From evidence pointers to experiment-grounded evidence — a worked example on Familial Hypercholesterolemia

Date: 2026-07-30 Status: design exploration / proposal (not yet a schema change) Scope: the dismech EvidenceItem model, illustrated on kb/disorders/Familial_Hypercholesterolemia.yaml

---

1. The question

Is the dismech evidence model about evidence, or is it mostly a set of guardrails against AI hallucination? Today an EvidenceItem carries seven fields:

evidence:
- reference: PMID:1301956            # a real, resolvable ID
  reference_title: "..."
  supports: SUPPORT                  # polarity: SUPPORT/REFUTE/PARTIAL/NO_EVIDENCE/WRONG_STATEMENT
  evidence_source: HUMAN_CLINICAL    # provenance-of-organism, coarse
  snippet: "..."                     # an EXACT quote, validated as a substring of the cited text
  explanation: "..."                 # curator prose
  images: [...]

The load-bearing machinery is reference + snippet + the linkml-reference-validator pipeline, and its job is to answer "is this citation real, and does the quoted text actually exist in it?" That is citation integrity. It is necessary and it is dismech's primary defense against fabrication — but it is not evidence appraisal. Two fields gesture at evidence semantics and both are thin:

• supports encodes direction, not strength. A single case report and a human natural-knockout study both collapse to SUPPORT.
• evidence_source encodes study type, not study design or inferential power. Every human observation from an n=1 case report to a 27,000-patient trial is one value.

Nothing records what experiment was run, what was measured, what resulted, or how the mechanistic claim was inferred from that result. That is the gap this document explores, using FH — arguably the best-evidenced disease mechanism in all of human biology.

---

2. Two real anchors already in the FH file

2a. The mis-targeting problem (evidence pointer ≠ evidence for this edge)

Two monogenic head-nodes — APOB-LDLR Binding Defect and PCSK9 Gain-of-Function — are each supported by the same GeneReviews sentence (PMID:24404629, which appears four times in the file: node-level and edge-level on both nodes):

"The molecular diagnosis of FH can be established by identification of heterozygous or biallelic pathogenic variants in APOB (variants that impair binding of LDL-C to the LDL receptor), LDLR, or PCSK9 (gain of function); or rarely, identification of biallelic pathogenic variants in LDLRAP1."

That snippet validates every time — it is a real quote. For the APOB node it is arguably on-target: the parenthetical "(variants that impair binding of LDL-C to the LDL receptor)" states exactly the binding-defect mechanism that node claims. But for the PCSK9 Gain-of-Function → PCSK9-mediated LDLR degradation edge the same sentence is a diagnostic-classification statement — it says PCSK9 gain-of-function is a gene you sequence to diagnose FH, and says nothing about degradation, the actual mechanism on that edge. That is the mis-targeting: a valid quote standing in for mechanism evidence it does not contain. (The LDLRAP1-Related LDL Uptake Defect node, by contrast, does not use this sentence — it cites LDLRAP1-specific evidence, DOI:10.3390/ijms24043224, and is correctly targeted.) The PCSK9 edge is the clean, unambiguous example, and the one this document builds on.

2b. The experiment is already in the prose — just not structured

On the PCSK9 genetic entry (its genetic: block) sits a genuinely mechanistic, already-validated snippet:

- reference: DOI:10.1073/pnas.0409736102
  supports: SUPPORT
  evidence_source: IN_VITRO
  snippet: "Overexpression of PCSK9 in HepG2 cells caused a decrease in whole-cell and
    cell-surface LDLR levels. PCSK9 overexpression had no effect on LDLR synthesis but
    caused a dramatic increase in the degradation of the mature LDLR"

Read what that one string actually contains:

Experiment dimension	Value carried in the snippet
System	HepG2 cells (hepatocyte-like)
Perturbation	PCSK9 overexpression — a gain-of-function manipulation
Readout	whole-cell + cell-surface LDLR level; LDLR synthesis; LDLR degradation
Result	LDLR ↓; synthesis unchanged; degradation ↑
Inference	PCSK9 acts post-translationally to increase LDLR degradation → sufficiency + localizes the mechanism to degradation, not synthesis

Every one of those dimensions is present — but the model stores the whole thing as an opaque string with supports: SUPPORT. None of it is queryable. You cannot ask the KB "show me every edge established by a gain-of-function perturbation", or "every claim whose only evidence is correlative", because that structure lives only inside free text.

