菜单
Identify tissue regions and spatial niches from preprocessed spatial transcriptomics data using Leiden, Louvain, SpaGCN, STAGATE, GraphST, BANKSY, or CellCharter.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills- --skill spatial-domains命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Operate MedSAM2 for promptable segmentation of 3D medical images and medical videos, including CT lesion propagation, MRI volumes, RECIST-guided prompts, efficient CPU-oriented variants, training, and 3D Slicer integration. Use when generating or validating volumetric masks from sparse prompts or propagating masks through image slices or video frames.
Build reproducible healthcare imaging pipelines with Project MONAI for DICOM, NIfTI, pathology, and multidimensional imaging tasks including preprocessing, augmentation, training, sliding-window inference, evaluation, model bundles, labeling, and deployment. Use when implementing medical image classification, segmentation, registration, detection, generative, or foundation-model workflows in PyTorch.
Operate Google TxGemma prediction and chat models for therapeutic property prediction across small molecules, proteins, nucleic acids, diseases, targets, and cell lines. Use when formatting Therapeutics Data Commons tasks, choosing TxGemma model size or variant, running local or Model Garden inference, fine-tuning on private therapeutic data, or evaluating TxGemma in drug-discovery workflows.
| name | spatial-domains |
| description | Identify tissue regions and spatial niches from preprocessed spatial transcriptomics data using Leiden, Louvain, SpaGCN, STAGATE, GraphST, BANKSY, or CellCharter. |
| version | 0.5.0 |
| author | OmicsClaw |
| license | MIT |
| tags | ["spatial","domains","niche","tissue-region","clustering","leiden","louvain","spagcn","stagate","graphst","banksy","cellcharter"] |
| metadata | {"omicsclaw":{"domain":"spatial","allowed_extra_flags":["--auto-k","--auto-k-min","--auto-k-max","--dim-output","--epochs","--k-nn","--lambda-param","--method","--n-domains","--n-layers","--num-neighbours","--pre-resolution","--rad-cutoff","--refine","--resolution","--spagcn-p","--spatial-weight","--stagate-alpha","--use-rep"],"param_hints":{"leiden":"[Truncated]","louvain":"[Truncated]","spagcn":"[Truncated]","stagate":"[Truncated]","graphst":"[Truncated]","banksy":"[Truncated]","cellcharter":"[Truncated]"},"legacy_aliases":["domains"],"saves_h5ad":true,"requires":{"bins":"[Truncated]","env":"[Truncated]","config":"[Truncated]"},"emoji":"🗺️","homepage":"https://github.com/TianGzlab/OmicsClaw","os":["macos","linux"],"install":["[Truncated]","[Truncated]"],"trigger_keywords":["spatial domain","tissue region","niche","SpaGCN","STAGATE","CellCharter"]}} |
You are Spatial Domains, a specialised OmicsClaw agent for tissue region and spatial niche identification. Your role is to partition spatial transcriptomics tissue sections into biologically meaningful domains using graph-based clustering methods that incorporate both gene expression and spatial coordinates.
figures/figure_data/ contains domain-level plotting contracts for downstream customizationr_visualization/ templates consume figure_data/ without recomputing domains| Format | Extension | Required Fields |
|---|
基于 SOC 职业分类
| Example |
|---|
| AnnData (preprocessed) | .h5ad | X (log-norm), obsm["spatial"], obsm["X_pca"], raw (counts), layers["counts"] | preprocessed.h5ad |
| AnnData (raw, demo mode) | .h5ad | X, obsm["spatial"] | demo_visium.h5ad |
After the standard spatial-preprocess pipeline, the AnnData object holds
multiple representations of the expression data. Each domain identification
method selects the appropriate layer automatically:
adata.layers["counts"] # raw integer counts (preserved before normalization)
adata.raw.X # raw counts (scanpy convention)
adata.X # log-normalized expression (normalize_total + log1p)
adata.obsm["X_pca"] # PCA embeddings (from scaled HVGs)
adata.obsp["connectivities"] # k-NN neighbor graph (from PCA)
adata.obsm["spatial"] # spatial coordinates (x, y)
adata.var["highly_variable"] # HVG annotation
Different methods consume the preprocessed.h5ad data differently. The skill
automatically selects the correct data layer for each method:
| Method | Primary input consumed | What the method actually operates on | Extra requirements |
|---|---|---|---|
| Leiden | obsp["connectivities"] (neighbor graph) | Pre-built k-NN graph from log-normalized + PCA | Spatial coords for spatial-weighted graph |
| Louvain | obsp["connectivities"] (neighbor graph) | Pre-built k-NN graph from log-normalized + PCA | Spatial coords optional |
| SpaGCN | X (log-normalized expression) | Full gene matrix; internal PCA by SpaGCN | Spatial coords required; current wrapper uses coordinates only (histology=False) |
| STAGATE | X (log-normalized, HVG subset) | Auto-filters to var["highly_variable"] before training | Spatial coords for radius-based adjacency |
| GraphST | raw.X or layers["counts"] (raw counts) | Internal log1p -> normalize -> scale -> HVG(3000) | Spatial coords required |
| BANKSY | layers["counts"] or raw.X -> normalize_total | Non-negative library-size normalized expression | Spatial coords required |
| CellCharter | obsm["X_pca"] or X (log-normalized) | Fast neighbor aggregation over embeddings | Spatial coords for Delaunay graph |
Why raw counts for GraphST? GraphST's
preprocess()internally doeslog1p+normalize_total+scale+ HVG selection. Passing already log-normalized data would cause a double log-transform (log(log(x+1)+1)), which distorts the expression distribution. The skill automatically restoresadata.rawwhen available.
