Re-ground 9 GENE_OR_PROTEIN-object enables edges by idiom (#334)

[realmarcin]

biolink declares enables range 'biological process or activity' — only
BIOLOGICAL_PROCESS, PATHWAY and MOLECULAR_FUNCTION satisfy it, so the 33 edges
pointing elsewhere each entail a false type.

#334 asks for group-by-idiom rather than one sweep, because the 33 mix distinct claims.
This takes the GENE_OR_PROTEIN-object group's unambiguous half.

What reading them showed

The 13 protein-object edges are roughly seven idioms, not one. Only those with an exact
and range-safe upstream term are migrated:

target	edges	idiom
encodes (biolink:encodes)	1	a gene cluster encodes its product
part of (biolink:part_of)	3	a subunit is part of its complex
contributes to (RO:0002326)	5	an energy source or acquired repertoire

All three declare no rdfs:domain or rdfs:range — checked, not assumed — so none can
reintroduce the class being removed. All three already had mapping rows, so nothing new
was minted.

Four deliberately left

Each needs a modelling decision, not a relabel:

• molecular_oxygen → cytochrome bd and cytoplasmic_na → Mrp complex — substrate
  availability; the honest form is probably the complex taking the chemical as has input,
  which reverses the edge.
• flagellar_motor_torque → flagellar_motor — the subject is a property of the object;
  the edge is confused.
• Hfq → RpoS — "enables sRNA-dependent translation of RpoS": the object should be that
  translation process, which isn't currently a node.

Result

ENABLES_RANGE_VIOLATION: 33 → 24. The remaining 24 stay baselined under #334.

A nice confirmation

This change touched 8 trait files and rewrote 8 pages. Before #304 landed earlier today
it would have rewritten 477 — the reviewability fix paying off on the very next data PR.

What was checked

just qc green · just validate-strict 477 / 0 errors · 434 tests pass ·
audit-graphs unchanged at 1535 · snippet baseline steady at 2738 · curation events on all
8 records with histories verified in order.

Partial burn-down of #334 (stays open for the remaining 24).

🤖 Generated with Claude Code

[culturebot-reviewer]

Claude finished @realmarcin's task in 4m 29s —— View job

---

Review

The per-idiom split is the right call, the baseline/report/page regeneration is consistent (33 → 24 rows, verified in conf/predicate_domain_audit_baseline.tsv), and two of the nine edges are unambiguously improved (flii_atpase_complex part of ft3ss_export_system, virulence_gene_acquisition contributes to human_virulence_factors — that one fits biolink's "contributes to the ... generation of the other" exactly). Three problems below.

---

🟡 biolink:encodes is not a term in the pinned biolink model.
data/traits/metabolism/methanogenesis.yaml:165

data/raw/biolink-model.yaml is version 4.4.0 and has no slot encodes — the gene→product slot is has gene product (line 3421, domain: gene, range: gene product mixin, exact_mappings: RO:0002205). The repo's own checker already records this: reports/biolink_coverage.tsv:28 is applied encodes biolink:encodes with both biolink_via_label and biolink_via_exact_mapping empty — the only applied biolink: CURIE in that report with no backing slot. mappings/predicate_grounding.tsv:28 calls it an "exact label match against biolink slot (RO:0002205 inverse: encoded_by)"; neither encodes nor encoded by is a slot, and RO:0002205 belongs to has gene product.

This is pre-existing — 8 other trait files already use it — so it is not a defect this PR introduced. But the PR asserts "All three declare no rdfs:domain or rdfs:range — checked, not assumed", and for this one the check is vacuous: a term absent from the model trivially declares nothing. Either ground the edge to RO:0002205 / biolink:has_gene_product (noting that a gene cluster → protein complex is not a clean fit for range: gene product mixin, which is presumably why encodes got minted), or keep biolink:encodes and correct the mapping-row note to say plainly that it is a local coinage with no upstream slot. What should not stand is the note claiming an exact biolink match that biolink_coverage.tsv contradicts.

🟡 mreB part of rod_complex contradicts its own evidence and its own node description.
data/traits/morphology/rod_shaped.yaml:108, data/traits/morphology/bacillus_shaped.yaml:106

The snippet on the edge is "MreB filaments interact with Rod complexes". The target node in the same file is described as "Cell-wall elongation machinery associated with MreB". The edge description still reads "MreB organizes rod-shape elongation machinery." Three statements of association, and the predicate now asserts parthood.

Whether MreB is a constituent of the Rod complex or a filament that guides it is genuinely unsettled, and the corpus is split on it — cell_shape.yaml:64 types rod_complex as PATHWAY and lists MreB inside it, while these two files type it GENE_OR_PROTEIN and put MreB outside it. That is exactly the kind of modelling decision the PR says it is deferring elsewhere. interacts with (biolink:interacts_with, already mapped, range-unconstrained) says what the quoted sentence says and needs no decision.

