Background

Gene correction technologies aimed at treating diseases at their root are continuously advancing. First-generation CRISPR-Cas9 editors have demonstrated excellent performance but are limited in their delivery into human cells due to their large size. Adeno-Associated Virus (AAV) vectors are widely recognized as representative delivery vehicles for gene therapies, but they have limited capacity for packaging genetic material. Scientists have therefore turned their attention to Fanzor2 (Fz2), an ultra-small eukaryotic gene editor with fewer than 500 amino acids.

However, naturally occurring Fz2 exhibits very low activity in human cells. To elevate its performance to a therapeutic level, extensive structural modifications of the protein are required. Artificial Intelligence (AI) has increasingly been used in protein engineering as a tool to predict the effects of mutations. The challenge, however, is that newly discovered protein families like Fz2 lack sufficient experimental data for reliable AI predictions. This has led to a continuous demand for novel design strategies to efficiently optimize unknown proteins under data-scarce conditions.

Key Discovery

The research team developed EvoMax, a machine learning framework capable of selecting high-performance mutant proteins using only a small amount of data. EvoMax is an algorithm that integrates Gaussian Process Regression (GPR), a protein language model (PLM) named ESM-2 with 650 million parameters, and a protein structure inverse folding (IF) model called ESM-IF. The team utilized 209 existing mutation data points from a similar protein, NlovFz2, to train the model. As a result, the fine-tuned GPR model accurately predicted activity changes due to amino acid substitutions with an R2 value of 0.693 and an RMSE of 0.232.

Using EvoMax, the researchers optimized NaloFz2, a eukaryotic Fanzor2 derived from Naegleria lovaniensis, in three sequential rounds. In each round, the model's contribution was finely adjusted to flexibly expand the search space. Simultaneously, the structure of the guide omega RNA (ωRNA), a key determinant of editing performance, was improved. By removing unstable terminal loops and enhancing base-pairing strength, the team successfully reduced the RNA sequence from 120 to 92 nucleotides, creating an improved version called enωRNA v2. Additionally, to protect the easily degradable 3' end, the team fused the human La RNA-binding protein (hLa) to the C-terminus of the protein.

The final product, FanzMAX v3-hLa, achieved a maximum double-strand break (DSB) efficiency of 97% at specific gene targets in human cells. On average across 19 target genes, it demonstrated an editing efficiency of 33%. This performance is more than 2.6 times better than previously reported ultra-small gene editors such as enNlovFz2 and enCnCas12f1. Even previously inactive Fz2 variants like M2 and M3 regained editing capability after key regions were substituted using EvoMax.

Implications and Outlook

This study demonstrates that even novel proteins with sparse data can be rapidly optimized to a commercial level by combining structural information with AI prediction models. It also proves that synergistic effects are maximized when protein language models and physical structure prediction models work in complementary ways. The platform technology has the potential to be extended for the rapid improvement of various unknown gene scissors candidates, making it highly significant.

The team also confirmed the therapeutic potential of the editor in animal models. In a humanized mouse model, administration of an AAV 1.0 therapeutic targeting the human PCSK9 gene resulted in 25% editing of the target gene in the liver and a 34% reduction in serum PCSK9 protein levels. However, clear limitations were also observed. Mice receiving more potent AAV 2.0 and AAV 3.0 versions died within a short period due to immune responses and acute toxicity. To achieve safe and effective gene correction, follow-up research must focus not only on improving target efficiency but also on precisely controlling dosing and enhancing vector stability in vivo.