Adaptive model-guided protein evolution with sparse data optimizes compact eukaryotic genome editors

In plain English

The team used a small number of lab measurements to teach models which protein changes to try next, then repeated the cycle until the editor became much more active.

How the study worked

A plain-language walk through the work behind the result.

Screened more than 1,600 Fanzor2 candidates and experimentally tested a prioritized subset.
Combined Gaussian-process regression, the ESM-2 protein language model, and ESM-IF inverse folding to rank mutations from sparse experimental data.
Ran three engineering rounds, optimized the guide-RNA scaffold in parallel, and tested the resulting editor in mammalian cells and humanized mice.

What they found

FanzMAX v3-hLa reached up to 97% editing at its best endogenous target and averaged about 33% across 19 loci in bulk mammalian cells.
The optimized editor outperformed two established compact editors by more than 2.6-fold in the authors’ comparison.
A single-AAV experiment edited human PCSK9 in humanized mice.

Why it matters

The study shows how language-model and structure priors can narrow a large protein search space when experimental data are scarce.

The catch

The workflow does not explicitly model higher-order interactions among multiple mutations.
Performance has not yet been established across unrelated protein families.
The mouse experiment supports feasibility but does not establish clinical safety or efficacy.

Methods at a glance

Study type: Model-guided protein-engineering study with cell and mouse experiments
Biological system: Fanzor2 RNA-guided nucleases
Organism: Mammalian cells and humanized mice
Dataset: More than 1,600 Fanzor2 ortholog candidates plus iterative mutation screens
Sample size: More than 1,600 candidates screened computationally; 332 prioritized; selected orthologs and 10–20 candidates per engineering round tested
Model architecture: Transfer-learning Gaussian-process regression with ESM-2 and ESM-IF ranking channels
Training objective: Prioritize Fanzor2 mutations that improve genome-editing activity from sparse measurements
Code: Source code is available in the linked EvoMax GitHub repository.
Data: The article links source data and an archived software record.

Tasks and metrics

Tasks: Endogenous-locus editing; Target-adjacent-motif compatibility; In vivo hPCSK9 editing

Metrics: Editing efficiency; Mean activity across loci; Relative performance against compact-editor baselines

Baselines: enNlovFz2; enCnCas12f1

Principal findings

FanzMAX v3-hLa reached up to 97% editing at the best-performing endogenous locus and about 33% mean editing across 19 loci.
The authors report more than 2.6-fold higher activity than the selected compact-editor baselines.

What remains unresolved

External validation: The final system was tested at endogenous mammalian loci and in a humanized-mouse hPCSK9 experiment.

Single-site mutation ranking does not explicitly model higher-order epistasis.
Transfer to unrelated protein families remains untested.
In vivo results are preclinical.