BAGEL2 Essentiality - Usage Guide

Overview

Decision-grade essentiality calling for CRISPR-Cas9 fitness screens using BAGEL2 (Kim & Hart 2021 Genome Medicine). Computes Bayes Factors per gene by log-likelihood ratios over per-sgRNA fold changes, anchored against CEGv2 reference core-essentials (684 genes) and NEGv1 reference non-essentials (927 genes). Improvements over BAGEL1: linear extrapolation enables tumor-suppressor detection; multi-target off-target correction reduces false positives. BF >=6 corresponds to ~90% posterior probability of essentiality (Hart 2017); stricter BF cutoffs are BAGEL convention rather than a published FDR mapping.

Prerequisites

# BAGEL2 (Python)
git clone https://github.com/hart-lab/bagel
# Run BAGEL.py directly from the clone; the repo has no setup.py and no PyPI release

# Reference gene sets
wget https://raw.githubusercontent.com/hart-lab/bagel/master/CEGv2.txt
wget https://raw.githubusercontent.com/hart-lab/bagel/master/NEGv1.txt

Required inputs: - Count matrix (tab-separated, columns: sgRNA, GENE, then sample columns) - Reference essential genes (CEGv2.txt) - Reference non-essential genes (NEGv1.txt) - Control sample (Day 0 or plasmid) for fold-change baseline

Quick Start

Tell the AI agent what to do: - "Run BAGEL2 on my Brunello screen: compute fold changes then Bayes factors using CEGv2/NEGv1; pick BF threshold from precision-recall analysis" - "Identify tumor-suppressor candidates from BAGEL2 negative BF values; cross-check against COSMIC tumor suppressor list" - "Compare BAGEL2 BF vs MAGeCK FDR on the same dataset; reconcile disagreements at BF 5-7 boundary" - "Diagnose BAGEL2 calling no hits despite known essentials -- is it a reference-set issue?" - "Calibrate BF threshold for clinical-grade essentiality: BF >12 vs BF >30, using BAGEL.py pr on my own screen"

Example Prompts

Standard Workflow

"Run BAGEL.py fc then bf on counts.txt with controls Plasmid and treatment samples Drug_r1,Drug_r2,Drug_r3. Output bayes_factor.txt sorted by BF descending."

"Run BAGEL.py pr after to generate precision-recall curve against CEGv2; pick the BF threshold that reaches 95% precision on my screen's own PR curve."

Tumor Suppressor Detection

"From BAGEL2 output, identify genes with BF <-6 (significantly negative). Cross-reference against COSMIC tumor suppressor gene list. Output candidate tumor suppressors with their BF, fold change, and sgRNAs."

"BAGEL2 calls 80 tumor-suppressor candidates in my dropout screen. Many don't replicate in literature -- investigate whether my screen design supports tumor-suppressor calling (drug-modifier vs simple dropout)."

Calibration and Comparison

"Compare BAGEL2 BF >6 hits vs MAGeCK RRA neg|fdr <0.05 hits on the same screen. Compute Jaccard similarity. Investigate top 10 disagreements."

"Adjust BF threshold based on screen quality. High-quality screen (PR-AUC against CEGv2 >0.85) supports BF >6. Lower-quality may need BF >12 for same precision."

Diagnostics

"My BAGEL2 returns no genes with BF >0. Diagnose: wrong reference set file, library coverage too low, or genuinely flat screen?"

"Per-sgRNA LLR contributions show one guide dominating BF for several hits. Apply second-best-sgRNA rule from [[hit-calling]] to filter these guide-of-one hits."

