Protein-Nucleic Acid Constrained Language Model Assisted Design of Precise

Adenine base editors (ABEs) enable efficient A-to-G base editing in genomic DNA, serving as powerful tools for gene therapy of 48% of heritable diseases. However, currently existing efficient ABEs (e.g. ABE8e) induce excessive bystanders, off-targets, and are large in size. Those limitations of ABEs hinder their clinical application for disease treatment. In this study, we leveraged a pre-trained Protein-Nucleic Acid Constrained Language Model to design TadA-8e variants with high activity, reduced editing window and decreased size. Among the top 20 truncated variants tested in HEK293T cells, 3 showed comparable editing efficiency to ABE8e, with only 147-152 aa deletion narrowing ABE8e's major editing window. By combining these truncations and linker deletions, the smallest ABE8e (with a 27% reduction in size), named PNLM-pcABE, was obtained. It exhibits high activity, precise 3nt editing window, and reduced DNA off-targets near background level. Moreover, PNLM-pcABE achieved up to 134-fold precision improvement in pathogenic mutation correction compared to ABE8e.