Chapter 14
Downloaded by PENNSYLVANIA STATE UNIV on August 19, 2012 | http://pubs.acs.org Publication Date: May 5, 1996 | doi: 10.1021/bk-1995-0613.ch014
Hybrid Process Modeling for Advanced Process State Estimation, Prediction, and Control Exemplified in a ProductionScale Mammalian Cell Culture M. Dors, R. Simutis, and A. Lübbert Institut für Technische Chemie, Universität Hannover, Callinstrasse 3, 30167 Hannover, Germany Indirect measurements, process state predictions and optimization can be largely improved by exploiting all available knowledge and data about a given production process. This concept has been implemented in an industrial process producing a recombinant protein in mammalian cell culture. The model has been used for indirect measurements and optimization of feeding profiles and was directly used for process control.
As is well-known in practice, the protein folding properties and, hence the product quality is dependent on the cultivation conditions of the host organism. Thus, in order to guarantee products within narrow specification limits, the production process must be kept under tight control (4). An essential prerequisite is accurate monitoring of the process' state and an optimization of the trajectories of the key process variables. Advanced control strategies require to predict the process behaviour at least over time horizons which are needed to influence the process so that the state variables will not escape from the acceptable intervals. Prediction, however, means that the process has to be modelled. In this article we describe a model used to supervise a process in which a recombinant protein is produced with mammalian cell culture. This model has been used to optimize the process trajectories for the production process. Structure of the model Production processes using genetically modified cell systems are not so thoroughly understood that they could be modelled comprehensively with physically based mathematical models only. On the other hand there is much heuristical knowledge available from many productions run in industrial practice. This can be described by correlations or at least by simple rules-of-thumb. The main idea behind the model to be described is to exploit all the knowledge and information which is available, may it be provided by means of mathematical models, heuristic knowledge or even by simple 0097-6156/95/0613-0144$12.00/0 © 1995 American Chemical Society
In Biosensor and Chemical Sensor Technology; Rogers, K., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 1996.
Downloaded by PENNSYLVANIA STATE UNIV on August 19, 2012 | http://pubs.acs.org Publication Date: May 5, 1996 | doi: 10.1021/bk-1995-0613.ch014
14.
145 Hybrid Process Modeling and Mammalian Cell Culture
DORS E T AL.
data sets from successful production runs only. The main problem is how to represent and process that knowledge and how to effectively combine the information from different sources so that the different pieces of a priori knowledge can be processed simultaneously. Shortly, the proposed approach fits the techniques used to monitor, predict and optimize a particular production process to the knowledge available. What is known fuzzily is treated accordingly, while what is known accurately will be treated by means of conventional mathematical models. The basic idea is first of all, to provide a frame for the model, making use of wellestablished knowledge and then to fill the gaps with mathematical, heuristical and data driven approaches, depending on the mode the a priori knowledge is available. The frame we use is the overall mass balance, which is usually represented by a set of ordinary differential equations for the main state variables. An example of a balance equation system for the fed batch culture of mammalian cells discussed here is: dX/dt
=r -^ X x
Bicmass, X , production - Dilution effect through feed dS/dt
= - r + γ (S - S) + ^ (S - S) - ^ S s
f
z
Glucose, S, consumption of the cells + Dilution effect of fresh medium + Addition through glucose-glutamin feed - Dilution effect through base addition dG/dt
=-r
G
+ Ç(G -G) + ^ ( G - G ) - f f G f
z
- K 0 Total feed, F, into the vessel = Feed, Ff, of fresh medium + Feed, F , of glucose-glutamine + Feed, Fi, of base addition. w
f
z
Downloaded by PENNSYLVANIA STATE UNIV on August 19, 2012 | http://pubs.acs.org Publication Date: May 5, 1996 | doi: 10.1021/bk-1995-0613.ch014
z
The most challenging terms in such a balance equation system are the specific rate μ for growth and the rates η for production and/or consumption of the i-th component. These are issues of current discussion and cannot be provided on the same level of accuracy as the balance equations themselves. Hence we use two sources of knowledge simultaneously: (i) The textbook approach of mammalian cell kinetics in form of modified Monod-type correlations and (ii) artificial neural networks trained on the data from the on-going production process. Figure 1 presents a scheme of the model architecture. The Monod-type approach used considers two limitation and two inhibition terms: μ = μη^χ Maximum specific growth rate S * S+K Limitation of glucose cone. Q * G + KG Limitation of glutamine cone. s
A K
A+ K
L
L + K
r
=
x
_
rs
-
Inhibition by ammonia A
Inhibition by lactate L
m X
