Bio Crispr Screens Drugz Chemogenomic

Version Compatibility

Reference examples tested with: drugZ Aug-2019+ (hart-lab/drugz; Python 3.6+), MAGeCK 0.5.9+, pandas 2.2+, numpy 1.26+, scipy 1.12+, statsmodels 0.14+, matplotlib 3.8+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: python drugz.py --help (the repo has no setup.py, so there is no drugz console script)
• GitHub: install via git clone https://github.com/hart-lab/drugz

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

drugZ Chemogenomic Analysis

"Identify genes that sensitize or confer resistance to my drug in a CRISPR screen" -> Compare drug-treated vs vehicle-treated arms (NOT Day-0 baseline) using bidirectional Z-scores per sgRNA, sum to per-gene normalized Z, and rank genes for sensitizer (synthetic lethal) vs suppressor (resistance) phenotype.

• CLI: python drugz.py -i counts.txt -o drugz.txt -c Vehicle_r1,Vehicle_r2 -x Drug_r1,Drug_r2
• Python: programmatic via drugz.drugZ_analysis(args) (takes an argparse Namespace)
• Workflow: vehicle-anchored counts -> Z-scoring -> per-gene summation -> direction-specific FDR

Why drugZ for Drug Screens (not MAGeCK)

Why MAGeCK is suboptimal for drug screens: MAGeCK's RRA was designed for two-condition essentiality; drug-vs-vehicle screens often have small effect sizes (10-30% sgRNA shift) that RRA rank-based aggregation under-detects. drugZ uses parametric Z-scoring tuned for these small effects.

Benchmark (Colic et al. 2019): On DNA-damage-response chemogenomic screens, drugZ hits were far more strongly enriched for the expected pathway (DDR) than STARS, MAGeCK, edgeR or RIGER hits across FDR thresholds, reflecting better sensitivity to the moderate fitness defects typical of drug-gene interactions. Compare methods on expected-pathway enrichment, not raw hit count.

The drugZ Algorithm (under the hood)

1. For each sgRNA, compute log2-fold-change drug vs vehicle: LFC_drug_vs_veh
2. Compute an empirical-Bayes Z per sgRNA: Z = LFC / eb_std, where eb_std is the standard deviation of a sliding window of guides with similar control abundance (--half_window_size, default 500), smoothed monotonically
3. Per gene, sum Z across all sgRNAs targeting it: sumZ = sum(Z_sgRNA)
4. Normalize and re-standardize across genes: normZ = zscore(sumZ / sqrt(numObs))
5. Compute a one-sided p-value per direction: synth (sensitizer = negative normZ) and supp (resistance = positive normZ)
6. Benjamini-Hochberg FDR correction per direction

Critical: Vehicle vs drug, NOT Day 0 vs drug. Day-0 baseline conflates proliferation effects with drug effects.

Run drugZ on a Drug-Modifier Screen

Goal: Quantify per-gene sensitizing and suppressor effects from a chemogenomic screen.

Approach: Run drugz.py with vehicle and drug sample columns; output per-gene sumZ, normZ, and direction-specific p-values + FDR.

git clone https://github.com/hart-lab/drugz
cd drugz

# Standard drug screen comparison:
# Vehicle (DMSO or carrier) replicates: Veh_r1, Veh_r2, Veh_r3
# Drug-treated replicates: Drug_r1, Drug_r2, Drug_r3

python drugz.py \
    -i counts.txt \                       # input read-count file (tab-separated)
    -o drugz_output.txt \                  # output file
    -c Veh_r1,Veh_r2,Veh_r3 \              # control samples (comma-separated)
    -x Drug_r1,Drug_r2,Drug_r3 \           # treated samples (comma-separated)
    -r RPS3,RPL11,EIF3A \                  # OPTIONAL: comma-delimited GENE NAMES to exclude (not a file)
    -p 5                                   # pseudocount (default 5)

# Output: drugz_output.txt with columns:
#   GENE, sumZ, numObs, normZ, pval_synth, rank_synth, fdr_synth, pval_supp, rank_supp, fdr_supp

Output columns:

Interpretation:

• Sensitizers (synthetic lethal): fdr_synth < 0.05 -- loss of these genes makes cells more sensitive to drug. Examples: PARPi targets BRCA1/2; cisplatin sensitizes ERCC.
• Suppressors (resistance): fdr_supp < 0.05 -- loss of these genes confers resistance. Examples: drug-efflux genes; drug target itself paradoxically.

Vehicle vs Day-0 Reference: Critical Decision

Why this matters: Drug screen analysis can compare drug to:

1. Vehicle (DMSO / carrier) -- isolates drug-specific effect; correct anchor.
2. Day 0 (initial library) -- conflates proliferation, drug, and vehicle effects.

counts at Day 0          (no perturbation; cloning baseline)
    |
    v
counts at Day 7 - Vehicle (proliferation only; what survives in normal culture)
counts at Day 7 - Drug    (proliferation + drug effect)
    |
    v
Drug effect = LFC(Drug vs Vehicle)         # CORRECT
Wrong:       LFC(Drug vs Day 0)            # confounds drug with general proliferation

drugZ specifically requires -c to name the vehicle samples. Always include matched vehicle controls in drug screens.

Drug-Dose and Time-Course Designs

drugZ for dose-response: Not natively designed for dose; instead, run drugZ separately at each dose vs vehicle, then look for genes with consistent direction across doses.

for DOSE in low mid high; do
    python drugz.py \
        -i counts.txt \
        -o drugz_${DOSE}.txt \
        -c Veh_r1,Veh_r2 \
        -x Drug${DOSE}_r1,Drug${DOSE}_r2
done

# Then aggregate: genes significant at high dose AND consistent direction at mid/low dose

For multi-condition drug-screens (time × drug × cell-line), use MAGeCK MLE with explicit design matrix instead -- MLE handles multi-factorial; drugZ does not.

Comparison: drugZ vs MAGeCK MLE for Drug Screen

Goal: When to use each method.

Reconciliation: For simple drug-modifier screens with one drug and one vehicle, run both drugZ and MAGeCK MLE; hits called by both are high confidence; drugZ-only hits at low LFC need orthogonal validation (drug + arrayed validation).

Removing Genes from Null Distribution

Goal: Exclude reference essential or control genes from the Z-score null distribution.

Approach: Provide -r with a file listing gene symbols whose sgRNA-level Z scores should not influence the null. Useful when CEGv2 essentials would otherwise inflate the null distribution.

# Pass a file with one gene per line
cat > remove_essential.txt <<EOF
RPS3
RPL11
EIF3A
POLR2A
CDK1
EOF

python drugz.py \
    -i counts.txt \
    -o drugz_clean.txt \
    -c Veh_r1,Veh_r2 \
    -x Drug_r1,Drug_r2 \
    -r remove_essential.txt

When to use: If pilot drugZ runs show many essential genes appearing as "sensitizers" purely because they drop out under any condition, removing them gives a cleaner drug-specific signal.