Repository navigation
Add VP1840 system trait - #1426
Conversation
There was a problem hiding this comment.
Structurally this is a clean copy of the DS-24 pattern: identity (traitmech:000450 is free corpus-wide, parent traitmech:000209 is phage defense system), both predicates resolve against mappings/predicate_grounding.tsv with * domains (RO:0002326 contributes to, METPO:2007700 confers), the 4-node graph is a chain with every node reachable from the trait node, attaches_to matches the declared graph_id, the v326→v327 / METPO:1040300→METPO:1040400 / 000449→000450 continuity claims in the proposal all check out, and the history/ record is present. Two things need fixing.
Caveat on what I could check: this review environment has no network and no Python, so I could not open the two Springer .xlsx files, run just qc, or run just verify-snippets. The DefenseFinder table rows are internally consistent in column padding with the rows other records quote from the same pinned files, which is a decent sign they were really copied. The findings below are ones I could establish without the sources.
🟡 The two Supplementary Table 9 notes: lines cannot both be right, and they carry the whole experimental basis of the record.
snippet: "VP1840\tBas21\t2e-09" notes: ... a 2e-09 fold-change phage-plating readout ...
snippet: "VP1840\tLL1\t0" notes: ... a zero fold-change phage-plating readout ...
These two rows are the only quantitative support for reduced vibriophage plaquing — the definition's "experimentally linked to reduced vibriophage plaquing", the vp1840_locus --contributes to--> reduced_vibriophage_plaquing edge, and the confers edge all rest on them. Under the "fold change" reading the notes assert, 2e-09 is a ~billion-fold reduction and 0 is no change at all, so one of the two notes is describing the column wrongly. If the column is actually EOP, then 0 means complete block and "zero fold-change" says the opposite of what the datum means; if it really is a fold reduction, then 2e-09 does not mean what the note says either. Neither snippet quotes a header or enough columns for a reader to tell which, so nothing in the record resolves it.
Please open the sheet, name the readout column, quote the row with enough columns (or the header row) to identify it, and correct whichever note is wrong. The playbook's #621 precedent is exactly this: a notes: line asserting something the source does not support. Compare ds_24_system.yaml, which quotes the full 18-column Table S7 row and names the statistic (-log(EOP) value of 7.477).
🟡 README.md corpus counts were not updated, and now contradict pages/ in this same PR.
pages/browse.html and pages/index.html correctly regenerate to 845 records / GENOMICS 259, while README.md still says 844 / 258. Verified against the tree: data/traits holds 845 records, data/traits/genomics 259, of which 240 are PROPOSED and 259 carry causal_graphs. So README needs:
- L23 GENOMICS row: PROPOSED 239→240, causal_graphs 258→259, Total 258→259
- L29 TOTAL row: PROPOSED 329→330, causal_graphs 600→601, Total 844→845
- L33 prose: "239 genomics records"→240
- L37 prose: "599"→"600" is already stale as 600; should be 601
- L95 tree comment: "843 curated TraitRecords" reads 844; should be 845
Every one of the last eight trait-addition commits (#1412 through #1425) updated this table, so this is the established convention, not an optional extra. There is no automated check on it, which is precisely why it needs to be in the PR.
🔵 Optional: the Supplementary Data 1 snippet "Vibrio parahaemolyticus\t1\t3e-121\t100\tWP_229652597.1" contains no VP1840 identifier, so the quote alone cannot be tied to VP1840 — a PSI-BLAST homolog table for nine systems could put that row under any of them. A couple more columns, or the row's system key, would make it self-identifying. Also worth a sentence in the note on why the homolog accession (WP_229652597.1) differs from the cassette protein in Table 8 (WP_005483293.1) despite "100%".
The scope discipline elsewhere is good — scope_status: NONMECHANISTIC, the explicit "no pinned DefenseFinder rule row" in scope_notes, the empty canonical_examples with a logged reason, and the KNOWLEDGE_GAP discussion enumerating what is unresolved are all the right calls for a single-cassette system with this little mechanism known.
