CRISPR Screen QC - Usage Guide

Overview

Decision-grade quality control for pooled CRISPR screens. Covers six bottleneck stages (plasmid pool, Day-0 infection, selection, endpoint, biological signal, copy-number artifact), with stage-specific Gini, skew, replicate Pearson/Spearman, sequencing depth, MOI verification, and essentialome PR-AUC against CEGv2 (Hart 2017). Outputs a DepMap-style composite score and recommends remediation for each failing stage.

Prerequisites

conda install -c bioconda mageck   # not on PyPI
pip install pandas numpy scipy matplotlib seaborn scikit-learn
# MAGeCK QC dashboard
conda install -c bioconda -c conda-forge mageck-vispr
# R dashboard (optional)
R -e "remotes::install_github('WubingZhang/MAGeCKFlute')"   # removed from Bioconductor at 3.22

Required inputs: MAGeCK count output (screen.count.txt), plasmid-pool counts (separate file or first sample), known copy-number profile per cell line (from WGS / SNP-array / ASCAT / matched cell-line database), and CEGv2 / NEGv1 reference gene sets (CEGv2 from Hart 2017, NEGv1 from Hart 2014; hart-lab/bagel repository).

Quick Start

Tell the AI agent what to audit: - "Audit my Brunello-screen counts: plasmid Gini, Day-0 to endpoint dropout, replicate Pearson and Spearman, CEGv2 PR-AUC, depth in reads per sgRNA, MOI verification" - "Diagnose why hits include ERBB2 in HER2+ SK-BR-3 -- is this copy-number bias or real essentiality" - "Decide whether my screen passes DepMap quality thresholds, fails, or is salvageable" - "Recommend which hit-calling method (MAGeCK RRA / MLE / BAGEL2 / Chronos / drugZ) my screen quality grade supports" - "Generate a composite quality score across all my replicate pairs to gate which conditions enter hit calling"

Example Prompts

Library and Plasmid Pool QC

"Compute Gini coefficient and skew ratio (p90/p10) on my plasmid-pool sequencing. Pass at Gini <0.1 (MAGeCK-VISPR), zero-count <0.5% (Joung 2017), and skew <2 as a stricter modern convention than Joung 2017's <10. Decide whether to proceed."

"My plasmid Gini is 0.18 and skew is 4.2. Diagnose: PCR over-amplification, synthesis defect, or cloning bottleneck? Recommend remediation."

Replicate and Depth Audit

"Compute pairwise Pearson on log10(counts+1) and Spearman on raw ranks between all replicates within each condition. Flag pairs below Pearson 0.85 or Spearman 0.7."

"One of my treatment replicates shows Pearson 0.78 vs the other two replicates which are >0.95 with each other. Decide whether to drop or rescue the outlier replicate."

"Verify sequencing depth: reads per sgRNA per sample. Fail below 100 (Joung 2017 plasmid-QC floor), warn between 100 and 300 (MAGeCK-VISPR), pass at 300+, and treat 500+ as ideal for screening (Joung 2017)."

Biological Signal (Essentialome Recovery)

"Compute PR-AUC against CEGv2 essentials (Hart 2017) and NEGv1 non-essentials (Hart 2014). My screen passes only if PR-AUC >0.7. Tell me whether this screen has interpretable biology before I run hit calling."

"PR-AUC is 0.45 even though Gini and Pearson pass. Diagnose: Cas9 not selected pre-screen, premature timepoint, or library targets wrong TSS?"

Copy-Number Artifact Diagnostic

"Run the copy-number bias diagnostic. Compute Spearman ρ between gene-level LFC and copy number from the matched WGS profile. Flag CN bias if abs(ρ) >0.1 and recommend CRISPRcleanR / CERES / Chronos correction."

"Hits include ERBB2 in SK-BR-3 and FGFR1 in head-and-neck lines. Confirm whether these are copy-number artifacts before publication."

