(531f) Efficient Bioprocess Development across Scales Using Multi-Fidelity Batch Bayesian Optimization

Bioprocess model for scale- and clone-dependent analysis

A key component of our approach is a customized bioprocess model for Chinese hamster ovary cells that enables systematic evaluation of process conditions across scales. The model, adapted from Craven et al. [2], captures scale-dependent process dynamics and allows us to test different scale-up scenarios, particularly cases where small-scale systems such as microtiter plates fail to represent pilot-scale performance (Figure 1).

By adjusting noise levels, process kinetics, and feeding strategies, the model simulates realistic scale-up challenges. This enables a more data-driven comparison of different experimental strategies, helping to identify conditions where small-scale results are misleading and how it might affect the efficiency of process optimization. Additionally, the mechanistic nature of the model allows for realistic implementation of cell line variability, such as differences in lactate metabolism or inhibition by side products. The combination of bioprocess modeling and in silico data generation provides a robust framework for comparing traditional bioprocess development strategies with our BO-based approach.

Multi-fidelity batch Bayesian optimization for bioprocess development across scales

Bayesian optimization (BO) is particularly suited to optimize bioprocesses because real-world screening with cell lines is costly and time-consuming, typically requiring days to weeks per experiment. To further address the challenges of experimentation across scales, we introduce a multi-fidelity batch BO framework for bioprocess development. Our approach extends classical BO by explicitly considering multiple experimental scales (multi-fidelity), allowing for dynamic selection of the scale at which experiments should be conducted (Figure 1A). In addition, batch BO enables consideration of different batch sizes, which is important to integrate parallelized screening stages such as microtiter plates vs. individual bioreactors at later stages. We leverage in silico data generated from the aforementioned customized bioprocess model for mammalian cell lines to evaluate the method. In an iterative approach, the bioprocess model is used to simulate data for suggested cell lines, process conditions and scales. Critically, the model is used to incorporate different noise levels and feeding strategies across scales, thus better reflecting real laboratory scenarios.

Case studies and comparison with heuristic methods

Throughout the work, different case studies are considered, including the influence of different batch sizes, the number of clones and the assumed changes in process behavior between different scales. Using these case studies to mimic different scenarios in industrial bioprocess development, we evaluate how the BO strategy compares to conventional heuristics, where most cell lines are discarded early. Specifically, we analyze the impact of the multi-fidelity batch BO approach on the simulated product, evaluating the number of experiments that are required to reach certain titers and confidence levels in the prediction.

Figure 1 shows one of these case studies, in which multi-fidelity batch BO for experiments at three scales (microtiter plate, medium-scale reactor and a pilot scale) is compared to a classical, linear screening where the scale-up is done sequentially after a fixed number of experiments. Our results show how the usage of experiments across all scales early on reduces cost and increases the product titer with fewer experiments. By investigating different scenarios, our findings highlight how early-stage use of diverse fidelities can streamline optimization and improve decision-making in bioprocess development.

Industrial relevance and future applications

The combination of cell line screening and process optimization across scales remains a challenge with high industrial relevance, which is currently not sufficiently addressed in research. Current heuristic-driven approaches often lead to inefficient experimental allocation and suboptimal scale-up decisions. This work provides a new approach that takes various features of real-world experiments into account: mixed-integer problems (cell lines and process conditions), different fidelities of experimental setups and batch-wise experimentation.

By reducing the number of required experiments while maintaining or improving product titers, our approach offers significant potential for industrial application. Future work could further explore integrating additional information about cell lines, including their genetic characteristics, using embeddings to enhance predictive accuracy. This study marks a step towards more efficient, systematic screening strategies, moving beyond the current heuristic-driven status quo in bioprocess development.