TL;DR: Penn Engineers used a machine-learning method called EvoMax to build FanzMAX, an engineered Fanzor2 editor that reached 97% editing efficiency at its best genomic target.

University of Pennsylvania engineers have built a far more active compact gene editor by asking artificial intelligence to rank a small set of protein changes for laboratory testing. The result, described in a study published in Nature Biotechnology, is called FanzMAX, an engineered version of the RNA-guided DNA-cutting protein Fanzor2, according to Penn Engineering.

The team was led by Xue Sherry Gao, Presidential Penn Compact Associate Professor in Chemical and Biomolecular Engineering and in Bioengineering, with postdoctoral fellow Shijie Wan and other collaborators, according to Penn Engineering. Their machine-learning framework, named EvoMax, prioritizes candidate mutations using small amounts of experimental data. At its best-performing genomic target, FanzMAX achieved an editing efficiency as high as 97%.

That outcome addresses two linked problems in genome editing. Many editors are too large to move easily into cells through current delivery vehicles, and the compact alternatives have shown only modest activity in mammalian cells. The Penn group kept the small size of Fanzor2 while lifting activity, and the compact size allowed the team to design single-AAV editor systems, according to Penn Engineering.

A delivery bottleneck for powerful editors

Gene editing has been discussed as a way to treat disease at its genetic roots, including inherited genetic disorders, blood conditions and cancer, according to Penn Engineering. CRISPR tools, which allow scientists to alter DNA sequences and modify gene function with high precision, have become increasingly powerful.

A practical challenge remains around getting an editor inside a cell. Cell membranes prevent large molecules from simply entering, so a gene-editing system must use a carrier before it can make a desired genetic change, according to Penn Engineering.

Researchers can use lipid nanoparticles to deliver editor proteins or messenger RNA, or leverage engineered viruses to deliver the DNA encoding the editor, according to Penn Engineering. When the editor or its genetic instructions are larger, this delivery process becomes more difficult.

One possible solution described by the Penn team is to make editors smaller. Smaller instructions and smaller proteins place fewer demands on the carrier and leave more room for the other components needed to direct and control editing.

Why AAV capacity makes size matter

Adeno-associated viruses, or AAVs, are commonly used to deliver gene therapies, but they can carry only a limited amount of genetic material, according to Penn Engineering. That limit shapes how engineers assemble a therapy.

A compact gene-editing tool could help overcome this constraint by leaving more room for the other components needed to direct and control editing, according to Penn Engineering. A smaller editor leaves more space in a delivery vehicle for parts such as the RNA that guides the editor to the right genomic site.

Xue Sherry Gao explained the trade in similar terms in the Penn Engineering report:

“Fanzor2 is unusually small compared with many gene-editing proteins, and that matters because a smaller editor leaves more space in a delivery vehicle for the other components needed to make the system work, such as the RNA that guides the editor to the right genomic site.”

The compact size of Fanzor2 allowed the team to design single-AAV editor systems, according to Penn Engineering. The report notes, however, that small size alone is not enough, because an editor also needs to be active, accurate and safe.

What Fanzor2 is and why engineers selected it

FanzMAX starts from Fanzor2, a small, RNA-guided DNA-cutting protein that is being explored as a potential alternative to CRISPR-Cas systems, according to Penn Engineering. The protein family drew interest because it combines programmability with a much smaller footprint.

CRISPR-associated proteins such as Cas9 originated in bacteria, where they evolved as part of microbial immune systems, according to Penn Engineering. Fanzors are different because they are RNA-guided DNA-cutting proteins also found in eukaryotic organisms, the same broad category of organisms that includes humans.

Fanzor2 proteins in particular are fewer than 500 amino acids long, making them considerably smaller than Cas9, which can be well over 1,000 amino acids long, according to Penn Engineering. That size difference makes Fanzor2 potentially well suited to delivery by AAVs.

Earlier work had left an important gap. Fanzor2 proteins studied previously have generally shown only modest DNA-editing activity when tested in mammalian cells, according to Penn Engineering. The Penn project therefore treated Fanzor2 as a promising starting point that needed protein engineering to become more useful.

The engineering problem: too many possible changes

One option for improving a protein is to test mutations experimentally. For a protein roughly 500 amino acids long, the number of possible sequences is enormous, making it prohibitively burdensome to test them all, according to Penn Engineering.

Artificial intelligence can help narrow that search by ranking which changes look most worth testing. Yet AI models also need data, usually large amounts of data to make accurate predictions, according to Penn Engineering.

Fanzor2 presented a sparse-data case. Because Fanzor2 is a relatively new protein family with little experimental information available, even an AI approach would need to be innovative, according to Penn Engineering.

Postdoctoral fellow Shijie Wan described the starting point as about 200 measured single-mutation activities, according to Penn Engineering:

“We had only about 200 measured single-mutation activities. That is a tiny sample compared with the full mutational space of a protein that is roughly 500 amino acids long.”

That figure explains why the team did not rely on a conventional large training set. About 200 measurements cannot cover the combinations available to a chain roughly 500 units long. The approach summarized by BioCentury as adaptive model-guided protein evolution was framed as a way to work when experimental training data are limited.

How EvoMax learns from limited measurements

To tackle the sparse-data problem, the researchers developed EvoMax, according to Penn Engineering. The method does not test every possibility. It prioritizes candidate mutations using small amounts of experimental data.

