The space of possible protein changes is immense, while a biology laboratory can test only a tiny fraction of variants. The team therefore built EvoMax, an iterative workflow combining laboratory measurements with Gaussian-process regression, a protein language model and an inverse-folding model. Training began with only 209 point mutations from related Fanzor2 editors.

Across three rounds, the model selected 10–20 candidates at a time for validation. Together with engineering of the guide ωRNA, this produced the compact FanzMAX v3-hLa editor. At the best tested site it edited 97% of sequences in cultured cells; the mean across 19 sites was about 33%. The 97% figure is therefore not a universal efficiency for any gene.

The authors tested predicted high-risk off-target sites and measured less than 1% editing at most of them, although two sites for the CXCR4 target reached roughly 2.2% and 5%. They also delivered a single AAV vector against human PCSK9 in humanized mice, but this was a small preclinical experiment with three animals per group, not a human treatment.

If the approach transfers to other protein families, it could shorten the search for enzymes used in gene therapy, diagnostics and industrial biotechnology where huge mutation databases do not exist. A smaller editor is also easier to fit into one AAV carrier, potentially simplifying delivery of the complete editing system.

Clinical use would still require independent replication, broader unbiased off-target mapping, long-term safety and immune testing, and comparisons with established editors. Optimistically, further target-specific preclinical demonstrations could arrive within 2–4 years; a first limited clinical study might follow in roughly 5–8 years if those results are favourable. Broad deployment cannot yet be estimated.