ACBD3 (GCP60, PAP7) — review notes

PAINT no-IBA project review, using the affinage deep-research provider
(gates passed, pairwise: win) plus UniProt Q9H3P7, the GOA TSV and the primary literature.

Headline, corrected: the ACB domain does bind acyl-CoA

(Rounds 1–8 of this review said otherwise. This section is the retraction.)

GO:0000062 fatty-acyl-CoA binding (IEA, InterPro) comes from the ACB domain at residues
83–174. Rounds 1–8 of this review called it an over-annotation and asserted that no acyl-CoA
ligand had ever been reported for ACBD3. That was false, and the correction is now
KEEP_AS_NON_CORE.

The source that refutes it was already cited here, in a sentence I quoted the second half of.
PMID:38134218's introduction reads:

PMID:38134218

I quoted the tail of that same sentence — "the GOLD domain and its extended UR interact with
multiple different golgins…" — for the domain map, and never read the head of it, which names the
ligands. Rule 1 (quote to the end of the interpreting clause) has a mirror image: read the
start of the sentence too. The primary source it points to is Soupene and Kuypers 2015
(PMID:26290611), whose abstract makes the ACBD3 claim by contrast:

PMID:26290611

That paper is titled for ACBD6, which is why it never surfaced in any ACBD3-keyed search — and
why the campaign rule about not treating a silent record as evidence of absence exists.

And the ACB domain has a gene-specific function too. PMID:23166793: ACBD3 binds SREBP1 directly
and blocks its maturation PMID:23166793. The ACB-containing N-terminus plays an important part in that effect — but only a
part, since deleting it attenuates FASN-promoter suppression from 76% to 40% rather than abolishing
it, and the co-IP itself used full-length ACBD3 PMID:23166793. The lipogenic readout is real but weaker than the
abstract implies — unchanged at 48 h, reduced only at 72 h PMID:23166793.

Why non-core rather than accepted as core. The binding is demonstrated, so the InterPro
inference is corroborated. But no acyl-CoA-dependent step in ACBD3's Golgi activity is
known. This is a statement about the ligand, not the domain — the ACB domain has its own
functions (SREBP1 restraint, and the FAPP2 attribution; see the domain map). What is absent is
any demonstrated consequence of acyl-CoA binding. Deleting the ACBD domain does not prevent Golgi
localisation PMID:38134218, and the domain is
dispensable for 3A-mediated PI4KB recruitment — the enteroviral assay, not host recruitment on
its own
PMID:30755512, with Q+GOLD sufficient PMID:30755512. A real molecular
function; not the core one.

How the error survived eight rounds

Three separate guards failed, and all three failed the same way — by treating an absence as a
finding:

  1. The affinage record's silence was used as evidence. "Synthesises 30+ primary papers and
    never mentions acyl-CoA once" was quoted as support for the over-annotation call, in four
    places. It is a coverage gap in one provider's summary, nothing more. The campaign brief says
    this in as many words — an empty or silent affinage record is NOT evidence that literature is
    absent     — and I applied that rule to other genes' records while breaking it here.
  2. A domain-name-is-not-an-activity heuristic was applied without checking the converse. The
    ABHD8 analogy (fold name propagating into GO as activity) was apt in form and wrong in fact:
    here the domain name is backed by biochemistry. Pattern-matching to a known failure mode is
    not evidence either.
  3. The papers that had the answer were read only in the parts that concerned other claims.
    PMID:38134218 was read for the golgins; PMID:30755512 was read from its abstract, whose ACB
    sentence is the dispensability result — while the results paragraph that names the two ACB
    functions, with citations, sits at line 164. The reviewer found it by reading the paragraph.

The generalisable rule, added below: a negative claim about the literature needs a positive
search, not a silent source. Before writing "no X has been reported", search for X by name — and
search under the paralog names too, since the paper that had this answer is titled for ACBD6.

