{
  "governance_schema": "nmd-vcell-metric-governance/1.0",
  "interface_build": "EA-20260730-36",
  "resource_release": "v1.0.0-database-resource",
  "aggregate_score_policy": "No single composite score defines model success.",
  "metric_families": [
    {
      "family": "mean_fidelity",
      "metrics": [
        "MAE",
        "RMSE",
        "correlation"
      ],
      "purpose": "Average-expression fidelity and simple-baseline comparison"
    },
    {
      "family": "differential_response",
      "metrics": [
        "DES",
        "signed_log_fold_change",
        "AUPRC"
      ],
      "purpose": "Recovery of perturbation-responsive genes and direction"
    },
    {
      "family": "perturbation_identity",
      "metrics": [
        "PDS_raw",
        "PDS_truth_norm_matched",
        "PDS_prediction_norm_matched"
      ],
      "purpose": "Identity discrimination with explicit amplitude controls"
    },
    {
      "family": "population_geometry",
      "metrics": [
        "MMD",
        "energy_distance",
        "Wasserstein_or_optimal_transport"
      ],
      "purpose": "Distribution-level agreement"
    },
    {
      "family": "state_composition",
      "metrics": [
        "state_proportion_error",
        "composition_divergence"
      ],
      "purpose": "Responder fraction and state-occupancy shifts"
    },
    {
      "family": "uncertainty",
      "metrics": [
        "coverage",
        "Brier_score",
        "abstention_quality"
      ],
      "purpose": "Calibration, OOD detection and useful refusal"
    },
    {
      "family": "dmd_utility",
      "metrics": [
        "functional_hit_rate",
        "replication_rate",
        "toxicity_false_negative_rate"
      ],
      "purpose": "Disease-relevant function, reproducibility and safety"
    }
  ],
  "pds_safeguards": [
    "Report raw and norm-matched variants together",
    "Run amplitude scaling and normalization sensitivity analyses",
    "Freeze preprocessing and distance definition before scoring",
    "Never interpret PDS without DES, mean-fidelity and distribution metrics"
  ],
  "baseline_rule": "Every advanced model is compared with zero, train-mean, pseudobulk ridge and nearest-neighbour baselines on the same holdout.",
  "sources": [
    {
      "label": "Arc Virtual Cell Challenge metric explainer",
      "source_type": "official",
      "url": "https://arcinstitute.org/news/behind-the-data-virtual-cell-challenge"
    },
    {
      "label": "PDS scale-sensitivity analysis",
      "source_type": "preprint",
      "url": "https://arxiv.org/abs/2511.16954"
    }
  ]
}