ADIG (adipogenin / SMAF1, Q0VDE8) — review notes

2026-09-26. PAINT no-IBA backlog. Deep research provider: affinage.

Headline

Every one of ADIG's 16 GO annotations is non-experimental (5 IBA, 2 IEA, 9 ISS), and all of
them descend from mouse Adig (UniProtKB:Q8R400 / MGI:MGI:2675492). So this is a pure
propagation review. What makes it worth doing carefully is that the donor set was
re-adjudicated in 2025 and GOA has not caught up.

The recall gap in the deep-research file

The affinage record's own trust gates passed (accession matches Q0VDE8, no non-human organism
token, pairwise self-evaluation win, faith 100%). Its citations are all real and on-topic —
precision is fine. Recall is not. Its most important dated finding, the seipin cryo-EM
work, is listed with PMIDs: — and journal bioRxiv:

[file:genes/human/ADIG/ADIG-deep-research-affinage.md "Adig directly interacts with seipin to form a rigid complex"]

That preprint (PMID:39211078, bioRxiv 2024-07-26 — resolved via NCBI esearch and now cached, so
the identifier is checkable rather than asserted) published in Science on 2025-11-06 as
PMID:41196993, doi 10.1126/science.adr9755 — resolved via the NCBI esearch API, not from
memory. Note the dates: the affinage record's own affinage_run_date is 2026-06-09, seven
months after the Science paper appeared, so this is not a stale-snapshot problem — the
published version was available and was not retrieved. This is a recurring pattern in the
campaign: the gates certify that what affinage returned is real, never that it returned
everything, so partners, family members and paralogs have to be searched independently.

Consequence: affinage's mechanism_profile still says

[file:genes/human/ADIG/ADIG-deep-research-affinage.md "localization: GO:0005811 lipid droplet"]

which this review declines, for the reasons below.

What PMID:41196993 settles

It is unusual in that it explicitly set out to adjudicate ADIG's contested localization:

PMID:41196993

Three orthogonal assays — EGFP fused at either terminus against ER-mCherry-KDEL; APEX2
electron microscopy at near-endogenous levels in Adig-KO adipocytes; digitonin protease
protection — converge:

PMID:41196993

UniProt has already absorbed this (Q0VDE8: "Endoplasmic reticulum membrane ... Single-pass
membrane protein", TRANSMEM 14..38, and SUBUNIT "Forms a dodecameric complex with seipin
(BSCL2)"). GOA has not.

Mechanism:

PMID:41196993
PMID:41196993
PMID:41196993

Nucleus: removed

Both nucleus rows trace to one 2005 result:

PMID:15567149

ADIG is 80 aa / ~10 kDa — far below the passive nuclear-pore diffusion limit — so an EGFP
fusion in the nucleoplasm says nothing about the endogenous protein, and the authors
themselves wrote "suggestive". PMID:41196993 named this claim and replaced it. These are
ISS/IBA rows, not experimental ones, so REMOVE here is not overruling a curator who read a
full text I cannot see.

Lipid droplet: modified, not accepted

This one is more interesting than it looks. The LD location rests on

PMID:26427354

i.e. an overexpressed tagged construct. But PMID:41196993 goes further than simply not
reproducing it — live-cell tracking shows the Adig-bound species is specifically the one
that does not dock at droplets:

PMID:41196993

So ADIG acts on droplet formation from the ER, in trans through seipin, rather than residing
on the droplet. MODIFY → GO:0005789.

The missing half of the annotation set

There is not one molecular function term among the 16 annotations. For a protein whose
entire biology is a structural protein–protein interaction, that is the largest gap. Added as
NEW (all ISS from PMID:41196993, since the structures are mouse/human seipin–Adig):

GO:0044877 protein-containing complex binding — and note it is oligomer-state selective,
which generic protein binding could never express.
GO:0031334 positive regulation of protein-containing complex assembly — bridging adjacent
subunits is doing the work of assembly, so the participation test is met.
GO:0050821 protein stabilization — the definition's anti-aggregation clause is met literally.

GO has no seipin complex term (OLS returns nothing for "seipin complex"), so in_complex
had to be left empty; filed under proposed_new_terms with parent GO:0140534.

Tension worth flagging

Acute siRNA knockdown gave a negative result:

PMID:26427354

whereas germline knockout does:

PMID:33691105
PMID:33691105

Recorded on the GO:0019915 row rather than suppressed. Plausibly developmental vs. acute
requirement, or slow seipin oligomer turnover — captured in suggested_experiments as a degron
experiment.

Human relevance

All the functional work is rodent, but PMID:33691105 notes a GWAS association of the ADIG
locus with BMI-adjusted leptin levels, which is the one piece of direct human evidence and the
reason the ISS transfers are worth making rather than leaving the human gene bare.

Outcome

19 rows: 11 ACCEPT, 3 MODIFY, 3 NEW, 2 REMOVE. just validate human ADIG passes.

Round 2 (review feedback)

The PR reviewer found two provenance defects that no validator could catch, and both were mine:

An unsourced phenotype. I wrote "blunts cold tolerance" into description and
core_functions. PMID:41196993's cached full text contains zero occurrences of "cold" or
"thermogen" — I checked, and the reviewer was right. The claim came from the affinage summary
of the bioRxiv row, i.e. from the very preprint this review argues is superseded. Having just
criticised affinage for framing non-primate biology as human, I imported one of its uncited
claims myself.

The interesting part: the claim is real, but only in the preprint —
PMID:39211078 and
PMID:39211078
Neither survives into the published Science paper. A claim dropped between preprint and
publication should not be curated as established, so it is out of description and
core_functions and recorded instead as a provenance note on the preprint reference plus a
suggested question.

An identifier without provenance. PMID:39211078 was asserted with no cached record. Now
cached; its title confirms it is the same work.

Other changes from the review:

GO:0031334 → GO:0065003. My reason argued participation while the term said
regulation — the enum-drifts-from-prose pattern again. The source says Adig is "actively
participating in the assembly", and a regulator acts on an assembly it is not part of, whereas
ADIG is a constituent of the product. Switched to the participation term.
"defined stoichiometry" was wrong, and it was load-bearing for the proposed complex term.
The paper says the opposite: PMID:41196993
and enumerates coexisting species. A dynamically exchanging, variable-occupancy subunit is a
harder case for a complex term. The justification now says so and argues anyway.
A question I posed was already answered in a paper I had cited and cached:
PMID:33691105
and PMID:33691105.
Folded into the GO:0019915 reason instead of being asked.
The GO:0140042 reason omitted the sign asymmetry the description got right. ADIG suppresses
nucleation and makes droplets fewer and larger; a curator reads the reason, not the description.
GO:0050821 was defended with its weakest evidence (co-overexpression gel filtration — the
same class this review discounts elsewhere). Re-anchored on the MD result
PMID:41196993,
noting stabilization is mutual.
core_functions[1] said of itself that it was not a separate function. Folded into one entry.
Added the three cached-but-uncited references (PMID:37249025, PMID:27766294, PMID:42005040).
Added a sentence reconciling "binds a complex" (GO:0044877) with "is a subunit of one": the
association is dynamic, with FRAP exchange on a ~10 s timescale.

Final: 19 rows, 11 ACCEPT / 3 MODIFY / 3 NEW / 2 REMOVE. 33 distinct quotes (91 instances) verified.