• It is based on the lack of knowledge of the potential and limits of model-based/
model-assisted methods and “bad experiences” with them (e.g., due to unrealistic
expectations).
• There is a lack of method integration for a consistent development strategy, which
can be adapted to existing work processes, a suspected high (modeling) effort and
doubts about the transferability of methods and models to other processes.
• The qualification profile of the involved personnel does often not fit (required:
biotechnology, process technology, modeling, and statistics).
An additional challenge is the application of models on complex metabolic
pathways of mammalian cells regarding cell growth and product formation. In
addition, models targeting the metabolism of cell cultures demand more effort than
those applied in chemical or microbial processes. Even if mathematical models are a
promising tool for the development of stable processes that comply with the principles of QbD, examples have so far only been published in the field of product
purification and polishing [8]. Nevertheless, Möller et al. (2018) and Abt et al.
(2018) showed that model-assisted and model-based DoE methods have great
potential for the development of process strategies and makes the process development more knowledge-based [7].
General differences between model-based and model-assisted DoE methods are
due to the aim of the recommended experiments. Model-based DoE (MBDoE) [53–
55] is used to supply valid experimental data for a precise model structure and model
parameter identification, where the conventional statistical DoE could fail
[56]. Uncertainties are key information in MBDoE, as model and data imperfections
cause undesirable variations in model parameters and simulation results. This
variation drives the MBDoE methodology, where it manifests itself as optimal
experimental settings (e.g., measurement principle, sampling rate, inputs/stimuli)
and informative data [55, 57]. However, uncertainties cause a discrepancy between
computed and experimental outputs leading to suboptimal or even meaningless
experimental designs for model parameter adaptation. To overcome these problems,
a sequential approach, as shown in [7], has proven to be very effective by increasing
the robustness of the MBDoE against parametric uncertainties [58–60].
In model-assisted DoE, a process-related target (i.e., product titer) is optimized,
and the model supports in the evaluation and recommendation of DoE designs. A
structure for a model-assisted DoE concept is shown in Fig. 2. At first, a mathematical process model is used to describe, e.g., the growth, the substrate, and metabolite
concentrations as well as the productivity of a specific cell line. Therefore, the model
is adapted to first cultivation data (Fig. 2, Box 1), e.g., based on literature and/or
existing knowledge. The evaluated data should be used to cover typical known
effects, e.g., inhibitions or limitations. Certainly, the number of experiments that can
be performed at this stage, preferably in small scale, such as shaking flasks or deep
well plates [11–13], is usually limited. However, only a few experiments are
required to generate the mathematical model, as shown in the case study (see Sect.
4). Accordingly, the number of experiments in mDoE is still less than the number of
experiments to be performed in statistical DoE. Based on these data, model
Digital Twins and Their Role in Model-Assisted Design of Experiments
39
model-assisted methods and “bad experiences” with them (e.g., due to unrealistic
expectations).
• There is a lack of method integration for a consistent development strategy, which
can be adapted to existing work processes, a suspected high (modeling) effort and
doubts about the transferability of methods and models to other processes.
• The qualification profile of the involved personnel does often not fit (required:
biotechnology, process technology, modeling, and statistics).
An additional challenge is the application of models on complex metabolic
pathways of mammalian cells regarding cell growth and product formation. In
addition, models targeting the metabolism of cell cultures demand more effort than
those applied in chemical or microbial processes. Even if mathematical models are a
promising tool for the development of stable processes that comply with the principles of QbD, examples have so far only been published in the field of product
purification and polishing [8]. Nevertheless, Möller et al. (2018) and Abt et al.
(2018) showed that model-assisted and model-based DoE methods have great
potential for the development of process strategies and makes the process development more knowledge-based [7].
General differences between model-based and model-assisted DoE methods are
due to the aim of the recommended experiments. Model-based DoE (MBDoE) [53–
55] is used to supply valid experimental data for a precise model structure and model
parameter identification, where the conventional statistical DoE could fail
[56]. Uncertainties are key information in MBDoE, as model and data imperfections
cause undesirable variations in model parameters and simulation results. This
variation drives the MBDoE methodology, where it manifests itself as optimal
experimental settings (e.g., measurement principle, sampling rate, inputs/stimuli)
and informative data [55, 57]. However, uncertainties cause a discrepancy between
computed and experimental outputs leading to suboptimal or even meaningless
experimental designs for model parameter adaptation. To overcome these problems,
a sequential approach, as shown in [7], has proven to be very effective by increasing
the robustness of the MBDoE against parametric uncertainties [58–60].
In model-assisted DoE, a process-related target (i.e., product titer) is optimized,
and the model supports in the evaluation and recommendation of DoE designs. A
structure for a model-assisted DoE concept is shown in Fig. 2. At first, a mathematical process model is used to describe, e.g., the growth, the substrate, and metabolite
concentrations as well as the productivity of a specific cell line. Therefore, the model
is adapted to first cultivation data (Fig. 2, Box 1), e.g., based on literature and/or
existing knowledge. The evaluated data should be used to cover typical known
effects, e.g., inhibitions or limitations. Certainly, the number of experiments that can
be performed at this stage, preferably in small scale, such as shaking flasks or deep
well plates [11–13], is usually limited. However, only a few experiments are
required to generate the mathematical model, as shown in the case study (see Sect.
4). Accordingly, the number of experiments in mDoE is still less than the number of
experiments to be performed in statistical DoE. Based on these data, model
Digital Twins and Their Role in Model-Assisted Design of Experiments
39