SPKW ViralZone Upstream-Source Subproject

SPKW ViralZone Upstream-Source Subproject

Parent project: SPKW.md · Sibling: SPKW-VIRUS.md

Overview

Every other SPKW subproject audits a downstream product: annotations that a
keyword-to-GO mapping produced on particular proteins. This one audits the
upstream source those verdicts quietly depend on.

SPKW-VIRUS.md established VZ-primary as a confidence flag: a
Swiss-Prot keyword that is the primary keyword of a ViralZone page is much less
likely to be one of the generic keyword artifacts driving the virus-wide SPKW
burden, so the flag raises the prior on the annotation. That is a claim about
ViralZone's editorial quality — and nobody had checked it.

Checking it turns out to matter more than expected, because ViralZone does not
feed GO once. It feeds GO twice, by two routes that meet again downstream:

  1. Definition route. GO cites ViralZone pages as definition cross-references
    under the prefix VZ — GO:0039713 viral factory carries VZ:1951. Where the
    citation marks real textual reuse, the ViralZone page is effectively the term's
    authored definition.
  2. Annotation route. A ViralZone page may declare one primary UniProt keyword
    in its header. That keyword maps to a GO term and generates the SPKW
    (GO_REF:0000043) IEA annotations the parent project reviews.

So a single paragraph of ViralZone prose can determine both what a GO term
means and which proteins are annotated to it. When both routes run from the
same page, an error upstream is not caught by cross-checking the annotation
against the term definition — because the definition came from the same place.
That compounding is the object of this subproject.

Scope and Method

All figures on this page are produced by
viralzone/build_viralzone_audit.py, which
enumerates ViralZone from its public sitemap.xml, fetches each page, reads GO's
VZ: cross-references from the local OAK GO build, and writes three tables:

Table Contents
vz-go-xrefs.tsv every GO term citing ViralZone, with a text-reuse score
vz-dead-xrefs.tsv GO terms citing a page absent from the live sitemap
vz-dual-path.tsv pages on both routes (GO definition xref and primary keyword)

Reuse scoring. For each sentence of a GO definition, the best-matching window
of the cited page's prose is scored with difflib.SequenceMatcher on normalized
text; the term's score is its best sentence. Two cautions are load-bearing:

Page prose is cut at the first navigation/UniProt-table marker: ViralZone appends
the full list of matching UniProt entries to keyword pages (VZ:982 is 161 KB of
HTML for ~450 characters of definition), which would otherwise swamp the text.

Cross-check on the keyword extraction. This script reads the primary keyword
from the (kw:KW-####) string rendered in the page header and finds 147
pages carrying one. The 2026-05-21 audit behind SPKW-VIRUS.md
read the same field from the header's data-about attribute and also found 147.
Two independent readings agreeing is the reason to trust either.

Results

Propagation footprint

Measure Count
ViralZone pages in the live sitemap 1,170
...with prose successfully fetched 1,158
...declaring a primary UniProt keyword (annotation route) 147
GO terms carrying a VZ: dbxref (definition route) 152
...scorable (cited page still live) 148
...by aspect 126 BP / 22 CC
ViralZone pages cited by GO 144
Pages on both routes 107
...keyword only 40
...definition xref only 37

There is no molecular-function term in the set: ViralZone's influence on GO is
entirely on process and component.

Text reuse is the exception, not the rule

Reuse band Best-sentence similarity Terms Share
VERBATIM_ISH ≥ 0.85 10 7%
CLOSE_PARAPHRASE 0.70–0.85 12 8%
REWORDED 0.55–0.70 24 16%
INDEPENDENT < 0.55 102 69%

This is the subproject's first substantive finding, and it is reassuring: in
roughly seven cases out of ten, GO curators used ViralZone as a reference and
wrote their own definition. The naive worry — that GO's viral vocabulary is a
transcription of ViralZone — is wrong as a general claim.

It is right for a minority, and that minority is where the two routes overlap
most:

keyword present keyword absent
VERBATIM_ISH 9 1
CLOSE_PARAPHRASE 8 4
REWORDED 21 3
INDEPENDENT 75 27

17 of the 22 high-reuse terms also sit on the annotation route. For those, the
GO definition and the SPKW annotations trace to one unreferenced paragraph, and a
VZ-primary flag on the annotation adds no independent confirmation of the term
it is annotated to. This is the concrete form of the compounding risk, and it is a
sharpening of SPKW-VIRUS's working rule rather than a contradiction of it.

