NMD-VCell Neuromuscular Virtual Cell Research Platform Module: Data · Evidence → perturbation → experiment → outcome NMD = neuromuscular disorders
For Reviewers Current Release

Release 2026.08DMD context observedDMD candidate-conditioned prediction not yet eligible

View scientific status
Evidence freeze: 3 August 2026 Resource: v1.2.0-measured-dmd-evidence Schema: 1.1 Open release status →

Research Landscape · intake audited 2026-08-12

Know what exists, what is competitive and what should run next.

This is the missing operating map across the website, manuscript, figures, data and models. It separates released local assets from preserved history, external opportunities and decision-critical missing target-context outcome.

Audit local assetsCompare modelsFind dataOpen full JSON
Top-line decision

NMD-VCell should invest next in source recovery, donor-resolved data objects and one baseline-first public benchmark adapter—not in another unvalidated DMD predictor label.

Inspect the DMD prediction lock →

Release authority · 2026-08-18

The source candidate is auditable; its own bundle cannot prove a later deployment.

This authority object is a release and submission status contract. It does not convert pending metadata, a production receipt, external references or opportunity registries into scientific validation.

Authority JSON →
Source candidatev1.2.0-measured-dmd-evidenceEA-20260817-57UNVERIFIED_AFTER_SOURCE_BUILD
Latest historical receiptSites v16791e7e8ff973dSUCCEEDED
Anonymous route QABLOCKED_HTTP_403Browser-visible production verification remains required.
Submission freezeNOT_FROZENDOI, licence and maintainer metadata remain external gates.
Next required gate. Build and verify EA-20260817-57, capture its external hosting receipt after deployment, then complete Stage 2.5 integrity review and obtain author, licence, maintainer and DOI authorization.

Data intake contract · audited 2026-08-12

Found, pilot-only and model-ready are now separate states.

Use GSE277637 as the line-aware organoid context object and GSE288958 as a bounded, read-only, sample-aware biopsy reference. GSE288958 independent doublet and annotation audits are complete as sensitivity/disagreement layers; ambient RNA and author-exact reprocessing remain explicitly open, and disease-effect/model-label uses remain locked. Keep GSE293514 as a bounded muscle screen until exact feature and guide identity is restored.

Intake JSON →Download TSV
3datasets audited
2context objects ready
1existing-reference pilots ready
None qualifiedperturbation-model status
Not availabledirect DMD candidate sets
GSE288958 operational status. Existing processed reference: AVAILABLE READ ONLY; read-only matrix QC: COMPLETE WITH NORMAL 2 SENSITIVITY; independent doublet audit: COMPLETE DONOR WISE SCRUBLET SENSITIVITY; independent annotation audit: COMPLETE BOUNDED DISAGREEMENT AUDIT; ambient RNA: OPEN UNFILTERED DROPLETS UNAVAILABLE. Proceed with bounded existing-reference context analysis and retain the open ambient-RNA and author-exact reprocessing boundaries.
GSE288958 QC gate. REFERENCE CONTEXT READY INDEPENDENT AUDITS COMPLETE AMBIENT OPEN Use the reference context with visible Normal_2, Scrublet and annotation-disagreement sensitivities. Seek unfiltered droplets only for ambient-RNA closure and the six exact author artifacts only before an author-exact reprocessing claim. Open visual audit · Open gate JSON · Download audit brief
Sample-aware context object. 11 samples · 6 stable modules · context-only alignments: ADAM10, CPEB1. Open context JSON · Download context brief · Module table
GSE277637LOCAL LINE AWARE OBJECTS READY

DMD iPSC-derived skeletal-muscle organoid scRNA-seq

Donor/line-aware DMD disease-context and organoid heterogeneity stress test

Local objects
13 checked · 0 required missing
Reference state
CONTEXT OBJECT READY
Reference QC
CONTEXT READY
SRA route
NOT APPLICABLE
Context analysis
READY · context only
Perturbation model
NOT READY
Inferential unit
iPSC-derived line/sample; cells are nested observations, not independent biological replicates
Next gate

Freeze line-aware sample metadata, preserve one-line-one-unit aggregation, publish source checksum and license record before public matrix release.

GSE288958REFERENCE CONTEXT READY · AMBIENT OPEN

Human DMD/BMD/control muscle snRNA-seq and spatial context

Donor-resolved DMD disease-state and cell-context validation

Local objects
27 checked · 0 required missing
Reference state
OPENED CHECKSUMMED IDENTIFIER RECONCILED
Reference QC
COMPLETE READ ONLY MATRIX QC WITH NORMAL 2 SENSITIVITY
SRA route
OPTIONAL DUPLICATE DOWNLOAD ROUTE CLEANED
Context analysis
READY · context only
Perturbation model
NOT READY
Inferential unit
independent biopsy/sample; donor, site and age must remain explicit covariates
Next gate

Use the read-only sample-aware reference with Normal_2, Scrublet and annotation-disagreement sensitivities. Unfiltered droplets or author contamination estimates are needed only to close ambient RNA; exact author artifacts are needed only for an author-exact reprocessing claim.

GSE293514LOCAL PARTIAL BOUNDED SCREEN IDENTITY BLOCKED

Human myoblast fusion CRISPR screen and CROP-seq context

Healthy-myoblast muscle-context safety/stress test only

Local objects
11 checked · 4 required missing
Reference state
NOT READY
Reference QC
NOT COMPLETE
SRA route
NOT APPLICABLE
Context analysis
NO · intake incomplete
Perturbation model
NOT READY
Inferential unit
screen-level or selected validation object; deposited cell metadata is not a guide-resolved perturbation truth object
Next gate

Recover author-ordered feature table, guide reference, exact cell metadata and processing contract; checksum every returned file before any expression transfer.

Unit structure rule. Cells are nested observations. Donor, line, biopsy, culture or independent perturbation unit must remain the inferential unit for disease-level claims. Disease-state, organoid, correction and healthy-myoblast screen data remain context references and do not create direct candidate-level DMD perturbation truth.

Citable objects · checked 2026-09-09

Two bounded disease-context objects are citable now; perturbation truth remains separate.

GSE277637 preserves line-aware organoid context. GSE288958 now adds a sample-aware biopsy reference with independent doublet and annotation-disagreement audits.

