Bio Workflows Crispr Screen Pipeline

Version Compatibility

Reference examples tested with: MAGeCK 0.5.9+, BAGEL2 1.0.5+, drugZ Aug 2019+, JACKS 0.2.0+, Chronos 2.0+, CRISPRcleanR 3.0+ (R), Pertpy 0.6+, PRIDICT2, CRISPResso2 2.2.14+, MAGeCKFlute 2.0+, pandas 2.2+, numpy 1.26+, matplotlib 3.8+.

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

• CLI: mageck --version, BAGEL.py fc --help, drugz -h, CRISPResso --version
• Python: pip show pertpy scanpy anndata (mageck-vispr via conda mageck --version; JACKS/Chronos are GitHub installs — check their repos)
• R: packageVersion('CRISPRcleanR'), packageVersion('MAGeCKFlute')

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

CRISPR Screen Pipeline

"Analyze my pooled or single-cell CRISPR screen end-to-end" -> Pick the screen design branch, run guide counting, audit six QC stages, apply copy-number and batch correction as needed, run the design-matched hit-calling method, and consolidate across methods for high-confidence hits.

This is a workflow skill: it owns the chaining decisions and hand-offs, not the internals of any one step.

The governing principle

Every LFC, QC gate, and hit call is computed against a reference that is committed once at library-order time; a wrong-but-silent commitment invalidates the endpoint with no error thrown.

1. The guide LIBRARY definition (guide->gene map + control classes) is the denominator, the calibrator, and the training reference — committed once. The library must carry non-targeting controls (NTCs, ~1%, the null distribution) AND CEGv2 reference essentials + NEGv1 non-essentials (the positive/negative calibrators for PR-AUC and BAGEL2/Chronos priors). NTCs calibrate the null/FDR; CEGv2/NEGv1 calibrate PR-AUC — swapping or dropping a class silently breaks FDR or QC.
2. The baseline choice has a right answer and rescales every hit. Dropout/enrichment LFC is against a baseline: plasmid pool for the cloning-bottleneck baseline, Day-0/T0 for the biology baseline, and vehicle (NOT Day-0) for drug screens — drug-vs-Day-0 conflates drug effect with normal proliferation.
3. Copy-number correction MUST precede hit calling in cancer cell lines. Multiple simultaneous Cas9 cuts at amplified loci trigger a gene-independent DNA-damage/G2 arrest (Aguirre 2016; Munoz 2016; the effect appears in both TP53-mutant and TP53-wild-type lines, though Aguirre 2016 found TP53 status correlates with its magnitude -- separately, Ihry 2018 / Haapaniemi 2018 report p53-dependent toxicity of Cas9 cutting generally), so amplified regions look essential regardless of gene function; calling hits first yields false essentials at ERBB2/MYC/FGFR1. Run CRISPRcleanR/Chronos BEFORE hit calling, or use CRISPRi to bypass the DSB. This is a pipeline step, not a post-hoc interpretation. Verifying abs(rho(LFC,CN)) < 0.1 afterwards needs a matched CN profile: CRISPRcleanR corrects unsupervised without one, so the check happens in crispr-screens/copy-number-correction (or use Chronos, which consumes CN directly).
4. CEGv2 essential-gene depletion is the screen's built-in positive control. If known essentials do not deplete (CEGv2 PR-AUC below ~0.7), the screen failed selection and NO novel hit is trustworthy regardless of its p-value — the seam analog of a spike-in. Normalize -> QC -> (CN correct) -> hit-call, never hit-call first; add batch as an MLE covariate, never pre-corrected with ComBat on counts (distorts the NB mean-variance the caller assumes).

Made-once commitments

Pipeline Branches by Screen Design

                    Library Design ([[library-design]])
                              |
                              v
                FASTQ Files -> mageck count -> count matrix
                              |
                              v
                Six-Stage QC ([[screen-qc]])
                              |
        +---------------------+---------------------+
        |                                            |
        v                                            v
  Cancer cell line?                          Non-cancer?
  Apply CN correction                        No CN correction needed
  ([[copy-number-correction]])
        |                                            |
        +---------------------+---------------------+
                              v
                  Multi-batch? Apply batch covariate
                  ([[batch-correction]])
                              |
                              v
                 Pick hit-calling method by design ([[hit-calling]])
                              |
        +-----------+---------+---------+-----------+-----------+
        |           |         |         |           |           |
        v           v         v         v           v           v
    2-cond       Time      Drug      Essential   Multi-       Specialized
    MAGeCK RRA   MAGeCK    drugZ     BAGEL2      screen       (PE/BE/SC/
                 MLE                              JACKS or     in vivo/
                                                  Chronos      combinat)
        |           |         |         |           |           |
        +-----------+---------+---------+-----------+-----------+
                              v
                   Tier-based consensus
                              v
                Orthogonal validation

Step 1: Library Design and Pre-Screen Validation

Reference [[library-design]] for full library composition. Verify before sequencing:

• Plasmid pool Gini <0.1 (Li W et al 2015 MAGeCK-VISPR, Genome Biol 16:281)

• =99% guides detected at >25 reads/guide

• Skew (p90/p10) <2
• NTCs comprise ~1% of library; CEGv2 reference essentials + NEGv1 non-essentials included

Step 2: Guide Counting

Goal: Turn raw FASTQ into a per-guide count matrix with consistent sample labels.

