mims-harvard/tooluniverse-crispr-screen-analysis

ToolUniverse CRISPR Screen Analysis

RULE ZERO — Check for pre-computed results FIRST

Before following any instruction below, scan the data folder for:

• *_executed.ipynb → read with tu run read_executed_notebook '{"data_folder":"<path>","search":"<keyword>"}' and cite its cell outputs as the authoritative answer
• Pre-computed result files (CSV/TSV with names like *results*, *deseq*, *enrich*, *stats*, *_simplified.csv) → read directly and report the requested value
• Canonical analysis scripts (analysis.R, run_*.py, find_*.R, *.Rmd) → execute as-is and read the output

Only follow this skill's re-analysis recipe below if none of the above exist. Re-running from raw data produces different numbers than the published answer and is much slower (often 5-10× turn count).

---

Comprehensive skill for analyzing CRISPR-Cas9 genetic screens to identify essential genes, synthetic lethal interactions, and therapeutic targets through robust statistical analysis and pathway enrichment.

Overview

CRISPR screens enable genome-wide functional genomics by systematically perturbing genes and measuring fitness effects. This skill provides an 8-phase workflow for:

• Processing sgRNA count matrices
• Quality control and normalization
• Gene-level essentiality scoring (MAGeCK-like and BAGEL-like approaches)
• Synthetic lethality detection
• Pathway enrichment analysis
• Drug target prioritization with DepMap integration
• Integration with expression and mutation data

---

Core Workflow

Phase 1: Data Import & sgRNA Count Processing

Load sgRNA count matrix (MAGeCK format or generic TSV). Expected columns: sgRNA, Gene, plus sample columns. Create experimental design table linking samples to conditions (baseline/treatment) with replicate assignments.

Phase 2: Quality Control & Filtering

Assess sgRNA distribution quality:

• Library sizes per sample (total reads)
• Zero-count sgRNAs: Count across samples
• Low-count filtering: Remove sgRNAs below threshold (default: <30 reads in >N-2 samples)
• Gini coefficient: Assess distribution skewness per sample
• Report filtering recommendations

Phase 3: Normalization

Normalize sgRNA counts to account for library size differences:

• Median ratio (DESeq2-like): Calculate geometric mean reference, compute size factors as median of ratios
• Total count (CPM-like): Divide by library size in millions

Calculate log2 fold changes (LFC) between treatment and control conditions with pseudocount.

Phase 4: Gene-Level Scoring

Two scoring approaches:

• MAGeCK-like (RRA): Rank all sgRNAs by LFC, compute mean rank per gene. Lower mean rank = more essential. Includes sgRNA count and mean LFC per gene.
• BAGEL-like (Bayes Factor): Use reference essential/non-essential gene sets to estimate LFC distributions. Calculate likelihood ratio (Bayes Factor) for each gene. Higher BF = more likely essential.

Phase 5: Synthetic Lethality Detection

Compare essentiality scores between wildtype and mutant cell lines:

• Merge gene scores, calculate delta LFC and delta rank
• Filter for genes essential in mutant (LFC < threshold) but not wildtype (LFC > -0.5) with large rank change
• Sort by differential essentiality

Query DepMap/literature for known dependencies using PubMed search.

Phase 6: Pathway Enrichment Analysis

Submit top essential genes to Enrichr for pathway enrichment:

• KEGG pathways
• GO Biological Process
• Retrieve enriched terms with p-values and gene lists

Phase 7: Drug Target Prioritization

Composite scoring combining:

• Essentiality (50% weight): Normalized mean LFC from CRISPR screen
• Expression (30% weight): Log2 fold change from RNA-seq (if available)
• Druggability (20% weight): Number of drug interactions from DGIdb

Query DGIdb for each candidate gene to find existing drugs, interaction types, and sources.

Phase 8: Report Generation

Generate markdown report with:

• Summary statistics (total genes, essential genes, non-essential genes)
• Top 20 essential genes table (rank, gene, mean LFC, sgRNAs, score)
• Pathway enrichment results (top 10 terms per database)
• Drug target candidates (rank, gene, essentiality, expression FC, druggability, priority score)
• Methods section

---

ToolUniverse Tool Integration

Key Tools Used:

• PubMed_search_articles - Literature search for gene essentiality and drug resistance
• ReactomeAnalysis_pathway_enrichment - Pathway enrichment (param: identifiers newline-separated, page_size)
• enrichr_gene_enrichment_analysis - Enrichr enrichment (param: gene_list array, libs array)
• DGIdb_get_drug_gene_interactions - Drug-gene interactions (param: genes as array)
• DGIdb_get_gene_druggability - Druggability categories
• STRING_get_network - Protein interaction networks
• kegg_search_pathway - Pathway search by keyword
• kegg_get_pathway_info - Pathway details by ID

Cancer Context (essential for drug resistance screens):

• civic_search_evidence_items - Clinical evidence for drug resistance/sensitivity
• COSMIC_get_mutations_by_gene - Somatic mutation landscape
• cBioPortal_get_mutations - Mutations in specific cancer cohorts
• ChEMBL_search_targets - Structural druggability assessment

Expression & Variant Integration:

• GEO_search_rnaseq_datasets / geo_search_datasets - Expression datasets
• ClinVar_search_variants - Known pathogenic variants
• gnomad_get_gene_constraints - Gene constraint metrics (pLI, oe_lof)
• UniProt_get_function_by_accession - Protein function for hit validation