Why non-negative expression for BANKSY? BANKSY constructs neighborhood- averaged features by summing neighboring expression values. The official tutorial uses library-size normalization without log or z-score scaling. Z-score scaling (
sc.pp.scale) introduces negative values which distort the neighborhood aggregation. The skill restores raw counts and appliesnormalize_total(withoutlog1p) to produce a non-negative matrix.
figure_data/*.csv and figure_data/manifest.json for downstream customizationreport.md, result.json, processed.h5ad, figures, tables, output guide, and reproducibility bundle including a notebook and optional R helperOmicsClaw treats spatial-domains visualization as a two-layer system:
figure_data/ and does not recompute domain labels.The standard gallery is declared as a recipe instead of hard-coded plotting branches. Current gallery roles include:
overview: spatial and UMAP domain mapsdiagnostic: PCA domain view and domain neighbor-mixing heatmapsupporting: domain size summaryuncertainty: local-purity spatial map and histogramThis keeps the default OmicsClaw story stable while leaving room for later method-specific visualization extensions.
# Standard usage (Leiden, default)
oc run spatial-domains --input <preprocessed.h5ad> --output <report_dir>
# Specify method and parameters
oc run spatial-domains \
--input <preprocessed.h5ad> --method leiden --resolution 0.8 --spatial-weight 0.3 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method louvain --resolution 1.0 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method spagcn --n-domains 7 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method stagate --n-domains 7 --k-nn 6 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method graphst --n-domains 7 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method banksy --lambda-param 0.8 --resolution 1.0 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method cellcharter --n-domains 7 --n-layers 3 --output <dir>
oc run spatial-domains \
--input <preprocessed.h5ad> --method cellcharter --auto-k --auto-k-min 4 --auto-k-max 12 --output <dir>
# Apply spatial refinement
oc run spatial-domains \
--input <preprocessed.h5ad> --method leiden --refine --output <dir>
# Demo mode
oc run spatial-domains --demo --output /tmp/domains_demo
# Note: 'oc run' is an alias for 'python omicsclaw.py run'
python omicsclaw.py run spatial-domains --demo
Every successful standard run also writes reproducibility/analysis_notebook.ipynb
and a top-level README.md to make the output directory easier to inspect.
adata.obsm["spatial"])adata.obsp["connectivities"] (preserves downstream UMAPs).sc.tl.leiden(adjacency=adjacency, resolution=resolution)adata.obs["spatial_domain"]Key parameters:
resolution: Controls granularity (default 1.0; higher = more domains)spatial_weight: Weight of spatial graph (0.0-1.0, default 0.3)adata.obsm["spatial"])adjacency=adjacency.sc.tl.louvain(adjacency=adjacency, resolution=resolution)adata.obs["spatial_domain"]pip install louvainKey parameters:
resolution: Controls granularity (default 1.0)spatial_weight: Weight of spatial graph (0.0-1.0, default 0.3)normalize_per_cell + log1p before training)adata.X (log-normalized full gene matrix) — SpaGCN performs internal PCASpaGCN.train() with n_domains target clustersadata.obs["spatial_domain"]Key parameters:
n_domains: Target number of spatial domainsspagcn_p: Spatial neighborhood contribution (default 0.5)epochs: Training epochs (default 100 in OmicsClaw CLI wrapper)Current OmicsClaw behavior: the wrapper uses
histology=False, so this execution path is coordinate-aware but not image-aware.
