PhysioMap: assessment and relevance to dismech (2026-08-23)

Subject. Hoehndorf R, Schofield PN, Gkoutos GV. PhysioMap: an ontology-grounded causal knowledge graph of human physiology. bioRxiv preprint, posted 2026-08-07. doi:10.64898/2026.08.03.742471 · code + OWL release · interactive demo. Code BSD-3-Clause; map, benchmark data, and docs CC BY 4.0. To appear in a PSB-style proceedings volume (© 2027).

How to read this document. It has three layers, kept deliberately separate because conflating them is misleading:

Part	Source	What it is
I (§1)	the preprint only	What the authors claim and how they say they built and evaluated it
II (§2)	their public repo	Their actual working artifacts — curation files, evidence records, the archived expert review
III (§3)	my own computation	Analyses I ran over the v1.1.1 release that are not in the paper
IV–V (§4–5)	this repo	Comparison with dismech and recommendations

Nothing here modifies a KB entry, schema file, or cache.

Bottom line. PhysioMap and dismech are close to structural duals, and the difference is not subject matter — it is what an edge means and what a node names. dismech has over 11× PhysioMap's edge count but cannot run PhysioMap's central inference, because dismech causal edges carry directness without polarity and dismech nodes are file-scoped strings rather than a shared variable namespace. Those are two separable gaps, and the first is much cheaper to close than the second.

Their curation discipline is worth studying independently of the schema question: every imported edge carries a machine-checkable provenance record, failed checks are recorded rather than dropped, and the expert review is archived down to the sha256 of the returned spreadsheet (§2).

---

Part I — What the paper says

1. The claims and method as published

1.1 There are five distinct graph objects, not one

The paper is careful about this and it is the part most easily blurred. One knowledge base is built, and several different graphs are derived from it:

Object	What it is	Size
K	OWL 2 EL TBox. Not a graph — a terminology of trait classes and axioms	1,699 traits + relation axioms
CKG	The projected typed causal knowledge graph, produced from K by a versioned pattern registry Π	1,699 nodes, 2,387 typed relation instances
G∂ = (V, E∂)	Signed dependency graph. Causal influence + production + quantitative identity only. Defines the signed Jacobian	cyclic; largest SCC = 213 traits
Gκ = (V, C)	Constitutive dependency graph. Shares vertices with G∂ but deliberately does not enter the causal Jacobian	acyclic by construction
H ⊆ V × Ec	Modulation — (node, edge) pairs, i.e. hyperedges. The first-order solver does not read these	19

At query time two more are derived: G∂^do(k), which is G∂ with the intervention target's incoming edges cut, and the condensation of that into a DAG of strongly connected components, processed in topological order.

Why constitution is split into its own graph: an intervention on a whole must replace the whole's defining equation, not propagate backwards into its parts. Body weight rising must not imply adipose mass rose. Determination therefore flows along Gκ in topological order only, and because scales are strictly hierarchical, Gκ has no cycles.

1.2 Trait semantics

A trait is the entity–quality (EQ) pattern used by HPO/MP: a PATO quality borne by an entity or process in a recorded context — Trait ⊓ ∃hasPart.(E ⊓ ∃κ.C ⊓ ∃hasQuality.Qᴱ), where κ is contextPartOf for continuants and occursIn for processes (occursIn ∘ partOf ⊑ occursIn, so process location propagates along parthood). Entity classes come from ChEBI, PR, CL, Uberon, GO; qualities from PATO. The context conjunct is omitted when no context is recorded.

Causal claims are asserted type-level via the collection pattern (T_tᵃˡˡ ⊑ ∃hasMember.(∃causedBy.T_s)) rather than the stronger T_t ⊑ ∃causedBy.T_s, which would claim every instance of the target is so caused. The naive subclass axiom is the wrong claim, not merely an expensive one; the collection form also stays in OWL 2 EL.

1.3 The five relation types and their SCM readings

The paper's core claim is that typing a relation determines what an intervention replaces:

Relation	Quantitative constraint	Enters causal Jacobian?
Causal influence	target mechanism Ft depends non-constantly on xs; σ = sign(∂Ft/∂xs)	yes
Production	Ft = Rt + εpt·qpt(xp), q′ > 0, ε ∈ {−1,+1}	yes
Quantitative identity	Xq = Qq(Xa₁…Xaℓ) — a named variable is a function of its arguments	yes
Constitution	Xw = Cw(Xp₁…Xpr) — a whole determined by its constituents	no (separate Gκ)
Modulation	mixed partial: sign(∂²F̃t/∂η∂θ) = µ	no (retained in H, unread)

Sign values are {+, −, 0, ?}. ? means non-constant dependence with no sign constraint — it is a positive claim, not a curation gap.

1.4 Every example edge the paper actually gives

This is a short list — the paper contains roughly twenty example edges in total.

Table 1, one per relation type:

• Causal influence — "More plasma ACE directly raises plasma angiotensin II: D_ACE,AngII > 0"
• Production — "A higher adrenal aldosterone-secretion rate raises plasma aldosterone"
• Constitution — "Adipose-tissue mass contributes positively to total body weight"
• Quantitative identity — "Mean arterial pressure = cardiac output × total peripheral resistance; both partial derivatives are positive"
• Modulation (§2.3 text) — "higher cortisol makes a fixed increase in norepinephrine produce a larger increase in total peripheral resistance"

Figure 1 is the only neighbourhood the paper draws: an intervention do(plasma ACE↓) with incoming edges cut, split into a whole-body loop and an intracellular arm. (The PDF figure text extracts out of order, so the signs below were verified against the release rather than read off the figure.)

Whole-body RAAS / baroreflex loop:

renal perfusion pressure                --[-]--> plasma renin activity
plasma renin activity                   --[+]--> plasma angiotensin II
plasma angiotensin II                   --[+]--> plasma aldosterone
plasma angiotensin II                   --[+]--> renal tubular sodium and water reabsorption
plasma aldosterone                      --[+]--> renal tubular sodium and water reabsorption
renal tubular sodium/water reabsorption --[+]--> blood volume
blood volume                            --[+]--> venous return
venous return                           --[+]--> cardiac output
mean arterial pressure                  --[+]--> renal perfusion pressure
mean arterial pressure                  --[+]--> baroreceptor afferent firing
baroreceptor afferent firing            --[-]--> baroreflex sympathetic outflow
baroreflex sympathetic outflow          --[+]--> plasma renin activity

Intracellular effector arm setting TPR:

plasma angiotensin II              --[+]--> AT1 receptor activation
AT1 receptor activation            --[+]--> cytosolic free calcium
cytosolic free calcium             --[+]--> myosin light-chain phosphorylation
myosin light-chain phosphorylation --[+]--> vascular smooth muscle contractile tone
vascular smooth muscle tone        --[+]--> total peripheral resistance
plasma ACE                         --[-]--> plasma bradykinin
plasma bradykinin                  --[+]--> endothelial nitric oxide production
vascular smooth muscle cGMP        --[-]--> cytosolic free calcium

ACE is a well-chosen example precisely because inhibition lowers angiotensin II and raises bradykinin — two routes to TPR pulling the same way by different mechanisms.

Figure 1 also carries the two non-causal types explicitly: a double line for the quantitative identity MAP = CO × TPR, and a dash-dot line for a modulation on the inhibitory baroreceptor→sympathetic edge (in the release: modulator = plasma angiotensin II, mixed-derivative sign +).

