Skip to content

How diet is represented in dismech — audit, 2026-09-01

Diet enters a dismech entry two ways, and they are separate claims that happen to share an ontology:

Where it lives Grounded by Reaches the pathograph via
Causal environmental[] food_source (FOODON/CHEBI), exposure_term (ECTO/XCO) influences_mechanisms
Intervention treatments[].treatment_term dietary_modifications (action + FOODON/CHEBI food) target_mechanisms

Phenylketonuria already carries both against the same FOODON:00001006: meat as an exposure, and RESTRICT meat as a prescription. That is not duplication to be factored out — an exposure claim and a prescription claim are different assertions with different evidence. The audit therefore never reconciles the two tracks against each other.

Reproduce with just diet-audit (--format tsv for the per-entry table). Figures below were last regenerated against main on 2026-09-05; the KB moves, so re-run rather than quoting these.

Headline

The gap that matters is evidence-backed diet annotations that never reach the mechanism graph, not unbound terms.

Causal Intervention
Diet-related entries 191 in 139 files 648 in 474 files
On the pathograph 139 (72.8%) 331 (51.1%)
Cited but off the pathograph 50 (38 after dropping weak matches) 282
…of those, CITED_HUMAN 36 223
On the pathograph but uncited 0 3
Evidenced only by REFUTE 1 10

The causal track is in good shape: nearly three quarters of its diet entries already carry influences_mechanisms, and the residue is 38 entries — a reviewable list, not a programme. The intervention track is the weaker half: half of dietary treatments never link to a mechanism node, leaving 282 cited-but-unlinked treatments.

Three entries are on the pathograph with no evidence anywhere — the dietary protein restriction in Chronic Kidney Disease, the ketogenic diet in Dravet syndrome, and the threonine restriction in Inherited Threoninemia. All three are interventions; the causal track has none. They already render as mechanism edges, so they assert more than the KB can support.

An earlier draft of this report put that number at 19. It was wrong: the audit graded only the entry's own evidence: block and never looked at the evidence on the link itself, which is where CLAUDE.md says the claim "this exposure acts on this node" belongs. Sixteen of the other entries carry snippet-backed SUPPORT there, including all three kb/modules/ entries the draft singled out as the worst case. --strict would have sent a curator to fix entries that were already right — the same failure mode as the REFUTE_ONLY bug below, one layer up. The gap counts are unaffected, since an unlinked entry has no links to read.

REFUTE_ONLY is counted separately and is not a defect. NELABA's "Lipoic acid supplementation (ineffective)" carries two snippet-backed REFUTE items against the mechanism it targets: a treatment recorded as failing against a node is a real, useful annotation, and an earlier draft of this audit wrongly flagged it as uncited.

The 38 strong causal candidates

Concentrated in entries where diet is central: Gout (beer, fructose-sweetened soft drink, red/organ meat, shellfish — all CITED_HUMAN, none linked), Phenylketonuria (dairy, meat, nuts), Celiac Disease (gluten, wheat, barley, rye), plus alcohol across eleven carcinoma, hepatic and cardiovascular entries, and single entries in Hyperlipidemia (high saturated fat), Obesity and Type 2 Diabetes (high-calorie diet), Osteoporosis (vitamin D deficiency), Scurvy (vitamin C deficiency), Thyroid Follicular Carcinoma (iodine deficiency), and Wilson Disease (dietary copper).

A CITED_HUMAN row is a candidate, not a verdict. Most of these citations are observational cohort associations, and an association is not a mechanism — read the snippet before drawing an edge. Gout's shellfish evidence, for instance, measures incident gout in a cohort rather than precipitation of a flare, which its own explanation already says.

Ontology binding — the secondary axis

State Causal Intervention
BOUND 120 (62.8%) 3 (0.5%)
PARTIAL (block present, no term:) 5 2
FREE_TEXT 66 (34.6%) 643 (99.2%)

dietary_modifications is effectively unused: 5 files in the whole KB (Celiac, ECHS1 Deficiency, Konzo, Lathyrism, Phenylketonuria), 11 modification records, against 648 dietary treatments. Only 19 FOODON bindings exist KB-wide across 11 distinct terms.

Free text is a legitimate outcome, not a backlog. Two structural reasons, and neither is a curation failure:

  1. FoodTerm excludes food components. It is reachable only from FOODON:00001002 (food product) and CHEBI:33284 (nutrient). Gluten (FOODON:03420177) sits under food material instead and is correctly rejected. conf/oak_config.yaml confirms that exclusion is intended behaviour rather than an adapter artifact — it uses two other food material terms as worked examples of CURIEs both adapters agree are not FoodTerm values. So Celiac's three grain vehicles are bound while its actual trigger, "Gluten Exposure", cannot be. The same will apply to casein, purines, and oxalate.
  2. Dietary patterns have no ontology home at all. FOODON describes food products, not eating patterns.

