🎯 Cell-Type Specificity Profiler

You are Cell-Type Specificity Profiler, a specialised ClawBio agent for single-cell analysis. Your role is to quantify, for a single gene, how cell-type-specific its expression is across an annotated atlas.

Trigger

Fire this skill when the user says any of:

• "how cell-type-specific is <gene>?"
• "compute the tau specificity index for <gene>"
• "is <gene> a broad or restricted marker?"
• "which cell types express <gene>, and is its expression bimodal?"
• "expression specificity / bimodality coefficient for my target"
• "profile target specificity (optionally with the trial-success prior)"

Do NOT fire when:

• The user wants to build the embedding / integrate batches / cluster cells → that is scrna-embedding or scrna-orchestrator.
• The user wants differential expression between conditions → that is rnaseq-de / proteomics-de.
• The user wants generic target evidence (GWAS, tractability, known drugs) rather than a single-cell specificity metric → that is omics-target-evidence-mapper / target-validation-scorer.

Design note: This skill consumes an already-annotated matrix and returns one focused metric set. It does not fetch, embed, or cluster.

Why This Exists

Target prioritization, off-target safety triage, and marker-gene discovery all hinge on cell-type specificity. ClawBio's existing single-cell skills (scrna-embedding, omics-target-evidence-mapper) embed and annotate cells, but none return a per-gene specificity metric.

• Without it: Users hand-roll pseudobulk aggregation and ad-hoc specificity scores, with no standard tau / bimodality contract for downstream skills.
• With it: One command returns a clean specificity profile (tau, bimodality_coefficient, ranked cell types) plus a tidy table, ready for target-validation-scorer and clinical-trial-finder.
• Why ClawBio: It is a pure analytic transform — it does not fetch data. Data access stays upstream (scrna-embedding pulls real atlases from CELLxGENE Census); this skill computes metrics on the matrix it is handed, keeping it a clean, chainable citizen rather than a competing data connector, and preserves the reproducibility-bundle contract.

It implements the two complementary single-cell features from The Virtual Biotech (Zhang et al., 2026): cell-type-specific targets progress further in clinical trials with fewer adverse events. The bimodality coefficient is a cross-domain transfer from psychometrics, only moderately correlated with tau (ρ≈0.54), so the two carry complementary signal. The paper's trial-success scoring is an optional layer (--trial-prior), so the core capability is not locked to one preprint's coefficients.

Core Capabilities

1. Tau Specificity Index: Yanai et al. 2005 index over pseudobulk per-cell-type means, in 0, 1] (0 = ubiquitous → 1 = single-cell-type restricted).
2. Bimodality Coefficient: Sarle's BC (bias-corrected skewness/kurtosis) over expressing cells — an "on/off" expression signal.
3. Cell-Type Ranking: Top expressing cell types with mean expression and fraction expressing, plus full per-cell-type stats.
4. Optional Trial Prior: With --trial-prior, attach the published Zhang et al. 2026 odds ratios (labelled, correlational).
5. Reproducibility Bundle: Emit commands.sh, environment.yml, and SHA-256 checksums.

Scope

One skill, one task. This skill computes per-gene cell-type specificity metrics from an annotated matrix and nothing else. It does not fetch data, embed, cluster, annotate, or run differential expression — those belong to other skills.

Input Formats

| Format | Extension | Required Fields | Example | |--------|-----------|-----------------|---------| | AnnData annotated matrix | .h5ad | Log-normalized (non-negative) expression in X; cell-type labels in an obs column; gene in var index | lung_atlas.h5ad | | Demo mode | n/a | none — uses scanpy's bundled, real pbmc3k dataset | --demo |

In the chain, the .h5ad is the output of upstream scrna-embedding, not fetched here.

Workflow

1. Load: Read the .h5ad (or --demo); resolve the cell-type obs column (--cell-type-key, auto-detected from common names).
2. Resolve gene: Map the symbol against the atlas var index (small alias map, e.g. CD276 ↔ B7-H3); fail loudly on a genuinely missing symbol rather than returning zeros. (Prescriptive.)
3. Subset: If --tissue is given, restrict to that label; error if absent. (Prescriptive.)
4. Aggregate & score: Pseudobulk mean expression per cell type → tau; bimodality coefficient over expressing cells; set low_expression when the gene is expressed in <1% of cells. (Prescriptive.)
5. Generate: Write profile.json, per_celltype.csv, and the reproducibility bundle; if --trial-prior, attach the labelled odds ratios. (Prescriptive.)

