Development and Implementation of Process Analytical Technology (PAT) Systems in the Biopharmaceutical Sector – Challenges and Returns
Posted on August 15,2017
The Food and Drug Administration’s (FDA) process analytical technology (PAT) initiative encourages pharmaceutical manufacturers to use innovative analytical tools to improve the understanding and control of their manufacturing processes. In-line and on-line process analytics constitute a vital element of the PAT strategy. The utilization of PAT tools to monitor media composition and control specific feeding regimens for cell culture processes can lead to improvements in process understanding, process performance, as well as drug substance yield and product quality. In order to implement PAT as a part of the commercial manufacturing strategy for bio-processes, it is essential to understand the method development challenges as well as demonstrate return on investment from a business and product performance perspective.
GlaxoSmithKline is evaluating the implementation of Raman spectroscopy as a key PAT tool at the R&D stage including method development and prediction capabilities across varying process conditions, as well as media and different cell expression systems. The method development is based around utilizing datasets collected with widespread process variability and spiking studies, use of algorithms to delineate and orthogonalize spectral regions of interests for multiple constituents, and testing the feasibility of global and local calibrations. The resulting methods are being utilized for real-time monitoring of media components and control of feeding regimens for bio-processes. There is evidence that the implementation of spectroscopy as a process control toolbox gives us the capability to achieve improvements in process performance.
This extension of PAT for monitoring biological processes has demonstrated development and implementation challenges quite unique in comparison to small molecule-based drug product and processes. Challenges have been observed associated with the use of complex media systems, spectral overlap of nutrients and metabolites, balance of model accuracy and robustness, and a limited global modeling capability.
Mammalian cell culture processes are widely used in the pharmaceutical industry for the manufacture and production of a wide variety of biological drug products, such as antibodies, therapeutic proteins, and growth factors. A typical cell culture process is broken down into upstream and downstream components, wherein the upstream part represents the cell culture process i.e. inside the fermenter/reactor, while the downstream component involves processes such as clarification, purification, virus inactivation, sterilization, concentration, purification, and lyophilization. Intuitively, a successful cell culture process understandably depends on successful execution of the cell culture component. A cell culture process intended to manufacture bio-therapeutic proteins require balanced and consistent metabolic and cellular states to enable stable production.1 The efficacy of therapeutic proteins depend on certain post-translational modifications, for example Glycation.2 It represents the non-enzymatic addition of a reducing sugar to an amino acid residue of the protein, mostly at the N-terminal amine of proteins and the positive amine group of lysine.1 Other operational and mechanistic issues with the manufacture of therapeutic proteins include the culture productivity and controlling the amount of secreted lactate due to its impact on cell growth and culture’s performance.2
Data Informatics and PAT in Upstream Bioprocess Development: Current and Future
The quality by Design (QbD) initiative by the FDA maintains that quality be built into the process rather than being tested in the final product. QbD philosophy identifies, monitors, and controls all critical process parameters (CPP) which impacts the final quality of the product. The FDA’s PAT initiative encourages pharmaceutical manufacturers to use innovative analytical tools to monitor media composition and control specific feeding regimens for cell culture processes. These initiatives can lead to improvements in process understanding, process performance, and higher protein yield and quality. Concurrent with this need is the development of process analyzers, data management systems, and advanced process control solutions.
Common CPPs with mammalian bioprocess include the pH, culture temperature, dissolved oxygen (DO), nutrient set point, feed composition, and feed frequency. Also included are chemical attributes such as Glucose, Lactate, Carbon Dioxide, and Ammonium concentration. The biochemical properties such as viable cell count (VCC) and Viability are CPP’s indicating the state of cell culture health.3 An improved degree of control over these CPP’s can reduce process variability and improve therapeutic protein drug product quality as indicated by Titer and Glycosylation. Glycosylation has shown to be impacted by glucose, ammonia, and DO levels.1,4 Similarly, cell growth and culture longevity depend on temperature and pH.5
Within the biopharmaceutical space, the monitoring and control of CPPs is limited to a few basic parameters, such as pH, temperature, and DO utilizing established in-situ sensor technology. Other data sets are measured offline at a very low frequency (1-2 time a day), such as cell properties , nutrients, and metabolites. Such low frequency and long intervals between sampling and analysis times make the offline measurements unsuitable for process control. However, in the last decade or so, significant improvements have occurred with respect to biosensors, chemometric, HPLC, spectroscopy, capacitance probes, and in-situ microscopy. These soft sensors enable operational flexibility such as preserving sterility and high sensitivity to low level of analyte, and stability over long time periods.
