GPTomics/bio-crispr-screens-base-editing-analysis

Version Compatibility

Reference examples tested with: CRISPResso2 2.2.14+, BE-Hive 1.0+ (BE prediction), pandas 2.2+, biopython 1.83+, numpy 1.26+, scipy 1.12+, scikit-learn 1.4+; Broad be-validation-pipeline notebooks (repo HEAD).

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

• CLI: CRISPResso --version
• Python: pip show CRISPResso2; BE-Hive is a GitHub clone (maxwshen/be_predict_bystander), not a PyPI package

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

Base Editing Screen Analysis

"Analyze my base-editor variant-function screen" -> Quantify per-sgRNA target-base conversion, bystander rate, and indel byproducts from amplicon sequencing; filter on editing efficiency; map each sgRNA to its intended SNV (target + bystander pattern); compute per-variant fitness from the screen log-fold change; reconcile target vs bystander variant attribution; annotate against ClinVar / COSMIC.

• CLI: CRISPResso --base_editor_output for per-amplicon BE quantification
• CLI: Broad be-validation-pipeline for end-to-end pooled-screen analysis with editing-efficiency filtering
• Python: BE-Hive (Arbab 2020) for editing-efficiency prediction; clone maxwshen/be_predict_bystander and import via sys.path
• Web: BE-Designer (Hwang 2018, RGEN Tools) for variant-encoding sgRNA design

Base Editor Chemistry Selection

| Editor | Reaction | Editing window | Indel byproduct rate | When to use | |--------|----------|----------------|----------------------|-------------| | BE3 (Komor 2016) | C->T (also G->A on opposite strand) | Pos 4-8 from PAM-distal end | 5-10% | Original; superseded | | BE4 / BE4max (Koblan 2018) | C->T | Pos 4-8 | <5% | CBE standard | | eA3A-BE3 | C->T narrow specificity | Pos 5-7 | <5% | Specifically TC contexts (eA3A prefers TC) | | ABE7.10 (Gaudelli 2017) | A->G (T->C opposite strand) | Pos 4-7 | <2% | First ABE; slow at non-TA contexts | | ABE8.20 (Gaudelli 2020) | A->G | Pos 4-8 | <2% | Modern ABE; high activity | | ABE8e (Richter 2020) | A->G | Pos 4-8 | <2% | Highest editing activity; more processive than ABE7.10 | | evoCDA-BE | C->T (broader) | Pos 1-9 | 5-10% | Larger editing window; more bystander | | CGBE1 (Kurt 2021) | C->G | Pos 5-7 | 5-10% | C-to-G transversion; rare use | | GBE (Zhao 2021) | C->G or C->A | Pos 4-7 | 5-10% | Transversions; less mature |

Decision rule: For a target SNV at position 4-8 of a candidate spacer with no bystander Cs/As in the same window, BE3-BE4 or ABE7.10 is sufficient. For high-throughput variant scanning where bystander tolerance must be minimized, use eA3A-BE3 (TC contexts only) for C->T, or ABE7.10 rather than ABE8e/ABE8.20 for A->G -- its 4-7 window is the narrowest ABE.

Editing Window Math

Why this matters for postdoc-level use: Base editors are tethered to dCas9 (or nCas9) and the deaminase acts on the displaced ssDNA "R-loop" formed when Cas9 binds. The deaminase has a fixed reach -- positions 4-8 from the PAM-distal end of the protospacer for canonical BE3/BE4, and 4-7 for ABE7.10. Outside this window, editing efficiency drops by 10-50x.

PAM-distal end                                                            PAM-proximal
   |                                                                          |
   1  2  3  4  5  6  7  8  9  10 11 12 13 14 15 16 17 18 19 20    NGG
                  ^^^^^^^^^^^
                  Canonical editing window (positions 4-8)

   For BE4max: positions 4-8 are 5-50x more efficient than positions 1-3 or 9-13 (ABE7.10: 4-7)
   For SpABE8e: positions 4-8 (Richter 2020), matching the corresponding CBEs rather than ABE7.10's narrower 4-7
   For evoCDA-BE: window 1-9 (broader; more bystander)

Critical implication for variant interpretation: If the intended edit is at position 5 and there is an additional editable C/A at position 7, both will be edited in the same molecule. The screen scores the combination of edits, not the intended one alone. This is bystander confounding.

sgRNA Library Design for BE Screens

Goal: Tile editing-window-positioned spacers across a protein region of interest to enable variant scanning.

