ABI3BP (Q7Z7G0, TARSH / NeshBP) — review notes

Human, UniProt Q7Z7G0 (TARSH_HUMAN), 1068 aa precursor, HGNC:17265, chromosome 3q12.

1. What the GOA record actually contains

The whole GOA record is ten rows and it is entirely about where the protein is, plus one
molecular function repeated three times:

Term Aspect Evidence Reference Qualifier
GO:0005576 extracellular region CC IEA GO_REF:0000044 located_in
GO:0140149 non-collagenous component of interstitial matrix CC TAS PMID:36399478 located_in
GO:0031012 extracellular matrix CC HDA PMID:28675934 located_in
GO:0031012 extracellular matrix CC HDA PMID:27559042 located_in
GO:0031012 extracellular matrix CC HDA PMID:20551380 colocalizes_with
GO:0005576 extracellular region CC HDA PMID:27068509 located_in
GO:0005576 extracellular region CC HDA PMID:20551380 located_in
GO:0005201 extracellular matrix structural constituent MF RCA PMID:28675934 enables
GO:0005201 extracellular matrix structural constituent MF RCA PMID:27559042 enables
GO:0005201 extracellular matrix structural constituent MF RCA PMID:20551380 enables

Facts worth stating plainly:

Meanwhile the primary literature contains a worked-out receptor-level mechanism, a knockout
mouse, and human loss-of-function tumour data. None of it is in GO.

2. Protein architecture (checked directly against ABI3BP-uniprot.txt)

Sequence checks I ran on the canonical sequence in ABI3BP-uniprot.txt:

3. The mechanism the GO record is missing

Integrin-β1 is the receptor. In mesenchymal stem cells, Abi3bp binds integrin-β1 and this
is what switches cells out of proliferation and into differentiation
[PMID:23666637, "Upon Abi3bp binding to integrin-β1 Src associated with paxillin which inhibited
proliferation."]. The binding is shown two ways in that paper — co-IP of integrin-β1 with a
myc-tagged Abi3bp from conditioned medium, and co-IP of integrin-β1 with endogenous Abi3bp
[PMID:23666637, "immunoprecipitation of endogenous Abi3bp from MSC-GFP-Akt1 cells resulted in the co-precipitation of integrin-β1"] — plus a function-blocking antibody screen in which only the β1 antibody
had an effect [PMID:23666637, "incubation with the β1 blocking antibody increased phospho-ERK1/2 levels ~3.5-fold."].

The downstream logic is explicit: [PMID:23666637, "In the absence of Abi3bp, the integrin-β1 is
maintained in a non-active state and the lack of phosphorylated paxillin prevents sequestration of
Src and ERK at the plasma membrane, leaving these kinases to activate cyclin-d1 and drive
proliferation."]. So ABI3BP is a positive regulator of integrin activation and a
negative regulator of the ERK cascade and of proliferation.

Loss of function, in vivo. [PMID:23666637, "MSCs from Abi3bp knockout mice displayed severe
deficiencies in osteogenic and adipogenic differentiation."] and
[PMID:23666637, "In vivo, Abi3bp knockout increased MSC number and proliferation in bone marrow,
lung, and liver."].

The same axis in cardiac progenitors. [PMID:25296984, "In vivo, genetic ablation of the Abi3bp
gene inhibited CPC differentiation, whereas CPC number and proliferative capacity were increased.
This correlated with adverse recovery after myocardial infarction."] and
[PMID:25296984, "Abi3bp controlled CPC differentiation via integrin-β1, protein kinase C-ζ, and
v-akt murine thymoma viral oncogene homolog."]. The receptor dependence was tested directly:
[PMID:25296984, "Integrin-β1 blocking antibodies completely abrogated the positive effects of
re-expression of Abi3bp (Figure 5B)"].

Human loss-of-function evidence. Re-expression in human thyroid carcinoma lines
[PMID:18559958, "Re-expression of ABI3BP in thyroid cells resulted in a decrease in transforming
activity, cell growth, cell viability, migration, invasion, and tumor growth in nude mice."];
overexpression in human NSCLC lines [PMID:40092729, "Overexpression of ABI3BP in NSCLC cells
resulted in a substantial reduction in cell growth and motility and induced cell cycle arrest."];
and in gallbladder cancer the gene is epigenetically silenced by MALAT1/EZH2
[PMID:31174563, "the expression of MALAT1 was up-regulated while that of the ABI family member 3
binding protein (ABI3BP) was down-regulated in GBC tissues and cell lines"].

