Most gene therapies built on adeno-associated virus get exactly one shot. Inject the vector once and the immune system learns the shell it rides in, builds neutralizing antibodies against that shell, and any second dose gets swarmed and cleared before it reaches a cell. Neutralizing antibodies are the immune proteins that grab a virus and stop it from getting inside. Worse, many people already carry them from ordinary AAV exposure years earlier, which can lock them out of the therapy before they ever get a first dose. Roughly a third to a half of adults test positive for anti-AAV antibodies at baseline, according to a review of pre-existing anti-AAV immunity ↗.
A group at the Institute of Hematology and Blood Diseases Hospital in Tianjin, part of the Chinese Academy of Medical Sciences, working with AI researchers at the Harbin Institute of Technology, tried a workaround: build a decoy. Their method, published in Molecular Therapy Advances ↗, designs short peptides that look, to an antibody, like the outer surface of the virus. Give the antibodies something else to grab, and the real vector slips past.
How they picked the decoys
The design was the point, not a screen of thousands of candidates. The team ran two protein language models (BERT plus Evolutionary Scale Modeling, tuned with a lightweight adjustment method called LoRA) and slid a window across human and viral proteins to find stretches that are distinctly viral, sit on the exposed outer surface, and do not resemble anything human. That last filter matters: a decoy that looks human-made-strange could itself become a new immune target. From the shells of two AAV variants, the common AAV2 and an engineered capsid called AAV843, the models flagged the surface fragments worth turning into peptides.
In a dish, the chosen peptides latched onto AAV-specific antibodies tightly (nanomolar affinity, meaning a strong and selective grip), pulled those antibodies off the virus, and restored more than 60 percent of the vector's normal delivery even with neutralizing antibodies present.
The mouse arm needed help
In mice that either already carried anti-AAV antibodies or had built them after a first dose, the decoys alone did not clear the block. Paired with an IgG-degrading enzyme, a bacterial protein that cuts antibodies apart, the combination knocked high antibody levels down to about a 1-to-4 titer and got the liver making the delivered gene's protein again, with no sign of inflammation or a fresh immune reaction against the peptides. A second AI model, a variational autoencoder, generated entirely synthetic peptides covering a wider spread of antibody targets.
Here is the honest read. This is cells and mice, not patients. A titer cut to 1:4 is a reduction, not a clearance. And the part that did the heavy lifting in the animals was the enzyme, not the peptide.
IgG-degrading enzymes are already the leading tactic against this problem. Imlifidase, an enzyme from Streptococcus that cleaves human antibodies, has stripped anti-AAV antibodies in primates and is being tested in people. Its weakness is that a large share of the population already carries antibodies against the enzyme itself, which blunts it and can trigger reactions. A decoy peptide that lets a smaller enzyme dose do the same job, or works in the window where the enzyme stalls, would be worth having. That case has not been made in a human, and "no immune response to the peptide" over a short mouse study is a thin guarantee for a molecule meant to be given more than once.
The design logic is what stands out. Instead of fishing through a library, the group let sequence models decide which fragments of a viral shell are safe to mimic, then invented new ones that never existed in nature. That is the same design-first approach behind a growing shelf of computationally proposed peptides, several of which peptidemodel tracks against the immune system ↗. Whether AI-drawn decoys hold up against a human immune system is the open question, and it hangs over most of them.