ARGLU1 bioinformatics results

Four analyses supporting the human ARGLU1 GO annotation review. Each is a
standalone script in this directory writing a JSON artefact; this file is a
hand-written summary of those artefacts, not a generated report.

Run any of them with uv run python <script>.py from this directory. All but
composition_and_features.py and sibling_row_verdicts.py require network
access (QuickGO, UniProt, IntAct); composition_and_features.py reads the
committed ../ARGLU1-uniprot.txt and sibling_row_verdicts.py reads the
repository's own review YAMLs (it derives the repo root rather than hardcoding
a worktree path).

1. The mitochondrion IBA rests on one Candida albicans Mediator-tail paralog

panther_node_audit.py compares the two cellular-component terms human ARGLU1
receives by IBA from the same PANTHER node, PTN001271790:

The node's reach is identical for the two terms, so the difference is entirely in
the donor pool. Every one of the 111 mitochondrion rows names CGD:CAL0000179812
in its WITH/FROM field — i.e. the whole family's mitochondrial annotation, across
64 taxa including plants, green algae, Selaginella, Marchantia, Trichoplax,
Daphnia and all vertebrates, traces to that single protein.

What that protein is matters. InterPro assigns A0A1D8PRM3 three signatures:

interpro IPR021017 | Mediator complex, subunit Med2, fungi | family
interpro IPR027267 | AH/BAR domain superfamily         | homologous_superfamily
interpro IPR033371 | Arginine and glutamate-rich protein 1 | family
pfam     PF11214   | Mediator complex subunit 2
panther  PTHR31711 | ARGININE AND GLUTAMATE-RICH PROTEIN 1

Human ARGLU1 (Q9NWB6) carries only IPR033371 / PF15346 / PTHR31711. TLO16
is therefore a fungal Med2-family protein that also matches the ARGLU1 family
signature — an architecturally distinct occupant of the node, not an ARGLU1
ortholog.

The source paper is explicit that the mitochondrial localisation is not even a
property of the TLO family as a whole: the TLO family is a C. albicans-specific
expansion "from one or two copies in other CUG clade members to 14 expressed
copies", and only the γ clade reaches mitochondria while clades α and β "encode
proteins that localize primarily to the nucleus".

Conclusion: a clade-restricted localisation of a lineage-specific fungal
gene-family expansion has been placed at a pan-eukaryotic node and distributed to
111 gene products. This is a node-placement defect, not evidence about ARGLU1.

Divergence from the corpus, stated deliberately. sibling_row_verdicts.py
finds 213 merged reviews in this repository carrying a GO:0005739 IBA row, and
their modal verdict is ACCEPT (127), with KEEP_AS_NON_CORE 68,
MARK_AS_OVER_ANNOTATED 12, MODIFY 2 and REMOVE 2. This review's REMOVE is
therefore against the corpus trend. That is intended: for most of those 213 genes
the donor pool is genuinely mitochondrial, whereas here the entire donor pool is
one Candida TLO protein whose own source paper restricts the localisation to one
clade of one lineage-specific family. The corpus trend is not an argument about
this node.

2. The cadherin-binding row is a proximity-labelling screen read as binding

reference_scope_audit.py counts, per (reference, term) pair, how many distinct
gene products the reference annotates. Annotation counts and entity counts are
reported separately because they are different numbers; where the result set is
too large to enumerate, the entity count is reported as unavailable rather than
extrapolated from one page.

The control discriminates: the paper that actually studied ARGLU1 annotates four
gene products; the screens annotate 121–272+ each.

For the cadherin row specifically, the assay measures proximity, not binding —
the paper reports "561 proteins in the vicinity of the cytoplasmic tail of
E-cadherin" — and ARGLU1's every curated location is nuclear (nucleus,
nucleoplasm, nuclear speck, chromosome).

Cross-check against merged reviews: 38 gene reviews already in this
repository carry this same GO:0045296 / PMID:25468996 row. Their verdicts are
MARK_AS_OVER_ANNOTATED ×21, KEEP_AS_NON_CORE ×11, REMOVE ×3, UNDECIDED ×1,
ACCEPT ×1, PENDING ×1. This review follows the modal verdict.

3. Partner audit: three checks, and what each returned

intact_partner_audit.py. ARGLU1 has 250 IntAct binary rows across 222 distinct
partner accessions — it is a well-connected disordered protein, so partner count
alone discriminates nothing.

Note on counting: IntAct lists several identifier forms for one study (PubMed
id, DOI, MINT ac, IntAct ac), so the raw identifier count overstates the number of
studies — 10 forms for SRPK2's 3 studies. The table counts distinct PubMed ids.