Approach: For each amino acid in the target region, find NGG-adjacent spacers where the SNV-of-interest base falls in editing positions 4-8 with minimal bystander C/A in the same window. Annotate each spacer with the predicted amino acid changes (target + bystander).

import pandas as pd
import re
from Bio.Seq import Seq

def find_be_spacers(cds_sequence, cds_protein_start, target_aa, target_base='C', editor='BE4max'):
    '''Find sgRNAs that place target_base in editor-specific window at target_aa.
    Returns spacers with bystander annotation.

    Args:
        cds_sequence: nucleotide CDS (translated frame 1)
        cds_protein_start: amino acid number of CDS start (usually 1)
        target_aa: amino acid number to install variant (e.g., 130 for residue 130)
        target_base: 'C' (CBE) or 'A' (ABE)
        editor: 'BE3', 'BE4max', 'eA3A-BE3', 'ABE7.10', 'ABE8.20', 'ABE8e', 'evoCDA-BE'

    Returns: DataFrame with spacer, position-in-cds, target-base-position-in-spacer,
             bystander_positions, predicted_aa_changes
    '''
    # Editor-specific editing window (positions from PAM-distal end of spacer)
    window_by_editor = {
        'BE3': (4, 8),       'BE4max': (4, 8),    'eA3A-BE3': (5, 7),
        'ABE7.10': (4, 7),   'ABE8.20': (4, 8),   'ABE8e': (4, 8),     # SpABE8e matches CBE window (Richter 2020)
        'evoCDA-BE': (1, 9),
    }
    window_lo, window_hi = window_by_editor[editor]
    aa_index = target_aa - cds_protein_start  # 0-indexed in protein
    aa_start_nt = aa_index * 3                # nt offset in cds
    candidates = []
    spacer_len = 20
    pam_pattern = re.compile(r'(?=([ACGT]GG))')
    for strand, seq in [('+', cds_sequence), ('-', str(Seq(cds_sequence).reverse_complement()))]:
        for pam_match in pam_pattern.finditer(seq):
            pam_pos = pam_match.start()
            spacer_start = pam_pos - spacer_len
            if spacer_start < 0:
                continue
            spacer = seq[spacer_start:pam_pos]
            # Editor-specific window from PAM-distal end (1-indexed)
            # Find all editable bases in window
            edit_bases_in_window = []
            for i, b in enumerate(spacer[window_lo-1:window_hi], start=window_lo):
                if b == target_base:
                    edit_bases_in_window.append(i)
            if not edit_bases_in_window:
                continue
            # Annotate which edits hit the target_aa codon
            target_codon_start = aa_start_nt
            target_codon_end = target_codon_start + 3
            target_position_in_spacer = []
            for i in edit_bases_in_window:
                genomic_pos = spacer_start + i - 1
                if target_codon_start <= genomic_pos < target_codon_end:
                    target_position_in_spacer.append(i)
            bystander_positions = [i for i in edit_bases_in_window if i not in target_position_in_spacer]
            candidates.append({
                'spacer': spacer,
                'strand': strand,
                'spacer_start': spacer_start,
                'target_positions': target_position_in_spacer,
                'bystander_positions': bystander_positions,
                'n_bystanders': len(bystander_positions),
            })
    return pd.DataFrame(candidates).sort_values('n_bystanders')

Decision rule: Select spacers with target_positions != empty AND n_bystanders minimized. For variant-by-variant scanning, accept up to 1-2 bystanders if biology of those positions is interpretable; flag for downstream variant attribution.

Editing Efficiency Filtering (Critical Pre-Hit-Calling)

Goal: Drop sgRNAs that do not edit efficiently, since unedited reads represent no biological perturbation.

Approach: From CRISPResso2 output, compute target-base-conversion percentage per sgRNA; filter library to sgRNAs with >50% target editing in a pilot or co-screened control.

def filter_by_editing_efficiency(crispresso_outputs_dir, target_pos, target_base, efficiency_threshold=0.5):
    '''Drop sgRNAs that edit <efficiency_threshold of reads at target position.
    crispresso_outputs_dir: directory containing CRISPResso per-sample outputs.'''
    from pathlib import Path
    results = []
    for sample_dir in Path(crispresso_outputs_dir).glob('CRISPResso_on_*'):
        sgrna_id = sample_dir.name.replace('CRISPResso_on_', '')
        quant_file = sample_dir / 'Quantification_window_nucleotide_percentage_table.txt'
        if not quant_file.exists():
            continue
        df = pd.read_csv(quant_file, sep='\t')
        # Find target position in the quantification window
        target_row = df[df['Position'] == target_pos]
        if target_row.empty:
            continue
        # Editing = sum of non-original bases at target position
        original_pct = target_row[target_base].values[0]
        editing_pct = (100 - original_pct) / 100
        results.append({'sgrna_id': sgrna_id, 'editing_pct': editing_pct,
                         'pass_filter': editing_pct >= efficiency_threshold})
    return pd.DataFrame(results)

Convention: Drop sgRNAs below 50% editing for variant-function screens. A common working split is a 30% editing floor for primary screening and a 50% floor for confirmed hits. Below 30%, the screen has insufficient power; above 70%, results approach saturation editing.

