GPTomics/bio-crispr-screens-prime-editing-screens

Version Compatibility

Reference examples tested with: PRIDICT2 v1.0+ (https://github.com/uzh-dqbm-cmi/PRIDICT2), CRISPResso2 2.2.14+, pandas 2.2+, biopython 1.83+, numpy 1.26+.

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

• CLI: python pridict2_pegRNA_design.py single --help; python pridict2_pegRNA_design.py batch --help
• Web: PRIDICT2 web interface at https://pridict.it/

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

Prime-Editing Screen Analysis

"Design or analyze a pooled prime-editor screen" -> Design pegRNAs (spacer + scaffold + PBS + RTT) for intended edits, predict efficiency with PRIDICT2, filter pre-synthesis to efficient candidates, install variants in the screen, quantify intended-edit vs scaffold-incorporation vs indel via CRISPResso2, and aggregate to per-variant fitness scores.

• Python: PRIDICT2 for pegRNA efficiency prediction
• Python: ePRIDICT for chromatin-context prediction; pair with PRIDICT2 rather than replacing it
• CLI: CRISPResso --prime_editing_pegRNA_* for amplicon-level analysis
• Workflow: pegRNA library design -> PRIDICT2 filtering -> screen execution -> CRISPResso2 quantification -> per-variant scoring

Prime Editor Chemistry Comparison

| Editor | Year | Mechanism | Indel rate | Use when | |--------|------|-----------|------------|----------| | PE2 (Anzalone 2019) | 2019 | nCas9-RT fusion + pegRNA | 1-3% | Standard PE; lowest indel rate | | PE3 | 2019 | PE2 + nick of opposite strand by additional sgRNA | 2-5% | Higher editing efficiency, slightly more indels | | PE3b | 2019 | PE3 with edit-blocking ssgRNA | 1-3% | When PE3's added nick risks unwanted indels | | PEmax (Chen 2021) | 2021 | Engineered RT + nCas9 | 1-2% | Higher editing rate per pegRNA | | PE5max (Chen 2021) | 2021 | PE3 plus MMR inhibition (MLH1dn) on the PEmax architecture | 1% | Highest efficiency at favorable sites | | PE6 / dual-pegRNA (2023) | 2023 | Engineered compact PE; twin-pegRNA systems | Variable | Specific applications |

Decision rule: For pooled screens at scale, PE2 or PEmax (single-guide architecture) is preferred over PE3, whose additional nicking sgRNA complicates library architecture. For specific high-efficiency edits, PEmax + PRIDICT2-optimized pegRNA.

pegRNA Architecture

A pegRNA contains four critical elements that determine efficiency:

5'  SPACER (20 nt)  -- standard sgRNA spacer; defines target locus via NGG PAM
    +
    SCAFFOLD (~80 nt) -- canonical or recoded scaffold (Chen 2021 recodes it to cut scaffold-incorporation byproducts)
    +
    PBS (Primer Binding Site, 8-15 nt) -- complements protospacer downstream of cut site
    +
    RTT (Reverse Transcription Template, 10-30 nt) -- encodes intended edit; copied by RT
3'

Key design parameters:

• PBS length: 11-13 nt typical; longer for high-GC contexts; PBS GC fraction critical (35-65% target)
• RTT length: 10-20 nt typical; longer for distant edits (10+ bp away from cut)
• RTT-edit position: intended edit at position 4-30 from cut site
• Scaffold: standard sgRNA scaffold OR the Chen 2021 recoded scaffold, which removes homology with the genomic target and cuts scaffold-derived byproducts. The recoding does not itself raise editing efficiency; that comes from MLH1dn and the PEmax architecture.

PRIDICT and PRIDICT2 pegRNA Efficiency Prediction

Mathis N et al 2023 Nat Biotechnol 41:1151 (PRIDICT v1) / 2025 Nat Biotechnol 43(5):712 (PRIDICT2; published online June 2024) developed deep-learning predictors of per-pegRNA editing efficiency. PRIDICT2 is the current state of the art.

