ADPRS (ARH3 / ADPRHL2, Q9NX46) — review notes

Identity: which symbol is current

The worklist, UniProt and GOA do not all use the same name, so this was settled first.

So ADPRS is current; ADPRHL2 is the retired symbol and ARH3 the protein
nickname used throughout the literature. Most primary papers, and UniProt's own
DE RecName, still say ARH3, and the two 2018 disease papers are titled for
ADPRHL2 — worth knowing when searching.

The worklist file is named human-no-IBA-simple.csv and ADPRS carries six IBA rows
(GO:0004649, GO:0005634, GO:0005739, GO:0071451, GO:0140290, GO:0140292),
matching UniProt's DR PAN-GO; Q9NX46; 6 GO annotations based on evolutionary models.
Another instance of the stale-snapshot problem; the IBA rows turned out to be the
highest-yield rows on the gene.

Row reconciliation, done before reviewing

GOA TSV data lines                57
distinct (term, ev, ref, qual, with/from)  57
fetch-gene stub `- term:` entries 56

One collapse, and it is the documented GOAValidator.seed_missing_annotations key
omission: the two GO:0005515 IPI rows from PMID:32296183 differ only in WITH/FROM
(UniProtKB:O95271 vs UniProtKB:Q9NQX1-2), which is not part of the seeding key. Both
were restored so each partner gets its own verdict. The review has 57 entries and
no NEW proposals, so the counts reconcile exactly.

Verdicts

Counted from the emitted YAML, not by hand: 47 ACCEPT + 7 MODIFY + 3
MARK_AS_OVER_ANNOTATED = 57, and 26 MF + 19 CC + 12 BP = 57.

aspect ACCEPT MODIFY MARK_AS_OVER_ANNOTATED total
MF 22 2 2 26
CC 19 0 0 19
BP 6 5 1 12
all 47 7 3 57

The seven MODIFY rows: GO:0004553 → GO:0140292; GO:0004649 IMP PMID:33769608 →
GO:0140292; GO:0071451 ×3 → GO:0070301; GO:0060546 ×2 → GO:0062099.
The three MARK_AS_OVER_ANNOTATED rows: GO:0006287 TAS, and GO:0005515 ×2.
No REMOVE, and no NEW proposals, so existing_annotations equals the GOA row count
exactly.

Finding 1 — the annotated ROS species is the wrong one

GO:0071451 cellular response to superoxide sits on ADPRS three times (IMP, IBA, ARBA
IEA). Every experiment behind it used hydrogen peroxide.

GO:0071451 and GO:0070301 cellular response to hydrogen peroxide are siblings
under GO:0034614 cellular response to reactive oxygen species — verified by fetching
both ancestor closures; neither contains the other. So this is not a broad parent that
quietly covers the data, it names a different chemical species.

Reach: PANTHER:PTN008564042 was fully paginated — 132 annotations over 33 gene
products, all ADPRS orthologues, every one carrying GO:0071451. Fixing the single
mouse MGI row would retract the error from all 33 at once. Filed as a question to MGI
and GO Central.

Finding 2 — parthanatos annotated as necroptosis

GO:0060546 negative regulation of necroptotic process reaches human ADPRS twice (ISS
from Q8CG72, plus the Compara IEA mirror of the same mouse row). QuickGO confirms the
mouse row is an IDA from PMID:30830864.

GO:0060546 regulates GO:0070266, whose definition requires "activation of
receptor-interacting serine/threonine-protein kinase 1 and/or 3 (RIPK1/3 …) and …
critical dependence on mixed lineage kinase domain-like (MLKL)". What ARH3 restrains
is parthanatos, which uses none of those:
PMID:34479984, and
PMID:24191052.

GO itself records the ambiguity rather than asserting the identity — the comment on
GO:0097527 necroptotic signaling pathway says PARP-dependent cell death "is sometimes
referred to as PARP-dependent cell death or parthanatos; it is still being debated if it
constitutes an independent cell death modality." That is why the action is a one-step
generalisation to GO:0062099 negative regulation of programmed necrotic cell death
(verified to be an ancestor of GO:0060546, so nothing new is asserted) rather than a
removal, with a GO term request for parthanatos filed alongside.

Finding 3 — a paper with no poly(ADP-ribose) in it, annotated to PARG activity

GO:0004649 poly(ADP-ribose) glycohydrolase activity IMP PMID:33769608. Reading the
full text: the paper synthesises mono-ADP-ribosylated Ser/Thr/Cys peptides and asks
which hydrolase reverses each linkage. The only ARH3 result is
PMID:33769608. No polymer is
used anywhere.

