ADNP2: did the ADNP NAP-peptide annotation defect propagate to the paralogue?

Generated by analyze_adnp2_propagation.py. Every number below is computed at run time from UniProt, QuickGO, PubMed and the two committed GOA tables; nothing is hardcoded except the merged ADNP review's published PxVxL figures, which are asserted as a precondition.

A. The NAP octapeptide is absent from ADNP2

Positive controls ['Q9H2P0', 'Q9JKL8', 'Q9Z103'] all carry the octapeptide, so a negative for ADNP2 is a result rather than a broken scan. Human and mouse ADNP2 contain no NAPVSIPQ and not one NAP tripeptide in 1131 / 1165 residues. No experiment on the NAP peptide can be an experiment on ADNP2.

B. PxVxL, with its null

Precondition satisfied: this scan reproduces the merged ADNP review's published figures exactly (1 hit at 820-824, null 0.758).

One occurrence in ~1100 aa is not enrichment for either paralogue: the compositional null is 0.758 for ADNP and 2.705 for ADNP2. What distinguishes ADNP2 is not the count but that PMID:38960717 mutated the motif and lost HP1beta binding.

C. Which of ADNP2's PxVxL candidates is the homologue?

ADNP2 has 2 P-x-V-x-[LMIV] candidates (PPVLV@662-666, PSVLL@1107-1111) against a compositional expectation of 2.705. Presence of the motif is therefore uninformative on its own, and the count cannot be used as evidence.

The discriminating test is projective. Global BLOSUM62 alignment of ADNP vs ADNP2 (27.1% identity) maps ADNP's motif start (820) onto ADNP2 position 1107. Does that land on one of ADNP2's candidates? True ({'match': 'PSVLL', 'start': 1107, 'end': 1111}).

ADNP   IASHFSNKRKKCVRDCEKYKPGVLLGFNMKELNKVKHEMDFDAEW
ADNP2  VASFFGKRRYICMKAIKNHKPSVLLGFDMSELKNVKHRLNFEYEP

The absolute positions differ only because ADNP carries 278 residues after the motif against ADNP2's 20 -- the poorly conserved C-terminal extension described in PMID:38960717. The second ADNP2 candidate has no ADNP counterpart and no experimental support.

D. Which route carried each ADNP2 row, and what did not cross

ADNP2 GOA: 18 rows (18 distinct), 8 terms. ADNP GOA: 53 rows, 33 terms.

Sequence-similarity donors for ADNP2 (GO_REF:0000107 Ensembl Compara and GO_REF:0000024 UniProt ISS): UniProtKB:Q8CHC8
Sequence-similarity donors for ADNP: UniProtKB:Q9JKL8, UniProtKB:Q9Z103

ADNP's rat-Compara block covers 19 terms. Terms from that block that also appear on ADNP2: none.

E. What molecule was assayed behind each IBD seed

4 of 4 seed/term pairs carry their own experimental annotation in the propagated term's subtree.

Partial confirmation. 1 seed annotation(s) rest on a reference whose own title names the NAP peptide. This is inside the donor set of an ADNP2 row, but it is not the whole donor set -- see the table above for the co-seed's evidence, which is a gene-product experiment.

F. Logical-opposite citation cross-product

NEGATIVE -- ADNP2 carries no logically opposed term pair, so the cross-product defect found on ADIPOQ cannot occur here. (opposed pairs found: 0)

G. TFClass node reach, and the import's own exclusion set

tfclass:3.1.8 reaches 14 human gene products (28 annotations). Positive control: the subject (UniProtKB:Q6IQ32) and its paralogue (UniProtKB:Q9H2P0) are both present, so this is a census of the right set. Per-entity term signatures: {'GO:0000785,GO:0000981': 14}.

So every gene the node reaches receives the identical pair, which means GO:0000981 on ADNP2 is a property of class membership rather than a judgement about ADNP2.