EvoMax combines three types of information, according to Penn Engineering. One part learns from mutations researchers have already tested in the laboratory. A second part considers what evolution has already explored in natural protein sequences. A third part asks whether a proposed mutation appears compatible with the protein’s three-dimensional shape.

The design reflects a shift in how the laboratory used computation. Gao said computational tools had previously helped the group compare protein sequences, predict protein structures and generate ideas for experiments, according to Penn Engineering. In this project, computational and machine-learning tools were used more directly to help decide which mutations to test.

Gao also emphasized the division of labor between prediction and experiment in the Penn Engineering report:

“The computer does not give the final answer. Instead, it helps us choose a smaller and better set of candidates to test experimentally.”

That framing is consistent with a September 4 summary from Penn Engineering’s AI account, which said the approach helped researchers more efficiently identify promising ways to improve the system while noting that AI does not replace scientists.

Three rounds, 10 to 20 tests at a time

Instead of testing thousands of possibilities, EvoMax generated candidates and the researchers selected a small number, around 10 to 20, to test experimentally, according to Penn Engineering. That narrow batch size is central to the reported advance.

The best-performing variant from each round became the starting point for the next round, according to Penn Engineering. While moving forward, the researchers adjusted how they scored the different traits they were looking for.

After three rounds of optimization, FanzMAX v3 paired with the optimized RNA scaffold produced more than 13-fold higher reporter activation than the naturally occurring Fanzor2 enzyme paired with the original RNA scaffold, according to Penn Engineering. Reporter activation here served as the laboratory readout used to compare engineered and natural configurations.

The stepwise method meant that limited data grew more useful over time. Each round added new laboratory results, and the next predictions could build on the winner of the prior round. The team therefore did not need to solve the full sequence space at once.

Wan cautioned that the predictions remain imperfect and that laboratory confirmation remains required, according to Penn Engineering. The Penn Engineering AI summary likewise described EvoMax as a way to find promising candidate mutation sites from a very limited set of data, rather than as a replacement for testing.

What FanzMAX achieved in the reported tests

The headline laboratory result was editing efficiency as high as 97% at the best-performing genomic target for the engineered Fanzor2-based editor FanzMAX, according to Penn Engineering. Efficiency in this context refers to performance at that target, not an average across all possible targets.

The more than 13-fold higher reporter activation for FanzMAX v3 with the optimized RNA scaffold provides a second measure of improvement, according to Penn Engineering. The comparison was against the naturally occurring Fanzor2 enzyme paired with the original RNA scaffold.

Both the protein and its RNA partner mattered in that comparison. The engineered system combined changes to the protein with an optimized RNA scaffold, while the baseline combined the natural enzyme with the original RNA scaffold, according to Penn Engineering.

The Penn Engineering news index for September 3, 2026 summarized the same advance as transforming a compact gene-editing protein into a more powerful tool despite limited experimental data. It also noted that the work addresses why the best gene editor is not necessarily the most powerful one, a point that points to tradeoffs among activity, accuracy and safety.

Why activity, accuracy and safety must advance together

The Penn report states that a useful compact editor needs to be active, accurate and safe, not merely small, according to Penn Engineering. Activity measures whether editing happens. Accuracy and safety concern whether it happens at the intended place without unwanted effects.

That distinction matters for delivery-focused engineering. A single-AAV editor system is attractive because it fits the delivery constraint, but fitting inside AAVs does not by itself establish clinical usefulness. The team’s scoring adjustments across rounds reflect an effort to weigh different desired traits, according to Penn Engineering.

The study’s venue also signals the intended audience. Publication in Nature Biotechnology places the protein-engineering method and the FanzMAX editor before researchers who develop genome-editing tools, according to Penn Engineering. BioCentury’s September 2, 2026 summary likewise described the adaptive model-guided approach in the context of compact eukaryotic genome editors.

For readers tracking gene-therapy delivery, the relevant contrast is between lipid nanoparticles, engineered viruses and other carriers. Lipid nanoparticles can deliver editor proteins or messenger RNA, while engineered viruses can deliver the DNA encoding the editor, according to Penn Engineering. AAVs belong in the second group and face a limited cargo capacity, so a protein under 500 amino acids long has a structural advantage over a protein well over 1,000 amino acids long.

What remains a laboratory result

The facts reported by September 4, 2026 describe a laboratory-engineered editor and a machine-learning workflow, not a therapy ready for patients. Fanzor2 is being explored as a potential alternative to CRISPR-Cas systems, and FanzMAX is the engineered Fanzor2-based editor that emerged from this project, according to Penn Engineering.

The sparse-data claim is specific. The project began with only about 200 measured single-mutation activities for a protein roughly 500 amino acids long, used EvoMax to select around 10 to 20 candidates per round, ran three rounds, and retained the best variant as the next starting point, according to Penn Engineering. Those details define the scope of the reported efficiency gain.

Independent corroboration by the cutoff was limited to summaries and coverage of the same Penn result. BioCentury listed the University of Pennsylvania in connection with adaptive model-guided protein evolution for compact editors from sparse data on September 2, 2026. Penn Engineering published its detailed account on September 3, 2026, and Penn Engineering research accounts amplified the Nature Biotechnology study on September 4, 2026.

The practical takeaway is therefore narrow and concrete. A small eukaryotic RNA-guided protein can be made substantially more active without losing its size advantage, and a three-part model that uses laboratory data, natural sequence history and structural compatibility can guide that work when measurements are scarce. Whether that laboratory gain translates into active, accurate and safe editing in broader therapeutic settings remains a separate question that the reported facts do not answer.

Sources referenced