ADAMTSL3 (P82987, human) — review notes

Working journal for the PAINT + affinage review of human ADAMTSL3 / punctin-2.

1. What the protein is

UniProt P82987, ATL3_HUMAN, 1691 aa, PE 1: Evidence at protein level, HGNC:14633,
chromosome 15q25.2. Cleaved signal peptide (FT SIGNAL 1..26), chain 27–1691,
heavily N-glycosylated (17 annotated CARBOHYD sites), secreted into the extracellular
matrix.

Domain content, read off the UniProt feature table rather than the family name: ten
PROSITE TSP type-1 domains (SMART counts 12, Pfam TSP1_ADAMTS 11 — the counts differ
by method, and the original paper says 13), three Ig-like C2-type domains at
896–992 / 1185–1279 / 1296–1378, an ADAMTS-type cysteine-rich region
(InterPro:IPR045371 ADAMTS_CR_3, Pfam:PF19236), and a C-terminal PLAC domain at
1655–1691. Two splice isoforms (P82987-1, P82987-2) differing only in the last 35
residues, which in isoform 2 replaces the PLAC-domain C terminus.

There is no metalloprotease domain and no disintegrin-like domain. This is stated
three independent ways:

So the "ADAMTSL proteins lack the catalytic domain" premise is true for ADAMTSL3 —
established here from the record, not assumed. The consequence for this review is the
opposite of the usual campaign pattern, see §3.

Product size in cells: PMID:14667842

Expression is broad: PMID:17597111

2. The GOA record is tiny and lopsided

18 rows in ADAMTSL3-goa.tsv:

n term evidence source
1 GO:0031012 extracellular matrix IBA GO_REF:0000033, node PTN000347317
1 GO:0030198 extracellular matrix organization IEA InterPro:IPR013273
15 GO:0005515 protein binding IPI 2 interactome papers, 13 distinct partners
1 GO:0071953 elastic fiber TAS PMID:23962539 (a review)

The fetch-gene stub collapsed the 15 GO:0005515 rows into 2, so the review had to be
expanded back to one row per partner (the ACTR5 lesson). Coverage is asserted
programmatically by ADAMTSL3-bioinformatics/analyze_adamtsl3.py --only coverage, keyed on
(term, evidence, reference, with/from).

Notable absence: there is no IDA row for extracellular matrix localisation even
though UniProt records SUBCELLULAR LOCATION: Secreted, extracellular space, extracellular matrix
with ECO:0000269|PubMed:14667842. The direct experimental localisation exists and has
simply never been transferred to GOA.

3. The catalytic-loss question, and where the real defect is

The campaign hypothesis was that a peptidase-flavoured annotation might have reached
ADAMTSL3 by domain-name propagation. It has not. No peptidase, metallopeptidase,
proteolysis or hydrolase term appears anywhere in ADAMTSL3's GOA. That hypothesis is
refuted for this gene.

But resolving why it has not turned up a genuine, fixable PAINT defect in the other
direction. From PANTHER's primary files (IBD.gaf for node annotations,
gene_association.paint_uniprot.gaf.gz for the leaf projections — both parsed by the
analysis script, not read off QuickGO):

Streaming the leaf GAF shows PTN002673039 projects onto exactly 22 leaves, every one of
them an ADAMTSL2 orthologue (human ADAMTSL2 Q86TH1, mouse Adamtsl2 Q7TSK7, and 20
vertebrate orthologues). So PAINT has already made the key-residue judgement that the
catalytic domain is lost, and has attached it to the ADAMTSL2 orthology group alone.

