Competitive benchmark to implementation blueprint

Learn, borrow and apply without overclaiming.

This roadmap turns the 31 July 2026 competitive review into a machine-readable implementation contract for NMD-VCell. It records what the platform should learn from Arc, CZI, CELLxGENE, Open Targets, DepMap, VCell and CMap, and exactly what must be true before any prediction claim changes.

Open blueprint JSON Download TSV Prediction gate

Disease-specific pathCompete by DMD specificity, prospective freezing, failed-result return and object governance; competitor lessons are implementation gates, not inherited evidence.

Current maturity translation

The benchmark report becomes five execution gates.

Competitor lessons encoded6

Implementation axes5

DMD prediction unlock0

Outcome statusNot measured yet

Learn · borrow · apply

Competitor practices are translated into NMD-specific contracts.

Reference pattern	Learn	Borrow	Apply to NMD-VCell	Acceptance gate
Arc STATE and Virtual Cell Challenge	A model claim must be bound to task, split, hidden labels, frozen submission and permanent simple baselines.	Prediction-freeze, hidden-test and baseline-retention rules.	Create a DMD Prospective Benchmark where Study Card, Prediction Card and Outcome object are separate immutable records.	No DMD prediction label is shown until a timestamped prediction exists before outcome access and simple baselines run on the same split.
CZI Virtual Cells Platform	Data, model, benchmark and inference should use one ModelRun contract across web, CLI and local execution.	Model adapter, run log, dataset digest, split manifest and reusable benchmark package.	Preserve ridge, transfer, GEARS, scGPT, TxPert and MORPH history in one ledger, then route every future baseline and population model through the same ModelRun object.	Every run records model id, revision status, container status, data SHA status, split, seed, status, metrics and failure reason; missing migrated provenance remains explicit.
CELLxGENE and Vitessce-style browsers	A Cell Browser must let users operate real cell objects, not only read concept cards.	Dataset, donor, disease, state, batch, expression, DE, cell-proportion and spatial-neighbour filters.	Upgrade Cell Browser v1 around real DMD scRNA/snRNA/spatial data with donor-level pseudobulk as the default statistical unit.	At least two DMD datasets can be filtered by donor and disease, with expression, cell count, pseudobulk DE and exportable selections.
Open Targets and DepMap	Long-lived platforms are entity graphs plus stable programmatic file indices, not only pages.	Entity relation graph, API version metadata, file index, historical releases and provenance by source.	Expose Gene, Dataset, Comparison, Study, ModelRun, Prediction and Outcome as linked objects under one release context.	Each page answers: what object is studied, what evidence supports it, and which next object would change the conclusion.
VCell mechanism simulation platform	Mechanistic models separate model definition, parameters, solver, simulation run and result object.	Solver/run/result separation and explicit parameter documentation.	Keep Computability Gate, ModelRun and future mechanistic simulation objects distinct so the name virtual cell does not imply unavailable numerical simulation.	UI never presents evidence checks as simulation output; future solver objects must cite equations, parameters, run logs and result files.
CLUE/CMap service continuity lesson	A valuable scientific portal can lose utility if data are only reachable through an active website.	Portable archive, checksum ledger, mirror plan, historical manifest and exit strategy.	Keep RO-Crate, release manifest, file registry, checksums and deprecated archive policy as first-class release objects.	A release can be reconstructed from manifest, checksums, schemas and archived files without relying on the interactive site.

Applied workstreams

Where this changes the platform.

IMPLEMENTED_INTERNAL_GATE

Release engineering

Health, route manifest and overlay identity are generated and checked.

Apply: Keep manifest, downloads, overlay, health and route QA as a single generated release context.

api/health api/v1.1/release_manifest.json api/v1.1/overlay_archive_identity.json api/v1.1/route_manifest.json

PARTIAL_IMPLEMENTED

Deep-link research objects

Planner can restore gene state and object trace server-side.

Apply: Extend parent inheritance from Comparison to Study Card, Prediction and Outcome object URLs.

/resource/planner?gene=ZNF133 api/v1.1/study-cards/ZNF133.json

PLANNED_REQUIRES_DATA_IMPORT

Real Cell Browser

Cell Browser is a disease ecology and computability map.

Apply: Connect real DMD single-cell, single-nucleus and spatial datasets with donor-level filters and pseudobulk DE.

api/v1.1/dataset_registry.json

PARTIAL_LEDGER_RELEASED_RUNNER_NOT_UNIFIED

Executable ModelRun system

Six historical ModelRuns are frozen; negative and stopped results are visible; calibrated DMD model count is zero.

Apply: Route future runs through the immutable schema and backfill commit, container and data digests where migrated records identify a provenance gap.

api/v1.1/model_registry.json api/v1.1/model_run_registry.json api/v1.1/model_run.schema.json api/v1.1/benchmark_suite.json

PLANNED_REQUIRES_EXPERIMENT_AND_WRITE_PATH

Prospective validation loop

Study Cards are drafts; predictions and measured outcomes are zero.

Apply: Create append-only Study -> Prediction -> Experiment -> Outcome chain with null, toxic, failed-QC and abstained outcomes preserved.

api/v1.1/study_card_manifest.json api/v1.1/prediction_registry.json

Research object 2.0 scaffold

The next release should promote ModelRun, Dataset and Outcome from text to contracts.

Research object envelope

object_id
object_type
schema_version
lifecycle_status
release_context
identity.parents
biological_context
perturbation
endpoint
evidence_state
provenance
files
validation

Dataset v2 required fields

accession
permanent_identifier
license
species
disease
donor_count
cell_count
modality
perturbation_type
endpoint
raw_sha256
processed_sha256
benchmark_roles
permitted_model_uses
prohibited_uses

ModelRun required fields

model_id
model_version
source_commit
checkpoint_sha256
container_digest
data_release
feature_schema
split_id
random_seed
status
failure_reason
predicted_object_ids
metric_object_ids
stdout_log_sha256
stderr_log_sha256

Twelve-month route

Progression is gated by evidence, not branding.

Horizon	Rule	NMD status
0-30 days	Release context, health, route contract, deep-link SSR and missing-data/reproducibility inventory.	MOSTLY_IMPLEMENTED_IN_V93_INTERNAL_GATE
1-3 months	DMD Cell Browser v1, Dataset Registry v2, ModelAdapter baselines and Study Card state machine.	NEXT_SOFTWARE_PHASE
3-6 months	Minimum Perturbome freeze, hidden prospective roles, Prediction Card and Outcome alpha.	REQUIRES_GOVERNED_STUDY_DESIGN
6-12 months	At least one independent DMD Study -> Prediction -> Outcome chain with failures preserved.	REQUIRES_EXPERIMENTAL_OUTCOME

Boundary. This page does not assert data-scale parity with Arc, model-platform parity with CZI, CELLxGENE-level browsing, validated DMD perturbation prediction, clinical decision support or therapeutic target ranking. v1.0 is a database and evidence-governance release; it is not a validated disease-prediction or clinical decision-support release.