Bystander Edit Attribution

Why this matters: When a sgRNA's editing window contains the target base AND a bystander base, the screen scores the combination. To attribute screen signal to the target variant alone, either (a) include sgRNAs that edit only the target (no bystander) -- often impossible -- or (b) deconvolute via parallel measurements.

Strategies for variant-by-variant attribution:

1. Tile multiple sgRNAs with different bystander patterns: If 5 different sgRNAs all hit the target base but have different bystanders, common signal across them is target-attributable (Hanna 2021 approach).

2. Use orthogonal chemistry: Run the same variant scan with prime editor (no bystanders); cross-validate. See [[prime-editing-screens]].

3. Bystander stratification: From CRISPResso2 allele table, partition reads by exact edit pattern (target only, target+bystander_1, target+bystander_2, etc.); separately score each pattern's contribution to the phenotype.

4. Restrict library: Use only sgRNAs with zero bystanders in the editing window (rare; may exclude most candidate spacers).

def deconvolute_bystander(allele_table_path, target_pos, bystander_pos_list):
    '''From CRISPResso2 allele table, partition reads by edit pattern at target + bystanders.
    Returns: per-pattern frequency for each combination of target/bystander edits.'''
    alleles = pd.read_csv(allele_table_path, sep='\t', compression='zip')
    # Mark target_edited and per-bystander_edited
    alleles['target_edited'] = alleles['Aligned_Sequence'].str[target_pos-1] != alleles['Reference_Sequence'].str[target_pos-1]
    for bp in bystander_pos_list:
        alleles[f'bystander_{bp}_edited'] = alleles['Aligned_Sequence'].str[bp-1] != alleles['Reference_Sequence'].str[bp-1]
    return alleles.groupby(['target_edited'] + [f'bystander_{bp}_edited' for bp in bystander_pos_list])['Reference_pct'].sum().reset_index()

Hit Calling for Variant-Function Screens

Goal: Score per-variant fitness from a base-editor screen.

Approach: Filter library to efficiency-passing sgRNAs (>50% editing), then run MAGeCK MLE or drugZ on the sgRNA-level counts; map each significant sgRNA to its predicted variant + bystander pattern; aggregate to per-variant scores.

def aggregate_variant_scores(mageck_sgrna_summary, variant_annotation_df):
    '''Aggregate sgRNA-level scores to per-variant scores.
    variant_annotation_df: per-sgRNA -> predicted variants (target + bystanders).'''
    df = mageck_sgrna_summary.merge(variant_annotation_df, on='sgRNA')
    # Target-only contribution: sgRNAs with no bystanders
    target_only = df[df['n_bystanders'] == 0]
    target_only_scores = target_only.groupby('target_variant')['LFC'].agg(['mean', 'std', 'count'])
    # Mixed signal: sgRNAs with bystanders
    mixed = df[df['n_bystanders'] > 0]
    return target_only_scores, mixed

Hanna 2021 BRCA1/2 Variant-Function Screen Methodology

Hanna et al 2021 Cell 184:1064 benchmarked CBE variant scanning at scale, screening 68,526 sgRNAs covering 52,034 ClinVar variants across 3,584 genes, with BRCA1 and BRCA2 as the positive/negative-selection benchmark:

1. Design the CBE library from predicted variant impact (ClinVar annotation), covering each variant with the sgRNAs that install it
2. Run drug-modifier screens (PARPi sensitivity) with vehicle vs drug
3. Score per variant by aggregating over all sgRNAs that install it; cross-check against bystander-controlled sgRNAs

Standard surrounding practice: verify editing efficiency at a control timepoint via amplicon sequencing, drop low-efficiency sgRNAs (see the editing-efficiency convention above), and call sensitizers with a bidirectional method such as drugZ.

Quantified result: Recovered known loss-of-function variants in BRCA1 and BRCA2 with high precision, and identified PARP1 variants conferring resistance to PARP inhibitors.