The same loss is shared, identically, by ADAMTSL1, ADAMTSL3, ADAMTSL4, ADAMTSL5, PAPLN,
THSD4 and the invertebrate members (madd-4, loh, papilin). None of them carries the
NOT. The projection matrix (recomputed by the script) is:

member GO:0031012 GO:0030198 GO:0004222 GO:0006508
ADAMTS1 / 9 / 10 / 17 (catalytic control) ✓ ✓ ✓ ✓
ADAMTSL2 (human + mouse) ✓ ✓ NOT –
ADAMTSL4 (human + mouse) ✓ ✓ – –
THSD4 (human + mouse) ✓ ✓ – –
ADAMTSL3 (human + mouse) ✓ – – –
ADAMTSL5 (human) ✓ – – –
ADAMTSL1 (human) – – – –
PAPLN (human + mouse) – – – –
madd-4 (worm) – – – –

Two things follow, and they are different in kind:

  1. The NOT GO:0004222 sits at the wrong node — the ADAMTSL2 orthologue node rather
    than the ancestral node of the non-catalytic ADAMTSL clade. This is the AADACL2/3/4
    "right term, wrong node" pattern, in its too-deep form. Moving or duplicating that
    IKR to the clade LCA would state the catalytic loss for six more human genes in one
    edit. This belongs in suggested_questions once, naming all affected genes.
  2. GO:0030198 reaches some non-catalytic members and not others, on no visible
    biological criterion: ADAMTSL2, ADAMTSL4 and THSD4 receive it, ADAMTSL3 and ADAMTSL5 do
    not. ADAMTSL3 has more direct evidence for matrix organisation than several members
    that do receive it (§4, §5). Its GO:0030198 arrives instead from InterPro. The two
    pipelines happen to agree here; the point is that PAINT's coverage is ragged, not that
    the InterPro term is wrong.

The NOT GO:0006508 IRD row at PTN002673039 produces zero leaf annotations — even
ADAMTSL2 has no NOT-proteolysis row. Recorded as an observation; there may be an export
rule suppressing a NOT where no positive was projected.

Positive control for all of the above: the four catalytic ADAMTS members do receive
GO:0004222, so absence in the ADAMTSL rows is a property of the projection and not of
the query. The script raises if that control fails.

4. Molecular partners that GOA does not have

The most important paper for this gene's molecular function is not in its GOA at all.

PMID:22242013
— surface plasmon resonance against recombinant fibrillin-1 polypeptide rF90.

Internal contradiction in that paper, which changes the evidence code. The Results
describe the reagents as recombinant human ADAMTSL-1, -2, -3, and mouse papilin polypeptides, but the Methods say
PMID:22242013
(and, in the same paragraph, that ADAMTSL2 came from a mouse RIKEN clone, contradicting
the same sentence). A RIKEN clone plus mouse lung cDNA is mouse. Only ADAMTSL1 is
explicitly human (Full-length ADAMTSL1 was obtained from human fibroblast cDNA). The
conservative reading is that the ADAMTSL-3 polypeptide was mouse, so the human
annotation should be ISS with the mouse orthologue as supporting entity, not IPI. The
discrepancy is recorded rather than resolved — it is a question for the authors.

Explicit negative in the same paper, which bounds how far the fibrillin link can be
taken: PMID:22242013
No published work localises ADAMTSL3 protein to a microfibril or an elastic fibre. That
matters for the GO:0071953 row (§6).

5. TGF-β, and the strongest human-cell experiment

PMID:36539599

Loss of function (mouse): PMID:36539599

Gain of function, and this one is human cells with the human protein:
PMID:36539599
with the readout
PMID:36539599
and
PMID:36539599

Adenoviral overexpression is a perturbation, so this is IMP, not IDA. The two halves
agree in sign, which is what makes the negative-regulation call safe:
PMID:36539599

The mechanism is consistent with §4: LTBP-1 and fibrillin-1 are the two proteins that
sequester the large latent TGF-β complex in the matrix, and ADAMTSL3 binds both. The same
paper frames the family this way: PMID:36539599

Not annotated: the cardiac phenotypes themselves. Those are mouse, and a
pressure-overload phenotype is a stress response rather than a normal-physiology process
for this protein; the reviewable claim is the TGF-β regulation, which has human-cell
support.

6. The elastic-fiber TAS row

GO:0071953 located_in, TAS, from PMID:23962539 — a review, not primary data, and
full_text_available: false in our cache. Its abstract does not name ADAMTSL3.

