Which dismech pathographs could seed AOPs: measurable node chains

Date: 2026-09-10 Scope: a census of kb/disorders/ and kb/modules/ (3,009 files) for causal chains whose every node carries a model-system measurement. One such chain is the raw material for one putative Adverse Outcome Pathway, so the census asks how many entries hold a chain long enough to be worth deriving — not how many AOPs the knowledge base amounts to. Companion: projects/AOP_EMOD_ALIGNMENT.md, which supplies the framework comparison this report screens against. Worked derivations: two entries from the tables below are taken all the way through into AOP form in projects/AOP_EMOD_ALIGNMENT/putative-aops-from-dismech-2026-09-10.md. Illustrated summary: Pathographs to Pathways — a web page summarising this report and those derivations, with two chain diagrams neither document contains. It was written by hand rather than generated, so the two Markdown files are the authority where they disagree.

Where the census started, and why that first cut failed

This began as a simpler question: which entries have at least two experimental model systems attached? That is the obvious first cut, because an entry with several models sounds like an entry whose mechanism has been measured. It returns 161 entries, and almost none of them are usable. The reason is worth setting out before the counts, because it governs how every table below should be read.

The alignment page records that an AOP Event is defined partly by how it is measured — Event.measured_or_detected is a deployed, populated v2.8 field, present on all 1,598 Events in the 2026-08-06 export. dismech has the structural counterpart in the wrong place for this purpose: Pathophysiology.assays mirrors Event.assays but is effectively unused (as counted on the alignment page on 2026-09-07: 2 of 564 assay entries sit on a pathophysiology node, and none bind an OBI term).

What is populated is measurement hanging off the model, not the node: ModelMechanismLink.readouts on experimental_models, animal_models and computational_models. So a node's measurability in dismech is a property of the links pointing at it, and an entry-level count of models says nothing about whether those models land on consecutive nodes.

That distinction is the finding. Counting models per entry returns a large, mostly unusable set, because the models can all sit on one node or scatter across unconnected branches; requiring them on consecutive nodes returns a small, directly usable one.

KB totals

Every count below is from just aop-chain-census, run against main on 2026-09-16. Regenerate it rather than trusting these figures: they move with every curation PR.

A model block is one entry in an experimental_models: / animal_models: / computational_models: list. An entry can hold several, so blocks and entries are different units and both are given:

Model section	Model blocks	Blocks with `modeled_mechanisms`	Entries with one	Entries with a linked one
`experimental_models` (NAM)	649	590	403	370
`animal_models`	1,687	1,259	917	698
`computational_models`	99	79	39	35

Entry-level filters, for the record:

161 entries carry ≥2 experimental_models blocks.
525 entries carry ≥2 pathograph-linked models of any kind.
885 entries carry at least one pathograph-linked model.

The chain screen

For each entry, build the causal graph from pathophysiology[].downstream[].target (bare names, matched verbatim — see CLAUDE.md, Pathograph Targets Are Bare Names), then find the longest path in the subgraph induced by nodes that are a modeled_mechanisms target.

Two tightenings:

Requirement on every node in the chain	≥3 nodes	≥4 nodes	≥5 nodes
≥1 model link carrying `readouts` (measured)	97	21	5
≥2 distinct models linked	13	4	1

A readout is what makes a node an Event rather than an assertion. Neither row says anything about the arrows between nodes, which is the defect The joint screen below repairs.

The five 5-node fully-measured chains

Entries in this table and the next link to their rendered dismech pages. The Joint screen column is the result from The joint screen below, which requires evidence on every edge as well as a readout on every node — the screen that actually selects derivable chains. Only one of these five survives it.

UI developer note — how these URLs are built

Two things govern them. They are absolute because this document is built by MkDocs into `elements/`, which does not contain `pages/`, so a relative link would not resolve. And the page slug comes from each entry's `name:` field, not its filename: | KB file | Rendered page | |---|---| | `Chemotherapy_Induced_Diarrhea.yaml` | `Chemotherapy-Induced_Diarrhea.html` | | `Mitochondrial_Complex_I_Deficiency_Nuclear_Type_1.yaml` | `Mitochondrial_Complex_I_Deficiency,_Nuclear_Type_1.html` | Deriving a URL from the filename gives a dead link for both. Nothing in the repository checks links in `docs/` against the files in `pages/`, so a renamed entry breaks these silently.

