study, quantitative metabolomics was used to monitor the consumption of amino
acids in scale-down cultivations of Bacillus megaterium expressing green fluorescent protein. The metabolomics results from the scale-down conditions were then
used to design a better feed composition for the process [106].
It is important to note that a strain’s response to heterogenous environments in a
scale-down bioreactor is the total sum of its molecular level responses. What if a
particular response characteristic, such as accumulation of non-conventional amino
acids or de-activation of acetate cycling (accumulation of acetate) in E. coli could be
traced to particular metabolic pathways and mechanisms? What if data from a scaledown bioreactor could be used to trace specific metabolic fluxes of a clone as it is
exposed to various concentration gradients? Such information would be useful, not
only in strain engineering, but also in designing efficient processes at industrial
scale. Advanced modeling of scale-down data, i.e. fitting mechanistic and dynamic
metabolic models to data from scale-down cultivations can reveal specific pathways
that are active under given heterogeneous conditions. This is offcourse assuming that
the parameterization of the model includes such metabolic and physiological indices.
The flux terms to be fitted to the scale-down data should be an integral part of the
building of the cell model (Fig. 1). Again, Anane et al. [29] fitted data from parallel
scale-down cultivations of E. coli under multiple glucose gradient conditions to a
mechanistic model describing the process and the strain. The authors found that the
different responses of the strain to the gradients translated directly into different
values of the parameters of the model. Therefore, the model expanded on the primary
data of the scale-down experiment, and expanded the interpretation of the available
data for better process and strain design. In a similar study, Janakiraman and
co-workers used multi-variate data analysis techniques to interpret scale-down data
[107]. Their aim was to establish comparability between scale-down cultivations in
Ambr15
® minibioreactors and cultivations in 15,000 L manufacturing scale, by
applying principal component analysis to the scale-down datasets. By employing
this model-based approach, they were able to clearly identify that the runs in both
scales were statistically similar to each other, a conclusion that would have been
difficult to draw by looking at the raw scale-down data.
For digital twins to be applicable in this sense, there are a few pre-requisites the
model of the bioprocess must fulfill: (1) the parameter estimates in the cell model
must be subjected to rigorous validity and uncertainty tests, as presented by Anane
et al. [18]. The reported parameter values should always be accompanied by
confidence intervals at valid significance levels, to be able to derive biological
meaning from the model results. (2) The model should be just as detailed as is
necessary for its application. As pointed out by Gábor and Banga [108], the
parsimony principle should always be applied in building the model: i.e., the number
of parameters should not be more than those required to describe the process in its
simplest form [108, 109], and (3) the model should be constantly updated to include
the most recent research findings in cell physiology and metabolism. The physiological accuracy of the model should be ascertained by subject matter experts in the
field, which may not necessarily be the modeler (mathematician).
20
P. Neubauer et al.
acids in scale-down cultivations of Bacillus megaterium expressing green fluorescent protein. The metabolomics results from the scale-down conditions were then
used to design a better feed composition for the process [106].
It is important to note that a strain’s response to heterogenous environments in a
scale-down bioreactor is the total sum of its molecular level responses. What if a
particular response characteristic, such as accumulation of non-conventional amino
acids or de-activation of acetate cycling (accumulation of acetate) in E. coli could be
traced to particular metabolic pathways and mechanisms? What if data from a scaledown bioreactor could be used to trace specific metabolic fluxes of a clone as it is
exposed to various concentration gradients? Such information would be useful, not
only in strain engineering, but also in designing efficient processes at industrial
scale. Advanced modeling of scale-down data, i.e. fitting mechanistic and dynamic
metabolic models to data from scale-down cultivations can reveal specific pathways
that are active under given heterogeneous conditions. This is offcourse assuming that
the parameterization of the model includes such metabolic and physiological indices.
The flux terms to be fitted to the scale-down data should be an integral part of the
building of the cell model (Fig. 1). Again, Anane et al. [29] fitted data from parallel
scale-down cultivations of E. coli under multiple glucose gradient conditions to a
mechanistic model describing the process and the strain. The authors found that the
different responses of the strain to the gradients translated directly into different
values of the parameters of the model. Therefore, the model expanded on the primary
data of the scale-down experiment, and expanded the interpretation of the available
data for better process and strain design. In a similar study, Janakiraman and
co-workers used multi-variate data analysis techniques to interpret scale-down data
[107]. Their aim was to establish comparability between scale-down cultivations in
Ambr15
® minibioreactors and cultivations in 15,000 L manufacturing scale, by
applying principal component analysis to the scale-down datasets. By employing
this model-based approach, they were able to clearly identify that the runs in both
scales were statistically similar to each other, a conclusion that would have been
difficult to draw by looking at the raw scale-down data.
For digital twins to be applicable in this sense, there are a few pre-requisites the
model of the bioprocess must fulfill: (1) the parameter estimates in the cell model
must be subjected to rigorous validity and uncertainty tests, as presented by Anane
et al. [18]. The reported parameter values should always be accompanied by
confidence intervals at valid significance levels, to be able to derive biological
meaning from the model results. (2) The model should be just as detailed as is
necessary for its application. As pointed out by Gábor and Banga [108], the
parsimony principle should always be applied in building the model: i.e., the number
of parameters should not be more than those required to describe the process in its
simplest form [108, 109], and (3) the model should be constantly updated to include
the most recent research findings in cell physiology and metabolism. The physiological accuracy of the model should be ascertained by subject matter experts in the
field, which may not necessarily be the modeler (mathematician).
20
P. Neubauer et al.