Rapid customization of base editors via machine learning-powered combinatorial mutagenesis.

Area of Science:

• Molecular Biology
• Genetics
• Bioengineering

Background:

• Precise control over DNA base editor protein interactions with target DNA is crucial for accurate, bystander-free gene editing.
• Current base editors can exhibit off-target edits, particularly at purine motifs, limiting their therapeutic potential.

Purpose of the Study:

• To engineer base editors with enhanced precision and specificity by analyzing DNA motif preferences and protein interactions.
• To develop a predictive machine learning model for identifying functional base editor variants at scale.

Main Methods:

• Combinatorial mutagenesis was used to generate large libraries of evoAPOBEC1 and TadA variants.
• Machine learning models were trained on DNA motif preferences from profiled variants in human cells.
• A structure-based deep learning model was developed to predict functional TadA variants across a vast sequence space.

Main Results:

• Identified base editor variants with motif-specific activity, eliminating residual adenine editing in cytosine base editors.
• Engineered variants demonstrated superior performance in correcting disease-associated mutations, significantly reducing unintended edits and achieving undetectable bystander edits in 50% of cases.
• The deep learning model successfully predicted functional TadA variants with 63% accuracy across an immense number of theoretical variants without prior experimental data.

Conclusions:

• The combined approach of mutagenesis and machine learning effectively re-engineers base editors for improved precision and reduced off-target effects.
• These advancements offer a streamlined method for developing tailored base editors for specific genetic targets and therapeutic applications.
• The predictive deep learning model demonstrates the potential for rapid, data-efficient discovery of novel protein variants.