Results of the three standing checks, including the nulls:

1. Canonical-partner check — negative (no defect). All four accessions resolve
   to reviewed Swiss-Prot entries at their canonical lengths. No unreviewed
   partial-ORFeome substitution of the kind found on ACRV1.
2. Topological-plausibility check — negative (no defect). All four partners
   are nuclear; SRPK2 is itself annotated to nuclear speckles, the compartment
   ARGLU1 occupies. The partner set coheres with ARGLU1's own biology rather than
   with a screen's bait panel.
3. NbExp inflation check — mixed. UniProt lists NbExp=3 for SRPK2 and
   NbExp=4 for the other three. Expanding the records shows SRPK2's three rows
   are three orthogonal methods from three independent laboratories
   (Varjosalo 2013 in vitro kinase assay; Hegele 2012 yeast two-hybrid; Huttlin
   2021 AP-MS) — genuine replication, not one screen logged three ways. JMJD6 is
   the opposite: all five rows are anti tag coip, and three of them are
   spoke-expanded BioPlex rows from two releases of the same pipeline, so its
   effective independent support is the primary paper's directed co-IP plus one
   AP-MS pipeline.

The single most informative row in the table is SRPK2's, which IntAct types
not as physical association but as phosphorylation, detected by
protein kinase assay, host In vitro. ARGLU1 is recorded there as an in vitro
substrate of SRPK2, not merely a binding partner.

3b. A term the review first said did not exist — and who it applies to

splicing_factor_eligibility.py. An earlier draft of this review removed all six
U2AF2 / PUF60 / JMJD6 GO:0005515 rows on the stated ground that GO had no
splicing-factor-binding molecular function term. That was false.
GO:1990935 splicing factor binding is active, is a molecular function, and is
defined as "Binding to a protein involved in the process of removing sections of
the primary RNA transcript to form the mature form of the RNA."

The error is instructive rather than careless: the draft searched for
"spliceosomal complex binding". GO's text search is token-based, and
spliceosomal complex shares no token with splicing factor, so that query could
not have returned the term however it was phrased. It was found by walking the
ontology (children of GO:0044877, plus compound-word searches) rather than by
searching harder.

Whether the recovered term applies is then a per-partner question, settled by
querying GOA for each partner's own annotations under GO:0008380 RNA splicing:

Consequences for the review:

• U2AF2 (×3 rows) and PUF60 (×2 rows) → MODIFY to GO:1990935.
• JMJD6 (×2 rows) → stays REMOVE, on a narrower ground than first stated. The
  first version of this section said JMJD6 had zero annotations under
  GO:0008380. That query used is_a/part_of descendants only, which
  structurally cannot reach regulation terms — GO relates regulation to its
  target by regulates, not is_a. Widening it (splicing_regulation_scope
  block in splicing_factor_eligibility.json) shows JMJD6 does hold
  GO:0048024 regulation of mRNA splicing, via spliceosome by IMP. The
  surviving distinction is finer: JMJD6 is annotated to the regulation of
  splicing, not to splicing, and it is an oxygenase acting on a splicing factor
  rather than a component of the machinery. Whether that satisfies "involved in
  the process of removing sections of the primary RNA transcript" is not settled
  by the term's definition, so the term is withheld rather than asserted.
• SRPK2 (×3 rows) → stays MODIFY to GO:0019901 protein kinase binding. SRPK2
  qualifies under both terms, but the in vitro kinase assay typed by IntAct
  as phosphorylation is the better-evidenced relationship, and the kinase term
  carries the mechanistic content.

The general rule this cost a round to relearn: a failed keyword search is not
evidence that a term is absent. Search the compound word as well as the
fragment, and enumerate children of the plausible parent rather than trusting a
text query.

4. The composition is a composition, and there is no domain

composition_and_features.py, parsing the committed UniProt flat file (parsed
length asserted equal to the declared 273 aa before anything is computed).

Three things follow.

• The two regions UniProt delimits by experiment have opposite charge
  character, and the boundary is sharp: E+D goes 9.5% → 27.1% and S goes 27.0%
  → 4.0% across it.
• The N-terminal RNA-binding region is not merely "arginine-rich": it is a
  genuine RS-repeat (SR-protein-like) region, with 12 RS and 12 SR dipeptides
  and an unbroken 8-dipeptide alternating run. This is the structural fact behind
  SRPK2 — an SR-protein kinase — phosphorylating ARGLU1 in vitro, and behind
  UniProt's phosphoserines at Ser-58, Ser-60 and Thr-61, all of which fall inside
  this region (the other three annotated sites, Ser-76, Ser-77, Ser-266, fall in
  the C-terminal region).
• The complete signature content of the entry is IPR033371 / PF15346 /
  PTHR31711(:SF1) — one family and its own PANTHER family, nothing else.
  PF15346 is typed by InterPro as coiled_coil, not domain. There is no
  catalytic, nucleotide-binding, or recognised nucleic-acid-binding fold anywhere
  in the protein.

That last point is the negative result that matters for curation: no molecular
function can be reverse-engineered from ARGLU1's domain content, because it has
none. "Arginine and glutamate rich" names a composition. Every molecular
function claim on this gene has to come from an experiment, and the review treats
it that way.