Landscape JSON →GSE288958 visual audit →
GSE277637CITABLE_LOCAL_CONTEXT_OBJECT_NOT_PUBLIC_MATRIX_RELEASE

Line-aware DMD organoid context

DMD organoid disease-axis heterogeneity stress test

Sample units
4 · DMD1, DMD2, DMD3 + WT shared control
Cells
10,480 nested cells
Evidence role
Context analysis only · no perturbation model
Source state
METADATA_QC_AND_CHECKSUMS_ONLY_SOURCE_MATRIX_NOT_REHOSTED
Claim ceiling

Supports donor/line-aware DMD disease context and organoid heterogeneity stress testing only. It does not support candidate perturbation response, therapeutic efficacy, clinical prediction or a calibrated DMD predictor.

GSE293514AUTHOR_CONTACT_OR_ARCHIVE_RECOVERY_REQUIRED

Author/source identity recovery package

Healthy-myoblast CROP-seq context remains bounded until feature order, guide identity and processing provenance are restored.

Current blocker
12 metadata columns · 62711 matrix features vs 62710 frozen inventory rows
Requested files
5 identity and processing artifacts
Biological transfer
BLOCKED
Unlocks
Exact technical reanalysis decision, not DMD candidate outcome
Next gate

Hash-freeze returned files, reconcile features and barcodes, validate guide assignment, then reproduce Fig. 6 counts.

GSE288958LOCAL_AUTHORITATIVE_RDS_OPENED_IDENTIFIERS_RECONCILED_INDEPENDENT_AUDITS_COMPLETE_AMBIENT_OPEN

Portable muscle-context import gate

The official processed matrices, inherited annotations and checksum-matched authoritative RDS are workspace-local. Object opening, identifier reconciliation, read-only matrix QC, donor-wise Scrublet sensitivity and independent marker-label disagreement audit are complete; ambient RNA and author-exact reprocessing remain explicitly open.

Pilot evidence
59,222 cells · 11 samples · selected endothelial
Official raw archive
359,823,360 bytes public · local validated 359823360 bytes
Local pilot objects
27 checksum-addressed source, metadata and audit artifacts; raw archive remains the matrix package
Import gate
4 portable source/metadata requirements
Next gate

Use the bounded read-only context route. Unfiltered droplets are needed only for ambient-RNA closure; exact author execution artifacts are needed only for author-exact reprocessing. The 1847703081-byte object remains SHA-addressed, and disease-effect testing remains locked.

Unchanged boundary. GSE277637 contributes a line-aware organoid stress test; GSE288958 contributes bounded read-only sample-aware biopsy context with independent doublet and annotation audits; GSE293514 remains technically identity-blocked. GSE288958 ambient RNA and author-exact reprocessing remain open, and direct DMD candidate perturbation outcomes remain Not measured yet.
Executable checks. GSE293514 recovery preflight is currently BLOCKED_MISSING_OR_INVALID_IDENTITY_ARTIFACTS; GSE288958 portable-import preflight is READY_FOR_BOUNDED_READ_ONLY_CONTEXT_INDEPENDENT_AUDITS_COMPLETE_AMBIENT_OPEN. Its context route is ready under the published sample-unit and claim-boundary contract.

Local asset ledger

The project did not lose its historical figures; it failed to expose their lineage.

The visible shortage was primarily an indexing and product-organization problem, not wholesale deletion. Historical assets remain preserved; current release objects and missing scientific outcomes must still be distinguished.

Inventory JSON →
161controlled routes
5formal datasets
6frozen runs
9current main figures
27current supplementary figures
773unique historical visual hashes
PUBLIC PRODUCTRELEASED

Controlled public website and typed API

161 controlled routes with release, route and checksum manifests

A route inventory existed, but research assets, external opportunities and adoption decisions were not presented in one navigable view.

Action

Keep this landscape page in the controlled route and release manifests.

COMPUTABLE REGISTRIESRELEASED LIMITED

Typed dataset, model-card and ModelRun registries

5 datasets · 5 model cards · 6 frozen runs · 0 calibrated DMD models

The execution ledger is strong, but disease-relevant perturbation outcomes and author-missing feature/metadata files remain absent.

Action

Acquire missing metadata first; evaluate new adapters on compatible non-DMD benchmarks before any disease claim.

PUBLICATION PACKAGECURATED NOT SUBMISSION READY

Curated v06 manuscript, main figures and supplement

9 main figures · 27 supplementary figures · 44 supplementary tables · 36 traced figure artifacts · 16 traced key numeric claims

Technical verification passes, while the pre-submission integrity state remains blocked and the package is explicitly working material.

Action

Expose the current package and its verification state without hiding v05 or claiming submission readiness.

HISTORICAL VISUAL LINEAGEHISTORICAL PRESERVED

All-version visual asset lineage

1,707 visual asset instances · 773 unique SHA-256 values · 1.187 GB audited source lineage

Historical assets were retained in the workspace but were not discoverable from the public product.

Action

Preserve every predecessor and publish selection lineage; do not replace historical assets with only the current render.

Open public object →curated_v06/source_registry/ALL_VERSION_VISUAL_ASSET_SUMMARY.json (workspace audit source)
DISEASE EVIDENCERELEASED WITH SCOPE DIFFERENCES

DMD, FSHD, DM1 and SMA evidence atlas

4 disease modules and 10 measured-or-missing cell-context records under one evidence vocabulary

Disease evidence depth differs and cannot be collapsed into a shared predictor or disease ranking.

Action

Extend donor-resolved cell objects only when source metadata and statistical units pass registration.

DECISION CRITICAL MISSING ASSETMISSING

Independent candidate-level DMD perturbation outcomes

No governed candidate has a qualifying independent DMD perturbation outcome yet

No released dataset can calibrate a candidate-level DMD response predictor.

Action

Keep DMD perturbation prediction locked; computational expansion cannot manufacture this missing outcomes.

Version policy. Curated v06 preserves v05 assets byte-identically and adds the current selected package. Historical and current assets stay distinct; neither is silently discarded.

Absorption layer · four publishing contracts

Data, models, benchmarks and experiments now publish through the same decision grammar.