---

3. What experiment-grounded evidence would encode

The proposal: keep the exact-snippet discipline (it stays the anti-hallucination floor), but let an evidence item optionally decompose the micropublication — the experiment and the inferential step from result to claim. A sketch (ExperimentalEvidence, extending EvidenceItem):

evidence:
- reference: DOI:10.1073/pnas.0409736102
  supports: SUPPORT
  evidence_source: IN_VITRO
  snippet: "Overexpression of PCSK9 in HepG2 cells caused a decrease in whole-cell and
    cell-surface LDLR levels. PCSK9 overexpression had no effect on LDLR synthesis but
    caused a dramatic increase in the degradation of the mature LDLR"
  experiment:
    design: OVEREXPRESSION                     # closed enum (see §5)
    system:
      organism:  {id: NCBITaxon:9606, label: Homo sapiens}
      cell_type: {id: CL:0000182, label: hepatocyte}   # HepG2, hepatocyte-like
      background: HepG2 cell line
    perturbation:
      role: GAIN_OF_FUNCTION
      target_gene: {id: hgnc:20001, label: PCSK9}
      method: transient overexpression
    readout:
      measured: cell-surface and whole-cell LDLR protein; LDLR synthesis; LDLR degradation
      method: immunoblot / surface labeling
    result:
      direction: DECREASED                     # of LDLR
      qualifier: synthesis unchanged; degradation increased
      snippet: "no effect on LDLR synthesis but caused a dramatic increase in the
        degradation of the mature LDLR"        # exact-quote grounds the RESULT specifically
    inference:
      role: SUFFICIENCY_GOF                     # what the result licenses about the edge
      supports_edge: "PCSK9 Gain-of-Function#PCSK9-Mediated LDLR Degradation"
      claim: PCSK9 is sufficient to drive post-translational LDLR degradation
      caveats: hepatocyte-like line, not primary human hepatocytes; overexpression is
        non-physiologic in level

The key move: the exact snippet now grounds a specific result, and a separate, typed inference records how the mechanistic edge is drawn from that result. Strength is no longer a subjective grade a curator invents (which would reopen the fabrication hole) — it is read off the experiment's design and the inference role, both of which are snippet-anchored.

---

4. Worked example — the PCSK9 → LDLR sub-graph, four experiments, one edge

FH is the gold standard because necessity, sufficiency, direct mechanism, and therapeutic rescue all converge on the same arrows. Here is that convergence made explicit on the PCSK9 → LDLR degradation → reduced LDL clearance sub-chain. (PMIDs marked ⚠ are canonical papers that must be fetched with just fetch-reference and snippet-validated before any commit; snippet text below is illustrative placeholder, not a validated quote.)

A. Sufficiency + direct mechanism — cell overexpression (real, already in repo) DOI:10.1073/pnas.0409736102 — HepG2 PCSK9 overexpression → LDLR degraded, synthesis unchanged. design: OVEREXPRESSION, inference.role: SUFFICIENCY_GOF.

B. Necessity — model-organism loss of function ⚠ Pcsk9 knockout mouse → increased hepatic LDLR, lower plasma LDL. design: KNOCKOUT, evidence_source: MODEL_ORGANISM, inference.role: NECESSITY_LOF. (Remove PCSK9, the effect on LDLR reverses — the necessity counterpart to A.)

C. Human natural knockout — the single most important missing citation ⚠ Cohen, Hobbs et al. (NEJM 2006): PCSK9 nonsense variants → ~28% lower LDL and ~88% reduction in coronary heart disease over 15-year follow-up (ARIC cohort). design: HUMAN_GENETIC_LOF, evidence_source: HUMAN_CLINICAL, inference.role: NECESSITY_LOF. A natural human loss-of-function experiment confirming both the edge direction and its disease relevance — currently absent from the entry. Its reciprocal, Abifadel et al. (Nat Genet 2003) ⚠, identified the gain-of-function variants that cause FH — the human sufficiency arm.

D. Therapeutic rescue — pharmacological perturbation in humans ⚠ FOURIER (Sabatine et al., NEJM 2017): anti-PCSK9 monoclonal antibody (evolocumab) → ~59% LDL reduction and fewer cardiovascular events in ~27,500 patients. Inhibiting PCSK9 raises LDLR and lowers LDL — a pharmacological confirmation of the entire PCSK9 → LDLR → LDL → ASCVD chain. dismech already has target_mechanisms (TreatmentMechanismTarget) linking the drug to the node; that link is confirmatory edge evidence and should be citable as such.

