{ "registry_schema": "nmd-vcell-technical-method-registry/1.0", "interface_build": "EA-20260817-57", "resource_release": "v1.2.0-measured-dmd-evidence", "checked_at": "2026-08-16", "route": "/resource/technology-methods/", "data_contract": { "object_shape": "AnnData-shaped typed dataset", "matrix": "X = declared analysis matrix; raw counts are retained in a named layer when available", "observations": [ "cell/sample identifier", "donor or culture", "batch/assay", "disease", "cell type/state", "perturbation", "dose", "time", "control class" ], "variables": [ "stable gene identifier", "symbol", "feature selection state", "reference genome" ], "layers": [ "raw/counts", "normalized", "model input only when transformation is named" ], "unstructured": [ "source accession", "license", "checksums", "processing lineage", "task eligibility", "split hash" ], "embedding_rule": "obsm embeddings are derived views and never replace the declared source matrix or provenance." }, "baseline_firewall": { "permanent_controls": [ "zero change", "control/train mean", "perturbed or matching mean where legal", "ridge/linear", "nearest neighbor where legal" ], "same_coverage_rule": "Every comparator uses the same training objects, held-out units, feature set, seeds and metric implementation as the advanced model.", "advancement_rule": "A complex model advances only when it adds perturbation-specific, distributional or calibrated value beyond simple controls on the frozen task.", "stop_rule": "If the task lacks matched truth, the correct result is an abstention plus a prospective study—not a proxy score." }, "architecture_layers": [ { "layer_id": "TECH-01", "title": "Identity & ontology", "question": "Are genes, cells, diseases, perturbations and assays named consistently?", "input": "source identifiers + metadata", "artifact": "canonical IDs, ontology terms and ambiguity receipts", "gate": "No silent alias, species or cell-type coercion", "local_route": "/resource/search" }, { "layer_id": "TECH-02", "title": "Dataset contract", "question": "Can a machine and a reviewer reconstruct what each matrix dimension means?", "input": "matrix + sample/cell + feature metadata", "artifact": "typed AnnData-shaped Dataset Card with provenance", "gate": "X, obs, var, layers and uns roles are explicit", "local_route": "/resource/datasets/" }, { "layer_id": "TECH-03", "title": "QC & harmonization", "question": "Which technical effects were measured, corrected or left unresolved?", "input": "raw/count layers + donor, batch and assay metadata", "artifact": "QC receipt, excluded units and transformation lineage", "gate": "Biological replicates remain the inferential unit", "local_route": "/resource/data-universe/" }, { "layer_id": "TECH-04", "title": "Task & split builder", "question": "What is predicted, and along which axis must it generalize?", "input": "context × perturbation × time × output contract", "artifact": "frozen estimand, split, leakage checks and holdout identity", "gate": "Target, context and output are unseen exactly as declared", "local_route": "/resource/benchmarks/generalization/" }, { "layer_id": "TECH-05", "title": "Permanent baseline lane", "question": "Does the task require a complex model at all?", "input": "the identical training set, genes, split and metric code", "artifact": "zero-change, mean, linear and compatible neighbor results", "gate": "Advanced models cannot advance without same-coverage controls", "local_route": "/resource/benchmarks/" }, { "layer_id": "TECH-06", "title": "Model adapter", "question": "Can a model family consume this contract and return the required object?", "input": "frozen task + versioned model/configuration", "artifact": "ModelRun with mean vector, population or uncertainty output", "gate": "Architecture capability is not inferred from model name", "local_route": "/resource/models/" }, { "layer_id": "TECH-07", "title": "Evaluation & gate", "question": "Did the model recover perturbation-specific and biologically useful signal?", "input": "predictions + held-out truth + permanent baselines", "artifact": "six-level scorecard, uncertainty and abstention decision", "gate": "One aggregate expression metric cannot unlock a