DESIGN OF ONLINE ESTIMATOR FOR CULTURE MONITORING AND MEDIA DEVELOPMENT FOR BORDETELLA PERTUSSIS

Abstract

Whooping cough, or pertussis, is a contagious respiratory infection. Sanofi Pasteur’s pertussis vaccine production involves fermentation using reactors of increasing sizes, where cells from each stage inoculate the next. The upstream product undergoes purification to extract the five antigens needed for the acellular vaccine. A major challenge is the variability in antigen yields, particularly for pertactin, which limits overall productivity. Despite controls for parameters like pH, temperature, agitation, aeration, and nutrient feed rates, there’s no real-time monitoring to confirm growth and productivity during the batch run. This often leads to undetected deviations and potential yield losses, making fault detection methods essential for identifying variations and ensuring productivity. The current work deals with the development of a method of bioprocess monitoring using soft sensors for Bordetella pertussis cultures and studying changes in media composition that impact the yield of antigens. Accordingly, the following objectives were pursued: 1) Development of a real time fluorescence-based monitoring system. 2)Investigation of possible sources of oxidative stress and its impact or correlation with antigen production. 3) Development of a protocol for simultaneous monitoring of oxidative stress and antigen produced. 4) Development of a mechanistic model for use in model-based filtering of an online fluorescence based sensor. This thesis presents the design and implementation of an in-line fluorescence spectroscopy system capable of real-time monitoring of critical bioprocess parameters. By utilizing dual excitation wavelength fluorometry, this method was able to track the dynamics of biomass, amino acids, and antigen production throughout the fermentation process. The fluorescence data, combined with statistical technique such as Partial Least Squares (PLS) regression was used for predicting key state variables. A significant enhancement in the predictive accuracy of the PLS models was observed when the models were calibrated for the bacterial strain and the media composition used in each cultivation process. Previous research by our group (Vitelli et al. [2023c], Zavatti et al. [2016]) indicated a potential link between oxidative stress and antigen yield variability, particularly for pertactin. We hypothesized that glutamate, the most abundant carbon source, could be causing this oxidative stress. To test this, we developed a multi-parametric flow cytometry method to simultaneously monitor intracellular ROS and pertactin surface expression. Given that pertactin is an auto-transporter protein, understanding the relation between expression and secretion was crucial. Our findings showed a negative correlation between oxidative stress and pertactin surface expression. Using a tailored protein quantification method via affinity chromatography, combined with multi-color flow cytometry, we confirmed that higher glutamate concentrations induced higher oxidative stress and result in reduced pertactin secretion. The studies inlcuded both batch and fed batch bioreactor experiments where the latter closely emulated the production environment at Sanofi. Hence, the findings demonstrated that variability in initial glutamate concentrations can have a major impact of productivity and may partially explain the variability observed in the manufacturing process. The model from (Vitelli et al. [2023c]) was adapted to formulate a hybrid model that was used for model-based filtering of an online sensor. This mechanistic model incorporated the interactions between glutamate, ROS, and NADPH in neutralizing ROS. After calibration, it accurately predicted key variables under different oxidative stress conditions. In parallel, PLS regression models were developed using in-line fluorescence spectra which could predict OD, glutamate, and NADPH. By using a hybrid model that combines the mechanistic and PLS regression models via an Extended Kalman Filter, more accurate real-time estimates of key variables were obtained. The Akaike Information Criteria (AIC) confirmed that this hybrid model achieved a superior balance between complexity and accuracy as compared to purely mechanistic or PLS models.