A few hundred labeled mutations — not millions — were enough for AI-guided evolution to turn a weak compact gene editor into one hitting 97% editing efficiency at its best endogenous locus and roughly 33% on average across 19 genomic sites. In a new Nature Biotechnology study (Wan, Gold, Vure et al., 2026), researchers combined homology discovery, ωRNA scaffold engineering and a sparse-data adaptive framework called EvoMax to optimize Fanzor2 (Fz2), a compact eukaryotic RNA-guided nuclease. The fitness rules learned from one protein even transferred across the family, reviving naturally inactive homologs. The catch: in mice, the most active versions proved toxic — a result that redefines what AI protein engineering should optimize for.
A Compact Editor Worth Engineering
Fz2 is interesting precisely because it is small: typically under 500 amino acids, far smaller than classic CRISPR nucleases, which gives it the potential to pack an entire editing system into a single AAV vector for in-vivo delivery. But natural Fz2 proteins are weak in mammalian cells, and there is almost no experimental mutation data for them — a typical "cold-start" family where most AI methods cannot even get going.
The team started by mining natural diversity: from more than 1,600 Fz2-like sequences they selected 332 eukaryotic homologs for phylogenetic analysis and functional screening, landing on NaloFz2 from Naegleria lovaniensis. A key structural insight emerged from sequence comparison: the catalytic RuvC and DNA-binding regions are conserved, but the ~100-amino-acid N-terminal domain (NTD) varies dramatically — deleting just 20–30 N-terminal amino acids eliminated editing activity, showing the NTD is essential for structure and ribonucleoprotein assembly, not merely nuclear localization.
Next came ωRNA engineering. Structure prediction showed the natural 120-nt ωRNA carries three stem-loops; engineering the SL3 region — replacing the distal loop with a GAAA tetraloop, truncating it, and strengthening base pairing — produced a compact 92-nt enωRNA v2 (about three-quarters of the natural length) with significantly higher activity. It boosted the related nuclease M7 8.7-fold and, combined with a NaloFz2 NTD transplant, revived the inactive M9 — evidence that optimized RNA scaffolds also transfer across homologs. Fusing human La protein (hLa) to the Fz2 C-terminus added another boost, completing the layered "nuclease + RNA + fusion" design.
EvoMax: Three Models, One Lean Loop
The core methodological contribution is how the team handled the data scarcity. Instead of training a large neural network on 209 labeled single-point mutations, EvoMax fuses three complementary signals:
- GPR (Gaussian process regression with BLOSUM62 encoding) learns local fitness patterns from the 209 measured mutations — the strongest single learner, with R² = 0.693;
- ESM-2 (650M parameters) supplies evolutionary priors learned from millions of natural protein sequences;
- ESM-IF (inverse folding) scores structural compatibility against the AlphaFold-predicted architecture, blocking mutations that would destroy the fold.
The three weights are not fixed: early rounds trust experimental data for local optimization, later rounds shift toward evolution and structure to explore farther sequence space. Each round, only 10–20 candidates are experimentally verified, with the best variant seeding the next round — a tight model–experiment–model loop rather than a one-shot prediction.
The hit rate is striking: in round one, 84% of EvoMax-recommended mutations were functional, versus 20% for EVOLVEpro and 35% for AiCE under the same starting background. Three rounds produced FanzMAX v1→v3, and v3 with enωRNA v2 lifted reporter activity more than 13-fold over native NaloFz2 with the native ωRNA. At the real CXCR4 locus, editing climbed from ~12% (native) to 66% (FanzMAX v3) to ~97% with the hLa fusion. Across 19 endogenous loci, FanzMAX v3-hLa averaged ~33% — about 10× the native system (~3.4%) and roughly 2.7× the best engineered compact editors (enNlovFz2 ~11%, enCnCas12f1 ~12.5%). It also widened TAM compatibility to more 5′-ANNG sequences (ACAG, ACTG, AATG), and at ~97% on-target editing at CXCR4, most candidate off-target loci stayed below 1%.
Fitness Rules That Travel: Reviving Dead Proteins
The authors then posed a harder question: can rules learned on NaloFz2 migrate to other Fz2s — even restoring naturally inactive ones? They transplanted the intact NTD into truncated or highly diverged dead homologs, applied enωRNA v2, and let EvoMax pick the most promising mutations, introducing only 2–5 per protein.
