ARBA00027234: search for additional phosphorylation demo candidates

ARBA generic phosphorylation audit

No defensible additions to the 48-gene demo shortlist from this screen.
This is a focused candidate search, not a completed formal rule review.
Screened 2026-09-28 PDT / 2026-09-29 UTC.

What was checked

ARBA00027234 predicts
GO:0006468 protein phosphorylation through 90 alternative condition sets.
Its current record is dated 2026-05-04. The
saved rule preserves the exact conditions.

Each condition set was translated into a UniProt search with AND between
conditions, OR within each condition's values, and its taxonomic constraints
retained. All 90 sets were queried against Swiss-Prot, both with and without
excluding GO:0004672 protein kinase activity and descendants. The searches
returned 370 distinct Swiss-Prot proteins, and zero lacking a positive
protein-kinase annotation. All result sets were completely retrieved.
The match table records proteins and their matching
condition-set numbers. audit.json contains the 180 query URLs,
counts, provenance checks, and the caution cases below.
This saved audit was assembled during the original analysis and enriched
outside scan_rule.py; it is not the script's verbatim output. The field
mapping and the separately selected evidence are documented below.

The PANTHER subfamily condition (set 9) needs an exact quoted text search
for PTHR11042:SF160. An xref:panther-"PTHR11042:SF160" search returned
zero even though the exact-text query found 10 proteins. The corrected
query is used in the saved results. InterPro conditions use structured
cross-reference queries; FunFam conditions use exact quoted identifiers.

Absence of a kinase annotation is a screening criterion, not the only test
for error. The returned records were also searched for explicit inactivity
and pseudokinase cautions. Additional exploratory samples included unreviewed
matches of unusual-looking conditions: the PH-domain condition matched
ROCK/MRCK kinases; the paired PKN regulatory-domain condition matched PKN
kinases; the amphibian condition matched insulin/IGF receptors; bacterial
conditions matched histidine kinases or Wzc/Etk tyrosine kinases. These were
not convincing regulator/substrate mistakes. The unreviewed proteome was
not exhaustively screened, and this audit does not establish that the rule
is universally correct.

Rule records versus emitted annotations

QuickGO queries with withFrom=ARBA:ARBA00027234 returned zero current
annotation rows. The same was true for ARBA00026648, ARBA00026662,
ARBA00085084, and ARBA00088043. A positive control, ARBA00089890, returned
92 rows, including a current GO_REF:0000117 IEA annotation explicitly
naming that rule. Requests and complete first-page responses are preserved
in audit.json.

Consequently, a rule record with a GO output is not sufficient evidence
that a particular protein currently carries an ARBA-derived assertion.
Domain-condition matches were not inserted into the demo's list of current
GOA annotations. These observations do not establish why the five rules
have no current rows in QuickGO, nor prove that they were never applied.

Caution case: IRAK3, condition set 68

Set 68 requires these three FunFams together:

It matches human IRAK3/Q9Y616
and mouse Irak3/Q8K4B2.
Both UniProt records explicitly flag replacement of the expected catalytic
Asp and disagreement about kinase activity. However, both also report low
autophosphorylation activity and carry experimentally supported kinase and
phosphorylation annotations: human PMID:10383454 and mouse PMID:12054681.
These are disputed/nuanced candidates, not obvious errors. No experimental
annotation has been rejected on the basis of the caution text. Human IRAK3
also already has repository reviews and is ineligible for the original demo.

Mouse Irak3 could be a separate future pseudokinase discussion case after
reading the conflicting studies. It is not counted as an ARBA-derived
misannotation or appended to the current shortlist.

Evidence provenance

The committed audit.json is the dated analysis record. The script reruns
the searches and writes a new audit.json with a simpler structure. The
original validation compared query counts and matched accessions, not the
complete JSON documents. Use the following mapping when comparing a rerun:

Committed audit field Source or corresponding script output
retrieved_at_utc, rule_id, condition_sets, distinct_swissprot_matches Same fields in the script output; the timestamp changes on each run.
reviewed_queries Script queries entries with exclude_kinases: false, with that flag removed. Compare by condition_set.
negative_queries Script queries entries with exclude_kinases: true, with that flag removed. Compare by condition_set.
quickgo_rule_queries Same field in the script output, including URLs and first-page API responses.
rule_modified_date Copied outside the script from modifiedDate in the saved rule JSON.
target_go Selected outside the script from the rule's GO output: GO:0006468.
swissprot_matches_without_protein_kinase_annotation Summary added outside the script. All saved negative-query counts are zero; the corresponding script field nonkinase_queries_with_hits is empty. For a future nonzero result, count distinct accessions rather than summing overlapping condition-set counts.
new_demo_candidates Manual triage outcome (zero), not a search result or an automated decision.
caution_cases Manually selected IRAK3 records, with verbatim UniProt CAUTION comments and selected GO cross-references; see extraction details below. The script does not select or adjudicate caution cases.

The two caution records came from UniProt entries
Q9Y616 and
Q8K4B2. For each record,
uniprot_id and entry copy primaryAccession and uniProtkbId;
condition_sets comes from the search membership table. caution copies
the comments entries with commentType: CAUTION. relevant_go copies
uniProtKBCrossReferences entries with database: GO and IDs
GO:0004672, GO:0004674, GO:0006468, or GO:0046777, where present.
These extracts preserve the source evidence; the decision to discuss these
two proteins and call them disputed remains a manual interpretation.

Each script run also writes matched-proteins.json, containing the full
UniProt records for all distinct matches, including the comments and GO
cross-references needed to inspect caution cases. That file was omitted
from the commit because the original rerun produced about 32 MB; the
selected IRAK3 extracts were retained in the committed audit instead.
The committed reviewed-matches.tsv was also enriched outside the script
with gene, organism, and protein_name columns. The script's table
contains accession, entry, taxon, and condition_sets; compare those
shared columns by accession, not by row order.

Reproduction

From the repository root:

python3 projects/PHOSPHORYLATION_REFACTOR/demo/arba-audit/scan_rule.py \
  --rule ARBA00027234 --output /tmp/phosphorylation-arba-recheck

This refetches the rule, reruns both Swiss-Prot searches for every condition
set, checks rule attribution in QuickGO with a positive control, and writes
fresh evidence to the chosen output directory. It fails on unknown condition
types or API errors rather than treating those failures as zero matches.
It never creates gene reviews or changes the demo candidate list.
The output files are the fetched rule JSON, search audit.json, full
matched-proteins.json, and reviewed-matches.tsv. Repeating the search
does not regenerate the manual additions described above, and live database
changes may alter the results. Keep reruns separate from the dated snapshot.

Back to the demo shortlist