Researchers have developed an adaptive computational pipeline called EvoMax to engineer compact eukaryotic Fanzor2 genome editors using sparse experimental datasets, according to a study published in Nature Biotechnology. The top-performing variant, named FanzMAX v3-hLa, achieved up to 97% editing efficiency at its most effective endogenous locus and averaged approximately 33% efficiency across 19 endogenous human targets. The system outperformed existing compact genome editors enNlovFz2 and enCnCas12f1 by more than 2.6-fold.
The researchers targeted Fanzor2 (Fz2) nucleases because the enzymes measure under 500 amino acids in length, making them compact enough for single adeno-associated virus (AAV) packaging. The team computationally filtered more than 1,600 Fz2-like sequences down to 332 eukaryotic orthologs for taxonomic and phylogenetic profiling. In vitro cleavage assays identified functional activity in an ortholog from Naegleria lovaniensis (NaloFz2) and an ortholog designated M7, both preferring a 5′-CCG-3′ target-adjacent motif. Sequence analysis demonstrated that deleting 20 to 30 residues from the N-terminal domain eliminated DNA cleavage in HEK293T reporter assays, showing that the intact domain is required for ribonucleoprotein assembly.
Guide RNA and fusion design
To stabilize the system, researchers rebuilt the wild-type 120-nucleotide guide RNA (ωRNA). Structural modeling using AlphaFold 3 and RNAfold led the team to truncate stem loop 3 and insert GAAA tetraloops, yielding 92-nucleotide variants called enωRNA v1 and enωRNA v2. The enωRNA v2 scaffold increased editing activity in M7 by 8.7-fold and restored editing in an inactive ortholog, M9, when combined with the NaloFz2 N-terminal domain. Structural predictions also revealed that the compact enzyme leaves the 3′ spacer RNA exposed to exonucleases. The authors addressed this by fusing human La protein (hLa) to the C-terminus of NaloFz2, which produced the highest editing levels under stringent screening at the CXCR4 site 2 locus.
Model-guided evolution
EvoMax identified beneficial protein substitutions without requiring large experimental training datasets. The authors trained a Gaussian process regression model on 209 single-point mutations from NlovFz2, recording an R2 score of 0.693 and a root-mean-squared error of 0.232 using BLOSUM62 matrix encoding. The pipeline shortlisted variants by combining regression predictions with scores from the 650-million-parameter ESM-2 protein language model, then reranked candidates using the ESM-IF inverse folding model against predicted structural backbones.
The team tested 10 to 20 candidates per round across three iterative engineering cycles, shifting model weights from empirical regression in round one toward evolutionary and structural predictions in rounds two and three. EvoMax produced an 84% hit rate for functional variants in round one, compared to 35% for AiCE and 20% for EVOLVEpro on the same starting background. The final variant, FanzMAX v3, improved reporter activation more than 13-fold over wild-type NaloFz2 when paired with enωRNA v2. To evaluate therapeutic utility, researchers packaged FanzMAX v3 into a single-AAV vector and delivered it to humanized mice, successfully editing the hPCSK9 gene in vivo.