claim", "local_route": "/resource/no-wet-lab-evaluation/" }, { "layer_id": "TECH-08", "title": "Registry & outcome return", "question": "Can the result be reproduced, challenged and updated by an experiment?", "input": "ModelRun, Study, Prediction and measured Outcome", "artifact": "immutable receipt and evidence-graph return", "gate": "Null, toxic and failed outcomes are retained", "local_route": "/resource/registry" } ], "model_families": [ { "family_id": "MF-01", "title": "Linear & non-parametric baselines", "core_mechanism": "Predict no change, a matched mean, a regularized linear response or a nearest observed neighbor.", "best_fit_question": "Is there learnable signal beyond systematic assay and context structure?", "input_contract": "matched feature space; frozen split; context labels for matching where permitted", "output_contract": "aggregate response vector or neighbor-based reference", "strengths": [ "transparent", "low variance", "fast leakage diagnostic", "permanent comparator" ], "failure_modes": [ "cannot represent complex interactions", "matching can leak held-out context", "good mean error may miss perturbation identity" ], "minimum_data": "one task-compatible training response set plus held-out truth", "permanent_baselines": [ "zero change", "train/control mean", "perturbed or matching mean when legal", "ridge", "nearest neighbor when legal" ], "evaluation_requirements": [ "mean fidelity", "perturbation-specific delta", "split integrity" ], "reference_methods": [ "Systema", "ridge", "matching mean" ], "local_state": "LOCALLY_EXECUTED_LIMITED", "local_runs": [ "MRUN-RIDGE-SAFE-2.3-G0-REPEATED-FOLD" ], "next_gate": "Retain ridge as a permanent same-assay comparator; do not extend it to DMD transfer." }, { "family_id": "MF-02", "title": "Factorized latent generative models", "core_mechanism": "Disentangle basal cell state, perturbation and covariates in a latent representation, then compose an unseen condition.", "best_fit_question": "Can known factors be recombined across dose, time or cell context?", "input_contract": "cell-level expression with perturbation, covariate, dose/time and control labels", "output_contract": "counterfactual cells or an expected post-perturbation distribution", "strengths": [ "compositional representation", "covariate conditioning", "counterfactual generation" ], "failure_modes": [ "disentanglement is not guaranteed", "performance declines as unseen covariates accumulate", "batch may be encoded as biology" ], "minimum_data": "factorial overlap across perturbation and covariate combinations", "permanent_baselines": [ "matching mean", "context-conditioned ridge", "nearest observed condition" ], "evaluation_requirements": [ "mean fidelity", "delta recovery", "distribution fidelity", "OOD covariate stress test" ], "reference_methods": [ "CPA", "scGen", "biolord" ], "local_state": "SOURCE_VERIFIED_NOT_RUN", "local_runs": [], "next_gate": "Run only after a factorial task exposes which covariate combinations are genuinely unseen." }, { "family_id": "MF-03", "title": "Graph-informed perturbation models", "core_mechanism": "Propagate gene and perturbation information over co-expression, ontology or learned graphs.", "best_fit_question": "Can structured gene relationships improve unseen-gene or combination response prediction?", "input_contract": "perturbation responses plus gene identities and a versioned graph or embedding source", "output_contract": "post-perturbation mean expression or effect vector", "strengths": [ "uses gene relationships", "supports structured inductive bias", "can represent combinations" ], "failure_modes": [ "graph mismatch", "single-perturbation data may not identify combinations", "relation priors can dominate sparse truth" ], "minimum_data": "single and preferably combinatorial perturbations with stable gene coverage", "permanent_baselines": [ "additive single-perturbation baseline", "ridge", "matching mean" ], "evaluation_requirements": [ "held-out genes", "held-out combinations", "sign and interaction recovery", "seed stability" ], "reference_methods": [ "GEARS", "TxPert" ], "local_state": "LOCALLY_EXECUTED_FAILED_OR_PENDING", "local_runs": [ "MRUN-GEARS-0.1.2-FIVE-SEED-20260713", "MRUN-TXPERT-CONFIG-GAT-SEED-20260712" ], "next_gate": "GEARS remains failed on five frozen splits; TxPert remains calibration-pending. Neither unlocks DMD prediction." }, { "family_id": "MF-04", "title": "Foundation transformer & embedding models", "core_mechanism": "Pretrain token or rank-based cell representations at scale, then adapt embeddings or decoders to a downstream task.", "best_fit_question": "Does broad pretraining improve a precisely frozen perturbation or cell-state task?", "input_contract": "gene-aligned expression plus the exact tokenizer, vocabulary, checkpoint and adaptation recipe", "output_contract": "cell embeddings, labels or decoded response vectors depending on the adapter", "strengths": [ "broad representation prior", "transferable embeddings", "large reference context" ], "failure_modes": [ "embedding quality is not perturbation accuracy", "vocabulary/context mismatch", "scale can obscure task leakage" ], "minimum_data": "checkpoint-compatible genes plus task-specific adaptation and held-out truth", "permanent_baselines": [ "PCA or linear embedding", "ridge", "train mean", "task-specific shallow model" ], "evaluation_requirements": [ "task-level baseline comparison", "OOD split", "ablation of pretraining", "reproduction receipt" ], "reference_methods": [ "scGPT", "Geneformer", "scFoundation", "CellFM" ], "local_state": "LOCALLY_EXECUTED_FAILED", "local_runs": [ "MRUN-SCGPT-0.2.5-FIVE-SEED-20260714" ], "next_gate": "The local scGPT adapter lost to same-coverage baselines; a new checkpoint is a new frozen ModelRun, not an inherited upgrade." }, { "family_id": "MF-05", "title": "Set-to-set & context-prompted models", "core_mechanism": "Represent a cell population as a set and condition one set of cells on another context or prompt.", "best_fit_question": "Can a model predict population transitions while using context examples at inference time?", "input_contract": "cell sets, context labels, perturbation identity and a task-compatible feature universe", "output_contract": "predicted cell set, state embedding or transition distribution", "strengths": [ "population-native", "context prompting", "heterogeneity-aware representation" ], "failure_modes": [ "prompt leakage", "set composition confounding", "source-reported scale may not transfer to disease context" ], "minimum_data": "multiple comparable cell populations per context and held-out population truth", "permanent_baselines": [ "matching population", "stratified mean", "optimal-transport baseline" ], "evaluation_requirements": [ "distribution fidelity", "composition recovery", "prompt ablation", "unseen-context holdout" ], "reference_methods": [ "Arc State", "Arc State/Stack" ], "local_state": "COMPATIBILITY_ROUTE_EXECUTED_NO_ADVANCEMENT", "local_runs": [ "NMD-B1-COMPAT-HEPG2-R1-20260816" ], "next_gate": "STATE: Do not rerun by reputation; require a corrected preregistration, untouched truth or materially different task-compatible checkpoint. Stack: Identify a task-compatible pretrained genetic checkpoint and preregister an untouched benchmark before any model-performance claim." }, { "family_id": "MF-06", "title": "Optimal-transport population maps", "core_mechanism": "Learn a transport map from an unpaired control population to a treated population.", "best_fit_question": "How does a distribution of control cells move under treatment when cells are not paired?", "input_contract": "unpaired control and treated cell populations with shared features and sufficient state coverage", "output_contract": "transported cells or a treatment-conditioned population distribution", "strengths": [ "distributional output", "unpaired design", "population geometry" ], "failure_modes": [ "rare states are unstable", "transport assumptions may not identify biology", "composition shifts can mimic state transitions" ], "minimum_data": "adequate cells across represented states in both conditions", "permanent_baselines": [ "identity map", "mean shift", "nearest-neighbor transport" ], "evaluation_requirements": [ "MMD or energy distance", "Wasserstein distance", "state composition", "rare-state stratification" ], "reference_methods": [ "CellOT" ], "local_state": "SOURCE_VERIFIED_NOT_RUN", "local_runs": [], "next_gate": "Use