Approach: Run mageck count with the library CSV, sample labels in column order, the vector adapter trimmed off the 5' end, and median normalization.

mageck count \
    --list-seq library.csv \
    --sample-label Plasmid,Day0,Veh_r1,Veh_r2,Drug_r1,Drug_r2 \
    --fastq Plasmid.fq.gz Day0.fq.gz Veh_r1.fq.gz Veh_r2.fq.gz Drug_r1.fq.gz Drug_r2.fq.gz \
    --norm-method median \
    --output-prefix experiment \
    --trim-5 5   # integer base-count (or AUTO), NOT an adapter sequence; 5 trims the CACCG scaffold

For Cas12a libraries (Inzolia, in4mer): see [[combinatorial-screens]]. For 10X single-cell direct capture: use cellranger-arc or pertpy-aware counting; see [[perturb-seq-analysis]].

Step 3: Six-Stage Quality Control

Goal: Decide whether the screen is analyzable before calling any hits, using six orthogonal QC stages.

Approach: Load the count matrix, compute per-sample Gini, zero-fraction, and depth plus replicate correlation against the hard gates below. Essential-gene recovery (CEGv2 PR-AUC) is a separate check computed once endpoint-vs-baseline LFCs exist (it needs CEGv2/NEGv1 labels) -- see screen-qc.

import pandas as pd
import numpy as np

counts = pd.read_csv('experiment.count.txt', sep='\t', index_col=0)
genes = counts['Gene']
count_matrix = counts.drop('Gene', axis=1)

def gini(x):
    x = np.sort(x[x > 0].astype(float))
    if x.size == 0:
        return np.nan
    n = x.size
    cumx = np.cumsum(x)
    return (n + 1 - 2 * np.sum(cumx) / cumx[-1]) / n

per_sample = pd.DataFrame({
    'pct_zero': (count_matrix == 0).sum() / len(count_matrix) * 100,
    'gini': count_matrix.apply(gini),
    'reads_per_sgrna': count_matrix.sum() / len(count_matrix),
})

log_counts = np.log10(count_matrix + 1)
pearson = log_counts.corr()
print(per_sample)
print('Replicate Pearson:', pearson.values[pearson.values < 1].mean())

Hard gates from [[screen-qc]]:

• Plasmid Gini <0.1; endpoint <0.3 (or <0.55 for heavy drug screens)
• Replicate Pearson on log-counts >0.85
• CEGv2 PR-AUC >0.7 against the Hart 2017 reference essential gene set (community convention, not a threshold defined in that paper)
• Reads per sgRNA per sample >=300 (DepMap convention)

Step 4: Copy-Number Correction (Cancer Cell Lines Only)

If screening in a cancer cell line, apply CRISPRcleanR (unsupervised, no CN profile needed) or Chronos (joint with CN profile). Required to remove Aguirre 2016 / Munoz 2016 amplicon artifact.

Goal: Strip the copy-number amplicon artifact that makes amplified regions look essential in cancer lines.

Approach: Run CRISPRcleanR unsupervised genome-wide LFC correction (no CN profile needed), then feed the corrected counts downstream; for DepMap-scale panels with matched CN, use Chronos instead.

library(CRISPRcleanR)
data(KY_Library_v1.0)
norm <- ccr.NormfoldChanges('experiment.count.txt', min_reads = 30, EXPname = 'screen',
                              libraryAnnotation = KY_Library_v1.0)   # arg 1 is the file PATH
gw_lfc <- ccr.logFCs2chromPos(norm$logFCs, KY_Library_v1.0)          # $logFCs, not $norm_fold_changes
cleaned <- ccr.GWclean(gw_lfc, display = TRUE, label = 'screen')
corrected_counts <- ccr.correctCounts('screen', norm$norm_counts, cleaned,
                                        KY_Library_v1.0)              # (CL, normalised_counts, correctedFCs, libraryAnnotation)
# ccr.correctCounts returns an in-memory frame; it does NOT write this file. Persist it, because the
# hit callers below read a count TABLE from disk -- the CN-correction commitment in rule 3 is only
# honored if that file, not the raw experiment.count.txt, is what MAGeCK / BAGEL2 / drugZ consume.
write.table(corrected_counts, 'screen_cleanr_corrected_counts.txt',
            sep = '\t', quote = FALSE, row.names = FALSE)

For DepMap-scale panels with longitudinal data + matched CN, use Chronos. See [[copy-number-correction]].

Step 5: Batch Correction (Multi-Batch Screens)

For multi-batch screens, add batch as a covariate in MAGeCK MLE rather than pre-correcting with ComBat. See [[batch-correction]] for full decision tree.

Step 6: Method-Matched Hit Calling

Goal: Call hits with the method that matches the experimental design, plus at least one orthogonal method for consensus.

Approach: Pick by design - RRA or BAGEL2 for two-condition essentiality, MLE for time course, drugZ for drug-modifier, JACKS for multi-screen, Chronos for cancer panels - and run two methods so the consensus step has something to reconcile.

6a. Two-condition essentiality (MAGeCK RRA or BAGEL2)

Cancer cell lines: pass the CRISPRcleanR-corrected count file (screen_cleanr_corrected_counts.txt from Step 4) as --count-table/-i below, NOT the raw experiment.count.txt — the CN-correction commitment is only honored if the corrected counts are what the hit caller reads.

The Day0/Day14_r* columns below illustrate a time-course dropout design; they must match the count step's --sample-label (the drug-screen count above uses Plasmid,Day0,Veh_r*,Drug_r*).

mageck test \
    --count-table experiment.count.txt \
    --treatment-id Day14_r1,Day14_r2,Day14_r3 \
    --control-id Day0 \
    --norm-method median \
    --output-prefix essentiality_rra

BAGEL.py fc -i experiment.count.txt -o experiment -c Day0 --min-reads 30   # -o is an output LABEL; fc writes experiment.foldchange
BAGEL.py bf -i experiment.foldchange -o bayes_factor.txt -e CEGv2.txt -n NEGv1.txt \
    -c Day14_r1,Day14_r2,Day14_r3   # add -b -NB 1000 for bootstrapping; -k is not a bf option