Promoteur : Prof. Alain VANDE WOUWER

Summary

Mammalian cell cultures are increasingly used for the bioproduction of therapeutic
components, such as antibodies. However, due to the limited on-line instrumentation
of bioreactors, particularly as regards the key components’ concentrations
(amino-acids, etc) in the culture medium, as well as the diculty of establishing
reliable dynamic models, monitoring and controlling the quality of these products
remains arduous. This thesis explores various aspects of this challenging context,
from macro-modelling to metabolic flux analysis and from the development of software
sensors to the optimization and control of the culture conditions.
In this work, macro-modelling concepts are illustrated on data originating from
Vero cell cultures used for the production of a dengue vaccine. Model identification
theory is applied to build a model capable of describing and predicting the process
dynamics in a context of limited measured variables. Simulations of experiments
under unstudied operating culture conditions using the identified model highlight
potential opportunities for process optimization or design of future experiments.
Microscopic insights of the main cell metabolic pathways are provided in a study of
HEKcell cultures. Metabolic networks are usually under-determined systems that do
not always provide feasible solutions under the constraints imposed by the external
flux measurements. The use of interval representation of external fluxes is usually
used in classical Flux Variability Analyses (FVA) to ensure problem feasibility. An
original tool called Adaptive Flux Variability Analysis (AFVA) is developed based on
this concept. The proposed methodology always provides the solutions associated
with the narrowest external flux intervals and, consequently, the tightest limits for
the internal fluxes. In that study, the eciency of AFVA is demonstrated using
external fluxes often disregarded due to the strong constraints that they impose on
the system, namely biomass growth, urea production, pyruvate release/consumption
etc. Both macro- and micro-modelling approaches are linked in a third study using
another original methodology. The latter allows macroscopic models to be built by
selecting appropriate reactions among the set of elementary flux modes associated
with a chosen metabolic network. The fluxes leading to the smallest deviations from
the available datasets are identified for each model candidates by solving a linear
programming problem. The combination providing the smallest total deviation is
then selected. This methodology is used to build macroscopic models capable of
accurately describing available datasets of perfused hybridoma cell cultures. Finally,
an original cascade control strategy is successfully implemented on a real HEK cell
culture to simultaneously and independently regulate biomass and glucose levels
with very limited prior process knowledge. The strategy consists of estimating the
specific growth and the glucose uptake rates by means of sliding mode observers to
linearize the process dynamics which are then controlled using simple PI controllers.
In this thesis, dierent methodologies are successfully applied to monitor, predict,
optimize and control dierent process variables. It also provides a model reduction
technique applicable to metabolic networks to infer macro-models representing the
eective cellular pathways under given operating conditions. In that respect, further
studies should assess whether the proposed model reduction methodology is relevant
for the design of macro-models suitable for the control of specific culture conditions.