Generic Raman-Based Calibration Models Enabling Real-Time Monitoring of Cell Culture Bioreactors Hamidreza Mehdizadeh, David Lauri, Krizia M. Karry, and Mojgan Moshgbar Advanced Manufacturing Technology, Pfizer Inc., Peapack, NJ
Renee Procopio-Melino and Denis Drapeau Bioprocess R&D, Pfizer Inc., Andover, MA DOI 10.1002/btpr.2079 Published online April 18, 2015 in Wiley Online Library (wileyonlinelibrary.com)
Raman-based multivariate calibration models have been developed for real-time in situ monitoring of multiple process parameters within cell culture bioreactors. Developed models are generic, in the sense that they are applicable to various products, media, and cell lines based on Chinese Hamster Ovarian (CHO) host cells, and are scalable to large pilot and manufacturing scales. Several batches using different CHO-based cell lines and corresponding proprietary media and process conditions have been used to generate calibration datasets, and models have been validated using independent datasets from separate batch runs. All models have been validated to be generic and capable of predicting process parameters with acceptable accuracy. The developed models allow monitoring multiple key bioprocess metabolic variables, and hence can be utilized as an important enabling tool for Quality by Design approaches which are strongly supported by the U.S. Food and Drug Administration. C 2015 American Institute of Chemical Engineers Biotechnol. Prog., 31:1004–1013, 2015 V Keywords: Raman spectroscopy, multivariate modeling, bioprocess monitoring, CHO cell culture, process analytical technology
Introduction Biopharmaceutical processes and production of monoclonal antibodies Mammalian cell culture processes are widely used in the pharmaceutical industry for production of a large variety of biological products.1 These products include, but are not limited to, antibodies, therapeutic proteins, and growth factors, all serving as life-saving medicines to address an increasing number of human health complications. Chinese hamster ovary (CHO) cells are the most commonly used mammalian host for commercial-scale production of biological medicines.2 Despite the long history of using cell culture processes in pharmaceutical industry, manual fed-batch bioprocessing methodologies and inefficient bioreactor operations are the common and prevailing manufacturing method. As a result, recent advancements in process monitoring and control have not yet been fully utilized in the field, often leaving the manufacturing process using traditional methods that are based on offline sampling and manual feed calculations implemented by operators. Many groups have focused on implementation of Process Analytical Technology (PAT) tools3–5 and soft sensors6–8 to circumvent offline sampling and displace the use of manual feed, so as to increase the yield of the manufacturing process and comply with tight regulatory controls.
Correspondence concerning this article should be addressed to H. Mehdizadeh at [email protected]. 1004
The main reason advanced automatic control techniques are absent in cell culture bioprocesses and bioreactor operations is that these techniques require robust and reliable measurement methods to be available in place.9 However, online measurement in bioreactors in most cases is limited to few key process parameters including dissolved oxygen (DO), temperature, agitation, and pH. This leaves many important parameters such as concentrations of nutrients and metabolites, cell densities, and viability, unmeasured and uncontrolled or only manually controlled with large sample intervals (12–24 h). As a result, possible process upsets may be detected only after long delays, making it difficult to take corrective actions and increasing the risk of batch loss. Moreover, manual control increases the risk of human error, negatively affecting the processes while adding to the operators’ workload. Controlling glucose concentration in cell cultures through optimized feeding strategies is highly important and has been shown to be critical for increasing cell growth, and hence productivity.10,11 Regulators continue to require increasingly detailed and complex measurement of processes. Underfeeding will deprive the cells of an important nutrient and their main source of energy, while overfeeding leads to conversion of excess nutrients to lactate, which is an inhibitor of cell productivity.12,13 Due to lack of online glucose measurements, current feeding strategies are commonly based on daily offline samples or methods that measure a variable that is indirectly affected by glucose concentration as a result of cellular metabolic pathways or changes in cell numbers. One successful method was developed by Gagnon C 2015 American Institute of Chemical Engineers V
Biotechnol. Prog., 2015, Vol. 31, No. 4
et al. for control of lactate accumulation in Chinese Hamster Ovary (CHO) cells based on the culture’s pH.10 Using an indirect method they were able to efficiently eliminate lactate accumulation during the growth phase of fed-batch CHO cell cultures at both bench scale and large scale, leading to improved cell growth and cell specific productivity, and consequently yielding higher titers for various cell lines. In another study, Lu et al. used online capacitance measurements as well as Nova BioProfile Flex autosampler measurements for dynamic feeding.11 Both works reported substantial product concentration increase as the result of media and feed strategy optimization, via automatic feed rate adjustments based on time varying culture behavior. These results indicate the value of real-time monitoring for cell culture optimization. Raman spectroscopy has received great attention in recent years as a multipurpose real-time analytical technique.14,15 Testing the feasibility of Raman spectroscopy for measurement of glucose in bioreactors dates back to the late 1990s where limits of measurement of lower than 1 g/L have been reported.16,17 In more recent investigations, application of Raman-based probes for in situ measurement of process parameters in mammalian cell culture bioreactors have been studied, analyzing its practicality as a PAT tool for manufacturing of biologics.18 Abu-Absi et al. reported a Raman probe developed for in-line monitoring of cell culture metabolic parameters (concentrations of nutrients and metabolites) as well as viable and total cell densities.19 They correlated Raman spectra from three similar large scale (500 L) batches with offline Nova measurements of glucose, glutamine, glutamate, lactate, ammonium, viable cell density (VCD), and total cell density (TCD) using partial least squares (PLS) calibration models. In a similar study, Moretto et al. developed Raman-based calibration models using one CHO cell culture batch conducted in a 200 L bioreactor over 18 days.20 They used a separate batch from the same product for model validation; achieving acceptable RMSEP values. In another work, Whelan et al. reported Raman-based multivariate models for quantitation of multiple cell culture components, developed using small scale (3 L) bioreactor runs. They validated the Raman prediction models using the same cell lines and process conditions, both at 3 L scale and at a larger 15 L scale.21 In all these published works, model calibration and validation has been performed using the same process and cell line, though. Furthermore, other spectroscopy-based methods such as such as NIR, MIR, and fluorescence spectroscopy have been used to monitor CHO culture process parameters.22–24 Results from preliminary studies indicate Raman is an excellent candidate for process monitoring and inline measurement of cell culture process parameters. In this study, we build upon previous work to generate generic calibration models using a comprehensive data pool from several mammalian cell culture batches. These batches were run for various manufacturing conditions with multiple processes and at different scales, so as to test and validate the robustness and scalability of the developed models. Current feeding strategies Current feeding strategy in mammalian cell culture processes usually includes manual sampling from the bioreactor once or twice a day and measuring concentration values using available offline analytical methods (such as Nova
