Scientists use generative AI to build better proteins for editing DNA

Researchers searched 31,565 eukaryotic genomes and identified over 116,000 potential PiggyBac transposase sequences, then tested 23 natural candidates in human cells, finding nine with detectable activity and two comparable to a previously optimized version, HyPB. They then fine-tuned the protein language model ProGen2 on over 13,000 bioprospected sequences, generating more than 100,000 synthetic variants of which 22 were tested; all showed excision activity and seven significantly outperformed HyPB. One variant, sequence 3277, doubled targeted DNA insertion when incorporated into the FiCAT gene-editing system in cultured cells. In primary human T cells, different proteins performed best at different tasks, with Poetur and sequence 136 achieving higher non-targeted integration than HyPB while sequence 3277 performed similarly to HyPB. The authors caution that improved activity does not guarantee better precision or safety, and that specificity and therapeutic effectiveness require further investigation.

(Source：The Brighter Side of News)