Composite Quality Gate

"Generate the DepMap-style composite quality score across all metrics: Gini-inverse, replicate Pearson minimum, PR-AUC, log-depth, detected-fraction. Use this to decide which conditions enter downstream hit calling."

MOI Verification

"Verify infection MOI: from the titration plate (8 wells with 1:2 serial dilution), interpolate the infection efficiency at the volume used in the screen. Compute Poisson P(≥2 sgRNAs/cell) at the resulting MOI."

What the Agent Will Do

Identify each sequencing-sample stage (plasmid, Day 0, selection, endpoint) from metadata
Run library_representation() per sample: % zero, % low-count, skew ratio
Compute Gini per sample with stage-specific thresholds
Compute replicate Pearson on log-counts and Spearman on raw ranks per condition
Compute CEGv2 PR-AUC at the endpoint timepoint, vs Day-0 or plasmid
Verify MOI from titration data or qPCR of integrated copies
Detect copy-number artifact via gene-level LFC vs copy-number correlation
Run PCA to visualize sample clustering (condition vs batch)
Generate composite quality score and per-sample grade
Recommend downstream hit-calling method based on quality grade (high quality -> MAGeCK MLE or Chronos; low quality -> RRA or drugZ; cancer-line -> Chronos with CN correction; in-vivo -> bottleneck-adjusted thresholds)

Tips

Plasmid-pool sequencing (Gini <0.1, ≥99% guide detection at >25 reads/guide) is non-negotiable; everything downstream is normalized against this baseline. A screen with un-sequenced plasmid pool is uninterpretable.
The single most diagnostic metric is CEGv2 PR-AUC; if it passes >0.7 the screen has biology even if individual sample metrics look weak. If it fails <0.5, no remediation in software will fix it -- the screen has no essentiality signal.
For drug screens, expect Gini to drift up (endpoint Gini 0.3-0.5) as biology drives selection; this is normal. Compare vs vehicle, not Day 0, for any chemogenomic interpretation.
Cancer-cell-line screens with focal amplification ALWAYS require copy-number correction. Apply CRISPRcleanR / Chronos / CERES preemptively; the Aguirre/Munoz copy-number artifact is universal, not conditional.
MOI 0.3 is non-negotiable; there is no analytical correction for high-MOI confounding. Re-run rather than try to model around it.
"Passing" QC at each stage is necessary but not sufficient; the final gate is essentialome PR-AUC.
High Pearson with low Spearman = a few outlier guides dominate; switch hit calling to RRA or drugZ which use ranks.
For CRISPRi/a, expect lower per-gene PR-AUC than Cas9 because not all essentials respond to knockdown the way they do to knockout; calibrate against the DepMap CRISPRi sub-essentialome rather than CEGv2.

QC Decision Reference

Stage	Primary metric	Pass	Action if fail
Plasmid	Gini, skew	<0.1, <2	Re-clone or re-amplify
Day 0	Pearson with plasmid	>0.9	Diagnose infection issue
Endpoint	Replicate Pearson	>=0.8 (MAGeCK-VISPR floor)	Drop outlier replicate
Biology	CEGv2 PR-AUC	>0.7	Cas9 selection / timepoint / TSS
CN	Spearman LFC vs CN	abs(ρ) <0.05 post-correction	CRISPRcleanR / Chronos
Depth	Reads/sgRNA	>300	Re-sequence
MOI	Poisson P(≥2)	<5%	Re-infect

crispr-screens/library-design - Design library with adequate skew margin
crispr-screens/mageck-analysis - Generate count files for QC
crispr-screens/copy-number-correction - Remediate CN artifact
crispr-screens/batch-correction - Address inter-batch drift
crispr-screens/hit-calling - Pick analysis method by quality grade
crispr-screens/in-vivo-screens - In-vivo-specific bottleneck QC
read-qc/quality-reports - General NGS QC upstream