What the Agent Will Do

Verify input file format: sgRNA, GENE, sample columns
Download/verify CEGv2 and NEGv1 reference files from hart-lab
Run BAGEL.py fc to compute fold changes from control samples
Run BAGEL.py bf to compute Bayes Factors (default: 10-fold cross-validation; -b -NB 1000 to bootstrap)
Run BAGEL.py pr to generate precision-recall curve
Apply BF threshold (default BF >6, ~90% posterior per Hart 2017)
Stratify genes: essential (BF >6), neutral (-6 to 6), candidate tumor suppressors (BF <-6)
For tumor suppressors, cross-check against COSMIC / published tumor suppressor lists
Compare to MAGeCK output (if also run); report consensus hits
Cross-check per-sgRNA LLR contributions for low-efficacy guides
Output essential genes ranked by BF, tumor suppressor candidates flagged, library calibration metrics

Tips

The single most common silent failure: wrong reference gene set file (truncated, gene-symbol mismatch). Always verify by counting genes in CEGv2.txt and NEGv1.txt before running.
BF >6 corresponds to ~90% posterior probability against CEGv2 in Hart 2017 (the ~5% FDR mapping is a rough BAGEL convention). For cell types outside cancer (iPSC, primary T cells), this calibration may not hold; use a cell-type-specific essentialome derived for the relevant lineage.
BAGEL2's tumor-suppressor sensitivity comes from linear extrapolation in the BF calculation. This is a real improvement over BAGEL1 and worth using for drug-modifier screens or screens expecting positive selection.
For copy-number-confounded cancer-line screens, pre-correct with CRISPRcleanR (see [[copy-number-correction]]) before BAGEL2; otherwise, amplified regions will appear as "essential" with high BF.
The resampling CI (STD column) is the diagnostic for guide-quality issues. Wide CI = low confidence; investigate per-sgRNA contributions.
For screens with only 3 sgRNAs/gene (some custom libraries), switch to bootstrapping with -b -NB to 5000+ for stable estimates.
When BAGEL2 and MAGeCK disagree, BAGEL2 typically calls more hits in screens with high background variance (it's robust due to reference-set anchoring) and fewer in screens with strong NTC null distribution.
For non-cancer cell types, cross-check that CEGv2 genes drop out at expected rate. If not, use cell-type-specific reference.

Decision Cheat Sheet

Question	Answer
Standard cancer-line essentiality	BAGEL2 with CEGv2/NEGv1; BF >6 threshold
Drug-modifier screen wanting both directions	BAGEL2 (tumor-suppressor sensitive) or drugZ
Multi-cell-line panel	Chronos (DepMap) preferred over BAGEL2
iPSC / primary cell / non-cancer	Use cell-type-specific essentialome
Custom library, <4 sgRNAs/gene	Bootstrap with `-b -NB`; consider MAGeCK alternative
Clinical-grade essentiality	BF >12 or higher

Thresholds

Precision and recall are screen-specific -- generate them with BAGEL.py pr rather than assuming fixed values. The BF ladder used in practice:

BF threshold	Use
0	Exploratory; highest recall
6	Standard call (~90% posterior, Hart 2017)
12	High-confidence
30	Ultra-stringent; near-certain essentials

Validation Checklist

[ ] CEGv2 and NEGv1 files downloaded and verified
[ ] Resampling left at the 10-fold cross-validation default, or -b -NB 1000 used deliberately
[ ] PR curve generated against CEGv2 for empirical threshold selection
[ ] Per-sgRNA LLR distributions checked for outliers
[ ] Hits cross-validated with MAGeCK or JACKS (consensus tier)
[ ] Tumor-suppressor calls (if any) cross-checked against COSMIC TS list

crispr-screens/mageck-analysis - MAGeCK as alternative
crispr-screens/drugz-chemogenomic - drugZ for drug screens (also tumor-suppressor sensitive)
crispr-screens/jacks-analysis - JACKS efficacy diagnostics for guide-of-one
crispr-screens/hit-calling - Cross-method consensus
crispr-screens/screen-qc - Pre-BAGEL CEGv2 PR-AUC
crispr-screens/library-design - 4-6 sgRNAs/gene library standard
crispr-screens/copy-number-correction - Pre-correction for cancer-line screens
pathway-analysis/go-enrichment - Downstream