Model interpretability methods provide an understanding of complex model decisions and verify that a meaningful difference in the data has been identified. We have applied model interpretability methods to our predictive model of genome truncation within adeno-associated virus (AAV) manufacturing and revealed that the model uses a set of DNA secondary structures predictive of truncation. These secondary structures provide a simple mechanism for understanding AAV truncations and a strong basis for independently validating our model’s predictions. Moreover, these structures have been well studied and are shown to be related to DNA replication errors; however, only one of these structures (i.e., hairpins) has been previously implicated in AAV manufacturing failures. Future research will concentrate on additional contributors to our truncation model to gain a deeper understanding of AAV manufacturing failures more generally.
We support the commitment of the Biden-Harris Administration to address global issues through biomanufacturing innovation. Learn more.
Continuing our blog series on Transformer architectures and their use in gene therapy applications, we explore the latest variants’ size and speed improvements.
Transformer based architectures are uniquely efficient and powerful when used to process structured data. Read this blog to learn how they can be customized, adapted and applied to AAV construct optimization.