🟡 The IMF edges are the same idiom the PR defers, migrated instead of deferred.
motile.yaml:148, motility.yaml:121, flagellated.yaml:142, gliding.yaml:107

Biolink defines contributes to (line 3853) as "holds between two entities where the occurrence, existence, or activity of one contributes to the occurrence or generation of the other." Ion motive force does not contribute to the occurrence or generation of the flagellar motor — the motor assembles without it; IMF powers its rotation. The descriptions say so: "Ion motive force powers rotation of the flagellar motor", "Ion flux through stator complexes powers torque generation".

That is the same energy-availability-powering-a-machine shape the PR sets aside for molecular_oxygen → cytochrome bd and cytoplasmic_na → Mrp, with the same honest form already named in the PR body: the motor has input the gradient, or IMF enables the rotation process — which is range-safe and would need a process node. As it stands these four pass the audit because contributes to declares no range, not because the claim became true, and that is the failure mode the audit exists to catch. Either defer them with the other four or model them as inputs.

🔵 Descriptions were left phrased for enables ("organizes", "powers") on most migrated edges. The rendered page puts predicate and description on adjacent lines, so the mismatch is now visible to a reader.

Curation events are on all 8 records and ordered correctly. No record was added under history/records/ — advisory per history/README.md, not blocking, and the inline curation_history entries carry the reasoning.

I could not run just qc / audit-predicate-domains in this environment (Python execution is not permitted here), so the green-CI claims are unverified by me; the baseline arithmetic and the page regeneration I checked by hand and they are correct.

---

Requesting changes on the three 🟡 items.
· fix/334-enables-range-proteins

[culturebot-reviewer]

Three 🟡 items — see the detailed review comment.

1. biolink:encodes (methanogenesis.yaml:165) is not a slot in the pinned biolink-model 4.4.0; reports/biolink_coverage.tsv:28 already records it as unmatched. The PR's "checked, not assumed" claim is vacuous for a term absent from the model. Pre-existing, but the mapping-row note asserting an exact biolink match should be corrected or the edge re-grounded to RO:0002205.

2. mreB part of rod_complex (rod_shaped.yaml:108, bacillus_shaped.yaml:106) contradicts its own snippet ("MreB filaments interact with Rod complexes"), its own edge description ("organizes"), and the target node's description ("associated with MreB"). interacts with matches the quoted evidence and needs no modelling decision.

3. The ion-motive-force edges (motile.yaml:148, motility.yaml:121, flagellated.yaml:142, gliding.yaml:107) are the same energy-availability idiom the PR deliberately defers for molecular_oxygen and cytoplasmic_na. Biolink's contributes to is about contributing to the occurrence or generation of the other entity; IMF powers the motor's rotation, it does not help generate the motor. These pass the audit because contributes to declares no range, not because the claim got truer.

The per-idiom approach is right and the baseline/report/page regeneration is consistent (33 → 24 verified). The FliH-FliI-FliJ part-of and the virulence-gene-acquisition contributes-to edges are both good.

[realmarcin]

All three were right, and the first exposed a flaw in how I verified the other two.

🟡 biolink:encodes is not a slot — and my check couldn't have caught that

You're right, and the way I got it wrong matters: my check was

s = slots.get(name) or {}
print(s.get('domain'), s.get('range'))

so an absent term returns None for both, and I recorded it as "declares no rdfs:domain or rdfs:range — checked, not assumed". A term absent from the model trivially declares nothing. The check could not distinguish absent from unconstrained, which is exactly the distinction the claim rested on. That's a bad verification, not just a bad conclusion.

And as you note, reports/biolink_coverage.tsv already recorded it — the only applied biolink: CURIE with both backing columns empty. The signal existed and nothing read it.

Kept the grounding (pre-existing across 9 files) because a gene cluster → protein complex edge doesn't fit has gene product's range: gene product mixin — repointing would trade a missing term for a wrong range. What's fixed is the mapping row's note, which claimed an exact biolink label match that biolink_coverage.tsv contradicts. Filed as #342, including your suggestion to check whether any other applied biolink: CURIE lacks a backing slot.

🟡 mreB part of rod_complex

Over-claimed, and your evidence is decisive: the snippet says MreB filaments interact with Rod complexes, the target node is "associated with MreB", and the corpus is split — cell_shape.yaml types rod_complex a PATHWAY with MreB inside it. That's precisely the modelling decision this work defers elsewhere, so I shouldn't have made it in passing.

Both edges now use interacts with (biolink:interacts_with, already mapped, range-safe), which says what the quoted sentence says.

