Computational-model curation backlog (2026-08-22)

Executive summary

The raw absence count is not a useful work queue. On the main snapshot used for this audit, the knowledge base contains 2,140 disorder files:

17 have one or more curated computational_models entries;
6 explicitly carry computational_models: [];
2,117 have no computational_models field.

The resumed curation series has eleven open, one-disorder PRs. If all merge, 28 disorders will have populated computational-model sections and 2,106 will still have no field. Most of those 2,106 are rare disorders for which no disease-level computational model is expected. They should not be converted mechanically to empty lists.

The useful remaining queue is therefore the ranked, evidence-led set below. It favors models that illuminate a named pathophysiology node, prioritizes recent patient-specific digital twins, and prefers runnable public deposits. It deliberately includes several model classes rather than turning the section into an index of diagnostic machine-learning classifiers.

This report complements the execution-focused Computational Model Execution: State of the Art.

Work already in flight

PR	Disorder	Distinctive model class
#9084	Glioblastoma, IDH-wildtype	reaction-diffusion and data-assimilation twins; multiscale ABM/RL
#9087	Advanced sleep phase syndrome	circadian ODE and physiological phase estimator
#9091	Multiple sclerosis	computational disease models
#9104	Hyperinsulinemic hypoglycemia	glucose-insulin physiology
#9119	Noonan syndrome	allele-specific signaling models
#9122	Ataxia-telangiectasia	DNA-damage/signaling models
#9123	Brugada syndrome	cardiac electrophysiology
#9140	Renal cell carcinoma	agent-based tumor model
#9142	Hepatitis C	within-host viral dynamics
#9145	Chronic myeloid leukemia	ecological ODE / treatment-response ensemble
#9263	Alzheimer disease	EEG-personalized brain digital twin

Operational state is deliberately kept out of the table because it changes faster than this report. As of 2026-08-22 19:11 UTC, nine approved PRs (#9087 through #9145 above) were marked ready for review and were conflict-free. #9084 remained a draft while its requested review fix was being completed; #9263 was also awaiting re-review after requested changes. Approved, ready, unassigned PRs become eligible for the repository's deterministic auto-merge sweep after its three-day cooling period; drafts are never eligible. Current state should be read from the linked PRs.

Tier 1: curate next

These have a strong disease/model fit and a credible path to mechanistic, repository-backed curation. Order within the tier is a judgment call.

Priority	Disorder file	Model lead and intended value	Principal guardrail
1	`Epilepsy.yaml`	Virtual Epileptic Patient / The Virtual Brain patient-specific seizure-network models; connect excitability and propagation parameters to the epilepsy mechanism chain	Separate a genuine patient-specific mechanistic twin from generic seizure classifiers
2	`Heart_Failure.yaml`	Med-Real2Sim and CircAdapt-style inverse hemodynamic twins; include pressure, flow, and contractility readouts	Confirm that the evaluated disease cohort is heart failure, not only healthy cardiac geometry
3	`Tuberculosis.yaml`	UISS-TB and other immune agent-based models with executable intervention simulations	Do not promote regulatory qualification or platform claims into clinical efficacy claims
4	`Asthma.yaml`	airway smooth-muscle, bronchoconstriction, and immune/airway multiscale models	Prefer mechanistic airway models over risk-prediction ML; verify any claimed asthma digital twin is more than a proposal
5	`Cystic_Fibrosis.yaml`	CFTR/airway-surface-liquid ion-transport ODEs and mucus-clearance models	Keep epithelial transport models distinct from organoid or animal models
6	`Chronic_Obstructive_Pulmonary_Disease.yaml`	patient-specific ventilation, airway-network, and emphysema mechanics models	Link to an actual COPD mechanism node; generic lung simulators alone are insufficient
7	`Sickle_Cell_Disease.yaml`	multiscale red-cell sickling, adhesion, and microvascular-flow simulations	State oxygenation, rheology, and vessel-geometry assumptions; avoid treating in-vitro calibration as clinical validation
8	`Duchenne_Muscular_Dystrophy.yaml`	calcium handling, energetics, and finite-element muscle mechanics models	Do not misclassify iPSC or mouse muscle systems as computational models
9	`Prostate_Adenocarcinoma.yaml`	patient-specific treatment-response and executable PI3K/AKT/mTOR/AR resistance models	Coordinate with the existing phenotype algorithm and mTOR-resistance conformance rather than duplicating them
10	`Non-Small_Cell_Lung_Cancer.yaml`	tumor-growth and immune-checkpoint treatment twins; executable Boolean/agent-based alternatives	Require subtype/cohort identity and a direct pathograph join point
11	`COVID-19.yaml`	within-host viral/immune ODE, QSP, and multiscale intervention models	Prefer models with parameter identifiability and external validation; avoid undifferentiated pandemic forecasting
12	`Chronic_Kidney_Disease.yaml`	nephron/hemodynamic and fibrosis/QSP models that complement, rather than repeat, the CKD-MBD exemplar	Keep CKD progression separate from the already-curated mineral-bone subsystem