This revived detectable editing activity in multiple homologs, most notably M2, M5 and M8. A second round on M2 and M3 added mutations (G298A, I368L) with ~1.9× and ~2.5× further gains, and overall 55–70% of model-recommended variants outperformed their parents. The implication is methodological: fitness constraints are at least partly family-level, so a small, high-quality dataset can bootstrap an entire protein family — not just one sequence.
The In-Vivo Warning: Stronger Is Not Safer
Fz2's compact size enabled a true single-AAV system. In HuH-7 cells, iterative AAV 1.0→2.0→3.0 development reached ~50–68% editing efficiency. In mice, the 3.7-kb AAV 1.0 produced ~25% liver editing and ~34% reduction in circulating PCSK9 with no obvious toxicity. But the more efficient AAV 2.0 and 3.0 caused acute morbidity within 7–14 days of systemic administration, with more large genomic deletions — evidence that excessive nuclease activity and DNA damage are themselves toxic to hepatocytes.
This flips the goal of therapeutic editor optimization: the fitness function should not be "maximize editing efficiency" but a multi-objective window over activity, targeting range, repair outcomes and safety. For next-generation compact editors aimed at single-AAV in-vivo gene editing, this is the difference between a paper result and a drug.
What This Means for AI Protein Engineering
Three structural signals are worth separating from the headline numbers. First, sparse-data bootstrap is becoming viable: a few hundred labels plus a language-model prior plus a structure filter can navigate vast sequence space without deep mutation scanning, which changes the economics of engineering newly discovered protein families. Second, scaffold-level transfer turns one dataset into a platform — the same playbook that revived dead Fz2 homologs could be applied to whole families of enzymes and editors. Third, the objective function itself is being rewritten: as AI-generated editors move toward the clinic, activity alone is no longer the metric, and in-vivo toxicity must be part of the model loop. These trends sit alongside the broader shift in AI-for-science — from world models of biology to judging AI scientists by real discovery — where capability and evaluation are both maturing at once.
Actionable Playbook
For teams doing AI-guided protein or editor engineering, the study is effectively a template:
- Do not wait for big datasets. A few hundred high-quality measured mutations can seed a working loop (209 here, R² = 0.693 for the GPR component).
- Fuse three signal types: empirical regression (GPR), a protein language model prior (ESM-2), and a structural compatibility filter (ESM-IF). No single model carries the loop.
- Shift weights during iteration — trust experiments early, broaden with evolution and structure later; keep each round to 10–20 experimental validations.
- Engineer beyond the protein: RNA scaffolds and fusion domains (ωRNA v2, hLa) contributed as much as the sequence changes — optimize the whole complex, not just the residues.
- For therapeutic editors, redefine the objective: track on-target/off-target ratios, repair outcomes and in-vivo toxicity endpoints; treat "most active" as a hypothesis to be screened, not a target to be reached.
FAQ
How can AI evolve a protein with only a few hundred data points?
EvoMax fuses three complementary signals: GPR on 209 measured single-point mutations (R² = 0.693), ESM-2 evolutionary priors from millions of natural sequences, and ESM-IF structural compatibility checks. Each round only 10–20 candidates are tested; round-one hit rate was 84%, versus 20% for EVOLVEpro and 35% for AiCE. Three rounds yielded FanzMAX v3 with ~97% editing at CXCR4 and ~33% average across 19 endogenous loci.
Is the most active gene editor the best choice for therapy?
No — the paper's mouse data says the opposite. Single-AAV AAV 1.0 achieved ~25% liver editing and ~34% reduction in circulating PCSK9 with no obvious toxicity, while the more active AAV 2.0 and 3.0 caused acute morbidity within 7–14 days, accompanied by more large genomic deletions. Therapeutic editors must balance activity, specificity, repair outcomes and safety.
What is Fanzor2 and why does it matter compared with CRISPR?
Fanzor2 is a compact eukaryotic RNA-guided nuclease, typically under 500 amino acids — small enough to fit the whole editing system into one AAV vector. The engineered FanzMAX v3-hLa widened TAM compatibility to 5′-ANNG sequences (e.g. ACAG, ACTG, AATG) and, at ~97% on-target editing at CXCR4, kept most candidate off-target loci below 1%.
