Comparison of strategies for iterative model-based upstream bioprocess development with single and parallel reactor set-ups

The bioprocess industry shows an increased interest to use model-based approaches for upstream bioprocess development. Iteratively, one or multiple experiments need to be designed with the objective to iteratively learn the process behavior and drive it towards a desired state. Due to the inherent dynamic nature of upstream bioprocesses, dynamic modeling approaches are used to describe the evolution of the process state. This provides the opportunities to design dynamic changes in the control variables (process parameters) and to understand the influence of those control inputs on the process dynamics, which is particularly important should the model be used for process control. In this contribution, we compare different strategies for iterative dynamic model-based process development with single and parallel reactor set-ups. Using a simulated bioprocess, we show that most of the strategies quickly converge (typically within 5–6 iterations) on sub-optimal process conditions with satisfactory product concentrations in relation to the global optimum. Our results reveal significant differences in the optimization outcome depending on the strategy used for single and parallel reactor set-ups. Overall, more sophisticated strategies that involve a model validity measure seem to outperform those that purely seek to maximize the quantity of interest. Simply maximizing the upper prediction interval level underperforms significantly when compared to maximizing the median or the median with consideration of the model validity. The insights obtained from this study allow selecting the strategy for single or multi-reactor model-based process development. •Maximum quantity and iterations vary significantly in function of strategy used for single and parallel reactor set-ups.•Strategies considering a model validity measure outperform those that purely seek to maximize the quantity of interest.•Upper prediction interval limit maximization underperforms compared to median maximization with(out) model validity.