and ADGRA3 says nothing about ADGRA2. What it removes is a
strengthener: the row cannot be argued down on the grounds that the clade as a whole is
uncoupled, only on ADGRA2's own evidence. An absence asserted at clade level, when two of three
members have since been assayed, is exactly the "an absence is not a finding" failure this
campaign has recorded before.

Its other cross-gene premise — "across all three human ADGRA paralogs there is no
molecular-function IBA at all" — I checked independently and it is correct (see the
node-reach table in RESULTS.md: no node in this family carries an MF term).

Checks run that came back negative (recorded so the next reviewer knows)

• Retraction / erratum / expression-of-concern sweep over all 15 cited PMIDs, reading
  CommentsCorrections/RefType on each cited article's own PubMed record (a Publisher
  Correction is not findable by a publication-type search). One hit: PMID:36115835 carries an
  Author Correction, PMID:36477203. Read it — it is figure formatting only ("errors in
  Fig. 2, Fig. 4 and Fig. 5 … PCC values were missing … axes labels were missing … the order of
  panels did not match"). It touches no data, and ADGRA1's data are in Supplementary Data 1
  regardless. Nothing rests on a corrected figure. No retractions anywhere.
• Crossref relation/update-to check for null-PMID corrigenda on the three load-bearing
  DOIs (10.1016/j.celrep.2026.117255, 10.1038/s41467-022-33018-0,
  10.1073/pnas.1312296111): no update-to on any; the only relations are has-preprint.
• Partner-accession discipline (the ACRV1 Q86WV8 check): all 21 resolve to reviewed
  Swiss-Prot entries at canonical length. No partial ORFeome constructs.
• IBA-precision check (ACRV1): negative, see above.
• Dead-accession check: every accession lookup prints its entry name and fails loudly on an
  empty one. None was empty.

Things affinage got right, and one it did not

gates_passed: True, faith_pct: 100.0, no PMID:bio_* pseudo-identifiers in the citation
list. Its narrative is broadly accurate and it surfaced PMID:28935861 and PMID:41961591,
which are the two most load-bearing references here.

Two corrections:

• It cites PMID:40766348 as a 2025 bioRxiv finding and PMID:41961591 as a separate 2026
  Cell Reports finding. They are the same study — PubMed records UpdateIn: 41961591 on the
  preprint. Counting them as two dated findings inflates the apparent evidence base. The review
  cites the published version and marks the preprint LOW/superseded.
• Its SST claim ("PV and SST interneurons", "deficits in learning and memory") is the
  preprint's abstract. The published Cell Reports abstract is narrowed to PV
  PMID:41961591.
  The SST work survives in the published paper's body, but the headline claim was scoped down
  between versions, so the preprint wording should not be quoted as the finding.

It also never mentions the PDZ interactome, which is the best-evidenced molecular fact about
this protein.

UniProt corrections to report (no GO row exists for these)

1. CC -!- FUNCTION: Orphan receptor. is the entire function annotation. It predates
   PMID:41961591 (Gα13/Gα11/Gα15 coupling of the mouse ortholog) and PMID:36115835
   (21 PDZ partners with measured affinities). "Orphan" is still right about the ligand;
   it is now wrong as a summary of what is known.
2. No FT MOTIF for the C-terminal PDZ-binding motif (residues 557–560, ETTV), despite
   the entry listing 21 PDZ-domain interactors in its own CC -!- INTERACTION block and the
   motif being predicted in the literature since 2007. This is the single most useful missing
   feature on the entry.
3. CC -!- SUBCELLULAR LOCATION: Membrane {ECO:0000255} is sequence-predicted only, while
   SynGO holds IDA GO:0014069/GO:0098978 on the mouse ortholog from PMID:28935861.
4. NbExp in the CC -!- INTERACTION block counts PDZ domains assayed, not experiments (see
   Finding 2). Every one of the 21 values is the IntAct record count for that partner.

Verdict summary

Over the 39 GOA rows: ACCEPT 11, MODIFY 28. Plus 2 NEW proposed rows, counted separately (41 existing_annotations entries in total).

MODIFY breakdown, summing to 28: 22 × GO:0005515 → GO:0030165; 4 × GO:0016020 →
GO:0005886; 1 × GO:0004888 → GO:0004930; 1 × GO:0007165 → GO:0007166.

ACCEPT breakdown, summing to 11: 4 IBA (GO:0005886, GO:0007166, GO:0014069,
GO:0098978); 3 InterPro IEA (GO:0004930, GO:0007166, GO:0007186); 2 NAS and 1 TAS
GO:0004930; 1 TAS GO:0007186.

NEW: GO:0032230 (positive regulation of GABAergic synaptic transmission) and GO:0098982
(GABA-ergic synapse), both ISS from PMID:41961591 with UniProtKB:Q8C4G9 as the supporting
entity — the process and the compartment halves of the same experiment.
GO:0004930 is deliberately not filed as NEW even though it needs a real evidence line —
the term is already in GOA four times, and action: NEW is for terms GOA lacks (the repo
validator enforces this, correctly). The request to attach PMID:41961591 as an ISS evidence
line to the existing term is made in suggested_questions and in each GO:0004930 row's reason.

No REMOVE anywhere. Nothing on this gene is demonstrably wrong — only under-specified
(protein binding, membrane, signal transduction) or, in the GO:0004930 case,
under-evidenced by references that a 2026 paper has since overtaken. These counts are
enforced against the YAML by ADGRA1-bioinformatics/audit_adgra1_claims.py.