AGTRAP (ATRAP, Q6RW13) — review notes

Human gene, HGNC:13539, 159 aa, UniProt PE 1: Evidence at protein level.
PANTHER PTHR16521 / InterPro IPR009436 / Pfam PF06396, a single-subfamily
family (PTHR16521:SF3) with 1380 member proteins but only 5 Swiss-Prot
entries (human, mouse, rat, orangutan, Xenopus — all AGTRAP orthologues, no
paralogues). Reviewed-member counts from interpro/panther/PTHR16521/ are the
Swiss-Prot subset, not the family.

The shape of this gene's GO record

AGTRAP-goa.tsv has 199 rows. The fetch-gene stub seeded 199
existing_annotations entries, so nothing was collapsed here (unlike ACTR5 /
ADAMTSL5). Counts reconciled before review began:

wc -l < AGTRAP-goa.tsv        -> 200 (199 + header)
grep -c '^- term:' *-ai-review.yaml -> 199

Composition of the 199 rows:

So 94% of this gene's GO record is protein binding, and the interaction
that the gene is named for — AGTR1 — is not in it at all.

Finding 1: GO:0004945 angiotensin type II receptor activity is a symbol confusion amplified by orthology projection

The human row is enables GO:0004945, IEA (ECO:0000265, GO_REF:0000107,
Ensembl Compara) with UniProtKB:Q9WVK0 = mouse Agtrap (Swiss-Prot,
ATRAP_MOUSE, a true 1:1 orthologue — resolved, not assumed).

Going one level deeper than the donor (QuickGO,
geneProductId=UniProtKB:Q9WVK0): mouse Agtrap holds GO:0004945 by IMP from
PMID:10425188, assigned by MGI. That paper is "Increased vasoconstrictor
response of the mouse lacking angiotensin II type 2 receptor" — an Agtr2
knockout study. Its abstract never mentions Agtrap/ATRAP.

Querying by reference rather than by gene makes the mechanism explicit:

QuickGO annotation/search?reference=PMID:10425188  -> 13 annotations, 3 entities
  Agt    (P11859)  5 rows
  Agtr2  (P35374)  5 rows  incl. GO:0004945 enables IGI   <- the correct holder
  Agtrap (Q9WVK0)  2 rows  incl. GO:0004945 enables IMP   <- the confusion

Agtr2 and Agtrap both received GO:0004945 from the same paper. Only one of
them is an angiotensin receptor.

Independent of provenance, the term is biologically false for this protein:

• GO:0004945 is defined as "An angiotensin receptor activity that acts via Gi
  protein coupling and cGMP (NO) generation". AGTRAP is a 159-aa protein with
  three TM helices (UniProt FT 24-44, 56-76, 87-107), not a 7TM GPCR, and
  has no G-protein coupling anywhere in its record.
• The gene's founding paper tested exactly this and excluded it:
  PMID:10358057 — "ATRAP interacts specifically with the carboxyl-terminal
  domain of the AT1a receptor but not with those of angiotensin II type 2 (AT2),
  m3 muscarinic acetylcholine, bradykinin B2, endothelin B, and
  beta2-adrenergic receptors."

So the annotation asserts, of the protein named for the type-1 receptor, an
activity of the type-2 receptor that was experimentally excluded for it.
Action: REMOVE.

The error has spread to three species. Rat Agtrap (Q642A2) carries
GO:0004945 twice — by Ensembl IEA and by RGD ISO — so human, rat and mouse
all currently assert it. Fixing the MGI row would clear all three.

Finding 2: GO:0008217 regulation of blood pressure — right term, wrong evidence

The human row is IBA from PANTHER:PTN002703102 + MGI:MGI:1339977. The PAINT
file (interpro/panther/PTHR16521/PTHR16521-paint.tsv) shows the family has
exactly two IBD annotations, both seeded from mouse Agtrap alone:

PTHR16521  PTN002314244  GO:0005886  C  IBD  MGI:MGI:1339977  taxon:6072
PTHR16521  PTN002703102  GO:0008217  P  IBD  MGI:MGI:1339977  taxon:117571

and mouse Agtrap's only experimental GO:0008217 row is the same
PMID:10425188 IMP from Finding 1. So the chain human IBA -> PAINT node ->
mouse IMP terminates in a paper about a different gene.