Application of Raman Spectroscopy as a PAT Tool
In order to implement PAT as a part of the commercial manufacturing strategy for bio-processes as an end goal, it is essential to understand the method development challenges as well as demonstrate return on investment from a business and product performance perspective. Raman was utilized as an analytical tool to investigate its potential of enabling real-time monitoring of a nutrient’s consumption (Glucose) and a key metabolite generation (Lactate). Several indicators of analytical suitability were tested, such as “global” vs. “local” modeling capability, management of variability across media and cell lines, and various chemometric approaches to maximize method specificity. Raman is now more commonly applied in the biopharmaceutical space, hence it becomes essential to understand the various method development attributes, and how they impact method applicability. Spectral data was collected using a commercial Raman spectrometer, using a 785 nm excitation laser and collecting over a range of 100 – 3425 cm-1. The workflow also included a data link and a distributed control system (DCS) to utilize the real-time predictions of Glucose to enable a glucose pump and feed the solution to maintain a target level in the system. The method development and application workflow was as follows: Raman spectra was collected in real-time during the culture using in-situ probes followed by matching with offline collected reference values; chemometric methods were used to develop multivariate prediction methods; models were applied in real-time for future batches to understand the model applicability; integrating with DCS to monitor and control the bioprocesses at a target Glucose set-point. The study followed an iterative pattern, where the method was first developed on ‘x’ number of batches followed by predicting the features of ‘y’ future batches. Later, the model dataset was updated by including the ‘y’ batches into the ‘x’ dataset. The cell culture batches were carried out with variability in batch media composition, the cell line/therapeutic protein, the set points of various CPPs (DO, Temperature, pH), feed set point, Raman spectrometer, and probes. For the study, two Raman spectrometers with four probes each were included, thus incorporating widespread method variability. The entire workflow is depicted in Figure 1.
The media were chemical defined and developed in house and the cell cultures were 2 L working volume. The basic modeling algorithm was based on Partial Least-Squares (PLS), followed by model updates and optimizations at each stage of model application.
As a part of investigating the applicability of Raman spectroscopy as a successful PAT tool to accurately and robustly predict biomarkers, several sub-studies were performed, as detailed here:
Study 1: Raman Model Development on a Specific Media/Cell Line System
The first study was based on routinely carrying out the cell culture processes with little process variability. This was intended to simulate a routine manufacturing process, where little run-to-run variations are incorporated typically. The data from the spectrometer probes was included in a single calibration set as shown in Figure 2.
The PLS model created displayed good linearity and high accuracy to predict Glucose and Lactate concentrations as shown in Figure 3.
Study 2: Raman Model Application
The created calibration model was used to predict the trends of Glucose and Lactate for two future batches and as depicted in Figure 4, excellent prediction capabilities were observed with prediction errors of 0.61 and 0.55 g/L for Glucose, and 0.18 and 0.20 g/L for Lactate.
Study 3: Set-Point Control of Glucose using Raman Spectroscopy
The developed calibration methods were also used for set-point control of Glucose as shown in Figure 5 at 6, 4, and 2 g/L. Set-point control is expected to improve batch-to-batch consistency and an improved control over cell health. For the 2 g/L set point, some poor predictability was observed at the end of the run. This was attributed to low Glucose levels, which led to model extrapolations.
Study 4: Improved Set-Point Control of Glucose using Raman Spectroscopy
The set-point control was improved by adding batches with lower Glucose concentration to update the calibration, as shown in Figure 6.
The updated model’s predictions are shown in Figure 7. The addition of low Glucose spectra improved model predictability at 2 g/L indicating the dependence of model applicability on the calibration range.
Study 5: Model Improvement Strategies for a Different Cell Line
The Raman model development and implementation strategy (as used in Study 1) was tested on a separate media and cell line system. Similar to study 1, this involved collecting spectral data from several repeat routine cell culture batches for this specific media and cell line combination. The model was developed and implemented for future batch Glucose monitoring. In contrast to what was observed in studies 1 and 2, this study demonstrated suboptimal model performance which presented a unique challenge to create accurate and robust prediction model. The prediction of the model on four future runs is shown in Figure 8. This demonstrated that a simple PLS model would not always work across different media/cell line systems and necessitate certain chemometric model optimizations.
Study 6: Raman-Based Glucose Model Improvements using Chemometric Approaches
Two approaches were used for improving the model performance: 1) increasing the number of latent variables (LV), 2) wavelength range truncation, performed by including variables with a correlation coefficient between their absorbance and Glucose > 0.05. Table 1 represents the model statistics obtained after these two model optimization approaches.
The model performance after these two optimizations was tested on future batches. It was observed that the increase in LVs (Figure 9) was a better strategy compared to wavelength truncation (Figure 10).
