Bio Crispr Screens Copy Number Correction

Version Compatibility

Reference examples tested with: CRISPRcleanR 3.0+ (R; github.com/francescojm/CRISPRcleanR), Chronos 2.0+ (https://github.com/broadinstitute/chronos), CERES (legacy, superseded by Chronos), pandas 2.2+, numpy 1.26+, scipy 1.12+.

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

• R: packageVersion('CRISPRcleanR'); ?ccr.GWclean
• Python: pip show crispr_chronos; python -c 'import chronos; print(chronos.__file__)'

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

Copy-Number Bias Correction in CRISPR Screens

"Correct copy-number artifacts in my cancer-cell-line screen" -> Identify gene-independent depletion at amplified loci, apply CRISPRcleanR (pre-hoc, unsupervised, position-based) or Chronos (joint model, supervised with CN profile) to remove the artifact, then proceed to hit calling on corrected data.

• R: CRISPRcleanR::ccr.GWclean() for unsupervised pre-hoc correction (no CN profile required)
• Python: Chronos (crispr_chronos) for joint cell-population dynamics + CN modeling
• Python: CERES (legacy, superseded by Chronos)

The Copy-Number Artifact (Mechanism)

Aguirre AJ et al 2016 Cancer Discov 6:914 and Munoz DM et al 2016 Cancer Discov 6:900 demonstrated that focal amplification regions in cancer cell lines appear systematically "essential" in CRISPR-Cas9 screens, independent of the gene's actual biology. The mechanism:

1. A focal amplification creates 4-50+ copies of a genomic region.
2. Each sgRNA targeting a gene in that region cuts at all copies simultaneously.
3. Multiple cuts trigger a DNA-damage response and G2 arrest, in both TP53-mutant and TP53-wild-type lines but with larger magnitude in wild-type (Aguirre 2016).
4. Cells arrest in G2 phase; the sgRNA appears depleted because its bearer cells don't proliferate.
5. The depletion is proportional to the number of simultaneous cuts, not the gene's essentiality.

Consequence: ERBB2 appears essential in HER2-amplified SK-BR-3. MYC appears essential in MYC-amplified colorectal lines (10+ copies). FGFR1 appears essential in FGFR1-amplified head-and-neck lines. These are all false positives.

Affects: All Cas9-KO screens in cancer cell lines. Universal, not conditional. Cannot be remediated by sequencing depth, library size, or replicate count. Requires explicit correction.

The p53-dependence of Cas9-cut toxicity in general was characterized later, by Haapaniemi 2018 and Ihry 2018.

Bypassed by:

• CRISPRi (catalytically dead Cas9, no DNA damage) -> no artifact
• CRISPRa (catalytically dead Cas9) -> no artifact
• Base editing (single-strand nick + deaminase) -> reduced artifact
• Prime editing (nick + RT) -> reduced artifact

Correction Method Decision Tree

CRISPRcleanR (Iorio 2018) - Unsupervised Pre-Hoc

Goal: Correct copy-number bias without requiring matched CN profile by detecting position-based systematic enrichment / depletion patterns.

Approach: Order sgRNAs by chromosomal coordinate; detect segments where sgRNAs show systematic depletion (or enrichment) inconsistent with single-gene biology; shift these segments toward the global mean. The intuition: focal amplifications create depletion bands extending tens to hundreds of kb; non-amplified essential genes are punctate.

library(CRISPRcleanR)

# Load library annotation (sgRNA -> chromosomal coordinates)
data(KY_Library_v1.0)   # KY library; replace with your library annotation
# OR use ccr.PrepareAnnotations() to make custom

# Load count data with first 2 cols: sgRNA, gene, then sample counts
counts <- read.table('counts.txt', header=TRUE, sep='\t')

# 1. Normalize and compute logFC
norm_counts <- ccr.NormfoldChanges(filename='counts.txt', min_reads=30,
                                     EXPname='my_screen',
                                     libraryAnnotation=KY_Library_v1.0)

# 2. Compute genome-sorted sgRNA fold changes
gw_log_fc <- ccr.logFCs2chromPos(norm_counts$logFCs,
                                   KY_Library_v1.0)

# 3. Apply CRISPRcleanR correction
corrected <- ccr.GWclean(gw_log_fc, display=TRUE, label='my_screen')
# Output: corrected$corrected_logFCs and corrected$segments
# corrected_logFCs can replace LFCs downstream

# 4. Re-derive corrected counts for downstream MAGeCK
corrected_counts <- ccr.correctCounts('my_screen',
                                        norm_counts$norm_counts,
                                        corrected,
                                        KY_Library_v1.0,
                                        OutDir='./')

Key parameter: min_reads=30 is the lower-count threshold for inclusion. This must match the library-coverage strategy; too high removes legitimate guides, too low keeps noisy guides.

Output: Pre-corrected LFCs and counts that can be fed into MAGeCK / BAGEL2 / drugZ as if they were the original screen data. The correction is independent of CN profile (unsupervised) and works on cell lines without matched WGS.

Chronos (Dempster 2021) - Joint Population-Dynamics + CN Model

Goal: Estimate gene fitness while jointly accounting for copy-number-driven depletion, screen quality, and longitudinal cell-population dynamics.

Approach: Model the cell population over time as an ODE driven by per-gene fitness effects; add a separate term for copy-number-driven depletion; estimate all parameters via maximum-likelihood with regularization. Outputs a "gene effect score" normalized against the empirical distributions of essential and non-essential reference genes.