There was a problem hiding this comment.
Local checks all pass. traitmech:000450 is the next free ID and is unique across the corpus; parent traitmech:000209 resolves to phage_defense_system.yaml; RO:0002326, METPO:2007700 and rdfs:subClassOf are all in mappings/predicate_grounding.tsv, and the confers edge respects its declared BIOLOGICAL_PROCESS -> TRAIT domain/range; all four nodes are connected and reachable from the TRAIT node; METPO:1040400 follows v326 METPO:1040300 and METPO:1016300 is the same placeholder parent 195 sibling cohorts use; the README counts match the corpus (845 records, 259 genomics of which 240 PROPOSED, 330 PROPOSED overall, 601 graphs). History record present. The shape follows the DS-24 precedent closely.
This review environment has no network access, so I could not fetch the DOI or the pinned DefenseFinder files. Whether the snippets are verbatim rests on your verification, not mine. What I can review is what the notes: claim about them, and two of them assert things the quoted rows do not contain.
🟡 The two Supplementary Table 9 notes name a column the rows do not carry.
snippet: "VP1840\tBas21\t2e-09" notes: ... a 2e-09 fold-change phage-plating readout ...
snippet: "VP1840\tLL1\t0" notes: ... a zero fold-change phage-plating readout ...
Neither row includes a header, so "fold-change" is inference. The pairing 2e-09 / 0 is at least as consistent with a p-value column (0 = underflow) as with a fold change -- a literal fold change of 2e-09 is a ~5x10^8-fold reduction. Worse, read as written, "a zero fold-change readout" states no change in plaquing, which is the opposite of the edge it is attached to. These two cells are the only quantitative evidence anywhere in the record; they also carry the entire confers edge, reused verbatim from the contributes to edge, and the definition claim "experimentally linked to reduced vibriophage plaquing" rests on them. Please quote the header row alongside them or name the column in the notes, so the direction is not left to the reader to guess.
🟡 Same problem in the Supplementary Data 1 note. "Vibrio parahaemolyticus\t1\t3e-121\t100\tWP_229652597.1" -> "a full-length 100% PSI-BLAST hit". "full-length" appears nowhere in the row. And if 100 is percent identity over the full length, the hit should not be a different accession from Table 8 WP_005483293.1 -- RefSeq WP_ records are non-redundant by sequence, so an identical protein would share one. Either 100 is coverage rather than identity, or the alignment is not full-length; say which.
The rest is the right call for a single-paper cloned-cassette result: NONMECHANISTIC scoping, scope_notes that state plainly what is not asserted, the open knowledge gap for native host breadth and the absent rules-table row, and empty canonical_examples with a recorded reason rather than a guessed exemplar.
There was a problem hiding this comment.
Grounding, graph structure and identity all check out. RO:0002326 is "contributes to", METPO:2007700 is the corpus's confers placeholder (and mappings/predicate_grounding.tsv already admits BIOLOGICAL_PROCESS→TRAIT), traitmech:000209 really is phage defense system, traitmech:000450 is the next free id with no duplicate, and METPO:1016300/METPO:1040400 are the right parent and next cohort block. The graph has no dangling edges and all four nodes are reachable from the TRAIT anchors. node_id: phage_defense_system is typed TRAIT consistently with the other 200+ system records. README/data.js counts match the tree (845 records, 601 graphs, 259 genomics).
🟡 app/dashboard/priority.{json,html} was not regenerated
app/dashboard/priority.json:3 still reads "records": 844, and the new record is absent from the queue. Every one of the six preceding sibling PRs refreshed it in the same commit (e735270b1a DS-24, fd1d53cc11 DS-23, 983748fed9 DS-22, …); this branch's file set is identical to DS-24's except for these two. Nothing in just qc gates on them — audit-derived-reports covers reports/ and pages/ but not app/dashboard/ — so CI will stay green while the dashboard README links at line 121 silently goes stale.
Fix: just gen-priority-dashboard and commit both files.