Reference-scope check (the ACTR8/ACTRT3 discriminator): querying QuickGO by reference
returns 66 annotations over 62 distinct entities, all TAS, all GO_Central, split
GO:0071953 × 41 entities, GO:0001527 × 15, GO:0140149 × 8, GO:0140144 × 2. So it is
a bulk curation of an elastic-fibre proteome from a specialist review.

But it is not an indiscriminate family sweep, and that is the point that decides the
action. The GO:0071953 set contains ADAMTSL3, ADAMTSL4, ADAMTSL5 and THSD4 (plus mouse
orthologues) and excludes ADAMTSL1, ADAMTSL2 and PAPLN. A curator applying the term by
family membership would have taken all seven. The selection therefore reflects the
review's actual content, so this is a real traceable statement rather than a projection
artefact.

Weighed against that: no direct localisation of ADAMTSL3 to an elastic fibre or microfibril
has ever been published (§4, and the antibodies did not exist when the fibrillin work was
done). The honest position is KEEP_AS_NON_CORE — retain a specialist curator's traceable
statement, consistent with the fibrillin-1 binding, but do not treat it as an established
location. REMOVE would be overruling a curator on an inference I cannot check.

7. The 15 GO:0005515 rows

All 15 IPI rows, over 13 distinct partners, come from two CCSB systematic yeast
two-hybrid interactome maps: PMID:25416956 (HI-II-14) and PMID:32296183 (HuRI).

Querying IntAct directly (findInteractions/P82987) reproduces the ACRV1 finding exactly:
the detection methods across ADAMTSL3's records are two hybrid array ×16,
two hybrid prey pooling approach ×16 and validated two hybrid ×16 — three sub-methods
of the same Y2H pipeline, not three independent assays. So UniProt's NbExp=3 (and
NbExp=6 for the two partners found in both screens) counts Y2H variants, not orthogonal
confirmation. MI-scores are 0.56, or 0.72 for the two that appear in both maps.

Corrected after review — the aggregate does not answer the question. The first version
of this analysis counted detection methods across ADAMTSL3's whole IntAct record and then
asserted that no orthogonal assay existed for these pairs. That was a non sequitur, and the
committed results.json refuted it at the aggregate level: the same counter records
anti tag coip ×15 and one BioID. Disaggregating per pair is what settles it, and it
does so in favour of the claim: 0 of the 13 GOA partners carries any non-Y2H detection
method, while the anti tag coip and BioID methods belong entirely to the 17 IntAct
partners that are not in GOA (from PMID:40205054 and PMID:39232006, separate publications
that are not the source of any row reviewed here).

Writing that per-pair check exposed a second, quieter defect. IntAct returns partner
identifiers with a database suffix — A8MQ03 (uniprotkb), not A8MQ03 — so a lookup keyed
on the bare accession matches nothing, and "no non-Y2H method recorded" would have been
indistinguishable from "partner not found". The script now strips the suffix and raises if
any GOA partner is absent from the IntAct record, because a silent zero is exactly what this
check exists to rule out. Normalising the identifiers also merged a few duplicate labels, so
the degree numbers moved by one apiece: ADAMTSL3's own degree is 30 (was 31) and the
partner median is 188 (was 189). Small, but they were wrong, and the numbers quoted in
the review are the corrected ones.

Partner identity, all resolved to reviewed Swiss-Prot entries at canonical length (no
TrEMBL/ORFeome substitutions — the ACRV1 Q86WV8 trap checked and negative):

Quantitatively: the median number of distinct IntAct partners across the 13 is 188,
against 30 for ADAMTSL3 itself; six of them (CYSRT1 516, MDFI 483, KRT40 448,
KRTAP10-8 415, NOTCH2NLA 276, GLRX3 190) are extreme interactome hubs. Only KRTAP2-3 (10)
is not.

Compartment argument: ADAMTSL3 has a cleaved signal peptide and is a secreted matrix
protein. Hair-keratin-associated proteins are intracellular intermediate-filament matrix
proteins of the hair cortex; GLRX3 is cytosolic; MDFI is nuclear/cytoplasmic. In a Y2H
assay both partners are forced into the yeast nucleus, where the signal peptide is inert,
so the compartment mismatch is not tested by the assay that produced these rows. Nothing
in the ADAMTSL3 literature follows any of them up.