only for a population-output task with explicit rare-state and composition stress tests." }, { "family_id": "MF-07", "title": "Probabilistic sparse-effect models", "core_mechanism": "Estimate interpretable perturbation effects with a probabilistic prior and calibrated uncertainty.", "best_fit_question": "Which gene-level effects are supported, and where should the model abstain?", "input_contract": "replicated perturbation responses with biological units, covariates and stable features", "output_contract": "effect posterior, uncertainty interval and sparse active set", "strengths": [ "uncertainty-aware", "interpretable effects", "appropriate for sparse signals" ], "failure_modes": [ "prior sensitivity", "poor scaling", "cell-level pseudo-replication produces false confidence" ], "minimum_data": "independent biological replication, not cells treated as replicates", "permanent_baselines": [ "regularized linear model", "empirical Bayes shrinkage", "no-effect model" ], "evaluation_requirements": [ "interval coverage", "calibration", "effect-sign recovery", "donor-level resampling" ], "reference_methods": [ "GPerturb" ], "local_state": "SOURCE_VERIFIED_NOT_RUN", "local_runs": [], "next_gate": "Prioritize when replicated DMD perturbation outcomes exist and uncertainty is decision-critical." }, { "family_id": "MF-08", "title": "Distributional flow & diffusion generators", "core_mechanism": "Learn a conditional generative process that samples heterogeneous post-perturbation cells.", "best_fit_question": "Can the full conditional response distribution be generated rather than only its mean?", "input_contract": "large cell-level perturbation datasets with context, dose/time and robust controls", "output_contract": "sampled post-perturbation cell population", "strengths": [ "multimodal distributions", "heterogeneity", "sample-level counterfactuals" ], "failure_modes": [ "plausible-looking hallucinated states", "mode collapse", "weak calibration", "high compute burden" ], "minimum_data": "large, balanced and context-rich perturbation populations", "permanent_baselines": [ "matching population", "CellOT or transport baseline", "conditional Gaussian baseline" ], "evaluation_requirements": [ "distribution metrics", "mode coverage", "calibration", "biological state validity", "prospective function" ], "reference_methods": [ "conditional flow matching", "diffusion perturbation models" ], "local_state": "WATCHLIST_NOT_RUN", "local_runs": [], "next_gate": "Do not adopt until the population task, compute budget and prospective validation route are frozen." } ], "evaluation_ladder": [ { "level_id": "EVAL-01", "title": "Mean fidelity", "question": "Is the average predicted profile numerically close?", "measures": [ "RMSE", "MAE", "mean correlation" ], "required_baselines": [ "zero change", "train/control mean", "ridge" ], "failure_caught": "gross reconstruction error", "cannot_prove": "perturbation identity, heterogeneity or biological function" }, { "level_id": "EVAL-02", "title": "Perturbation-specific signal", "question": "Did the model recover the change caused by this perturbation rather than systematic variation?", "measures": [ "delta correlation/cosine with amplitude checks", "signed DE recovery", "precision/recall of responsive genes", "rank recovery" ], "required_baselines": [ "matching mean", "perturbed mean", "additive effect baseline" ], "failure_caught": "mean predictors that ignore perturbation identity", "cannot_prove": "cell-population fidelity or function" }, { "level_id": "EVAL-03", "title": "Population distribution", "question": "Do predicted cells occupy the held-out treated distribution?", "measures": [ "MMD", "energy distance", "Wasserstein", "classifier two-sample test" ], "required_baselines": [ "matching population", "identity transport", "stratified mean" ], "failure_caught": "correct mean with wrong spread or geometry", "cannot_prove": "correct cell-state composition or mechanism" }, { "level_id": "EVAL-04", "title": "State composition & heterogeneity", "question": "Are rare states, proportions and response modes preserved?", "measures": [ "cell-state proportion error", "rare-state recall", "mode coverage", "stratified