Systematic profiling of PE6 variants establishes generalizable principles for efficient and specific prime editing

Technical Summary: Systematic Profiling of PE6 Variants for Prime Editing Optimization

Problem Statement
Prime editing offers a versatile platform for precise genome engineering without inducing double-strand breaks or requiring donor DNA templates. However, the field lacks generalizable principles for selecting optimal prime editor variants. While the PE6 family (comprising variants a through g) and related systems (PE2max, PE2max-ΔRNaseH) have demonstrated robust activity, prior studies were limited by small-scale evaluations of pegRNA sets. Consequently, the factors determining when specific PE6 variants outperform others—particularly under therapeutically relevant pegRNA designs—remain poorly defined. This gap forces researchers to rely on empirical testing rather than data-driven design principles for therapeutic genome editing.

Methodology
The authors employed a massively parallel, high-throughput screening approach to systematically compare nine prime editors: PE2max, PE2max-ΔRNaseH, and seven PE6 variants (PE6a–g).

• Library Design: A comprehensive library of 18,297 unique pegRNA–target sequence pairs was constructed. This included:
  • 5,815 pairs modeling 1–3 nt mutations from ClinVar.
  • 5,233 pairs designed for insertions longer than 3 nt.
  • 2,058 pairs with systematic single-nucleotide mismatches to profile tolerance.
  • Additional controls and random-mismatch sets.
  • The library utilized engineered pegRNA (epegRNA) scaffolds, randomly assigning 3' motifs (tevopreQ1 or tmpknot) for on-target analysis.
• Experimental Setup: HEK293T cell lines were generated to stably express each of the nine prime editors at a low multiplicity of infection (MOI) to ensure single-copy integration. These cells were transduced with the pairwise library at an MOI of 0.5. Editing efficiencies were quantified via deep sequencing at 7 and 14 days post-transduction.
• Computational Modeling: The authors developed DeepPrime6, a deep learning model to predict editing activity. Three training strategies were compared: (i) zero-shot application of a pre-trained DeepPrime model, (ii) training from scratch on the new PE6 datasets, and (iii) fine-tuning the pre-trained DeepPrime model with the PE6 data.
• Enhancement Strategies: The study evaluated the efficacy of specific 3' RNA motifs (tmpknot vs. tevopreQ1) and the fusion of the RNA-binding La protein to PE variants (generating PE7, PE7b, and PE7c) to stabilize the reverse transcriptase–pegRNA interaction.

Key Results

• Variant Performance: Across 12,481 pegRNAs, PE6b (average efficiency 6.6%) and PE6c (6.1%) exhibited significantly higher activity than other variants, including PE6a (5.0%) and PE2max (2.4%). Notably, PE2max-ΔRNaseH (4.6%) showed substantially higher activity than full-length PE2max, a distinction not apparent in earlier limited studies.
• Generalizability: PE6b and PE6c demonstrated the highest or equivalent activity in 89.0% of pegRNAs with ≥5% editing efficiency. PE6b was the dominant editor in 74.8% of these cases. Other variants (PE6d–g) showed advantages in less than 3% of cases, indicating their utility is restricted to specific sequence contexts.
• Mismatch Tolerance and Specificity: High-throughput profiling of position-dependent mismatches revealed that prime editors are highly sensitive to mismatches in the PBS and RTT regions. PE6b achieved the most favorable balance between general activity and mismatch tolerance, making it a robust, broadly applicable editor. PE6a exhibited the lowest mismatch tolerance, positioning it as a specificity-maximized alternative. All tested variants maintained indel frequencies below 0.05%.
• Motif and Fusion Optimization: Contrary to previous recommendations based on limited sets, the tmpknot motif consistently outperformed tevopreQ1 across most PE variants and edit types. Furthermore, fusing the La protein to PE6b and PE6c (creating PE7b and PE7c) yielded consistent, albeit modest, activity gains (1.1-fold and 1.4-fold, respectively).
• Predictive Modeling: The DeepPrime6 model, utilizing a fine-tuning strategy, achieved the highest predictive accuracy (Pearson's r = 0.66–0.71) for PE6 variants, outperforming both the zero-shot and scratch-trained models.

Significance and Claims
The study establishes practical, data-driven design principles for the selection and optimization of PE6-based prime editors. The authors conclude that PE6b serves as a robust standard editor for therapeutic applications, with PE6c as a comparable alternative, while PE6a is optimal for contexts requiring maximized specificity.

The authors assert that their findings provide a quantitative basis for therapeutic genome editing, moving the field away from empirical trial-and-error. They further suggest that the activity profiles and scaffold preferences identified here are likely shared by next-generation editors (such as the recently reported PE8 variants), as these are derived from similar catalytic cores. Finally, the development of the DeepPrime6 model offers a computational framework to guide variant selection and optimize prime editing strategies for diverse therapeutic targets.