Every public asset must resolve to a typed card, every card must name its claim boundary and next gate, and every experiment-capable path must return through the outcome registry.

Contract JSON →Download TSV
01 · Dataset CardRELEASED PARTIAL STANDARDIZATION

What was measured, in which biological units, and what can this dataset validly support?

Turn a stored matrix or source accession into a citable, task-qualified biological data object.

Required sections
  • Identity and version
  • Disease, tissue, cell state and perturbation context
  • Sample, donor and statistical-unit design
  • Matrix, feature, metadata and checksum inventory
  • Access, licence and source provenance
  • Readiness, permitted use and prohibited use
  • Linked ModelRuns, benchmarks and studies
Required links
source record · download or access object · readiness gate · downstream run or study
Learned from
CELLxGENE · Tahoe-100M · DepMap
Next gate

Render the same required sections for every dataset and reject model-ready status when donor, feature or perturbation identity is unresolved.

BoundaryDataset scale or availability does not establish independent biological replication or task suitability.

02 · Model CardRELEASED AUDITED

What did this exact model version receive, return, beat or fail, and where must it abstain?

Bind a model identity to one task, one execution history and one bounded claim surface.

Required sections
  • Model identity, version and provider
  • Task, input and output object
  • Training and evaluation context
  • Split, leakage and baseline controls
  • Metrics, seeds and local execution result
  • Failure, abstention and prohibited claims
  • Artifacts, run receipts and reproducibility links
Required links
ModelRun · dataset digest · benchmark task · artifact or source · failure analysis
Learned from
CZI Virtual Cells Platform · X-Cell · NVIDIA BioNeMo
Next gate

Expose a uniform adapter and quickstart only after the same model identity is reproducible across documentation, API and run receipts.

BoundaryArchitecture reputation and external performance are not local NMD or DMD validation.

03 · Benchmark CardRELEASED TASK BOUNDED

On which sealed outcomes, split, baselines and metrics did a model advance or stop?

Make model evaluation a frozen scientific task rather than a movable leaderboard claim.

Required sections
  • Scientific question and task identifier
  • Frozen data, target context and outcome object
  • Split, leakage groups and hidden-label policy
  • Permanent simple baselines
  • Metric families, seeds and uncertainty
  • Advance or no-advance decision gate
  • Immutable result and negative-run history
Required links
benchmark schema · split manifest · baseline runs · ModelRuns · result download
Learned from
Arc Virtual Cell Challenge · CZI Virtual Cells benchmarks
Next gate

Add external adapters only on compatible frozen tasks and retain every stopped or negative result.

BoundaryA same-assay benchmark result cannot be transferred to DMD muscle or a different output distribution.

04 · Workflow / Outcome CardSCHEMA RELEASED OUTCOME REGISTRY EMPTY

What experiment can change the conclusion, and how will every positive, null, toxic or failed outcome return?

Connect a biological question to a frozen experiment, returned outcome and explicit evidence update.

Required sections
  • Question, estimand and hypothesis
  • Intervention, cell context, dose, time and comparator
  • Biological unit, controls, endpoints and QC
  • Lifecycle, preregistration and immutable decision rule
  • Prediction freeze when applicable
  • Measured outcome, provenance and failure state
  • Evidence transition and claim update
Required links
Study Card · Prediction Card when applicable · Outcome object · source evidence · release update
Learned from
Recursion OS · Cellular Intelligence · VCell
Next gate

Register one immutable DMD study and return its outcome through the same identifiers without suppressing null, toxic or failed branches.

BoundaryA study design is prospective infrastructure, not a measured intervention effect.

What this changes. A page is no longer considered complete because it has a title and description. It must expose identity, biological context, inputs, outputs, provenance, failure state, claim boundary, linked objects and the next gate appropriate to its product type.

Competitive map · 14 official-source references

Borrow product contracts, not unverified performance claims.

Compete on neuromuscular disease specificity, source-to-decision traceability, baseline-first model audits and prospective evidence governance—not on unsupported scale parity.

Landscape JSON →Download TSV
FULL STACK VIRTUAL CELLP0 · DIRECT

Arc Virtual Cell Initiative · STATE · Stack · VCC

Primary userML developers and computational biologists
Entry taskTrain or benchmark a perturbation model
MoatDATA MODEL BENCHMARK
Open / lab loopOPEN RESEARCH · PARTIAL

Connects large observational and perturbational atlases, open models, standardized evaluation and a held-out challenge. Stack adds in-context single-cell modeling; the 2025 challenge retained simple baselines and exposed generalization failures.

Adopt
Bind every prediction to a frozen task, hidden or sealed outcomes, permanent simple baselines and a ModelRun record.
Local equivalent
Data Universe · ModelRun registry · Benchmark dashboard
NMD-VCell edge
Disease-specific evidence governance, explicit missing outcomes, candidate Study Cards and preservation of negative runs.
Current gap
No Arc-scale training corpus, no locally executed STATE or Stack adapter and no disease-relevant held-out perturbation outcomes.

Do not copyDo not present Arc-scale data, leaderboards or generalization as local capability.

MODEL DATA BENCHMARK PLATFORMP0 · DIRECT

CZI / Biohub Virtual Cells Platform

Primary userBiologists and ML developers
Entry taskFind data, select a model, run or compare
MoatMODEL DATA BENCHMARK WORKSPACE
Open / lab loopOPEN PLATFORM · NO

Publishes model cards, datasets, benchmarks, CLI access and hosted workflows in one ecosystem, including TranscriptFormer and the context-specific scLDM.CD4 perturbation model.

Adopt
Use one adapter and metadata contract across data discovery, local execution, benchmark reporting and web presentation.
Local equivalent
Dataset registry · Model cards · ModelRun registry · Virtual Cell Studio
NMD-VCell edge
A narrower neuromuscular decision loop with explicit claim ceilings and source-to-experiment traceability.
Current gap
NMD-VCell has a run ledger but not a unified executable adapter layer or hosted inference workspace.

Do not copyDo not build a broad model marketplace before disease workflows and adapters are stable.

