Crispr Screen Triage

CRISPR Screen Triage

You are CRISPR Screen Triage, a specialised ClawBio agent for ranking gene-level CRISPR screen hits from supplied guide counts and annotations.

Trigger

Fire this skill when the user says any of:

• "triage CRISPR screen hits"
• "rank guide-level CRISPR counts"
• "rank depleted CRISPR genes"
• "score genes from a knockout screen"
• "which CRISPR hits should I follow up"

Do NOT fire when:

• The user asks for variant interpretation.
• The user asks for single-cell clustering.
• The user asks for clinical actionability.

Why This Exists

• Without it: Users sort fold changes manually and ignore follow-up feasibility.
• With it: Depletion, essentiality, and druggability are combined deterministically.
• Why ClawBio: The score is transparent, local, and reproducible.

Core Capabilities

1. Count validation: Requires guide ID, gene, control count, treatment count, essentiality, and druggability; rejects non-finite values and negative counts.
2. Guide aggregation: Computes guide-level log2 fold change and aggregates by gene using the median.
3. Local triage: Computes a fixed gene triage score from depletion plus user-supplied essentiality and druggability.
4. Report pack: Writes report, JSON, gene/guide CSVs, and reproducibility command.

Scope

One skill, one task. This skill ranks gene hits from guide-level screen counts and does not design guides, perform statistical screen calling, fetch external annotations, or claim therapy suitability. The essentiality and druggability columns must already be present in the input table. They are not fetched from DepMap, Open Targets, ChEMBL, or any other service.

Input Formats

essentiality and druggability are user-supplied downstream annotations. This skill only averages and weights them after guide-level depletion is calculated.

Workflow

1. Validate: Confirm required columns; reject non-numeric, non-finite (nan, inf), or negative counts with exit 2.
2. Compute: Calculate guide-level log2((treatment + 1) / (control + 1)).
3. Aggregate: Collapse guides to genes using median log2 fold change and mean annotations.
4. Triage: Score depletion, druggability, and essentiality with fixed weights.
5. Report: Write ranked markdown, JSON, gene table, guide table, and command trace.

CLI Reference

python skills/crispr-screen-triage/crispr_screen_triage.py --input counts.csv --output /tmp/crispr
python skills/crispr-screen-triage/crispr_screen_triage.py --demo --output /tmp/crispr
python clawbio.py run crispr-triage --demo

Demo

python clawbio.py run crispr-triage --demo

Expected output: a synthetic twelve-guide, six-gene ranked report with BRCA1 as the top hit.

Algorithm / Methodology

1. Guide depletion: Convert each treatment/control guide count pair to log2 fold change.
2. Gene aggregation: Use median guide log2FC per gene so one noisy guide cannot dominate.
3. Score: 0.55 * max(0, -median_log2FC) + 0.25 * druggability + 0.20 * essentiality.
4. Priority: High requires score >= 1.35 and median log2FC <= -1.0.
5. Non-goal: This is not a canonical statistical screen caller. It does not model negative-binomial counts, copy number, or Bayesian essentiality.

Example Queries

• "Rank these CRISPR hits"
• "Triage depleted genes from this screen"
• "Which knockout hits are most follow-up ready?"

Example Output

# CRISPR Screen Triage Report

| Rank | Gene | Guides | Median log2FC | Priority |
|---:|---|---:|---:|---|
| 1 | BRCA1 | 2 | -2.66 | high |

Output Structure

output_directory/
├── report.md
├── result.json
├── tables/
│   ├── triaged_genes.csv
│   └── guide_metrics.csv
└── reproducibility/
    └── commands.sh

Dependencies

• Python 3.10+ standard library only.

Gotchas

• Do not treat the priority score as validation: It is a triage score only.
• Do not call external databases: Demo and tests must remain deterministic.
• Do not mix guide-level and gene-level semantics: Input uses guide-level counts plus user-supplied gene annotations.

Safety

• Local-first: No external APIs or uploads.
• Disclaimer: Every report includes the ClawBio medical disclaimer.
• Audit trail: Commands are written to reproducibility/commands.sh.

Agent Boundary

The agent dispatches and explains. The Python skill scores and writes outputs.

Integration with Bio Orchestrator

Trigger conditions: CRISPR screen, depleted genes, knockout hit ranking.

Chaining Partners

• target-validation-scorer: downstream target evidence synthesis.
• omics-target-evidence-mapper: cross-omics support for top hits.

Maintenance

• Review cadence: Re-evaluate weights quarterly.
• Staleness signals: Repo adds guide-level statistical screen-calling support.
• Deprecation: Archive if replaced by a full screen-analysis workflow.

Author & Attribution

Prepared by Mrinal Joshi, Imperial College London and UK Dementia Research Institute, using his functional-genomics and bioinformatics background to scope a local deterministic CRISPR hit triage helper. The implementation is intentionally a transparent downstream ranker over supplied counts and annotations, not a canonical screen-scoring method.

Citations

• ClawBio local deterministic triage rules in crispr_screen_triage.py; this skill does not claim a method-paper implementation.