The informative/uninformative inversion

Worth naming because it inverts the usual pattern in this campaign:

Term Source Informative?
GO:0043495 protein-membrane adaptor activity IBA ✅ exactly right
GO:0034237 PKA regulatory subunit binding IPI ✅ specific partner
GO:0005515 protein binding ×6 IPI, from the focused mechanistic papers ❌ says nothing

The bare rows are the ones from the papers that did the work — PI4KB structural NMR, the
picornavirus 3A studies. The phylogenetic inference is more informative than six experimental
annotations. Three PI4KB rows are MODIFYed to GO:0043495.

Partners, resolved

Partner Papers Verdict
PI4KB (Q9UBF8) 4 core — Q-domain recruitment, NMR of residues 241–308
Picornaviral 3A (Aichi O91464, polio P03300, …) 2 the same adaptor activity, hijacked
PRKAR1A (P10644) 1 already informatively annotated
TBC1D22A/B 3 bind the same Q domain, mutually exclusively with PI4KB → MODIFY
PPM1L 1 topically coherent, no follow-up → over-annotated

Core/non-core consistency

Two annotations are kept non-core, and in both cases the term is right while the evidence is
not core-strength:

Correction: PI4KB binds the Q domain, not GOLD

An earlier draft of this review stated seven times that PI4KB is recruited through the GOLD
domain. That is wrong, and UniProt says so explicitly:

Interacts (via Q domain) with PI4KB (via N-terminus)
Interacts (via Q domain) with TBC1D22A and TBC1D22B;
  interactions with PI4KB and with TBC1D22A and TBC1D22B are mutually exclusive
-!- SUBUNIT: (Microbial infection) Interacts (via GOLD domain) with 3A proteins

FT DOMAIN 384..526 GOLD; the NMR structure of the PI4KB complex is of residues 241–308,
which is the Q domain.

How the error got in: the affinage narrative says PI4KB is recruited "through its GOLD
domain", and I used that sentence as supporting_text without checking it against UniProt —
on rows that listed the UniProt file in the same supported_by block. The campaign's own rule
is that a provider sentence is a lead, not evidence; I did not apply it. The quote is removed
everywhere and the affinage reference_review is marked MISCITED for that claim.

Two things the correction bought:

  1. The Q domain is a switch. PI4KB and TBC1D22A/B compete for the same surface. So the
    TBC1D22 rows are not "uncharacterised screen hits" as I first called them — they engage the
    very surface that defines the protein's molecular function, and the competition is
    presumably how the Golgi arbitrates between PI4P synthesis and Rab-directed traffic. Those
    rows are now MODIFY rather than MARK_AS_OVER_ANNOTATED.
  2. Viral 3A uses a different surface from the host kinase — GOLD, not Q. That explains how
    picornaviral 3A can clamp ACBD3 onto replication organelles without displacing PI4KB, which
    is the whole point of the hijack.

The PKA regulatory subunit: compartment-specific, not contested

An earlier version of these notes and of the review called this "genuinely contested" and said
"the two sources disagree". That was over-dramatised, and a reviewer was right to push back: the
2023 paper reconciles the split itself. The only claim that survives as a genuine problem is a
negative one, and it belongs to UniProt.

Source Evidence Claim
GOA GO:0034237 IPI, WITH/FROM UniProtKB:P10644 IPI binds PRKAR1A, i.e. RIα
UniProt SUBUNIT line (By similarity), ECO:0000250 binds RI-alpha; does not bind RI-beta or RII-alpha
PMID:37044218 (2023, human) experimental GOLD domain binds RIIα; RIα not detected in the same pull-down, but explicitly not excluded

The 2023 pull-down detects RIIα and not RIα PMID:37044218, but the authors do not read that as excluding RI.
Two sentences in the same paragraph do the reconciling for us:

So "which subunit" is not a live dispute: the Golgi pool is RII-anchored, the mitochondrial pool
RI-anchored, and both records can be right. Per the project rule about not overruling curators, the
curated RIα IPI stands — I have not read PMID:17911601's full text, which is abstract-only in the
cache and whose title is about ezrin.