1005
BioProfile FLEX and YSI 2700 Glucose & Lactate Analyzer). These measurements are then used for manual adjustment of the feed values. Application of Raman for cell culture measurements Spectroscopic techniques such as Raman or near infrared (NIR) are analytical methods based on interaction of light and matter, and are increasingly used in the pharmaceutical industry.25 These methods are noninvasive and nondestructive while providing rich information about multiple process/metabolic variables, and therefore are excellent candidates for measurement of cell culture bioprocesses.26 Due to the complexity of collected spectra, chemometrics and multivariate analysis techniques are required to extract useful information from spectral signals and to remove unwanted instrument noise and reduce the inherent fluorescence. A number of recent studies have investigated Raman spectroscopy for online measurement of nutrients and metabolites in cell cultures,19,21 however, calibration models are developed and validated using the same process with the same product and at similar scales. Complexity of the matrix to be measured in cell culture processes results in interfering signals from present analytes, and hence it is required to develop large data pools for development of robust models, combined with chemometrics and multivariate data analysis techniques. Multivariate calibration is used extensively in the industry for fast on-line determination of important process parameters and critical quality attributes, enabling nondestructive measurement, on-line monitoring, and process control. These methods replace traditional reference methods that, despite high accuracy, are slow and labor intensive and commonly require offline sampling which can introduce systematic errors. Accurate multivariate calibration models based on Raman spectra can substitute current reference methods and enable a more efficient bioprocess. In addition, multivariate models allow identification of outliers and nonlinearities, both critical to assuring consistent and robust manufacturing strategies within statistical control. This publication reports recent advancements to develop a generic platform technology for in situ measurement of nutrients, metabolites, and cell densities, with a focus on glucose and lactate concentrations and viable cell density (VCD) within mammalian cell culture bioreactors. The constituents included in the study had been previously identified as key parameters for development of advanced dynamic feeding strategies.27 Generic multivariate models were developed to measure glucose and lactate concentrations as well as VCD during fed-batch mammalian cell culture processes. The final Raman models were generic in the sense that they are scale and process independent, and thus can be used at multiple scales and for various processes, regardless if the cell line had been included in the initial calibration dataset. In effect, these models are an enabling factor for moving from traditional quality-by-inspection methodology to a Quality-by-Design (QbD) methodology where critical process parameters (CPPs) are monitored and tightly controlled during the process. QbD approaches are based on gaining a thorough understanding of the effect of important process parameters on the critical quality attributes (CQAs) and identification of the sources of variability and controlling them within acceptable ranges to meet desirable product quality attributes.28 This methodology is strongly supported by the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA), and pharmaceutical companies
1006
Figure 1.
Biotechnol. Prog., 2015, Vol. 31, No. 4
Amplified Raman spectra by Savitzky-Golay second derivative showing relevant (A) glucose peaks for the CAO stretching and CAOAH bending modes, and (B) lactate peak for CH3 deformation. Color bar represents corresponding glucose concentrations (g/L).
Irvine, CA) that included pump and gas mass flow controller modules. Bioreactor pH was controlled near 7.0 by either a sodium carbonate solution or carbon dioxide. Dissolved oxygen was maintained in the range of 25 to 50% of air saturation by microbubble sparging of pure oxygen, and an overlay of 7% carbon dioxide/air was continuously delivered to the bioreactor. Two impellers, one Ruston and one axial-blade, supplied agitation to the bioreactor at a calculated power per unit volume of approximately 50 W/m3. Electric heating blankets maintained culture temperatures between 31 and 37 C. Bioreactors were fed a concentrated nutrient solution on a semicontinuous basis at rates calculated based on working volume. Additionally, a supplemental glucose solution was supplied in specific quantities to the culture in order to provide a wide range of glucose and lactate values to be used as calibration and validation sample sets. Samples collected during bioreactor runs included variation in process conditions of interest, i.e., different media types and constituent concentrations were used. The concentrations of measured constituents were varied systematically and simultaneously so as to avoid correlations between these variables. To address batch-to-batch variability and also matrix differences, data from various batches using different cell lines and media were incorporated in the calibration model. This approach increased flexibility and efficiency in applying developed models to new processes, minimizing model redefinition required and enhancing applicability as a robust analytical tool which could be utilized in a GMP environment. Analytical methods
Seven CHO monoclonal antibody (mAb) secreting cell lines derived from the same host were used for all experiments. Cell lines were stably transfected with proprietary DNA vectors to express different mAbs. All media and feeds used are proprietary solutions developed by Pfizer. Variations of two different platform media have been used in the calibration dataset. The experiments were carried out in a laboratory fed-batch cell culture bioreactor setup specifically developed for enabling systematic variation of process parameters of interest following carefully designed experimental protocols. Data from seven lab scale batches and one pilot scale batch were used as the calibration dataset for model development. Independent lab and pilot scale batches were used for model validation. To enhance model performance, process samples and spiked glucose and lactate samples were added to the calibration dataset.
Offline gas and pH values were quantified using RapidLab 248 blood gas analyzers (Siemens Healthcare Diagnostics, Inc., Deerfield, IL). Cell counts, viability, osmolality, and metabolic and electrolyte concentrations were analyzed using a BioProfile Flex Analyzer (Nova Biomedical, Waltham, MA). An in-line automated sampling system (Nova Biomedical, Waltham, MA) was used to analyze culture samples as frequently as every 30 min. Using the autosampler resulted in building large data matrices enabling robust calibration models. In parallel with in-line autosampling, a stainless steel immersion probe was used to collect and transfer Raman signal through fiber optic MR probe to a RXN2 Raman analyzer (Kaiser Optical Systems, Inc., Ann Arbor, MI), collecting Raman spectra used for in situ analysis of metabolite concentrations. A laser with 785 nm excitation wavelength was used with an approximate power of 200 mW at probe tip. Raman spectra were acquired implementing cosmic ray removal and dark spectrum subtraction with an exposure time of 10 s, adding 75 scans consecutively to result in a collection time slightly above 750 s (considering instrument overhead time). The interval for collecting Raman spectra was variable between 30 min and 1 h, depending on corresponding intervals for collecting offline samples using Nova remote valve machines (RVM). iCRaman software version 4.1 (Kaiser Optical Systems, Inc., Ann Arbor, MI) was used for spectral signal acquisition and export in the format of spc files.
