Science published results showing AI-designed SynTnpB nucleases stayed functional across bacterial, plant and human cells, with many variants matching or surpassing natural TnpB activity.
The team combined Meta's ESM-IF1 inverse-folding model with evolution-based residue constraints to tackle a long-standing design problem in multi-domain RNA-guided enzymes.
Designed binding regions were far more divergent than sequence-only model outputs—83% and 72% identity to the closest natural counterparts versus more than 99% for reference-like designs.
Cryo-EM and microscopy linked that performance to new electrostatic and hydrogen-bonding networks that stabilized RNA-DNA interactions across different conformations.
The work suggests AI can broaden the CRISPR toolbox beyond natural proteins, opening a path to custom RNA-guided nucleases and other conformationally active nucleic-acid binders.