SINGLE CELL DATA PLATFORMP1 · ADJACENT

CZ CELLxGENE Discover · Explorer · Census

Primary userSingle-cell biologists
Entry taskFind and explore a cell dataset
MoatSTANDARDIZED DATA DISCOVERY
Open / lab loopOPEN PLATFORM · NO

Provides versioned, ontology-harmonized single-cell data with low-latency metadata queries, source H5AD access and interoperable AnnData, Seurat and SingleCellExperiment slices.

Adopt
Make dataset, donor, disease, state, batch and gene selections resolve to stable source objects and exportable analysis units.
Local equivalent
Cell Context Map · Dataset registry
NMD-VCell edge
Neuromuscular disease interpretation and donor-aware claim ceilings rather than a general atlas browser.
Current gap
The current Cell Context Explorer is a typed context registry, not yet a donor-level cell browser over imported matrices.

Do not copyDo not treat cell counts as independent biological replication.

TARGET DISEASE ENTITY GRAPHP0 · ADJACENT

Open Targets Platform

Primary userTarget and translational researchers
Entry taskInspect a target–disease relationship
MoatENTITY RELATION EVIDENCE
Open / lab loopOPEN PLATFORM · NO

Exposes source-provenanced target, disease, drug and association entities through a web interface, downloads, GraphQL and a versioned official MCP server.

Adopt
Keep stable entity identifiers, release metadata, source-level provenance and machine interfaces synchronized.
Local equivalent
Gene records · Disease atlas · Evidence graph
NMD-VCell edge
Cell-context, perturbation-run and prospective-study objects specialized for neuromuscular research.
Current gap
NMD-VCell has linked objects but lacks a general graph query or agent interface across every object type.

Do not copyDo not collapse heterogeneous evidence into one opaque target score.

FUNCTIONAL DEPENDENCY PLATFORMP2 · ADJACENT

DepMap Portal

Primary userFunctional genomics researchers
Entry taskQuery dependency in a model context
MoatRECURRING FUNCTIONAL DATA
Open / lab loopOPEN DATA PORTAL · PARTIAL

Combines recurring public CRISPR dependency releases, molecular characterization, model metadata, downloads and experimental APIs.

Adopt
Version data releases, model contexts and mapping files together, and expose historical release identities.
Local equivalent
External context evidence · Source buckets
NMD-VCell edge
Disease-specific evidence boundaries prevent cancer-cell dependency from being relabeled as muscle efficacy.
Current gap
DepMap can add contextual dependency evidence, but it is not direct DMD muscle perturbation outcome.

Do not copyDo not relabel cancer-cell dependency as neuromuscular efficacy.

LINKED SINGLE CELL VISUALIZATIONP2 · INFRASTRUCTURE

Vitessce

Primary userSingle-cell and spatial analysts
Entry taskCompose linked views over registered objects
MoatLINKED VISUAL CONFIGURATION
Open / lab loopOPEN SOURCE · NO

Uses JSON view configurations to coordinate embeddings, expression matrices, spatial images and controls over static or object-store data.

Adopt
Add linked gene, cell-state, donor, heatmap and spatial views after qualified AnnData-Zarr objects are registered.
Local equivalent
Future linked cell browser
NMD-VCell edge
The visual layer would inherit NMD-VCell evidence states and donor-aware statistical rules.
Current gap
No production AnnData-Zarr or Vitessce view configuration is registered for DMD tissue data.

Do not copyDo not add a viewer before donor-aware, checksum-addressed objects exist.

MECHANISTIC SIMULATIONP2 · ADJACENT

VCell Modeling & Analysis Software

Primary userSystems and computational biologists
Entry taskBuild and simulate a mechanistic model
MoatMECHANISTIC SIMULATION
Open / lab loopOPEN SOURCE · NO

Separates biological model definition, applications, parameters, geometry, numerical solvers, simulation runs and downloadable results across deterministic and stochastic methods.

Adopt
Keep evidence checks, learned response models and future mechanistic solvers as distinct object classes with equations, parameters and run logs.
Local equivalent
DMD process map · Future mechanistic solver lane
NMD-VCell edge
NMD-VCell begins from disease evidence and the experiment needed to establish a modelable transition.
Current gap
No calibrated mechanistic DMD model, parameter set or numerical simulation result is released.

Do not copyDo not relabel an evidence graph as a calibrated mechanistic simulation.

PERTURBATION DATA ENGINEP1 · DIRECT

Tahoe · Mosaic · Tahoe-100M

Primary userVirtual-cell model builders and drug discovery teams
Entry taskAccess a large chemical perturbation atlas
MoatPROPRIETARY DATA ENGINE WITH OPEN RELEASES
Open / lab loopMIXED · YES

Builds large perturbational single-cell maps and publishes Tahoe-100M as a reusable data product with manuscript, download, community and model entry points.

Adopt
Treat every important disease dataset as a versioned product with a scientific question, access path, limitations, examples and downstream model links.
Local equivalent
Measured DMD evidence release · Dataset cards
NMD-VCell edge
NMD-VCell can make donor structure, neuromuscular context and claim ceilings more visible than a scale-first perturbation atlas.
Current gap
No flagship NMD perturbation dataset currently connects a public Dataset Card to measured outcomes, model runs and follow-up studies.

Do not copyDo not use cell-count scale as a substitute for donor, disease and perturbation relevance.

CAUSAL PERTURBATION MODELP1 · DIRECT

Xaira Therapeutics · X-Cell

Primary userAI drug discovery researchers
Entry taskEvaluate or apply a cross-context perturbation model
MoatPROPRIETARY PERTURBATION DATA AND MODEL
Open / lab loopMIXED · YES

Packages a virtual-cell release around a named model, a large perturbation training asset, a technical report and a cross-context prediction task.

Adopt
Give each locally evaluated model a permanent release page that binds model version, training context, target context, artifacts, benchmark results and limitations.
Local equivalent
Model card audit · ModelRun release objects
NMD-VCell edge
NMD-VCell publishes negative runs, abstentions and disease-specific transfer boundaries instead of relying on a broad model launch claim.
Current gap
Model releases are audited, but the public product does not yet present every ModelRun as one linked release package with quickstart and failure analysis.

Do not copyDo not claim cross-context DMD generalization without prospective disease-relevant validation.

