SPKW Methodology Notes

Parent project: SPKW.md

Overview

This document describes the analytical methods used to identify SPKW-unique annotations for review. The key insight is that naive queries overestimate the problem - using ontology closure to filter subsumed annotations reduces false positives by 70%+ in well-curated organisms.

The Problem: Naive vs True SPKW-Unique

A naive query finds all genes with SPKW annotations but no other evidence for the same term:

Gene X has:
  - GO:0006915 (apoptotic process) from SPKW
  - No other annotation to GO:0006915
→ Naive query flags as "SPKW-unique"

But this misses cases where experimental evidence exists for a more specific child term:

Gene X has:
  - GO:0006915 (apoptotic process) from SPKW
  - GO:0097300 (programmed necrotic cell death) from IDA  ← child of GO:0006915!
→ The SPKW annotation is REDUNDANT, not unique

Solution: Closure-Filtered Queries

The TRUE SPKW-unique query uses the ontology's is_a/part_of closure to exclude annotations that are subsumed by more specific experimental evidence:

An SPKW annotation to term X is NOT unique if there exists a non-SPKW annotation to term Y where Y is_a X

This dramatically reduces false positives:
- Human apoptosis: Naive query returns ~500 genes, closure-filtered returns ~280 (44% reduction)
- Well-curated organisms: 70%+ reduction in candidates
- Poorly-curated organisms: Minimal reduction (closure filtering has less impact)

SQL Queries

Core Query: Find SPKW-Only Annotations (Naive)

-- Find proteins with annotations ONLY from UniProt Keywords (no corroboration)
-- Database: ~/repos/go-db/db/goa_human.ddb (or organism-specific)

WITH annotations_to_term AS (
    SELECT a.db_object_id, a.db_object_symbol, a.supporting_references
    FROM gaf_association a
    INNER JOIN isa_partof_closure ipc ON a.ontology_class_ref = ipc.subject
    WHERE ipc.object = 'GO:0006915'  -- or other term of interest
      AND a.is_negation = false
),
genes_with_other_evidence AS (
    SELECT DISTINCT db_object_id
    FROM annotations_to_term
    WHERE supporting_references != 'GO_REF:0000043'
),
spkw_only_genes AS (
    SELECT DISTINCT db_object_id, db_object_symbol
    FROM annotations_to_term
    WHERE supporting_references = 'GO_REF:0000043'
      AND db_object_id NOT IN (SELECT db_object_id FROM genes_with_other_evidence)
)
SELECT * FROM spkw_only_genes ORDER BY db_object_symbol;

TRUE SPKW-Unique Query (Closure-Filtered)

For well-curated organisms, filter out SPKW annotations that are subsumed by more specific experimental evidence:

-- An SPKW annotation to term X is NOT unique if there exists
-- a non-SPKW annotation to term Y where Y is_a X

WITH spkw_annotations AS (
    SELECT a.db_object_id, a.db_object_symbol, a.ontology_class_ref, a.aspect
    FROM gaf_association a
    WHERE a.supporting_references = 'GO_REF:0000043'
),
gene_term_with_other_evidence AS (
    SELECT DISTINCT db_object_id, ontology_class_ref
    FROM gaf_association
    WHERE supporting_references != 'GO_REF:0000043'
),
-- Check if non-SPKW evidence exists for child terms
spkw_subsumed AS (
    SELECT DISTINCT s.db_object_id, s.ontology_class_ref
    FROM spkw_annotations s
    INNER JOIN isa_partof_closure ipc ON ipc.object = s.ontology_class_ref
    INNER JOIN gene_term_with_other_evidence g
        ON g.db_object_id = s.db_object_id
        AND g.ontology_class_ref = ipc.subject
),
true_spkw_unique AS (
    SELECT s.*
    FROM spkw_annotations s
    WHERE NOT EXISTS (
        SELECT 1 FROM gene_term_with_other_evidence o
        WHERE o.db_object_id = s.db_object_id
          AND o.ontology_class_ref = s.ontology_class_ref
    )
    AND NOT EXISTS (
        SELECT 1 FROM spkw_subsumed sub
        WHERE sub.db_object_id = s.db_object_id
          AND sub.ontology_class_ref = s.ontology_class_ref
    )
)
SELECT * FROM true_spkw_unique;

Evidence Source Distribution

-- Count annotations by evidence source
SELECT
    supporting_references,
    COUNT(*) as annotation_count,
    COUNT(DISTINCT db_object_id) as gene_count
FROM gaf_association
GROUP BY supporting_references
ORDER BY annotation_count DESC;

When to Use Each Query

Key Insight

Use closure-based queries as the default - they reduce false positives by 70%+ in well-curated organisms. The naive query is only useful when:
1. Exploring poorly-curated organisms
2. Wanting to see the full scope of SPKW coverage
3. Debugging or validating the closure-filtered results

For viruses, use both views when working across many clades. The virus-wide analysis in SPKW-VIRUS.md found 135,117 naive SPKW-unique annotations and 80,218 closure-filtered SPKW-unique annotations.