Equipment
Chemometrics modeling
Cells were cultured in 3 L working volume Applikon bioreactors (Applikon, Inc., Schiedam, Netherlands) operating with BioNet modular controllers (Broadley-James Corp.,
Exported spectral files where analyzed and processed using Umetrics’ SIMCA software version 13.0.3.0. Calibration spectra were fundamentally analyzed for peak
are highly encouraged to invest in advanced methods that enable eliminating batch-to-batch variations leading to robust processes with consistent product quality and increased productivity.
Materials and Methods Mammalian cell culture and bioreactor runs
Biotechnol. Prog., 2015, Vol. 31, No. 4
1007
Figure 2. (A) Raw Raman data showing large amount of baseline drift. (B) Pretreated Raman data in glucose finger print spectral region. Baseline drift can be removed using pretreatment. Color bar represents glucose concentration values.
assignment, preprocessed and modeled via partial least squares (PLS) regression. First, the vibrational frequencies of the functional groups of the individual constituents, namely glucose and lactate, were identified in spectra with amplified signals. This amplification was carried out by computing a Savitzky-Golay (SG) second derivative with 25 smoothing points and a second order polynomial. For glucose, the stretching of the CAO carbon monoxide bond at 1072 cm21 as well as the CAOAH bending at 1128 cm21 were monitored.29 As seen in Figure 1A, there was a shift of the CAO bond to lower wavenumbers while the CAOAH bending peak at 1128 cm21 greatly correlated with changes in glucose concentration. It is known that anaerobic conditions within the cells promote the reduction of pyruvate, a glucose product, to lactate and NAD1. It can be seen from Figure 1B that Raman spectroscopy allows tracking of these redox reactions and further enables the quantitation of lactate in the media by monitoring the CH3 deformation peak of lactate at 1455 cm21.29 Following peak assignment, spectral data were modeled via Principal Component Analysis (PCA) and PLS so as to validate the spectral regions that accounted for most of the
variation and to identify the outliers. The pretreatment combination of a second-order polynomial Savitzky–Golay derivative over 25 points (1 point 5 1 cm21) followed by standard normal variate (SNV) and Savitzky–Golay smoothing applied over 25 points enhanced chemical differences between the samples whilst reducing backscattered fluorescence. These pretreatments resulted in the most accurate calibration models. All spectral data were mean centered before model development. Figure 2 shows Raman spectral data from a number of collected samples, before (A) and after (B) preprocessing. Figure 2B includes the pretreated Raman spectra in the identified glucose region of 1128 cm21 and a color bar that shows the direct relationship between the height of the peak and the constituent concentration. The final PLS quantitation model included 900 data points. Samples were mainly acquired from seven small scale (1 and 3 L) bioreactor runs and one pilot scale (500 L) bioreactor run. Various CHO-based cell lines and proprietary media were used in fed-batch runs to develop models that were independent of specific cell line or media composition. Twenty data points included fresh media with spiked amounts of glucose and lactate (no cells in the samples). Separate PLS models were developed for glucose, lactate, and viable cell density. An independent validation dataset from a batch not used for model calibration was used to evaluate model predictions and to optimize model parameters such as number of principal components, spectral range, and pretreatment method used. After determination of optimal model parameters, the developed models were tested using two additional validation datasets, one from a batch in 500 L Pilot scale to test for model scalability, and the other from a small scale bioreactor run with a CHO cell line not used previously in model development, to verify generic models are applicable to new CHO cell lines used for cell culture processes. The results of model calibration and validation are presented in Results and Discussion section.
Results and Discussion Model calibration and validation Model identification and validation results are summarized in Table 1. Specifically, number of samples and principal components, values of coefficient of determination (R2Y), as well as values for errors of calibration and prediction are reported for glucose, lactate, and VCD models. Careful attention has been made to ensure the calibration range encompasses the operating range of process parameters as multivariate regression models are not suitable for extrapolating. For all three modeled parameters, the values of the root mean square error of evaluation (RMSEE), cross validation (RMSECV), and prediction (RMSEP) are comparable and in the same range, indicating stability of developed models. The root mean square error (RMSE) is a frequently used measure of the difference between model predictions and actual or reference measurement values. Root mean square errors were calculated using the following equation in which
Table 1. Summary of Model Calibration and Validation Component Glucose (g/L) Lactate (g/L) VCD (E5 cells/mL)
N
# Factors
R2 Y
RMSEE
RMSECV
Range
RMSEP
900 896 884
6 5 9
0.938 0.954 0.936
0.32 0.19 13.61
0.33 0.20 14.17
0.0–8.45 0.0–3.95 0–300
0.43 0.26 19.82
1008
Biotechnol. Prog., 2015, Vol. 31, No. 4
Figure 3. Model calibration plots comparing reference measurements versus model prediction (A, C, and E), as well as time series plots of reference measurements and model predictions (B, D, and F) for the calibration dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models. Arrows indicate high concentration lactate samples for lactate model and cell-free standard media samples for VCD model.