GO:0004649 is correct for this gene from eleven other rows; this citation does not
substantiate it, and the same experiment substantiates GO:0140292 exactly. Hence
MODIFY, not REMOVE. This is a deliberate same-term/different-action divergence and the
repo validator warns about it; the divergence is per-reference rather than per-term and
suppressing it would have meant discarding the finding.

Ten other GO:0004649 rows were checked for the same defect and all are genuine PAR
studies (PMID:34321462, PMID:34019811, PMID:34479984, PMID:33894202,
PMID:22433848, PMID:16278211, PMID:17075046, PMID:30830864, plus the IBA, the
ARBA IEA and the Reactome TAS). Reported as a mostly-negative sweep, so the next
reviewer knows all twelve were looked at rather than one.

Finding 4 — a mitochondrial reaction inheriting a nucleoplasmic BER term

GO:0006287 base-excision repair, gap-filling TAS Reactome:R-HSA-110373. Traced end to
end in Reactome's own data:

GO:0006287's definition requires an apurinic endonuclease degrading bases and a
polymerase synthesising a patch. ARH3 does neither. Reference-projection test, fully
paginated: R-HSA-110373 gives GO:0006287 to exactly five entities — POLB, LIG1, FEN1,
PARG and ADPRS — the first three being the actual gap-filling machinery.

Not REMOVE: ARH3 does participate in the ADP-ribose signalling that accompanies
single-strand-break repair, and GO:0006281 DNA repair (IMP, human, PMID:30045870)
already carries that at the right granularity.

Contrast the other Reactome reference on this gene: R-HSA-8952903 projects to
1 entity, 2 annotations — a gene-specific curated reaction, correctly used for
GO:0004649 and GO:0005759. The same database is both right and wrong on this gene,
and the discriminator is reaction-level versus pathway-level export.

Finding 5 — a GO branch-placement question, with its own positive control

term acceptor atom under GO:0016799 N-glycosyl under GO:0004553 O-glycosyl
GO:0003875 ADP-ribosylarginine-[protein] hydrolase guanidino N yes no
GO:0140292 ADP-ribosylserine-[protein] hydrolase serine hydroxyl O yes no
GO:0140293 ADP-ribosylglutamate-[protein] hydrolase carboxylate O yes no
GO:0004649 poly(ADP-ribose) glycohydrolase ribose 1''–2' O no yes

GO:0003875 — ADPRH's term — is the internal control: arginine's acceptor really is a
nitrogen, so its placement is right. The serine and glutamate terms share that placement
while their acceptor atoms are oxygens, and UniProt/RHEA name the substrates
O-(ADP-D-ribosyl)-L-seryl-[protein] (RHEA:58256) and
5-O-(ADP-D-ribosyl)-L-glutamyl-[protein]. The literature is explicit:
PMID:29907568.

Practical consequence on this gene: GO:0004553 is annotated from PMID:30045870, which
measured only serine-linked hydrolysis, and because the two terms sit in disjoint
branches the generalisation a curator would expect does not hold. Raised as a question
for GO rather than curated around; the MODIFY to GO:0140292 stands on the simpler
ground that it is the activity the paper measured.

Caveat recorded honestly: UniProt also assigns EC 3.2.2.- (an N-glycosylase subclass)
to the serine activity, so the convention may be deliberate rather than an error. Hence
a question, not an assertion.

Finding 6 — both protein-binding rows are one two-hybrid screen

GO:0005515 ×2, both IPI PMID:32296183 (HuRI). Expanding IntAct:

partner records from PMID:32296183 methods MI score
TNKS (O95271) 3 two hybrid array; two hybrid prey pooling approach; validated two hybrid 0.56
PRDM5 (Q9NQX1-2) 3 two hybrid array; two hybrid prey pooling approach; validated two hybrid 0.56

Third recorded instance of NbExp=3 is one screen counted three ways (after ACRV1 and
ADAMTSL5). Both partners resolve to reviewed canonical Swiss-Prot entries at full
length (TNKS1_HUMAN 1327 aa; PRDM5_HUMAN 630 aa), so neither is a TrEMBL/ORFeome
substitution — a check reported as negative rather than skipped. Likewise the topology
check is negative: both partners and ADPRS are nuclear/cytoplasmic, so there is no
compartment objection of the ACRV1 kind.