The convergence view for a single edge

Experiment	Design	Source	Perturbation	Inference role
A. HepG2 overexpression	OVEREXPRESSION	IN_VITRO	GoF	SUFFICIENCY_GOF
B. Pcsk9-null mouse	KNOCKOUT	MODEL_ORGANISM	LoF	NECESSITY_LOF
C. Human nonsense carriers	HUMAN_GENETIC_LOF	HUMAN_CLINICAL	natural LoF	NECESSITY_LOF
D. Evolocumab (FOURIER)	RANDOMIZED_TRIAL	HUMAN_CLINICAL	pharmacologic LoF	THERAPEUTIC_RESCUE

Four orthogonal method classes agreeing on one arrow. That agreement is the evidence — and today the model records none of it, only causal_link_type: DIRECT (topology) and supports: SUPPORT (polarity) four times over. With the structure above, the KB can compute "this edge is supported by convergent necessity + sufficiency + human-genetic + therapeutic evidence" — the actual reason a biologist believes it.

---

5. Proposed vocabularies (small, closed, dismech house-style)

Two enums do most of the work. Both are ~6–10 values, matching the register of existing enums like biological_scale (4) and evidence_source (5).

experiment.design — how it was shown (the observable protocol): CORRELATIVE_OBSERVATION, OVEREXPRESSION, KNOCKDOWN, KNOCKOUT, HUMAN_GENETIC_LOF, HUMAN_GENETIC_GOF, RESCUE_COMPLEMENTATION, EPISTASIS, DIRECT_BIOCHEMICAL (binding/structure/reconstitution), RANDOMIZED_TRIAL, PHARMACOLOGIC_PERTURBATION.

inference.role — what the result licenses about the causal edge (the reasoning): CORRELATIVE, NECESSITY_LOF, SUFFICIENCY_GOF, RESCUE, EPISTASIS_ORDERING, DIRECT_PHYSICAL, THERAPEUTIC_RESCUE.

design is largely extractable from the reference metadata / snippet; inference.role is the appraisal axis — but it is constrained by design (an overexpression cannot establish necessity), so the pair is mutually checkable rather than freely asserted.

---

6. How this composes with what already exists

• causal_link_type (DIRECT / INDIRECT_KNOWN_INTERMEDIATES / INDIRECT_UNKNOWN_INTERMEDIATES / UNKNOWN) stays — it is topology (how many steps). inference.role is epistemics (how well established). Orthogonal, complementary.
• target_mechanisms (TreatmentMechanismTarget) already links treatment → node. Allowing that link to surface as confirmatory edge evidence turns experiment D into first-class support without new plumbing.
• association_signals.statistics is the existing structured-effect model for comorbidity pairs. The experiment.result block generalizes the same idea (direction/magnitude, snippet-grounded) to any evidence item.
• HUMAN_MODEL_MISMATCH discussions already flag model-to-human translation gaps; experiment.system + inference.caveats make the species/system distance structured rather than a prose note.
• Not ECO. ECO types the assay behind an annotation of an entity to a term (its GO heritage); dismech's unit is a supported assertion in a causal graph, and the axis it needs is the inferential role of the experiment for that edge — which ECO does not model. A bespoke closed enum fits both the modeling target and the house style.
• SEPIO is the correct ontology-backed shape for assertion ↔ evidence-line ↔ evidence-item ↔ provenance, and is Monarch-lineage — but it is heavy for a curation surface. Treat it as an export-layer target (as BioLink already is), not the thing curators hand-edit.

---

7. Guardrails preserved

1. Exact snippets stay mandatory — experiment.result.snippet is still validated as a substring, so the granular claim is grounded, not narrated.
2. Strength is derived, not authored — from design + inference.role, both constrained and snippet-anchored, so the appraisal layer does not become a new fabrication surface (the reason a curator-typed certainty: HIGH was rejected).
3. Claim-type-local — the causal-inference vocabulary applies to pathophysiology edges (the mechanistic backbone). Comorbidity, prevalence, and treatment-efficacy claims keep their own strength models rather than being forced through one column.

---

8. Next steps

1. Float experiment.design + inference.role on the design register (issue) before any schema change.
2. Prototype the block on the FH PCSK9 sub-graph only, fetching and validating the experiment B–D references first — the Pcsk9-null mouse (Rashid et al.), Abifadel, Cohen-Hobbs, and FOURIER (experiment A, the HepG2 overexpression, is already in the repo). Brown & Goldstein anchors the separate LDLR Functional Defect edge and would come in when the prototype extends beyond the PCSK9 sub-graph.
3. Re-point the mis-targeted PCSK9 GeneReviews snippet (§2a) to a real degradation-mechanism paper as a current-schema fix, independent of the extension.

Companion slide deck: From evidence pointers to experiment-grounded evidence. Conceptual overview: The Evidence Model.