What does survive. UniProt's (By similarity) statement that ACBD3 does not interact with
RII-alpha is contradicted head-on by a human pull-down that does detect RIIα. A by-similarity
negative standing against a human experimental positive is the weakest configuration in the
record, and that specific point remains a UniProt correction to report.

GO:0034237 is unaffected either way: the term is agnostic about which R subunit binds.

Why this matters beyond the isoform. The 2023 paper places the RII interaction in the GOLD
domain PMID:37044218. Combined with the Q-domain
correction above, the UR/GOLD split collapses — leaving three functional regions, not four:

An earlier version of these notes called this map "clean and non-overlapping" and set UR
(golgins) against GOLD (3A, PKA, SEC22B). That was wrong, and a reviewer caught it using the
very paper I had just cited: PMID:38134218 places 3A recruitment at "the UR of the GOLD
domain" PMID:38134218, its Figure 1D
puts I380/K381 four residues downstream of T376 PMID:38134218, and the study's premise is competition
PMID:38134218. Having adopted the convention that "the UR of the GOLD domain" is
one locus, I could not then use UR and GOLD as contrasting labels.

Cross-referencing the two papers makes it sharper still — but the residue has to be read off the
right experiment. An earlier version of this section said PMID:37044218 "maps RIIα docking to
K381", citing PMID:37044218. That over-reads a
proline scan. The mutants are prolines chosen as helix breakers PMID:37044218, and the authors read the result structurally, not as a contact
PMID:37044218. Two sentences later they nominate a contact from the crystal structure:
PMID:37044218, with Phe383 buried.

Second correction, same sentence, opposite direction. A later version of this paragraph
then asserted that the RIIα contact residue is I380. That over-reads too, and the
reviewer caught it in the half of the sentence I had again cut away: I380P had almost no
effect. Quoting the sentence whole, as the rule below demands:

PMID:37044218

And the Ile380 nomination is not a measurement. It comes from inspecting the apo GOLD
structure — PDB 5LZ1, ACBD3 alone, no R subunit PMID:37044218 — it is hedged ("likely
involved in binding"), and it is offered to explain why the buried Phe383 does not make
ACBD3 RI-selective. Its only direct test in that paper is the null I380P.

What the two papers jointly support is the helix, not a residue. The GOLD domain has one
α helix PMID:37044218, K381P breaks it and abolishes RIIα binding, and both 3A contact residues
PMID:38134218 lie inside it. So residues 374–381 remain
a shared hub for the golgins, PKA RII and viral 3A — as overlapping surfaces, which is a
weaker and correct claim. Whether PKA and 3A share a contact residue is exactly what
suggested_experiments[3] is designed to test, so it must not be written as settled.

What survives unchanged is the Q-domain conclusion: 3A does not displace PI4KB, because neither
the UR nor GOLD is the Q domain. What changes is the mechanism of the hijack — it is most
likely displacement of ACBD3's own anchors, giantin/golgin-45 and PKA, rather than occupancy
of a spare site. That is now the leading suggested_experiment, with 3A-vs-golgin competition
foregrounded over 3A-vs-PKA because the paper states the golgin premise itself.

A gap the GOA record has, found by reading the paper the record cites

ACBD3 has no retrograde-transport annotation at all. The GO:0006888 IMP cites
PMID:37044218, but that paper's subject is Golgi-to-ER retrograde recycling: anterograde cargo
arrival is the trigger, and what ACBD3 controls is whether retrograde transport runs
constitutively PMID:37044218.

So the record captured the trigger and missed the regulated process. GO:2000156 regulation of
retrograde vesicle-mediated transport, Golgi to ER is proposed as a NEW IMP — it has no
children, so it is the most specific term available, and the regulation parent is right because
ACBD3 is not transport machinery: it sets whether the pathway is cargo-gated. The existing
GO:0006888 IMP is left ACCEPT, since the addition is additive rather than a correction to a
curator's experimental call.