The TNKS hit is superficially attractive — an ADP-ribose eraser meeting a PARP writer —
which is why it was checked rather than accepted. Tankyrase-1's UniProt catalytic
activities install ADP-ribose on aspartate and glutamate side chains
(4-O-(ADP-D-ribosyl)-L-aspartyl-[protein], 5-O-(ADP-D-ribosyl)-L-glutamyl-[protein]),
a linkage class reversed by the macrodomain hydrolases (GO:0140293), not by ARH3, whose
measured scope is Ser/Thr/Tyr, PAR and O-acetyl-ADP-ribose. Promiscuity: TNKS has 186
IntAct partners and PRDM5 96, against ADPRS's 35. PubMed returns no study of an
ARH3–tankyrase or ARH3–PRDM5 interaction. MARK_AS_OVER_ANNOTATED, not REMOVE — the
interactions are unreplicated and uninformative, not demonstrated false.

Checks that came back negative (recorded so they are not re-run blindly)

Affinage record

gates_passed: True, 18 citations, faith_pct: 100. All 18 were resolved against
PubMed and every one is genuinely about ARH3/ADPRS — none of the cross-gene
miscitation seen on the sibling ADPRH. Two are preprints and were excluded from anything
load-bearing: PMID:bio_10.1101_2024.08.28.610034 is a bioRxiv DOI in a PMID-shaped
field (not a PubMed id at all), and PMID:36945462 is a real PubMed id whose record
type is Preprint (bioRxiv, the Arh3 cardiac study).

Recall assessment: the record found the mechanistic literature well, but it missed the
two papers that decide the review's largest findings — PMID:24191052 (the mouse
donor experiment behind both the superoxide and the necroptosis terms) and
PMID:33769608 (the mis-attributed GO:0004649 row). Both were found from the GOA
WITH/FROM chain and from UniProt's own reference list, not from the provider. Consistent
with the campaign note that gates_passed measures precision, not recall.

No affinage sentence is used as supporting_text anywhere in the review; the leads it
supplied (α-NAD+, Tyr-ADPr) were re-grounded on PMID:31599159 and PMID:39342999
directly. The validator therefore warns "No annotations reference available deep research
files", and that warning is left standing deliberately.

GO gaps proposed

  1. ADP-ribosylthreonine-[protein] hydrolase activity. Measured
    ([PMID:33769608, above]); UniProt's FUNCTION says "proteins ADP-ribosylated on serine
    and threonine"; no GO term exists between the arginine, serine and glutamate terms.
  2. poly(ADP-ribose) catabolic process. ADPRS degrades protein-free PAR, which
    GO:0051725 protein de-ADP-ribosylation cannot cover by its own definition, and free
    PAR is a distinct signalling species (the AIF-releasing death signal). Searched
    QuickGO and OLS for "poly(ADP-ribose) catabolic", "ADP-ribose catabolic" and
    "ADP-ribose metabolic process": absent. Positive control that the searches worked:
    they do return GO:1990966 ATP generation from poly-ADP-D-ribose.
  3. parthanatos. See Finding 2.

Also raised as questions rather than proposed, because each turns on a curation
convention rather than on missing data: whether GO:0072570 ADP-D-ribose binding should
be annotated (measured micromolar affinity —
PMID:17015823 — nine
ADP-ribose co-crystals, and recruitment to lesions depending on it, against the
convention that an enzyme is not annotated to binding its own product); and whether the
α-NAD+ hydrolase activity (RHEA:68792) needs an anomer-specific term, since GO:0003953 NAD+ nucleosidase activity is unqualified and the enzymes tested hydrolyse α- but
explicitly not β-NAD+.

A superseded negative, flagged so it is not re-imported

The 2006 structure paper states that hARH3 "efficiently de-ADP-ribosylates poly- but not
monoADP-ribosylated proteins" (PMID:17015823). That was overturned in 2017 —
PMID:28650317 — and the two results are compatible
once the substrate is specified: the 2006 assay used arginine-linked mono-ADP-ribose,
on which ARH3 genuinely is inactive — UniProt's FUNCTION comment states that it does not
hydrolyse ADP-ribosyl-arginine, -cysteine, -diphthamide or -asparagine bonds, citing
PubMed:16278211 and PubMed:33769608 (the sentence spans a CC continuation line in
ADPRS-uniprot.txt, so it is paraphrased rather than quoted). Reference marked DISPUTED with the
reconciliation in review_notes, so nobody turns it into a NOT annotation.