Highest-reuse definitions (audit queue)

All 22 are live (non-obsolete) terms. kw marks the annotation route.

GO term Label Page best mean kw
GO:0099045 viral extrusion VZ:3951 0.99 0.81 KW-1249
GO:0075520 actin-dependent intracellular transport of virus VZ:991 0.99 0.73 KW-1178
GO:0099009 viral genome circularization VZ:3968 0.91 0.75 KW-1253
GO:0099000 symbiont genome ejection through host cell envelope, contractile tail mechanism VZ:3950 0.90 0.63 KW-1242
GO:0060141 symbiont-mediated induction of syncytium formation VZ:5957 0.88 0.62 KW-1180
GO:0099001 symbiont genome ejection through host cell envelope, long flexible tail mechanism VZ:3952 0.87 0.64 KW-1243
GO:0075525 viral translational termination-reinitiation VZ:858 0.86 0.66 KW-1158
GO:0039592 symbiont-mediated arrest of host cell cycle VZ:876 0.85 0.67 KW-1079
GO:0039690 positive stranded viral RNA replication VZ:1116 0.85 0.67 —
GO:0075526 cap snatching VZ:839 0.85 0.81 KW-1157
GO:0099018 symbiont-mediated evasion of host restriction-modification system VZ:3966 0.85 0.64 KW-1258
GO:0039698 polyadenylation of viral mRNA by polymerase stuttering VZ:1916 0.84 0.72 —
GO:0039704 viral translational shunt VZ:608 0.84 0.84 —
GO:0039708 nuclear capsid assembly VZ:1516 0.83 0.69 —
GO:0098931 virion attachment to host cell flagellum VZ:3949 0.83 0.71 KW-1240
GO:0099002 symbiont genome ejection through host cell envelope, short tail mechanism VZ:3954 0.83 0.67 KW-1244
GO:0039713 viral factory VZ:1951 0.82 0.74 —
GO:0098669 superinfection exclusion VZ:3971 0.79 0.59 KW-1260
GO:0046755 viral budding VZ:1947 0.78 0.75 KW-1198
GO:0039666 virion attachment to host cell pilus VZ:981 0.78 0.54 KW-1175
GO:0039587 symbiont-mediated-mediated suppression of host tetherin activity VZ:665 0.77 0.57 KW-1084
GO:0098671 adhesion receptor-mediated virion attachment to host cell VZ:3943 0.76 0.60 KW-1233

Two rows connect directly to an existing verdict. GO:0099018 (R-M evasion) and
GO:0099000 (contractile-tail genome ejection) are both named in
SPKW-BPT4.md as the good phage-specific replacement targets;
GO:0099000 carries an explicit ACCEPT there for T4 18/P13332, 19/P13333
and 20/P13334, reasoned from the tail-contraction mechanism. Those calls still
look right — the sheath/tube/portal argument stands on UniProt's own description,
not on ViralZone. But the definitions of both terms score in the high-reuse band,
so "the GO definition supports this" and "ViralZone supports this" are one
argument, not two, and the ACCEPT should rest on the mechanism rather than on
their agreement.

Worked case: the viral-factory branch

GO:0039713 viral factory scores 0.82/0.74 against VZ:1951. The branch
beneath it — spherule, double-membrane vesicle, tube viral factory, virogenic
stroma, peristromal region — appears to be drawn from the same page's sections,
but only the parent carries a VZ: dbxref, so none of the children is in the
scored table above: that reading is manual, not measured, and is flagged as
such in the module's scope note. Sizes ("50-400nm diameter membrane invagination",
"200-300nm... derived from the endoplasmic reticulum or Golgi") and family lists
travel verbatim into the definitions.

GO's curator did filter: VZ:1951 states that viroplasms are produced by
"ssRNA(-) viruses like filoviridae", and GO:0039716 viroplasm viral factory
drops that clause. So the definition route is not a blind copy even at its most
derivative. What is not known is whether other ViralZone claims were filtered as
carefully — which is what the audit queue is for.

This is recorded as scope note (5) in modules/viral_factory.yaml,
so a reader of that module does not mistake its GO groundings and ViralZone for
independent support.