Note the species split: the mechanism (integrin-β1, paxillin, Src, PKCζ, Akt) is mouse;
the growth-suppression phenotype is reproduced in human cells by three independent groups
in three tissues. That is why the mechanistic core function here is grounded as ISS from mouse
with human corroboration, not asserted as a human experimental result.

4. The senescence literature genuinely contradicts itself

This is not a reading error on my part; the field says so out loud
[PMID:23666637, "Abi3bp has been reported to have both a positive and negative role in senescence [13, 14]."].

Both signs are supported by knockdown/knockout experiments in different cell types, so this is
a real biological conflict (cell-type-dependent, or stress-induced vs replicative senescence),
not a citation problem. No senescence GO term should be asserted for this gene until the
conflict is resolved. I have deliberately kept senescence out of core_functions and put it
in suggested_questions / suggested_experiments instead.

5. The naming problem

The gene is named for a binding partner the gene has never been shown to bind as a full-length
protein. The original clone was a yeast two-hybrid hit against the Nesh (ABI3) SH3 domain
[PMID:11501947, "By using a conventional two-hybrid technique with an Src homology 3 (SH3) domain
of Nesh as the bait protein, a novel full-length cDNA was isolated and sequenced from a human
placenta cDNA library."], and it was the 486-aa short isoform (see §2). Twelve years later the
caveat was still standing [PMID:23666637, "However, in vivo binding activity between full-length
Abi3bp and Abi3 awaits confirmation."]. UniProt encodes exactly this hedge:
CC -!- SUBUNIT: Probably interacts with ABI3. — a Probably, not an assertion.

Consequently GOA carries no protein-binding annotation for ABI3BP–ABI3, and it is correct
that it does not. I have not proposed one.

6. Where GO:0005201 comes from, and why it is an over-annotation

All three GO:0005201 extracellular matrix structural constituent rows are RCA from BHF-UCL,
each citing an ECM-proteomics dataset or protocol:

What these datasets show is that ABI3BP peptides are recovered from ECM-enriched fractions of
human tissue. That supports the cellular component (and the same three papers are indeed
cited for the HDA GO:0031012/GO:0005576 rows, which I accept). It does not support the
molecular function claim, whose definition is specific:
"The action of a molecule that contributes to the structural integrity of the extracellular
matrix" (verified against QuickGO). Nobody has measured whether removing ABI3BP changes the
mechanical integrity of any matrix. The one assembly-relevant experiment is second-hand and is a
fragment, not the protein: [PMID:23666637, "Computational screening followed by in vitro assays
identified that a partial fragment of Abi3bp, containing one of the two Fibronectin type-III
domains found in the full length protein, promoted cell attachment and was capable of assembling
into an extracellular matrix"] — citing PMID:18757743, whose full text is not retrievable and
whose abstract does not name ABI3BP.

Contrast that with what has been measured: receptor binding, receptor activation, and a
signalling output. ABI3BP behaves as a matricellular protein — matrix-resident, cell-instructive,
not load-bearing.

The structural reading is also what the ontology forces. GO:0005201's children are all
mechanical — tensile strength (GO:0030020), compression resistance (GO:0030021), elasticity
(GO:0030023), lubricant (GO:0030197) — and GO has no term for the matricellular class at all
(confirmed: OLS search_all_ontologies(query="matricellular", ontologies="go") returns an empty
result set, and the QuickGO ontology search for "matricellular" likewise returns nothing). So a
pipeline that knows only "this is a core-matrisome ECM glycoprotein" has exactly one ECM-specific
molecular function available to it, and it is the wrong one. That is the gap recorded in
proposed_new_terms.

7. Other things worth flagging

8. Decisions taken

On the ten existing rows:

Eight NEW rows proposed, all coded ISS with supporting_entities: MGI:MGI:2444583 (mouse
Abi3bp, MGI id taken from the UniProt cross-reference on F7B3T6, not from memory), because every
perturbation experiment behind them is mouse:

