sccoda.json

{
  "tool_id": "sccoda",
  "name": "scCODA Compositional Analysis",
  "category": "composition",
  "framework": "sccoda",
  "function": "sccoda.util.cell_composition_data.from_scanpy + sccoda.model",
  "valid_after": [
    "annotation"
  ],
  "valid_before": [],
  "paradigms": [
    "disease_vs_healthy",
    "temporal_longitudinal"
  ],
  "parameters": {
    "cell_type_key": {
      "type": "string",
      "required": true,
      "guidance": "obs column with cell type labels."
    },
    "sample_key": {
      "type": "string",
      "required": true,
      "guidance": "obs column identifying biological samples."
    },
    "condition_key": {
      "type": "string",
      "required": true,
      "guidance": "obs column with condition labels."
    },
    "reference_cell_type": {
      "type": "string",
      "default": "automatic",
      "guidance": "Reference cell type for compositional analysis. Choose a cell type believed to be unaffected by the condition. If 'automatic', scCODA selects one."
    },
    "fdr_threshold": {
      "type": "float",
      "default": 0.05,
      "guidance": "FDR threshold for credible changes."
    }
  },
  "outputs": {
    "credible_effects": "Which cell types show credible compositional changes",
    "effect_sizes": "Log-fold change in composition per cell type"
  },
  "validation": {
    "check": "Model should converge. At least one cell type should be testable."
  },
  "notes": "Compositional analysis is necessary because cell type proportions sum to 1. An increase in one type necessarily decreases others. Simple proportion tests (e.g., t-test on fractions) are WRONG because they ignore this constraint. scCODA uses a Bayesian model that properly handles the compositional nature of the data.",
  "provenance_captures": [
    "cell_type_key",
    "sample_key",
    "condition_key",
    "reference_cell_type",
    "fdr_threshold",
    "n_cell_types",
    "n_samples"
  ]
}