🔵 BAC60103.1 is asserted in a note with nothing quoted behind it
data/traits/genomics/vp1840_system.yaml:49 says the PSI-BLAST hit came "from the BAC60103.1 query", but that accession appears in no snippet and nowhere else in the record, and it is a different accession from the WP_005483293.1 the record attributes to the cassette elsewhere. The chain is probably fine (old locus tag → old protein accession → RefSeq), but a reader cannot check it from the record. If Supplementary Data 1's VP1840 worksheet names the query in a cell, quoting that cell would close the loop; otherwise drop the accession from the note.
What I could not check
No network access in this environment, so I could not fetch the two Springer .xlsx files or the DOI text to confirm the snippets are verbatim. What I could check internally is consistent with verbatim copying: the two pinned DefenseFinder markdown rows have column padding byte-identical in width to the DS-23/DS-24 rows quoted from the same fixed-width tables (VP1840__VP1840 + 35 spaces = the same 49-char field as DS-24__DS-24 + 37), and the trailing space after WP_005483293.1 in the Table 8 row is the signature of a raw cell copy rather than a retype. Worth a second pair of eyes on the Table 9 Fold Change rows specifically, since the two snippets (2e-09, 0) are the only support for the reduced-plaquing edge and the notes deliberately do not interpret the column's direction.
There was a problem hiding this comment.
Structurally this is a clean instance of the established DefenseFinder-system pattern, and I checked that part carefully:
traitmech:000450is unique in the corpus and sequential after000449(DS-24);parent_traits: traitmech:000209resolves todata/traits/genomics/phage_defense_system.yaml.- All three predicates match
mappings/predicate_grounding.tsvon both CURIE and label —contributes to/RO:0002326,confers/METPO:2007700,is a/rdfs:subClassOf— and theBIOLOGICAL_PROCESS → TRAITuse ofconfersis within its declaredsubject_types/object_types. - The graph is one component: all four nodes are wired, every edge references a declared node, and the trait node is reachable. Nothing new lands in
causal_graph_audit.tsvorcausal_graph_connectivity.tsv, correctly. - README counts verified live, not taken from the diff: 845 record files, 601 with
causal_graphs, 330PROPOSED, 259 genomics — all match, and the narrative sentence still sums to 330/477/845. - History record present,
attaches_toanchor matches the graph id,app/discussions/data.jsmetrics increment consistently, andscripts/add_vp1840_system_trait.pydefaults to dry-run and goes throughwrite_validated_trait/record_curation_event.
Two things need to change before merge.
🟡 The "vibriophage" claim is not carried by any quoted snippet, and there is a concrete reason to doubt it.
The record is otherwise scrupulous about host: the definition says only "expressed from a VSV105 plasmid", the node description says "in cells expressing cloned vp1840", and scope_notes explicitly declines to "assert native host breadth". But "vibriophage" is then asserted four times — in definition, in the graph title, in the reduced_vibriophage_plaquing node label, and in its description ("Reduced plaquing of Vibrio-infecting bacteriophages") — plus in the discussion prompt.
Nothing quoted supports it. The only phage names in evidence are Bas21 and LL1, and the notes say the column header is just "Phage". Bas21 matches the BASEL E. coli collection naming that this same corpus already cites for E. coli MG1655 assays — ds_24_system.yaml (Bas1), ds_7_system.yaml (Bas60), ds_14_system.yaml (Bas50) — and those records deliberately write "bacteriophage Bas50", with no host-clade adjective. pVSV105 doesn't settle it either: its pES213 replicon functions in E. coli as well as Vibrio, so the vector choice is not evidence of a Vibrio plating host. Note also that gaps4_system.yaml in this corpus already handles this exact trap correctly, separating a V. parahaemolyticus origin locus from an E. coli expression host.
Please confirm from Supplementary Table 9 what the plating host was and what Bas21/LL1 actually are. If the assay host was E. coli, or if the table doesn't say, drop to "bacteriophage" everywhere. This matters more than usual because the same sentence is copied verbatim into the definition column of proposals/metpo_traitmech_v327/metpo_proposal_classes_robot.tsv, so a wrong host clade ships upstream to METPO with the class.
