Predicting prime editing efficiency and product purity by deep learning.

Area of Science:

• Molecular Biology
• Genomics
• Bioinformatics

Background:

• Prime editing is a powerful genome editing technology.
• Optimizing prime editing guide RNA (pegRNA) is crucial for high editing efficiency.
• Predicting pegRNA performance is challenging but essential for applications.

Purpose of the Study:

• To identify sequence features influencing prime editing outcomes.
• To develop a predictive model for pegRNA efficiency.
• To validate the model's performance in various settings.

Main Methods:

• Conducted a high-throughput screen of 92,423 pegRNAs across 13,349 human pathogenic mutations.
• Utilized an attention-based bidirectional recurrent neural network to train the PRIDICT model.
• Validated PRIDICT predictions on endogenous editing sites and an external dataset.

Main Results:

• Identified key sequence context features affecting prime editing.
• PRIDICT achieved high prediction accuracy (Spearman's R of 0.85 for intended edits, 0.78 for unintended edits).
• High PRIDICT scores correlated with significantly increased prime editing efficiency (12-fold in vitro, tenfold in vivo).

Conclusions:

• PRIDICT accurately predicts prime editing efficiency based on sequence context.
• The tool enhances prime editing outcomes, offering value for basic and translational research.
• PRIDICT facilitates optimization of pegRNAs for improved genome editing applications.