Model the acetate kinase step instead of a chemical producing ATP (#331)

[realmarcin]

acetyl_phosphate --produces--> atp had a CHEMICAL subject on METPO:2007800, whose
definition is that the activity or state of the subject brings the object into
existence. A chemical has neither in that sense.

RO:0001001 (derives into) didn't fit either, and #327 already said why: ATP's matter
comes mostly from ADP and only the phosphoryl group transfers, so it fails that
relation's inherits-the-significant-portion-of-the-matter test. That's exactly why it was
filed separately rather than swept into #327.

The answer was already in the record

The edge's own description says "The acetate kinase step generates ATP", and its evidence
says "ACK convert acetyl-phosphate to acetate, yielding ATP via SLP". The process was
named but never modelled.

Added it as a BIOLOGICAL_PROCESS node and split the edge:

acetate_kinase_step --has input--> acetyl_phosphate	RO:0002233
acetate_kinase_step --has output--> atp	RO:0002234

Both range-correct, and this is option 1 from the issue — the one that lets the gate tighten.

Wired on both sides deliberately: atp had exactly one edge, so replacing it without an
input edge would have orphaned the new node and atp together. audit-graphs confirms
no new findings (1535, unchanged).

The gate now matches its definition

This was the last CHEMICAL-subject produces edge, so METPO:2007800's
subject_types drops CHEMICAL — the tightening #331 predicted. Also added a has input
mapping row (RO:0002233), gated to the three activity node types for the same reason
has output is.

The one honest cost

Evidence-snippet baseline moves 2737 → 2738, and the delta is exactly explainable:

• the pre-existing MISSING_SNIPPET re-keys onto the new has-output edge, and
• one genuinely new finding, because the has-input edge reuses the same evidence, which
  carries notes but no verbatim snippet.

A snippet cannot be invented — the schema specifies a verbatim quote and the source text
isn't to hand — so it joins the known backlog (2586 such findings) rather than being
fabricated.

What was checked

just qc green · just validate-strict 477 / 0 errors · 431 tests pass ·
audit-graphs unchanged · zero CHEMICAL-subject produces edges remain · curation event
added with history verified in order.

Closes #331.

🤖 Generated with Claude Code

[culturebot-reviewer]

Claude finished @realmarcin's task in 6m 1s —— View job

---

Review

The modelling call is right and the reasoning in the PR body holds up. Two things I checked that the body asserts, both confirmed independently:

• Reachability is safe. _reachable() in scripts/audit_causal_graphs.py:95 is undirected, so acetate_kinase_step — acetyl_phosphate — acetyl_coa — wood_ljungdahl_pathway — acetogenesis_trait keeps both the new node and atp anchored. No new UNREACHABLE_FROM_TRAIT.
• The gate claim is true. I enumerated the subjects of all 84 produces edges: every one is GENE_OR_PROTEIN, TRAIT, STATE or BIOLOGICAL_PROCESS. Dropping CHEMICAL from subject_types orphans nothing.

RO:0002233 / RO:0002234 are correct, the new mapping row is well-formed and mirrors the has output row exactly, the snippet-baseline delta (+2 rows, −1) matches the claimed 2737→2738, and no snippet was invented — the new edge carries notes only, which is the honest thing to do.

I could not run just qc or the test suite (uv run is blocked in this sandbox), so the green-CI claims are unverified by me.

---

🟡 data/traits/upper/biological_process.yaml:123 — the new mapping row covers this edge, but it was left ungrounded.

That edge is molecular_function --has input--> chemical_input with no predicate_id. Its sibling three lines below carries predicate_id: RO:0002234, and its own evidence note literally reads Has input [RO:0002233]. Before this PR the label was unmapped so the gap was inevitable; now it isn't — MOLECULAR_FUNCTION → CHEMICAL passes the new row's gate, which is exactly why the row disappeared from reports/predicate_grounding_residual.tsv.