ymodel;i are the values predicted by the model and yref;i are the actual values known or obtained from reference measurement method. RMSE ¼
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Xn 2 ðy 2yref;i Þ i¼1 model;i n
It is observed that developed models are performing with acceptable error, having high R2Y values (above 0.90 for all models) and low RMSEP values (0.43 g/L, 0.26 g/L, and 19.82 E5 cells/mL for glucose, lactate, and VCD predictions, respectively), indicating good model fit. Figure 3 illustrates model predictions plotted against reference measurements as well as time series plots of these parameters for the calibration dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
In general, model fit to the data is good. Exceptions are high concentration lactate samples for the lactate model and cellfree standard media samples for the VCD model. Although model fit was less favorable for these samples, our investigations show that including these samples in the dataset improved accuracy of model predictions as reflected by lower RMSEP values. Models were validated with data from an independent small scale batch using a mAb-producing cell line previously used in model development. Figure 4 shows model validation results, comparing model predictions with reference measurements and corresponding RMSEP values as well as time series plots of these parameters over batch time for three models developed. Glucose model predictions were in good agreement with reference measurements throughout the run, closely following the trend of reference
Biotechnol. Prog., 2015, Vol. 31, No. 4
1009
Figure 4. Model validation plots. Data generated from an independent batch run in a 3 L bioreactor using a mAb-producing cell line previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
Table 2. Comparison of RMSEP Values for Glucose, Lactate, and VCD Models RMSEP 1 RMSEP 2 RMSEP 3 Glucose (g/L) Lactate (g/L) VCD (E5 cells/mL)
0.43 0.26 19.82
0.28 0.072 30.87
0.44 0.41 13.95
RMSEP 1, 2, and 3 refer to standard errors in prediction of the original validation batch (accuracy validation), the batch with previously used cell line in large scale (500 L) bioreactor (scalability validation), and the batch using a cell line not included in model development in small scale (5 L) bioreactor (generic model validation).
measurements, with the exception of larger residuals towards the end of the batch (Figures 3A,B). This effect is commonly observed in batch runs due to increase in florescence excitation towards the end of the batch as a result of
increased dead cell population and accumulation of cell debris in the culture media. The lactate model was less accurate at the beginning of the batch run when higher lactate concentrations were present, predicting higher than actual values (Figures 4C,D). This is explained by the fact that most lactate data used in model calibration are in the lower end of the concentration range (Figure 3D). The difference between model predictions and reference values decreased towards the end of the run, resulting from accurate predictions at lower concentrations. Although the lactate calibration dataset included the full range of lactate concentrations, having a lower number of samples in the upper end of this range affected model predictions. This should be considered for future improvements of the models. Furthermore, careful attention should be made regarding accuracy of the models when predicting lower
1010
Biotechnol. Prog., 2015, Vol. 31, No. 4
Figure 5. Model scalability validation plots. Data generated from an independent batch run in a 500 L stainless steel bioreactor using a mAb producing cell line previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
concentrations. This is due to the fact that the model prediction error is in the range of 60.3 (g/L) and therefore model prediction percent errors are high in lower ranges (below 0.5 g/L). The VCD model over-predicted at the beginning of the run and under-predicted towards the end, while still predicting the trend in VCD changes with acceptable accuracy with an approximate RMSEP of 2.0 E5 cells/mL (Figure 4F). A number of VCD values obtained by the reference measurement method were outliers (around index number 80), so careful attention was given in data analysis to remove such measurements from calibration datasets. The VCD model uses a larger number of factors (9 compared with 6 and 5 for glucose and lactate models, respectively), which is due to increased complexity of predicting living cell counts compared with concentrations of individual sub-
stances present in the culture. This leads to noisier measurements, as evident from higher amplitude and frequency of the noise (Figure 4F). Model scalability and process independence validation In order to validate the generic capacity of the models, i.e. the models are applicable to various cell lines based on CHO host cells and their corresponding products, media compositions, and/or scales, they were used to predict glucose and lactate concentrations and VCD of two cell culture bioreactor batch runs. One of the batch runs was conducted using a cell line previously used in model development but now in a stainless steel 500 L pilot scale bioreactor, and the other was using a CHO-based cell line not included in model development and conducted in a 5 L glass bioreactor. Table
Biotechnol. Prog., 2015, Vol. 31, No. 4
1011
Figure 6. Generic model validation plots. Data generated from an independent batch run in a 5 L glass bioreactor using a protein producing cell line not previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
2 provides a comparison of RMSEP values for glucose, lactate, and VCD models using the three validation datasets, with RMSEP 1, 2, and 3 referring to standard errors in prediction of the original validation batch (accuracy validation), the batch with previously used cell line in stainless steel 500 L bioreactor (scalability validation), and the batch using a cell line not included in model development in 5 L glass bioreactor (generic model validation), respectively. Prediction errors of all three models are comparable and in the same range for all three validation batch runs, demonstrating consistency, stability, and scalability of the developed models. In case of the large scale bioreactor run, RMSEP 2 values are smaller for glucose and lactate models compared with the original validation batch run (RMSEP 1), indicating excellent model scalability. Figures 5 and 6 depict model predictions versus reference measurements and corresponding RMSEP values of
individual models for these validation batches. In Figure 5 it is observed that models accurately predict concentrations of glucose and lactate throughout the batch. VCD model predictions in this case are less accurate at the beginning of the batch, resulting in higher RMSEP 2 values for the VCD compared to the original validation batch. Figure 6 shows opposite behavior in the validation run using a new cell line not used in model calibration, where glucose and lactate models are predicting less accurately compared with VCD model. Figure 6D illustrates that the lactate predictions are specifically less accurate when high lactate concentrations are present, consistent with what observed previously with original calibration dataset (Figure 4D). In contrast, the VCD model predicts with excellent accuracy throughout the run, closely following trends of VCD changes observed using reference measurement method (Figure 6F).