Tier 2: strong candidates after focused literature and repository checks

Disorder file	Candidate model family	Why it is second tier
`Amyotrophic_Lateral_Sclerosis.yaml`	motor-neuron excitability, protein-aggregation, and progression models	Many models are mutation- or subsystem-specific; the disease-level join point needs careful scoping
`Dilated_Cardiomyopathy.yaml`	patient-specific electromechanics and ventricular-remodeling twins	Repositories often supply generic cardiac infrastructure but not the disease parameterization
`Hypertrophic_Cardiomyopathy.yaml`	sarcomere-to-organ electromechanics and sudden-death risk twins	Separate mechanistic simulation from morphology/risk classifiers
`Essential_Hypertension.yaml`	closed-loop circulation and renal-pressure control models	Broad physiological models can be relevant without being disease-specific
`Osteoarthritis.yaml`	cartilage finite-element, mechanobiology, and cell-based degradation models	Public code and model-to-node evidence are uneven
`Rheumatoid_Arthritis.yaml`	immune QSP and synovial agent-based treatment models	Many deposits are proprietary or treatment-development platforms
`ER_Positive_Breast_Cancer.yaml`	endocrine-resistance Boolean/QSP and patient-specific response models	Curate the molecular subtype, not generic breast-cancer prediction
`HER2_Positive_Breast_Cancer.yaml`	HER2 signaling and combination-treatment models	Check that repositories preserve the exact published parameterization
`Triple_Negative_Breast_Cancer.yaml`	executable signaling networks and tumor-immune ABMs	High volume of ML papers, low fraction of interpretable mechanism models
`Metastatic_Colorectal_Cancer.yaml`	liver-metastasis growth/therapy twins and agent-based models	Primary and metastatic disease are often conflated in search results
`Acquired_Immunodeficiency_Syndrome.yaml`	classic within-host HIV dynamics and treatment models	Strong literature, but repository and modern reproducibility metadata need work
`Major_Depressive_Disorder.yaml`	neural-circuit and treatment-response models	Most available work is predictive ML; accept only models with a defensible mechanism link
`Schizophrenia.yaml`	cortical excitation/inhibition and dysconnection models	Heterogeneity makes disease-level fidelity claims particularly fragile

Explicit empty sections

The following six entries on main explicitly declare computational_models: []:

ATTR_Amyloidosis.yaml
Alsahan-Harris_Syndrome.yaml
Aminoacylase_1_Deficiency.yaml
Amyloidosis.yaml
Anterior_Spinal_Artery_Syndrome.yaml
BEST1-Related_Dominant_Retinopathy.yaml

Treat these as "reviewed but no model curated," not as evidence that no model can exist. Reopen one only when a concrete paper or repository lead appears. The broad Amyloidosis.yaml entry is especially prone to subtype conflation; a subtype-specific model belongs on the corresponding subtype entry.

Defer and skip rules

A disorder can remain without computational_models when any of the following applies:

No disease-specific or mechanism-specific computational model was found.
The only results are diagnostic/prognostic classifiers that do not illuminate a curated pathophysiology node.
A paper proposes a future digital twin but does not instantiate or evaluate one.
The system is an organoid, cell line, organ-chip, or animal model and therefore belongs under experimental_models or animal_models.
A generic platform is not parameterized or evaluated for the disorder.
Disease identity is ambiguous or a paper actually concerns a sibling subtype.
The source is too poorly specified to support honest provenance, limitations, or a model-to-mechanism link.

An inaccessible repository is not automatically disqualifying: a paper-only model may still be curated when its equations, variables, findings, and limits are recoverable. Conversely, a GitHub repository is not automatically relevant.

Per-disorder completion standard

Each future PR should modify one disorder file and should, where the source permits:

classify the model type accurately;
pin a repository commit or stable accession;
record software, format, and license/reproducibility caveats;
connect the model to exact pathophysiology or phenotype names through modeled_mechanisms;
distinguish link-level evidence from readout-level evidence;
represent negative results and fidelity limitations explicitly;
run the deposited model or at least a bounded smoke test when feasible;
use deep-research/OpenScientist output only as a lead, after disease-identity and reference verification;
add a history record and regenerate app/models/data.js with just gen-models-data (or uv run python -m dismech.export.models_export).

Reproducing the census

The snapshot counts use only top-level fields on origin/main:

git ls-tree -r --name-only origin/main -- kb/disorders/ | grep -c '\.yaml$'
git grep -l '^computational_models:' origin/main -- 'kb/disorders/*.yaml'

The second command identifies 23 files. Inspection of those fields separates 17 populated sections from 6 explicit empty lists. Open PRs were then audited individually rather than counted from branch names.