🟡 The Table 9 snippets can't establish the direction of the effect they are the sole support for.
"VP1840\tBas21\t2e-09" and "VP1840\tLL1\t0" are bare three-field fragments. The entire causal claim — reduced plaquing — rests on the notes: line asserting that column 3 is "Fold Change" and that a small value means protection. 0 isn't a well-formed fold change under the opposite reading, which is what makes the intended direction guessable rather than evidenced, and these two rows are the only quantitative support in the record.
This record already solved the same problem once: CLARIFIED_EVIDENCE_SNIPPETS and GROUNDED_PSIBLAST_QUERY_ACCESSION fixed the PSI-BLAST item by quoting the header row above the data row in the snippet itself. Do the same here instead of leaving it in notes:.
One disclosure on scope. This environment has no network access, so I could not open the Springer Nature supplementary workbooks (MOESM2/MOESM3) or the pinned DefenseFinder tables at afb0e5a8. I verified the two DefenseFinder snippets are formatted consistently with the dozens of existing quotes from those same pinned files, but I have not confirmed any snippet is verbatim against its source, including the main-text sentence. just verify-snippets --record data/traits/genomics/vp1840_system.yaml is not in the validation list in the PR description; per the playbook's #623 section it is worth running, with the caveat that it only catches edits to a quote, not a rewrite of one.
There was a problem hiding this comment.
Two blocking issues. The grounding and graph structure are clean.
🔴 Two tracked derived reports were not regenerated after the vibriophage → bacteriophage rename
The last commit (18e2f0cc40, "Clarify VP1840 phage plating scope") renamed the graph and its process node, but two derived TSVs still carry the old strings:
reports/node_grounding_residual.tsv:1718—reduced vibriophage plaquing, whereas the node label indata/traits/genomics/vp1840_system.yaml:86is nowreduced bacteriophage plaquingreports/graph_protein_taxon_coverage.tsv:407—vp1840_locus_reduces_vibriophage_plaquing, whereasgraph_idis nowvp1840_locus_reduces_bacteriophage_plaquing
node_grounding_residual.tsv is byte-compared against scripts/ground_causal_nodes.py output by audit-derived-reports, which is in qc (justfile:701, justfile:1128), so just qc fails as committed. The PR body's "temp regeneration of tracked derived TSVs and pages matched the working tree" was true before that last commit; it needs to be redone after it. app/discussions/data.js and the rendered pages did get refreshed, so it is only these two.
🔴 The Supplementary Table 9 snippets are header rows stitched to non-adjacent data rows
data/traits/genomics/vp1840_system.yaml:39 and :43 (repeated on the edges at :119/:123, :149/:153, and in the discussion at :203):
"Strain\tPhage\tFold Change\tNotes\nVP1840\tBas21\t2e-09"
"Strain\tPhage\tFold Change\tNotes\nVP1840\tLL1\t0"
Both present the table's header line immediately followed by a data row. The Bas21 and LL1 rows are different rows, so at most one of them can actually be adjacent to the header — at least one of these two is a non-contiguous stitch presented as a single verbatim quote. docs/CURATION_PLAYBOOK.md:649 and :673 require a contiguous span, and curation_history at :265 records the header as having been prepended deliberately ("Quoted the Supplementary Table 9 Fold Change header in the Bas21 and LL1 snippets").
The same snippets also drop a cell: the quoted header declares four columns including Notes, and each data row supplies three. That is an unmarked elision of the Notes value — the docs/CURATION_PLAYBOOK.md:699 (#620) defect class, and it matters here specifically because a Fold Change of 0 for LL1 is the kind of value whose Notes cell would carry the qualification.
Please split these into a snippet that is one contiguous span (quote the row, and put the column meaning in notes:, which has no verbatim requirement), or quote header and row as two separate evidence items. :51/:131/:211 (Scientific Name\t…\nVibrio parahaemolyticus\t1\t3e-121\t100\tWP_229652597.1) has the same construction; it is plausibly the top PSI-BLAST hit and therefore genuinely adjacent, but it should be confirmed against Supplementary Data 1 rather than assumed.
