Predicting adenine base editing efficiencies in different cellular contexts by deep learning

Area of Science:

• Gene editing technologies
• Molecular biology
• Bioinformatics

Background:

• Adenine base editors (ABEs) facilitate A•T to G•C conversions.
• Predictive models for base editing efficiency are limited by in vitro data.
• In vivo predictive power for primary cells remains uncertain.

Purpose of the Study:

• To evaluate adenine base editing efficiency in vitro and in vivo.
• To develop a predictive computational model for base editing outcomes.
• To assess the potential of ABEs for correcting pathogenic mutations.

Main Methods:

• Conducted base editing screens using SpRY-ABEmax and SpRY-ABE8e.
• Targeted 2,195 pathogenic mutations across cell lines and murine liver models.
• Developed BEDICT2.0, a deep learning model for predicting editing efficiencies.

Main Results:

• Observed strong correlations between in vitro and in vivo base editing datasets (Spearman R = 0.83-0.92).
• BEDICT2.0 accurately predicted adenine base editing efficiencies in cell lines (R = 0.60-0.94) and liver (R = 0.62-0.81).
• Demonstrated high on-target editing with minimal bystander effects.

Conclusions:

• Adenine base editing shows significant potential for correcting numerous pathogenic mutations.
• BEDICT2.0 is a robust computational tool for optimizing sgRNA-ABE combinations.
• The findings support the use of ABEs for in vitro and in vivo therapeutic applications.