distribution metrics" ], "required_baselines": [ "observed control composition", "matched-context composition" ], "failure_caught": "mode collapse and composition confounding", "cannot_prove": "uncertainty reliability or disease relevance" }, { "level_id": "EVAL-05", "title": "Calibration, OOD & abstention", "question": "Does uncertainty increase where the task leaves the training support?", "measures": [ "interval coverage", "calibration error", "selective risk", "OOD-stratified performance", "seed stability" ], "required_baselines": [ "empirical residual interval", "no-model abstention rule" ], "failure_caught": "confident extrapolation and unstable wins", "cannot_prove": "therapeutic efficacy or clinical utility" }, { "level_id": "EVAL-06", "title": "Biological & prospective validation", "question": "Does molecular fidelity translate into a reproducible, disease-relevant functional result?", "measures": [ "independent donor replication", "prespecified functional endpoint", "safety/toxicity", "prospective outcome return" ], "required_baselines": [ "negative and positive experimental controls", "prespecified null decision" ], "failure_caught": "molecular proxies that do not alter function", "cannot_prove": "clinical benefit without a separate clinical design" } ], "knowledge_tracks": [ { "track_id": "KNOW-01", "title": "Data truth", "outcome": "Read a matrix as a biological measurement contract, not an anonymous tensor.", "lessons": [ { "title": "What is the inferential unit?", "mental_model": "cells are observations nested in donors or cultures", "failure_mode": "pseudo-replication", "route": "/resource/datasets/" }, { "title": "What does each AnnData slot mean?", "mental_model": "X + obs + var + layers + uns form one typed object", "failure_mode": "opaque transformed inputs", "route": "/resource/data-universe/" }, { "title": "Which context is missing?", "mental_model": "storage presence is not task qualification", "failure_mode": "using DMD context as DMD perturbation outcome", "route": "/resource/prediction_readiness" } ] }, { "track_id": "KNOW-02", "title": "Model mechanisms", "outcome": "Choose an architecture from its assumptions, input and output—not from its reputation.", "lessons": [ { "title": "Mean vector or cell population?", "mental_model": "the output object determines the model family", "failure_mode": "calling an aggregate vector a virtual-cell population", "route": "/resource/models/" }, { "title": "Where does context enter?", "mental_model": "covariate, graph, prompt and transport encode different assumptions", "failure_mode": "architecture-name inference", "route": "/resource/technology-methods/#model-families" }, { "title": "What was actually executed?", "mental_model": "paper capability and local ModelRun are separate ledgers", "failure_mode": "inheriting external performance", "route": "/resource/model-run/mrun-ridge-safe-2-3-g0-repeated-fold/" } ] }, { "track_id": "KNOW-03", "title": "Evaluation without shortcuts", "outcome": "Detect trivial mean predictors, leakage, mode collapse and overconfident extrapolation.", "lessons": [ { "title": "Why simple baselines are permanent", "mental_model": "systematic variation can make trivial predictors look strong", "failure_mode": "complexity bias", "route": "/resource/benchmarks/" }, { "title": "Why one metric is insufficient", "mental_model": "mean, delta, distribution and function answer different questions", "failure_mode": "metric substitution", "route": "/resource/no-wet-lab-evaluation/" }, { "title": "How OOD axes change the claim", "mental_model": "unseen perturbation, donor and disease are different tasks", "failure_mode": "generalization collapse", "route": "/resource/benchmarks/generalization/" } ] }, { "track_id": "KNOW-04", "title": "From prediction to experiment", "outcome": "Convert uncertainty or abstention into a study whose outcome can update the evidence graph.", "lessons": [ { "title": "Freeze the estimand", "mental_model": "context × perturbation × time × endpoint is the scientific unit", "failure_mode": "post-hoc endpoint switching", "route": "/resource/planner" }, { "title": "Return every outcome", "mental_model": "positive, null, toxic and failed