LAB IN THE LOOP DRUG DISCOVERYP0 · DIRECT

Recursion OS · Predict–Explain–Discover

Primary userDrug discovery programs
Entry taskMove from perturbation maps to a therapeutic program
MoatAUTOMATED LAB DATA MODEL LOOP
Open / lab loopPROPRIETARY · YES

Connects automated perturbation experiments, phenomics and transcriptomics, learned maps, design workflows and downstream therapeutic programs in a physical-to-digital feedback loop.

Adopt
Show the complete lifecycle from experimental material through assay, data object, model, decision, new experiment and evidence update.
Local equivalent
Evidence-to-experiment loop · Study Cards · Outcome registry
NMD-VCell edge
NMD-VCell can expose each evidence transition and missing link publicly even without proprietary lab scale or a drug pipeline.
Current gap
Study Cards and an outcome registry exist, but no measured candidate-level DMD outcome has yet completed the loop.

Do not copyDo not imply an automated wet-lab or therapeutic pipeline that NMD-VCell does not operate.

MULTISCALE WORLD MODELP2 · DIRECT

GenBio AI · AIDO

Primary userMultiscale biological AI researchers
Entry taskUse or combine biological foundation models
MoatMULTISCALE MODEL SYSTEM
Open / lab loopMIXED · NO

Frames DNA, RNA, protein, structure and single-cell models as interoperable modules on a roadmap toward a multiscale biological world model.

Adopt
Publish a layered capability roadmap that distinguishes released modules, interfaces between scales and future simulation goals.
Local equivalent
Meaning map · Capability atlas
NMD-VCell edge
NMD-VCell can make every currently supported scale and unsupported transition explicit rather than presenting a universal world-model claim.
Current gap
Gene, pathway, cell-state and experiment objects are linked conceptually but do not yet share one typed cross-scale relation contract.

Do not copyDo not use world-model language where only bounded evidence objects are available.

TEMPORAL SIGNALING ACTIVE LEARNINGP1 · DIRECT

Cellular Intelligence

Primary userRegenerative medicine and cell engineering teams
Entry taskDesign temporal signaling interventions
MoatSEQUENTIAL SIGNALING DATA ENGINE
Open / lab loopPROPRIETARY · YES

Emphasizes sequential signaling, dose, temporal order, cell-fate control and active learning over static cell-state representation.

Adopt
Represent intervention sequence, dose and time as first-class variables, then prioritize experiments by the uncertainty they can resolve.
Local equivalent
DMD process map · Experiment Planner
NMD-VCell edge
The DMD process map and Study Cards can connect temporal signaling hypotheses to source-linked disease evidence and explicit outcome branches.
Current gap
Current studies record time and endpoint, but the planner does not yet compare sequential interventions or information gain across an experiment portfolio.

Do not copyDo not inherit company-reported scale or efficiency claims as validated evidence.

SINGLE CELL DATAOPS MLOPSP0 · INFRASTRUCTURE

Broad Cellarium AI

Primary userSingle-cell computational teams
Entry taskRun a task-specific single-cell tool
MoatDATAOPS MLOPS AND TOOLING
Open / lab loopOPEN RESEARCH · NO

Presents annotation, denoising, perturbation interpretation and cloud infrastructure as separate tools with task-specific identities.

Adopt
Turn existing NMD-VCell pages into a tool catalog only when each module has declared inputs, outputs, failure states, examples and machine interfaces.
Local equivalent
Resolver · Compare · Process inspector · Model auditor · Planner
NMD-VCell edge
NMD-VCell can bind each tool output to disease evidence, a claim ceiling and the next experiment instead of ending at a generic analysis artifact.
Current gap
Resolver, comparator, process inspector, model auditor and planner exist but do not yet share a visible tool contract.

Do not copyDo not label an informational page as a tool without executable inputs and outputs.

MODEL DEPLOYMENT INFRASTRUCTUREP1 · INFRASTRUCTURE

NVIDIA BioNeMo

Primary userAI developers and enterprise research teams
Entry taskBuild, adapt or deploy a biology model
MoatCOMPUTE PACKAGING AND DEPLOYMENT
Open / lab loopMIXED · NO

Packages biomolecular AI capabilities as frameworks, web interfaces, APIs and deployable inference microservices.

Adopt
Keep one model identity across web documentation, API records, local adapters, artifacts and deployment-specific run receipts.
Local equivalent
API · Model adapters · Run receipts
NMD-VCell edge
NMD-VCell can provide stronger disease-specific evidence governance around externally executed or locally adapted models.
Current gap
The public API exposes ModelRuns, but there is no uniform adapter package or deployment receipt across external model families.

Do not copyDo not build enterprise infrastructure before a reproducible disease adapter is needed.

Source rule: every record was checked on 2026-08-16. External scale and performance remain reference facts and are never inherited as NMD-VCell evidence.

Model radar

Executed failures and unrun opportunities belong in different lanes.

Prefer task compatibility, complete metadata, reproducible code/checkpoints and baseline-comparable evaluation over novelty or parameter count.

Model registry JSON →

12records shown

MODEL · LINEAR BASELINEEXECUTED LOCAL LIMITED PASS

Same-context ridge residual baseline

NMD-VCell · NMDVCELL-RIDGE-SAFE-2.3
Processed HepG2 CRISPRi substrate→Same-context perturbation-response prediction

Local evidence: Small mean-RMSE improvement over train mean; raw direction unsupported.

Recommended role: Permanent comparator on every compatible frozen task.

Next gate

Retain without expanding beyond the same-assay scope.

MODEL · GRAPH PERTURBATION MODELEXECUTED LOCAL GATE FAIL

GEARS

External model · local NMD-VCell run · EXTERNAL-GEARS-0.1.2
Five frozen HepG2 same-coverage splits→Genetic perturbation response

Local evidence: 0/5 splits advanced; ridge won all five same-coverage comparisons.

Recommended role: Official negative control.

Next gate

No rerun without a new preregistration, data view or untouched test set.

MODEL · SINGLE CELL FOUNDATION MODELEXECUTED LOCAL GATE FAIL

scGPT

Bowang Lab · local NMD-VCell run · EXTERNAL-SCGPT-0.2.5
Restricted gene coverage on five frozen HepG2 splits→Checkpoint-supported genetic perturbation response

Local evidence: 0/5 splits advanced; higher RMSE than same-coverage ridge.