With wavelength truncation, method accuracy was improved (table 1), however the model suffered on robustness aspect. An increase in LVs improved model’s accuracy as well as robustness to predict a future batch.
Study 7: Interference amongst Glucose and Lactate: Genetic Algorithms
The higher LV model as created in Study 6 when utilized to monitor Glucose for a system where Lactate was also controlled at a set point, depicted reduced model performance as shown in Figure 11.
For the instances where Lactate was desired at a specific set point, the Glucose predictions suffered. This indicated the need for a right balance of model’s specificity and impact of interferences. In a typical cell culture profile (Figure 12), lactate first accumulates and then consumed so its levels eventually approach zero.
For the studies where lactate was maintained at a constant level throughout the run, the presence of lactate caused this interference and hence reduced Glucose prediction. The test batches where the lactate content was higher were used for model correction.
Model correction was carried by the use of Genetic Algorithm (GA), which allows identification of a variable subset which maximizes model accuracy. From the dataset generated (containing high lactate data as well), the algorithm was executed to obtain the set of wavelengths which would maximize Glucose prediction in the presence of interferences as well. Figure 13 shows the wavelengths selected after the execution of the algorithm. The locations where a peak is observed were included in the model.
The Glucose predictions in a cell culture batch after model update is shown in Figure 14. After the use of GA, the prediction capability improved significantly even with lactate control.
Study 8: Global Calibration Capability: Model Applicability across Different Media/Cell Line System
The global capability of the Raman model for Glucose prediction was tested by extending the model created on a specific media/cell line (study 1) to predict the trends of : 1) same media/different cell line and 2) different media/different cell line. Figure 15 shows the implementation of the model with the change in cell line only with same media composition.
The set point control of Glucose was carried out at different targets and the model universality was evident based on the accurate predictions. In contrast, when the media composition was modified (Figure 16), the model applicability suffered, both for set point as well as bolus Glucose feed. This indicated that method is significantly sensitive to the media design, which limits its global applicability.
Raman spectroscopy is projected to be an analytical tool to enable real-time monitoring and control over bioprocess media components. Its projected to be integral to a future control strategy at GSK as an improved control is expected to yield improved product quality. It was observed that within the same media/cell line system, accurate models can be generated by including variability from a couple batches only, and that the model demonstrate good global/universal capability within the same media platform. With change in media; however, a reduced model performance was observed for predicting Glucose concentrations. It was also demonstrated that an appropriate balance of model specificity and interference from other constituents must be achieved, as shown by implementation of the GA for model optimizations. Overall, the method development for large molecules is typically more challenging relative to small molecule, since the extent of variability encountered is higher.
1. Berry BN, Dobrowsky TM, Timson RC, Kshirsagar R, Ryll T, Wiltberger K 2016. Quick generation of Raman spectroscopy based in-process glucose control to influence biopharmaceutical protein product quality during mammalian cell culture. Biotechnology Progress 32(1):224-234.
2. Matthews TE, Berry BN, Smelko J, Moretto J, Moore B, Wiltberger K 2016. Closed loop control of lactate concentration in mammalian cell culture by Raman spectroscopy leads to improved cell density, viability, and biopharmaceutical protein production. Biotechnology and Bioengineering 113(11):2416-2424.
3. Whelan J, Craven S, Glennon B 2012. In situ Raman spectroscopy for simultaneous monitoring of multiple process parameters in mammalian cell culture bioreactors. Biotechnology Progress 28(5):1355-1362.
4. Restelli V, Wang M-D, Huzel N, Ethier M, Perreault H, Butler M 2006. The effect of dissolved oxygen on the production and the glycosylation profile of recombinant human erythropoietin produced from CHO cells. Biotechnology and Bioengineering 94(3):481-494.
5. Trummer E, Fauland K, Seidinger S, Schriebl K, Lattenmayer C, Kunert R, Vorauer-Uhl K, Weik R, Borth N, Katinger H, Müller D 2006. Process parameter shifting: Part I. Effect of DOT, pH, and temperature on the performance of Epo-Fc expressing CHO cells cultivated in controlled batch bioreactors. Biotechnology and Bioengineering 94(6):1033-1044.
1. Manufacturing Simulation and Modeling Group, GlaxoSmithKline, King of Prussia, Pennsylvania, U.S.A. 19406
2. Process Analytical Technology (PAT) Group, GlaxoSmithKline, King of Prussia, Pennsylvania, U.S.A. 19406
3. Microbial Cell Culture Development (MCCD), GlaxoSmithKline, King of Prussia, Pennsylvania, U.S.A. 19406