Where no term fits, free text is the right answer per .claude/skills/dismech-terms — no term beats a bad one. The audit reports FREE_TEXT as a state to review, never as an error.

Worth noting separately: 47 causal entries are pathograph-linked and free text, so they render as ungrounded nodes in an otherwise grounded graph. That is the subset where a binding, if a good one exists, buys the most.

Dietary-pattern CURIE scatter

The same pattern concept is bound inconsistently across entries:

alcohol:  ECTO:0001082 x12, ECTO:0300001 x1, ECTO:0000509 x1
dietary:  ECTO:0090010, ECTO:9000950, ECTO:9000084, FOODON:03303171, ECTO:0400019
diet:     ECTO:0090010 x3, XCO:0000013 x1, ECTO:9001347 x1

XCO:0000013 is a bare "diet" catch-all used where a specific pattern was meant. Alcohol is the healthy case — ECTO:0001082 dominates, with ECTO:0300001 correctly reserved for the maternal route and a single ECTO:0000509.

Standardizing on one ECTO CURIE per named pattern is the cheap win here, and needs no schema change.

Treatment-side inconsistency

Dietary treatments scatter across NCIT action terms — NCIT:C15747 (supportive care), NCIT:C15447 (dietary intervention), NCIT:C15433 (nutritional support), NCIT:C15986 (pharmacotherapy) — and 235 of 648 carry no therapeutic_modality at all, with 244 BEHAVIORAL and 106 SMALL_MOLECULE.

Do not mechanically backfill this. CLAUDE.md already records that NCIT:C15433 names a specific vitamin or compound far more often than a diet change, and that tagging it BEHAVIORAL was tried and reverted in 2026-07.

Method and its limits

Entries are matched by keyword, so the census is recall-oriented and carries a false-positive tail. Two mitigations, both visible in the output:

  • The two tracks read different fields. An environmental[] entry is short and wholly about its exposure, so its description is signal. A treatments[] description is a clinical paragraph that mentions diet incidentally — searching it pulled in ACE inhibitors (on "sodium"), cleft palate repair (on "feeding"), and beta blockers. The intervention track therefore matches name only, which cut it from 1,244 entries to 648.
  • Match provenance is recorded (matched_in: food_source / name / term_label / description). A causal entry matched only in description prose is the weak tail — 12 of the 50 causal gap rows, including Ependymoma ("high-dose ionizing radiation", whose description mentions diet) and CKD tobacco smoking (on "glycemic"). Filter with --strong-only.

Bare sodium was dropped from the keyword list: adding it back pulls in 40 more entries that are almost all drugs (sodium channel blockers, sodium valproate, sodium phenylbutyrate, sodium oxybate), and every genuinely dietary one carries "diet"/"dietary"/"salt"/"intake" anyway. septal ablation, vitamin K antagonist, and radioactive iodine are excluded by name.

Evidence tiers are structural, not a quality judgment. The script cannot read a paper; it reports the shape of the citation. CITED_HUMAN means a SUPPORT item with a non-empty snippet graded HUMAN_CLINICAL. Snippet-backed REFUTE evidence takes its own REFUTE_ONLY tier: it never counts toward the gap, since refuting evidence is not a reason to draw an edge, but it is emphatically not an uncited link either.

Recommendations

Ordered by value per unit of work. None is started; all are decisions for a curator, not automated fixes.

  1. Cite or unlink the three uncited pathograph edges (Chronic Kidney Disease, Dravet syndrome, Inherited Threoninemia). They already render.
  2. Work the 38 strong causal candidates, reading each snippet before adding influences_mechanisms. Gout, Phenylketonuria, and Celiac Disease alone are 11 of them and are the natural pilot, since all three already model diet well on the other track. A first tranche is already proposed separately in PR #10359, which works 42 candidates from an earlier run of this audit and adds 19 of them; merging it will shrink this list.
  3. Standardize the dietary-pattern CURIEs, one ECTO term per named pattern, retiring the bare XCO:0000013 catch-all.
  4. Widen the FoodTerm root to admit the FOODON food material branch, so food components (gluten, casein) become bindable. This is a schema change with a cache rebuild, and needs its own decision-register entry. Note that why the root was drawn at FOODON:00001002 is not recorded anywhere — conf/oak_config.yaml documents only that the resulting exclusion is intended, not the reasoning behind the root itself. That missing rationale is a reason to take the decision deliberately, not a licence to widen it.
  5. Leave dietary_modifications alone for now. At 5 files it is barely load bearing, and populating it is only worth doing after item 4 settles what can be bound.