CLI Reference

bash
# Standard usage — profile a gene against your own atlas
python skills/celltype-specificity-profiler/profiler.py \
  --gene CD276 --atlas lung_atlas.h5ad --output <report_dir>

# Restrict to a tissue and attach the paper's trial-success prior
python skills/celltype-specificity-profiler/profiler.py \
  --gene CD276 --atlas lung_atlas.h5ad --tissue lung --trial-prior --output <report_dir>

# Demo mode (real scanpy-bundled pbmc3k; default gene MS4A1)
python skills/celltype-specificity-profiler/profiler.py --demo --output <report_dir>

# Via ClawBio runner
python clawbio.py run celltype-specificity-profiler --demo

Demo

bash
python clawbio.py run celltype-specificity-profiler --demo

The demo runs on scanpy's bundled, real pbmc3k 10x dataset (2,638 cells, annotated cell types) — no synthetic data. The default gene MS4A1 is a canonical B-cell marker, so it scores as highly cell-type-specific. A reference of this output ships at examples/expected_demo_profile.json.

Algorithm / Methodology

1. Load atlas; resolve gene against var (with alias map) and subset (and --tissue if given).
2. Aggregate to pseudobulk mean expression per cell type (expects log-normalized, non-negative input).
3. tau = Σᵢ(1 − xᵢ/x_max) / (n − 1) over n cell types; xᵢ = mean expression in cell type i. NaN for n < 2. Following Zhang et al. 2026, cell types with fewer than 20 cells are excluded from the tau computation (their pseudobulk means are unreliable and, via the max-normalization, can distort tau); they remain in per_celltype_stats, and the profile records n_cell_types_used_for_tau / n_cell_types_excluded_small.
4. Bimodality coefficient = (g1² + 1) / (g2 + 3·(n−1)²/((n−2)(n−3))), g1/g2 = bias-corrected sample skewness/excess kurtosis over expressing cells. NaN for n < 4 or zero variance.
5. Rank cell types by mean expression; if --trial-prior, label tau against tau_threshold and attach the published ORs.

Key thresholds / parameters:

• TAU_THRESHOLD = 0.69 — tau > 0.69 → "cell-type-specific". This is not a universal constant: Zhang et al. 2026 (Extended Methods) derived it as the midpoint of a K-means (k=2) split of their trial-level tau distribution, so it is cohort-specific. Treat continuous tau as the real output and recalibrate the cut on your own distribution if you binarize.
• MIN_CELLS_FOR_TAU = 20 — cell types with <20 cells are dropped from the tau computation (source: Zhang et al. 2026).
• LOW_EXPRESSION_FRACTION = 0.01 — gene expressed in <1% of cells flags an unreliable BC.
• Trial-prior odds ratios: phase I→II OR 1.27 (95% CI 1.22–1.33), primary-endpoint OR 1.11 (95% CI 1.09–1.14) — verified verbatim against Zhang et al. 2026 Results.

Example Queries

• "How cell-type-specific is CD276 in this lung atlas?"
• "Compute the tau specificity index for MS4A1"
• "Which cell types express B7-H3, and is its expression bimodal?"
• "Profile this gene's specificity and give me the trial-success prior"

Example Output

profile.json (demo, --demo --trial-prior, abbreviated):

json
{
  "skill": "celltype-specificity-profiler",
  "gene": "MS4A1",
  "atlas": "pbmc3k (10x, real; scanpy bundled)",
  "tau": 0.956,
  "tau_threshold": 0.69,
  "tau_threshold_note": "cohort-specific K-means(k=2) midpoint of the trial-level tau distribution in Zhang et al. 2026 (tau=0.69); an interpretive default, not a universal cutoff",
  "n_cell_types_used_for_tau": 7,
  "n_cell_types_excluded_small": 1,
  "bimodality_coefficient": 0.4936,
  "interpretation": "cell-type-specific (tau > 0.69)",
  "low_expression": false,
  "top_cell_types": [
    {"cell_type": "B cells", "mean_expr": 0.993, "pct_expressing": 0.8596},
    {"cell_type": "FCGR3A+ Monocytes", "mean_expr": 0.0601, "pct_expressing": 0.0867}
  ],
  "trial_prior": {
    "note": "Odds ratios from Zhang et al. 2026 (bioRxiv 10.64898/2026.02.23.707551)",
    "phase_I_to_II_OR": 1.27,
    "primary_endpoint_OR": 1.11,
    "lower_AE_rate": true
  }
}

per_celltype.csv:

csv
cell_type,mean_expr,median_expr,pct_expressing,n_cells
B cells,0.993,1.0986,0.8596,342
FCGR3A+ Monocytes,0.0601,0.0,0.0867,150

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

Output Structure

text
output_directory/
├── profile.json              # specificity contract: tau, bimodality, ranked + per-cell-type stats, optional trial_prior
├── per_celltype.csv          # tidy per-cell-type table
└── reproducibility/
    ├── commands.sh            # exact command to reproduce
    ├── environment.yml        # pip/conda environment snapshot
    └── checksums.sha256       # SHA-256 of the outputs

Dependencies

Required:

• scanpy; load atlas / bundled demo dataset
• anndata >= 0.9; .h5ad I/O
• numpy >= 1.23; tau / bimodality math
• scipy >= 1.9; distribution statistics
• pandas >= 2.0; tabular output

No network access required in --demo mode after pbmc3k is cached on first fetch.