🟡 The motive-force edges

Also right, and the inconsistency is the tell: I deferred molecular_oxygen → cytochrome bd and cytoplasmic_na → Mrp for being energy/substrate-availability, then migrated four edges of the same shape. biolink's contributes to is about contributing to the occurrence or generation of the object; a motive force doesn't generate the motor, it powers rotation — as the descriptions say.

Reverted all four to enables and re-deferred them. They passed the audit only because contributes to declares no range, which is the failure mode the audit exists to catch.

🔵 Descriptions

Realigned the two mreB descriptions, which still read "organizes" from the enables phrasing.

On the events

Added correction events to all six records rather than editing the originals — a correction is its own event, and rewriting history would hide that the first pass was wrong.

Corrected result

ENABLES_RANGE_VIOLATION: 33 → 28, not the 24 first claimed. Five edges migrate: 1 encodes, 1 part of (the FliI ATPase, which you didn't challenge and I believe holds), 2 interacts with, 1 contributes to (acquired genes → repertoire, which does fit "generation of").

just qc green · validate-strict 477/0 · 434 tests pass · audit-graphs unchanged.

[culturebot-reviewer]

Claude finished @realmarcin's task in 3m 28s —— View job

---

Re-review

All three 🟡 items are fixed, and I re-ran the domain/range check properly this time rather than trusting either of our earlier claims:

predicate	in biolink 4.4.0	domain / range
part of	✓ line 4806	none declared
contributes to	✓ line 3853	none declared
interacts with	✓ line 2965	named thing / named thing — GENE_OR_PROTEIN satisfies both
encodes	✗ absent	—

• mreB → interacts with on both files, descriptions realigned to the snippet. Correct.
• The four motive-force edges → reverted to enables/RO:0002327 and re-baselined. The four new baseline rows match the four reverted edges exactly.
• Arithmetic: ENABLES_RANGE_VIOLATION is 28 in both conf/predicate_domain_audit_baseline.tsv and reports/predicate_domain_audit.tsv (byte-identical), down from 33 on main. Five migrations: encodes ×1, part of ×1, interacts with ×2, contributes to ×1. Pages are regenerated in sync — the only surviving contributes to renders in gliding.html and rod_shaped.html are pre-existing edges this PR never touched.

One item left.

---

🟡 The retracted verification is still asserted in the record it was wrong about.
data/traits/metabolism/methanogenesis.yaml:290

The correction events went to the six records whose edges changed. methanogenesis.yaml carries the biolink:encodes edge — the one the flawed check actually covered — and got no correction event, so its history still reads:

The replacements are chosen per idiom rather than swept: a gene cluster ENCODES its product … All three declare no rdfs:domain or rdfs:range, so none can reintroduce the class of defect being removed.

That sentence is now contradicted inside this same PR by mappings/predicate_grounding.tsv:28, which says biolink:encodes is a LOCAL COINAGE, not an upstream slot. The record points at a justification the corpus no longer holds. Since the stated principle here was that a correction is its own event, the record where the finding landed is the one that most needs it. human_pathogen.yaml:290 carries the same enumeration and is now partly stale too (the "energy source … machine it powers" half was retracted), though its own surviving edge is fine.

Fix this →

🔵 On predicate_grounding.tsv:28, the note now says local coinage but the machine-readable columns still say skos:exactMatch / source=biolink. scripts/ground_causal_predicates.py:126 reads source, so anything that breaks grounding down by source will keep counting this as biolink-derived. There is no local value in the column's current vocabulary (METPO/biolink/RO/RDFS), so introducing one is a bigger call than this PR — worth folding into #342 rather than doing here.

---

Keeping biolink:encodes rather than repointing to has gene product is the right call and well argued: range: gene product mixin would not accept a protein complex, so repointing trades a missing term for a wrong one. The FliI ATPase part of and the virulence-repertoire contributes to both hold. Curation events are ordered correctly on all eight records, and no snippet was altered.

I still can't execute Python here, so just qc / audit-predicate-domains remain unverified by me; the baseline and page consistency I checked by hand and both are correct.

--- · fix/334-enables-range-proteins

[culturebot-reviewer]

All three earlier items are fixed and verified: mreB moves to interacts with with realigned descriptions, the four motive-force edges are reverted to enables and re-baselined, and the biolink:encodes mapping note now states the coinage plainly. Re-ran the domain/range check properly — part of (4806) and contributes to (3853) declare none, interacts with (2965) is named thing/named thing, encodes is absent. Baseline and report are byte-identical at 28 ENABLES_RANGE_VIOLATION rows, down from 33; pages are in sync.

One 🟡 left: methanogenesis.yaml carries the biolink:encodes edge but got no correction event, so its curation history still asserts 'All three declare no rdfs:domain or rdfs:range' — contradicted by predicate_grounding.tsv:28 in this same PR. human_pathogen.yaml has the same now-partly-stale enumeration. The six records whose edges changed all got correction events; the record the finding was actually about did not. See the detailed comment.

