メニュー
Cell type annotation for spatial transcriptomics data using Scanpy marker-gene overlap scoring, Tangram mapping, scANVI transfer, or CellAssign probabilistic models.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills- --skill spatial-annotateコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SOC 職業分類に基づく
SKILL.md を表示中
| name | spatial-annotate |
| description | Cell type annotation for spatial transcriptomics data using Scanpy marker-gene overlap scoring, Tangram mapping, scANVI transfer, or CellAssign probabilistic models. |
| version | 0.4.0 |
| author | OmicsClaw |
| license | MIT |
| tags | ["spatial","annotation","cell-type","tangram","scanvi","cellassign","marker-genes","scanpy"] |
| metadata | {"omicsclaw":{"domain":"spatial","allowed_extra_flags":["--batch-key","--cell-type-key","--cellassign-max-epochs","--cluster-key","--layer","--marker-n-genes","--marker-overlap-method","--marker-overlap-normalize","--marker-padj-cutoff","--marker-rank-method","--method","--model","--reference","--scanvi-max-epochs","--scanvi-n-hidden","--scanvi-n-layers","--scanvi-n-latent","--species","--tangram-device","--tangram-num-epochs","--tangram-train-genes"],"param_hints":{"marker_based":"[Truncated]","tangram":"[Truncated]","scanvi":"[Truncated]","cellassign":"[Truncated]"},"legacy_aliases":["annotate"],"saves_h5ad":true,"requires_preprocessed":true,"requires":{"bins":"[Truncated]","env":"[Truncated]","config":"[Truncated]"},"emoji":"🏷️","homepage":"https://github.com/TianGzlab/OmicsClaw","os":["macos","linux"],"install":["[Truncated]"],"trigger_keywords":["cell type annotation","annotate cell types","Tangram","scANVI","CellAssign","marker genes","label transfer","spatial annotation"]}} |
You are Spatial Annotate, the cell type annotation skill for OmicsClaw. Your role is to assign biologically meaningful labels to spatial spots or cells using a fast marker-based baseline, deep-learning reference transfer, or probabilistic count-based models.
rank_genes_groups + marker_gene_overlap, no external reference requiredreport.md, result.json, processed.h5ad, annotation tables, figures/manifest.json, figure_data/, and reproducibility commands| Format | Extension | Required Fields | Example |
|---|---|---|---|
| Spatial AnnData (preprocessed) | .h5ad | X, obsm["spatial"], clusters for marker-based mode, preferably layers["counts"] | preprocessed.h5ad |
| Reference AnnData (Tangram / scANVI) | .h5ad | X, obs["cell_type"], preferably for scANVI |
layers["counts"]reference_sc.h5ad |
| Marker JSON (CellAssign, optional) | .json | {cell_type: [marker genes...]} | tumor_panel.json |
| Demo | n/a | --demo flag | Runs the default marker-based workflow on demo data |
Different annotation methods make different assumptions about the input matrix:
| Method | Input Matrix | Rationale |
|---|---|---|
marker_based | adata.X (log-normalized) | scanpy.tl.rank_genes_groups and marker-overlap scoring compare continuous expression magnitudes across clusters |
tangram | adata.X (log-normalized) for both spatial and reference | Tangram maps cells to space by matching expression profiles on a shared normalized scale |
scanvi | adata.layers["counts"] or another raw-count layer | scVI/scANVI assume count-based likelihoods in setup_anndata(..., layer=...) |
cellassign | adata.layers["counts"] or another raw-count layer | CellAssign models count data and computes size factors from the same raw matrix |
Data layout requirement: Preprocessing should preserve raw counts before normalization:
adata.layers["counts"] = adata.X.copy() # before normalize_total + log1p
adata.X = lognorm_expr # after normalize_total + log1p
If layers["counts"] is missing, count-based methods try adata.raw first and only fall back to adata.X with an explicit warning.
report.md, result.json, processed.h5ad, tables, the standard Python gallery, plot-ready figure_data/, and the reproducibility bundle.# Marker-based baseline (default)
oc run spatial-annotate \
--input <preprocessed.h5ad> --output <dir>
# Marker-based tuning with Scanpy overlap scoring
oc run spatial-annotate \
--input <preprocessed.h5ad> --method marker_based \
--cluster-key leiden --marker-rank-method wilcoxon \
--marker-n-genes 50 --marker-overlap-method overlap_count \
--marker-overlap-normalize reference --output <dir>
# Marker-based p-adjusted marker selection (no fixed top-N)
oc run spatial-annotate \
--input <preprocessed.h5ad> --method marker_based \
--marker-n-genes 0 --marker-padj-cutoff 0.01 --output <dir>
# Tangram transfer
oc run spatial-annotate \
--input <preprocessed.h5ad> --method tangram \
--reference <reference_sc.h5ad> --cell-type-key cell_type \
--tangram-num-epochs 500 --tangram-train-genes 2000 --output <dir>
# scANVI transfer
oc run spatial-annotate \
--input <preprocessed.h5ad> --method scanvi \
--reference <reference_sc.h5ad> --cell-type-key cell_type \
--layer counts --batch-key sample_id \
--scanvi-n-hidden 128 --scanvi-n-latent 10 --scanvi-max-epochs 100 \
--output <dir>
# CellAssign with built-in marker panel
oc run spatial-annotate \
--input <preprocessed.h5ad> --method cellassign \
--species human --layer counts --cellassign-max-epochs 400 --output <dir>
# CellAssign with a custom marker-panel JSON
oc run spatial-annotate \
--input <preprocessed.h5ad> --method cellassign \
--model <marker_panel.json> --layer counts --output <dir>
# Demo mode (marker-based baseline)
oc run spatial-annotate --demo --output /tmp/annotate_demo
# Direct script entrypoint
python skills/spatial/spatial-annotate/spatial_annotate.py \
--input <preprocessed.h5ad> --method marker_based --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.