Figure 2a, the one worked prediction — HFE haemochromatosis:

do(hepcidin ↓)
plasma hepcidin      --[-]--> ferroportin (SLC40A1) iron-export activity   ⇒ ferroportin ↑
ferroportin activity --[+]--> plasma iron                                  ⇒ plasma iron ↑   ✓ HPOA
plasma iron          --[+]--> transferrin saturation (TSAT)                ⇒ TSAT ↑          ✓ HPOA

The release also holds TSAT --[+]--> hepcidin, closing the loop; the figure omits it because the intervention cuts hepcidin's incoming edges.

1.5 How the paper says the content was built

Four routes (§2.5):

1. Direct expert curation.
2. Extraction from quantitative models — the cardiovascular core used 17 Guyton BioModels modules and CellML integrators. "Available equations supplied derivatives" — the sign is read off a model's partial derivative rather than judged.
3. Curated databases — SIGNOR and AOP-Wiki, "directional sources [that] supplied constraints without complete functions": a sign but no function.
4. Checked language-model proposals from textbooks. An LLM proposes edges from textbooks; the proposals are then checked.

Trait creation is a separate pipeline: each candidate gets an entity identifier, a PATO determinable, an optional context, a stable label, and a scale; identifiers are resolved against pinned ontology releases; measurement kind is classified as extensive / ratio / rate / intensive; bearer and quality categories are checked for compatibility; duplicates are reconciled. Measurement kind has teeth — only composable kinds admit part-to-whole rules, which is why volume composes over parts and pressure does not. Unresolvable traits stay "explicitly primitive".

Release checks cover schema, references, identifiers, bearers, constitution, relation typing, and regression against a frozen behavioral baseline. The authors state these "assess formal consistency, not the scientific correctness of every axiom."

1.6 The evidence model is an admission gate, not content

§2.6, and the design decision is stated twice in the paper: evidence governs whether a relation is admitted, and is then used by nothing downstream. Neither the reasoner nor the solver reads evidence class or provenance. A relation's semantics comes from its type, arguments, and context; its sign constrains a derivative.

Six admissible classes for causal influence, in two groups:

• Interventional (a manipulation or valid instrument): perturbation, pharmacological intervention, genetic loss/gain of function, Mendelian randomization
• Mechanistic: mechanistic-model evidence (sign derived from a curated model), curated-mechanistic evidence (a curated functional or mechanistic account)

Explicitly insufficient alone: association, coexpression, binding. Support not yet mapped to a controlled class stays "unclassified" rather than being backfilled. Each stable relation id links to its support and a projection trace.

1.7 How they evaluate

(a) Expert content review. A co-author (P.N.S.) reviewed a fixed-seed stratified sample of 83 relations across all five types, marking TRUE (accepted) / FALSE? (flagged) / FALSE (rejected). Unequal sampling rates preclude a map-wide accuracy estimate.

(b) Forward prediction against HPO. The reference construction is the interesting part:

• Per gene, specify one or more primary lesion variables from gene-product identity plus disease mechanism.
• HPO classes follow the same EQ pattern, so a class is directional when its quality states an increase or decrease. PATO:0002300 → +, PATO:0002301 → −. Hypoglycemia = blood glucose + decreased concentration.
• HPOA annotations propagate upward only through the subclass hierarchy — a specific annotation contributes an ancestor's mapping, never a descendant's.
• A gene–variable direction is retained only when the resulting set is a singleton.
• The intervened variable is not scored.
• Two deoxycortisol terms were hand-blocked because their asserted HPO ancestry would otherwise map them onto cortisol.

Metrics: precision = C/(C+W), coverage = (C+W)/(C+W+A); a missing prediction and a ? both count as abstention. Two baselines run on the same axioms and interventions, so the comparison isolates the inference rule: shortest signed path (pick one shortest directed path, multiply edge signs) and signed diffusion (r = (1−α)(I−αS)⁻¹e, α = 0.85, S the column-normalized signed adjacency, sign(rᵢ) the prediction, |rᵢ| the confidence, swept for a precision–coverage curve).

(c) Abduction. A closed pool of 175 single-lesion hypotheses, always containing the true lesion. Agreement +1, contradiction −1, abstention 0; rank by net agreement, then fewer contradictions, then more agreements. Note the scoring rewards abstaining — withholding beats guessing wrong.

(d) Relation-type ablation. Both tasks re-run while cumulatively adding production, quantitative identity, and constitution to causal influence.

1.8 Reported results

v1.1.1: 1,699 traits; 2,270 causal-influence, 85 production, 4 constitution, 9 quantitative-identity, 19 modulation axioms → 1,699 nodes and 2,387 relation instances. Largest SCC 213 traits.

• Expert review: 69 accepted, 12 flagged, 2 rejected (both causal) of 83.
• Forward prediction: before exclusions, 175 determinate, 171 agreeing (97.7%). Discrepancy review found 3 of the 4 disagreements conflicted with the cited primary literature and 1 varied across studies; all 4 dropped uniformly for every method, leaving 866 adjudicated pairs. On that set: 171 determinate, 171 correct, 0 wrong, 695 abstentions — 100% precision at 19.7% coverage. Excluding the 15 highest-overlap genes left 151 determinate, all concordant.
• Baselines: shortest signed path 375 directions, 292 correct, 83 wrong (77.9%). Signed diffusion retained errors at matched prediction counts.
• The abstention result. On the 171 pairs where the solver committed, the path rule gave the same answer (0% error). On 204 pairs where the solver abstained, the path rule was right 121 times (40.7% error). An exact within-gene conditional permutation test — holding each gene's counts of path errors and abstentions fixed, permuting abstention labels within gene — rejected independence at P = 2.05 × 10⁻⁷. The claim is that abstentions concentrated the hard cases, not that the solver wins when both commit.
• Abduction: solver 40 unique top-1, 127 best-tied top-3, 161 top-10, MRR 0.707; shortest path 30/94/105/0.535; diffusion 30/99/110/0.558; chance 1/3/9/0.033.
• Ablation: the full first-order relation set changed 4 of 866 forward calls (no sign flip) and 12 of 163 inverse ranks, with no aggregate improvement.

1.9 Limits the authors themselves state

Unusually forthright, and not to be softened in retelling:

• 100% precision is at 19.7% coverage, on an adjudicated reference. Four reference directions were removed post hoc after literature review. The exclusion was applied uniformly and reported as post hoc — the right procedure — but the headline is not "PhysioMap is never wrong."
• The reference is not independent. PhysioMap curation and HPOA may draw on the same literature; the 15-gene exclusion mitigates but does not remove this.
• Abduction used a closed pool always containing the true lesion, and top-k/MRR use the best position within a tie — optimistic for every rule compared.
• Relation typing barely mattered on these tasks (the ablation above). The richer typing is justified by semantics and by what it will enable, not by measured benchmark gain.
• Proof of concept, not clinical readiness.

---

Part II — What their repository shows

2. Working artifacts

The repo contains substantially more than the paper describes, and the curation discipline on display is the part most directly transferable to dismech.

2.1 Where the content actually lives

Not in the OWL. The source of truth is hand- and agent-curated YAML under benchmarks/: 56 files in human/systems/ (one per physiological system — thyroid.yaml, iron_hepcidin.yaml, calcium_pth.yaml, respiratory_acidbase.yaml, purine_urate_metabolism.yaml, …), 6 in human/curated/ (the bulk imports), plus guyton/ and multiscale/. 2,358 causal edges across the YAML, projecting to the 2,268 in the released graph.

benchmarks/human/curated/README.md, in full:

"Merged curator contributions land here (see scripts/curation_merge.py). Each is a DRAFT FOR DOMAIN REVIEW that passed the full pre-deploy gate suite before being committed."

Every import file carries a *** DRAFT FOR REVIEW *** banner in its header.

2.2 IEM enzyme edges are Rhea-grounded, with a documented sign convention

e1a_iem_reactions.yaml — the largest content block — states its rule in the header:

# SIGN CONVENTION (comparative statics, d[metabolite]/d[enzyme_activity]):
#   enzyme_activity -(-)-> SUBSTRATE  (LoF => substrate ACCUMULATES)
#   enzyme_activity -(+)-> PRODUCT    (LoF => product FALLS)
# Physiological flux direction was curated per enzyme (can reverse the canonical Rhea
# equation, e.g. CPT2 vs CPT1A on plasma_palmitoylcarnitine -> opposite signs).

That last line is the non-mechanical part: the same Rhea equation yields opposite signs for CPT1A and CPT2 depending on which way flux actually runs in vivo. A record:

- source: cyp21a2_activity
  target: plasma_17ohp
  sign: '-'
  causal_evidence: curated_mechanistic
  mechanism: 'RHEA:50308 (EC 1.14.15.15): 17alpha-hydroxyprogesterone + reduced
    [NADPH--hemoprotein reductase] + O2 = 11-deoxycortisol + ...;
    17alpha-hydroxyprogesterone is the substrate'
  evidence: '21-hydroxylase deficiency (congenital adrenal hyperplasia), OMIM:201910
    — LoF causes massive 17-hydroxyprogesterone accumulation in plasma'

2.3 Textbook extraction is an LLM fan-out with verbatim quotes

williams_extracted.yaml describes its own method:

"DRAFT within-scale signed causal edges + variables extracted from Williams Textbook of Endocrinology, 14th ed. via per-chapter multi-agent fan-out. Signs adversarially verified; every edge carries a verbatim quote + chapter/page anchor. New nodes de-duplicated against the existing map (physiomap_core.reconcile)."

- source: type_3_deiodinase_activity
  target: free_t3
  sign: -
  causal_evidence: genetic_lof_gof
  mechanism: Type 3 iodothyronine deiodinase (DIO3) removes an inner-ring iodine from T3,
    inactivating it; overexpression (GoF) in infantile hepatic hemangiomas consumes T3
    faster than the thyroid can produce it, causing consumptive hypothyroidism
  evidence: "Large infantile hepatic hemangiomas express high D3 levels, causing
    'consumptive hypothyroidism,' because thyroid hormone is inactivated at a more rapid
    rate than it can be produced"
    — Williams Textbook of Endocrinology, 14th ed., ch 1, p 11

Verbatim quote plus page anchor — structurally the same discipline as dismech's snippet rule, applied to a book rather than an abstract. Same companion files exist for Guyton & Hall (hall_extracted.yaml) and West (west_extracted.yaml).

Some mechanism fields spell out an explicit quality→disposition→process→quality chain, e.g. for phosphorylase kinase: "quality: PHK activity → disposition: phosphotransfer competence → process: phosphorylation of Ser-14 on glycogen phosphorylase → quality: glycogen phosphorylase catalytic activity".

2.4 AOP-Wiki edges cite Key Event Relationships by UUID

- source: hepatic_t4_clearance_rate
  target: plasma_t4
  sign: -
  causal_evidence: curated_mechanistic
  evidence: AOP-Wiki KER 035747ab-b591-4283-acfd-90ba868d5fc4; AOPs: 'Upregulation of
    Thyroid Hormone Catabolism via Activation of Hepatic Nuclear Receptors...' — KER:
    increased T4 clearance from serum (liver/organ) -> decreased serum T4 (organ_system);
    do(hepatic T4 clearance rate up) -> serum T4 down (sign -) | No PMID in batch |
    *** DRAFT FOR REVIEW (imported) ***

Note | No PMID in batch | — a recorded gap rather than a silent one.

2.5 The SIGNOR import uses precursor of, not signalling

SIGNOR is a signalling resource (protein→protein activation/inhibition), so it is reasonable to wonder what it is doing supplying metabolite→metabolite edges. The answer is that the signalling content was not used at all. The file is signor_metabolism_import.yaml and describes itself as "steroidogenesis / amino-acid / pyrimidine / sterol metabolism". All 22 edges come from SIGNOR's precursor of relation — a chemical-conversion mechanism SIGNOR carries alongside its protein interactions. Every edge is CHEBI→CHEBI except one (PR:P23109 AMPD1 → CHEBI:16027 AMP).

- source: plasma_17_hydroxypregnenolone
  target: plasma_17ohp
  sign: +
  causal_evidence: curated_mechanistic
  mechanism: 'precursor of: 3beta-HSD/HSD3B isomerase converts 17alpha-hydroxypregnenolone
    (delta-5) to 17alpha-hydroxyprogesterone (delta-4)'
  evidence: 'SIGNOR-268639: ... PMID 2139411 (Lorence et al. 1990 Endocrinology 126:2493):
    resolves OK; abstract explicitly confirms ''catalyzed the conversion of
    17 alpha-hydroxypregnenolone to 17 alpha-hydroxyprogesterone'' — concordant.
    Both nodes organ_system scale. INDRA: unreachable. *** DRAFT FOR REVIEW (imported) ***'

Three practices worth copying, all visible in that one record: the SIGNOR id is kept, the PMID is independently checked to resolve and to be concordant with the claim, and a failed cross-check (INDRA: unreachable) is recorded rather than omitted.

The header also records what was discarded — "Cross-scale/duplicate/pituitary-mediation- conflicting candidates were dropped" — and the import was "validated by fan-out agents (ontology IDs vs OBO, PMID existence+concordance vs PubMed, sign, scale)."

One honest failure is preserved in place: two of the 22 edges cite NBK536726, an NCBI Bookshelf ID rather than a PMID. The file says pmid_ok=false and keeps the edge, justified as "well-established biochemistry (PAH, phenylalanine hydroxylase)". Recorded, not hidden — but it is a retained failure, and it covers the Phe→Tyr edge.

2.6 Evidence classes, as actually used

Counting causal_evidence across all 2,358 source edges:

Class	Group	Edges	Share
genetic_lof_gof	interventional	915	38.8%
curated_mechanistic	mechanistic	552	23.4%
pharmacological	interventional	319	13.5%
(none — legacy unclassified)	—	206	8.7%
mechanistic_model	mechanistic	206	8.7%
perturbation	interventional	160	6.8%
mendelian_randomization	interventional	0	0%

Mendelian randomization is defined in the paper and in physiomap_core/causal_evidence.py and used zero times. Interventional evidence covers 59% of edges.

causal_evidence.py also enumerates the refused classes with reasons — binding_only ("ChIP-seq / motif occupancy: physical binding is not a demonstrated functional effect"), coexpression, observational_association, unknown — and adds a rule not in the paper: a modulation edge always requires a class, with no equivalent of the hand-curated "no class" pass an ordinary edge may take. The stated reason is that a modulation asserts a two-factor interaction, and "a correlation of slopes is not" one.

2.7 The expert review is archived in full

benchmarks/results/expert_gold_review.tsv holds all 83 rows with verdicts and comments. The summary records reviewer (Paul N. Schofield), sample sent 2026-07-28, returned 2026-07-30, sampling seed 20260728, sha256 of both the sent template and the returned workbook, the reviewer's email Message-ID, and an integrity check confirming identity_columns_unchanged: true, 83/83 verdicts, 39 comments.

The sampling rates make the "no map-wide accuracy estimate" caveat concrete:

Type	Reviewed	In map	Rate
constitution	4	4	100%
quantitative	9	9	100%
modulation	10	19	53%
production	10	85	12%
causal	50	2,268	2.2%

(Raw verdict strings are slightly inconsistent — TRUE 68, True 1, FALSE? 11, False? 1, FALSE 2 — normalized to the published 69/12/2.)

Both rejections:

DHCR7 activity --[-]--> plasma cholestane-triol. "The edge is indirect (mediated by 7-DHC accumulation and non-enzymatic oxidation), and it's not the dominant determinant of plasma C-triol. Canonical DHCR7 readouts remain 7-DHC, 8-DHC and the 7-DHC/cholesterol ratio… Considerable disagreement and inter lab variability. Err on side of caution."

mesangial cell contraction --[-]--> renin secretion. "The source is actually the calcium signal not the contraction although they correlate. Contractile activation is kind of OK but drop the mesangial contraction."

Selected flags — these show the class of error only a physiologist catches:

melatonin --[-]--> cortisol. "Not causative. Melatonin and cortisol are in near-perfect antiphase… strongly inversely correlated. But both are driven by the same upstream pacemaker — the suprachiasmatic nucleus." A textbook confounder that had passed with an evidence class of perturbation.

GLUD1 activity --[+]--> plasma glutamate. "Sign is the wrong way round for normal conditions. GDH catalyses a reversible reaction… In vivo the flux runs predominantly in the oxidative deamination direction."

thyrotroph TSH secretion --[+]--> TSH secretion. "Tautology. No evidence."

DPYD --[-]--> plasma 5-hydroxymethyluracil. "Correct target is dihydrouracil/uracil."

CoQ10 --[+]--> complex III. "if target is electron flux then this is true but if target is enzymic activity of Complex III then its not true."

TPO/NIS organification --[+]--> T4 secretion. "Source fuses two sequential and separable steps — NIS-mediated trapping and TPO-mediated organification… the perchlorate discharge test exists precisely to distinguish a trapping defect from an organification defect."

Two flags land on the Bohr-effect modulations: the reviewer's point is that pH→SaO₂ is correct as a level effect, but under the strict cross-partial definition the sign flips depending on where you sit on the dissociation curve — "A left shift at PaO₂ ≈ 100 mmHg pushes you further onto the plateau."

The hematocrit flag is treated in §3.4, because it contradicts released content.

---

Part III — My own analysis of the release

3. Analyses not in the paper

Everything in this part is computed by me from the published artifact — the repo cloned from bio-ontology-research-group/physiomap, with figures derived from web/physiomap-1.1.1.json and benchmarks/results/. The release holds 1,699 nodes, 2,268 causal, 85 production, 4 constitution, 9 quantitative-identity and 19 modulation relations, largest SCC 213, 1,448 SCCs.

Those reconcile with the paper on every count except causal influence, where the release has two fewer than the paper reports: §1.8 gives 2,270 causal-influence axioms and 2,387 projected relation instances (2,270 + 85 + 4 + 9 + 19 = 2,387, so the paper is internally consistent), while the release sums to 2,385. A two-edge drift between a preprint and its release is small and unremarkable in itself — text and artifact are cut at different moments — but it is worth stating precisely rather than rounding to "identical", particularly in a document whose subject is provenance discipline. Every other figure below matches.

3.1 A node is an (entity, quality, context, scale, system) tuple

{ "id": "renin", "label": "plasma renin activity",
  "entity_iri": "CHEBI:50266", "quality_iri": "PATO:0001414",
  "scale": "organ_system", "system": "Cardiovascular–renal",
  "scc": 0, "in_big_scc": true, "source": "Guyton core (curated)" }

The EQ decomposition is machine-readable on every node — the thing dismech pathophysiology nodes lack, where name is free text and ontology terms live in descriptor lists describing the node's participants rather than its identity.

Entity ontologies: CHEBI 671, PR 510, GO 216, UBERON 139, CL 35, none 128. Predominantly chemicals and proteins — a metabolite-and-enzyme map far more than an anatomy map.

Quality reuse is concentrated. Three PATO terms cover 86% of traits: PATO:0000033 concentration (854), PATO:0001414 activity (389), PATO:0000161 rate (221). The tail (mass, volume, pressure, functionality…) is 235.

Scale is a 7-value enum — molecular, subcellular, cellular, tissue, organ, organ_system, organism — against dismech's 4-value biological_scale. Distribution: organ_system 933, cellular 463, organ 180, subcellular 46, molecular 43, tissue 29, organism 5. PhysioMap's molecular tier is thinner than dismech's; it is more physiological, not more molecular.

3.2 The content is dominated by metabolism, not the cardiovascular showpiece

System	Traits		Edge provenance	Edges
Metabolic / hepatic	935		IEM-enzyme connections	639
Cardiovascular–renal	197		Curated system fragment	273
Endocrine	175		Williams (endocrinology)	247
Neuro / thermal	114		Curated bridges	184
Hematologic	83		HPO gap-fill (lab analytes)	151
Respiratory / acid–base	80		Textbook extraction (A&P/…)	136
Immune / inflammation	68		Connect-isolated (curated)	111
Mineral / bone	31		West (respiratory)	106
Fluid / electrolyte	11		Guyton & Hall extraction	101
Unassigned	5		Molecular/cellular module	84
			Guyton-Hall textbook	71
			Other	45
			Phenotype-connection fan-out	43
			Guyton core (curated)	26
			AOP-Wiki import	26
			SIGNOR import	22
			Curated (G. Gkoutos)	3

The systems column is the complete 10-value enum; the provenance column is all 17 lines. Note that Guyton-derived textbook content is split across two labels — Guyton & Hall extraction (101) and Guyton-Hall textbook (71), so 172 edges in total — distinct from the 26-edge hand-curated Guyton core.

55% of traits are metabolic/hepatic and the largest single edge source is IEM enzyme connections, against 26 edges for the Guyton core the paper's figure showcases.

Two provenance lines deserve a curator's eye given the evaluation is against HPOA: "HPO gap-fill (lab analytes)" (151 edges) and "Phenotype-connection fan-out" (43). Edges added to cover HPO analytes, then scored against an HPO-derived reference, are the concrete form of the circularity the authors flag in prose. Their 15-gene leakage control is the right mitigation; it is worth knowing it is mitigating this.

3.3 Signs are nearly always committed; ? is rare and well-chosen

+ 1,265, − 982, ? 21 — 99.1% of edges carry a definite sign. All 21 ? edges are genuine non-monotonicities:

• left-ventricular end-diastolic volume (preload) → ejection fraction
• lung volume → pulmonary vascular resistance (the classic U-shaped curve)
• plasma epinephrine → total peripheral resistance (α1 vasoconstriction vs β2 vasodilation)
• arterial PaCO2 → total peripheral resistance
• free T4 → plasma triglyceride; plasma testosterone → HDL-cholesterol

Direct empirical support for R1's design note: an authored ? is a positive claim, used 21 times out of 2,268, and must not be collapsed with "not yet curated".

3.4 The complete non-causal relation sets — and a flagged inconsistency

All 4 constitutive edges (micro → macro, all +): adipose tissue mass and lean body mass constitute body weight; red cell mass and plasma volume constitute blood volume.

All 9 quantitative identities, each argument carrying its own derivative sign, with role distinguishing factor / summand / numerator / denominator / argument:

Result	Kind	Arguments (with derivative sign)
mean arterial pressure	product	cardiac output (+) × total peripheral resistance (+)
cardiac output	product	heart rate (+) × stroke volume (+)
blood volume	aggregation	red cell mass (+) + plasma volume (+)
body weight	aggregation	adipose tissue mass (+) + lean body mass (+)
hematocrit	ratio	red cell mass (+) / blood volume (−) — see below
arterial O₂ content	structural-function	Hb O₂ saturation (+), hematocrit (+), arterial PO₂ (+)
oxygen delivery	product	cardiac output (+) × arterial O₂ content (+)
arterial pH	structural-function	arterial PCO₂ (−), plasma bicarbonate (+)
alveolar PO₂	structural-function	inspired PO₂ (+), alveolar PCO₂ (−)

The last two are the Henderson–Hasselbalch and alveolar gas equations reduced to derivative signs — a good illustration of signs as an abstraction of a quantitative constraint.

The hematocrit row is not sound, and their own reviewer said so. From the archived review (§2.7):

"Identity rather than a causal relationship. Note Blood volume = red cell volume + plasma volume, so red cell mass is itself a component of the denominator. Taking the partial derivative with respect to red cell mass 'holding blood volume constant' isn't physically coherent, since increasing red cell mass necessarily increases blood volume. The correct formulation is Hct = RCV / (RCV + PV)."

The release makes this sharper than the comment alone suggests: it holds both blood_volume = red_cell_mass + plasma_volume (aggregation) and hematocrit = red_cell_mass / blood_volume with independent partial signs. The two identities contradict each other within the same file. This row was flagged 2026-07-30 and is still in v1.1.1, which is consistent with the paper's position that the review "isolated a minority for correction or further investigation" rather than that corrections had landed.

All 19 modulation edges are (modulator, edge) pairs with a mixed-derivative sign. Four reconstruct the oxyhaemoglobin dissociation curve's shift factors as modulations of the single arterial PO₂ → arterial Hb O₂ saturation edge:

[-] arterial PaCO2                    on  arterial PO2 -> Hb O2 saturation
[+] arterial blood pH                 on  arterial PO2 -> Hb O2 saturation
[-] body core temperature             on  arterial PO2 -> Hb O2 saturation
[-] erythrocyte 2,3-BPG               on  arterial PO2 -> Hb O2 saturation

Four more are cortisol's permissive effects (on glucagon→hepatic glucose production, epinephrine→glycogenolysis, epinephrine→lipolysis, norepinephrine→TPR) — permissive hormone action really is a mixed partial and really is awkward as an ordinary edge. This is the clearest argument in the release for modulation as a distinct type, even though the first-order solver does not read it. Note the reviewer flagged two of the four Bohr modulations on cross-partial grounds (§2.7).

3.5 Worked edges, by provenance

Quoted verbatim from web/physiomap-1.1.1.json, labels truncated.

Guyton & Hall extraction (101) — renal haemodynamics and acid–base:

renal afferent arteriolar resistance (RA) --[-]--> glomerular capillary hydrostatic pressure
renal afferent arteriolar resistance (RA) --[-]--> renal blood flow
plasma angiotensin II                     --[+]--> renal efferent arteriolar resistance (RE)
arterial PaCO2                            --[+]--> renal tubular H+ secretion / bicarb reabs.
arterial blood pH                         --[-]--> renal proximal tubule ammonia (NH3) secretion
arterial blood pH                         --[+]--> total peripheral resistance
blood colloidal osmotic pressure          --[+]--> glomerular capillary colloid osmotic pressure

West (106) — ventilation mechanics, essentially a textbook derivation chain:

tidal volume                     --[+]--> total (minute) ventilation
respiratory rate                 --[+]--> total (minute) ventilation
total (minute) ventilation       --[+]--> alveolar ventilation
dead space ventilation           --[-]--> alveolar ventilation
tidal volume                     --[-]--> dead space fraction
alveolar ventilation             --[-]--> alveolar PCO2
whole-body CO2 production rate   --[+]--> alveolar PCO2
alveolar surface tension         --[-]--> alveolar stability against collapse
alveolar surfactant concentration--[+]--> alveolar stability against collapse
blood-gas barrier thickness      --[-]--> rate of respiratory gas diffusion

Williams (247) — axis control, including the KNDy pulse generator:

KNDy neuron neurokinin B secretion    --[+]--> hypothalamic kisspeptin secretion
KNDy neuron dynorphin secretion       --[-]--> hypothalamic kisspeptin secretion
hypothalamic kisspeptin secretion     --[+]--> hypothalamic GnRH secretion
tuberoinfundibular dopaminergic (TIDA)--[-]--> anterior-pituitary prolactin secretion
tissue type-III deiodinase (DIO3)     --[-]--> free plasma triiodothyronine (free T3)
plasma myostatin (GDF8)               --[-]--> lean body mass
plasma melatonin                      --[-]--> suprachiasmatic nucleus (SCN) output

IEM-enzyme connections (639) — the stereotyped substrate/product shape:

fumarate hydratase (fumarase) activity          --[-]--> plasma fumarate concentration
fumarate hydratase (fumarase) activity          --[+]--> urinary malic acid level
succinate dehydrogenase complex (SDH) activity  --[-]--> plasma succinate concentration
mitochondrial aconitase 2 (ACO2) activity       --[-]--> urinary citrate level

AOP-Wiki (26) — the most molecular content in the map:

thyroid NIS (SLC5A5) sodium/iodide symporter --[+]--> follicular cell intracellular iodide
follicular cell intracellular iodide         --[+]--> follicular cell thyroxine (T4) secretion
follicular-cell DUOX2 H2O2 generation        --[+]--> thyroid peroxidase iodide organification
endothelial AKT1/eNOS-Ser1177 phosphorylation--[+]--> endothelial nitric oxide production
renal proximal-tubule OAT1 activity          --[-]--> plasma urate concentration

Cross-scale. The transition matrix is dominated by cellular → organ_system (709) and organ_system → organ_system (654). Overall 986 edges run finer→coarser, 975 within a scale, and 307 coarser→finer — genuine top-down control, not only bottom-up composition:

plasma (ionised) calcium concentration --[-]--> parathyroid chief-cell PTH secretion rate
CSF glucose concentration              --[+]--> intracellular ATP concentration
thrombin (FIIa) activity               --[+]--> activated platelet procoagulant surface

3.6 The topology: a coupled core bolted to a shallow fan

Classifying all 1,699 nodes by in/out degree over the causal edges:

Role	Nodes	Share
Pure source (in-degree 0) — candidate lesion variables	461	27.1%
Pure sink (out-degree 0) — terminal observables	710	41.8%
Internal (both in and out) — can be a waypoint in a chain	452	26.6%
Isolated (no causal edges at all)	76	4.5%

And 593 causal edges (26.1%) run directly from a pure source to a pure sink — one-hop terminal fans that can never participate in a longer chain:

CYP11B2 (aldosterone synthase) activity        --[+]--> adrenal zona glomerulosa aldosterone secretion
lysosomal acid beta-galactosidase (GLB1)       --[-]--> plasma tissue-non-specific alkaline phosphatase
hepatocyte apolipoprotein E secretion rate     --[-]--> plasma triglyceride concentration

The archetype is MCAD/ACADM activity: out-degree 25, in-degree 0 — an enzyme fanning out to two dozen analytes, caused by nothing. The next top sources are mostly the same shape (VLCAD 11, BCKDH 10, DPYD 10, transaldolase 9, CPT2 8, LCAT 8) — with one exception worth noting: Bruton's tyrosine kinase (BTK) catalytic activity also sits at out-degree 11, and is an immunodeficiency gene rather than a metabolic enzyme, so the fan pattern is not exclusively IEM.

So PhysioMap is two graphs sharing a namespace:

1. A genuinely coupled physiological core — the 213-trait SCC: 53 cardiovascular–renal, 43 endocrine, 32 respiratory/acid–base, 32 metabolic, 21 haematologic, 15 mineral/bone. Renin, angiotensin II, aldosterone, ADH, ANP, baroreflex outflow, GFR, ECF volume, cardiac output, heart rate. Feedback is real here and the Cramer's-rule machinery earns its place.
2. A broad one-hop enzyme→analyte fan — the IEM content, topologically trivial, where a "prediction" is a single edge lookup.

That split explains §3.7, and it also explains why the relation-type ablation changed almost nothing (§1.8): constitution, quantitative identity, and modulation live almost entirely in graph 1, while the benchmark mostly exercises graph 2. The authors say as much in a drug-panel benchmark comment: "Organ targets embedded in the whole-body feedback SCC return '?' (reported separately, never wrong)." PhysioMap's sophisticated inference and its broad coverage are largely in different parts of the map.

The 76 isolated nodes are traits with no causal edges at all — mostly secretion rates (intestinal L-cell GLP-1 secretion rate, kidney interstitial fibroblast erythropoietin secretion rate, gastric G-cell gastrin secretion rate). The "Connect-isolated (curated)" provenance line (111 edges) shows this is a known worklist already being chipped at.

3.7 The benchmark result is shallower than the headline suggests

Taking the 866 adjudicated pairs in benchmarks/results/e1b_forward_pairs.tsv and running BFS from each lesion variable over causal + production edges:

	Pairs	Share
Unreachable from the lesion at any depth	492	56.8%
Reachable	374	43.2%
Lesion id not resolvable	0	—

Among the 374 reachable, 161 (43.0%) are a single hop, 53.5% within two, 70.1% within four.

Parsing note. 27 of the 866 rows carry two semicolon-separated lesion variables in primary_intervention — e.g. ARSA → arsa_activity -; arylsulfatase_a_activity -, and likewise CTNS, KYNU, PKLR. These are exactly the four genes the paper excludes from the inverse benchmark for having two primary lesion variables (§1.8). Every constituent id resolves in the graph. The figures above split on ; and take a pair as reachable at the minimum distance over its lesion variables, which matches the paper's "one or more primary lesion variables" specification. Counting only the first lesion instead gives 369 reachable / 497 unreachable / 157 one-hop — the conclusions are identical either way. Two of these multi-lesion rows exist because of duplication: ARSA is one of the five duplicate-trait groups in §3.8, where arsa_activity and arylsulfatase_a_activity are the same trait under two ids.

1. Most abstentions are absence of a path, not feedback-induced indeterminacy. The solver abstained on 695 pairs; 492 have no directed route at all.
2. The determinate set is plausibly concentrated at depth 1. The solver committed on 171 pairs; there are 161 one-hop reachable pairs — consistent with determinacy being mostly the near-definitional enzyme→analyte edges of §3.2.

The reachability count corroborates itself against the paper: the shortest-path rule returns a direction exactly when a path exists, and the paper reports 375 such directions against the 374 reachable pairs found here — a one-pair gap, which is about as close as an independent reconstruction from the web payload can be expected to land.

Consistent-with, not proven. The release does not publish per-pair determinacy labels, so I cannot confirm the 171 determinate pairs are the 161 one-hop pairs plus 10; the counts permit it but do not establish it. My BFS also reconstructs reachability from the projected web payload rather than from the solver's own graph, and does not traverse quantitative-identity edges.

This does not undercut the abstention result (§1.8): that permutation test was restricted to pairs where the shortest-path rule returned a direction — i.e. reachable pairs — so unreachable pairs never entered it. What the depth analysis changes is the interpretation of the coverage figure, not the significance of the concentration result.

3.8 Other content defects visible in the release

Minor, and recorded because dismech would hit the same class of problem if it built a shared trait namespace (§4.2).

• Exact duplicate traits. Grouping the 1,568 nodes that carry both an entity_iri and a quality_iri by that pair finds 195 groups with more than one member (grouping all 1,699 nodes regardless gives 202; requiring only entity_iri gives 196 — the conclusion below is identical under all three). Nearly all are legitimate context distinctions the web payload does not expose — plasma_phenylalanine and csf_phenylalanine share CHEBI:28044 + PATO:0000033 and differ only in context, which is what the trait definition's context conjunct is for. Filtering to collisions that also share an identical label leaves 5 genuine duplicate groups (6 redundant nodes):

Label	Duplicate ids
plasma 11-deoxycorticosterone (DOC) concentration	plasma_deoxycorticosterone, plasma_doc, plasma_11_deoxycorticosterone
erythrocyte pyruvate kinase (PKLR) catalytic activity	pyruvate_kinase_activity, erythrocyte_pyruvate_kinase_activity
alpha-galactosidase A (GLA) activity	alpha_galactosidase_a_activity, gla_activity
arylsulfatase A (ARSA) activity	arylsulfatase_a_activity, arsa_activity
tissue glucosylceramide content	tissue_glucosylceramide, tissue_glucosylceramide_content

0.35% of nodes — the "reconciled duplicates" step evidently worked, with a residue. Not cosmetic under SCM semantics: each of the three DOC nodes independently carries a − edge into renin, so an intervention on "DOC" reaches a third of its edges and the feedback around renin is understated. Three of the five are ARSA/GLA/PKLR — genes that also appear in the rare-disease benchmark.

• One contradictory edge pair. Exactly one (source, target) pair carries two edges with opposing signs: plasma insulin → hepatic VLDL secretion rate, once + and once −. Biologically a real tension (acute insulin suppresses VLDL secretion; chronic hyperinsulinaemia with insulin resistance raises it), so the honest encoding is a context distinction or a ?, not two unqualified edges. One conflict in 2,268 edges is a good consistency record.

---

Part IV — Comparison with dismech

4. Structural comparison

Counts measured from this repo at c0250519 (2026-08-28), the branch's merge base.

An earlier revision of this report measured at 4b2f90c and mislabelled that column 2026-08-23; 4b2f90c is in fact dated 2026-08-11, and the KB grew substantially in the intervening fortnight (11,923 → 15,296 nodes, 22,564 → 27,532 edges). Every figure below has been recomputed at the current base so the document carries one vintage. No conclusion changed: the cyclic-file share moved 1.4% → 1.5%, the description-regex share 18.3% → 18.5%, and both self-loops are still present. The one materially different row is the gain/loss-of-function modifier census in §4.1, which was 5/4 at the older commit and is now 37/106 — no longer a rounding error, and worth knowing before citing it.

	PhysioMap v1.1.1	dismech (c0250519, 2026-08-28)
Organizing axis	one shared physiology, disease-agnostic	per-disease entries + reusable modules
Nodes	1,699 traits	15,296 pathophysiology nodes
Causal edges	2,385 projected relations	27,532 downstream edges
Node identity	global EQ-pattern OWL class	free-text name, scoped to its file
Edge polarity	signed (+/-/0/?), semantically load-bearing	absent
Edge typing	5 SCM-distinct types	causal_link_type = directness only
Feedback	first-class; SCC + Cramer's rule; largest SCC 213	not represented as feedback
Inference	intervention prediction + abduction, abstains	rendering, compliance, export; no sign propagation
Evidence model	6 classes, admission-gating, not used by solver	per-edge EvidenceItem, snippet-verified against cached sources
Cross-scale composition	constitution in a separate acyclic graph	conforms_to module anchors (manual, per-node)

4.1 dismech signs treatment and environmental edges but not mechanism→mechanism edges

dismech already commits to edge polarity in two places:

• TreatmentMechanismTarget.treatment_effect → INHIBITS, ACTIVATES, MODULATES, BYPASSES, RESTORES
• EnvironmentalMechanismTarget.environmental_effect → TRIGGERS, EXACERBATES, PREDISPOSES, PROTECTS_AGAINST, MODULATES

But CausalEdge — the 27,532-edge backbone — carries only target, description, evidence, hypothesis_groups, causal_link_type, intermediate_mechanisms. There is no slot for whether A raises or lowers B.

So a drug edge into a node is signed, an exposure edge into the same node is signed, and the mechanism edge out of it is not. Any signed traversal of a dismech pathograph dies at the first downstream hop.

Polarity is not absent from the KB, just not in a queryable slot. Node-level descriptors carry modifier heavily — DECREASED 4,795, INCREASED 4,186, ABNORMAL 2,916, DYSREGULATED 856 (plus LOSS_OF_FUNCTION 106, GAIN_OF_FUNCTION 37, ABSENT 30) — counting modifier: on the three ontology-bound descriptor lists that hang off a pathophysiology node (biological_processes, molecular_functions, cell_types) and nowhere else. That scope matters: descriptor lists elsewhere in an entry also carry modifier:, so a whole-file count is a larger number measuring a different thing, and the claim here is specifically about polarity attached to pathograph nodes. Regenerate with uv run python scripts/physiomap_release_analysis.py --dismech .. INCREASED/DECREASED are already bound to PATO:0002300/PATO:0002301, the exact two terms PhysioMap uses for +/−. A crude regex over edge description text (inhibit|suppress|decreas|reduc|block|impair|loss of|deplet|antagoni) matches 5,085 of 27,532 edges (18.5%). Treat that only as evidence that polarity language is pervasive in the prose — the regex cannot distinguish "A inhibits B" from "A causes impaired B", and the two imply opposite edge signs. A reason to look, not a measurement.

Note the modifier counts describe node states, not edge signs, and the two are not interchangeable: a node marked DECREASED says nothing about whether its outgoing edge raises or lowers its target. This is why the information cannot simply be derived.

4.2 The deeper gap is the namespace, not the sign

CausalEdge.target is range: string, resolved by convention within the same file. "Hepatic Stellate Cell Activation" in Liver_Cirrhosis.yaml and the corresponding node in fibrotic_response.yaml are related only through a hand-authored conforms_to anchor, and CLAUDE.md is explicit that conformance is not inheritance — disorder entries fully duplicate module content.

PhysioMap's traits are globally identified classes in one namespace, which is what lets 1,699 nodes form a single coupled system with a 213-trait feedback component.

Consequence: dismech's 27,532 edges do not compose into one graph. They are ~2,500 disjoint per-disease graphs joined by manual anchors. Signing the edges would make each entry traversable; it would not by itself produce a whole-body physiology. That is a much larger undertaking and should not be smuggled in as a side effect of adding a sign slot.

§3.8's duplicate-trait residue is a preview of what that undertaking costs: PhysioMap reconciles duplicates as an explicit pipeline step and still ships six.

4.3 dismech pathographs are almost entirely acyclic — which makes signing them easier

PhysioMap's solver is elaborate because feedback is unavoidable in whole-body physiology: a 213-trait SCC forces the Cramer's-rule test and the 16-trait expansion cap. Would dismech inherit that complexity? Measured over the 2,517 KB files with a pathophysiology block, resolving downstream targets within each file:

	Count	Share
Files with a genuine multi-node feedback cycle	37	1.5%
Self-loop edges (a node listing itself as downstream)	2	—
Files whose pathograph is acyclic	2,480	98.5%

The 37 cyclic files are where you would expect real feedback — Alzheimer_Disease, Chronic_Kidney_Disease, Idiopathic_Pulmonary_Fibrosis, Abdominal_Aortic_Aneurysm, Aortic_Valve_Stenosis, Epilepsy, Alopecia_Areata — self-amplifying vicious circles, not modeling accidents.

This substantially strengthens R1. For 98.5% of entries, signed propagation is plain topological sign algebra over {+, -, 0, ?} — no Jacobian, no SCC decomposition, no determinant test. PhysioMap's hard machinery would be needed for a couple of dozen entries, and its conservative fallback (abstain when unresolved) is acceptable for those in a first implementation.

The two self-loops are almost certainly curation errors and worth fixing independently:

• Periventricular_Nodular_Heterotopia.yaml — "Progressive Lung Disease" → itself
• Autoimmune_Polyendocrine_Syndrome_Type_1.yaml — "Chronic Mucocutaneous Candidiasis" → itself

Nothing currently rejects a node listing itself as its own downstream target; a one-line assertion in tests/test_data.py would close that. Flagged as an observation, not fixed here.

---

Part V — Recommendations

5. What to do about it

Ordered by value-to-cost. Only the first is proposed as near-term work.

R1 — Add an optional polarity slot to CausalEdge (recommended)

The smallest change that unlocks the most. Sketch, deliberately mirroring the existing effect enums rather than inventing a new vocabulary:

CausalEdge:
  slots:
    - target
    - causal_effect        # NEW: INCREASES | DECREASES | MODULATES | UNKNOWN
    - causal_link_type     # unchanged: directness
    ...

Design notes, now grounded in the release rather than just the paper:

• Keep it optional and default to absent, not to UNKNOWN. PhysioMap's ? means "non-constant dependence with no sign constraint" — a positive claim. §3.3 shows they used it 21 times in 2,268 edges, on genuine non-monotonicities like preload→EF and lung volume→PVR. That is not "nobody has curated this yet", and conflating the two turns a coverage gap into a false knowledge claim.
• Do not derive it from the target node's modifier. Node state and edge sign are different quantities (§4.1).
• Sign is a separate claim, so it takes separate evidence — the discipline dismech already applies to phenotype frequency: and to model-link vs. readout evidence. PhysioMap goes further and makes evidence class an admission gate (§1.6/§2.6): associational evidence cannot license a causal edge at all. That is worth considering for a signed dismech edge.
• Retrofitting 27,532 edges is not proposed. Curate forward and backfill opportunistically; a mechanically-inferred backfill from description text would be exactly the plausible-looking fabrication the repo's evidence SOP exists to prevent.