Verdict: MARK_AS_OVER_ANNOTATED on all 15, not REMOVE. These are real observations
recorded correctly by IntAct curators; what they do not license is a functional claim, and
bare protein binding conveys none anyway. Decided per partner (the brief's rule), and the
two that appear in both screens are not rescued by it: HI-II-14 and HuRI share the CCSB
ORFeome and search space, and both replicating partners (KRTAP10-8, MDFI) are among the
highest-degree hubs in the set.

Two newer screens are in IntAct but not in GOA and so are out of scope here: a
15-partner anti tag coip spoke-expanded set (PMID:40205054) and one BioID hit with CDH5
(PMID:39232006).

8. Neural biology — real, but mouse

Two strong mouse papers, both from the Tyagarajan group:

The perineuronal net is a specialised extracellular matrix, so the second paper is
independent corroboration that ADAMTSL3 is involved in extracellular matrix organisation —
cited in the GO:0030198 row for that reason rather than as a new term.

Synapse organisation is proposed as a NEW row with ISS, not IMP: the conditional
knockouts are mouse. The orthology is well founded — the C. elegans orthologue is a
synaptic organiser too (PMID:24896188 "C. elegans Punctin specifies cholinergic versus
GABAergic identity of postsynaptic domains") and MADD-4 signals through the DCC orthologue
UNC-40 (PMID:22014523), which is the same
receptor as in the mouse hippocampus. Human relevance is supported by ADAMTSL3 protein
being present in human cortical and ganglionic neurons (§1, PMID:17597111).

One caution worth writing down: MADD-4/Ce-Punctin is the orthologue of both ADAMTSL1
and ADAMTSL3 — PMID:22014523 — so worm synaptic phenotypes cannot be assigned to
ADAMTSL3 alone. The mouse conditional knockouts can.

9. Human genetics — real signal, but not GO-annotatable

A GWAS association is a statement about a locus, not about a gene product's activity, and
none of these identifies a molecular mechanism. None is used to support a GO term. They are
recorded because they are the reason this gene is studied and because they motivate the
suggested experiments.

10. Cancer and glaucoma claims — weaker than they read

11. Provenance and quality of the affinage record

gates_passed: False. The tripped gate is a suspected symbol collision flagged because the
narrative opens on C. elegans. On inspection that gate is a false positive: the
narrative correctly identifies human ADAMTSL3/punctin-2 and is discussing the worm
orthologue Ce-Punctin/madd-4, which is a legitimate part of this gene's story. Every claim
used from the record was nonetheless re-verified against the cited PMIDs, per the rule.

Defects found in the record itself:

  1. Mis-joined citation. It attributes "functions as a synaptic organizer specifying
    postsynaptic neurotransmitter receptor domain identity" to [PMID:14667842, PMID:24896188].
    PMID:14667842 is the 2003 ECM cloning paper and contains no synaptic content whatever;
    PMID:24896188 is about the worm protein. The claim about the human protein is assembled
    across two sources neither of which makes it. Every constituent statement is individually
    defensible; the error is in the join.
  2. A — in the PMID column. The 2025 CAR-T/TGF-β finding is sourced to bioRxiv with no
    PubMed id. It has since been published as PMID:42497245 (Sci Immunol 2026), but the
    claim is a germline variant-burden association in a trial cohort and supports no GO term.
  3. Over-reading of PMID:18474779 (§10).
  4. Misses the two most informative papers for this gene: PMID:22242013 (the fibrillin-1 /
    ADAMTS-10 / LTBP-1 binding, §4) and PMID:36539599 (the TGF-β knockout and human-cell
    rescue, §5). Its "the biochemical mechanism ... has not been characterized in the
    available corpus" is a statement about its corpus, not about the literature.

Retraction / erratum check: all 16 PMIDs relied on were fetched from PubMed and inspected
for RetractionIn, ErratumIn, ExpressionOfConcernIn and CorrectedandRepublishedIn in
their own CommentsCorrections records. All clean. Recording the negative so the next
reviewer knows the check was run.