[realmarcin]

Right, and the place I stopped was the wrong one.

🟡 The record the finding actually landed in

I applied "a correction is its own event" to the six records whose edges changed, and skipped methanogenesis.yaml — which carries the biolink:encodes edge, i.e. the one the vacuous check actually covered. So its history still asserted the retracted justification, contradicted inside this same PR by predicate_grounding.tsv calling that term a local coinage.

Correction event added there, recording that the justification was vacuous because the term is absent from the pinned model, that the grounding is retained because a gene cluster → protein complex edge doesn't fit has gene product's range, and that the edge is still an improvement on enables whose declared range it genuinely violated.

human_pathogen.yaml also got one — you're right that its enumeration is stale on a second count too, since the "energy source … the machine it powers" half was retracted with the motive-force reverts. Its own surviving edge is unaffected, and the event says so explicitly.

🔵 source=biolink / skos:exactMatch

Agreed on both the substance and the scoping. The prose and the machine-readable columns disagree, and ground_causal_predicates.py reads source — but introducing a local value to a vocabulary of METPO/biolink/RO/RDFS affects every consumer of that column, which is more than this PR should decide.

Folded into #342, along with your other suggestion: sweep biolink_coverage.tsv for any further applied biolink: CURIE with both backing columns empty, since nothing currently reads that report and the same gap could reopen silently.

just qc green · validate-strict 477/0 · 434 tests pass.

[culturebot-reviewer]

Claude finished @realmarcin's task in 2m 13s —— View job

---

Re-review (round 3) — approved

• Verify the correction events on methanogenesis.yaml / human_pathogen.yaml
• Re-check baseline / report / page consistency
• Check nothing regressed elsewhere

The last 🟡 is fixed. methanogenesis.yaml:295 — the record carrying the biolink:encodes edge, the one the vacuous check actually covered — now has a correction event stating the justification was vacuous because the term is absent from the pinned model, why the grounding is retained anyway (has gene product's range: gene product mixin would not accept a protein complex), and that it is tracked in #342. human_pathogen.yaml:299 got one covering both stale halves of its enumeration and noting its own surviving edge is unaffected. Originals untouched; both events append at 09:30:00Z, after the 08:00 and 09:00 events, so ordering holds on all 8 records.

Re-checked the rest rather than trusting the earlier rounds:

check	result
ENABLES_RANGE_VIOLATION rows	28 in baseline and report, files byte-identical
predicate deltas across data/traits	5× enables/RO:0002327 out; encodes ×1, part of ×1, interacts with ×2, contributes to ×1 in
snippet lines touched	0
page stamps	the two regenerated pages carry the new events and moved their own per-record stamp; the six untouched trait pages did not move
CI	qc, validate-strict, pr-sanity, curation-history, label-correspondence all ✓

🔵 Residual, not worth another round: all 8 records carry the same boilerplate three-idiom enumeration, and only these two now carry the encodes retraction — the other six correct only their own edges. The substantive correction is recorded where the edge lives, plus in predicate_grounding.tsv and #342, so spraying six more identical events would add noise rather than accuracy.

Scoping source=biolink / skos:exactMatch to #342 is the right call — a local value in that column's vocabulary affects every consumer and is not this PR's decision to make.

Approved.
· fix/334-enables-range-proteins

[culturebot-reviewer]

The last 🟡 is fixed. methanogenesis.yaml:295 — the record carrying the biolink:encodes edge, the one the vacuous check actually covered — now has a correction event stating the justification was vacuous because the term is absent from the pinned model, why the grounding is retained anyway (has gene product's range: gene product mixin would not accept a protein complex), and that it is tracked in #342. human_pathogen.yaml:299 got one covering both stale halves of its enumeration and noting its own surviving edge is unaffected. Originals untouched; both events append at 09:30:00Z, after the 08:00 and 09:00 events, so ordering holds on all 8 records.

Re-checked the rest rather than trusting the earlier rounds: baseline and report are byte-identical at 28 ENABLES_RANGE_VIOLATION rows; the five migrations are exactly 5× enables/RO:0002327 out and encodes ×1, part of ×1, interacts with ×2, contributes to ×1 in; zero snippet lines changed anywhere in the diff; the two regenerated pages carry the new events and their own per-record stamps moved while the six untouched trait pages did not. CI is green on qc, validate-strict, pr-sanity, curation-history and label-correspondence.

🔵 Residual, not worth another round: all 8 records carry the same boilerplate three-idiom enumeration, and only these two now carry the encodes retraction — the other six correct only their own edges. The substantive correction is recorded where the edge lives, plus in predicate_grounding.tsv and #342, so spraying six more identical events would add noise rather than accuracy.

Scoping source=biolink/skos:exactMatch to #342 is right — a local value in that column's vocabulary affects every consumer and is not this PR's call. Approving.