The conclusion nevertheless survives on evidence nobody curated: PMID:20093357
(Oppermann/Castrop, JASN 2010) — "Mean systolic BP was significantly higher in
Atrap-/- mice compared with wild-type mice." Kept as KEEP_AS_NON_CORE
(organism-level physiology downstream of AT1R restraint, not AGTRAP's own
molecular function), with a suggested_question asking MGI to re-reference.

This is the inverse of the usual campaign finding: the propagation is sound and
the donor's citation is the defect.

Finding 3: GO:0001666 response to hypoxia rests on one rat IEP

Human row: IEA from rat Q642A2 (Swiss-Prot ATRAP_RAT, true orthologue).
Rat's own evidence: a single IEP from PMID:17033689, "Vascular response to
hypoxic preconditioning in the immature brain". That reference annotates
exactly 2 entities in all of GOA (rat Agtrap and Angptl4), both IEP — i.e. a
transcript-abundance change in neonatal rat brain. Same shape as AFF4's
GO:0034976. MARK_AS_OVER_ANNOTATED.

Finding 4: 94% of the interaction record is one assay class, and the informative partners are in the other one

IntAct (/intact/ws/interaction/findInteractions/Q6RW13, all 617 evidences
retrieved; len(results) == totalElements asserted):

92 x 3 and ~68 x 3 is the arithmetic of one screen logged three ways. This
is the NbExp=3 artefact (ACRV1, ADAMTSL5) confirmed by exact counts rather
than inferred: UniProt's NbExp=3 for most AGTRAP partners is one experiment.
Overall 583/617 = 94.5% of evidences are two-hybrid, and only 23 of 213
distinct pairs carry any non-two-hybrid method.

The GOA import is drawn entirely from the Y2H side: all 188 binding rows trace
to 9 CCSB-lineage publications. Partner accessions resolved (132 distinct; all
resolved, entryType.startswith("UniProtKB reviewed") used so TrEMBL is not
silently promoted): 121 Swiss-Prot, 11 TrEMBL, and several TrEMBL ones are
partial ORFeome clones rather than the canonical protein —

Topology check (scratchpad/topology.py): AGTRAP's only substantial
cytosol-facing surface is TOPO_DOM 108-159 — 52 residues, 20 of them
disordered. Of the 131 non-self GOA partners, 28 are confined to the
mitochondrial matrix/inner membrane, the nucleoplasm, or the extracellular
space and cannot reach that surface; 17 more are mixed; 67 are cytosol-facing;
19 have no recorded location. Reported as it came out — this is a supporting
argument, not the decisive one, because a majority of partners are nominally
accessible. The decisive argument is the assay structure above.

The property that every screen row's verdict rests on, stated exactly. An
earlier draft said, on all 178 rows, that "searching this partner independently
of the AGTRAP symbol found no follow-up study". I did not run 131 separate
partner searches, so that was an over-claim, and it was replaced by what was
actually measured (scratchpad/nony2h_vs_goa.py, which asserts
len(results) == totalElements and then asserts the property per partner):

GOA partner accessions (base): 132
GOA partners with no IntAct record under Q6RW13: 0
IntAct partners with any non-two-hybrid method: 23
   of those, present in GOA: 0
ASSERTION HELD: every GOA partner found in IntAct has two-hybrid methods only

Every one of the 132 GOA partners resolves in IntAct, every one is
two-hybrid-only, and the 23 partners that do have orthogonal support are
disjoint from the GOA set.

The contrast is the finding. The 23 pairs with non-Y2H (AP-MS / co-IP /
BioID) evidence are almost all membrane proteins — CYB5D2, ENPP6, MBLAC2,
TMEM47, TMEM106B, LAMP5, SLC15A3, SLC15A4, TMEM63A, CFTR, CHRNB1, SLC1A1,
SLC18A2, PAFAH2, TNFRSF10C — a coherent ER/endolysosomal membrane neighbourhood
for a 3-TM ER/Golgi protein. None of them is in GOA. GOA imported the assay
class whose partners are largely implausible and none of the class whose
partners fit the protein's own cell biology.

A count that refused to add up, and what it turned out to be

First draft of the PMID:16189514 reference note said "3 binding rows". The TSV has
2. The 3 was an IntAct evidence count, and the extra evidence is a canonical
Q6RW13-Q6RW13 self-pair that GOA did not import. Chasing it corrected a
second claim: the GO:0042802 row had been written as a self-interaction "detected
once, in the HuRI screen", which is true of the GOA row and false of IntAct —
PMID:16189514 and PMID:25416956 also report a canonical self-pair. All three are
the same CCSB clone collection in the same two-hybrid system, so the verdict
(MARK_AS_OVER_ANNOTATED) is unchanged, but it now rests on "recurrence on one
platform, never tested by another method" rather than on "seen once", which was
simply wrong.