Cuella-Martin 2021 DDR-Gene Variant Screening

Cuella-Martin et al 2021 Cell 184:1081-1097 screened ~86 DNA-damage-response (DDR) genes (including BRCA1/2) with CBE saturation mutagenesis:

• Saturation CBE design across 86 DDR genes (not BRCA1/2 alone)
• Identified pathogenic/likely-pathogenic variants in critical protein domains
• Combined with biochemical and genetic validation (for example the 53BP1-USP28 interaction surface)
• Demonstrated saturation mutagenesis is feasible at protein-domain scale

Relationship to Hanna 2021: the two studies appeared back-to-back in the same Cell issue and apply the same CBE variant-scanning strategy to complementary targets -- Hanna benchmarks against ClinVar-annotated variants genome-wide, Cuella-Martin saturates 86 DDR genes. Treat them as complementary methodology references, not as cross-validations of each other.

Cas9 vs Base Editor vs Prime Editor for Variant Installation

| Approach | What it does | Bystander | Indels | When to use | |----------|--------------|-----------|--------|-------------| | Cas9 + HDR template | Installs precise edit + template | None | High (NHEJ competition) | When precise edit needed; high indel byproduct | | Cas9 (no template) | Random indels at cut site | None | 70%+ | Loss-of-function; not variant-specific | | CBE (BE3/BE4) | C->T at editing window | Yes (multiple Cs) | <5% | C->T variants with manageable bystanders | | ABE (ABE7.10/ABE8e) | A->G at editing window | Yes (multiple As) | <2% | A->G variants; clean for single-A spacers | | CGBE / GBE | C->G or C->A | Yes | 5-10% | Transversions; rare use cases | | Prime editor (PE2/PE3) | Templated edit; any base change | None | 1-3% | Precise variants; lower efficiency |

Decision: For C->T or A->G with available editing window: base editor is preferred (higher efficiency than PE). For other transitions/transversions, multi-base edits, or zero-bystander requirements: prime editor.

Broad be-validation-pipeline

The Broad Institute's be-validation-pipeline (https://broadinstitute.github.io/be-validation-pipeline/) is a CRISPResso2 post-processing and validation toolkit for BE amplicon data -- a set of Jupyter notebooks, not a workflow-engine pipeline. Run CRISPResso2 first, then execute the notebooks in order:

git clone https://github.com/broadinstitute/be-validation-pipeline
cd be-validation-pipeline
pip install -r requirements.txt

# Step 1: run CRISPResso2 in batch mode (or use the BEV tool on GPP LIMS).
# The batch file is tab-delimited with columns: name, fastq_r1, amplicon_seq, guide_seq
# (plus optional -w, -wc, --exclude_bp_from_left/right).
docker run -v ${PWD}:/DATA -w /DATA -i pinellolab/crispresso2 \
    CRISPRessoBatch --batch_settings batch_file.txt --skip_failed --base_edit

# Step 2: run the notebooks in order against the CRISPResso2 output
#   notebooks/01_BEV_allele_frequencies.ipynb
#   notebooks/02_BEV_nucleotide_percentage_plots.ipynb
#   notebooks/03_BEV_editing_efficiency.ipynb
# Outputs: allele-frequency tables, nucleotide-percentage plots, editing-efficiency heat maps

The notebooks cover allele-frequency tabulation, nucleotide-level editing quantification and editing-efficiency summaries. Hit calling is NOT part of this toolkit -- score the screen separately with drugZ or MAGeCK.

Failure Modes

Mostly indels in BE sample

Trigger: Cas9 contamination, wrong vector (e.g., used pCas9-BE3 plasmid but selected on Cas9 line), or evoCDA-BE / broader-window chemistry. Mechanism: Cas9 cuts dsDNA; BE relies on nicked-ssDNA deamination. Cas9 expression in the same cell creates indels. Symptom: Substitution-vs-indel ratio <3 in CRISPResso output. Fix: Verify vector (nCas9-BE3 not Cas9-BE3); confirm cell line lacks Cas9 background; restrict to specifically engineered BE-cell lines.

High editing but no biological signal

Trigger: Bystander C/A is dominating; intended variant is not the perturbation driving phenotype. Mechanism: When target is at position 5 and bystander is at position 7, the molecule carries both; phenotype is from the bystander. Symptom: Strong screen signal but variant attribution unclear. Fix: Run orthogonal prime-editor scan of the same intended variants; restrict library to bystander-free spacers when possible; deconvolute via allele-frequency table.

sgRNA shows perfect editing but no fitness signal

Trigger: Intended variant is silent or compensatory; the protein function is unchanged. Mechanism: Variants can be tolerated; not all variants are LoF or GoF. Symptom: High editing efficiency (>70%) but per-sgRNA LFC near zero. Fix: Expected outcome for many variants; flag silent / compensatory variants in the report.