1012
Biotechnol. Prog., 2015, Vol. 31, No. 4
Conclusion Raman spectroscopy provides great potential for real-time in situ measurement of key process parameters in fed-batch bioreactor operations. A set of generic calibration models based on Raman spectroscopy were developed to measure glucose and lactate concentrations and viable cell density within mammalian cell culture bioreactors. PLS regression models were used to correlate spectral peaks with reference measurements. To develop generic models with increased robustness, the calibration dataset included data from multiple batches of different products produced by CHO-based cell lines, and included changes in proprietary media. Models were validated using three different validation datasets: (1) using a cell line that had previously been included in model calibration dataset, (2) using the same cell line but predicting in a 500 L pilot scale bioreactor, and (3) a dataset from a cell line not used in model development. Model predictions using validation datasets were within acceptable ranges in general, providing evidence that generated models were independent of scale and applicable to new cell lines derived from CHO host cells. Model predictions of glucose and lactate concentrations are therefore suitable for on-line control of the rates at which glucose and other nutrients are fed to the culture. It is important to note that model accuracies are limited in lower concentration ranges (
Renee Procopio-Melino and Denis Drapeau Bioprocess R&D, Pfizer Inc., Andover, MA DOI 10.1002/btpr.2079 Published online April 18, 2015 in Wiley Online Library (wileyonlinelibrary.com)
Raman-based multivariate calibration models have been developed for real-time in situ monitoring of multiple process parameters within cell culture bioreactors. Developed models are generic, in the sense that they are applicable to various products, media, and cell lines based on Chinese Hamster Ovarian (CHO) host cells, and are scalable to large pilot and manufacturing scales. Several batches using different CHO-based cell lines and corresponding proprietary media and process conditions have been used to generate calibration datasets, and models have been validated using independent datasets from separate batch runs. All models have been validated to be generic and capable of predicting process parameters with acceptable accuracy. The developed models allow monitoring multiple key bioprocess metabolic variables, and hence can be utilized as an important enabling tool for Quality by Design approaches which are strongly supported by the U.S. Food and Drug Administration. C 2015 American Institute of Chemical Engineers Biotechnol. Prog., 31:1004–1013, 2015 V Keywords: Raman spectroscopy, multivariate modeling, bioprocess monitoring, CHO cell culture, process analytical technology
Introduction Biopharmaceutical processes and production of monoclonal antibodies Mammalian cell culture processes are widely used in the pharmaceutical industry for production of a large variety of biological products.1 These products include, but are not limited to, antibodies, therapeutic proteins, and growth factors, all serving as life-saving medicines to address an increasing number of human health complications. Chinese hamster ovary (CHO) cells are the most commonly used mammalian host for commercial-scale production of biological medicines.2 Despite the long history of using cell culture processes in pharmaceutical industry, manual fed-batch bioprocessing methodologies and inefficient bioreactor operations are the common and prevailing manufacturing method. As a result, recent advancements in process monitoring and control have not yet been fully utilized in the field, often leaving the manufacturing process using traditional methods that are based on offline sampling and manual feed calculations implemented by operators. Many groups have focused on implementation of Process Analytical Technology (PAT) tools3–5 and soft sensors6–8 to circumvent offline sampling and displace the use of manual feed, so as to increase the yield of the manufacturing process and comply with tight regulatory controls.
Correspondence concerning this article should be addressed to H. Mehdizadeh at [email protected]. 1004
The main reason advanced automatic control techniques are absent in cell culture bioprocesses and bioreactor operations is that these techniques require robust and reliable measurement methods to be available in place.9 However, online measurement in bioreactors in most cases is limited to few key process parameters including dissolved oxygen (DO), temperature, agitation, and pH. This leaves many important parameters such as concentrations of nutrients and metabolites, cell densities, and viability, unmeasured and uncontrolled or only manually controlled with large sample intervals (12–24 h). As a result, possible process upsets may be detected only after long delays, making it difficult to take corrective actions and increasing the risk of batch loss. Moreover, manual control increases the risk of human error, negatively affecting the processes while adding to the operators’ workload. Controlling glucose concentration in cell cultures through optimized feeding strategies is highly important and has been shown to be critical for increasing cell growth, and hence productivity.10,11 Regulators continue to require increasingly detailed and complex measurement of processes. Underfeeding will deprive the cells of an important nutrient and their main source of energy, while overfeeding leads to conversion of excess nutrients to lactate, which is an inhibitor of cell productivity.12,13 Due to lack of online glucose measurements, current feeding strategies are commonly based on daily offline samples or methods that measure a variable that is indirectly affected by glucose concentration as a result of cellular metabolic pathways or changes in cell numbers. One successful method was developed by Gagnon C 2015 American Institute of Chemical Engineers V
Biotechnol. Prog., 2015, Vol. 31, No. 4
et al. for control of lactate accumulation in Chinese Hamster Ovary (CHO) cells based on the culture’s pH.10 Using an indirect method they were able to efficiently eliminate lactate accumulation during the growth phase of fed-batch CHO cell cultures at both bench scale and large scale, leading to improved cell growth and cell specific productivity, and consequently yielding higher titers for various cell lines. In another study, Lu et al. used online capacitance measurements as well as Nova BioProfile Flex autosampler measurements for dynamic feeding.11 Both works reported substantial product concentration increase as the result of media and feed strategy optimization, via automatic feed rate adjustments based on time varying culture behavior. These results indicate the value of real-time monitoring for cell culture optimization. Raman spectroscopy has received great attention in recent years as a multipurpose real-time analytical technique.14,15 Testing the feasibility of Raman spectroscopy for measurement of glucose in bioreactors dates back to the late 1990s where limits of measurement of lower than 1 g/L have been reported.16,17 In more recent investigations, application of Raman-based probes for in situ measurement of process parameters in mammalian cell culture bioreactors have been studied, analyzing its practicality as a PAT tool for manufacturing of biologics.18 Abu-Absi et al. reported a Raman probe developed for in-line monitoring of cell culture metabolic parameters (concentrations of nutrients and metabolites) as well as viable and total cell densities.19 They correlated Raman spectra from three similar large scale (500 L) batches with offline Nova measurements of glucose, glutamine, glutamate, lactate, ammonium, viable cell density (VCD), and total cell density (TCD) using partial least squares (PLS) calibration models. In a similar study, Moretto et al. developed Raman-based calibration models using one CHO cell culture batch conducted in a 200 L bioreactor over 18 days.20 They used a separate batch from the same product for model validation; achieving acceptable RMSEP values. In another work, Whelan et al. reported Raman-based multivariate models for quantitation of multiple cell culture components, developed using small scale (3 L) bioreactor runs. They validated the Raman prediction models using the same cell lines and process conditions, both at 3 L scale and at a larger 15 L scale.21 In all these published works, model calibration and validation has been performed using the same process and cell line, though. Furthermore, other spectroscopy-based methods such as such as NIR, MIR, and fluorescence spectroscopy have been used to monitor CHO culture process parameters.22–24 Results from preliminary studies indicate Raman is an excellent candidate for process monitoring and inline measurement of cell culture process parameters. In this study, we build upon previous work to generate generic calibration models using a comprehensive data pool from several mammalian cell culture batches. These batches were run for various manufacturing conditions with multiple processes and at different scales, so as to test and validate the robustness and scalability of the developed models. Current feeding strategies Current feeding strategy in mammalian cell culture processes usually includes manual sampling from the bioreactor once or twice a day and measuring concentration values using available offline analytical methods (such as Nova
