AAV-ML for Experimentalists #1: ML That Proposes New Sequences

TL;DR Generative AI is a tool that proposes new AAV sequences based on patterns it learned from relevant data.

You've heard "generative AI" everywhere: ChatGPT writes your emails, Midjourney makes images, and now... it designs AAV capsids?

What does that actually mean for you?

This post breaks it down. No jargon. No hype. Just what you need to understand as an experimentalist.

The Core Concept

Traditional capsid engineering works like this: you design a library, make it, screen it, and pick the winners, then screen them in a second-round screen to validate and mature them.

Generative AI proposes new sequences for you to make in the second-round or even in the first-round.

You're no longer just searching through what you built. You're asking the model: "What should I build next?"

Why Bother?

Here's the problem: sequence space is astronomically large.

A 7-amino-acid insertion has 20⁷ = 1.28 billion possible combinations. Your library, no matter how cleverly designed, samples a tiny fraction of that.

You might screen 10,000,000 variants in an NNK library and find 1,000 hits. Great. But what about the millions of sequences you never made? Some of them are almost certainly better than your best hit.

Generative models let you explore that space without synthesizing everything. They learn patterns from what worked and propose new sequences that fit those patterns — sequences you would never have included in your original library.

How It Works (No Math, Just Intuition)

Think of it like this:

The model reads thousands of sequences that "worked": capsids that packaged, transduced, or bound a target.

It doesn't memorize them. Instead, it learns the underlying patterns: which amino acids tend to appear together, which positions tolerate variation, what combinations are compatible.

It learns the grammar of viable capsids.

Then, when you ask it to generate new sequences, it writes new "sentences" that follow that grammar. Novel sequences, but ones that fit the patterns of what worked before.

The output isn't a guarantee. It's an educated proposal. You still validate.

Two Ways to Use Generative AI

This is where it gets practical. There are two distinct scenarios, and they require different expectations.

Scenario 1: Zero-Shot or Transfer Learning (You don't have AAV data for your specific problem yet)

Some generative models are pretrained on massive datasets: millions of protein sequences, or large public AAV datasets from other serotypes or properties.

These models have learned general "protein grammar." You can use them to generate starting points without running your own screens first.

Example: A model trained on public AAV2 packaging data can generate capsids for AAV9. Published work shows ~50% viability at 9-10 mutations — from a model that never saw AAV9 training data.

When to use this: You're starting a new campaign and want viable candidates without waiting 6-12 months to generate your own training data.

Limitation: The model only knows what it was trained on. If it learned packaging fitness, don't expect it to magically give you CNS tropism.

Scenario 2: Iterative Design (You have screening data; now amplify it)

This is where generative AI shines for experimentalists.

Here's the workflow:

Round 1: Screen your library. Identify hits.

Train: Feed those hits to a generative model.

Generate: The model proposes new candidates that resemble your winners but explore further — novel combinations you didn't test.

Round 2: Screen your original hits plus the generated candidates together.

The generated sequences aren't random guesses. They're informed proposals based on patterns in your hits. And critically, they can outperform your original winners.

Example: In work I led at the Broad Institute, we had only ~1,200 sequences that bound a specific receptor on human brain endothelial cells. We trained an autoregressive model on just those hits. The model generated novel variants —sequences not in our training set — that performed as well as the best known binders we had in the first round. It expanded our hit pool computationally, without expanding the screen (Read the full story here).

When to use this: You've done a screen, found hits, but want more diversity or better performers without brute-force scaling.

What Generative Models Need

For any of this to work, you need:

Training data: Examples of what worked. Either your own screening hits or public datasets.

A way to evaluate outputs: The model proposes; something has to judge. That's either a predictive model (scoring the proposals computationally) or your next experiment (most common).

Realistic expectations: Outputs are suggestions, not guarantees. Hit rates will be higher than random, but not 100%.

What Generative AI Can and Cannot Do

It can:

Explore sequence space beyond your starting library

Propose non-obvious combinations you wouldn't have designed

Amplify a small set of hits into a larger, diverse candidate pool

Accelerate your campaign by reducing wasted screens

It cannot:

Guarantee anything works; it's probabilistic

Invent biology it hasn't seen; if tropism patterns weren't in training, it can't learn them

Replace your experimental judgment about what matters

Skip validation; you still need to make and test the sequences

Where the Field Stands

Generative AI in AAV has moved from academic curiosity to standard practice. Here's what's happening now (based on literature and ASGCT abstracts up to 2025):

Integrated pipelines: Most groups combine generative models with predictive filters. Generate candidates, score them computationally, validate the top tier experimentally.

LLMs entering the space: Foundation models pretrained on billions of proteins are being fine-tuned for AAV. Early results are promising, but the jury is still out on whether massive pretraining beats AAV-specific models.

Zero-shot is real but limited: Cross-serotype transfer works for some properties (packaging), less proven for others (tissue tropism).

Architecture diversity: VAEs, autoregressive models, diffusion models, transformers — different groups use different approaches. Do not get scared about the model names, they basically do the same generation process under different problem circumstances.

The trend: Generative is no longer a "nice to have." It's becoming a default step in capsid engineering workflows.

How It Fits Your Workflow

If you're starting fresh (zero-shot): Use pretrained models or public-data-trained generators to get viable starting libraries. You skip the cold-start problem and jumpstart your campaign.

If you have screening data (iterative): Train on your hits. Generate new candidates. Screen both together. Your ROI per experiment goes up because you're not just re-screening; you're expanding intelligently.

The Questions to Ask

When someone tells you they're using generative AI for AAV, ask:

"What was it trained on?" The model can only remix what it learned. AAV2 packaging data won't teach it liver detargeting.
"Zero-shot or fine-tuned?" Zero-shot means no task-specific training; broader but less precise. Fine-tuned means it saw data for your exact problem; narrower but more relevant.
"How were the generated candidates validated?" Computational scores alone aren't proof. Did they actually make and test the sequences? How many? What was the hit rate?

Final Thought

Generative AI is a tool that proposes new sequences based on patterns it learned from data.

It doesn't replace your experiments; it makes them smarter. Fewer wasted screens. Broader exploration. Higher hit rates.

The key is knowing when to use it, what to feed it, and what questions to ask.

Next in the series: ML That Scores Your Variants (how predictive models work and when to trust them).

PS: This is what The AIxAAV Interpreter is for: translating ML methods into actionable AAV engineering strategies. Follow me on LinkedIn for more practical insights that accelerate bio-innovation.

Search This Blog

The AI × AAV Interpreter