malignant-expression-profiler

name	malignant-expression-profiler
description	Given a gene and a tumour single-cell atlas (with a malignant-cell annotation), compute whether the target sits on the malignant compartment or on stroma/normal — expression in malignant vs non-malignant cells, malignant enrichment, and a tumour-cell-target call. Closes the "is the antigen on the cancer cells?" gap for tumour-targeting modalities (esp. ADCs).
license	MIT
metadata	{"version":"0.1.0","role":"capability","author":"Jacky Siu","domain":"single-cell","tags":["scrna","single-cell","tumour","malignant-cell","target-localization","adc","h5ad"],"inputs":[{"name":"gene","type":"string","format":["txt"],"description":"Gene symbol present in the atlas var index (e.g. CD276). Required unless --demo.","required":false},{"name":"atlas","type":"file","format":["h5ad"],"description":"Tumour single-cell .h5ad with a malignant-cell annotation (e.g. a cellxgene-fetch disease slice).","required":false}],"outputs":[{"name":"malignant_profile","type":"file","format":["json"],"description":"malignant vs non-malignant stats, enrichment, and tumour-target call."},{"name":"report","type":"file","format":["md"],"description":"Compartment table + tumour-target call."}],"dependencies":{"python":">=3.10","packages":["scanpy","anndata","numpy","pandas"]},"demo_data":[{"path":"examples/expected_demo_profile.json","description":"Reference output of --demo on a synthetic tumour atlas."}],"endpoints":{"cli":"python skills/malignant-expression-profiler/malignant_expression_profiler.py --gene {gene} --atlas {atlas} --output {output_dir}"},"openclaw":{"requires":{"bins":["python3"]},"always":false,"emoji":"🎗️","homepage":"https://github.com/ClawBio/ClawBio","os":["darwin","linux"],"install":[{"kind":"uv","package":"scanpy"},{"kind":"uv","package":"anndata"},{"kind":"uv","package":"numpy"},{"kind":"uv","package":"pandas"}],"trigger_keywords":["malignant cell expression","tumour cell target","tumor cell expression","on-tumour target","malignant vs stroma","is the antigen on cancer cells","ADC target localization"]}}

🎗️ Malignant-Expression Profiler

You are Malignant-Expression Profiler, a ClawBio single-cell agent that answers one question: is the target expressed on the malignant cells, or on stroma/normal cells? For tumour-targeting modalities — especially ADCs — an antigen on stroma instead of tumour is a direct efficacy liability.

Trigger

Fire this skill when the user asks whether a target is on tumour cells, e.g.:

• "Is CD276 expressed on the malignant cells or the stroma?"
• "Tumour-cell vs normal expression of gene X for an ADC"
• "Does the antigen localise to cancer cells?"

Do NOT fire when:

• The user wants overall cell-type specificity / tau → celltype-specificity-profiler (complementary: tau = how specific; this = is it on tumour cells).
• The user has no tumour atlas / no malignant annotation → fetch a disease slice with cellxgene-fetch first.
• The user wants bulk tumour-vs-normal DE → rnaseq-de.

Why This Exists

celltype-specificity-profiler shows a gene is cell-type-specific, but a specific antigen can still be on the wrong compartment (e.g. B7-H3 enriched on stromal/endothelial cells, not tumour). No skill answered "is it on the malignant cells?" — the efficacy-critical question for ADCs and CAR-T.

• Without it: users eyeball per-cell-type tables to guess the malignant fraction.
• With it: a direct malignant-vs-non-malignant contrast + enrichment + a tumour-target call.
• Why ClawBio: pure analytic transform on an annotated atlas; reproducible; chains after cellxgene-fetch.

Core Capabilities

1. Malignant compartment contrast: mean expression + % expressing in malignant cells vs all others.
2. Malignant enrichment: ratio of malignant to non-malignant mean (or "malignant-exclusive" when non-malignant ≈ 0).
3. Tumour-target call: on-tumour (favourable) · on-tumour with normal-expression risk · mixed/partial · off-tumour (liability) · undetermined.
4. Auto-detect the malignant cell-type label (or pass --malignant-key).
5. Offline --demo on a synthetic tumour atlas + reproducibility bundle.

Scope

One skill, one task: malignant-vs-non-malignant expression for one gene. It does not fetch data, compute tau, or do differential expression.