12. Checks run, including the ones that came back negative

check result
WITH/FROM resolution on the IBA row 17 tokens; 16 protein donors + 1 PANTHER node; all 16 resolved
Donors' own experimental evidence for GO:0031012 16/16 carry their own IDA/HDA to the term or a descendant — so "these sources only carry the same family-level inference" is false here
Self-referential WITH/FROM tokens 0
Donor heterogeneity donors hold GO:0001527 microfibril, GO:0005604 basement membrane, GO:0005614 interstitial matrix and GO:0031012 — the general term is the genuine LCA, so no specificity upgrade is warranted (AADACL4 rule)
IBA less precise than its donors? (ACRV1 pattern) No. The donors disagree about the specific compartment, so the parent is correct; no downward MODIFY
Partner accessions vs canonical entries 13/13 reviewed Swiss-Prot at canonical length; no ORFeome/TrEMBL substitution
IntAct method audit one Y2H pipeline logged as three sub-methods, as on ACRV1
Reference-scope / projection test on the TAS row 62 entities, but selective within the family — not a blanket projection
Retraction and erratum scan 16/16 clean
Paralog agreement ADAMTSL4's merged review (genes/human/ADAMTSL4) reached GO:0050840 extracellular matrix binding as a NEW term for the same fibrillin-binding biology; this review uses the same term, so the two are consistent

A third check, run because the campaign brief insists on it, caught something nothing else
would have. Reconciling the raw count of reference_id: occurrences in the review YAML
against the parsed count gave 32 versus 60. The tempting reading — "the 15 identical
two-hybrid rows must be collapsing somehow" — would have been a rationalisation. The real
cause: the fifteen GO:0005515 rows shared one Python list object, so PyYAML emitted it
once as supported_by: &id001 followed by fourteen *id001 aliases. Valid YAML, parses
identically, and invisible to checkquotes.py and to the repo validator, both of which walk
the parsed document. It is the mirror image of the duplicate-key trap found on ACTMAP:
there, parsing destroys data the raw text contains; here, parsing creates data the raw
text does not show. Fixed by deep-copying per row and by dumping through a
SafeDumper subclass whose ignore_aliases returns True, with an assertion that no
&id/*id survives into the file. Final reconciliation is 60 = 60, with a strict
duplicate-key loader passing as well.

Two identifier traps hit and fixed while doing this, both of the "silent wrong answer"
kind:

12a. Three decisions taken under review, recorded so they are not re-litigated

  1. GO:0002020 protease binding added. The 2 nM SPR interaction with ADAMTS-10 was
    described in the review but carried no term of its own. It does now, as ISS with the
    same reagent-species caveat as GO:0050840. Deliberately not an inhibitor term: the
    contact is with ADAMTS-10's non-catalytic Tsp1 region
    (PMID:22242013),
    so nothing has been shown about its activity. The merged ADAMTSL4 review also carries
    GO:0002020, so the paralogues stay consistent.

  2. P35555 removed from the GO:0050840 WITH/FROM. For ISS the field takes the
    sequence-similar entity — mouse G3UXC7 — and the binding partner belongs there only on
    an IPI upgrade. Worth flagging as a general trap: supporting_entities is an
    unconstrained string list in the schema, so validation passes either way and this can
    only be caught by reading.

  3. The neuronal role is non-core, and the document now says so once. An earlier draft
    listed it in core_functions while the annotation row called it non-core — and the
    core_functions entry justified its inclusion with the same premise ("the evidence is
    entirely non-human") that the row used to exclude it. The core_functions entry has been
    removed and the annotation row is the single statement of the position: the direct
    evidence is entirely mouse and worm, and the molecular activity behind the DCC effect is
    unknown and recorded as an MF_DARK gap. The biology is still fully present in the gene
    description, in the annotation row, and in §8 above; its matrix component is carried by
    GO:0030198 in core function 1. Note the deletion alone would not have been enough — the
    contradiction lived in two passages, and both were touched.

13. For whoever reviews ADAMTSL1

Derived independently here, offered as claims to check rather than facts to inherit: