Skip to content

Boolean Modeling and dismech Pathographs (2026-08-28)

Scope. Three questions:

  1. What is the state of the art in Boolean/logical modeling of disease, and what is the Disease Maps community actually doing with it?
  2. Is Boolean modeling restricted to SIGNOR-style protein–protein / signaling models, or can it carry the broader, multi-scale node vocabulary dismech pathographs use?
  3. Should Boolean models simply be another model_type under computational_models, alongside ABMs, ODEs and SBML — or do they need separate treatment?

Companion documents: reports/computational-model-execution-landscape-2026-08-01.md (the COMBINE/BioSimulators/CoLoMoTo survey this builds on), explanation/computational-models.md (the in-repo dismech-perturb ODE runner), and pathographs.md.

0. Answers up front

On question 2 — no, Boolean modeling is not restricted to PPI/signaling networks. The formalism places no constraint on what a node denotes. Published disease Boolean models routinely mix molecular species with cell populations, physiological states, environmental exposures and clinical phenotypes in a single network. The canonical MaBoSS prostate-cancer model has nine input nodes that are mostly not proteins at all (Nutrients, Hypoxia, Acidosis, Carcinogen presence) and six output nodes that are phenotypes rather than molecules (Proliferation, Apoptosis, Invasion, Migration, Metastasis, DNA repair). Multiple-sclerosis multiscale networks carry immune cell populations (Th17, Th1, CD8 subsets, memory B) as nodes, and connect them upward to retinal damage and disability. That node vocabulary is close to a one-to-one match for dismech's own — which is unsurprising, since both are abstractions of the same causal biology.

So the node vocabulary is not the obstacle. The obstacles are edge polarity and network topology, and dismech currently has a problem with both (§2).

On question 3 — Boolean models are already a category, and that is the right answer for half the problem, but it is the wrong frame for the other half. There are two distinct objects that both get called "Boolean modeling", and conflating them is the main design risk here:

Boolean model as an artifact Boolean semantics as an interpretation of the pathograph
What it is A specific published or authored network with named nodes and update rules A view over curated dismech content, mechanically derived
Where it belongs computational_models entry, model_type: BOOLEAN_NETWORK — exactly like an ABM or ODE An exporter, next to CX2 and KGX in src/dismech/export/
Analogue in repo models/urate_homeostasis.xml + .config.yaml pathographs/MONDO_*.json, output/kgx/*.jsonl
Curated by hand? Yes — a curator cites or authors it No — it is derived; curating it into YAML would be duplicating the pathograph
Status today Enum value exists, one inert entry Does not exist

The user's instinct — "treat Boolean models separately, categorize them under computational models the same way we treat ABMs, ODEs, SBML" — is correct for the left column and already implemented: ComputationalModelTypeEnum.BOOLEAN_NETWORK has existed all along. What is missing there is not a category but plumbing (§4).

The right column is the more interesting and more dismech-specific opportunity, and it is explicitly not a computational_models entry. A compiled Boolean model of a pathograph is a derived artifact of the same kind as the CX2 export. Putting it in the KB YAML would mean hand-maintaining a second copy of the causal graph.

1. State of the art

1.1 The Disease Maps pipeline is the closest precedent

The Disease Maps community has a well-established route from a curated static map to an executable model, and it is the single most relevant prior art for dismech:

CellDesigner map → CaSQ → SBML-qual → GINsim / Cell Collective / MaBoSS.

CaSQ ("CellDesigner as SBML-qual") infers parameter-free preliminary Boolean rules from network topology plus semantic annotations, with no hand-written update functions, and retains the source map's references, annotations and layout in the generated model. That last property is what makes the output reviewable rather than a black box, and it is the property dismech would most want to replicate.

Three flagship applications:

This is mainstream in that community, not a niche. At DMCM 2025 the programme includes Adrien Rougny on "A query-driven framework for constructing Boolean networks from disease maps: application to Parkinson's disease", Philippe Castera on "Boolean modeling of immune responses to vaccines", Othmane Hayoun-Mya on drug synergies through multiscale modeling, and a Joaquín Dopazo keynote on causal modeling of disease maps for target discovery. The query-driven framing in Rougny's talk is notable: it treats Boolean-network construction as a query over a map rather than a one-shot conversion — which is closer to what a dismech exporter would be than CaSQ's batch translation is.

1.2 Tooling is consolidated and cheap to adopt

The CoLoMoTo consortium ships ~20 tools (GINsim, bioLQM, MaBoSS, Pint, Cell Collective, pyBoolNet, mpbn) in one Docker image with a Jupyter interface and a Python API per tool; the 2025 Interface Focus tutorial is the current reference. bioLQM converts SBML-qual to every other tool's native format. Our own execution-landscape report already classed this as Tier 1 — laptop, seconds: attractor computation on FA/BRCA-scale networks is trivial. There is no HPC story to worry about at dismech's network sizes.

SBML-qual is the interchange standard (L3 package spec, paper): QualitativeSpecies with initialLevel/maxLevel, and Transition elements carrying function terms. Critically it is multi-valued, not merely Boolean — a species may have maxLevel > 1, with successor states differing by at most 1 per step. That matters for dismech because ModifierEnum already distinguishes graded states (INCREASED/DECREASED) from qualitative ones (GAIN_OF_FUNCTION/LOSS_OF_FUNCTION), and a three-level species (0 = absent, 1 = normal, 2 = elevated) expresses that far more naturally than a Boolean.

1.3 Semantics: most-permissive is the right fit for dismech's abstraction level

This is the most technically important point in the survey.

Classical Boolean semantics (synchronous, asynchronous) are known to miss behaviors that a quantitative system compatible with the same logic would show. Most Permissive Boolean Networks (Paulevé et al., Nat Commun 2020) fix this: MPBN carries a formal guarantee not to miss any trajectory achievable by any quantitative refinement of the network, specifically including transitions arising from heterogeneous time scales and concentration scales. Reachability is also dramatically cheaper to decide than in the classical modes.

Why this matters here: 4,695 dismech backbone edges are explicitly typed INDIRECT_UNKNOWN_INTERMEDIATES and another 3,994 INDIRECT_KNOWN_INTERMEDIATES. A dismech causal edge is, by construction, a coarse abstraction over omitted intermediates with unknown kinetics. That is precisely the class of object MPBN was designed to over-approximate soundly. Under asynchronous semantics, a compiled dismech model would be making timing claims the curation does not support; under most-permissive semantics it makes only the claims the curation does support. Any pathograph-compilation work should target MPBN (mpbn/bioLQM) as its default semantics, and should say so explicitly.

1.4 Multiscale coupling is a solved-enough problem

PhysiBoSS 2.0 embeds MaBoSS Boolean networks as the intracellular layer inside PhysiCell agent-based simulations — Boolean node states drive agent phenotype decisions, and the microenvironment feeds back into Boolean input nodes. This is directly relevant because dismech already catalogues nine AGENT_BASED PhysiCell models, and because it demonstrates the composition pattern our own landscape report identified as the endpoint: "a molecular ODE node feeding a cellular Boolean node feeding a tissue-scale ABM". It also validates the taxonomy question — the field does treat Boolean as a peer modality to ABM and ODE, and composes them.

2. What dismech has today (measured 2026-08-28)

Whole-KB scan of kb/disorders/ + kb/modules/:

Quantity Value
Files with a pathophysiology block 2,516
Pathophysiology nodes 15,288
downstream (CausalEdge) edges 27,525
Median pathograph size 5 nodes, 5 edges
Largest pathograph 116 edges (Fanconi_Anemia)
Targets with exactly one incoming edge 22,578 / 24,653 (91.6%)
Targets needing an AND/OR decision (>1 in-edge) 2,075 (8.4%)
Pathographs containing any feedback loop 39 / 2,351 (1.7%)
Nodes with a sign-bearing descriptor modifier 7,041 (46.1%)
Nodes with only sign-neutral modifiers (ABNORMAL/DYSREGULATED) 2,118 (13.9%)
Nodes with no modifier at all 6,129 (40.1%)
Nodes whose name matches a directional lexicon 6,377 (41.7%)
Nodes tagged biological_scale 6,720 (44.0%)
Edges carrying hypothesis_groups 1,715 (6.2%)
computational_models objects across the KB 67 in 26 entries (71 in 28 after the stage-3 curation below)
…of which model_type: BOOLEAN_NETWORK 1 at survey time (Fanconi anemia, FA/BRCA, PMID:22267503); 5 after stage 3

2.1 The two structural findings that shape everything else

Finding 1 — CausalEdge has no sign. The backbone edge class carries target, description, evidence, hypothesis_groups, causal_link_type and intermediate_mechanisms. causal_link_type encodes directness (DIRECT / INDIRECT_KNOWN_INTERMEDIATES / …), not polarity. Meanwhile the peripheral link classes already have exactly the vocabulary a logical model needs: TreatmentMechanismTarget.treatment_effect (INHIBITS/ACTIVATES/…) and EnvironmentalMechanismTarget.environmental_effect (TRIGGERS/EXACERBATES/PROTECTS_AGAINST/…). The causal spine is the one unsigned layer in the graph. That asymmetry is the central schema fact.

Sign is not absent from the KB — it is encoded in node names and node descriptor modifiers ("Calcitriol Deficiency", "Decreased BMD", modifier: INCREASED) rather than on edges. That convention works for human readers and for narrative rendering. It does not survive mechanical compilation, and it is silently fragile: a "Decreased X" node feeding a "Reduced Y" node is a double negative that no validator currently checks.

Finding 2 — pathographs are trees, not networks. 91.6% of targets have in-degree 1 and only 1.7% of pathographs contain any feedback loop at all. This is the finding that should temper expectations most, and it needs stating plainly:

A Boolean model compiled from a typical dismech pathograph today would be dynamically trivial. A DAG has exactly one fixed point per input configuration. No multistability, no oscillation, no interesting attractor landscape, no drug-synergy prediction. Every headline result the field gets from Boolean models — the RA drug synergies, the PD subtype separation, the MaBoSS phenotype probabilities — comes from feedback loops and combinatorial logic, and dismech has almost none of either.

This is not a reason to abandon the idea. It is a reason to be precise about what the near-term payoff actually is (§5), and to recognise that the causal-spine curation style — short linear chains from etiology to phenotype — is a deliberate and reasonable modeling choice that happens to be orthogonal to what makes Boolean dynamics informative.

2.2 What is already in place and reusable

  • ComputationalModelTypeEnum.BOOLEAN_NETWORK exists. No new enum needed.
  • ModelMechanismLink is shared across experimental_models, animal_models and computational_models, and already links a model to a pathophysiology node with relationship / fidelity / limitations / readouts. A Boolean model needs nothing new here.
  • perturbations is typed GeneDescriptor — which is a better fit for Boolean models (node knock-out / knock-in are the native Boolean perturbation) than it is for the ODE models it was written for.
  • hypothesis_groups on CausalEdge is a ready-made selector for compiling alternative logical models from one pathograph, at zero schema cost (§5.3).
  • biological_scale (44% tagged, MOLECULAR/CELLULAR/TISSUE/ORGANISM) is the natural axis for deciding which layer of a composed model a node belongs to.
  • The models/<model_id>.{xml,ant,config.yaml} convention plus perturb/__main__.py:_find_model_config gives a working pattern for attaching an executable artifact to a computational_models entry via model_id.

2.3 What is missing

  1. model_format is a free-text string, 34/67 unset, with 22 distinct prose values in use. Any design that keys off model_format: SBML-qual needs the range tightened first.
  2. No slot for an in-repo model artifact path — the ODE path gets away with a filename convention off model_id.
  3. ModelVariableDescriptor.threshold is float-valued and extract_model_variables hard-requires threshold + threshold_direction + an HP: term. For a Boolean model the node is the phenotype; threshold: 0.5, threshold_direction: above is a semantic hack that should not be enshrined.
  4. The one existing BOOLEAN_NETWORK entry is inert — no model_format, no model_id, no repository_url, no modeled_mechanisms. It is a bare literature citation, and is exactly the "disconnected list entry" failure mode CLAUDE.md warns about. (Resolved 2026-08-28: it now carries two modeled_mechanisms links — see the stage-3 note below.)

3. Design position

Boolean modeling enters dismech along two independent tracks that share almost no machinery. Do not merge them.

  • Track A — Boolean models as catalogued and executable artifacts. Low risk, follows the ODE precedent exactly, unblocks curation immediately. This is the "same as ABM/ODE/ SBML" framing, and it is right.
  • Track B — pathograph → logical-model export. Higher value, higher risk, requires one schema change, and its near-term payoff is validation and reachability queries, not simulation.

Track A should not wait for Track B, and Track B's schema change is worth making on its own merits regardless of whether any Boolean model is ever compiled (§5.1).

4. Track A — Boolean models as computational_models entries

The goal is that a curator can cite the RA-FLS model or the FA/BRCA model, and that where an SBML-qual file exists we can actually run it and map attractors onto HP phenotypes, using the same ModelVariable machinery the ODE path uses.

A1. Tighten model_format to an enum. Introduce ModelFormatEnum with at least SBML, SBML_QUAL, SBML_FBC, CELLML, ANTIMONY, BNET, GINML, MABOSS_BND_CFG, PHYSICELL_XML, ONNX, MATLAB, OTHER. Migrate the 33 populated free-text values; this is the mechanical backfill item 1 of the execution-landscape report already asked for. Without this, "is this model runnable and by what engine" stays unanswerable.

A2. Adopt the artifact convention for Boolean models. models/<model_id>.sbmlqual.xml as the canonical stored form (bioLQM converts to everything else), with .bnet permitted as the human-editable source in the way .ant is for ODE models. No new slot needed — reuse model_id.

A3. Relax extract_model_variables for discrete models. Allow a ModelVariable whose dataset_identifier names a Boolean node to map to an HP term with no threshold, and add threshold_kind: DISCRETE_STATE (or equivalent) so the intent is explicit rather than encoded as 0.5 / above. Note threshold_kind already exists as a derived field in perturb/results_export.py but not in the schema — promoting it is a small, well-motivated change.

A4. Add a CoLoMoTo execution path to dismech.perturb. Mirror the structure of the proposed COBRApy path (item 3 of the landscape report, which established the precedent that a new execution class needs no new schema): load SBML-qual via bioLQM/mpbn, apply perturbations as node fixings, compute attractors under most-permissive semantics, read the output-node states, and emit the same exports/model_runs/<model_id>.json shape the ODE path emits. Scenarios come from models/<model_id>.config.yaml as today, with gene_effects mapping a gene to a node fixing (LoF → 0, GoF → 1) instead of to a parameter multiplier — a natural fit for the existing GeneEffect dataclass.

A5. Repair and extend the FA/BRCA entry as the pilot. It is already curated, the network is small, and the paper is a CoLoMoTo-era Boolean model. Add model_format, repository_url, modeled_mechanisms linking to the Fanconi anemia pathograph's DNA-repair nodes, and — if the published network can be obtained or reconstructed — a models/ artifact. Then add 3–5 more from the Disease Maps corpus (RA-FLS, a COVID-19 C19DMap submodel, the PD-map cohort models), which are all deposited and citable.

None of A1–A5 requires a change to Disease, Pathophysiology or CausalEdge.

5. Track B — pathograph → logical model

5.1 The one schema change: sign on CausalEdge

Add a polarity slot to CausalEdge, mirroring the existing and well-established treatment_effect / environmental_effect precedent:

  causal_effect:
    description: >-
      Whether the source mechanism promotes or suppresses the target mechanism's state.
      Note that dismech node names frequently already encode a direction ("Decreased Bone
      Mineral Density"); this slot describes the influence of source on target as named,
      so an edge onto a negatively-named node is normally PROMOTES.
    range: CausalEffectEnum

  CausalEffectEnum:
    permissible_values:
      PROMOTES:     # source state increases the likelihood/degree of the target state
      SUPPRESSES:   # source state decreases it
      REQUIRED_FOR: # target cannot occur without source (necessity, not just promotion)
      UNKNOWN:

Why this is cheap. Because dismech encodes sign in node names, the overwhelming majority of backbone edges are PROMOTES — the edge "PTH Excess → Decreased BMD" is a promoting edge onto a negatively-named state. So the backfill is mostly a default, and genuine curation effort falls only on the minority of genuinely suppressive edges (compensatory mechanisms, negative feedback, protective responses). That is a much smaller ask than "sign 27,525 edges".

Why it is worth doing anyway, independent of Boolean modeling. An unsigned causal edge is lossy for the KGX/BioLink export too — BioLink distinguishes causes from predicates carrying directionality, and the environmental exporter already special-cases PROTECTS_AGAINST to reach biolink:associated_with_decreased_likelihood_of. The backbone cannot currently express the same thing. Sign also enables a double-negation lint: flag a PROMOTES edge whose source and target descriptor modifiers are both DECREASED, which today is an undetectable curation error.

REQUIRED_FOR is included deliberately: it is the distinction that lets a compiler emit AND rather than OR at the 8.4% of targets with multiple in-edges, and it corresponds to the necessary/sufficient logic-type distinction that the causal-logic-inference literature uses to reduce manual Boolean-model construction effort.

Per the process in explanation/design-decisions.md, this needs an issue, maintainer sign-off, and a design-decision entry in the same PR as the schema change.

5.2 The exporter

src/dismech/export/logical_export.py, sitting alongside cx2_export.py and kgx_export.py, reusing graph.build_causal_graph so it inherits every edge kind for free (treatment targets, environmental triggers, biochemical readout, genetic contributes_to) rather than reading YAML afresh.

Compilation rules, in brief:

  • Each graph node → one QualitativeSpecies. Default Boolean; allow maxLevel: 2 where a node carries graded INCREASED/DECREASED modifiers.
  • Node role from in/out degree: zero in-degree → input node (etiology, exposure, genotype); phenotype-typed nodes → output/read-out nodes, directly mirroring the MaBoSS input/output convention (§1.1). This is where the answer to question 2 becomes concrete: dismech's node typing already is the input/phenotype-output structure that published Boolean disease models hand-build.
  • Edge → term in the target's transition function, sign from causal_effect.
  • Multiple in-edges → OR by default; AND over the REQUIRED_FOR subset.
  • treatment_effect: INHIBITS → a suppressing input node, giving in-silico treatment perturbation for free on any pathograph with curated target_mechanisms.
  • Emit SBML-qual, so the entire CoLoMoTo stack applies with no bespoke tooling.
  • Carry evidence: like CaSQ, retain PMIDs and node/edge provenance as SBML annotations so the generated model is reviewable against the curation it came from.
  • Declare most-permissive semantics (§1.3) in the export and in any analysis.

5.3 What this buys us in the near term — and what it does not

Given §2.1 Finding 2, be honest about the payoff ordering:

  1. Consistency checking (real, immediate). Compiling to a logical model turns curation defects into type errors: double negations, edges whose sign contradicts the target's modifier, phenotype nodes unreachable from any etiologic input, nodes that no input can ever activate. This is analogous in spirit to check-entity-refs and would catch a class of error nothing currently detects.
  2. Reachability and perturbation queries (real, immediate). "If treatment T inhibits node X, which phenotype nodes become unreachable?" is answerable on a DAG, is cheap under MPBN, and is exactly the question the curated treatment.target_mechanisms edges were built to support. This works without feedback loops.
  3. Competing-hypothesis comparison (real, and distinctively dismech). hypothesis_groups lets us compile one logical model per MechanisticHypothesis from the same pathograph and ask which hypothesis's model reaches the observed phenotype set. No other disease-map resource has hypothesis-tagged edges to do this with. Only 6.2% of edges are tagged today, so this starts as a pilot on the entries that are.
  4. Attractor landscapes, multistability, drug synergy (not yet). These need feedback loops. 1.7% of pathographs have any. Claiming these as near-term deliverables would be overselling.

A useful side effect: running the exporter across the KB produces a ranked list of pathographs that would benefit from feedback curation — the entries where the literature plainly describes a vicious cycle that the current linear chain does not capture. That is a concrete curation-quality signal derived from the modeling work, even before the modeling work pays off.

6. Staged plan

Stage Work Depends on Payoff
0 Open the design-decision issue for CausalEdge.causal_effect; tag @cmungall — Unblocks Track B
1 A1 ModelFormatEnum + backfill; A5 repair the FA/BRCA entry — Fixes the known model_format mess; the one Boolean entry stops being inert
2 A2 artifact convention, A3 discrete thresholds, A4 CoLoMoTo runner 1 Boolean models become runnable in-repo, joining the four ODE models
3 ✅ Curate 3–5 Disease Maps Boolean models (RA-FLS, C19DMap submodel, PD-map) with modeled_mechanisms — Real content; validates the pipeline against externally published models
4 Schema: causal_effect + CausalEffectEnum; double-negation lint; default backfill 0 Signed causal spine — valuable for KGX/BioLink independently
5 logical_export.py → SBML-qual, MPBN semantics; consistency + reachability checks over the KB 4 Curation-defect detection; treatment-perturbation queries
6 Hypothesis-group model comparison pilot on the 6.2% tagged edges 5 The distinctive research contribution
7 (exploratory) feedback-loop curation worklist; PhysiBoSS-style composition using biological_scale 5 Where genuine Boolean dynamics would start to pay off

Stage 3 status — done (2026-08-28)

Five BOOLEAN_NETWORK models are now curated and pathograph-linked, ahead of stages 1–2 (the curation does not depend on the ModelFormatEnum or the runner):

Entry Model PMID Links
Rheumatoid_Arthritis RA-FLS large-scale Boolean model 37454172 3
Parkinsons_Disease PD map cohort-specific probabilistic Boolean models 39429779 2
COVID-19 C19DMap SBML-qual model collection (FAIRDOMHub 714) 34664389 1
COVID-19 Type 1 interferon signalling Boolean model 38414974 2
Fanconi_Anemia FA/BRCA pathway Boolean model (A5 repair) 22267503 2

All five record model_format: SBML-qual except the 2012 FA/BRCA model, where the format and repository are left unset with the reason recorded in notes rather than guessed. Four of the five carry a repository_url. Ten modeled_mechanisms links were added in total, four of them PARTIALLY_RECAPITULATES with LOW/MODERATE fidelity where the Boolean model addresses a molecular pathway rather than the tissue- or organism-scale node it is linked to — the honest grading matters more here than link count.

This also produced the first worked instances of the §5.2 claim that dismech node typing already matches the input/output structure of published Boolean disease models: the type 1 IFN model's four output nodes (viral replication, antiviral response, inflammation, IFNA1 secretion) and the RA-FLS model's five phenotype submodels are curated as readouts on the links, grounding phenotype-level Boolean outputs against dismech pathophysiology nodes.

Follow-on: ModelMechanismLink.model_scale (2026-09-02)

The stage-3 curation surfaced a gap the schema could not express. Four of the ten links are PARTIALLY_RECAPITULATES because the Boolean model addresses a molecular pathway while the node it attaches to sits at tissue scale — but that caveat existed only as prose in limitations, and fidelity compresses it into a tier that also absorbs species divergence, expression level and everything else, so LOW never says which problem it is.

ModelMechanismLink now carries an optional model_scale, ranged on the existing BiologicalScaleEnum, recording the scale the model observes. Compared against the target's own biological_scale the gap becomes derivable (just model-scale-audit) rather than stored — and directional, because a model below its target's scale is extrapolating upward and cannot observe the outcome it is cited for, whereas a model above it contains that scale and is unremarkable.

On the ten pilot links the derived gap independently reproduced the hand-assigned fidelity tiers: every 2-step upward extrapolation had been graded LOW, every aligned link MODERATE. The Fanconi anemia links are the informative negative — both ALIGNED, one still PARTIALLY_RECAPITULATES, because its caveat is pathway-activation-versus- recombination-fidelity rather than scale. The slot is orthogonal to relationship and fidelity, not a restatement of them.

Recorded as enacted in design decision 3b; it needs an issue and maintainer sign-off before it is enacted. 80.5% of the KB's 1,131 model→mechanism links already have a scale-tagged target, so the comparison becomes computable across the KB as model_scale is populated.

Follow-on: typed divergences (2026-09-02)

model_scale answered one question — is the model observing below its target's scale — and immediately exposed the limit of answering it by derivation alone. The Parkinson's link is a scale gap and a different-quantity gap; the Fanconi anemia link is PARTIALLY_RECAPITULATES with no scale gap at all. Neither is recoverable from an enum comparison.

So ModelMechanismLink also carries divergences: a multivalued list, each entry naming a kind from ModelDivergenceTypeEnum, explaining in curator prose why that kind applies here, and optionally recording materiality — whether it bears on this link's claim.

The taxonomy is evidenced rather than invented: all 50 computational-model limitations strings in the KB were read and clustered, with the animal and NAM sets probed to see which kinds are shared. Full survey and the ASME V&V 40 / FDA prior art in the taxonomy spec; the decision is recorded as enacted in 3c.

The single most useful value is PROXY_QUANTITY — a model reporting a different quantity at the same scale. It is what the scale comparison structurally cannot see, and it is the reason both of this tranche's INVALIDATING divergences exist.

Stages 1–3 are Track A and can proceed immediately and independently. Stage 4 is worth doing on its own merits. Stages 5–7 are where the research value is, and are gated on 4.

7. Risks and open questions

  • The trivial-dynamics risk (§2.1) is the main one. Mitigation: frame Track B's stage-5 deliverable as validation and reachability, not simulation, and let the feedback-curation worklist (stage 7) be the path to dynamics rather than assuming it.
  • Sign-in-node-name vs sign-on-edge is a genuine modeling ambiguity. The proposed causal_effect semantics ("influence of source on target as named") must be stated unambiguously in the slot description and in CLAUDE.md, or curators will split evenly on how to sign an edge into a "Decreased X" node. The double-negation lint is the safety net.
  • Do we ever want multi-valued rather than Boolean? SBML-qual supports it and ModifierEnum's graded/qualitative split maps onto it well. Recommendation: emit Boolean first, keep maxLevel in the exporter's design so the door stays open.
  • Should modules carry logical models? kb/modules/ is where reusable mechanism structure lives and is the most plausible home for curated feedback loops (a fibrotic response is a vicious cycle). A module-level logical model that conforming disorders inherit is attractive — but note the modules design is explicitly not DRY, so this would need care to avoid contradicting the conformance model.
  • Naming. "Boolean" vs "logical" vs "qualitative": the field uses logical model as the superset (Boolean + multi-valued). Prefer logical_export.py and "logical model" in prose, reserving BOOLEAN_NETWORK for the existing enum value.

Sources

Reviews and methodology

Semantics

Standards and tooling

Disease Maps applications

Non-molecular nodes and multiscale composition