Before implementing, confirm no existing issue covers this. My GitHub semantic search returned zero results on two phrasings, which I read as inconclusive (the queries also failed to surface anything adjacent) rather than as confirmation none exists.

R2 — Adopt principled abstention wherever dismech infers a direction

The most valuable result in the paper is not the 100% precision, it is that abstentions were provably concentrated on cases a naive path rule got wrong 40.7% of the time. Any future dismech feature that propagates direction along the pathograph — a signed-traversal query, a treatment-response predictor, a compliance heuristic — should be able to return "undetermined" and be evaluated on precision and coverage, never precision alone.

R3 — Reuse the HPO-EQ directional reference construction (methodology, no schema change)

Their recipe: decompose directional HPO classes into (variable, direction) via the EQ pattern and PATO:0002300/0002301, propagate annotations only upward through the subclass hierarchy, retain a gene–variable direction only when the resulting set is a singleton. dismech already binds those two PATO terms and relates model variables to HP terms with threshold_direction. Directly reusable if dismech ever wants to validate a signed pathograph against HPOA — including both guardrails. Note they had to hand-block two deoxycortisol terms whose asserted HPO ancestry would have mis-mapped them to cortisol: the recipe needs spot-checking, not blind trust.

R4 — Steal the curation-provenance practices, not the relation types

The relation types are semantically well-motivated but changed almost nothing on the evaluated tasks (§1.8), and dismech has no solver that would read them. Adding four more relation types now would add curation burden with no consumer. Revisit if R1 lands and a signed-traversal consumer exists.

What is worth taking now is §2's provenance discipline, most of which dismech either already does or could adopt cheaply:

• Record failed checks in place. INDRA: unreachable, pmid_ok=false, No PMID in batch — recorded on the edge rather than causing a silent drop.
• Record what was discarded and why ("cross-scale/duplicate/pituitary-mediation-conflicting candidates were dropped").
• Check the citation resolves and is concordant with the claim, as two separate checks.
• Archive human review as a first-class artifact — sha256 of the sent and returned workbooks, the sampling seed, an integrity check that identity columns were not edited. dismech's history/ records are the natural home for this shape.
• A *** DRAFT FOR REVIEW *** banner that survives into the committed file, marking content that passed automated gates but not yet domain review.

The one exception on relation types: quantitative identity (MAP = CO × TPR) has an obvious dismech analogue already in the tree — ComputationalModel / ModelVariable / ModelVariableDescriptor, where variables carry units, LOINC/CHEBI mappings, and thresholds. If dismech ever links a pathograph node to a model variable, that link is closer to a quantitative identity than to a causal edge and should not be recorded as one. §3.4's hematocrit case is the cautionary example: an identity whose arguments are not independent is incoherent under partial-derivative semantics, and it shipped anyway.

R5 — Not recommended: mirroring PhysioMap content into dismech

PhysioMap is CC BY 4.0 and technically ingestible, but its content is normal physiology, disease-agnostic. dismech entries are disease-scoped pathographs with per-edge, snippet-verified evidence against cached references. There is no clean way to import a PhysioMap trait into a dismech disorder file without either inventing disease-specific evidence for a disease-agnostic claim, or creating nodes no entry owns. If a bridge is ever wanted, the right shape is a mapping (dismech node ↔ PhysioMap trait IRI), not a copy — and that presupposes R1 and the §4.2 namespace work.

---

6. Open questions

1. Is edge polarity already tracked in a dismech issue? (search inconclusive — needs a human check)
2. Would causal_effect be better on CausalEdge, or as a fourth mechanism-link class alongside TreatmentMechanismTarget / EnvironmentalMechanismTarget / ModelMechanismLink? The existing three all sign their edges; CausalEdge is the odd one out, which weakly argues for putting it on CausalEdge itself.
3. ~~Does the dismech pathograph have feedback cycles today?~~ Answered in §4.3: 37 of 2,517 files (1.5%), plus 2 self-loops that look like curation errors.
4. Should a self-loop downstream edge be a test failure? (§4.3 — two exist today.)
5. conforms_to anchors are the only existing cross-file node identity. Are they dense enough to seed a shared trait namespace, or is §4.2 genuinely a from-scratch effort?
6. Worth asking the PhysioMap authors for the per-pair determinacy labels (which of the 866 pairs the solver committed on). That single file would settle §3.7 — whether the 171 determinate predictions are essentially the one-hop enzyme→analyte edges, or whether the solver is genuinely resolving multi-hop and feedback cases. The question is friendly, not adversarial: the release is unusually complete and this is the one artifact missing to reproduce the coverage claim end to end.

7. Provenance

Part I is the preprint, read in full (16 pp., text extracted locally with pypdf). No supplementary material was retrieved; statements attributed to Suppl. sections are as described in the main text. Figures quoted from the paper — the expert-review counts, the baselines, the abduction table, the permutation test — are as printed and were not recomputed. The one place Part I departs from the paper is the Figure 1 edge signs, which were verified against the release because PDF extraction garbles figure layout.

Part II is the repo at github.com/bio-ontology-research-group/physiomap, cloned 2026-08-23: benchmarks/human/curated/*.yaml (the import files and their headers), benchmarks/human/systems/*.yaml, physiomap_core/causal_evidence.py (the class definitions and admission rules), and benchmarks/results/expert_gold_review.tsv + expert_gold_review_summary.json (the 83 verdicts, comments, and review metadata). Evidence- class counts are computed over all 2,358 causal_edges in the benchmark YAML. Reviewer comments are quoted verbatim, lightly truncated where marked.

Part III is reproducible: scripts/physiomap_release_analysis.py regenerates every figure in it from a PhysioMap clone (uv run python scripts/physiomap_release_analysis.py <clone>). That script exists because the first revision of this report got §3.7 wrong — it read primary_intervention as a single identifier when 27 of the 866 rows carry two semicolon-separated lesion variables — and a committed script makes that class of error checkable rather than trusted. It follows the precedent in experiments/README.md: scripts that compute metrics are committed alongside so numbers can be regenerated rather than trusted.

Part III computes from web/physiomap-1.1.1.json (node/edge inventories, sign and provenance distributions, scale/system/PATO/entity breakdowns, the complete constitutive/quantitative/modulation sets, duplicate-trait and sign-conflict checks, and the degree/topology analysis) and benchmarks/results/e1b_forward_pairs.tsv + e1b_forward.json. The §3.7 depth analysis is my own BFS over causal + production edges; its limits are stated inline. The release's headline counts reproduce the paper's to within the two-edge causal drift noted in §3, which is a good integrity signal.

Part IV figures are measured from this repo at commit c0250519 (2026-08-28), the branch's merge base, and regenerate with scripts/physiomap_release_analysis.py --dismech . (the same script as Part III), which carries the directory list, the modifier-census scope, and the negative-language regex as named constants. Computed by walking kb/{disorders,modules,comorbidities}/*.yaml; the §4.3 cycle analysis resolves each downstream target against pathophysiology[].name within the same file, so it measures intra-entry structure only. Schema claims verified against src/dismech/schema/dismech.yaml (CausalEdge, TreatmentEffectEnum, EnvironmentalEffectEnum, CausalLinkTypeEnum, ModifierEnum, ModelVariableDescriptor).