Recommended role: Foundation-model negative control with coverage visible.

Next gate

Only reassess with a frozen adapter contract and untouched compatible task.

MODEL · KNOWLEDGE GRAPH PERTURBATION MODELEXECUTED LOCAL GATE FAIL

TxPert public-STRING config-gat

Recursion / Valence Labs · local NMD-VCell run · EXTERNAL-TXPERT-CONFIG-GAT
Frozen HepG2 advancement gate→Out-of-distribution transcriptomic perturbation response

Local evidence: Median delta cosine 0.346196 was below train mean 0.373307.

Recommended role: Stopped advanced comparator.

Next gate

Remaining seeds stay locked unless a new prospective protocol is frozen.

MODEL · GENE EMBEDDING PERTURBATION MODELEXECUTED LOCAL NO GO

MORPH DepMap validation pilot

External model · local NMD-VCell run · EXTERNAL-MORPH-1AD06D4
HepG2 with DepMap Public 25Q3 embeddings→Validation-only perturbation pilot

Local evidence: Engineering passed; the model did not beat train mean.

Recommended role: Sealed negative control.

Next gate

Keep the v27 test partition sealed; revisions need a fresh contract.

MODEL · SET BASED STATE TRANSITION MODELEXECUTED 8 LEGACY CONFIGS ALL GATE FAIL

STATE 1.0

Arc Institute · ARC-STATE-1.0
Large public observational and perturbational corpora→Population-level perturbation response across contexts

Local evidence: Eight CPU/GPU and held-out/seen-perturbation receipts were migrated; none beat both zero and train mean on every split.

Recommended role: Stopped comparator unless a materially new preregistered task or untouched outcome exists.

Next gate

Do not rerun by reputation; require a corrected preregistration, untouched truth or materially different task-compatible checkpoint.

MODEL · IN CONTEXT SINGLE CELL FOUNDATION MODELLOCAL NVME CORE SMOKE PASS NO COMPATIBLE CHECKPOINT

Stack

Arc Institute · ARC-STACK-2026
Reported pretraining on 149 million uniformly processed human cells→In-context representation and condition transfer

Local evidence: After one 180-second dependency-path timeout, a bounded local-NVMe retry passed the official StateICLModelBase forward path on a V100 with finite outputs. No pretrained checkpoint or task benchmark was run.

Recommended role: Runtime- and checkpoint-gated context-model hypothesis generator.

Next gate

Identify a task-compatible pretrained genetic checkpoint and preregister an untouched benchmark before any model-performance claim.

MODEL · CROSS SPECIES GENERATIVE FOUNDATION MODELSOURCE VERIFIED REPRESENTATION ONLY

TranscriptFormer

Chan Zuckerberg Initiative · CZI-TRANSCRIPTFORMER-0.6.0
Up to 112 million cells across 12 species→Cell embeddings, classification, disease-state and regulatory inference

Local evidence: No NMD-VCell execution; not registered as a perturbation-response model.

Recommended role: Cross-species and disease-context representation benchmark.

Next gate

Test whether embeddings improve a frozen donor- or disease-held-out task over PCA and scVI.

MODEL · LATENT FLOW MATCHING PERTURBATION MODELSOURCE VERIFIED CONTEXT MISMATCH

scLDM.CD4

CZ Biohub Chicago · BIOHUB-SCLDM-CD4-0.1
Primary human CD4+ T cells→Conditional perturbed profile generation

Local evidence: No NMD-VCell execution; CD4 context does not match muscle disease tasks.

Recommended role: Architecture reference, not a direct adapter priority.

Next gate

Require a compatible muscle perturbation corpus before adaptation is considered.

MODEL · NEURAL OPTIMAL TRANSPORTSOURCE VERIFIED NOT RUN

CellOT

ETH Zurich and collaborators · CELLOT-2023
Unpaired treated and untreated single-cell populations→Distribution-to-distribution perturbation response

Local evidence: No frozen NMD-VCell run.

Recommended role: Distribution-modeling comparator on compatible public drug/cytokine data.

Next gate

Compare against identity, mean-shift, linear and nearest-neighbour transport baselines on a sealed split.

MODEL · COMPOSITIONAL GENERATIVE MODELSOURCE VERIFIED NOT RUN

Compositional Perturbation Autoencoder

Meta Research / scverse collaborators · CPA-2023
High-throughput single-cell perturbation screens→Dose, drug and combination response

Local evidence: No frozen NMD-VCell run.

Recommended role: Permanent compositional baseline candidate for chemical tasks.

Next gate

Use only where dose, covariate and perturbation metadata are complete.

BENCHMARK FRAMEWORK · STANDARDIZED EVALUATIONSOURCE VERIFIED EVALUATION CANDIDATE

PerturBench

Altos Labs and collaborators · PERTURBENCH-2025
Multiple datasets, model families and metric views→Cellular perturbation model benchmarking

Local evidence: Not yet adopted as the NMD-VCell evaluation runtime.

Recommended role: Benchmark adapter and metric cross-check, not a predictor.

Next gate

Map NMD-VCell ModelRun inputs and outputs into PerturBench while preserving the existing metric firewall and simple baselines.

Baseline-first rule. A new architecture advances only if it beats prespecified simple baselines on the same frozen data, split, genes, seeds and metrics. Publication prestige is not a gate result.

Data opportunity registry

Recover identity first, then add scale.

Recover author-missing feature/metadata first. In parallel, standardize GSE288958 and GSE277637 as donor/line-aware disease-context objects, keep GSE293514 as a bounded muscle screen, prepare a VCC/Replogle-compatible benchmark adapter, and query CELLxGENE for additional donor-resolved neuromuscular datasets.

Data registry JSON →

13records shown

DECISION CRITICAL SOURCE RECOVERYP0

Author-missing feature and metadata files

Source authors / existing DMD perturbation study · DATA-AUTHOR-MISSING-FEATURE-METADATA
Scale
Unknown until files are obtained
Context
DMD myoblast / source-study context under audit
Perturbation
Potentially disease-relevant; identity contract incomplete
Access
AUTHOR CONTACT OR ARCHIVE RECOVERY REQUIRED
Feature / metadata
MISSING

Use: Recover identifiers, columns, sample mapping and feature semantics before any reanalysis.