results all reduce uncertainty", "failure_mode": "success-only evidence", "route": "/resource/registry" }, { "title": "Know the final claim ceiling", "mental_model": "functional validation is not clinical validation", "failure_mode": "therapeutic overclaim", "route": "/resource/claim_boundary" } ] } ], "source_matrix": [ { "source_id": "SRC-SYSTEMA-2025", "title": "Systematic evaluation of perturbation-response predictors and simple matching baselines", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2025, "url": "https://www.nature.com/articles/s41587-025-02777-8", "supports": [ "mandatory perturbed/matching mean baselines", "perturbation-specific evaluation", "warning that common metrics reward systematic variation" ], "does_not_support": [ "a universal winning architecture", "local NMD-VCell performance" ], "local_execution_state": "METHOD_RULE_ABSORBED" }, { "source_id": "SRC-NM-BENCHMARK-2025", "title": "Benchmarking 27 perturbation-response methods across 29 datasets", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2025, "url": "https://www.nature.com/articles/s41592-025-02980-0", "supports": [ "task- and context-dependent evaluation", "multiple datasets and metrics", "importance of cellular context" ], "does_not_support": [ "one method as best for every generalization axis" ], "local_execution_state": "METHOD_RULE_ABSORBED" }, { "source_id": "SRC-SIMPLE-BASELINES-2025", "title": "Deep learning perturbation models do not consistently beat deliberate simple baselines", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2025, "url": "https://www.nature.com/articles/s41592-025-02772-6", "supports": [ "permanent simple baselines", "architecture-neutral benchmarking" ], "does_not_support": [ "that deep learning can never be useful" ], "local_execution_state": "METHOD_RULE_ABSORBED" }, { "source_id": "SRC-CZI-BENCHMARKS", "title": "CZI Virtual Cell Models benchmark ecosystem", "source_type": "OFFICIAL_DOC", "evidence_grade": "B", "publication_year": 2026, "url": "https://virtualcellmodels.cziscience.com/benchmarks", "supports": [ "standardized task definitions", "shared metrics", "package, CLI and no-code entry points" ], "does_not_support": [ "NMD-VCell task compatibility or performance" ], "local_execution_state": "PRODUCT_PATTERN_ABSORBED" }, { "source_id": "SRC-ARC-STATE", "title": "Arc State virtual-cell architecture", "source_type": "OFFICIAL_PRODUCT_PREPRINT", "evidence_grade": "B", "publication_year": 2025, "url": "https://arcinstitute.org/news/virtual-cell-model-state", "supports": [ "set-of-cells representation", "separate state embedding and transition tasks", "evaluation beyond mean expression" ], "does_not_support": [ "advancement beyond permanent baselines", "validated DMD prediction" ], "local_execution_state": "HISTORICAL_RECEIPTS_MIGRATED" }, { "source_id": "SRC-ARC-STACK", "title": "Arc foundation model stack and in-context learning", "source_type": "OFFICIAL_PRODUCT_PREPRINT", "evidence_grade": "B", "publication_year": 2026, "url": "https://arcinstitute.org/news/foundation-model-stack", "supports": [ "cell-set prompting as a model pattern", "source-reported large-scale pretraining" ], "does_not_support": [ "checkpoint benchmark after runtime timeout", "NMD-VCell performance inheritance" ], "local_execution_state": "LOCAL_NVME_CORE_BASE_FORWARD_PASS_NO_CHECKPOINT" }, { "source_id": "SRC-PERTPY-2025", "title": "Pertpy perturbation-analysis framework", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2025, "url": "https://www.nature.com/articles/s41592-025-02909-7", "supports": [ "end-to-end perturbation analysis", "MMD, energy and Wasserstein distances", "typed analysis modules" ], "does_not_support": [ "a disease-response prediction result" ], "local_execution_state": "METHOD_REFERENCE" }, { "source_id": "SRC-ANNDATA", "title": "AnnData annotated matrix and on-disk specification", "source_type": "OFFICIAL_DOC", "evidence_grade": "A", "publication_year": 2026, "url": "https://anndata.readthedocs.io/en/stable/generated/anndata.AnnData.html", "supports": [ "X/obs/var/layers/obsm/uns data contract", "typed annotated matrices" ], "does_not_support": [ "biological correctness of any uploaded