Per-reference GOA binding-row counts, re-derived from the TSV rather than from prose:
32296183 91 (90 + the self row) · 25416956 66 · 31515488 10 · 26871637 8 ·
21516116 4 · 25910212 3 · 16189514 2 · 19060904 2 · 29892012 2 = 188.

Three Y2H partners that survive scrutiny

Searching the partners rather than the symbol (the recall gap affinage cannot
close) turned up orthogonal validation for three. "Three" is what this search
found, not a proof that no fourth exists — the exhaustive statement available is
the IntAct one above, that no GOA partner has non-two-hybrid evidence:

• PITPNC1 / RdgBβ (Q9UKF7, 4 GOA rows) — PMID:21728994, titled for
  RdgBβ, not for AGTRAP: "the PITP domain of RdgBβ interacts with the integral
  membrane protein ATRAP ... causing membrane recruitment". This is the one
  partner that yields an informative MF for AGTRAP itself: it is the membrane
  anchor. MODIFY -> GO:0043495 protein-membrane adaptor activity (a term
  already used for integral-membrane tethers such as SUN1/SUN2 and GPR179).
• FAM114A1 (Q8IWE2, 1 GOA row) — PMID:35671117, titled for FAM114A1.
  It explicitly starts from the Y2H hit and validates it: "we validated the
  interaction between AGTRAP and FAM114A1 in mouse heart lysates and cultured
  PMCFs by IP and Western blotting". KEEP_AS_NON_CORE.
• PBX3 (Q96AL5 / Q5JS98, 4 GOA rows) — PMID:35414770: "we performed
  co-IP assays in MDA-MB-453 and T47D cells transfected with Flag-tagged ATRAP,
  which revealed that ATRAP can directly bind with PBX3". Corroborated but in a
  breast-cancer over-expression context, and both GOA accessions are unreviewed
  non-canonical clones. KEEP_AS_NON_CORE.

The remaining 178 GO:0005515 rows and the GO:0042802 self-row get
MARK_AS_OVER_ANNOTATED — not REMOVE, matching the ADAMTSL5 precedent for
comparable unreplicated Y2H rows (and diverging deliberately from ADAMTSL4's
REMOVE; the divergence is flagged rather than silent).

Finding 5: the gene's defining interaction is absent from human GOA

AGTR1 binding is measured, curated by UniProt on the human entry
(SUBUNIT: Interacts with RACK1, and with the C-terminal region of AGTR1.,
ECO:0000269|PubMed:11733189, ECO:0000269|PubMed:12960423), given a feature
(FT REGION 110..122 /note="Interaction with AGTR1"), and present as
GO:0005515 IPI with P29754 (mouse Agtr1a) on the mouse orthologue — yet
it is nowhere in the human GO record.

GO:0031702 type 1 angiotensin receptor binding exists and is exactly right
(Daviet showed the binding is AT1a-specific and excludes AT2). Proposed as NEW.

A claim I had to retract here. The first draft said its 197 holders were all
angiotensinogen orthologues by IEA — read off page 1 of a paginated QuickGO result,
which is exactly the failure the brief warns about. Paging the whole thing (197
annotations over 188 gene products, len(results) asserted against numberOfHits)
gives 175 IEA, 10 ISO, 9 IPI, 3 IDA, and 16 entities with non-IEA evidence:

So the term is already used for cytoplasmic-face binders — AGTRAP's exact class —
which makes the proposal stronger than the retracted version claimed, not weaker. What
is genuinely thin is human coverage: five human entities, only AGT and BDKRB2 by IPI,
the rest Ensembl projections.

Note the composite-claim trap in UniProt's SUBUNIT line: the RACK1 half comes
from PMID:11733189 and the AGTR1 half from PMID:12960423. PMID:11733189
reports the human AGTRAP-RACK1 interaction (Y2H + GST pull-down + co-IP +
SPR) and does not itself assay AGTR1. Cited accordingly.

RACK1 (P63244) is likewise absent from GOA despite four orthogonal assays on
the human protein. Recorded as a coverage gap, not proposed as a bare
protein binding row.

