Stop Trying to Fix Toxic AAV Variants (with or without ML)

At The AI×AAV Interpreter, we dissect how computational tools transform capsid design. Often, we focus on building better models. But a recent preprint on muscle-targeting AAVs presents a powerful counterpoint: sometimes, the most intelligent solution isn't a more complex algorithm, but a smarter problem formulation. Here’s why this story is a critical lesson for every team using ML to engineer multi-trait viruses.

TL;DR Before training another model to predict multi-trait capsids, ask if the problem can be decomposed. PackGene didn't ML their way to a muscle-tropic, liver-detargeted capsid; they engineered a blank-slate backbone and plugged in a targeting peptide. Same muscle transduction as MyoAAV.4A, no hepatotoxicity in NHPs.

You spent months screening capsid libraries.

You finally found it: a variant with beautiful, enhanced targeting for your tissue of interest. It works. You’re excited.

Then you run the full biodistribution.

The signal in a toxic off-target organ is through the roof. Your elegant, tissue-tropic capsid now carries a dangerous side effect. So you try to "fix" it.

You mutate, you swap loops, you add shields. Or try to use ML to find the best of two worlds. Every change risks breaking the precious targeting you just found. Most of the time, it breaks.

This is the painful, backward struggle of multi-trait AAV engineering.

The Flawed Playbook

The standard approach goes like this:

Screen for your primary trait (e.g., muscle tropism).
Discover your hit has a dangerous secondary trait (e.g., high liver tropism).
Spend months trying to engineer out the bad trait using rational or ML methods without breaking the good one.
Often fail, because the traits are mechanistically entangled.

This happens because improved transduction often leverages pathways that are also highly efficient in toxic off-target tissues; when you enhance one, you frequently enhance the other.

The Inversion: Start Safe, Then Add Targeting

What if you flipped the script? Instead of starting with a toxic variant and trying to fix it, what if you began with a safe, "blank slate" capsid, then added your desired targeting?

This is exactly what a recent paper demonstrates, using muscle targeting as a powerful example. Researchers at PackGene Biotech created AAV.Zero3: a rationally designed capsid with globally reduced tropism. It barely transduces anything, liver, muscle, or otherwise. It’s a clean chassis.

Then, they added muscle targeting by inserting a known myotropic peptide. The result: AAV.eM, a capsid that delivered strong muscle transduction in mice and NHPs without the dangerous liver toxicity seen in earlier muscle-targeting capsids.

The liver de-targeting wasn’t an afterthought; it was built into the backbone from the start.

What this means for industry

The team showed this was a modular strategy. They inserted different targeting peptides into the same safe AAV.Zero3 backbone, and all showed reduced liver transduction while maintaining muscle targeting across species.

This is a design platform, not a one-off: start with a safe, low-background chassis, then plug in your targeting module.

The bigger picture for AAV engineering

This muscle story is a template for any multi-trait engineering challenge where desired and toxic traits are linked:

CNS targeting without peripheral toxicity: Start with a backbone that doesn’t transduce peripheral organs, then add brain tropism.
Immune evasion + efficient transduction: Start with a stealth capsid, then add your tissue-targeting element.
High yield + specific targeting: Start with a stable, manufacturable backbone, then add your functional loops.

The principle is universal: Begin with the hard-to-achieve "safe" trait engineered into your backbone. Then add the "targeting" trait modularly.

What This Means for ML-AAV

This approach reframes the computational challenge entirely. Instead of asking our models to perform a high-wire act, searching vast sequence space for a needle-in-a-haystack variant that miraculously combines perfect targeting with zero toxicity, we can decompose the problem.

First, use rational design or directed evolution for a single, critical trait: creating a clean, safe, "blank slate" backbone. Then, apply screening or ML to find modules that confer the second trait (e.g., targeting) on that safe foundation.

The lesson isn't that ML multi-trait engineering is useless. It's that ML is most powerful when applied to a well-structured problem. The paper's strategy simplifies the fitness landscape, making the remaining optimization task more tractable for both computational and library-based methods.

Before you train your next foundation model on complex multi-trait fitness data, ask: could we engineer the most dangerous trait out of the system first? Stop trying to fix toxic variants with ML. Start by building safe backbones; then let your models go to work.

Sometimes the smartest use of computation is recognizing when you don't need it.

Final Thought

At least four patients have died from AAV-related complications in gene therapy trials.

Every muscle-targeting program monitors liver enzymes obsessively. Every dose escalation balances efficacy against hepatotoxicity. Every clinical hold for elevated transaminases delays therapies reaching patients.

We keep doing this because we start from the wrong place.

We find capsids that do one thing beautifully (transduce muscle, cross the BBB, evade antibodies, etc) and then we fall in love with them. We patent them. We optimize them. We try to "fix" their toxic properties without breaking what we love.

It's like trying to change a toxic partner. It rarely works.

The paradigm shift is simple: start with a partner who doesn't hurt you. Then build the relationship you want.

Start with capsids that don't have the dangerous properties. Add the beneficial properties via rational, modular design.

This isn't just one paper. It's a different way to think about multi-trait AAV engineering.
Stop trying to fix toxic variants. Start with safe backbones.

Credit:

The muscle-targeting work discussed is from: "Engineering novel AAV capsids by global de-targeting and subsequent muscle-specific tropism in mice and NHPs" (2025). Preprint: [Link]

PS: If you want to go deeper on translating ML predictions into actionable AAV biology, that's what The AIxAAV Interpreter is for. I've spent two decades bridging ML theory and application.

Follow me on LinkedIn (#AIxAAV #AIxAAVinterpreter #TheBioMLClinic #TheBioMLPlayBook) for more practical insights that accelerate bio-innovation.

Search This Blog

The AI × AAV Interpreter