Maintenance defects found in passing

Neither of these needs a biology judgement; both are filable now.

Four dead cross-references. GO cites pages absent from the live sitemap:

GO term Label Page
GO:0039525 symbiont-mediated perturbation of host chromatin organization VZ:899
GO:0039628 T=169 icosahedral viral capsid VZ:10117
GO:0052150 symbiont-mediated perturbation of host apoptosis VZ:1518
GO:0039659 (obsolete) suppression by virus of host TBK1-IKBKE-DDX3 complex activity VZ:719

The first three are live terms whose stated provenance no longer resolves.

A doubled-word label. GO:0039587 reads
"symbiont-mediated-mediated suppression of host tetherin activity" — a
straightforward GO label typo, surfaced by reading 152 labels in one pass.

Nine obsolete terms still carry VZ: cross-references (GO:0019048,
GO:0019051, GO:0039511, GO:0039526, GO:0039553, GO:0039586,
GO:0039589, GO:0039655, GO:0039667). Harmless, and useful as provenance for
why the terms were once created; noted so the audit queue can skip them.

Source-Quality Assessment

Read as an evidence source rather than a vocabulary, ViralZone has one
disqualifying property and several minor ones.

No inline references, anywhere. VZ:695 asserts that "Vaccinia K1 inhibits
NF-kappa-B by preventing I-kappa-B degradation" and that "A238L from African swine
fever virus interacts with free NF-kappa-B to prevent its nuclear translocation"
with no PMID attached to either. Every ViralZone claim is therefore tertiary and
untraceable without independent lookup. This is why ViralZone is registered in
conf/oak_config.yaml as VZ: null — an explicitly
unvalidated prefix — and why a ViralZone quote should never be the sole support
for a curation action in this repository.

Copy quality is below journal standard. "Attachement", "neo-synthetized",
"NF-kBdependent", "occulsion", "african swine virus" for African swine fever
virus, and a page titled "Poylprotein splitting". None of this is disqualifying on
its own; it is a signal about editorial process, and it is a reason to read a
quoted sentence rather than trust it.

Prose is short and useful. Median ~1,300 characters after the UniProt table is
stripped. The process pages read like GO definitions already; the family pages are
structured mini-records (VIRION / GENOME / GENE EXPRESSION / ENZYMES /
REPLICATION). The whole audit queue is 24,824 characters of prose — a day's close
reading, not a programme.

The useful content is largely already harvested. GO has absorbed the
definitional material into terms with taxon constraints and a hierarchy.
ViralZone's remaining value to this repository is breadth on families GO has not
termed, plus provenance for why terms say what they say — not rigour GO lacks.

Working Rules (refines SPKW-VIRUS)

  1. VZ-primary remains a positive prior, with one subtraction. Where the
    annotated term is also a high-reuse term (the 17 in the crosstab above), the
    flag and the term definition are the same evidence. Do not count them twice.
  2. Never cite ViralZone as sole support. Use it to locate a claim, then cite
    the primary literature that establishes it.
  3. Cite pages as VZ:<id>, GO's own dbxref spelling and the numbering used
    throughout this project — not the bioregistry viralzone prefix.
  4. A VZ: citation carries no verified quote. The repository's
    supporting-text validator resolves PMID:/DOI: only, so a supporting_text
    on a VZ: source passes unchecked. Prefer statement over supporting_text
    for ViralZone evidence items.

Work Queue

# Task Scope Status
1 File the four dead VZ: cross-references with GO 4 terms Not started
2 File the GO:0039587 doubled-word label typo 1 term Not started
3 Audit the 22 high-reuse definitions against primary literature 24,824 chars Not started
4 For each, record divergences both ways: a ViralZone error that propagated into GO, and a ViralZone correction GO missed 22 terms Not started
5 Grade the 107 dual-path pages, so SPKW inherits a per-page confidence rather than a flat VZ-primary boolean 107 pages Not started
6 Decide whether VZ: should become quote-checkable (needs a page cache; see the PMCID/Reactome gap in the same validator) infra Open question

Task 3 is the only one requiring literature work, and the filovirus/viroplasm case
is the model: establish what the page claims, what GO carried over, and whether the
difference was a judgement or an accident.

Status