Finding 6: GO:0038166 asserts pathway membership where the data show negative regulation

The row is IEA from InterPro:IPR009436 — the AGTRAP family signature, i.e. a
term derived from the family's name. Every functional measurement on this
protein runs the other way: PLC inhibition (PMID:10358057), reduced inositol
lipids / c-fos / proliferation (PMID:12960423), elevated BP and increased
surface AT1R on knockout (PMID:20093357). GO keeps regulation terms outside
the process they regulate, so MODIFY -> GO:0110062 negative regulation of angiotensin-activated signaling pathway, which carries the direction the data
actually support.

Finding 7: a second, AT1R-independent molecular activity

PMID:27015675 (titled for SERCA2a, not for AGTRAP): pull-down + MALDI-MS,
co-IP and SPR place Atrap on the cardiac Ca2+-ATPase SERCA2a, and "Atrap
enhanced the SERCA-dependent Ca(2+) uptake in isolated SR membrane vesicles";
Atrap-/- myocytes show prolonged Ca2+-transient decay and the mice a reduced
maximum ventricular filling rate. All of it is mouse, so proposed for human as
ISS: GO:0141109 transporter activator activity, the sibling of
GO:0141110 transporter inhibitor activity that phospholamban holds by IDA.
Not in GOA in any form.

Finding 8: the coverage gap is symmetric, and mouse shows it is fixable

Every finding above looks at AGTRAP's own record. Querying each partner's GO
record instead (scratchpad/reciprocity.py, paging asserted against
numberOfHits) shows the missing curation is not a quirk of this gene's entry —
the targeted papers produced nothing on either side:

Five targeted interaction papers, ten opportunities across both partners, zero
human annotations. The single counter-example is the one that matters: MGI
curated PMID:35671117 reciprocally — mouse Fam114a1 IPI with Q9WVK0 and
mouse Agtrap IPI with Q9D281 — and also gave Fam114a1 GO:0038166 by IMP. So
the same paper yields a curated, reciprocal, experimentally-grounded pair in
mouse and nothing at all in human.

Two consequences worth stating. First, this is a curation defect rather than
a biology one, which is the distinction AFF4's review drew; the diagnosis for
AGTRAP is "uncharacterised in GO", not "over-annotated", even though 181 rows are
marked over-annotated — those two coexist here because the 188 screen rows and
the missing targeted rows come from different places. Second, human AGTRAP's only
GO:0038166 row is an InterPro family-name mapping, while mouse Fam114a1 — its
partner — holds GO:0038166 by IMP. The experimental grounding for
angiotensin-signalling involvement exists in the neighbourhood; it just never
reached this gene.

The three PITPNC1 back-references that do exist on the partner side
(PMID:16189514, PMID:19060904, PMID:25416956, all IntAct GO:0005515 IPI)
are the same CCSB screens already counted here — reciprocity of the screen, not
of the targeted study.

Round 2: MeSH organism indexing settles a species question the abstracts do not

The reviewer flagged a real internal inconsistency: GO:0031702 was coded IPI
on PMID:12960423 while GO:0002092 was coded ISS on PMID:11162453
"because the abstracts do not state the species" — the same situation, two
different codes. And it added a fact I had seen and failed to act on: that
abstract numbers the protein 14-36 / 55-77 / 88-108 with a cytoplasmic tail at
109-161, and 161 aa is mouse Agtrap, not the 159-aa human protein.

The tiebreaker neither of us had used is NLM's own MeSH organism indexing,
which discriminates these papers cleanly:

Reproduce the table with one call (the cached publications/PMID_*.md records
carry title, authors and abstract but no MeSH headings, so this cannot be
checked from the repo alone — worth capturing in the cache now that the indexing
has done real curatorial work here):

# Per-PMID, because the load-bearing entries are ABSENCES: a single efetch over
# all five ids yields an unattributed union in which 'Animals' appears (from the
# other four papers), so a merged query cannot show that 12960423 lacks it.
for id in 12960423 10358057 11162453 15757644 27015675; do
  printf 'PMID:%s\t' "$id"
  curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=$id&retmode=xml" \
    | grep -o '<DescriptorName[^>]*>[^<]*</DescriptorName>' \
    | sed 's/<[^>]*>//g' \
    | grep -Ex 'Humans|Animals|Mice|Rats|HEK293 Cells' \
    | paste -sd, -
  sleep 1
done

Output when run (this is the table above, and it was run before being recorded):

PMID:12960423   Humans
PMID:10358057   Animals,Mice
PMID:11162453   Animals,Rats
PMID:15757644   Animals,Mice
PMID:27015675   Animals,HEK293 Cells,Humans

So the two codes are consistent once the indexing is consulted rather than the
abstract alone: 12960423 is indexed as human work, 11162453 as rat work. NLM
essentially always adds Animals when animal material is used, so its absence
from 12960423 is meaningful, and it agrees with UniProt's ECO:0000269 on the
human entry.