Two consequences: the mapping table and the corpus disagree, and the next just ground-predicates --apply will silently modify a file this PR didn't touch. The precedent set by #327 is the other way — the derives into row landed with predicate_id: RO:0001001 written onto all four covered edges (disproportionation.yaml:175,184, cellulolysis.yaml:195, starch_degradation.yaml:90). Add predicate_id: RO:0002233 to that edge. Fix this →

🔵 acetate_kinase_step is label-only. mappings/node_grounding.tsv:59 already carries acetate kinase → GO:0008776 (acetate kinase activity), and typing the node MOLECULAR_FUNCTION instead would both ground it and match the GO-CAM shape data/traits/upper/biological_process.yaml uses for exactly this has input/has output pair. Against that: this trait's own research report (rendered at pages/traits/metabolism/acetogenesis.html:493) says to "prefer a process label plus explicit reaction edges" for the SLP step. Your call — I'd only ask that it be a decision rather than an omission.

🔵 The step has has output atp but not has output acetate, while acetyl_phosphate --is converted to--> acetate still runs alongside it. Half the reaction is on the process node and half on the chemical chain. Not wrong, just asymmetric.

No history/records/acetogenesis/… record — advisory only, and consistent with the two other curation events added to this file today, which have none either.

---

Verdict: requesting changes for the one 🟡. Everything else here is sound; that edge is a one-line addition.

• Read repo conventions (README.md, docs/CURATION_PLAYBOOK.md, docs/GROUNDING_POLICY.md)
• Review the substantive diff
• Verify ontology grounding (RO:0002233/RO:0002234, GO:0008776, METPO:2007800 gate)
• Check causal-graph structure & reachability
• Check evidence integrity, snippet baseline, curation history
• Post verdict
  · branch fix/331-acetyl-phosphate-atp

[culturebot-reviewer]

The modelling call is right and the reasoning holds up. Verified independently: reachability in audit_causal_graphs.py:95 is undirected so nothing is orphaned; all 84 produces edges have GENE_OR_PROTEIN/TRAIT/STATE/BIOLOGICAL_PROCESS subjects, so dropping CHEMICAL from the gate is safe; the RO CURIEs are correct and the snippet-baseline delta matches.

One blocker: the new has input mapping row now covers data/traits/upper/biological_process.yaml:123, but that edge was left without predicate_id: RO:0002233 while its sibling three lines below carries RO:0002234. The mapping table and the corpus disagree, and the next just ground-predicates --apply will silently touch a file this PR didn't. #327 set the opposite precedent — the derives into row landed with RO:0001001 written onto all four covered edges.

Details and two optional suggestions in the review comment.

[realmarcin]

Fixed, and you're right that #327 set the opposite precedent — I should have followed it.

🟡 The ungrounded has input edge

This is the same failure #328 had: adding a mapping row made an existing edge look mapped, so it dropped out of reports/predicate_grounding_residual.tsv, but nothing grounded it — leaving it in neither the grounded set nor the backlog, with the next ground-predicates --apply silently touching a file this PR hadn't.

Grounded to RO:0002233, which its own evidence note already named ("Has input [RO:0002233]") and which its sibling has-output edge has carried all along. Curation event appended by the tool.

That's twice now that adding a mapping row has stranded a pre-existing edge — worth remembering as the standard follow-up step when a row lands, which is exactly what the #327 precedent encodes.

🔵 Node type — decided, not omitted

Fair ask. Kept as BIOLOGICAL_PROCESS, and the reason is now in the curation event: MOLECULAR_FUNCTION would indeed ground to GO:0008776 via node_grounding.tsv, but this trait's own research report asks for "a process label plus explicit reaction edges" for the SLP step — which is precisely what this is. Your pointer to that line is what settled it; I'd otherwise have taken the grounding.

🔵 Half-reaction asymmetry — also decided

Kept, with the reasoning recorded: the existing acetyl_phosphate --is converted to--> acetate edge already states that half of the reaction on the chemical chain. Adding has output acetate would assert the same fact twice at two levels of description — which is the specific thing #303 criticised in dissimilatory_iron_reduction. So the asymmetry is the lesser of the two.

history/records/

Still none, consistent with the other events on this file; the convention gap is #325.

