Primary source bioRxiv: Machine learning-assisted directed evolution yields dramatic improvement on novel AAV engineering task
This is a preprint posted on bioRxiv on 7 October 2026. It has not been peer reviewed, and the points below come from the abstract alone, the only part of the paper we opened.
The authors work on adeno-associated virus (AAV) capsids, the protein shells used as vectors. Directed evolution finds improved protein mutants through repeated rounds of selection. Machine learning (ML) can rank mutants so that fewer need to be tested in the lab. The abstract says it is unclear whether that benefit outweighs the added cost of building ML-designed libraries (the sets of mutants made for testing).
What the authors measured
The team looked at improvements in the manufacturing efficiency of AAV capsids in the context of exosomal encapsulation. They built a directed evolution dataset of 53,974 mutants across three rounds of selection. They also built a separate assessment dataset of 472 ML-designed mutants with detailed profiling.
Directed evolution alone gave a 3-fold improvement in manufacturing efficiency. ML-assisted directed evolution gave a 41-fold improvement. The abstract does not say which baseline these folds are measured against, and we did not read the full text.
Data quality over model choice
The authors report that data quality had a greater impact on model performance than model architecture. They name two factors: the number of selection rounds and the statistical power of the mutant counts. They also say they provide practical guidance on experimental design and on ML tuning strategies for protein engineering.
What the abstract does not say
It gives no scale, titre or cost figures. It reports no manufacturing run in a production process. Results for other capsid tasks or other vectors are not covered.
