"All" pathway semantics behave as "any": HAVING compares a per-gene group count to a total row count

[jbrestel]

In Model/lib/wdk/model/questions/queries/pathwayQueries.xml, the any_or_all_pathway "all" branch does not implement "all". It is a no-op in at least one search and probably close to one in the others.

The defect

Take PathwaysByGeneList:

GROUP BY tp.pathway_source_id, tp.pathway_source, tp.gene_source_id, tp.ec_number_gene, tp.ec_number_pathway
HAVING (
  ('$$any_or_all_pathway$$' = 'any')
  OR (count(*) <= (SELECT count (*)
                   FROM apidbtuning.TranscriptPathway tp , dsGene ds
                   WHERE tp.gene_source_id = ds.gene
                  )
  ))

The GROUP BY includes tp.gene_source_id, so each group is one gene in one pathway with one EC pair — count(*) per group is typically 1. The subquery counts every pathway-gene row for the entire input list. So the comparison is roughly 1 <= (large number), which is true for every group. The all branch filters nothing.

For "pathways containing all of these genes", the comparison needs to be at the pathway level — something like count(distinct gene_source_id) = <number of input genes> after grouping by pathway alone, not per gene.

Evidence

EXPLAIN ANALYZE of PathwaysByGeneList against genomicsdb_071n, 5-gene input (PF3D7_0103700, PF3D7_0109500, PF3D7_0109850, PF3D7_0110600, PF3D7_0202900), pathways_source=1, exclude_incomplete_ec=0, exact_match_only=0:

mode	rows out
any	3687
all	3687

Identical. With any the planner constant-folds ('any' = 'any') and drops the subquery entirely; with all it evaluates it as an InitPlan and every group still passes.

Affected sites

Query	Line (approx)	Comparison
PathwaysByGeneList	~234	count(*) <= — never filters
PathwaysByGeneIds	~414	count(*) = — same grouping problem, also grouped per gene
PathwaysByCompounds	~307	count(*) = — same shape, worth checking against compound grouping

The <= in PathwaysByGeneList is the most clearly wrong; the = variants may filter something, but not the intended thing, since the group is still per-gene.

Why this wasn't caught

It fails open. "All" silently returns the "any" result set, which is a superset containing the correct answer — so results look plausible and nothing errors. Only a count comparison between the two modes reveals it.

Not in scope of #227

Noticed while fixing the gene-alias resolution in these same queries (#227), and deliberately left untouched there — it is pre-existing and needs a semantics decision, not a mechanical fix. Someone should confirm what "all" is meant to mean in the UI before rewriting the grouping.

🤖 Generated with Claude Code