What I checked and found clean
- Groundings.
RO:0002326= contributes to (mappings/predicate_grounding.tsv:10, exact label match against RO);METPO:2007700= confers (:102), and BIOLOGICAL_PROCESS subject → TRAIT object is inside its declared domain;rdfs:subClassOffine.METPO:1016300is the same unminted phage-defense-system placeholder prior cohorts use. The record grounds only totraitmech:IDs otherwise, so there is no external CURIE to mismatch. - Graph structure. Four declared nodes, three edges, every edge endpoint declared, single connected chain through the trait node. No orphans.
- Identifiers.
traitmech:000450is unused elsewhere and next after000449;METPO:1040400collides with nothing underproposals/. - README counts. Verified live: 845 records, 601 with causal graphs, 240 PROPOSED genomics.
- History record present at
history/records/vp1840_system/, and the mutation script follows the--apply/write_validated_trait/record_curation_eventcontract. - The two pinned DefenseFinder snippets match the fixed-width column layout and the escaped-DOI convention of sibling records (
ds_21_system.yaml:68,old_exonuclease_system.yaml:77) exactly, which is good evidence they were genuinely copied. I had no network access in this review, so I could not open the DOI, the two.xlsxfiles, orDefenseFinder_rules.tsvdirectly; the snippet finding above rests on the internal contradiction between the two Table 9 quotes, not on having read the table.
There was a problem hiding this comment.
Identity and structure check out: traitmech:000450 is free and follows 000449, the parent traitmech:000209 is phage defense system, RO:0002326/METPO:2007700 match their labels in mappings/predicate_grounding.tsv (lines 10 and 102), METPO:1040400 is unused and METPO:1016300 as the proposal parent matches v130/v141/v326, all four graph nodes are declared and the trait node is reachable by a directed path, and the README/report/dashboard counts match a live count of 845 records / 259 genomics. The history record and the --apply + write_validated_trait + record_curation_event contract are all in place.
One defect.
🟡 The Supplementary Table 9 snippets are stitched, not contiguous quotes — data/traits/genomics/vp1840_system.yaml:39, :43 (repeated at :119, :123, :149, :153, :203).
"Strain\tPhage\tFold Change\tNotes\nVP1840\tBas21\t2e-09"
"Strain\tPhage\tFold Change\tNotes\nVP1840\tLL1\t0"
Both present the table header row as immediately preceding a data row. They cannot both be true — the Bas21 row and the LL1 row are not the same row, so at most one of these is a contiguous span, and in practice neither is. docs/CURATION_PLAYBOOK.md:649 and :673 require a contiguous span copied from the source with no stitching, and the playbook severity table rates the stitched case ERROR. audit-snippets passes only because its detector keys on a literal .../…, so the pipeline will not catch this.
Secondary evidence of the splice: the header carries four columns (Strain, Phage, Fold Change, Notes) but each data row carries three, so the quoted row is also silently truncated relative to the header it is glued to.
data/traits/genomics/vp1840_system.yaml:51 / :131 / :211 has the same shape for Supplementary Data 1 (Scientific Name…Accession header plus the WP_229652597.1 hit row). I cannot disprove that one from the file alone, but it came out of the same pass and should be checked against the sheet.
The fix is the pattern the sibling record already uses: data/traits/genomics/ds_24_system.yaml quotes each supplementary row on its own, unstitched, and names the column in notes: — which is where the playbook puts that reasoning. The notes: lines here already do the job ("identifies the VP1840/Bas21 phage-plating readout as a Fold Change value of 2e-09"), so dropping the header prefix from snippet: loses nothing.
