Machine Learning and Directed Evolution of Base Editing Enzymes

As we enter the era of CRISPR medicines, base editors (BEs) emerged as one of the most promising tools to treat genetic associated diseases. However, unintended bystander editing beyond the target nucleotide poses a challenge to their translation into effective therapies. While many efforts have been made in the design of a universal enzyme with minimal bystander editing, the context dependent activity represents a major challenge for base editing-based therapies. In this work, we designed a sequence-specific guide RNA library with 3-extensions and detected guides that were able to reduce bystander and increase editing efficiency in a context dependent manner. The best candidate was later used for phage assisted non-continuous evolution to find a new generation of precise base editors. Simultaneously, we use protein language models trained on massive protein sequence datasets to find the evolutionarily plausible mutational patterns that can improve deaminase activity and precision. Both strategies provide a collection of precise TadA variants that not only drastically reduced bystander edits, but also was not in detriment of on-target activity. Our findings introduce a guide/enzyme parallel engineering pipeline, which lays the foundation for the development of new personalized genome editing strategies, ultimately enhancing the safety and precision of this groundbreaking technology.