Bio Crispr Screens Hit Calling

Custom z-score Hit Calling (when standard tools don't fit)

Goal: Compute gene-level z-scores when neither MAGeCK nor BAGEL2 fits the experimental design.

Approach: RPM-normalize, compute per-sgRNA log2 fold-changes, aggregate to gene level, derive z-score from the null distribution of non-targeting controls (cleanest) or all genes (assumes <40% changing), apply BH correction.

import pandas as pd
import numpy as np
from scipy import stats
from statsmodels.stats.multitest import multipletests

def custom_zscore_hit_calling(counts_df, ctrl_cols, treat_cols, genes_series, ntc_genes=None):
    '''Z-score gene-level hit calling. If ntc_genes provided, null derived from NTCs only;
    otherwise from all genes (assumes <40% changing).'''
    def rpm(df):
        return df.div(df.sum(axis=0), axis=1) * 1e6
    ctrl_rpm = rpm(counts_df[ctrl_cols])
    treat_rpm = rpm(counts_df[treat_cols])
    lfc_per_sgrna = np.log2((treat_rpm.mean(axis=1) + 1) / (ctrl_rpm.mean(axis=1) + 1))
    gene_lfc = pd.DataFrame({'gene': genes_series, 'lfc': lfc_per_sgrna}).groupby('gene')['lfc'].agg(['mean', 'std', 'count'])
    gene_lfc.columns = ['mean_lfc', 'std_lfc', 'n_sgrnas']
    if ntc_genes is not None:
        null = gene_lfc.loc[gene_lfc.index.isin(ntc_genes), 'mean_lfc']
        null_mean, null_std = null.median(), null.std()
    else:
        null_mean = gene_lfc['mean_lfc'].median()
        null_std = gene_lfc['mean_lfc'].std()
    gene_lfc['z'] = (gene_lfc['mean_lfc'] - null_mean) / null_std
    gene_lfc['p'] = 2 * stats.norm.sf(np.abs(gene_lfc['z']))
    gene_lfc['fdr'] = multipletests(gene_lfc['p'], method='fdr_bh')[1]
    return gene_lfc.sort_values('z')

Failure Modes

MAGeCK and BAGEL2 disagree by 200+ hits at FDR 0.05

Trigger: Heavy-selection screen (>40% guides change) or cancer-line CN bias. Mechanism: MAGeCK median normalization breaks; BAGEL2 is robust due to reference-set anchoring. Symptom: MAGeCK hit list inflated; BAGEL2 list closer to expected size. Fix: Run MAGeCK with --norm-method control; apply CN correction; trust BAGEL2 for essentiality.

Chronos and MAGeCK disagree at the top 10 in a cancer line

Trigger: Top hits are at amplified loci. Mechanism: Chronos models CN bias; MAGeCK does not. Symptom: ERBB2 in HER2+, MYC in MYC-amplified, etc. are top hits in MAGeCK but not Chronos. Fix: Apply [[copy-number-correction]] before MAGeCK or switch to Chronos.

drugZ and MAGeCK disagree on small-effect drug-modifier screen

Trigger: Effect size is small; MAGeCK rank-based test is less sensitive than drugZ bidirectional Z. Mechanism: drugZ specifically optimized for small effects in drug screens (Colic et al. 2019); MAGeCK RRA loses sensitivity at small effects. Symptom: At matched FDR, drugZ calls small-effect chemogenomic interactions (e.g. DDR genes) that MAGeCK RRA misses, with stronger expected-pathway enrichment. Fix: Use drugZ as primary for chemogenomic; MAGeCK as confirmatory. See [[drugz-chemogenomic]].

JACKS down-weights efficiency, MAGeCK doesn't, disagreement

Trigger: A gene has one or two strong sgRNAs and 2-3 weak ones; MAGeCK averages them, JACKS down-weights the weak. Mechanism: JACKS variational Bayes correctly identifies low-efficacy guides; MAGeCK aggregates without this prior. Symptom: Gene is JACKS hit but not MAGeCK. Fix: Inspect per-sgRNA LFC; if strong guides are consistent, JACKS is correct. Validate gene orthogonally.

Consensus across 3 methods is empty (no hits)

Trigger: Either no real biology, or each method has different failure mode being triggered. Mechanism: Screen quality is low; signal-to-noise across all methods is poor. Symptom: Tier 1 consensus list is empty. Fix: Re-audit QC. Check Cas9 selection, MOI, timepoint, library positioning. Re-run screen if QC fails.

Quantitative Thresholds

Common Errors

References

• Li W et al. 2014. Genome Biol 15:554. MAGeCK alpha-RRA.
• Li W et al. 2015. Genome Biol 16:281. MAGeCK MLE.
• Kim E & Hart T. 2021. Genome Med 13:2. BAGEL2.
• Colic M et al. 2019. Genome Med 11:52. drugZ.
• Allen F et al. 2019. Genome Res 29:464. JACKS.
• Dempster J et al. 2021. Genome Biol 22:343. Chronos.
• Meyers R et al. 2017. Nat Genet 49:1779. CERES.
• Hart T & Moffat J. 2016. BMC Bioinformatics 17:164. BAGEL Bayes factor framework.
• Hart T et al. 2017. G3 7:2719. CEGv2/NEGv1 calibration.

Related Skills

• crispr-screens/mageck-analysis - Full MAGeCK RRA + MLE detail
• crispr-screens/bagel-essentiality - Full BAGEL2 detail
• crispr-screens/drugz-chemogenomic - Full drugZ detail for drug screens
• crispr-screens/jacks-analysis - Full JACKS detail and library calibration
• crispr-screens/copy-number-correction - Chronos, CERES, CRISPRcleanR
• crispr-screens/screen-qc - Quality gates that drive method choice
• crispr-screens/library-design - Library type dictates analysis method
• crispr-screens/combinatorial-screens - GI scoring (synthetic lethality)
• crispr-screens/perturb-seq-analysis - SCEPTRE for single-cell screens
• crispr-screens/batch-correction - Multi-batch normalization upstream of hit calling
• pathway-analysis/gsea - Downstream pathway enrichment
• pathway-analysis/go-enrichment - GO enrichment of hit lists