# Chronos (pip install crispr_chronos, or pip install git+https://github.com/broadinstitute/chronos)
import chronos
from chronos.hit_calling import get_probability_dependent

# Inputs
# 1. Counts: rows = sgRNA, columns = samples (per-timepoint per-cell-line)
# 2. Sequence map: sgRNA -> cell line -> sample timepoint
# 3. Guide-gene map
# 4. Copy-number profile per cell line (applied AFTER training, not at construction)

# All three inputs are dicts of DataFrame keyed by library name, not bare DataFrames.
model = chronos.Chronos(
    sequence_map={'screen': sequence_map},
    guide_gene_map={'screen': guide_gene_map},
    readcounts={'screen': counts_df},
)
model.train(nepochs=301)
gene_effects = model.gene_effect                      # attribute, not a method call

# Copy-number correction is a separate post-hoc step, not a constructor argument
gene_effects_cn = chronos.alternate_CN(gene_effects, copy_number_df)
gene_probabilities = get_probability_dependent(gene_effects_cn, negative_control_genes, positive_control_genes)

DepMap convention: A gene-effect score <-1 corresponds to "essential" in that cell line; <-0.5 is "depleting." Each DepMap release (quarterly) provides Chronos gene effects and probabilities.

Critical: Chronos benefits most from longitudinal data (multiple timepoints per cell line) but can run with multiple cell lines at a single timepoint. Copy number is optional: Chronos trains without it and alternate_CN applies the correction afterwards. For a single screen (one line, one timepoint) without a matched CN profile, use CRISPRcleanR instead.

CERES (Legacy, Superseded by Chronos)

Meyers RM et al 2017 Nat Genet 49:1779 introduced the first formal CN-correction method at DepMap scale. CERES decomposes per-sgRNA LFC as sgRNA_efficacy * gene_effect - CN_term(copy_number), fitting jointly. Superseded by Chronos at DepMap in 2021 due to Chronos' better handling of screen quality and longitudinal data. CERES remains useful for cross-validation.

Detect Uncorrected CN Bias

Goal: Verify that copy-number bias is corrected (or detect it in raw data).

Approach: For genes with matched CN profile, compute Spearman ρ between gene-level LFC and copy number. A negative correlation (-ρ) indicates amplified genes are depleted, i.e., CN artifact.

import pandas as pd
from scipy.stats import spearmanr

def detect_cn_bias(gene_lfc_df, cn_df):
    '''Test whether gene-level LFC negatively correlates with copy number.
    A bias-free screen has Spearman rho near zero between CN and LFC.'''
    merged = gene_lfc_df.merge(cn_df, on='gene')
    rho, p = spearmanr(merged['copy_number'], merged['lfc'])
    return {
        'cn_lfc_rho': rho,
        'p_value': p,
        'amplified_mean_lfc': merged[merged['copy_number'] > 4]['lfc'].mean(),
        'diploid_mean_lfc': merged[(merged['copy_number'] >= 1.5) & (merged['copy_number'] <= 2.5)]['lfc'].mean(),
        'bias_present': rho < -0.1 and p < 0.01,
    }

Threshold (operational convention): Spearman ρ <-0.10 between LFC and CN indicates significant CN bias. Even modest amplifications generate detectable artifact. Run this diagnostic before AND after correction.

Reconciliation: When CN Correction Fails

If post-CRISPRcleanR or post-Chronos the CN-LFC Spearman is still significantly negative, the correction is incomplete. Possible causes:

1. Insufficient CN resolution: A specific 4-copy region went undetected. Refine CN profile with deeper WGS.
2. CRISPRcleanR position-based correction missed it: The amplification is small relative to the segmentation algorithm's resolution. Use Chronos with matched CN profile.
3. Genomic rearrangement creates a "ghost" amplification: A complex rearrangement appears as normal CN but Cas9 cuts at multiple sites due to translocation breakpoints. Combine WGS structural variants with the analysis.
4. Cell line has an unusually strong cut-toxicity response: The artifact may persist; use CRISPRi screens for that line.

Apply CN Correction to Pipeline

Workflow:

1. mageck count (raw counts)
2. screen-qc verification
3. CN diagnostic: Spearman of LFC vs CN (if CN profile available)
4. If bias detected:
   a. CRISPRcleanR (pre-hoc) -> corrected counts -> MAGeCK / BAGEL2 / drugZ
   OR
   b. Chronos (joint model with CN profile) -> gene effects directly
5. Re-diagnose: Spearman of CORRECTED LFC vs CN should be near zero
6. Hit calling

For DepMap-style large panels:

Chronos handles batch + CN + screen quality in one step; no pre-correction needed.

For Project Score-style panel (Behan 2019):

CRISPRcleanR was used historically; cross-check with Chronos when CN profile available.

Failure Modes

CRISPRcleanR removes legitimate essential signal

Trigger: A genuine essential gene happens to lie in a region with adjacent uncorrected non-essential signal; the segment-based correction includes the essential. Mechanism: CRISPRcleanR's ccr.GWclean() segments sgRNAs by position; segments containing multiple genes with directional consistency are corrected as a unit. Symptom: A known essential drops out of post-correction hit list. Fix: Inspect segments manually; if a known essential was within a corrected segment, investigate. Cross-check with non-CN-corrected MAGeCK + BAGEL2 to see if essential was a hit pre-correction.