Next action

Request the exact feature table, sample metadata, perturbation identity mapping and processing description; checksum every returned file.

HUMAN DMD SINGLE NUCLEUS AND SPATIALP1

GSE288958 human DMD/BMD/control muscle snRNA-seq and spatial context

Jeon et al. / NCBI GEO · DATA-GSE288958-DMD-HUMAN-MUSCLE
Scale
11 human muscle samples: 5 control, 3 BMD and 3 DMD
Context
Human quadriceps/abdomen muscle biopsies with DMD, BMD and control groups
Perturbation
Disease and treatment-mechanism context; no candidate perturbation outcome
Access
PUBLIC GEO AND LOCAL PROCESSED ASSETS
Feature / metadata
PUBLIC MATRIX AND SAMPLE METADATA

Use: Donor-resolved DMD cell-state signatures, context validation and treatment-mechanism stratification.

Next action

Preserve donor/sample units, harmonize cell-state labels and keep disease context separate from candidate perturbation response.

HUMAN DMD ORGANOID SINGLE CELLP1

GSE277637 DMD iPSC-derived skeletal-muscle organoid scRNA-seq

Kindler et al. / NCBI GEO · DATA-GSE277637-DMD-ORGANOID
Scale
4 10x samples: 1 healthy iPSC control and 3 DMD patient-derived iPSC lines
Context
Human iPSC-derived skeletal-muscle organoids with myogenic progenitor/satellite-cell focus
Perturbation
DMD disease context; no candidate perturbation outcome
Access
PUBLIC GEO LOCAL INGEST AVAILABLE
Feature / metadata
PUBLIC RAW MATRIX WITH SAMPLE LABELS

Use: Organoid-specific DMD/control baseline and myogenic progenitor state validation.

Next action

Analyze as an organoid layer with line-aware units; do not pool it with patient biopsy muscle or call it a candidate screen.

HUMAN MYOBLAST FUNCTIONAL CROPSEQP1

GSE293514 human-myoblast CROP-seq fusion screen

Bi et al. / NCBI GEO · DATA-GSE293514-HUMAN-MYOBLAST-CROPSEQ
Scale
57 samples; 250-hit mini-library; 3.7 GB count matrix and 36.3 MB deposited metadata
Context
Human myoblast fusion and early myogenic differentiation; not a DMD experiment
Perturbation
Split-pool CROP-seq with GM, DM2d and DM6d myoblast differentiation states
Access
PUBLIC GEO TECHNICAL RECONSTRUCTION AVAILABLE
Feature / metadata
AUTHOR FEATURE AND GUIDE IDENTITY RECOVERY REQUIRED

Use: Muscle-context safety/stress testing and prospective perturbation assay design, not DMD efficacy validation.

Next action

Resolve feature/guide identity before expression-level transfer; retain current gene-level fusion results only as a bounded safety filter.

MOUSE DMD SATELLITE CELL SINGLE CELLP2

GSE273343 DMD mouse satellite-cell scRNA-seq

Granet et al. / NCBI GEO · DATA-GSE273343-DMD-MOUSE-SATELLITE
Scale
4 sorted satellite-cell samples: B10, mdx, DBA and D2-mdx
Context
Mouse muscle satellite cells and myogenic progenitors
Perturbation
DMD genotype and regenerative-capacity context; no human candidate outcome
Access
PUBLIC GEO RAW MATRIX
Feature / metadata
PUBLIC SAMPLE AND CELL MATRIX

Use: Mechanistic satellite-cell module and cross-species state-direction stress test.

Next action

Use only as a species-specific layer with ortholog mapping and explicit mouse-to-human transfer limits.

MOUSE DMD MACROPHAGE SINGLE CELLP2

GSE265803 DMD mouse resident-macrophage scRNA-seq

Wang et al. / NCBI GEO · DATA-GSE265803-DMD-MOUSE-MACROPHAGE
Scale
4 quadriceps/diaphragm samples across mdx5cv and Ccr2-related contexts
Context
Mouse dystrophic skeletal-muscle immune niche
Perturbation
DMD inflammatory-niche and resident-macrophage activation context
Access
PUBLIC GEO RAW MATRIX
Feature / metadata
PUBLIC SAMPLE AND CELL MATRIX

Use: Inflammatory-niche module and cell-state composition stress test.

Next action

Keep as a mouse immune-context layer; do not convert macrophage activation into candidate rescue evidence.

OBSERVATIONAL SINGLE CELL CENSUSP1

CZ CELLxGENE Census LTS 2025-11-08

Chan Zuckerberg Initiative · DATA-CELLXGENE-CENSUS-2025-11-08
Scale
1,845 datasets; 162,025,130 human cells (99,633,637 unique)
Context
Human, mouse and three primate species with standardized metadata
Perturbation
Predominantly observational; dataset dependent
Access
PUBLIC API AND SOURCE H5AD
Feature / metadata
STANDARDIZED WITH SOURCE LEVEL VARIATION

Use: Find and register donor-resolved neuromuscular datasets; build disease-context slices and source-level citations.

Next action

Query muscle, myoblast, neuromuscular disease and donor metadata; reject datasets that cannot preserve donor-aware units.

OBSERVATIONAL AND PERTURBATIONAL ATLASP2

Arc Virtual Cell Atlas

Arc Institute · DATA-ARC-VIRTUAL-CELL-ATLAS
Scale
More than 300 million cells across scBaseCount, Tahoe-100M and VCC resources
Context
Broad public cell contexts
Perturbation
Mixed observational, genetic and chemical resources
Access
PUBLIC OPEN RESOURCES WITH DATASET SPECIFIC LICENSES
Feature / metadata
RESOURCE SPECIFIC REVIEW REQUIRED

Use: Pretraining, representation and benchmark discovery—not DMD efficacy evidence.

Next action

Register only the exact constituent dataset used by a run, including version and license.