Entry	Linked models	Readouts on chain	Joint screen
`Left_Ventricular_Noncompaction_8`	2 NAM / 7 animal	15	5 nodes
`Wiedemann-Rautenstrauch_Syndrome`	5 NAM / 2 animal	11	2 nodes (2/38 edges)
`Mitochondrial_Complex_I_Deficiency_Nuclear_Type_1`	2 NAM / 1 animal	10	1 node (0/16 edges)
`Autosomal_Dominant_Nonsyndromic_Hearing_Loss_25`	2 animal	11	1 node (0/11 edges)
`Liver_Cirrhosis`	1 NAM	6	2 nodes (3/8 edges)

The chains themselves:

Left_Ventricular_Noncompaction_8 — Loss of Compact Myocardium Transcriptional Identity → TGF-beta Signaling Dysregulation → Impaired Cardiomyocyte Proliferation → Left Ventricular Dilation and Systolic Dysfunction → Arrhythmia and Ventricular Pre-excitation
Wiedemann-Rautenstrauch_Syndrome — Aberrant POLR3A Transcript Processing → Reduced Wild-Type POLR3A Expression → RNA Polymerase III Transcriptional Hypofunction → Nucleolar Disruption, p53 Activation and Premature Senescence → Impaired Mesenchymal Progenitor Proliferation and Differentiation
Mitochondrial_Complex_I_Deficiency_Nuclear_Type_1 — Arrest of Complex I Assembly at the CI-830 Subcomplex → Isolated Complex I Deficiency → Reductive Stress and Bioenergetic Failure → Leukocyte-Mediated Neuroinflammation → Symmetric Necrotizing Brainstem and Basal Ganglia Lesions
Autosomal_Dominant_Nonsyndromic_Hearing_Loss_25 — Inner Hair Cell Stereocilia Bundle Disruption → Reduced Inner Hair Cell Receptor Potential → Synaptic Ribbon Enlargement and Altered Sustained Exocytosis → Failure of Auditory Nerve Activation with Preserved Cochlear Amplification → Secondary Deafferentation of the Inner Hair Cell
Liver_Cirrhosis — Hepatocyte Injury and Death → Kupffer Cell Activation → Hepatic Pro-Inflammatory Mediator Release → TGF-beta Signaling in Fibrogenesis → Hepatic Stellate Cell Activation

Liver_Cirrhosis is the AOP 38 comparator already worked through on the alignment page, so it validates the screen rather than producing anything new: the whole chain is instrumented by one NAM, the Akura Twin microphysiological system, which its authors describe as "mimicking the key events of the liver fibrosis AOP".

The thirteen chains where every node has ≥2 independent models

Two models on one node is what a KER's empirical support wants, and — more usefully — what makes a disagreement between models visible.

Nodes	Entry	Linked models	Joint screen
5	`Left_Ventricular_Noncompaction_8`	2 NAM / 7 animal	5 nodes
4	`Cystic_Fibrosis`	3 NAM / 3 comp	0 (34/55 edges)
4	`cellular_senescence` (module)	1 NAM / 2 animal / 4 comp	1 (0/6 edges)
4	`genomic_instability_aging` (module)	2 animal / 2 comp	1 (0/3 edges)
3	`Chemotherapy_Induced_Diarrhea`	6 NAM	0 (3/18 edges)
3	`Hereditary_Spastic_Paraplegia_3A`	4 NAM / 4 animal	3 nodes
3	`Intellectual_Developmental_Disorder_Autosomal_Recessive_67`	1 NAM / 2 animal	1 (0/3 edges)
3	`Metabolic_Dysfunction-Associated_Steatotic_Liver_Disease`	4 NAM / 3 animal	3 nodes
3	`Primary_Ciliary_Dyskinesia`	11 NAM / 4 animal / 1 comp	3 nodes
3	`TRAPPC12-Related_Encephalopathy`	5 NAM / 1 animal	0 (9/38 edges)
3	`Type_2_Diabetes_Mellitus`	3 NAM / 1 animal / 1 comp	3 nodes
3	`deregulated_nutrient_sensing` (module)	1 animal / 3 comp	1 (0/12 edges)
3	`epigenetic_alterations` (module)	1 animal / 3 comp	1 (0/2 edges)

Primary_Ciliary_Dyskinesia is the KB's most model-rich entry (11 NAMs, 16 linked models total) and still yields only a 3-node chain — the models cluster rather than chain. That is the entry-level-count failure in one line.