🔵 On the confers edge (:153), Fold Change 0 for LL1 is doing evidentiary work for a reduction claim, and a bare 0 cannot distinguish "no plaques recovered" from "no change measured". The Bas21 2e-09 carries the edge on its own; consider whether the LL1 row belongs there without an interpretation you can source.
|
Why this PR: it was the assigned target, and it is the only open PR in a non-converging review loop — 6 commits and 6 State. Head The green is misleading, and that is the part worth human attention. The one blocker that survives. Two independent review rounds on the same head commit (03:39, 03:43) converge on it: the Supplementary Table 9 snippets at One header line cannot be immediately adjacent to two different data rows, so at least one of these is a splice presented as verbatim. I verified the part that decides whether CI will ever catch this, and it will not: Scope note: I could not read the record at the PR head from this environment, so the two snippet strings above are quoted from the reviews rather than checked against the file. They appear identically in two independent rounds. Most useful next action. Someone with the Springer MOESM workbook open settles the Table 9 snippet form in one edit: quote the bare data row and put the column meaning in This will not come from another autonomous round. The reviewer states it has no network access to the Already fixed, do not re-chase: the 03:39 review first finding (stale |
There was a problem hiding this comment.
Approving — I could not find a defect worth blocking on.
What I checked against the live tree rather than the description:
- Identity.
traitmech:000450is unique acrossdata/traits/and is the next ID after000449;parent_traits: traitmech:000209resolves todata/traits/genomics/phage_defense_system.yaml, labelphage defense system, which matches thephage_defense_systemnode's grounding and therdfs:subClassOfedge. - Predicate grounding. Both
RO:0002326(contributes to) andMETPO:2007700(confers) are reviewed rows inmappings/predicate_grounding.tsv, and theconfersedge respects that row's gating —BIOLOGICAL_PROCESSsubject,TRAITobject. - Graph structure. Four declared nodes, three edges, every subject/object is a declared
node_id;app/dashboard/priority.jsonreportscomponents: 1, orphans: 0, so nothing is stranded off the trait node. - Derived artifacts. The five tracked files that carry a per-record row for a sibling like
viperin_system—reports/graph_protein_taxon_coverage.tsv,reports/node_grounding_residual.tsv,reports/proposal_citation_audit.tsv,app/dashboard/priority.json,app/discussions/data.js— all gained the matching VP1840 entry. Nothing regenerable was missed. - README counts. Re-derived rather than trusted: 845 records, 427 REVIEWED / 330 PROPOSED / 50 DEPRECATED / 38 SEEDED, 601 with
causal_graphs, genomics 259 total and 240 PROPOSED. The prose breakdown also sums to 330. All match. - Contract.
scripts/add_vp1840_system_trait.pygates on--applyand goes throughwrite_validated_trait+record_curation_event, matching the ~500 sibling add scripts.
Two things to be explicit about:
What I could not verify. This review environment has no network, so I did not open 41564_2025_1927_MOESM2/3_ESM.xlsx, the DOI, or the pinned defense-finder-models files — audit_evidence_snippets.py only checks snippet shape, so nothing in CI verifies the text either. What I could do is an internal-consistency check: the Liste_hmm_system.md row's column padding is byte-identical in width to the same row quoted in ds_10_system.yaml (49/50/23/25/7), and the List_system_article.md row escapes exactly the periods a real row would. That is corroborating, not verification.
🔵 One snippet to confirm, not a defect: the Supplementary Data 1 evidence keeps a header-plus-row stitch (Scientific Name\t…\nVibrio parahaemolyticus\t…), while the last curation event (UNSTITCHED_PHAGE_PLATING_SNIPPETS) deliberately removed that same construction from the Table 9 Bas21/LL1 snippets as non-contiguous. The playbook's rule is a "contiguous span copied from the paper", so this is only clean if the WP_229652597.1 row is the first row under that header. Plausible for a top PSI-BLAST hit, and the fact that the >BAC60103.1 query line was quoted as a separate item suggests you were already tracking this — worth one look, not a change request.
Advisory only, per the usual: history/records/vp1840_system/ holds a single CREATE session at 02:27:53Z, but the record's curation_history has four later events (02:51, 03:13, 03:27, 03:45) with no matching repository history record. The history CI job is satisfied either way.
🤖 Generated with Claude Code
Summary
Validation