bio-crispr-screens-copy-number-correction

What this skill does

## Version Compatibility

Reference examples tested with: CRISPRcleanR 3.0+ (R/Bioconductor), Chronos 2.0+ (https://github.com/broadinstitute/chronos), CERES (legacy, superseded by Chronos), pandas 2.2+, numpy 1.26+, scipy 1.12+.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.CleanCN`
- Python: `pip show chronos-cn`; `chronos --help`

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.CleanCN()` for unsupervised pre-hoc correction (no CN profile required)
- Python: Chronos (`chronos-cn`) 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- Python: CERES (legacy, superseded by Chronos)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 p53-dependent DNA-damage response.
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 (24 copies). 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.

**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 (Hess 2016 demonstration)
- Prime editing (nick + RT) -> reduced artifact

## Correction Method Decision Tree

| Available data | Recommended method | Why |
|----------------|---------------------|-----|
| Cell-line panel without matched CN profile | CRISPRcleanR | Unsupervised; uses genomic position only |
| Single cell line with matched WGS/SNP-array CN | CRISPRcleanR or Chronos | Either works; Chronos more rigorous |
| DepMap-scale (1000+ cell lines, longitudinal) | Chronos | Population-dynamics + screen quality + CN; DepMap quarterly standard |
| Single cell line, multi-timepoint | Chronos | Leverages longitudinal counts |
| Need to integrate with downstream MAGeCK | CRISPRcleanR (pre-hoc) | Outputs corrected counts for any downstream tool |
| Multiple cell lines + multiple batches | Chronos | Joint modeling of all dimensions |

## 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.

```r
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(counts, 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$norm_fold_changes,
                                   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,
                                        correction=corrected,
                                        outprefix='cleanr_corrected',
                                        libraryAnnotation=KY_Library_v1.0)
```

**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.

```python
# Chronos via the chronos-cn package
import chronos

# Inputs
# 1. Counts: rows = sgRNA, columns = samples (per-timepoint per-cell-line)
# 2. Sequence map: sgRNA -> cell line -> sample timepoint
# 3. Copy-number profile per cell line per genomic region
# 4. Guide-gene map

model = chronos.Chronos(
    sequence_map=sequence_map,           # which samples are from which cell line + condition
    guide_gene_map=guide_gene_map,
    reads=counts_df,
    copy_number=copy_number_df,           # per-cell-line CN profile
    pretrained_offset=None,
)
model.train(
    n_steps=2000,
    learning_rate=0.1,
    verify_normalize=True,
)
gene_effects = model.gene_effect()        # cells x genes; standardized score
gene_probabilities = model.gene_probability()  # probability of being essential
```

**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 single timepoint; it cannot run with a single screen (one line, one timepoint) without matched CN profile. For single-cell-line, single-timepoint screens without matched CN, 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 c