# PRIDICT2 is invoked via CLI: pridict2_pegRNA_design.py
# Single sequence input:
python pridict2_pegRNA_design.py single \
    --sequence-name BRCA1_c5135 \
    --sequence "AGCAGCCT(C/T)CTGAATGCCC...60nt_context" \    # parens = intended edit
    --output-dir predictions/ \
    --use_5folds                                              # 5-fold ensemble averaging

# Batch input from CSV:
python pridict2_pegRNA_design.py batch \
    --input-fname variants_to_design.csv \                    # CSV: sequence_name, sequence
    --output-dir predictions/ \
    --cores 4 \
    --summarize                                               # generate summary table

# Output: per-pegRNA predictions in predictions/<sequence_name>/
# Columns: PBS_sequence, PBS_length, RTT_sequence, RTT_length, predicted_editing_efficiency,
#          predicted_indel_rate, deep_ensemble_score, etc.

Loading PRIDICT2 results in Python:

import pandas as pd
from pathlib import Path

def load_pridict2_predictions(prediction_dir):
    '''Load PRIDICT2 batch outputs from prediction_dir/'''
    summary = pd.read_csv(Path(prediction_dir) / '<timestamp>_summary_K562_batch_summary.csv')
    # summary has columns: sequence_name, PBS, RTT, predicted_efficiency, predicted_indel, etc.
    return summary

Key determinants of PE efficiency (Mathis 2025 PRIDICT2):

| Feature | Effect on efficiency | |---------|----------------------| | PBS GC content | 40-55% optimal; high GC slows annealing | | PBS length | 11-13 nt optimal; longer for high-GC PBS | | RTT length | 10-20 nt typical; trade-off between coverage and processivity | | Edit position in RTT | Closest to PBS = highest efficiency | | Chromatin context | Dominant locus effect; H3K9me3 heterochromatin ~0.8% vs ~2.2% elsewhere | | Cell line / Cas9 expression | Variable; piloting required | | Cell cycle phase | S/G2 = higher efficiency |

Critical insight from Mathis 2025: Chromatin context is a major locus-level determinant that sequence-only predictors miss, which is why ePRIDICT is designed to be combined with PRIDICT2.0 rather than replace it -- the pairing helps most in regions of lower chromatin accessibility. For genome-scale screens, validate predictions empirically at representative loci.

PRIME Pooled Screen Methodology

Ren X et al 2023 Mol Cell 83:4633 established the PRIME pooled prime-editing screen methodology (earlier 2023 bioRxiv preprint):

• pegRNA library covering thousands of intended variants, screened for specificity at design time (Ren 2023 used GuideScan2; add PRIDICT2 efficiency prediction for new designs)
• Lentiviral delivery in a PE-expressing cell line (Ren 2023: MOI 0.3 for the MYC-enhancer screen, MOI 0.5 for the variant screens)
• Selection on integration marker
• Time-course screen for variant function (e.g., drug sensitivity)
• Endpoint amplicon sequencing of each pegRNA target locus
• CRISPResso2 quantification of intended-edit %
• MAGeCK / drugZ-style hit calling on edit-efficient pegRNAs

Quantified scale: ~3,699 ClinVar variants installed in a single PRIME screen, alongside 1,304 breast-cancer GWAS variants.

MOSAIC In Situ Saturation Mutagenesis

MOSAIC (Hsu 2024, bioRxiv) is a high-throughput in-situ saturation-mutagenesis prime-editing method with multiplexed read-out:

• Tile pegRNAs across protein domains for systematic mutagenesis
• Saturation: every possible amino acid change in a region
• Identify drug-resistance variants in real-time
• Smaller per-variant cell numbers (more variants total)

Use case: Cancer-drug-resistance variant scanning; protein-domain function mapping.

Run PRIDICT2 on a Custom pegRNA Library

Goal: Predict editing efficiency for thousands of pegRNAs before library synthesis.

Approach: Build a CSV with one row per intended edit (sequence + edit notation), run PRIDICT2 in batch mode, parse the per-pegRNA efficiency summary, and filter to candidates above the chosen efficiency threshold.

