The AI × AAV Interpreter

TL;DR  Public AAV2 data → generative model → viable AAV9 capsids. No AAV9 data needed. 50% hit rate at 9+ mutations. You can transfer the same way to other AAV serotypes; the data bottleneck just got a shortcut.

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.

TL;DR: Your PLM is confusing "distant" AAVs with "broken" AAVs.   Left: Raw PLM scores unfairly penalize deep mutants, trapping high-fitness variants behind a "distance barrier." Right: Depth-normalization removes the bias, allowing heavily mutated sequences (like novel AAV capsids) to shine just as bright as near-WT ones. →  Don't let your model hide the best proteins just because they look different.