object" ], "local_execution_state": "DATA_CONTRACT_ABSORBED" }, { "source_id": "SRC-CELLXGENE-SCHEMA", "title": "CELLxGENE schema 5.2.0", "source_type": "OFFICIAL_DOC", "evidence_grade": "A", "publication_year": 2026, "url": "https://github.com/chanzuckerberg/single-cell-curation/blob/main/schema/5.2.0/schema.md", "supports": [ "ontology-backed biological and technical metadata", "schema as a search and integration contract" ], "does_not_support": [ "automatic suitability for a perturbation task" ], "local_execution_state": "ONTOLOGY_PATTERN_ABSORBED" }, { "source_id": "SRC-GEARS-2023", "title": "GEARS graph neural network for perturbation prediction", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2023, "url": "https://www.nature.com/articles/s41587-023-01905-6", "supports": [ "graph-informed gene and perturbation embeddings", "structured combination prediction task" ], "does_not_support": [ "success on the local frozen task", "reliable combinations from single perturbations alone" ], "local_execution_state": "LOCALLY_EXECUTED_FAILED_FIVE_SPLITS" }, { "source_id": "SRC-CELLOT-2023", "title": "CellOT neural optimal transport", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2023, "url": "https://www.nature.com/articles/s41592-023-01969-x", "supports": [ "unpaired control-to-treated population mapping", "distributional prediction" ], "does_not_support": [ "stable inference in sparse cell types", "local DMD validation" ], "local_execution_state": "SOURCE_VERIFIED_NOT_RUN" }, { "source_id": "SRC-CPA-2023", "title": "Compositional perturbation autoencoder", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2023, "url": "https://doi.org/10.15252/msb.202211517", "supports": [ "factorized perturbation and covariate representations", "dose/time/context composition" ], "does_not_support": [ "constant performance as unseen covariates accumulate", "local NMD-VCell calibration" ], "local_execution_state": "SOURCE_VERIFIED_NOT_RUN" }, { "source_id": "SRC-GPERTURB-2025", "title": "GPerturb probabilistic perturbation-effect model", "source_type": "PEER_REVIEWED_PRIMARY", "evidence_grade": "A", "publication_year": 2025, "url": "https://www.nature.com/articles/s41467-025-61165-7", "supports": [ "sparse interpretable effects", "uncertainty estimates" ], "does_not_support": [ "treating cells as independent biological replicates", "clinical prediction" ], "local_execution_state": "SOURCE_VERIFIED_NOT_RUN" }, { "source_id": "SRC-OPENPROBLEMS-PERTURBATION", "title": "Open Problems perturbation prediction benchmark", "source_type": "OFFICIAL_DOC", "evidence_grade": "B", "publication_year": 2024, "url": "https://openproblems.bio/benchmarks/perturbation_prediction/?version=v1.0.0", "supports": [ "versioned task and metric metadata", "automated score checking" ], "does_not_support": [ "generalization outside the declared competition task" ], "local_execution_state": "BENCHMARK_PATTERN_ABSORBED" } ], "canonical_model_status_registry": "/resource/api/v1.1/model_execution_status_registry.json", "design_rules": [ "Freeze task, biological unit, output object and generalization axis before selecting a model.", "Separate source-reported capability, local execution and disease validation in every registry.", "Use donor, culture or independently randomized sample as the inferential unit; cells are not independent biological replicates.", "Do not equate embedding quality with perturbation prediction.", "Do not let mean-profile accuracy substitute for perturbation-specific or population-level fidelity.", "Publish failed, null, stopped and abstained states with the same object identity as successful runs." ], "current_local_decision": "Keep ridge as a permanent same-assay baseline. TxPert: STOP_UNTIL_NEW_PREREGISTRATION_OR_UNTOUCHED_TRUTH. STATE: NO_ADVANCEMENT. Stack: NO_NMD_B1_BENCHMARK_NO_COMPATIBLE_PRETRAINED_GENETIC_CHECKPOINT. Tahoe-x1: DO_NOT_RUN_ON_LOG_NORMALIZED_INPUT; ACQUIRE_RAW_COUNTS_FIRST. SLIM: NOT_ELIGIBLE_FOR_EXECUTION. None is a locally validated NMD model.", "claim_boundary": "This registry is a technical decision aid. It does not claim that every model family was executed, that source-reported performance transfers to NMD-VCell, or that DMD candidate responses are calibrated." }