A hallmark of directed evolution of AAV capsids for gene therapy applications is their ex or in vivo selection from diverse libraries, which are typically composed of hundreds of different variants (in the case of pre-selected, barcoded capsids) or up to billions (in the case of e.g. shuffled capsids). Recently, though, it has become increasingly apparent that the type of selection model can significantly affect and bias the outcome, evidenced by capsids that do not translate from cultured cells to animals or humans, or that even perform differently between various strains from the same species, such as mice or monkeys.

Here, we have systematically investigated and confirmed this phenomenon, by analyzing and comparing the results of iterative or single-round AAV capsid screening ex vivo in cultured murine muscle cells, or in vivo in the musculature of mice or of monkeys.

To this end, we initially screened a shuffled capsid library composed of AAV serotypes 1, 6, 8 and 9 upon systemic delivery in mice or monkeys, with a focus on the striated musculature as our major target of interest in this study. In mice, we observed a gradual loss of AAV1, 6 and 8 together with a concurrent enrichment of AAV9, which eventually contributed to about 72% of all capsid sequences at the end of the selection. In striking contrast, AAV9 but also AAV8 were diminished in macaques, while AAV1 and 6 became enriched and ultimately made up roughly 57% of all capsid sequences.

Based on these data, we next prepared a barcoded pool of 54 capsids enriched in the previous study and including three clusters of similar sequences, i.e., the AAV1/6-derived clusters 1 and 2 as well as the AAV8-derived cluster 3. Moreover, we added the parental AAV1, 6, 8 and 9 as benchmarks, as well as the AAV9-based peptide display variant AAVMYO that exhibits a pronounced muscle tropism in rodents. 

Also in this secondary screen, we noted substantial differences in the composition and phylogeny of the capsids that performed best in each of three selection systems, i.e., murine C2C12 myoblasts or intravenously injected NMRI mice or macaques. Most notably, AAV9 and AAVMYO outperformed all other variants on the RNA level in the mouse musculature, where they were enriched 7.1- or 30.7-fold over the input library, respectively. In contrast, AAV6 and two derivatives in cluster 2 with ~97% identity to AAV6 performed much better in the macaques, where they were ~5-fold enriched over the input library. The lead benchmark in mice, i.e., AAV9, was enriched 1.8-fold in the macaque model and thus 3.9-fold less efficiently than in the mouse model.

Collectively, this work enables us to draw two major conclusions, the first being that AAV6 might represent a better scaffold than the commonly used AAV9 for capsid evolution in the primate musculature. Secondly, we validate the concern that the outcome of AAV library screenings is critically affected by the selection model, underscoring the need for concerted efforts across academia and industry to find solutions to this problem.