Abstract
Ultracentrifugation is an attractive method for separating full and empty capsids, exploiting their density difference. Changes of the serotype/capsid, density of loading material, or the genetic information contained in the adeno-associated viruses (AAVs) require the adaptation of the harvesting parameters and the density gradient loaded onto the centrifuge. To streamline these adaptations, a mathematical model could support the design and testing of operating conditions.Here, hybrid models, which combine empirical functions with artificial neural networks, are proposed to describe the separation of full and empty capsids as a function of material and operational parameters, i.e., the harvest model. In addition, critical quality attributes are estimated by a quality model which is operating on top of the harvest model. The performance of these models was evaluated using test data and two additional blind runs. Also, a "what-if" analysis was conducted to investigate whether the models' predictions align with expectations.It is concluded that the models are sufficiently accurate to support the design of operating conditions, though the accuracy and applicability of the models can further be increased by training them on more specific data with higher variability.