Spatial Condition

Input Formats

Input Matrix Convention

This skill has a multi-step pipeline with different statistical assumptions:

Core principle: pseudobulk always starts from raw counts, then the DE method operates on the aggregated matrix.

Data layout requirement:

adata.layers["counts"] = adata.X.copy()    # before normalize_total + log1p
adata.X = lognorm_expr                      # after normalize_total + log1p
adata.obs["condition"] = condition_labels   # e.g. "treated" / "control"
adata.obs["sample_id"] = sample_labels      # biological replicate IDs

If layers["counts"] is missing, OmicsClaw falls back to adata.raw or adata.X with a warning.

Workflow

1. Validate: check condition / sample columns and verify each sample maps to exactly one condition.
2. Aggregate: create pseudobulk matrices per sample x cluster from raw counts.
3. Filter: drop low-count genes and skip contrasts with too few biological replicates.
4. Test:
   • pydeseq2: fit NB GLM per cluster/contrast
   • wilcoxon: run rank-sum test on transformed pseudobulk profiles
5. Render the standard gallery: build the OmicsClaw narrative gallery with design-context, diagnostic burden, supporting summaries, and uncertainty panels.
6. Export figure-ready data: write figure_data/*.csv and figure_data/manifest.json for downstream customization.
7. Report and export: write report.md, result.json, processed.h5ad, figures, tables, and reproducibility helpers.

Visualization Contract

OmicsClaw treats spatial-condition visualization as a two-layer system:

1. Python standard gallery: the canonical result layer. This is the default output users should inspect first.
2. R customization layer: an optional styling and publication layer that reads figure_data/ and does not recompute pseudobulk DE.

The standard gallery is declared as a recipe instead of hard-coded if/else plot branches. Current gallery roles include:

• overview: condition labels on tissue and a global pseudobulk volcano
• diagnostic: cluster-level DE burden projected onto spatial and UMAP views
• supporting: per-contrast significant-gene counts and per-cluster burden
• uncertainty: adjusted p-value distributions, sample support, and skipped-contrast summaries when applicable

CLI Reference

# PyDESeq2 default run
oc run spatial-condition \
  --input <data.h5ad> --output <dir> \
  --condition-key condition --sample-key sample_id

# PyDESeq2 with explicit cluster labels and tuning
oc run spatial-condition \
  --input <data.h5ad> --method pydeseq2 \
  --condition-key condition --sample-key sample_id --cluster-key leiden \
  --reference-condition control \
  --min-counts-per-gene 10 --min-samples-per-condition 2 \
  --pydeseq2-fit-type parametric \
  --pydeseq2-size-factors-fit-type ratio \
  --pydeseq2-alpha 0.05 --pydeseq2-n-cpus 1 \
  --output <dir>

# Wilcoxon fallback mode
oc run spatial-condition \
  --input <data.h5ad> --method wilcoxon \
  --condition-key condition --sample-key sample_id \
  --reference-condition control \
  --wilcoxon-alternative two-sided \
  --output <dir>

# Demo mode
oc run spatial-condition --demo --output /tmp/cond_demo

# Direct script entrypoint
python skills/spatial/spatial-condition/spatial_condition.py \
  --input <data.h5ad> --method pydeseq2 --output <dir>

Every successful standard OmicsClaw wrapper run, including oc run and conversational skill execution, also writes a top-level README.md and reproducibility/analysis_notebook.ipynb to make the output directory easier to inspect and rerun.

Example Queries

• "Compare healthy vs disease slices with pseudobulk statistics"
• "Find treatment-responsive genes per spatial cluster"
• "Run a Wilcoxon fallback because I only have two replicates per condition"

Algorithm / Methodology

PyDESeq2 (default)

1. Input: pseudobulk raw integer counts built from sample x cluster aggregation
2. Design: OmicsClaw fits ~ condition using official DeseqDataSet(...) controls
3. Testing: OmicsClaw uses DeseqStats(...) with explicit contrast=["condition", other, reference]
4. Outputs: reports base_mean, log2fc, stat, pvalue, and adjusted pvalue_adj

Core tuning flags:

• --reference-condition: defines the sign direction of the contrast
• --cluster-key: chooses the cluster partition for per-cluster pseudobulk
• --min-counts-per-gene: wrapper-level gene filter before DE testing
• --min-samples-per-condition: wrapper-level replicate gate
• --pydeseq2-fit-type: official PyDESeq2 dispersion fit mode
• --pydeseq2-size-factors-fit-type: official normalization size-factor mode
• --pydeseq2-refit-cooks: whether Cook's outlier refitting is enabled
• --pydeseq2-alpha: official target FDR used by DeseqStats
• --pydeseq2-cooks-filter: whether Cook's filtering is applied in result extraction
• --pydeseq2-independent-filter: whether independent filtering is applied in result extraction
• --pydeseq2-n-cpus: CPU count passed to PyDESeq2

Wilcoxon

1. Input: pseudobulk raw counts
2. Transformation: OmicsClaw computes CPM/log-transformed profiles inside the wrapper
3. Testing: scipy.stats.ranksums() is used on per-gene transformed values
4. Outputs: reports a pseudobulk-derived log2fc, Wilcoxon statistic, raw pvalue, and BH-adjusted pvalue_adj

Core tuning flags:

• --reference-condition: defines the sign direction of the contrast
• --cluster-key: chooses the cluster partition for per-cluster pseudobulk
• --min-counts-per-gene: wrapper-level gene filter before DE testing
• --min-samples-per-condition: minimum replicate count before a contrast is attempted
• --wilcoxon-alternative: official SciPy alternative hypothesis

Output Structure

output_directory/
├── README.md
├── report.md
├── result.json
├── processed.h5ad
├── figures/
│   ├── condition_spatial_context.png
│   ├── pseudobulk_volcano.png
│   ├── condition_effect_burden_spatial.png
│   ├── condition_effect_burden_umap.png
│   ├── condition_de_barplot.png
│   ├── cluster_de_burden.png
│   ├── condition_pvalue_distribution.png
│   ├── sample_counts_by_condition.png
│   └── manifest.json
├── figure_data/
│   ├── pseudobulk_de.csv
│   ├── pseudobulk_volcano_points.csv
│   ├── per_cluster_summary.csv
│   ├── skipped_contrasts.csv
│   ├── cluster_de_metrics.csv
│   ├── top_de_genes.csv
│   ├── sample_counts_by_condition.csv
│   ├── condition_run_summary.csv
│   ├── condition_spatial_points.csv
│   ├── condition_umap_points.csv          # when UMAP is available
│   └── manifest.json
├── tables/
│   ├── pseudobulk_de.csv
│   ├── per_cluster_summary.csv
│   └── skipped_contrasts.csv
└── reproducibility/
    ├── analysis_notebook.ipynb
    ├── commands.sh
    ├── requirements.txt
    └── r_visualization.sh

README.md and reproducibility/analysis_notebook.ipynb are generated by the standard OmicsClaw wrapper. Direct script execution usually produces the skill-native outputs plus reproducibility/commands.sh.

The bundled optional R templates live under:

skills/spatial/spatial-condition/r_visualization/
├── README.md
└── condition_publication_template.R

Dependencies

Required:

• scanpy
• scipy

Optional:

• pydeseq2 for proper negative-binomial GLM inference

Optional (R):

• ggplot2

Safety

• Local-first: all processing remains local.
• No pseudoreplication: do not interpret spot-level observations as biological replicates.
• Reference transparency: always state which condition is the reference before interpreting log2 fold changes.
• Replicate caution: fewer than 3 samples per condition is workable but materially less stable.

Integration with Orchestrator

Trigger conditions:

• condition comparison
• pseudobulk
• PyDESeq2 / DESeq2
• treatment vs control

Chaining partners:

• spatial-preprocess for counts preservation and baseline clustering
• spatial-enrichment for pathway analysis on DE genes
• spatial-genes for comparing treatment-responsive genes to SVG programs

Citations

• PyDESeq2 documentation: DeseqDataSet
• PyDESeq2 documentation: DeseqStats
• SciPy documentation: scipy.stats.ranksums
• Squair et al. 2021 — pseudobulk best practices