The current spatial-annotate implementation also writes a standard Python
gallery manifest at figures/manifest.json plus figure_data/manifest.json
for downstream custom visualization layers. A companion R template lives in
skills/spatial/spatial-annotate/r_visualization/.
adata.X (log-normalized expression) plus an existing cluster column such as leidenscanpy.tl.rank_genes_groups(adata, cluster_key, method=..., n_genes=...)scanpy.tl.marker_gene_overlap(...) compares cluster markers against built-in human or mouse marker signaturesCore tuning flags:
| Parameter | Default | Description |
|---|---|---|
--cluster-key | leiden | Cluster column to annotate |
--marker-rank-method | wilcoxon | Passed to scanpy.tl.rank_genes_groups |
--marker-n-genes | 50 | Passed as top_n_markers; use 0 to rely on adjusted p-value filtering instead |
--marker-overlap-method | overlap_count | overlap_count, overlap_coef, or jaccard |
--marker-overlap-normalize | reference | Scanpy normalization mode for overlap_count; none disables normalization |
--marker-padj-cutoff | unset | Passed as adj_pval_threshold when --marker-n-genes 0 |
tg.pp_adatas(..., genes=...) from reference HVGs or all shared genestg.map_cells_to_space(..., mode="cells", num_epochs=..., device=...)tg.project_cell_annotations(...) creates per-spot cell-type predictionsCore tuning flags:
| Parameter | Default | Description |
|---|---|---|
--reference | required | Reference scRNA-seq AnnData |
--cell-type-key | cell_type | Label column in reference.obs |
--tangram-num-epochs | 500 | Passed to map_cells_to_space(..., num_epochs=...) |
--tangram-device | auto | Device string such as cpu, cuda:0, or mps |
--tangram-train-genes | 2000 | OmicsClaw control for the gene list passed into pp_adatas; 0 means all shared genes |
layer plus a labelled reference AnnDataSCVI.setup_anndata(..., layer=..., batch_key=..., labels_key=...)scvi.model.SCVI(..., n_hidden=..., n_latent=..., n_layers=...)SCANVI.from_scvi_model(...), then query adaptation with SCANVI.load_query_data(...)obs["scanvi_confidence"]Core tuning flags:
| Parameter | Default | Description |
|---|---|---|
--reference | required | Reference scRNA-seq AnnData |
--cell-type-key | cell_type | Label column in reference.obs |
--layer | counts | Raw-count layer passed to setup_anndata |
--batch-key | unset | Optional batch column passed to setup_anndata |
--scanvi-n-hidden | 128 | Hidden-layer width in the SCVI backbone |
--scanvi-n-latent | 10 | Latent-dimensionality parameter |
--scanvi-n-layers | 1 | Number of encoder/decoder layers in the SCVI backbone |
--scanvi-max-epochs | 100 | Current wrapper uses this for pretraining, finetuning, and query adaptation |
CellAssign.setup_anndata(..., size_factor_key=..., batch_key=..., layer=...)model.train(max_epochs=...)Core tuning flags:
| Parameter | Default | Description |
|---|---|---|
--model | unset | OmicsClaw JSON marker-panel path; omit to use built-in signatures |
--species | human | Built-in signature set when --model is not supplied |
--layer | counts | Raw-count layer for CellAssign |
--batch-key | unset | Optional batch covariate passed to setup_anndata |
--cellassign-max-epochs | 400 | Passed to model.train(max_epochs=...) |
output_dir/
├── README.md
├── report.md
├── result.json
├── processed.h5ad
├── figures/
│ ├── cell_type_spatial.png
│ ├── cell_type_umap.png
│ └── cell_type_barplot.png
│ ├── marker_overlap_heatmap.png # marker_based
│ ├── annotation_confidence_spatial.png # probabilistic methods
│ ├── annotation_confidence_histogram.png # probabilistic methods
│ ├── annotation_probability_heatmap.png # Tangram / scANVI / CellAssign
│ ├── custom/
│ │ └── annotation_spatial_publication.png # optional R layer output
│ └── manifest.json
├── figure_data/
│ ├── annotation_spatial_points.csv
│ ├── annotation_umap_points.csv
│ ├── annotation_cell_type_counts.csv
│ ├── annotation_probabilities.csv # probabilistic methods
│ ├── marker_overlap_scores.csv # marker_based
│ └── manifest.json
├── tables/
│ ├── annotation_summary.csv
│ ├── cell_type_assignments.csv
│ ├── cluster_annotations.csv # marker_based only
│ └── marker_overlap_scores.csv # marker_based only
└── reproducibility/
├── analysis_notebook.ipynb
├── commands.sh
└── 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 and an optional
reproducibility/r_visualization.sh helper that calls the bundled R template.
Required: scanpy, anndata, numpy, pandas, scipy, matplotlib
Optional:
tangram-sc — Tangram mappingscvi-tools — scANVI and CellAssigntorch — acceleration backend for Tangram / scvi-toolsmarker_based / tangram; count-based methods require raw counts.result.json, and reproducibility/commands.sh.figure_data/ for publication-style refinement without recomputing the science.Trigger conditions:
Chaining partners:
spatial-preprocess — Prepare log-normalized expression, counts layer, clusters, and coordinatesspatial-domains — Compare domain assignments with annotated cell typesspatial-communication — Use labels for ligand-receptor or neighborhood communication analysesspatial-deconv — Use reference-aware deconvolution when spot-level mixtures are expectedrank_genes_groupsmarker_gene_overlap