antibodies. A two-stage fractionated factorial design with six factors was
implemented, and various regression models were used to identify the active variables [49]. This resulted in 384 experiments to be performed, which was only
possible by using a deep well plate system. Nevertheless, there were variations in
statistical significance, and possible active variables have to be tested on a larger
scale [49].
The amount of experiments to be performed can be seen as the main challenge in
using optimization designs as well. The most commonly used optimization design is
CCD. Yang et al. (2014) used a CCD to optimize the concentration and timing of
valproic acid (VPA) addition to the cultivation of three different CHO cell lines
[50]. Even the investigation of two factors for one cell line results in eight experiments. Torkashvand et al. (2015) optimized the concentrations of four amino acids
(aspartic acid, glutamic acid, arginine, and glycine) in the feed using a BBD. The
factors were investigated at 3 levels, resulting in 29 experiments to be implemented
[51]. Duvar et al. (2013) developed a feeding protocol for a fed-batch CHO cultivation. The choice of a D-optimal experimental design resulted in 18 experiments
with 4 factors (feeding volume, starting point, time of shift in temperature, and
osmolality) [34].
For the previously mentioned studies, the planned experiments in statistical DoE
result in identifying active parameters and optimization of the bioprocess. However,
there are still challenges, and the implementation of statistical DoE can lead to timeconsuming and costly rounds of experiments, especially if they are implemented in
fed-batch mode. Furthermore, the heuristic selection of, e.g., the parameter settings
or the design selection is seen critically. These rely on user-defined settings and
mostly require a lot of time and experimental effort. But, as the investigation of
various studies has shown, no adequate justification for the choice of an experimental design is provided. In addition, in conventional DoE only the experimental
endpoints are examined, and therefore only the integral of it is judged. The entire
time trajectory, with, e.g., metabolite formation or substrate uptake, is hardly
reflected.
3 Model-Assisted Design of Experiments
The combination of statistical DoE with mathematical process models is a novel
tool – enabling a knowledge-driven bioprocess development in the context of QbD.
Using this method, the abovementioned limitations of DoE methods can be avoided,
and the design as well as the optimization of bioprocesses can be improved.
However, in contrast to the chemical industry, bioprocess design on the basis of
mathematical models is not yet well established in biopharmaceutical manufacturing
processes with mammalian cells [52]. According to experiences of the authors from
discussions and projects, the use of model-based innovative methods for process
development has so far failed due to different reasons:
38
K. B. Kuchemüller et al.
implemented, and various regression models were used to identify the active variables [49]. This resulted in 384 experiments to be performed, which was only
possible by using a deep well plate system. Nevertheless, there were variations in
statistical significance, and possible active variables have to be tested on a larger
scale [49].
The amount of experiments to be performed can be seen as the main challenge in
using optimization designs as well. The most commonly used optimization design is
CCD. Yang et al. (2014) used a CCD to optimize the concentration and timing of
valproic acid (VPA) addition to the cultivation of three different CHO cell lines
[50]. Even the investigation of two factors for one cell line results in eight experiments. Torkashvand et al. (2015) optimized the concentrations of four amino acids
(aspartic acid, glutamic acid, arginine, and glycine) in the feed using a BBD. The
factors were investigated at 3 levels, resulting in 29 experiments to be implemented
[51]. Duvar et al. (2013) developed a feeding protocol for a fed-batch CHO cultivation. The choice of a D-optimal experimental design resulted in 18 experiments
with 4 factors (feeding volume, starting point, time of shift in temperature, and
osmolality) [34].
For the previously mentioned studies, the planned experiments in statistical DoE
result in identifying active parameters and optimization of the bioprocess. However,
there are still challenges, and the implementation of statistical DoE can lead to timeconsuming and costly rounds of experiments, especially if they are implemented in
fed-batch mode. Furthermore, the heuristic selection of, e.g., the parameter settings
or the design selection is seen critically. These rely on user-defined settings and
mostly require a lot of time and experimental effort. But, as the investigation of
various studies has shown, no adequate justification for the choice of an experimental design is provided. In addition, in conventional DoE only the experimental
endpoints are examined, and therefore only the integral of it is judged. The entire
time trajectory, with, e.g., metabolite formation or substrate uptake, is hardly
reflected.
3 Model-Assisted Design of Experiments
The combination of statistical DoE with mathematical process models is a novel
tool – enabling a knowledge-driven bioprocess development in the context of QbD.
Using this method, the abovementioned limitations of DoE methods can be avoided,
and the design as well as the optimization of bioprocesses can be improved.
However, in contrast to the chemical industry, bioprocess design on the basis of
mathematical models is not yet well established in biopharmaceutical manufacturing
processes with mammalian cells [52]. According to experiences of the authors from
discussions and projects, the use of model-based innovative methods for process
development has so far failed due to different reasons:
38
K. B. Kuchemüller et al.