The MWT/giantin claim: checked against the primary paper, and it holds

The first three versions of this review stated the Golgi-targeting mechanism as "an MWT motif that
binds the golgins giantin and golgin-45" in the top-level description and in seven row summaries,
sourced only to ACBD3-deep-research-affinage.md:68. A reviewer flagged that as the same failure
mode as the GOLD/PI4KB error above: a provider domain-assignment sentence outrunning the UniProt
file cited beside it, since ACBD3-uniprot.txt:366-368 says "The C-terminal GOLD domain is
essential for giantin binding".

That objection was procedurally right and substantively wrong, which is the interesting part.
PMID:38134218 (Stalder et al. 2024, Mol Biol Cell) is now fetched and cached with full text,
and it supports the claim:

So the two sources are reconciled, not in conflict, and the review now says so with the primary
paper cited on every row that makes the claim. The real defect was citing a provider narrative for a
mechanistic claim instead of the primary paper it named - the campaign rule, again, and this time
the rule caught a claim that happened to be true. Worth recording: "the provider said it" is not
evidence even when the provider is right, because you cannot tell which case you are in without
fetching the paper.

What else the 2024 paper brought

Reading it added three things beyond the fix:

  1. SCFD1 (SLY1) is an essential upstream recruitment factor. Its CRISPR knockout strips ACBD3
    from the Golgi PMID:38134218, and takes PI4KIIIβ with it. SEC22B is a different case:
    ACBD3 binds it, but its knockout PMID:38134218 wrecks the Golgi generally, so a
    specific requirement for ACBD3 recruitment cannot be read off it. The review therefore
    attributes the upstream requirement to SCFD1 alone and annotates SEC22B only as a binding
    partner.
  2. The UR plus GOLD domain binds the longin domain of SEC22B PMID:38134218 — the mapped
    fragment is 328–528, so GOLD alone is not the right attribution. With the SNARE and
    transmembrane domains excluded by truncation. Annotated as a NEW GO:0000149 SNARE binding
    IPI. Deliberately not annotated as complex membership, which the paper does claim: the assays
    are binary co-IPs from overexpressing cells, and the cytosolic MWT374-376>AAA mutant binds
    SEC22B more strongly than wild type, so binding and Golgi residence are separable. Nor is it
    listed in core_functions: SEC22B binding serves ACBD3's own delivery to the cis-Golgi,
    upstream of what ACBD3 does for the cell, and core_functions records the two outputs (PI4KB
    recruitment, PKA anchoring). An informative MF term need not be a core function.
  3. ACBD3 is required for PI4KB to reach the Golgi, one-way. ACBD3 knockout loses Golgi
    PI4KIIIβ; PI4KIIIβ knockout leaves ACBD3 in place. The gene had no protein-localisation
    process term at all, so GO:0034067 protein localization to Golgi apparatus is proposed as a
    NEW IMP - the process counterpart of the GO:0043495 adaptor MF it already carries.

Together these mean the UR/GOLD surface is not viral-only: it has two host partners
(SEC22B longin domain, PKA RIIα) as well as picornaviral 3A — and, per the correction above,
they are not on private sub-sites but share residues 374–381 with the golgin motif.

Recruitment, as it now stands

SCFD1 (SM protein); ACBD3 binds SEC22B       <- step 1, UR+GOLD binds SEC22B longin domain
        |                                       SCFD1 KO -> ACBD3 cytosolic
        v
giantin  OR  golgin-45   (redundant)         <- step 2, both bind MWT374-376 in the UR
        |                                       double KO -> ACBD3 off the Golgi
        v
ACBD3 on cis/trans-Golgi membranes
        |
        v
PI4KB (via Q domain) / PKA RII (via the UR+GOLD α helix, ~379-383) / FAPP2 / PPM1L / STING

Sequential, not parallel: ACBD3-giantin binding drops in SCFD1-KO cells, while ACBD3-SCFD1 binding
is unchanged in the golgin double KO PMID:38134218.

Three rules, two of them learned on one sentence

The same sentence of PMID:37044218 produced two opposite errors in consecutive rounds, which is
where rules 1 and 2 come from —
first reading the loss-of-binding proline mutant K381P as a contact residue, then reading
I380 as the contact although the same sentence reports I380P as having almost no effect.
Both halves were verbatim; both readings were wrong.