This does not dissolve the numbering conflict, and the review now says so in
the row rather than leaving it to be found. The likeliest reconciliation is that
the authors kept the residue numbering of the sequence first published in
PMID:10358057 while working with human constructs — but PMC284805 yields no
extractable full text (XML restricted, HTML empty, no PDF link), so it cannot be
confirmed. IPI is retained on explicit deference to the curator, with the
condition stated: if the full text shows rodent constructs, the code becomes
ISS with Q9WVK0 and only the code changes, not the term.

PMID:27015675 cuts the other way and is recorded as such: it carries Humans
and HEK293 Cells alongside Animals, so the SERCA2a binding measurements may
have used human constructs even though the calcium-uptake and knockout work did
not. ISS is the conservative reading there, flagged for revisiting.

Also from round 2: both ISS rows transfer from a mouse orthologue that holds
neither term, i.e. ISS from nothing. Rather than silently recode them, the mouse-
side annotations they should descend from are now requested explicitly in
suggested_questions, and the GO:0043495 differentia problem is filed as an
ontology request rather than left as a per-gene caveat. The PITPNC1 rows now say
why that membrane-anchoring activity is deliberately not elevated to a core
function — one PMA-dependent experiment, no measured consequence for AGTRAP
itself — which the PR body had described as "the anchor" without the review ever
saying so.

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

• Retractions / errata / expressions of concern: all 21 PMIDs relied on were
  checked via CommentsCorrections/RefType on each cited article's own PubMed
  record and publication types. None is retracted, corrected, or under an
  expression of concern; the only hits are three benign CommentIn entries
  (25416956, 25910212, 26871637).
• Paralogue transfer: PTHR16521 contains no paralogues — all 5 reviewed
  members are AGTRAP orthologues — so WRONG_ORTHOLOG_OR_PARALOG does not apply
  to any row here. Both Compara donors (Q9WVK0 mouse, Q642A2 rat) and the
  PAINT seed (MGI:MGI:1339977) are true orthologues.
• Dead/obsolete accessions: all 132 GOA partner accessions resolved to live
  UniProt entries with a gene name, as did both distinct UniProtKB:
  WITH/FROM accessions on the 11 non-binding rows — Q9WVK0 (3 rows) and
  Q642A2 (1 row). (An earlier draft of this line said "all 4 WITH/FROM protein
  accessions", conflating 4 rows with 4 accessions; there are 2. The
  O15507 failure mode from ACTR10 — a dead accession that answers every query
  with a silent zero — is absent here.)
• IBA landing above its donor (ACRV1 pattern): mouse Agtrap holds
  GO:0005886 plasma membrane by IDA and the human IBA is at the same term, so
  no downward MODIFY is warranted.
• Sibling-review inconsistency: no paralogue reviews exist to cross-check;
  the nearest precedent is the treatment of CCSB Y2H rows in ADAMTSL5 (followed)
  vs ADAMTSL4 (diverged from, deliberately, and said so).

affinage recall

gates_passed: True, faith_pct: 100, 21 citations, all numeric PMIDs, no
bioRxiv-style ids. It supplied the decisive partner-titled papers
(PMID:21728994, PMID:27015675, PMID:35414770) — better recall than this
campaign's 52% average. It missed PMID:35671117 (FAM114A1, the only
orthogonal validation of a HuRI hit) and PMID:10425188 / PMID:17033689, the
two donor references that turned out to carry the review's biggest findings.
Neither of the latter is findable from an AGTRAP-keyed literature search; both
came from querying donor evidence in QuickGO.

Scripts

Reproducible helpers used for this review live in the session scratchpad, not in
the repo, because they are one-off queries against live services rather than
analyses whose outputs are committed: partners.py (resolve every WITH/FROM
partner, assert 199 GOA rows and 0 unresolved accessions), topology.py
(compartment classification), intact2.py (IntAct evidence/method/publication
tabulation with a len(results) == totalElements assertion).