Re-verified

just qc green · just validate-strict 477 / 0 · 431 tests pass · snippet baseline steady at 2738 with 0 new · audit-graphs unchanged.

[culturebot-reviewer]

Claude finished @realmarcin's task in 1m 38s —— View job

---

Re-review (commit 64716512)

The 🟡 is fixed correctly and completely.

• data/traits/upper/biological_process.yaml:126 now carries predicate_id: RO:0002233, matching its sibling's RO:0002234 three lines down and the CURIE its own evidence note already named. RO:0002233 is has input — correct.
• It's the only other has input edge in the corpus (the other is the new acetate_kinase_step → acetyl_phosphate, grounded at data/traits/metabolism/acetogenesis.yaml:226), so the new mapping row now has zero uncovered edges. The mapping table and corpus agree again, and ground-predicates --apply has nothing left to silently touch.
• reports/predicate_grounding_residual.tsv drops the has input 1 unmapped line — the edge left the backlog by being grounded, which is the state that was missing before.

The derived artifacts are consistent with each other, not just individually plausible:

• reports/node_grounding_residual.tsv gains acetate kinase step BIOLOGICAL_PROCESS — the honest consequence of keeping the node label-only rather than typing it MOLECULAR_FUNCTION. It's in the backlog rather than quietly absent.
• conf/evidence_snippet_baseline.tsv is +2/−1 exactly as claimed: acetyl_phosphate->atp re-keys onto acetate_kinase_step->atp, plus the one new acetate_kinase_step->acetyl_phosphate row. Still no invented snippet — notes only.
• biological_process.yaml curation history is chronological through the new 2026-08-06T09:03:55Z event.

Both 🔵s are now decisions with reasons recorded in the curation event rather than omissions, which is what I asked for. The BIOLOGICAL_PROCESS call cites the trait's own research report; the deliberate absence of has output acetate cites #303's duplication-across-levels criticism. I don't agree or disagree — they're defensible and, more importantly, they're now legible to whoever reads this file next.

One thing to note for accuracy: the acetogenesis curation event was amended in place rather than appended. That's fine here — it's the event this same unmerged PR created — but it does mean the recorded reasoning and the change landed in different commits.

As before, I could not run just qc, validate-strict, or the test suite (uv run is blocked in this sandbox), so the green-CI claims remain unverified by me. Everything I could check statically holds.

Verdict: approving.

• Verify the 🟡 fix (RO:0002233 on biological_process.yaml:126)
• Confirm no other edges the new mapping row leaves stranded
• Check derived reports & baselines regenerated consistently
• Check curation event records the two 🔵 decisions
• Post verdict
  · branch fix/331-acetyl-phosphate-atp

[culturebot-reviewer]

The 🟡 is fixed correctly and completely.

data/traits/upper/biological_process.yaml:126 now carries predicate_id: RO:0002233, matching its sibling's RO:0002234 and the CURIE its own evidence note already named. It's the only other has input edge in the corpus, so the new mapping row has zero uncovered edges — the mapping table and corpus agree again, and ground-predicates --apply has nothing left to silently touch. The residual report drops the has input line by grounding rather than by disappearing.

Derived artifacts are consistent with each other: node_grounding_residual.tsv gains acetate kinase step (the honest cost of the label-only node, in the backlog rather than quietly absent), and the snippet baseline is +2/−1 exactly as claimed with no invented snippet.

Both 🔵s are now decisions with reasons in the curation event rather than omissions. I don't need to agree with either call — they're defensible and legible to the next reader, which was the ask.

Note for accuracy: the acetogenesis curation event was amended in place rather than appended. Fine here since it's this PR's own event, but the reasoning and the change landed in different commits.

I still could not run just qc, validate-strict, or the tests (uv run is blocked in this sandbox), so the green-CI claims are unverified by me. Everything checkable statically holds.