Widening to the whole import: GO_REF:0000113 = 1436 annotations over 727 distinct entities, evidence codes {'ISA': 1436}. Of those entities, 709 receive GO:0000785+GO:0000981, 18 receive GO:0000785 alone, and 0 carry some other signature. Entity counts are derived as a distinct set of gene-product ids, not from the annotation total.

The 18 chromatin-only entities are the import's own negative control — the pipeline already withholds the molecular-function term where it does not apply. Listed as a set rather than a count, because this is the payload of the ask to NTNU_SB and a curator needs something diffable:

Which excluded entities carry a DNA-binding domain? Rendering the DOMAIN note alongside the span, because the notes are not uniform and the difference matters. Of the 18 excluded entities, 3 carry an annotated DNA_BIND feature at all:

ADNP2's own feature is 1043-1102 (Homeobox). So the fold-symmetry precedent is HMBOX1 — annotated with the identical note — and not HOPX, whose domain UniProt calls degenerate (3-62 (Homeobox; degenerate)). That asymmetry is stated here rather than left in the JSON for a reader to find, because it is the first thing a curator would raise: if HOPX were the only precedent, the reply would be that HOPX is excluded because its homeodomain is broken, which would not transfer to ADNP2's intact one. The script asserts that at least one excluded entity has a non-degenerate domain and refuses to report if none does.

HOPX still carries the biological half of the precedent — UniProt describes it as an atypical homeodomain protein that does not bind DNA, and it is in the exclusion set (asserted, not assumed: the script fails if it is not). What it does not carry is fold symmetry with ADNP2.

Is the exclusion a per-entity judgement or per-node coverage? This decides whether the precedent supports the ask at all: if GO_REF:0000113 could only withhold GO:0000981 for a whole node, then excluding ADNP2 would also exclude ADNP — whose sequence-specific binding is measured — and the correct request would be a different one. Answer, from the data: per-entity, demonstrated = True (an existential: it holds if any node retains the term for some members while withholding it from a strict subset).

The import uses both granularities, and that distinction has to be stated rather than filtered away. The excluded entities are spread across 11 nodes, of which 7 withhold the term from a strict subset (10 entities) while 4 withhold it from every member (8 entities). All 11 are printed:

2 of the single-exclusion strict-subset nodes sit in the same TFClass class as ADNP2: HOPX excluded alone out of 47 members of tfclass:3.1.3, and HMBOX1 excluded alone out of 19 members of tfclass:3.1.10 — while ADNP2's own node tfclass:3.1.8 currently has 14/14 members holding the term. So single-entity exclusion inside a populated homeodomain node is something this import already performs, twice within class 3.1, and the request needs no new mechanism and would not touch ADNP: it takes one node from 14/14 to 13/14. The wholly-excluded nodes are not the precedent ADNP2 needs — but they do show the import has the coarser granularity too, so naming which one the ask relies on matters.

That also settles what these 18 entities do not have in common. They are biologically heterogeneous: NFX1 "Binds to the X-box motif of MHC class II genes", which is sequence-specific binding at a cis-regulatory region, and DMRTC1 is named a transcription factor. Three successive drafts generalised the set as "the non-DNA-binding members", then as "none is a sequence-specific polymerase II transcription factor", then as "each is an exclusion inside a node whose other members keep the term"; all three are false, each refuted by a table in this same document. So the statement below is computed from the partition rather than written:

They share no property at all. 10 of the 18 sit in nodes whose other members keep the term, but the remaining 8 sit in 4 nodes where NO member keeps it, so not even the structural description holds across the set. What can be said is only per-node, which is why the table above is the claim and this sentence is not.

And the positive argument for ADNP2 is neither of those. It is the measured failure to find a motif: no sequence motif explains ADNP2's ChIP-seq distribution, its peaks avoid transcription start sites, and PxVxL mutation nearly abolishes its chromatin binding. The exclusion set shows only that this import has a mechanism for withholding GO:0000981 and already applies it to intact-domain proteins; it does not itself argue that ADNP2 belongs there.

Note what does not depend on any of this: the GO:0000981 verdict rests on the quoted three-clause failure against GO:0003700's definition. This section supports the upstream ask, not the annotation action.