Term Ref Why
GO:0005178 integrin binding (MF) PMID:23666637 Co-IP with tagged and endogenous protein; β1-specific blocking-antibody panel
GO:0033625 positive regulation of integrin activation PMID:23666637 Ligand changes receptor state, measured as phospho-paxillin/paxillin ratio (see outside-in caveat below)
GO:1902461 negative regulation of mesenchymal stem cell proliferation PMID:23666637 Germline KO with in vivo phenotype in three organs + independent shRNA lines
GO:0070373 negative regulation of ERK1 and ERK2 cascade PMID:23666637 Perturbed at three points in the chain; vinculin knockdown is the specificity control
GO:0045669 positive regulation of osteoblast differentiation PMID:23666637 Osteogenesis fails in KO cells across all three isolations
GO:0045600 positive regulation of fat cell differentiation PMID:23666637 Adipogenesis fails in the same KO cells
GO:2000727 positive regulation of cardiac muscle cell differentiation PMID:25296984 In vivo ablation + marker analysis + antibody demonstration that it runs through integrin β1
GO:0050774 negative regulation of dendrite morphogenesis PMID:19302145 Conditioned medium, purified protein at 500 ng/ml, and a cell-autonomous overexpression arm; expression time course matches the process

The GO:0050774 row was added during PR review (#2250). My first draft described the olfactory
dendritic-refinement role in the top-level description but proposed no annotation, on the grounds
that the work was single-laboratory and mouse-only and was "described in the gene description and
questions". The reviewer correctly pointed out that both halves of that fail: the same evidence
grade underpins the five rows drawn from PMID:23666637, PMID:25296984 is from the same laboratory
as PMID:23666637, and there was in fact no question about the nervous system anywhere in the file.
Withholding ISS from one body of mouse work while applying it to another of identical grade was
just inconsistent. Added, with two new questions (neuronal receptor; diffusible vs tethered).

It is kept out of core_functions for a stated reason rather than by omission: no receptor is known
for the neuronal effect, so there is no molecular function to attach the process to without assuming
that integrin β1 is involved, and it has not been reproduced outside the originating lab.

The outside-in caveat on GO:0033625

GO:0033622 integrin activation is defined as "The aggregation, arrangement and bonding together
of an integrin… that lead to the increased affinity of the integrin for its extracellular ligands",
and that branch is conventionally used for inside-out affinity modulation by talin and kindlin.
ABI3BP works from the opposite side — it is the extracellular ligand. The definition does not
restrict the term to inside-out signalling, and the paper's own wording ("the integrin-β1 is
maintained in a non-active state" without Abi3bp) supports the claim, but an outside-in ligand is
not the canonical filler and a curator may prefer to carry the point on GO:0005178 alone. Recorded
in review.reason on that row.

Two term-choice traps I fell into and had to back out of

GO:2000738 positive regulation of stem cell differentiation is the wrong term, despite
looking exactly right. My first draft used it as the lineage-agnostic way to say "knockout MSCs
fail to differentiate". But its target, GO:0048863 stem cell differentiation, is defined (QuickGO)
as "The process in which a relatively unspecialized cell acquires specialized features of a stem
cell" — differentiation into a stem cell. Annotating it would assert that ABI3BP promotes cells
becoming stem cells, which is the opposite of the observed biology. GO:2000741 positive regulation of mesenchymal stem cell differentiation inherits the identical inversion via
GO:0072497 ("…acquires specialized features of a mesenchymal stem cell"), so the
more-specific-looking term is wrong in the same way.

The whole GO <cell type> differentiation branch means "becoming that cell type", so the only
correct route is outcome-specific terms — hence GO:0045669 (→ GO:0001649 osteoblast
differentiation, "acquires the specialized features of an osteoblast" ✓) and GO:0045600
(→ GO:0045444 fat cell differentiation, "…of an adipocyte" ✓). GO:2000727 survives the same
test (GO:0055007: "a cardiac muscle precursor cell acquires specialized features of a cardiac
muscle cell" ✓), and so does GO:1902461 (GO:0097168 MSC proliferation = "multiplication or
reproduction of mesenchymal stem cells" ✓).

Lesson: fetch the definition of the target of every regulation term, not just the regulation
term itself. The label reads correctly in both directions; only the definition disambiguates.

GO:0098640 integrin binding involved in cell-matrix adhesion was also considered for the MF
and rejected — the demonstrated consequence of binding is receptor activation and downstream
signalling, not adhesion of the cell to the matrix. Plain GO:0005178 is the honest choice.

Also:

The proposed definition had the same flaw and was fixed in a second pass. It originally read
"…without contributing appreciably to the mechanical integrity of the matrix" — an unmeasurable
negative, and one that this review's own position ("never tested") means ABI3BP cannot be shown to
satisfy. Proposing a term whose defining clause the motivating example cannot be demonstrated to
meet would have been self-defeating in front of a GO editor. Reworded positively: the defining
feature is the cell-instructive action on a receptor, and the term now explicitly neither asserts
nor excludes a structural contribution.