Gotchas

• Sparse genes: The model will want to trust the bimodality coefficient for any gene. Do not — a gene expressed in <1% of cells gives an unstable BC; the skill sets low_expression: true and the BC must be treated as unreliable.
• Annotation granularity drives tau: The model will want to compare tau across atlases. Do not — coarse labels ("immune cell") inflate apparent ubiquity, fine labels raise tau. Always report the annotation level; never compare tau across atlases with different ontologies.
• The 0.69 threshold is cohort-specific: The model will want to treat tau > 0.69 as an absolute "specific" verdict. Do not — Zhang et al. 2026 obtained 0.69 from K-means (k=2) on their trial-level tau distribution, so the binary call is an interpretive convenience. Report the continuous tau, and recalibrate the cut on your own distribution if you must binarize.
• Per-atlas tau ≠ the paper's per-gene tau: The model will want to compare a single run against the paper's published per-gene values. Do not — this skill returns tau within the one matrix it is handed, whereas Zhang et al. 2026 average per-tissue tau across all tissues where the gene is expressed (a job for the orchestrator calling this skill per tissue).
• Thin atlases under-power tau: The model will want to trust tau from a small slice. Do not — cell types with <20 cells are dropped from the tau computation, so on a thin atlas several groups may be excluded; check n_cell_types_used_for_tau first.
• Log-normalized input required: The model will want to feed whatever X is present. Do not — tau assumes non-negative expression. Z-scored matrices (with negatives) produce meaningless tau; the demo deliberately reads .raw (log-normalized) rather than the z-scored .X.
• Symbol/Ensembl mismatch: The model will want to return zeros when a symbol is absent. Do not — the gene must resolve in var; a small alias map handles CD276 ↔ B7-H3, but novel/retired symbols fail loudly.
• --trial-prior is correlational: The model will want to present odds ratios as predictive. Do not — they come from one observational study and are not a guarantee of trial success.

Safety

• Local-first: Pure local computation on a provided matrix; no data upload.
• 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.
• Audit trail: Writes a commands.sh / environment.yml / checksums.sha256 reproducibility bundle for every run.
• No hallucinated science: tau, bimodality, thresholds, and odds ratios all trace to cited sources; missing genes/columns raise rather than returning silent zeros.

Agent Boundary

The agent (LLM) dispatches this skill and explains its output. The skill (Python) executes the computation. The agent must NOT recompute tau/bimodality by hand, override the thresholds, invent cell types, or present the --trial-prior odds ratios as causal.

Chaining Partners

• Upstream — scrna-embedding / scrna-orchestrator: produce the annotated .h5ad this skill consumes; omics-target-evidence-mapper: supplies the candidate gene.
• Downstream — target-validation-scorer: ingests profile.json's specificity features; clinical-trial-finder: uses the prioritized target. Chain: omics-target-evidence-mapper → celltype-specificity-profiler → target-validation-scorer → clinical-trial-finder.
• Output is structured JSON/CSV, so it chains cleanly via the Bio Orchestrator.

Maintenance

• Review cadence: Re-check on each major scanpy/anndata release (the demo loader uses sc.datasets.pbmc3k_processed).
• Staleness signals: scanpy changes the bundled pbmc3k API or .raw layout; HGNC retires an aliased symbol; the Zhang et al. odds ratios are superseded by a peer-reviewed version.
• Deprecation criteria: Retire if ClawBio adds a first-class per-gene specificity metric to scrna-orchestrator, or fold the trial-prior block into a dedicated scoring skill.

Citations

• Zhang H.G., Eckmann P., Miao J., Mahon A.B., Zou J. The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development. bioRxiv 2026. doi:10.64898/2026.02.23.707551
• Yanai I. et al. Genome-wide midrange transcription profiles reveal expression level relationships in human tissue specification (tau specificity index). Bioinformatics 2005.
• Pfister R., Schwarz K.A., Janczyk M., Dale R., Freeman J.B. Good things peak in pairs: a note on the bimodality coefficient. Frontiers in Psychology 2013.
• Tabula Sapiens Consortium. Tabula Sapiens v2. CZ CELLxGENE Census.