Incidental repository finding

genes/human/publications/PMID_12345.md exists on main (committed in 21c5e7489,
"batch9") — a stray, wrong-directory publication cache containing a 1976 paper about
tablet granulation. It is not in publications/, so nothing cites it, but any script
that locates the repo root by walking up for a publications/ directory resolves to
genes/human/ instead of the repo root. verify_adprs_claims.py hit exactly that and
failed loudly rather than reporting zeroes; the fix (anchor on publications/ and
genes/ together) is committed with a comment explaining why. Not deleted here because
it is outside this gene's scope.

Computed evidence

ADPRS-bioinformatics/verify_adprs_claims.py re-derives every number above from the
public APIs and the local publication cache, writes results.json, and renders
RESULTS.md from it (--check fails if the two disagree). It carries five break-tests,
including one that flips a single boolean to prove the report is sensitive to the
sibling-versus-descendant claim rather than merely to a blanked input. What it explicitly
does not mechanise is the reading judgement in Finding 3 — a phrase-presence check
cannot prove the absence of an experiment, and the file says so.

Round 2 — the reviewer found nine quotes that were verbatim and off-topic

The repo's reference validator checks that a supporting_text is a verbatim substring
of its cited paper. It does not check that the sentence is about the row it sits under.
Nine supported_by entries here passed every automated gate while citing something else —
a mitochondrial-matrix sentence under nucleus, one sequence-identity sentence reused
under cytoplasm and twice under PARG activity, a localisation sentence under an activity
row, and the paper's own title under both protein-binding rows. All nine are fixed.

Two of the fixes are not "find a better quote", because for four rows no better quote
exists in the cached record:

Those rows now carry a full_text_unavailable: true marker plus an explicit statement of
the limitation in reason, and the claim is anchored on a paper that does state it. A
recorded absence beats a verbatim quote about something else, which looks checked and is
not.

A fifth case is worth separating: GO:0005759 EXP PMID:34479984 has full text
available, and the word "matrix" occurs 0 times in it (against 89 for
"mitochondri"). The paper measures the organelle by MitoID imaging and fractionation;
UniProt's matrix assignment for this reference must rest on detail the text does not
spell out. Per CLAUDE.md an experimental call is not overruled from incomplete evidence,
so the row stays ACCEPT with the limitation stated, and the matrix claim is anchored on
PMID:17991898, which demonstrates PAR-degrading activity inside the matrix directly.

The guard, and its own three defects

ADPRS-bioinformatics/audit_row_quotes.py makes the rule executable: every row must have
a quote matching a topic pattern declared for its GO id (the stable entity), or a
full_text_unavailable marker and a stated limitation — both halves, or the escape
hatch becomes a bypass for the defect it guards.

Running it caught three more rows the reviewer had not listed, and three defects in the
guard itself:

  1. \bMg\b did not match MgA/MgB — the exact residues the GO:0000287 row is
    about. A word-boundary pattern that excludes the specific form of the word it is
    looking for.
  2. Ser-ADPr did not match Ser-linked, the Drosophila paper's wording.
  3. A MODIFY row was judged against the term it is moving away from. On
    GO:0004649 IMP PMID:33769608 the whole finding is that the quote is not about
    poly(ADP-ribose); the guard should test against the replacement term. Fixed by
    collecting topics from proposed_replacement_terms too — and by failing loudly if a
    replacement term has no declared pattern, rather than skipping it.

Break-test F runs the guard against the YAML at commit aa019d486 — the version that
actually shipped the defect — and it fires on precisely the GO:0005634 row. A self-test
proves the guards you thought of fire; running against the shipped defect is the stronger
claim.

A provider error inherited into core_functions

core_functions said the O-acetyl-ADP-ribose rate exceeds the poly(ADP-ribose) rate "by
orders of magnitude". PMID:17075046 says only
PMID:17075046 The 250-fold figure in the same
abstract is a different comparison —
PMID:17075046 — ARH3 against ARH1, not against PAR.

The affinage record makes exactly that conflation: "at a rate 250-fold faster than its
hydrolysis of poly(ADP-ribose)". So this is a provider error that reached the review
through background knowledge rather than through a quoted sentence — the campaign rule
"never quote an affinage sentence" does not protect against absorbing its arithmetic. Both
sentences are now quoted side by side so the distinction cannot collapse again.

Two suggestions judged, one taken and one declined