Automated Modeling and Identification of Bioprocesses
As in many other areas of application, model-based approaches of synthesis, control, optimization and supervision have been proposed as well for bioprocesses. However, decades after the first suggestion and the first lab trials the acceptance of model-based approaches is still rather limited in bio-industries. One of the main reasons is related to the models used. Up to a few years ago the majority of scientific papers devoted to bioprocess modeling dealt with unstructured models. This very simplistic picture of a cell describes growth and production as a function of the biomass concentration and the nutrients only. As only one biotic state is considered, the cell’s variable behavior as a result of the cell’s variable composition cannot be taken into account. Hence, model-based planning and optimization is bound to fail. Moreover, the results of hundreds or thousands of individual reaction steps have to be lumped into one kinetic expression for which no theoretical insight concerning its structure is available. On the other end, very detailed models have been built up in the area of systems biology recently. Starting with information about the genetic code, the protein composition, the metabolism, the regulation, etc., very detailed mechanistic models are formulated using simple kinetic expressions for the individual reaction steps. In the near future, however, these models will not be available for the enormous amount of organisms found in screening studies. For these cases mildly structured biological models in the form of compartment models are needed for model-based approaches which, at least partly, account for the variability in the cell’s composition.
Building up such models is a difficult task. Biology cannot answer the question which and how many compartments are needed for an approximate description of the growth and production behavior of a strain in a certain cultivation medium. Moreover, it is unknown from a theoretical point of view how individual reaction steps depend on these compartments. As a result, the modeler is faced with a combined structure and parameter identification problem. A systematic test, i.e., an identification of ‘all’ possible model candidates has to be ruled out due to the combinatorial character of the problem. As an alternative, a meaningful subset of possible model candidates should be tested automatically in an appropriate software tool. The selection and deselection of model candidates based on experimental data should exploit qualitative methods first. With such an approach significantly more models can be tested, i.e., ruled out as qualitative computations can run much faster compared to quantitative parameter identification. Only models which cannot be ruled out by qualitative features should then be transferred automatically to parameter identification. This approach resembles the way an experienced modeler would solve the task manually.
The talk will review some of the methods we have developed in this respect to automatically build up and identify structured and unstructured models of biological systems.