ARHGEF19 bioinformatics

Two questions, two scripts, both reproducible from this directory:

uv run python dh_competence_check.py        # -> RESULTS_raw.txt   (exit 0 = all assertions held)
uv run python selftest_guards.py            # mutation tests for the above
uv run python go_coverage_by_reference.py   # -> COVERAGE_raw.txt

Every sequence, profile, structure and annotation count is fetched live at run
time (UniProt REST, InterPro/Pfam, RCSB, QuickGO, PubMed E-utilities). Nothing
below is transcribed by hand from a database page; the raw output files are the
scripts' own stdout.

1. Is the ARHGEF19 DH–PH module a competent Dbl-family GEF?

dh_competence_check.py. The honest answer has to run in both directions, so
the panel carries both: five Dbl-family GEFs that have both a co-crystal
structure with their substrate GTPase and measured exchange activity, and
three negative controls.

Published residue claims, checked against the current UniProt sequence

Q8IW93 is 802 aa; the script aborts with exit 2 if that ever changes, because
every position above would then need re-checking.

The profile fires on the module — and on a module that cannot work

Pfam PF00621 (RhoGEF / Dbl homology) at the gathering threshold:

The truncation control is the useful one. Isoform 2 deletes residues 484–783 —
the C-terminal half of the DH domain plus all of PH and SH3 — and still clears
the Pfam gathering threshold on its surviving N-terminal half. A DH-profile hit
is therefore not evidence that a protein can catalyse exchange.

The exchange surface, taken from structures rather than asserted

GEF-side residues within 4.5 Å of the GTPase were computed from the deposited
coordinates of 1X86 (LARG DH–PH : RhoA) and 1LB1 (Dbs DH–PH : RhoA), the
two complexes whose GTPase is ARHGEF19's own demonstrated substrate. Author
numbering in both structures was verified residue-by-residue against the
corresponding UniProt entry before anything was mapped (40 and 44 positions
respectively). Contacts inside the DH envelope were projected onto PF00621 match
states; 25 states are contacts in both complexes.

The pass/fail threshold is taken from the data, not chosen: the weakest
verified-active comparator spans 24/25, and the script asserts (a) that the
target reaches that floor and (b) that the truncation control falls below it.
Both hold.

What this does and does not establish

It does establish that ARHGEF19 presents a complete, canonical Dbl-family
exchange surface: 24 of the 25 RhoA-contacting positions are present, the same
coverage as ITSN1 and one position short of LARG, Dbs, Tiam1 and TRIO. There is
no residue-level reason to suspect a pseudo-GEF.

It does not establish that ARHGEF19 exchanges nucleotide, and it emphatically
does not establish which GTPase. Identity at the exchange surface is 36% for
ARHGEF19 against LARG — and 36% for Tiam1, a GEF for a different GTPase
(Rac1) entirely. A measure that scores a Rac1 GEF exactly like ARHGEF19 cannot be
read as evidence that ARHGEF19 is a RhoA GEF. That question is settled by
biochemistry, not by this table: hWGEF activates RhoA and not Rac1 or Cdc42 in
side-by-side pulldowns (PMID:18256687, Fig. 2A–B), and binds RhoA strongly, Rac1
weakly and Cdc42 not at all (Fig. 2C).

The two directions therefore agree here — intact surface, demonstrated activity —
but they agree by coincidence of evidence, not because either implies the other.

Guards

selftest_guards.py breaks one thing at a time and asserts the matching guard
fires with its expected message. 17 cases — 12 against dh_competence_check.py
and 5 against verify_quotes.py — including two negative controls that must stay
silent and a duplicate-accession case (FASTA names are the join key for
hmmsearch/hmmalign, so a duplicated panel entry would silently merge two rows).
All 17 behave as specified.

The count in that sentence is the one the script prints, and the script derives it
rather than storing it (len(CASES) + qc_ran): an earlier version hardcoded
+ 5, and this document and the review YAML both went stale at 12 the moment the
quote-checker cases were added. Both groups of mutations are anchored by name —
module attribute for the first, reference id for the second, asserted to match
exactly once — after an index-anchored mutation silently moved onto the wrong row
when a reference was inserted ahead of it, and "failed" for the wrong reason.

2. How much of the ARHGEF19 literature has reached GO?

go_coverage_by_reference.py queries QuickGO by reference for every primary
paper on the gene and counts annotations in any species. The paper list comes
from two independent sources — the Affinage record's citations and a live PubMed
sweep over the gene's synonyms — so it does not depend on either being complete.
"GEF19" is one of the queries on purpose: PMID:21686262 is titled "Grhl3 and
GEF19 in the front rho" and no ARHGEF19 or WGEF query returns it.

Of 14 primary experimental papers:

• 11 have produced no GO annotation of any gene in any species. That includes
  the paper that first demonstrated exchange activity (PMID:15485661), the paper
  that mapped autoinhibition relief (PMID:18537266), the 2024 structural and
  activity study (PMID:38714795), all three cancer-pathway papers
  (PMID:29164615, PMID:32993957, PMID:34813497), the ciliogenesis paper
  (PMID:31469868), and the only published negative result (PMID:20810787).
• 12 have produced no GO annotation of ARHGEF19 in any species.
• 1 has produced an ARHGEF19 annotation to anything other than
  GO:0005515 protein binding — PMID:20643356, whose subject is GRHL3 and whose
  ARHGEF19 annotations are the mouse IMP and the human IGI. Even that paper is
  under-curated on the human side: its RhoGEF19 knockdown was done in human
  HaCaT keratinocytes, with a scratch-closure readout and a rescue, yet the human
  gene carries GO:0042060 only as an IBA projected from the mouse orthologue in
  the same paper.

The single richest paper on this gene, PMID:18256687 (EMBO J), yielded exactly
one ARHGEF19 annotation: GO:0005515 protein binding with RhoA. The same
figure panel that supports that IPI (Fig. 2C) sits beside a RhoA activation assay
(Fig. 2A), a Rac1/Cdc42 negative (Fig. 2B), a Dvl2-domain mapping (Fig. 5C), a
Daam1 co-IP (Fig. 5D), a human MCF-7 RNAi experiment (Fig. 5B) and an
autoinhibition mapping (Fig. 6) — none of which reached GO.

This is a coverage problem, not an over-annotation problem. The review's actions
follow from that: almost nothing to remove, a great deal missing.

What the deep-research record left out

The same script parses the committed Affinage record for the PMIDs it cites and
diffs them against the classified list, so "Affinage missed X" is a proposition
about a file in this repo rather than a recollection. Affinage cites 11 PMIDs,
all of which resolve to real papers about this gene — its precision is fine, and
its trust gates passed. Those 11 omit 3 of the 14 primary papers, plus the
commentary PMID:21686262.

The tempting generalisation is that every miss has an unfindable title. The script
computes that rather than asserting it — it reads each missed paper's cached
title and matches it against the gene's synonyms — because an earlier revision of
this document asserted it of all four and the table directly below contained a
counterexample:

So three of the four misses have a retrieval story and one does not. That last row
is the sharper finding, not a blemish on the generalisation: a record whose trust
gates passed omitted a paper whose title begins with the gene symbol, alongside
the single most annotation-productive paper on the gene. Gates certify the
citations given, not the ones withheld.