# Step 1: prepare batch input CSV (sequence_name, sequence with (REF/ALT) edit notation)
cat > variants.csv <<EOF
sequence_name,sequence
BRCA1_R71X,AGCAGCCT(C/T)CTGAATGCCC...
MLH1_c677,GAGCTGAGC(A/G)GAGGCTCTTGAAGC...
EOF

# Step 2: run PRIDICT2 batch
python pridict2_pegRNA_design.py batch \
    --input-fname variants.csv \
    --output-dir predictions/ \
    --cores 8 \
    --summarize

# Step 3: parse and filter
import pandas as pd
predictions = pd.read_csv('predictions/<timestamp>_summary_K562_batch_summary.csv')

# Filter to pegRNAs with predicted efficiency > 50% (library-inclusion convention)
filtered = predictions[predictions['predicted_editing_efficiency'] > 50]
print(f'pegRNAs passing PRIDICT2 >50%: {len(filtered)} / {len(predictions)}')

# Pick top 3 per intended edit
top3 = (filtered.sort_values(['sequence_name', 'predicted_editing_efficiency'],
                              ascending=[True, False])
                 .groupby('sequence_name').head(3))
top3.to_csv('peg_library_filtered.csv', index=False)

Cross-Validate PE with Base Editor Screens

Goal: Confirm variant-function calls from PE with orthogonal BE screens.

Approach: Design parallel BE library for the same variants; run both screens; intersect hits.

# BE screen output (target conversion + bystander)
be_hits = pd.read_csv('be_screen_hits.tsv', sep='\t')
# PE screen output (intended edit + scaffold-incorp + indel)
pe_hits = pd.read_csv('pe_screen_hits.tsv', sep='\t')

# Intersect on intended variant
concordant = be_hits.merge(pe_hits, on='variant_id', suffixes=('_be', '_pe'))
# Filter to high-confidence: both methods call variant + same direction
concordant['high_confidence'] = (concordant['be_fdr'] < 0.05) & (concordant['pe_fdr'] < 0.05) & \
                                 (np.sign(concordant['be_lfc']) == np.sign(concordant['pe_lfc']))

Critical: PE-only hits in BE-coverable variants are suspect (BE should detect them). PE-only hits in non-BE-coverable variants (e.g., transversions) are genuinely PE-unique.

CRISPResso2 for PE Quantification

CRISPResso \
    --fastq_r1 pe_sample.fq.gz \
    --amplicon_seq <amplicon_seq> \
    --guide_seq <20nt_spacer> \
    --prime_editing_pegRNA_spacer_seq <spacer> \
    --prime_editing_pegRNA_extension_seq <RTT+PBS> \
    --prime_editing_pegRNA_scaffold_seq <scaffold> \
    --quantification_window_size 25 \              # widen to cover edit
    --output_folder pe_results \
    --name sample_id

# Output: CRISPResso_quantification_of_editing_frequency.txt
# Prime-editing outcomes appear as extra amplicon ROWS (Reference / Prime-edited /
# Scaffold-incorporated), each with Unmodified%, Modified% and read counts.

Failure Modes

Low pegRNA efficiency despite high PRIDICT prediction

Trigger: Sequence-only prediction missed chromatin context. Mechanism: Closed chromatin reduces Cas9 binding and RT activity; PRIDICT2 only sees sequence. Symptom: PRIDICT2 predicts 60% efficiency; observed is 5%. Fix: Cross-reference target with chromatin accessibility data (ATAC-seq) in the cell line; flag pegRNAs at silenced loci; pilot before screen.

High scaffold incorporation

Trigger: RTT too short relative to PBS, or RT processivity issue. Mechanism: RT reads past edit into scaffold; resulting product is detectable but undesired. Symptom: Scaffold incorporation >5%; intended edit efficiency low. Fix: Re-design pegRNA with longer RTT; verify with PRIDICT2 score for scaffold_incorp; pilot at representative loci.

PE2 cell line lacks RT expression

Trigger: PE2 construct expressed at low level; insufficient RT for productive editing. Mechanism: PE2 requires high RT expression; some cell lines down-regulate. Symptom: Library-wide editing <10%; not locus-specific. Fix: Verify PE2 expression by Western blot; consider PEmax (higher activity); use better-validated cell lines (K562, HEK293T, U2OS).

Multi-base intended edit but only one base installed

Trigger: Long RTT designed for multi-base edit; RT prematurely terminates. Mechanism: RT processivity drops with longer RTT; multi-base edits often incomplete. Symptom: Allele table shows partial-edit alleles (some bases installed, not all). Fix: Re-design with shorter RTT covering only the closest edits; or use PE3 to nick opposite strand and force longer RT processivity.

Library missing intended variant

Trigger: No suitable PAM/PBS/RTT combination for the intended edit. Mechanism: PE requires NGG PAM within 30 nt of edit; rare edits cannot be installed. Symptom: Specific variants absent from library. Fix: Use SpRY-PE for relaxed PAM; accept that some variants cannot be PE-installed; consider BE if applicable.

Cas9 vs BE vs PE for Variant Installation

| Approach | Bystander | Indels | Coverage | When to use | |----------|-----------|--------|----------|-------------| | Cas9 + HDR | None | High | Variable (depends on template integration) | Precise edits at scale; high indel byproduct | | Base editor | YES | Low (<5%) | Limited by editing window | C->T or A->G at editable position | | Prime editor | NONE | Low (<3%) | NGG-PAM within 30 nt of edit | Precise variants; multi-base; transversions | | Cas9 (no template) | NONE | 70%+ | Anywhere with NGG | LoF only; not variant-specific |

Decision tree:

• C->T or A->G at editing-window position: BE (higher efficiency than PE)
• Multi-base / transversion / out-of-window: PE
• LoF without specifying variant: Cas9
• Random insertions: HDR (lower throughput than PE)

Quantitative Thresholds

| Threshold | Value | Source / Rationale | |-----------|-------|--------------------| | PRIDICT2 efficiency for library inclusion | >50% | Project-chosen cutoff; PRIDICT2 prescribes none | | Intended edit % for screen power | >5%; >20% at favorable sites | Field convention | | Scaffold incorporation | <2% (clean PE); <5% acceptable | Empirical | | Indel byproduct | <3% (PE2); <5% (PE3) | Anzalone 2019; Chen 2021 | | PBS GC content | 40-55% | PRIDICT2 | | PBS length | 11-13 nt | PRIDICT2 | | RTT length | 10-20 nt | PRIDICT2 | | Edit position from cut | 1-30 nt | Anzalone 2019 | | Cell line for PE | K562, HEK293T, U2OS validated | High RT expression |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | Low editing across library | Cell-line RT inactivity | Verify PE2 expression; switch to validated line | | Scaffold incorporation >10% | RTT too short | Re-design with longer RTT | | Partial multi-base edits | RT processivity limit | Shorter RTT or PE3 | | PRIDICT predicts but observes much lower | Chromatin context | Pilot at chromatin-aware sites | | Library missing variants | No NGG PAM | SpRY-PE; BE alternative | | PE concordant with BE on transitions, disagrees on transversions | PE handles transversions BE doesn't | Expected; trust PE |

References

• Anzalone AV et al. 2019. Nature 576:149. Original PE2/PE3 (foundational prime editing paper).
• Mathis N et al. 2023. Nat Biotechnol 41:1151. PRIDICT v1 deep-learning pegRNA prediction.
• Mathis N et al. 2025. Nat Biotechnol 43(5):712 (published online June 2024). PRIDICT2 + chromatin context (current state-of-the-art).
• Chen PJ et al. 2021. Cell 184:5635. PEmax + engineered RT.
• Hsu JY, Lam KC, Shih J, Pinello L, Joung JK 2024 bioRxiv (doi:10.1101/2024.04.25.591078). MOSAIC in situ saturation mutagenesis via prime editing.
• Ren X et al. 2023. Mol Cell 83:4633. PRIME pooled prime-editing screen (~3,699 ClinVar variants); variant-installation scale.

Related Skills

• crispr-screens/library-design - pegRNA library design
• crispr-screens/base-editing-analysis - Orthogonal BE for variant attribution
• crispr-screens/crispresso-editing - CRISPResso2 PE mode and quantification
• crispr-screens/hit-calling - Per-variant hit calling
• crispr-screens/screen-qc - Editing-efficiency QC
• variant-calling/variant-annotation - Annotate edited variants
• clinical-databases/clinvar-lookup - Variant pathogenicity