Low editing across all guides

Trigger: Wrong cell line for the BE; cell line has poor BE activity (some lines lack APOBEC or have low expression). Mechanism: BE efficiency depends on cell-line expression of TadA or APOBEC components. Symptom: Median editing <30% across library. Fix: Test in a BE-validated cell line (HEK293T, U2OS, K562 generally work); pilot before full screen.

Library missing intended-variant sgRNAs

Trigger: No NGG-adjacent spacer places target base in editing window for that codon. Mechanism: Editor window is fixed; some codons cannot be targeted with given chemistry. Symptom: Specific variants absent from screen. Fix: Use PAM-relaxed BE variants (SpRY-CBE, SpRY-ABE); use prime editor for variants outside BE accessibility; accept that some variants cannot be installed.

Quantitative Thresholds

| Threshold | Value | Source / Rationale | |-----------|-------|--------------------| | Editing window | Positions 4-8 from PAM-distal end (BE3/BE4); 4-7 (ABE7.10); 4-8 (SpABE8e) | Komor 2016; Gaudelli 2017; Richter 2020 | | Editing efficiency for screen power | >30% (primary); >50% (validation) | Field convention (BE variant screens) | | Indel byproduct (clean BE) | <5%; <2% for ABE | Koblan 2018 (BE4max); Gaudelli 2017 (ABE) | | Substitution-vs-indel ratio | >10 (clean BE); <3 (Cas9-like) | CRISPResso2 diagnostic | | Bystander rate (target attribution) | <10% acceptable; <5% ideal for clean attribution | Application-dependent | | Cell-line BE activity (pilot) | >30% editing at validated target | Below = wrong cell line for BE | | Per-amino-acid sgRNA density | 10-15 (saturation designs); 5-8 (smaller screens) | Tradeoff with library size |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | Substitution-vs-indel ratio <3 | Cas9 contamination or wrong BE | Verify vector / cell line; pilot first | | All edits at bystander positions | Target base outside window | Re-design spacer with target at pos 4-8 | | Variant attribution unclear | Bystander confounding | Run orthogonal PE; restrict library | | Library lacks intended variant | No NGG-PAM accessibility | SpRY-CBE; prime editor; accept exclusion | | Editing <30% library-wide | Cell-line BE inactivity | Re-validate cell line | | Hit list dominated by single sgRNA | Bystander-driven phenotype | Cross-check with bystander-free sgRNAs |

References

• Komor AC et al. 2016. Nature 533:420. BE3.
• Gaudelli NM et al. 2017. Nature 551:464. ABE7.10.
• Koblan LW et al. 2018. Nat Biotechnol 36:843. BE4max + improved CBE.
• Richter MF et al. 2020. Nat Biotechnol 38:883. ABE8e; phage-assisted evolution of ABE7.10.
• Lapinaite A et al. 2020. Science 369:566. ABE8e mechanism.
• Gaudelli NM et al. 2020. Nat Biotechnol 38:892. ABE8 series (ABE8.20).
• Hanna RE et al. 2021. Cell 184:1064. Massively parallel BRCA1/2 variant function via CBE.
• Cuella-Martin R et al. 2021. Cell 184:1081-1097. CBE saturation across 86 DDR genes (BRCA1/2 plus others).
• Arbab M et al. 2020. Cell 182:463. BE-Hive prediction of editing outcomes.
• Anzalone AV et al. 2019. Nature 576:149. Prime editing (PE2/PE3).
• Clement K et al. 2019. Nat Biotechnol 37:224. CRISPResso2.
• Kurt IC et al. 2021. Nat Biotechnol 39:41. CGBE1.

Related Skills

• crispr-screens/crispresso-editing - CRISPResso2 BE/PE mode and allele tables
• crispr-screens/library-design - base-editor library design
• crispr-screens/prime-editing-screens - Orthogonal PE for variant attribution
• crispr-screens/hit-calling - Variant-level hit aggregation
• crispr-screens/screen-qc - Editing-efficiency QC
• crispr-screens/drugz-chemogenomic - drugZ for BE drug-modifier screens
• clinical-databases/clinvar-lookup - Variant pathogenicity annotation
• variant-calling/variant-annotation - VEP for predicted amino acid changes