---

Quick Start

import pandas as pd
from tooluniverse import ToolUniverse

# 1. Load data
counts, meta = load_sgrna_counts("sgrna_counts.txt")
design = create_design_matrix(['T0_1', 'T0_2', 'T14_1', 'T14_2'],
                               ['baseline', 'baseline', 'treatment', 'treatment'])

# 2. Process
filtered_counts, filtered_mapping = filter_low_count_sgrnas(counts, meta['sgrna_to_gene'])
norm_counts, _ = normalize_counts(filtered_counts)
lfc, _, _ = calculate_lfc(norm_counts, design)

# 3. Score genes
gene_scores = mageck_gene_scoring(lfc, filtered_mapping)

# 4. Enrich pathways
enrichment = enrich_essential_genes(gene_scores, top_n=100)

# 5. Find drug targets
drug_targets = prioritize_drug_targets(gene_scores)

# 6. Generate report
report = generate_crispr_report(gene_scores, enrichment, drug_targets)

---

Domain Reasoning: Hits Are Statistical, Not Biological

Screen hits are statistical findings, not direct readouts of biological relevance. A gene scoring as essential might be essential for cell growth in general (housekeeping) or essential specifically for the phenotype you are screening for (interesting). Always compare your screen hits to public essentiality data — use DepMap pan-cancer dependency scores to filter genes that are broadly essential across all cell lines. A gene essential only in your specific context, but not pan-essential in DepMap, is a better candidate for follow-up than one that scores in every screen.

LOOK UP DON'T GUESS: DepMap dependency scores, known core essential gene sets (Hart et al., Blomen et al.), and DGIdb druggability data for your top hits. Do not assume a hit is context-specific without checking public essentiality databases.

Interpretation Framework

| Evidence Grade | Criteria | Validation Priority | |----------------|----------|---------------------| | A -- Strong hit | MAGeCK RRA p < 0.001, BAGEL BF > 5, >=3 sgRNAs with concordant LFC | Immediate validation (individual KO, growth assay) | | B -- Moderate hit | MAGeCK RRA p < 0.01, BAGEL BF 2-5, >=2 concordant sgRNAs | Secondary validation pool | | C -- Weak/ambiguous | p > 0.01, BF < 2, or discordant sgRNA effects | Deprioritize; check for copy-number bias or seed effects |

Interpreting screen results:

• A gene with mean LFC < -1.0 across replicates and >=3 concordant sgRNAs is a robust essentiality hit; single-sgRNA effects are more likely off-target and should be flagged.
• Essential gene thresholds are context-dependent: core fitness genes (e.g., ribosomal, spliceosomal) should deplete in any screen and serve as positive controls -- their absence from the hit list indicates a QC problem.
• Synthetic lethal hits (depleted in mutant but not wildtype) require delta-LFC > 1.5 and confirmation in an independent cell line before therapeutic target nomination.

Synthesis questions to address in the report:

1. Do the top hits cluster in known pathways (Reactome/KEGG), or are they scattered -- suggesting technical noise?
2. Are known essential genes (Hart et al. reference set) correctly identified, confirming screen quality? The full CEGv2 core-essential (~684) and NEGv1 non-essential (~928) reference sets are bundled in scripts/ — load them with from reference_gene_sets import core_essential, nonessential, recovery_rate. recovery_rate(top_depleted_genes) gives the fraction of core-essential genes recovered (a good genome-wide screen recovers >~0.8).
3. For drug target candidates: does DGIdb show existing compounds, and does DepMap confirm the dependency across multiple cell lines?

---

References

• Li W, et al. (2014) MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens. Genome Biology
• Hart T, et al. (2015) High-Resolution CRISPR Screens Reveal Fitness Genes and Genotype-Specific Cancer Liabilities. Cell
• Meyers RM, et al. (2017) Computational correction of copy number effect improves specificity of CRISPR-Cas9 essentiality screens. Nature Genetics
• Tsherniak A, et al. (2017) Defining a Cancer Dependency Map. Cell (DepMap)

---

Analysis conventions

Replicate Spearman correlation — match the paper's aggregation level

Papers differ on how replicate reproducibility is reported: sgRNA-level CPM vs gene-level summed CPM vs gene-level mean CPM. The expected GT is almost always the sgRNA-level Spearman (noisier, lower ρ), not the gene-level aggregate. If you get ρ ≈ 0.6+ you are probably at gene level; drop to per-sgRNA CPM pairs.

Reactome GSEA ranking column

For GSEA on a MAGeCK output, rank by the neg|lfc or equivalent effect-size column the paper specifies (not p-value). Check the MAGeCK xlsx for a beta / sgRNA_effect / neg|score column and rank descending.

Pathway naming

Reactome pathway names in the .gmt bundle are literal (e.g., "cGMP effects", "Signaling by Hippo"). Answers that should match a Reactome term must reproduce the exact label — do not paraphrase the pathway.

See Also

• ANALYSIS_DETAILS.md - Detailed code snippets for all 8 phases
• USE_CASES.md - Complete use cases (essentiality screen, synthetic lethality, drug target discovery, expression integration) and best practices
• EXAMPLES.md - Example usage and quick reference
• QUICK_START.md - Quick start guide
• FALLBACK_PATCH.md - Fallback patterns for API issues