That reword then went too far, and was tightened in a third pass. Dropping the negative clause
left the definition with no differentia at all: "acts on cells by engaging cell-surface receptors,
thereby modulating their signalling, adhesion, proliferation or differentiation state" is satisfied
in full by GO:0008083 growth factor activity, since plenty of growth factors are matrix-resident
(FGFs, HB-EGF, latent TGF-β complexes), and localization is a weak differentia for a molecular
function in any case. The criterion the siblings genuinely fail is receptor class: GO:0005125
cytokine activity is defined for a soluble extracellular product, growth factor activity for
stimulation of growth, whereas a matricellular protein signals through an adhesion receptor
(GO:0004895, whose own definition names integrins) and here restrains proliferation. The
definition now carries that, and the justification records all three rejected candidates — the
unmeasurable negative, matrix residence alone, and receptor class as the one that survives.

Fourth pass: the receptor-class version over-generalised in the other direction. Writing "its
output is characteristically restraint of proliferation" into the differentia was true of ABI3BP but
false of the class — tenascin-C, periostin and CCN1/CCN2 signal through αVβ3, αVβ5 and α6β1 and are
characteristically pro-proliferative, so the clause would have excluded most of the proteins the
same justification cites as the reason the term is needed. A differentia hedged with
"characteristically" is not usable as a differentia in any case, and receptor class alone does the
whole exclusion job. Removed, along with the sibling comparison itself, which is reasoning about
the ontology and belongs in justification (ProposedOntologyTerm has no comment slot).

Two things were added instead. GO:0004895 is not just a plausible parent class: production GO-CAM
689e7a5d00003515 assigns exactly GO:0004895 cell adhesion receptor activity to human ITGB1
(UniProtKB:P05556), alongside GO:0007229 and GO:0005925 — verified at gocams/index.tsv:8672.
So the differentia matches how GO already types the very receptor ABI3BP engages. And a
reasoner-checkable form is now offered for the editors: GO:0048018 and has-target some GO:0004895,
with matrix residence as a non-defining characteristic.

Four passes on one definition is worth recording as a pattern. Unmeasurable negative → no
differentia at all → a differentia that over-claimed at class level → receptor class alone. Each
individual change was correct, and two of the intermediate states were genuine regressions. The
lesson: when tightening a class definition, test it against the other members of the class you
cited as motivation, not only against the gene in front of you.
- Senescence, antiviral and blood-brain-barrier roles left unannotated and moved into
suggested_questions / suggested_experiments. The senescence literature contradicts itself
(§4) and no term should be asserted until that is resolved. Note the contrast with the neuronal
role, which is now annotated: the difference is not the number of laboratories but the kind of
evidence. Dendritic refinement has three concordant gain-of-function arms plus a matching
expression time course and an activity-independence control; the senescence claims contradict each
other in sign, the antiviral report is self-described as preliminary in its own title, and the BBB
arm depends on supraphysiological recombinant protein.
- The affinage record is cited once, on the GO:1902461 row, and only for its high-level summary
sentence, which is independently carried on the same row by two verbatim quotes from
PMID:23666637 and one from PMID:18559958. No mechanistic claim in this review rests on it —
every domain boundary, binding partner and mechanistic step is quoted from UniProt's own feature
table or from the primary paper.

Verification run before pushing: checkquotes.py — 100 quotes, 0 problems (it checks file:
quotes as well as PMIDs, so the UniProt lines are verified too); just validate human ABI3BP —
✓ Valid, zero warnings; cache/go/terms.csv — 4 insertions, 0 deletions.

Two cache/go/terms.csv traps hit along the way, both worth remembering. First, a just validate
run silently drops unrelated cached terms (18 the first time, 1 each on two later runs), so the file
has to be restored from origin/main and the new terms re-added after every validation. Second,
do not globally sort the file to re-insert them — origin/main currently has at least one
out-of-order entry (GO:0140312), so a global sort relocates it and shows up as a spurious deletion.
Insert each new line in position without touching the rest.

A validator quirk also worth noting: for PMID:19302145 the locally cached record has
full_text_available: true and does contain the full text, but the reference validator's own record
holds only the abstract, so full-text-only passages are rejected as supporting_text even though
checkquotes.py accepts them. All quotations from that reference are therefore taken from the
abstract.