adata.var["highly_variable"] exists, only HVGs are used for the autoencoder (reduces VRAM explosion, improves convergence).k_nn=6) to prevent graph fragmentation across vastly different coordinate scales (e.g. Visium vs Stereo-seq).adata.obs["spatial_domain"]Key parameters:
n_domains: Target number of domainsk_nn: Nearest neighbors for spatial topology graph (default 6)rad_cutoff: Optional fixed-radius graph alternative when coordinate scale is knownstagate_alpha: Cell type-aware module weight (default 0.0)pre_resolution: Pre-clustering resolution for the cell type-aware extensionadata.layers["counts"] or adata.raw)log1p -> scale -> HVG).GraphST.preprocess() + GraphST.construct_interaction()torch.cuda.empty_cache() sweeping to prevent VRAM memory locking).reg_covar=1e-3 to prevent ill-conditioned singular matrices) with graceful KMeans fallback.adata.obs["spatial_domain"]Key parameters:
n_domains: Target number of domainsepochs: Training epochs (default 100 in OmicsClaw CLI wrapper)dim_output: Embedding dimension (default 64)refine: Optional GraphST / KNN-style post hoc smoothinglayers["counts"], strictly enforcing normalize_total WITHOUT log1p to adhere to exact non-negative BANKSY augmentation rules.n_domains is requested: Leverages advanced tied-GMM fallback to guarantee exact cluster slicing.n_domains is NOT requested: Heuristic sc.tl.leiden discovery mode via graph density.adata.obs["spatial_domain"]Key parameters:
lambda_param: Spatial regularization (default 0.2)n_domains: Exact cluster quantity target (optional)resolution: Leiden resolution (default 1.0 in the OmicsClaw CLI wrapper, used if n_domains is omitted)num_neighbours: Neighbors for feature construction (default 15)Note: Z-score scaling (
sc.pp.scale) is intentionally not applied, as it introduces negative values that distort BANKSY's neighborhood feature aggregation. See prabhakarlab/Banksy.
adata.obsm["X_pca"]).n_layers (default 3) spatial neighbors.--auto-k): Performs stability analysis across multiple cluster counts to identify the optimal n_domains.adata.obs["spatial_domain"]Key parameters:
n_domains: Target number of domains (default 7 in the current CLI wrapper when --auto-k is false)auto_k: Enable stability analysis for optimal cluster countauto_k_min / auto_k_max: Search range for stability-based K selectionn_layers: Number of spatial neighborhood hops to aggregate (default 3)use_rep: Feature representation to use (defaults to X_pca)Key parameters:
threshold: Disagreement threshold (default 0.5)k: Number of spatial neighbors (default 10)output_dir/
├── README.md
├── report.md
├── result.json
├── processed.h5ad
├── figures/
│ ├── spatial_domains.png
│ ├── umap_domains.png
│ ├── pca_domains.png
│ ├── domain_sizes.png
│ ├── domain_neighbor_mixing.png
│ ├── domain_local_purity_spatial.png
│ ├── domain_local_purity_histogram.png
│ └── manifest.json
├── figure_data/
│ ├── domain_counts.csv
│ ├── domain_spatial_points.csv
│ ├── domain_umap_points.csv
│ ├── domain_method_embedding_points.csv # if method-specific latent space exists
│ ├── domain_neighbor_mixing.csv
│ └── manifest.json
├── tables/
│ ├── domain_summary.csv
│ ├── domain_assignments.csv
│ └── domain_neighbor_mixing.csv
└── reproducibility/
├── analysis_notebook.ipynb
├── commands.sh
└── r_visualization.sh
README.md: Human-friendly output guide summarising the method, parameters, and where to look first.report.md: Narrative analysis report with domain counts, parameter summary, and next-step guidance.result.json: Machine-readable summary and parameter record for downstream automation.processed.h5ad: AnnData with obs["spatial_domain"] and method-specific embeddings when available.figures/: Standard Python gallery with overview, diagnostics, supporting summaries, uncertainty panels, and figures/manifest.json.figure_data/: Figure-ready CSV exports for custom plotting in Python or R.tables/domain_summary.csv: Per-domain counts and proportions for downstream summaries.tables/domain_assignments.csv: Per-observation labels plus local purity when available.reproducibility/analysis_notebook.ipynb: Auto-generated notebook for code inspection, result loading, and reruns inside Jupyter.reproducibility/commands.sh: Minimal shell rerun command with the same core parameters.reproducibility/r_visualization.sh: Entry point for the optional R visualization layer.Additional files such as reproducibility/requirements.txt may appear when
environment capture is enabled, but they are not guaranteed for every run.
The bundled repo-side R helpers live under:
skills/spatial/spatial-domains/r_visualization/
├── README.md
└── domains_publication_template.R
Required:
scanpy >= 1.9 — single-cell/spatial analysissquidpy >= 1.2 — spatial extensionsmatplotlib — plottingnumpy, pandas — numericsOptional:
SpaGCN — spatially-aware graph convolutional clusteringSTAGATE_pyG — graph attention auto-encoder domains (requires PyTorch)GraphST — graph self-supervised contrastive learning (requires PyTorch)banksy — spatial feature augmentationcellcharter — neighborhood-aggregated clustering and Auto-K selectionlouvain — Louvain clustering algorithmadata.obs column, original data preservedTrigger conditions:
Chaining partners:
spatial-preprocess: Provides the preprocessed h5ad inputspatial-de: Downstream differential expression between domainsspatial-enrichment: Gene set enrichment per domainspatial-communication: Cell-cell communication across domain boundaries