Iranian researchers use AI to design kidney-targeted viral shells for gene therapy

An AI approach to a stubborn gene-therapy target

A pair of researchers in Iran has unveiled an artificial intelligence system that designs entire viral shells for gene therapy, tuned specifically for one of medicine’s most difficult targets: the kidney.

The framework, called AAVGen, was posted Feb. 21 as a preprint on the scientific server arXiv. It uses advanced protein language models, a reinforcement learning algorithm originally built for large language models, and Google DeepMind’s AlphaFold3 structure predictor to generate adeno-associated virus (AAV) capsid proteins that, in computer tests, appear better suited for kidney-targeted delivery than the sequences the system was trained on.

None of the AI-designed viruses have yet been synthesized or tested in animals or people. But the work highlights how quickly techniques refined on text and images are moving into the design of the viral vectors that carry gene therapies into the body.

What AAVGen aims to optimize

In the 22-page manuscript, co-authors Mohammadreza Ghaffarzadeh-Esfahani and Yousof Gheisari describe AAVGen as “a generative AI framework for de novo design of AAV capsids with enhanced multi-trait profiles,” arguing that it could accelerate the development of next-generation vectors with tailored characteristics.

The traits in question reflect major bottlenecks in gene therapy. AAVGen tries to optimize three properties of the AAV VP1 capsid protein at once:

• Production fitness, which affects how efficiently a virus can be manufactured
• Kidney tropism, or how strongly it targets kidney tissue
• Thermostability, a measure related to storage and handling

To do that, the authors start from a large protein language model—neural networks trained to predict amino acids in sequences much as language models predict words in a sentence. They fine-tune the model on existing AAV capsid data and connect it to three smaller models, each based on Meta’s ESM-2 protein language model family, trained to predict one of the three desired traits.

The outputs of those ESM-2-based “reward models” are combined into a single score. A reinforcement learning algorithm known as Group Sequence Policy Optimization (GSPO) then iteratively adjusts capsid sequences to raise that composite score. GSPO was originally developed to stabilize training for large language models; here it is used to steer protein sequences instead of text.

What the preprint reports—and what it doesn’t

According to the preprint, most of the AI-generated sequences scored higher than those in the training set on all three traits at once, within the limits of the predictive models.

To check whether the new capsid proteins still folded into plausible three-dimensional structures, the team turned to AlphaFold3, which can model complexes of proteins and other biomolecules. The authors report that AlphaFold3 predictions for their designs maintained the canonical AAV capsid fold, despite in some cases extensive changes in amino acid sequence.

They did not report specific measures of predicted stability—such as model confidence scores—and they did not claim experimental validation.

Why kidney targeting is hard

AAV vectors are among the workhorses of modern gene therapy. Several approved medicines—including Luxturna for an inherited eye disease, Zolgensma for spinal muscular atrophy and Hemgenix for hemophilia B—deliver therapeutic DNA using modified versions of naturally occurring AAV serotypes. In these therapies, the viral shell is stripped of its disease-causing genes and repurposed as a delivery vehicle.

Which organs those delivery vehicles reach depends heavily on the capsid proteins forming the virus’s outer shell. Native serotypes such as AAV2 and AAV9 tend to favor certain tissues, including the liver, muscle or central nervous system. That pattern, known as tropism, has spurred years of research into engineering capsids that seek out other tissues and avoid off-target organs.

The kidney has been one of the hardest targets. Chronic kidney disease affects more than one in 10 people worldwide, but the organ’s dense network of filters and tubules, along with its unusually high blood flow, makes it difficult for viral vectors to reach the right cells and stay there. Many vectors are swept away or sequestered in non-target tissues.

In 2025, a team led by AAV engineer Aravind Asokan reported in Nature Biomedical Engineering that they had evolved kidney-tropic AAV variants dubbed AAV.k13 and AAV.k20 by repeatedly screening large capsid libraries through mice, pigs, nonhuman primates and human kidney organoids. Those variants delivered genes more efficiently to proximal tubule cells than standard AAV9 in several models, but the work relied on elaborate, animal-intensive selection campaigns.

AAVGen, by contrast, does its evolution in silicon. The authors do not claim their virtual designs outperform laboratory-evolved variants like AAV.k13 or AAV.k20—only that they look promising according to computational predictors. They also acknowledge that “comprehensive experimental validation” will be needed before any capsids can be considered for therapeutic use.

The larger trend—and lingering risks

AAVGen sits within a broader trend of using machine learning to design viral capsids and other proteins. In recent years, academic groups and startups have reported diffusion models that generate AAV capsid sequences, graph-based neural networks that rewire the protein core of AAV2 while preserving function, and industrial platforms that use high-throughput in vivo screens plus machine learning to discover capsids with better performance in nonhuman primates.

At the same time, structural prediction tools such as AlphaFold2 and AlphaFold3 have made it routine to assess whether a designed protein is likely to fold correctly before committing to wet-lab experiments.

For all the technical progress, experts caution that surrogate models and structure predictors are only approximations. Predicting that a capsid will assemble correctly or favor kidney cells based on sequence alone is far easier than ensuring it will do so safely and effectively in a living organism.

Many AAV-based therapies have run into unexpected immune responses or toxicity in human trials, even when using relatively well-characterized capsids. AAVGen’s current objective functions do not explicitly optimize for immunogenicity or pre-existing antibody recognition, leaving open questions about how the immune system would react to highly novel capsid designs.

There are also broader safety and governance questions as generative models for viruses become more capable and more accessible. While AAV vectors used in therapy are engineered to be replication-defective and have a strong safety record compared with many other viruses, similar computational techniques could, in principle, be applied to pathogenic viral families.

Because AAVGen relies largely on publicly available components—open protein language models, an arXiv preprint, a Kaggle notebook and the noncommercial AlphaFold server—its emergence in an academic medical center in Isfahan also illustrates how global the tools of AI-driven bioengineering have become. That accessibility may help researchers in countries with limited wet-lab resources participate more fully in early-stage design work, but it also complicates efforts to track or control advanced viral engineering.

What comes next

For now, AAVGen remains a digital proof of concept. The next steps, the authors indicate, will involve selecting a subset of the AI-designed capsids, synthesizing their genes, and testing whether they assemble, package and deliver cargo DNA in cells and animal models as predicted. Only then will it be clear whether the algorithm’s virtual gains in kidney tropism and stability translate into real-world improvements.

Even if many candidates fail at the bench, the work underscores a shift in how viral vectors are conceived. Instead of tweaking natural serotypes a few mutations at a time, researchers are increasingly treating capsids as programmable objects that can be optimized across multiple traits using the same kinds of AI systems that transformed language and image generation. As regulators and ethicists debate how to oversee those tools, the viruses that carry tomorrow’s gene therapies may be taking shape first on a GPU.