Behind The Data: Leveraging Generative AI To Design Novel, Functional Deaminases For Adenine Base Editing​

Natural enzymes offer only a limited starting point for base editor development, while the scale of protein sequence space makes systematic experimental screening impractical. An AI-driven workflow combining de novo protein design, computational filtering, and experimental feedback provides a more efficient path to novel editing enzymes. Using 370,000 curated adenine deaminase sequences, researchers trained a generative model to produce diverse candidates beyond natural sequence space. Machine learning then identified mutations that progressively improved editing performance. After one round of guided engineering, the leading de novo candidate increased on-target A-to-G editing from less than 2% to more than 20%, while maintaining bystander editing below 1%. Testing across therapeutically relevant gene targets further demonstrated a favorable, generalizable precision profile.

Uncover how this end-to-end approach could expand base editing capabilities and support new therapeutic strategies.