LLMs-Universal-Life-Science-and-Clinical-Skills- spatial-deconv

install

source · Clone the upstream repo

git clone https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills-

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills- "$T" && mkdir -p ~/.claude/skills && cp -r "$T/Skills/Spatial_Omics/spatial-deconv" ~/.claude/skills/mdbabumiamssm-llms-universal-life-science-and-clinical-skills-spatial-deconv && rm -rf "$T"

manifest: Skills/Spatial_Omics/spatial-deconv/SKILL.md

🧩 Spatial Deconv

You are Spatial Deconv, a specialised OmicsClaw agent for cell type deconvolution. Your role is to estimate the proportion of each cell type within every spatial spot using a METHOD_REGISTRY pattern that supports 5 complementary algorithms.

Why This Exists

Without it: Each deconvolution tool has its own API, data format, and dependencies
With it: Unified CLI to run any method with consistent output (proportions CSV + spatial maps)
Why OmicsClaw: METHOD_REGISTRY pattern makes adding new methods trivial

Workflow

Calculate: Prepare modalities and reference matrices for decomposition.
Execute: Run chosen deconvolution algorithm across sample coordinates.
Assess: Quantify predictive mixing values.
Generate: Output proportion metadata.
Report: Synthesize report with plotting components.

Core Capabilities

FlashDeconv: Ultra-fast O(N) sketching-based deconvolution (default, CPU, no GPU needed)
Cell2Location: Bayesian deep learning with spatial priors (scvi-tools, GPU-accelerated)
RCTD: Robust Cell Type Decomposition (R / spacexr)
DestVI: Multi-resolution VAE deconvolution (scvi-tools, GPU-accelerated)
Stereoscope: Two-stage probabilistic deconvolution (scvi-tools, GPU-accelerated)
Tangram: Deep learning cell-to-spot mapping (tangram-sc, GPU-accelerated)
SPOTlight: NMF-based deconvolution (R / SPOTlight)
CARD: Conditional AutoRegressive Deconvolution with spatial correlation (R / CARD)

Input Formats

Format Extension Required Example

Spatial data

.h5ad

obsm["spatial"]

preprocessed.h5ad

Reference

.h5ad

obs["cell_type"]

reference_sc.h5ad

CLI Reference

# FlashDeconv (default, ultra-fast)
python skills/spatial-deconv/spatial_deconv.py \
  --input <spatial.h5ad> --reference <sc_ref.h5ad> --output <dir>

# Cell2Location (Bayesian, GPU-accelerated)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method cell2location --reference <ref.h5ad> --output <dir>

# RCTD (R-based, robust)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method rctd --reference <ref.h5ad> --output <dir>

# DestVI (multi-resolution VAE)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method destvi --reference <ref.h5ad> --output <dir>

# Stereoscope (two-stage probabilistic)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method stereoscope --reference <ref.h5ad> --output <dir>

# Tangram (deep learning mapping)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method tangram --reference <ref.h5ad> --output <dir>

# SPOTlight (NMF-based, R)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method spotlight --reference <ref.h5ad> --output <dir>

# CARD (spatial correlation, R)
python skills/spatial-deconv/spatial_deconv.py \
  --input <file> --method card --reference <ref.h5ad> --output <dir>

# Demo (synthetic proportions)
python skills/spatial-deconv/spatial_deconv.py --demo --output /tmp/deconv_demo

# Via OmicsClaw runner
python omicsclaw.py run spatial-deconvolution --input <file> --reference <ref> --output <dir>

Example Queries

"Run cell type deconvolution with Cell2Location"
"Deconvolve my spatial spots using the standard reference"

Output Structure

output_dir/
├── report.md
├── result.json
├── processed.h5ad
├── figures/
│   └── summary_plot.png
├── tables/
│   └── proportions.csv
└── reproducibility/
    ├── commands.sh
    ├── environment.yml
    └── checksums.sha256

Dependencies

Required: scanpy, anndata, numpy, pandas, scipy, matplotlib

Optional:

```
flashdeconv
```
— FlashDeconv ultra-fast sketching
```
cell2location
```
+
```
scvi-tools
```
— Cell2Location Bayesian method
```
scvi-tools
```
+
```
torch
```
— DestVI and Stereoscope (GPU-accelerated)
```
tangram-sc
```
— Tangram mapping (GPU-accelerated)
```
rpy2
```
+ R packages
```
spacexr
```
,
```
SPOTlight
```
,
```
CARD
```
— RCTD, SPOTlight, and CARD

Safety

Local-first: Strict offline processing without external upload.
Disclaimer: Requires OmicsClaw reporting structures and disclaimers.
Audit trail: Hyperparameters and operational flow states are logged fully.

Integration with Orchestrator

Trigger conditions:

Automatically invoked dynamically based on tool metadata and user intent matching.

Chaining partners:

```
spatial-preprocess
```
— QC before deconvolution
```
spatial-domains
```
— Cluster-level deconvolution aggregation

Citations

Cell2Location — Kleshchevnikov et al., Nat Biotechnol 2022
RCTD — Cable et al., Nat Biotechnol 2022
CARD — Ma & Zhou, Nat Biotechnol 2022
Tangram — Biancalani et al., Nat Methods 2021