Leading candidate: `Left_Ventricular_Noncompaction_8`

The only entry in the KB with a 5-node chain where every node carries ≥2 independent models, and the strongest AOP-derivation target for three reasons beyond the count.

It spans all four levels of biological organisation, already recorded. Every node carries biological_scale and every one of the nine linked models declares model_scale — the axis AOP-Wiki calls LoBO, and the one construct both frameworks already carry. The ≥2-model chain walks it cleanly:

Node	`biological_scale`
TGF-beta Signaling Dysregulation	`MOLECULAR`
Impaired Cardiomyocyte Proliferation	`CELLULAR`
Failed Ventricular Compaction	`TISSUE`
Left Ventricular Dilation and Systolic Dysfunction	`ORGANISM`
Arrhythmia and Ventricular Pre-excitation	`ORGANISM`

The readout-backed chain in the table above substitutes Loss of Compact Myocardium Transcriptional Identity (CELLULAR) at the head and drops Failed Ventricular Compaction, so it reaches three levels rather than four — that node is model-linked but carries no readout. Which of the two is the better AOP skeleton is a real choice, not a technicality.

It contains recorded model disagreement. Two links are FAILS_TO_RECAPITULATE:

the Prdm16Q187X knock-in mouse fails to recapitulate TGF-beta Signaling Dysregulation, which the cardiomyocyte-specific conditional knockout mouse does recapitulate (fidelity: MODERATE);
the same knock-in fails on Left Ventricular Dilation and Systolic Dysfunction, where the Prdm16cKO, the systemic monoallelic mouse, the zebrafish model and the sex-stratified conditional knockout all report the node with fidelities from HIGH to LOW.

Per CLAUDE.md a FAILS_TO_RECAPITULATE link requires both limitations and evidence, so these are substantiated negative claims. In AOP terms that is uncertainties-or-inconsistencies material — a first-class KER field — arriving pre-curated.

No AOP-Wiki counterpart. Neither PRDM16 nor noncompaction/non-compaction appears anywhere in the AOP-Wiki export — no Event, KER or AOP — checked by string search of the cached 2026-08-06 and 2026-09-03 snapshots (1,598 Events in both). The one compaction hit is chromosome super-compaction in the Genomic Instability Event, unrelated. So a derivation here would be a seeding candidate rather than a re-derivation of something AOP-Wiki already holds. This is a statement about those two strings in that corpus, not a survey of cardiac AOP coverage, which is substantial.

Two gaps the screen exposed

model_scale is set on 36 of the 134 links landing on the 20 four-plus-node chains. The node-side biological_scale is better at 68 of 84 chain nodes, but the four entries with no scale on any chain node are Myocardial_Infarction and the cellular_senescence, loss_of_proteostasis and inflammaging modules. The alignment page already found this axis empty on almost every node the liver-fibrosis use case touched; it is still the cheapest, schema-free thing to populate, and without it a derived Event has no LoBO.

A readout is not an assay. ExperimentalReadout gives the measured quantity and its direction, which is enough to say an Event was measured. It does not give the method — the alignment page's How each model records the way a claim was measured section records that Pathophysiology.assays exists for this, binds to an AssayTerm enum rooted at OBI:0000070, has no OAK adapter behind it, and holds no OBI CURIE anywhere in kb/. So a dismech-derived Event can populate AOP's what was measured but not its how it was measured or detected.

Method

scripts/aop_chain_census.py, run as just aop-chain-census. Regenerable from the KB alone; no network access is involved.

Only one kind of arrow is counted. "Causal edge" here means a pathophysiology[].downstream[] entry — a mechanism node pointing at another mechanism node. The rendered disorder page draws several other kinds of arrow into the same graph, and none of them is in these counts:

Arrow type	Counted?
`pathophysiology[].downstream[]`	yes
`environmental[].influences_mechanisms`	no
`treatments[].target_mechanisms`	no
`phenotypes[].sequelae`	no
`phenotypes[].reports_on`	no
`experimental_models[]` / `animal_models[]` / `computational_models[].modeled_mechanisms`	no

The restriction is deliberate: a Key Event Relationship is a step from one mechanism to the next, so an exposure link, a treatment link or a model link is not a KER. But it means an entry's edge count here is lower than the arrow count on its page. Chickenpox has 3 node-to-node edges and 3 further environmental links, all six cited; it counts as 3/3.