CHEMICAL PERTURBATION ATLASP2

Tahoe-100M

Tahoe Bio / Vevo Therapeutics / Arc Institute · DATA-TAHOE-100M
Scale
Over 100 million profiles; 50 cancer cell lines; 1,100 small-molecule perturbations
Context
Cancer cell lines
Perturbation
Small molecules with vehicle controls and drug metadata
Access
PUBLIC HUGGING FACE CC0
Feature / metadata
TABLES AVAILABLE REQUIRES PLATE AWARE PROCESSING

Use: Chemical-response scalability and context-generalization benchmark.

Next action

Start with streaming metadata and a small frozen subset; keep cancer-cell findings outside DMD claims.

HELD OUT GENETIC PERTURBATION BENCHMARKP1

Arc Virtual Cell Challenge H1 hESC benchmark

Arc Institute · DATA-ARC-VCC-H1-2025
Scale
Approximately 300,000 cells across 300 CRISPRi perturbations
Context
H1 human embryonic stem cells
Perturbation
CRISPRi; 150 train, 50 validation and 100 held-out test perturbations
Access
PUBLIC CHALLENGE DATA AND RULES
Feature / metadata
BENCHMARK CONTRACT AVAILABLE

Use: First adapter target for STATE or another external model because the split and metrics are externally defined.

Next action

Reproduce simple baselines and metric formulas before running a complex adapter; preserve held-out boundaries.

GENOME SCALE CRISPRI PERTURB SEQP1

Genome-scale Perturb-seq

Replogle et al. · DATA-REPLOGLE-2022
Scale
Screens covering 9,866 expressed genes and 2,057 common-essential genes
Context
K562 and RPE1 cell systems
Perturbation
CRISPRi with direct guide capture
Access
PUBLIC PROCESSED AND ARCHIVAL SOURCES
Feature / metadata
PUBLIC BUT ADAPTER HARMONIZATION REQUIRED

Use: Genetic perturbation pretraining and frozen cross-context benchmarks.

Next action

Use a versioned processed object with explicit gene universe, control definition and cell-line split.

HARMONIZED PERTURBATION COLLECTIONP1

scPerturb

Sander Lab and collaborators · DATA-SCPERTURB-2024
Scale
44 public single-cell perturbation-response datasets in the Nature Methods release
Context
Multiple assays and cell systems
Perturbation
Genetic, chemical and other targeted perturbations
Access
PUBLIC ZENODO AND SOURCE CODE
Feature / metadata
HARMONIZED WITH DATASET LEVEL LIMITATIONS

Use: Dataset discovery, preprocessing comparison and multi-dataset stress tests.

Next action

Screen every dataset for cell context, controls, replicate identity, licensing and leakage groups before use.

CRISPR DEPENDENCY AND MOLECULAR CONTEXTP2

DepMap Public 26Q1

Broad Institute DepMap · DATA-DEPMAP-26Q1
Scale
Genome-wide CRISPR dependency plus expression, mutation and copy-number releases
Context
Cancer cell models
Perturbation
CRISPR knockout dependency
Access
PUBLIC PORTAL DOWNLOADS AND EXPERIMENTAL API
Feature / metadata
VERSIONED MODEL AND MAPPING FILES AVAILABLE

Use: Contextual safety/dependency evidence and embedding input only.

Next action

Register exact release and model identifiers; never relabel cancer dependency as DMD muscle response.

Truth boundary. CELLxGENE, Arc Atlas, Tahoe-100M, VCC, Replogle, scPerturb and DepMap can support discovery, representation or benchmarking. Direct candidate-level DMD perturbation datasets in this registry: 0.

Decision-gated upgrade queue

Software can proceed while DMD prediction remains locked.

Each work item names its acceptance gate and the exact capability it can unlock.

Queue JSON →
01P0
IMPLEMENTED THIS BUILD

Publish the research landscape and asset ledger

One public route plus machine-readable asset, competitor, model and data opportunity registries.

Acceptance gate
Every external record has a source, checked date, local status, next gate and inheritance boundary.
Unlocks
Discoverability and a reproducible upgrade queue; no prediction claim.
02P0
IMPLEMENTED THIS BUILD

Standardize four product cards and four flagship entry points

Dataset, Model, Benchmark and Workflow/Outcome contracts plus Evidence Atlas, Virtual Cell Models, Experiment Engine and Outcome Registry homepage routes.

Acceptance gate
Each contract declares required sections, object links, current state, next gate and claim boundary in both the website and machine-readable registry.
Unlocks
A coherent product architecture and reusable publishing rules; no new biological claim.
03P0
NEXT EXTERNAL DEPENDENCY

Recover author-missing feature and metadata files

Checksummed feature, sample, perturbation and processing identity objects.

Acceptance gate
Column semantics, sample mapping, perturbation identities and provenance are complete and internally consistent.
Unlocks
A defensible reanalysis decision; not automatic model training.
04P1
READY FOR DATA DISCOVERY

Register donor-resolved neuromuscular single-cell objects

At least two source-linked datasets with donor, disease, tissue, state, batch, matrix and license fields.

Acceptance gate
Pseudobulk-capable donor identity is preserved; source H5AD or equivalent matrices are checksum-addressable.
Unlocks
A real cell browser with linked expression and state views.
05P1
READY FOR ENGINEERING

Implement one external model adapter on a compatible public benchmark

STATE or CellOT adapter plus identical frozen data, split, baseline and metric contracts.

Acceptance gate
The adapter beats prespecified simple baselines across required seeds and metric families; failures remain public.
Unlocks
A reproducible generalist benchmark result, not DMD transfer.
06P2
BLOCKED BY REGISTERED CELL OBJECTS

Add Vitessce-style linked cell-state views

Gene, embedding, donor, heatmap and optional spatial views over registered objects.

Acceptance gate
Selections resolve to source dataset and donor-aware analysis objects; no cell-level pseudo-replication.
Unlocks
Interactive biological exploration with auditable statistical units.
07LOCKED
NOT AUTHORIZED BY CURRENT TRUTH

Calibrated candidate-level DMD response model

No deliverable is claimed in the current release.

Acceptance gate
Requires independent disease-relevant perturbation outcomes, frozen predictions, calibration and external replication.
Unlocks
Only then could a bounded DMD prediction claim be reconsidered.

Scientific boundary

External models, datasets and platform practices are opportunities or design references. They are not inherited evidence, local execution results or validated DMD predictions.

Open local ModelRunsOpen registered datasetsOpen legacy blueprint