different cell-cycle-dependent growth rates, metabolic activity, and DNA replication
rates are modeled, and metabolic regulations were studied, but the degree of
complexity and computational power increases significantly from a few seconds to
multiple hours. Hence, they require a comprehensive knowledge of the mechanisms
and more data to estimate the model parameters. Segregated and structured models
describe the nature of cell cultures with individual single-cell metabolism and their
interaction with the medium. Sanderson et al. (1999) introduced a single-cell model
that describes the interaction of 50 components in the medium, cytoplasm, and
mitochondria for an antibody-producing CHO cell line [80]. Other examples could
be found for baculovirus-infected insect cell cultures [81] and the amino acid
metabolism of HEK293 and CHO cells [82]. However, the computational power
and amount of data required to estimate the model parameters are still considerable,
which can limit their industrial application.
3.2 Recommendations on the Selection of Designs for mDoE
The choice of an experimental design significantly influences the implementation of
DoE and mDoE. Usually, the selection of a design depends on basic settings
(number of factors, number of factor steps, the regression model, and the number
of test runs) and design-specific properties (block formation, orthogonality, and
rotatability). However, as the investigation of various studies has shown, in most
references less information for the choice of an experimental design is provided.
DoEs are mostly selected based on heuristics within a given scientific field, and there
is no guided decision-making workflow yet. Based on the author’s understanding, a
scheme (Fig. 5) is presented in the following to assist in the selection of appropriate
DoE designs in the field of bioprocess engineering. This scheme was developed
based on literature and is seen to assist in the selection of DoE designs within mDoE
[8, 11, 12].
Due to their favorable properties, CCDs and BBDs are most frequently used for
optimization [39]. If settings are adjusted individually, optimal designs should be
used. As a result of the low computational effort, the D-optimal design has become
generally accepted among the optimal designs [38]. Therefore, commercial software
tools for creating optimal designs are often limited to the D-optimal design
[83]. However, the I-optimal designs are sometimes recommended [39, 83]. If a
large area of the factor space is to be covered, it is recommended, e.g., to combine the
LHSD with an optimal design. This results in a better distribution of points across the
factor space and reduces both bias as well as noise [45].
In the first decision-making level, the number of investigated factors k is used.
Except for BBDs, which require at least three factors, the number of factors can be
selected as desired [38, 41, 84]. Typically, three to six factors are used for processes
optimization [38]. In the case of a high number of factors and the use of insignificant
factors, the LHSD is recommended. This is enabled by the random distribution of
experiments. Hence, if one or more factors appear not to be important, every point in
44
K. B. Kuchemüller et al.
rates are modeled, and metabolic regulations were studied, but the degree of
complexity and computational power increases significantly from a few seconds to
multiple hours. Hence, they require a comprehensive knowledge of the mechanisms
and more data to estimate the model parameters. Segregated and structured models
describe the nature of cell cultures with individual single-cell metabolism and their
interaction with the medium. Sanderson et al. (1999) introduced a single-cell model
that describes the interaction of 50 components in the medium, cytoplasm, and
mitochondria for an antibody-producing CHO cell line [80]. Other examples could
be found for baculovirus-infected insect cell cultures [81] and the amino acid
metabolism of HEK293 and CHO cells [82]. However, the computational power
and amount of data required to estimate the model parameters are still considerable,
which can limit their industrial application.
3.2 Recommendations on the Selection of Designs for mDoE
The choice of an experimental design significantly influences the implementation of
DoE and mDoE. Usually, the selection of a design depends on basic settings
(number of factors, number of factor steps, the regression model, and the number
of test runs) and design-specific properties (block formation, orthogonality, and
rotatability). However, as the investigation of various studies has shown, in most
references less information for the choice of an experimental design is provided.
DoEs are mostly selected based on heuristics within a given scientific field, and there
is no guided decision-making workflow yet. Based on the author’s understanding, a
scheme (Fig. 5) is presented in the following to assist in the selection of appropriate
DoE designs in the field of bioprocess engineering. This scheme was developed
based on literature and is seen to assist in the selection of DoE designs within mDoE
[8, 11, 12].
Due to their favorable properties, CCDs and BBDs are most frequently used for
optimization [39]. If settings are adjusted individually, optimal designs should be
used. As a result of the low computational effort, the D-optimal design has become
generally accepted among the optimal designs [38]. Therefore, commercial software
tools for creating optimal designs are often limited to the D-optimal design
[83]. However, the I-optimal designs are sometimes recommended [39, 83]. If a
large area of the factor space is to be covered, it is recommended, e.g., to combine the
LHSD with an optimal design. This results in a better distribution of points across the
factor space and reduces both bias as well as noise [45].
In the first decision-making level, the number of investigated factors k is used.
Except for BBDs, which require at least three factors, the number of factors can be
selected as desired [38, 41, 84]. Typically, three to six factors are used for processes
optimization [38]. In the case of a high number of factors and the use of insignificant
factors, the LHSD is recommended. This is enabled by the random distribution of
experiments. Hence, if one or more factors appear not to be important, every point in
44
K. B. Kuchemüller et al.