Halfway to Automated Feeding of Chinese Hamster Ovary Cells

Abstract

This paper presents a comprehensive study on the development of models and soft sensors required for the implementation of the automated bioreactor feeding of Chinese hamster ovary (CHO) cells using Raman spectroscopy and chemometric methods. This study integrates various methods, such as partial least squares regression and variable importance in projection and competitive adaptive reweighted sampling, and highlights their effectiveness in overcoming challenges such as high dimensionality, multicollinearity and outlier detection in Raman spectra. This paper emphasizes the importance of data preprocessing and the relationship between independent and dependent variables in model construction. It also describes the development of a simulation environment whose core is a model of CHO cell kinetics. The latter allows the development of advanced control algorithms for nutrient dosing and the observation of the effects of different parameters on the growth and productivity of CHO cells. All developed models were validated and demonstrated to have a high robustness and predictive accuracy, which were reflected in a 40% reduction in the root mean square error compared to established methods. The results of this study provide valuable insights into the practical application of these methods in the field of monitoring and automated cell feeding and make an important contribution to the further development of process analytical technology in the bioprocess industry.

Keywords: Raman; kinetic model; modelling; outliers; simulator; soft sensor; spectroscopy; variable selection.

Figure 1

Spectra obtained with Raman spectroscopy (from four different batches where only spectra, which are used for training and validation, are shown).

Figure 2

Raman spectra to which Savitzky–Golay filtering has been applied.

Figure 3

Raman spectra to which Savitzky–Golay filtering and SNV normalization are applied.

Figure 4

Finding the most appropriate number of latent variables in a PLS model.

Figure 5

Key variables determined with the methods VIP and CARS for the PLS model of glucose concentration.

Figure 6

The mean error and standard deviation for all spectra.

Figure 7

Validation of the PLS glucose model using VIP, CARS and outlier removal methods.

Figure 8

Validation of the PLS glucose model: comparison of experimental and predicted values using CARS and outlier removal methods.

Figure 9

Signal reconstruction of key process variables via PLS models.

Figure 10

Implementation of a simulator within the Simulink environment based on the CHO cell kinetics model.

Figure 11

Validation of the CHO cell kinetics model in the case of glucose concentration prediction for the entire batch run.

Figure 12

Validation of the CHO cell kinetics model: comparison of experimental and predicted glucose concentrations.