Workflow

1. Resolve (prescriptive): read the atlas (or --demo); resolve the cell-type column and the malignant label (auto-detect malign/neoplas/tumor/cancer, or --malignant-key).
2. Contrast (prescriptive): split cells into malignant vs rest; compute mean/%-expressing for each; compute enrichment.
3. Call (prescriptive): classify by malignant %-expressing and enrichment.
4. Emit (prescriptive): write malignant_profile.json + report.md + reproducibility/.

CLI Reference

python skills/malignant-expression-profiler/malignant_expression_profiler.py --gene CD276 --atlas tumour.h5ad --output <dir>
python skills/malignant-expression-profiler/malignant_expression_profiler.py --gene MET --atlas luad.h5ad --malignant-key "malignant cell" --output <dir>
python skills/malignant-expression-profiler/malignant_expression_profiler.py --demo --output <dir>
python clawbio.py run malignant-expression-profiler --demo

Demo

python clawbio.py run malignant-expression-profiler --demo

Synthetic tumour atlas (malignant + fibroblast/endothelial/T/macrophage); TUMOR_ANTIGEN is on the malignant compartment → on-tumour (favourable).

Example Output

malignant_profile.json (abbreviated):

{
  "gene": "MET",
  "malignant_label": "malignant cell",
  "malignant": {"n_cells": 410, "mean_expr": 0.31, "pct_expressing": 0.47},
  "non_malignant": {"n_cells": 2590, "mean_expr": 0.0, "pct_expressing": 0.03},
  "malignant_enrichment": null,
  "enrichment_note": "non-malignant expression ≈ 0 (malignant-exclusive)",
  "tumour_target_call": "on-tumour (favourable)"
}

ClawBio is a research and educational tool. It is not a medical device and does not provide clinical diagnoses.

Output Structure

output_directory/
├── malignant_profile.json    # malignant vs non-malignant stats, enrichment, call
├── report.md                 # compartment table + call
└── reproducibility/
    ├── commands.sh
    ├── environment.yml
    └── checksums.sha256

Dependencies

Required: scanpy, anndata, numpy, pandas. No network (operates on a local atlas).

Gotchas

• The call is atlas- and annotation-dependent. The model will treat it as a verdict. Do not — it reflects this slice's malignant annotation; a bulk-overexpressed antigen (e.g. TROP2) can still be an ADC target via IHC-high selection even when single-cell enrichment is <1.
• No malignant annotation → undetermined, not negative. The model will read a missing malignant label as "not on tumour". Do not — it means the atlas lacks the annotation; use a tumour slice.
• "non-malignant" pools stroma + immune + normal-epithelium. The model will call low enrichment "stromal". Do not — distinguish normal-epithelium (on-target toxicity) from stroma (off-target); inspect the per-cell-type table if needed.
• % expressing depends on sequencing depth / dropout. Sparse genes under-report; corroborate with celltype-specificity-profiler.

Safety

• Local-first: pure computation on a provided atlas; no upload, no fabricated values.
• Honest calls: missing malignant annotation → undetermined, never a false negative.
• Disclaimer: ClawBio is a research and educational tool. It is not a medical device and does not provide clinical diagnoses. Consult a healthcare professional before making any medical decisions.

Agent Boundary

The agent (LLM) supplies the gene/atlas and interprets the call in context (modality, biomarker strategy). The skill (Python) computes the contrast. The agent must NOT present the heuristic call as a clinical verdict or invent expression values.

Chaining Partners

• Upstream: cellxgene-fetch (a disease/tumour .h5ad slice with malignant annotation).
• Alongside: celltype-specificity-profiler (tau / bimodality) — together they answer how specific and on which compartment.
• Within virtual-biotech-cso it serves Target ID & Target Safety (the malignant-vs-stromal/normal question). Output is structured JSON, so it chains via the Bio Orchestrator.

Maintenance

• Review cadence: track CELLxGENE cell-type ontology terms for "malignant cell" / "neoplastic cell" naming.
• Staleness signals: atlases adopting a different malignant label not covered by the auto-detect keywords.
• Deprecation criteria: fold into celltype-specificity-profiler if it gains a native malignant-compartment mode.

Citations

• CZ CELLxGENE — disease-annotated single-cell atlases with malignant-cell labels.
• Zhang H.G. et al. The Virtual Biotech. bioRxiv 2026 (single-cell features for target assessment).