AARD (alanine- and arginine-rich domain-containing protein, C8orf85) — review notes

PAINT no-IBA project review, using the affinage deep-research provider
(AARD-deep-research-affinage.md, gates passed) plus UniProt Q4LEZ3, the GOA TSV and the
primary literature.

The darkest gene in this campaign so far

AARD's entire GO record is 15 rows, every one of them GO:0005515 protein binding, all IPI,
all from a single publication (PMID:32296183, the HuRI binary interactome). The 15 rows differ
only in their WITH/FROM partner, so the seeder correctly collapses them to one
existing_annotations entry with 15 interactors behind it. There is no molecular function
beyond binding, no cellular component, and no biological process. UniProt has no FUNCTION
line, no SUBCELLULAR LOCATION, and only two keywords ("Proteomics identification", "Reference
proteome").

Affinage agrees there is nothing to ground: its mechanism_profile reports
molecular_activity: (none), localization: (none), partners: (none), and the narrative ends
by saying so explicitly — "no molecular activity, interaction partners, or cellular function for
AARD have been characterized in the available corpus".

So the whole review turns on one question: should those 15 interactions be believed?

Testing the interactions rather than asserting

Rather than dismiss them by the usual argument (single publication, no follow-up), I tested a
specific, falsifiable alternative explanation — coiled-coil bias. Sticky or self-activating
preys in yeast two-hybrid are enriched for coiled-coil proteins, which associate promiscuously
through heptad-repeat surfaces. AARD-bioinformatics/analyze_partners.py reads the accessions
straight out of AARD-goa.tsv and fetches UniProt features at run time.

Result: 7 of 15 partners (47%) carry a coiled-coil FEATURE. Against a background of
2059 / 20431 reviewed human proteins with that feature (10.1%), fetched from UniProt for the
purpose, that is a 4.6-fold enrichment, binomial P(X≥7 | n=15, p=0.101) = 3.3×10⁻⁴.
Numerator and denominator deliberately use the same criterion: two more partners (KRT24, KRT27)
carry only the Coiled coil KEYWORD, which is assigned more liberally than the feature, so
scoring them against a feature-only background would inflate the result. On the consistent
feature-or-keyword pair it is 9/15 against a 10.6% keyword background — 5.6-fold,
P = 4.8×10⁻⁶ — the same conclusion either way. With n=15 this is a
descriptive statistic rather than a controlled test — the ideal comparison would be against
other single-publication prey sets from the same screen — but the enrichment is now measured
rather than asserted.

The partners also span 14 top-level subcellular compartments:

Three syntaxins — STX1A, STX2 and STX5 are SNAREs from three different membranes that share
a coiled-coil SNARE motif. Two keratins. Plus centrosomal and centromeric coiled-coil
proteins, a nuclear splicing factor, and a secreted protein. There is no compartment where these
could plausibly meet, and the shared feature across them is architecture, not biology.

That is the artefact signature, and it is a much stronger argument than "no follow-up". Action:
MARK_AS_OVER_ANNOTATED rather than REMOVE — the physical interactions may well have occurred
in the assay, and I have no positive evidence that any individual pair is false.

A general trap worth recording. When a null model is built by querying the same database
that supplies the numerator, the background query must use the identical predicate. Here the
numerator counted coiled-coil on feature-OR-keyword while the background query was feature-only
(ft_coiled:*) — and since UniProt applies the keyword more liberally than the positional
feature, the mismatch inflated the enrichment (6.0× reported where like-for-like is 4.6×). The
fix is not to pick one criterion but to compute each numerator against its own matching
denominator, which also shows the conclusion is robust to the choice.

The family route is closed too — but it exists

I originally wrote that AARD has "no recognisable domain beyond the compositional
alanine/arginine-rich description". That was wrong, and contradicted by the gene's own
UniProt file, which I had already read:

DR   InterPro; IPR051771; FAM167_domain.
DR   PANTHER; PTHR32289:SF2; ALANINE AND ARGININE-RICH DOMAIN-CONTAINING PROTEIN; 1.
DR   PANTHER; PTHR32289; PROTEIN FAM167A; 1.

AARD is a FAM167-family protein, sharing PANTHER PTHR32289 with FAM167A and FAM167B. For
a gene this dark, paralogs are the most tractable remaining inference route, so it mattered.

Enumerating the family (now part of the analysis script): 10 reviewed members across 5
species, and not one carries a UniProt FUNCTION statement.

So guilt-by-association is unavailable — not because AARD lacks a family, but because the
entire family is uncharacterised. That is a far stronger statement than "no family", and it is
independently corroborated by UniProt's own PAN-GO line:
PAN-GO; Q4LEZ3; 0 GO annotations based on evolutionary models.

It also raises a better question than any about AARD alone: this is a conserved vertebrate
family, present in fish through mammals, in which no member has a described function. Solving
any one of them would likely illuminate the rest, and no member is currently a better starting
point than another.

What is actually known about AARD

Only expression biology, and only in mouse:

• Expression is Sertoli-cell specific, up-regulated during testis differentiation
  [PMID:17486547, "The period of elevated mRNA expression coincides with early
  differentiation of the testis and is limited to Sertoli cells of the developing
  testis cords."]
• It is a direct androgen receptor target gene
  [PMID:27959439, "The present study identified Aard as a gene
  that is directly regulated by AR in mouse SCs, which is important in
  spermatogenesis."]

Neither supports a GO annotation for AARD. This is precisely the A1BG situation from earlier
in this campaign (PR #2217): being a transcriptional target of a pathway is a downstream
relationship, not participation in it. Annotating AARD to androgen-receptor signalling would
repeat exactly the ROLE_CONFLATION error I found in the mouse A1bg GH annotation. Recorded
in the description and suggested_questions; not annotated.

Outcome

No NEW terms proposed. AARD stays dark, and that is the correct result. The value added by
this review is:

1. Evidence-based grounds for discounting the only annotations the gene has.
2. An explicit record that the expression/AR biology must not be converted into process terms.
3. Targeted experiments aimed at the actual gap.

Note on annotation coverage

The GOA file has 15 rows but existing_annotations has one entry. That is not missing coverage:
all 15 rows carry the identical term, qualifier, evidence code and reference, differing only in
WITH/FROM, so they are one annotation supported by 15 interactors. All 15 are named and analysed
in the review summary and in the bioinformatics results.