Comparison of mechanistic and hybrid modeling approaches for characterization of a CHO cultivation process: Requirements, pitfalls and solution paths

Abstract

Despite the advantages of mathematical bioprocess modeling, successful model implementation already starts with experimental planning and accordingly can fail at this early stage. For this study, two different modeling approaches (mechanistic and hybrid) based on a four-dimensional antibody-producing CHO fed-batch process are compared. Overall, 33 experiments are performed in the fractional factorial fourdimensional design space and separated into four different complex data partitions subsequently used for model comparison and evaluation. The mechanistic model demonstrates the advantage of prior knowledge (i.e., known equations) to get informative value relatively independently of the utilized data partition. The hybrid approach displayes a higher data dependency but simultaneously yielded a higher accuracy on all data partitions. Furthermore, our results demonstrate that independent of the chosen modeling framework, a smart selection of only four initial experiments can already yield a very good representation of a full design space independent of the chosen modeling structure. Academic and industry researchers are recommended to pay more attention to experimental planning to maximize the process understanding obtained from mathematical modeling.

Hybrid modeling in bioprocess dynamics: Structural variabilities, implementation strategies, and practical challenges

“…A two‐norm regularized objective function and cross‐validation can be adopted to avoid overfitting and identify the optimal number of neurons and regularization parameters. Similarly, a single hidden layer ANN (with hyperbolic tangent and linear transfer functions) was chosen to model specific rates of biomass, viable cell, substrate and product formation in fed‐batch production of interleukin‐8 antibody by Chinese hamster ovary (CHO) DP‐12 cell (Bayer et al, 2023). Adopting a serial hybrid structure, the ANN models the specific rates utilizing four operating parameters and predictions of the previous time step of each response variable as input.…”

“…The regularization drives the values of those parameters that are not contributing to the model to zero, promoting a parsimonious structure (Stosch et al, 2018). The Levenberg–Marquardt method (LMM), is very effective in solving indirect training and identification, to estimate the respective specific rates (Bayer et al, 2021b; Bayer et al, 2023). The error back propagation can be performed by calculating the gradient of the squared errors ( E ) for the model state to the ANN weights, that is, $$\frac{\partial E}{\partial w}$$(Oliveira, 2004; Pinto et al, 2019).…”

“…Experimental conditions for fed‐batch cultivations can be formulated with model‐assisted DoE of feeding strategies for seamless integration into a hybrid model development routine. For example, Bayer et al, (Bayer et al, 2023) performed four initial experiments, and a full DoE with 29 experiments varying the glucose, and glutamine concentration in the feed, the bolus feed rate, and the feed start. The author adopted different data set labels in the training and validation exercise, using combinations of the initial four and fill DoE sets.…”

“…To fix an optimal number of neurons for three different empirical submodels in a hybrid modeling structure, Yang et al (2016) adopted a uniform design table for the combination of hidden neuron numbers and performed leave‐one‐out cross‐validation with 12 batches of modeling data. The random data partitioning in the cross‐validation process (except for the leave‐one‐out) can be repeated multiple times to average the cross‐validation error (Bayer et al, 2023).…”

“…Experimental conditions for fed-batch cultivations can be formulated with model-assisted DoE of feeding strategies for seamless integration into a hybrid model development routine. For example, Bayer et al, (Bayer et al, 2023) A moderate number of experimental batch/fed-batch studies can be performed in laboratories, where the generated data set can randomly be distributed (in a defined proportion) into training and testing subsets (Dong et al, 2014;Narayanan et al, 2019) to build the hybrid model. However, the dimension of the data set used to train and test the hybrid model can vary depending on the model complexity, data availability, and at the discretion of the modeler.…”