One node universe for every column. A chain is a simple path (no repeated node) over pathophysiology nodes only. A downstream target naming a phenotype ends the chain and does not count toward its length — phenotypes carry no downstream edges of their own, so admitting them would lengthen chains under one screen and not the other and make the columns incomparable. Dangling targets — bare names matching no node — are excluded by construction, so the counts are unaffected by the grandfathered backlog in tests/causal_target_baseline.txt.

The four screens, each applied to that same universe:

Column	A node qualifies when	An edge qualifies when
Every node measured	some `modeled_mechanisms` link targeting it has a non-empty `readouts`	always
≥2 distinct models	≥2 model blocks link to it	always
Every edge cited	always	its `downstream[]` entry has non-empty `evidence`
Both	as measured above	as cited above

measured = {link.target for model in models for link in model.modeled_mechanisms
            if link.readouts}
cited    = [(node.name, d.target) for node in doc.pathophysiology
            for d in node.downstream if d.evidence and d.target in patho_node_names]

# longest simple path in the subgraph induced by the chosen node set,
# over the chosen edge set

The derivations, and what they say about this screen

Two entries from the table have been worked through in full as putative AOPs: projects/AOP_EMOD_ALIGNMENT/putative-aops-from-dismech-2026-09-10.md.

Left_Ventricular_Noncompaction_8 transfers its Event and KER layers nearly for free — three of its seven Events already exist in AOP-Wiki, three are proposed new, and all six relationships carry evidence — but has no MIE and can have none, because a germline lesion is not a stressor-biomolecule interaction.

Skeletal_Fluorosis was then derived as the chemically initiated counterpart. It has two MIE candidates, and its downstream Events join an OECD-endorsed radiation AOP (AOP 482) whose stressor is entirely different — while fluoride appears in none of AOP-Wiki's 756 stressor records. But it also exposes a defect in the screen below.

The screen measures Events and ignores KERs. It ranks a chain by whether every node carries a readout, because an AOP Event is defined partly by how it is measured. That says nothing about the arrows. In Skeletal_Fluorosis the two run opposite ways: the entry carries evidence on 16 of its 20 causal edges, and all four that it does not are inside the chain this screen selected, so the chain first derived from it had 2 of 5 relationships supported. Re-deriving it along the entry's evidenced backbone instead — a 9-node path through the gut-microbiome arm — gives 8 of 8. The measurable nodes are the ones models instrument, and models measure states rather than transitions.

The joint screen

Requiring both — a readout on every node and evidence on every edge — is what actually selects AOP-derivable chains.

Longest chain	Every node measured	Every edge cited	Both
≥2 nodes	282	1,349	123
≥3 nodes	97	1,051	38
≥4 nodes	21	731	6
≥5 nodes	5	435	1

37,001 pathophysiology[].downstream[] edges KB-wide, 15,790 carrying evidence (43%). Restricted to the node-to-node subset these chains are built from — edges whose target resolves to another pathophysiology node, excluding those terminating on a phenotype — 18,058 edges, 7,386 cited (41%). Fully-cited chains reach 12 nodes, where fully-measured ones stop at 5.

Edge evidence is about thirty-five times more available than node measurability — 731 entries carry a four-node chain cited at every step, against 21 measured at every node. So the binding constraint on deriving an AOP from any one dismech entry is modeled_mechanisms coverage, not literature. The 38 entries clearing the joint screen at three nodes are the real candidate list — just aop-chain-census --list-joint 3 prints them — and the two candidate tables above are an upper bound.

Left_Ventricular_Noncompaction_8 is the only entry in the knowledge base clearing both requirements at five nodes, so the derivation nominated below survives the tightened screen — by luck, since nothing in the original ranking looked at edges.

What this report does not decide

Whether a derived AOP should be recorded anywhere in dismech — a Pathophysiology cross-reference slot carrying aop.events: CURIEs, an AOP-Wiki structured reference source — is open and belongs in its own issue. See Open schema questions on the alignment page.