  1. Quote to the end of the interpreting clause. If the sentence continues with
    suggesting, while, although, whereas or a comparison, the continuation is part of
    the result. Cutting at a comma inverts meaning as often as it shortens it.
  2. A mutant's phenotype names an element, not a residue, unless the substitution is
    conservative and the structure is holo.     Two corollaries, one for each error:
  3. a proline substitution that abolishes binding implicates the secondary structure, not
    the side chain;
  4. a substitution with no effect does not nominate its residue as a contact either — it is
    evidence against, or at best uninformative.
    Apo structures propose contacts; they do not map them.
  5. When two sources draw different boundaries for one region, resolve to residue numbers
    and treat the domain names as commentary. (Rounds 3–5 were all boundary-label
    collisions; see the UR/GOLD correction above.)

One caveat on my own numbers. The ≈379–383 span I use for the helix is an inference,
flagged here under rule 3 because it is the same class of thing rule 3 warns about. Neither paper
gives the helix endpoints. What is sourced is: the GOLD domain has one α helix; Q379, I380 and
K381 were the residues substituted as helix-breakers; and Phe383 is buried. 379–383 is bracketed
from those facts, which is why it is written as "roughly" everywhere it appears.

  1. A negative claim about the literature needs a positive search, not a silent source.
    "No X has been reported" is a claim about all of PubMed, and no summary, however dense,
    can support it. Search for X by name — and under the paralog names, since the paper
    carrying this gene's acyl-CoA data is titled for ACBD6 and never surfaced in an
    ACBD3-keyed search.

One sentence, three errors

publications/PMID_38134218.md:72 has produced a distinct error in each of rounds 4, 9 and 10:

The ACBD domain oligomerizes upon binding to C18:1-CoA or C16:0-CoA (Soupene and
Kuypers, 2015)  and recruits the membrane-shaping protein FAPP2 (Liao et al., 2019),
  ^-- round 9 missed this head       ^-- round 10 missed this middle
the CAR-Q domain recruits PI4KIIIβ (Klima et al., 2016), and the GOLD domain and its
extended UR interact with multiple different golgins ...
                                  ^-- round 4 quoted only this tail

Round 4 quoted the tail for the domain map. Round 9 read the head and inverted the acyl-CoA
call. Round 10 read the clause between them and found the ACB domain has a protein client.
Each time the fix was correct and each time I stopped at the clause I needed.

Rule 1 restated, generally: quote the whole sentence, then decide what to use. Not "to
the end of the interpreting clause" — to both ends. A sentence that enumerates domains is
a table in prose; reading one row of a table is not reading the table.

It is also worth naming why this sentence in particular: it is a secondary sentence, an
introduction compressing four primary papers into four clauses. High information density,
each clause a lead worth a fetch, and no redundancy to catch a partial read. Reviews should
treat an introduction's domain enumeration as a work list, not as a citation.

Round 11: the same shape one level further in — a figure legend for a paragraph

Round 10 annotated GO:2000639 and said the cached text carried no quotable loss-of-function
result, so the call rested on overexpression alone. It does carry one, three occurrences of it:

PMID:23166793 — with the authors' conclusion
PMID:23166793 and a specificity control
PMID:23166793.

What I read was the Figure 3 legend, which describes the experiment and not its outcome. The
result is the paragraph immediately above it. So the progression across rounds is:

Round What stood in for the evidence
8 the abstract stood in for the results paragraph
10 one clause stood in for the sentence
11 the figure legend stood in for the paragraph

Each is a shorter, tidier text that sits next to the real one and reads as if it were complete.
Rule 5: a caption is not a result. Figure legends, abstracts and introductions are all
summaries; when a claim turns on what an experiment showed, the sentence that reports the outcome
is in the body text, and it is the only one that counts. Corollary: this cut against my own
annotation — the correction strengthened GO:2000639 from one-directional to bidirectional — so
the habit is not conservative, it is just wrong in whichever direction the summary happens to lean.