Figure 6: The basic principles of the quality by design approach, where more flexibility is obtained by a characterized space; adapted from (Carta et al., 2010).
Ideally, the full design space is covered by the training, validation and test data. In this ideal situation interpolation within the design space will be very good (Figure 6). Rolinger et al. (Rolinger et al., 2021) five runs in their study. Four runs covered the design space, and the center point was used as an independent test set. It is not surprising that the model performed well. Sauer et al and Walch et al. (Sauer et al., 2019; Walch et al., 2019) on the other hand simulated the situation of a manufacturing run under clearly defined process conditions and used eight runs performed under identical process conditions for model building. The performance was then evaluated on purification runs performed under identical conditions but with biological material from a new fermentation and under different process conditions. It was shown that the model performed well on new biological material (Sauer et al., 2019) and that model extrapolations are possible but are limited to modifications in the column load and small changes in buffer composition (Walch et al., 2019).