1005
BioProfile FLEX and YSI 2700 Glucose & Lactate Analyzer). These measurements are then used for manual adjustment of the feed values. Application of Raman for cell culture measurements Spectroscopic techniques such as Raman or near infrared (NIR) are analytical methods based on interaction of light and matter, and are increasingly used in the pharmaceutical industry.25 These methods are noninvasive and nondestructive while providing rich information about multiple process/metabolic variables, and therefore are excellent candidates for measurement of cell culture bioprocesses.26 Due to the complexity of collected spectra, chemometrics and multivariate analysis techniques are required to extract useful information from spectral signals and to remove unwanted instrument noise and reduce the inherent fluorescence. A number of recent studies have investigated Raman spectroscopy for online measurement of nutrients and metabolites in cell cultures,19,21 however, calibration models are developed and validated using the same process with the same product and at similar scales. Complexity of the matrix to be measured in cell culture processes results in interfering signals from present analytes, and hence it is required to develop large data pools for development of robust models, combined with chemometrics and multivariate data analysis techniques. Multivariate calibration is used extensively in the industry for fast on-line determination of important process parameters and critical quality attributes, enabling nondestructive measurement, on-line monitoring, and process control. These methods replace traditional reference methods that, despite high accuracy, are slow and labor intensive and commonly require offline sampling which can introduce systematic errors. Accurate multivariate calibration models based on Raman spectra can substitute current reference methods and enable a more efficient bioprocess. In addition, multivariate models allow identification of outliers and nonlinearities, both critical to assuring consistent and robust manufacturing strategies within statistical control. This publication reports recent advancements to develop a generic platform technology for in situ measurement of nutrients, metabolites, and cell densities, with a focus on glucose and lactate concentrations and viable cell density (VCD) within mammalian cell culture bioreactors. The constituents included in the study had been previously identified as key parameters for development of advanced dynamic feeding strategies.27 Generic multivariate models were developed to measure glucose and lactate concentrations as well as VCD during fed-batch mammalian cell culture processes. The final Raman models were generic in the sense that they are scale and process independent, and thus can be used at multiple scales and for various processes, regardless if the cell line had been included in the initial calibration dataset. In effect, these models are an enabling factor for moving from traditional quality-by-inspection methodology to a Quality-by-Design (QbD) methodology where critical process parameters (CPPs) are monitored and tightly controlled during the process. QbD approaches are based on gaining a thorough understanding of the effect of important process parameters on the critical quality attributes (CQAs) and identification of the sources of variability and controlling them within acceptable ranges to meet desirable product quality attributes.28 This methodology is strongly supported by the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA), and pharmaceutical companies
1006
Figure 1.
Biotechnol. Prog., 2015, Vol. 31, No. 4
Amplified Raman spectra by Savitzky-Golay second derivative showing relevant (A) glucose peaks for the CAO stretching and CAOAH bending modes, and (B) lactate peak for CH3 deformation. Color bar represents corresponding glucose concentrations (g/L).
Irvine, CA) that included pump and gas mass flow controller modules. Bioreactor pH was controlled near 7.0 by either a sodium carbonate solution or carbon dioxide. Dissolved oxygen was maintained in the range of 25 to 50% of air saturation by microbubble sparging of pure oxygen, and an overlay of 7% carbon dioxide/air was continuously delivered to the bioreactor. Two impellers, one Ruston and one axial-blade, supplied agitation to the bioreactor at a calculated power per unit volume of approximately 50 W/m3. Electric heating blankets maintained culture temperatures between 31 and 37 C. Bioreactors were fed a concentrated nutrient solution on a semicontinuous basis at rates calculated based on working volume. Additionally, a supplemental glucose solution was supplied in specific quantities to the culture in order to provide a wide range of glucose and lactate values to be used as calibration and validation sample sets. Samples collected during bioreactor runs included variation in process conditions of interest, i.e., different media types and constituent concentrations were used. The concentrations of measured constituents were varied systematically and simultaneously so as to avoid correlations between these variables. To address batch-to-batch variability and also matrix differences, data from various batches using different cell lines and media were incorporated in the calibration model. This approach increased flexibility and efficiency in applying developed models to new processes, minimizing model redefinition required and enhancing applicability as a robust analytical tool which could be utilized in a GMP environment. Analytical methods
Seven CHO monoclonal antibody (mAb) secreting cell lines derived from the same host were used for all experiments. Cell lines were stably transfected with proprietary DNA vectors to express different mAbs. All media and feeds used are proprietary solutions developed by Pfizer. Variations of two different platform media have been used in the calibration dataset. The experiments were carried out in a laboratory fed-batch cell culture bioreactor setup specifically developed for enabling systematic variation of process parameters of interest following carefully designed experimental protocols. Data from seven lab scale batches and one pilot scale batch were used as the calibration dataset for model development. Independent lab and pilot scale batches were used for model validation. To enhance model performance, process samples and spiked glucose and lactate samples were added to the calibration dataset.
Offline gas and pH values were quantified using RapidLab 248 blood gas analyzers (Siemens Healthcare Diagnostics, Inc., Deerfield, IL). Cell counts, viability, osmolality, and metabolic and electrolyte concentrations were analyzed using a BioProfile Flex Analyzer (Nova Biomedical, Waltham, MA). An in-line automated sampling system (Nova Biomedical, Waltham, MA) was used to analyze culture samples as frequently as every 30 min. Using the autosampler resulted in building large data matrices enabling robust calibration models. In parallel with in-line autosampling, a stainless steel immersion probe was used to collect and transfer Raman signal through fiber optic MR probe to a RXN2 Raman analyzer (Kaiser Optical Systems, Inc., Ann Arbor, MI), collecting Raman spectra used for in situ analysis of metabolite concentrations. A laser with 785 nm excitation wavelength was used with an approximate power of 200 mW at probe tip. Raman spectra were acquired implementing cosmic ray removal and dark spectrum subtraction with an exposure time of 10 s, adding 75 scans consecutively to result in a collection time slightly above 750 s (considering instrument overhead time). The interval for collecting Raman spectra was variable between 30 min and 1 h, depending on corresponding intervals for collecting offline samples using Nova remote valve machines (RVM). iCRaman software version 4.1 (Kaiser Optical Systems, Inc., Ann Arbor, MI) was used for spectral signal acquisition and export in the format of spc files.
Equipment
Chemometrics modeling
Cells were cultured in 3 L working volume Applikon bioreactors (Applikon, Inc., Schiedam, Netherlands) operating with BioNet modular controllers (Broadley-James Corp.,
Exported spectral files where analyzed and processed using Umetrics’ SIMCA software version 13.0.3.0. Calibration spectra were fundamentally analyzed for peak
are highly encouraged to invest in advanced methods that enable eliminating batch-to-batch variations leading to robust processes with consistent product quality and increased productivity.
Materials and Methods Mammalian cell culture and bioreactor runs
Biotechnol. Prog., 2015, Vol. 31, No. 4
1007
Figure 2. (A) Raw Raman data showing large amount of baseline drift. (B) Pretreated Raman data in glucose finger print spectral region. Baseline drift can be removed using pretreatment. Color bar represents glucose concentration values.
assignment, preprocessed and modeled via partial least squares (PLS) regression. First, the vibrational frequencies of the functional groups of the individual constituents, namely glucose and lactate, were identified in spectra with amplified signals. This amplification was carried out by computing a Savitzky-Golay (SG) second derivative with 25 smoothing points and a second order polynomial. For glucose, the stretching of the CAO carbon monoxide bond at 1072 cm21 as well as the CAOAH bending at 1128 cm21 were monitored.29 As seen in Figure 1A, there was a shift of the CAO bond to lower wavenumbers while the CAOAH bending peak at 1128 cm21 greatly correlated with changes in glucose concentration. It is known that anaerobic conditions within the cells promote the reduction of pyruvate, a glucose product, to lactate and NAD1. It can be seen from Figure 1B that Raman spectroscopy allows tracking of these redox reactions and further enables the quantitation of lactate in the media by monitoring the CH3 deformation peak of lactate at 1455 cm21.29 Following peak assignment, spectral data were modeled via Principal Component Analysis (PCA) and PLS so as to validate the spectral regions that accounted for most of the
variation and to identify the outliers. The pretreatment combination of a second-order polynomial Savitzky–Golay derivative over 25 points (1 point 5 1 cm21) followed by standard normal variate (SNV) and Savitzky–Golay smoothing applied over 25 points enhanced chemical differences between the samples whilst reducing backscattered fluorescence. These pretreatments resulted in the most accurate calibration models. All spectral data were mean centered before model development. Figure 2 shows Raman spectral data from a number of collected samples, before (A) and after (B) preprocessing. Figure 2B includes the pretreated Raman spectra in the identified glucose region of 1128 cm21 and a color bar that shows the direct relationship between the height of the peak and the constituent concentration. The final PLS quantitation model included 900 data points. Samples were mainly acquired from seven small scale (1 and 3 L) bioreactor runs and one pilot scale (500 L) bioreactor run. Various CHO-based cell lines and proprietary media were used in fed-batch runs to develop models that were independent of specific cell line or media composition. Twenty data points included fresh media with spiked amounts of glucose and lactate (no cells in the samples). Separate PLS models were developed for glucose, lactate, and viable cell density. An independent validation dataset from a batch not used for model calibration was used to evaluate model predictions and to optimize model parameters such as number of principal components, spectral range, and pretreatment method used. After determination of optimal model parameters, the developed models were tested using two additional validation datasets, one from a batch in 500 L Pilot scale to test for model scalability, and the other from a small scale bioreactor run with a CHO cell line not used previously in model development, to verify generic models are applicable to new CHO cell lines used for cell culture processes. The results of model calibration and validation are presented in Results and Discussion section.
Results and Discussion Model calibration and validation Model identification and validation results are summarized in Table 1. Specifically, number of samples and principal components, values of coefficient of determination (R2Y), as well as values for errors of calibration and prediction are reported for glucose, lactate, and VCD models. Careful attention has been made to ensure the calibration range encompasses the operating range of process parameters as multivariate regression models are not suitable for extrapolating. For all three modeled parameters, the values of the root mean square error of evaluation (RMSEE), cross validation (RMSECV), and prediction (RMSEP) are comparable and in the same range, indicating stability of developed models. The root mean square error (RMSE) is a frequently used measure of the difference between model predictions and actual or reference measurement values. Root mean square errors were calculated using the following equation in which
Table 1. Summary of Model Calibration and Validation Component Glucose (g/L) Lactate (g/L) VCD (E5 cells/mL)
N
# Factors
R2 Y
RMSEE
RMSECV
Range
RMSEP
900 896 884
6 5 9
0.938 0.954 0.936
0.32 0.19 13.61
0.33 0.20 14.17
0.0–8.45 0.0–3.95 0–300
0.43 0.26 19.82
1008
Biotechnol. Prog., 2015, Vol. 31, No. 4
Figure 3. Model calibration plots comparing reference measurements versus model prediction (A, C, and E), as well as time series plots of reference measurements and model predictions (B, D, and F) for the calibration dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models. Arrows indicate high concentration lactate samples for lactate model and cell-free standard media samples for VCD model.
ymodel;i are the values predicted by the model and yref;i are the actual values known or obtained from reference measurement method. RMSE ¼
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Xn 2 ðy 2yref;i Þ i¼1 model;i n
It is observed that developed models are performing with acceptable error, having high R2Y values (above 0.90 for all models) and low RMSEP values (0.43 g/L, 0.26 g/L, and 19.82 E5 cells/mL for glucose, lactate, and VCD predictions, respectively), indicating good model fit. Figure 3 illustrates model predictions plotted against reference measurements as well as time series plots of these parameters for the calibration dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
In general, model fit to the data is good. Exceptions are high concentration lactate samples for the lactate model and cellfree standard media samples for the VCD model. Although model fit was less favorable for these samples, our investigations show that including these samples in the dataset improved accuracy of model predictions as reflected by lower RMSEP values. Models were validated with data from an independent small scale batch using a mAb-producing cell line previously used in model development. Figure 4 shows model validation results, comparing model predictions with reference measurements and corresponding RMSEP values as well as time series plots of these parameters over batch time for three models developed. Glucose model predictions were in good agreement with reference measurements throughout the run, closely following the trend of reference
Biotechnol. Prog., 2015, Vol. 31, No. 4
1009
Figure 4. Model validation plots. Data generated from an independent batch run in a 3 L bioreactor using a mAb-producing cell line previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
Table 2. Comparison of RMSEP Values for Glucose, Lactate, and VCD Models RMSEP 1 RMSEP 2 RMSEP 3 Glucose (g/L) Lactate (g/L) VCD (E5 cells/mL)
0.43 0.26 19.82
0.28 0.072 30.87
0.44 0.41 13.95
RMSEP 1, 2, and 3 refer to standard errors in prediction of the original validation batch (accuracy validation), the batch with previously used cell line in large scale (500 L) bioreactor (scalability validation), and the batch using a cell line not included in model development in small scale (5 L) bioreactor (generic model validation).
measurements, with the exception of larger residuals towards the end of the batch (Figures 3A,B). This effect is commonly observed in batch runs due to increase in florescence excitation towards the end of the batch as a result of
increased dead cell population and accumulation of cell debris in the culture media. The lactate model was less accurate at the beginning of the batch run when higher lactate concentrations were present, predicting higher than actual values (Figures 4C,D). This is explained by the fact that most lactate data used in model calibration are in the lower end of the concentration range (Figure 3D). The difference between model predictions and reference values decreased towards the end of the run, resulting from accurate predictions at lower concentrations. Although the lactate calibration dataset included the full range of lactate concentrations, having a lower number of samples in the upper end of this range affected model predictions. This should be considered for future improvements of the models. Furthermore, careful attention should be made regarding accuracy of the models when predicting lower
1010
Biotechnol. Prog., 2015, Vol. 31, No. 4
Figure 5. Model scalability validation plots. Data generated from an independent batch run in a 500 L stainless steel bioreactor using a mAb producing cell line previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
concentrations. This is due to the fact that the model prediction error is in the range of 60.3 (g/L) and therefore model prediction percent errors are high in lower ranges (below 0.5 g/L). The VCD model over-predicted at the beginning of the run and under-predicted towards the end, while still predicting the trend in VCD changes with acceptable accuracy with an approximate RMSEP of 2.0 E5 cells/mL (Figure 4F). A number of VCD values obtained by the reference measurement method were outliers (around index number 80), so careful attention was given in data analysis to remove such measurements from calibration datasets. The VCD model uses a larger number of factors (9 compared with 6 and 5 for glucose and lactate models, respectively), which is due to increased complexity of predicting living cell counts compared with concentrations of individual sub-
stances present in the culture. This leads to noisier measurements, as evident from higher amplitude and frequency of the noise (Figure 4F). Model scalability and process independence validation In order to validate the generic capacity of the models, i.e. the models are applicable to various cell lines based on CHO host cells and their corresponding products, media compositions, and/or scales, they were used to predict glucose and lactate concentrations and VCD of two cell culture bioreactor batch runs. One of the batch runs was conducted using a cell line previously used in model development but now in a stainless steel 500 L pilot scale bioreactor, and the other was using a CHO-based cell line not included in model development and conducted in a 5 L glass bioreactor. Table
Biotechnol. Prog., 2015, Vol. 31, No. 4
1011
Figure 6. Generic model validation plots. Data generated from an independent batch run in a 5 L glass bioreactor using a protein producing cell line not previously used in model development. A, C, and E compare reference measurements with model prediction, and B, D, and F represent time series plots of reference measurements and model predictions for the validation dataset for glucose (A and B), lactate (C and D), and VCD (E and F) models, respectively.
2 provides a comparison of RMSEP values for glucose, lactate, and VCD models using the three validation datasets, with RMSEP 1, 2, and 3 referring to standard errors in prediction of the original validation batch (accuracy validation), the batch with previously used cell line in stainless steel 500 L bioreactor (scalability validation), and the batch using a cell line not included in model development in 5 L glass bioreactor (generic model validation), respectively. Prediction errors of all three models are comparable and in the same range for all three validation batch runs, demonstrating consistency, stability, and scalability of the developed models. In case of the large scale bioreactor run, RMSEP 2 values are smaller for glucose and lactate models compared with the original validation batch run (RMSEP 1), indicating excellent model scalability. Figures 5 and 6 depict model predictions versus reference measurements and corresponding RMSEP values of
individual models for these validation batches. In Figure 5 it is observed that models accurately predict concentrations of glucose and lactate throughout the batch. VCD model predictions in this case are less accurate at the beginning of the batch, resulting in higher RMSEP 2 values for the VCD compared to the original validation batch. Figure 6 shows opposite behavior in the validation run using a new cell line not used in model calibration, where glucose and lactate models are predicting less accurately compared with VCD model. Figure 6D illustrates that the lactate predictions are specifically less accurate when high lactate concentrations are present, consistent with what observed previously with original calibration dataset (Figure 4D). In contrast, the VCD model predicts with excellent accuracy throughout the run, closely following trends of VCD changes observed using reference measurement method (Figure 6F).
1012
Biotechnol. Prog., 2015, Vol. 31, No. 4
Conclusion Raman spectroscopy provides great potential for real-time in situ measurement of key process parameters in fed-batch bioreactor operations. A set of generic calibration models based on Raman spectroscopy were developed to measure glucose and lactate concentrations and viable cell density within mammalian cell culture bioreactors. PLS regression models were used to correlate spectral peaks with reference measurements. To develop generic models with increased robustness, the calibration dataset included data from multiple batches of different products produced by CHO-based cell lines, and included changes in proprietary media. Models were validated using three different validation datasets: (1) using a cell line that had previously been included in model calibration dataset, (2) using the same cell line but predicting in a 500 L pilot scale bioreactor, and (3) a dataset from a cell line not used in model development. Model predictions using validation datasets were within acceptable ranges in general, providing evidence that generated models were independent of scale and applicable to new cell lines derived from CHO host cells. Model predictions of glucose and lactate concentrations are therefore suitable for on-line control of the rates at which glucose and other nutrients are fed to the culture. It is important to note that model accuracies are limited in lower concentration ranges (