Stratifying by Swiss-Prot vs TrEMBL

Why This Matters

Keywords are assigned differently based on reviewed status:
- Swiss-Prot (reviewed): Keywords are manually assigned by curators
- TrEMBL (unreviewed): Keywords are automatically assigned via RuleBase/Spearmint/ARBA

This distinction affects interpretation: over-annotations in Swiss-Prot entries indicate problems in the KW→GO mapping, while TrEMBL over-annotations could be from bad mappings OR bad automatic keyword assignment.

Distribution Across SPKW Subprojects

Query: Stratify by Reviewed Status

For databases with swissprot table:

WITH spkw AS (
    SELECT g.db_object_id,
           CASE WHEN sp.uniprot_accession IS NOT NULL
                THEN 'Swiss-Prot' ELSE 'TrEMBL' END as status
    FROM gaf_association g
    LEFT JOIN swissprot sp ON g.db_object_id = sp.uniprot_accession
    WHERE g.supporting_references = 'GO_REF:0000043'
)
SELECT status, COUNT(*) as annotations, COUNT(DISTINCT db_object_id) as proteins
FROM spkw GROUP BY status;

For databases without swissprot table (use ID length heuristic):

SELECT
    CASE
        WHEN LENGTH(db_object_id) = 6 THEN 'Swiss-Prot'
        WHEN LENGTH(db_object_id) = 10 THEN 'TrEMBL'
        ELSE 'Other'
    END as status,
    COUNT(*) as annotations
FROM gaf_association
WHERE supporting_references = 'GO_REF:0000043'
GROUP BY status;

Limitations

• MOD databases (PomBase, FlyBase, TAIR) use organism-specific IDs, not UniProt accessions
• ID length heuristic is ~95% accurate but not perfect
• Some entries promoted from TrEMBL to Swiss-Prot retain automatic keywords

Keyword-Level View: Reverse-Mapping GO Terms to Source Keywords

All SPKW analyses above work at the GO-term level, because the GAF stores only the GO
term (ontology_class_ref) and supporting_references='GO_REF:0000043' — not the source
UniProt keyword. To ask the more direct question — which UniProt keywords drive
over-annotation? — reverse-map each SPKW-unique GO term to its source keyword(s) using the
public GOA/UniProt keyword2go mapping:

# 744 keyword->GO mappings (version date in the file header)
curl -sL https://current.geneontology.org/ontology/external2go/uniprotkb_kw2go -o kw2go.txt
# Parse "UniProtKB-KW:KW-0067 ATP-binding > GO:ATP binding ; GO:0005524" -> go_id, kw_id, kw_name
grep -vE '^!' kw2go.txt \
  | sed -E 's/^UniProtKB-KW:(KW-[0-9]+) (.+) > GO:.+ ; (GO:[0-9]+)$/\3\t\1\t\2/' > kw2go.tsv

Then LEFT JOIN kw2go on ontology_class_ref and aggregate by keyword. One GO term can
map from several keywords (e.g. defense response ← Plant defense / Defensin /
Antimicrobial), so pick the keyword whose biology matches the organism set.

The keyword watch-list (non-Arabidopsis plants, ranked by SPKW-unique gene count)

Tiering follows SPKW-PLANTS.md. The over-annotation risk is concentrated in
a small set of process / pathway / role keywords; the high-count keywords are otherwise
broad-but-correct binding/enzyme-class terms.

Two over-annotation mechanisms made explicit at the keyword level

1. Enzyme-class keyword → bare process term. Methyltransferase (a molecular-function
   keyword) maps to the generic process methylation (GO:0032259) — uninformative, and
   sometimes wrong about the substrate (protein vs DNA vs small-molecule methyltransferases
   all collapse to one term). The same shape produced cell division from Cell division.
2. Pathway/role keyword → developmental or defense process. Differentiation,
   Flowering, Nodulation, Plant defense attach broad process terms to genes that merely
   act during (or are induced by) the process — the recurring "expression ≠ function" and
   "responsive ≠ component" patterns.

Cross-organism reuse. Because keyword2go is organism-independent, this watch-list of ~30
process/role keywords is the same lever in any taxon — the keywords flagged here (defense,
hormone signaling, differentiation, methylation, cell division/meiosis, killing) are exactly
those that drove over-annotation in the human, S. pombe, and Arabidopsis subprojects.