The FDA's process analytical technology (PAT) initiative encourages drug manufacturers to apply innovative ideas to increase understanding of their manufacturing processes (OPS-FDA, 2004). The development of in-process analytics for process monitoring are integral components of PAT. However, there are many challenges to applying methods amenable to PAT for monitoring of biologics processes compared to those used for chemical pharmaceutics. Primarily, the analytical method must be capable of determining concentrations of key components in a complex culture formulation without interference from the other components. The ideal method should also be capable of in-line analysis allowing for real time monitoring without the need for sampling. Such sensors must be able to withstand harsh clean in place (CIP) and sterilization in place (SIP) procedures.
Various chemical, biological, and optical sensors have been implemented for process monitoring and are reviewed elsewhere (Becker et al., 2006; Ulber et al., 2003). Of all these systems, optical sensors are most likely to be successfully implemented for in-line measurements. Optical sensors are non-invasive, take measurements continuously, and can be steam sterilized in bioreactors thus eliminating the need to remove samples from reactors. Infrared and Raman spectroscopic measurements appear to be the most promising for process monitoring due to their ease of sampling and potential to measure multiple analytes simultaneously.
Both near infrared (NIR) and mid-infrared (MIR) spectroscopy have been extensively applied for monitoring a variety of bioprocesses (Card et al., 2008; McGovern et al., 2002; Sivakesava et al., 2001a,b). While sampling is fast and easy with NIR, the resultant peaks are broad and overlapping, making interpretation of spectra difficult. Absorbances in the mid-infrared region are stronger and yield more distinct spectral features; however, strong interference from water can make sample preparation in aqueous systems difficult. Raman spectroscopy measures the amount of light scattered inelastically at different frequencies by molecular vibrations. This results in very detailed molecular fingerprints with high chemical specificity. In addition, Raman measurements do not require sample preparation and can be applied to aqueous systems with ease. Because of this, Raman has been used extensively for a wide variety of applications (reviewed by Lewis and Edwards, 2001; Mulvaney and Keating, 2000), and appears to be the most promising spectroscopic method for in-line analysis of complex cell culture systems. Raman has been successfully applied to off-line monitoring of glucose, glutamine, lactate and ammonium in animal cell culture supernatants (Xu et al., 1997), and to off-line monitoring of multiple components in oil palm bioreactor culture supernatants (Willis et al., 2008). Lee and coworkers constructed a probe interfaced with a Raman spectrometer through a fiber optic for in-line monitoring of multiple components in Escherichia coli cultures (Lee et al., 2004). However, no literature reports can be found that apply Raman spectroscopy for in-line monitoring of mammalian cell systems. Regardless of the spectroscopic method used, most reports focus on a small number of analytes or are restricted to nutrients and metabolites. In a few cases, cell densities were predicted using spectroscopy without yielding information on culture viability.
The goal of this work was to determine if Raman spectroscopy could be applied for in-line monitoring of mammalian cell culture bioreactors. The Raman spectra were correlated with at-line measurements of glucose, glutamine, glutamate, lactate, ammonium, viable cell density (VCD), and total cell density (TCD) using partial least squares (PLS) modeling. This is of particular significance to the biotechnology industry since culture growth and viability are estimated in real time in addition to cell nutrients and metabolites. The results reported here demonstrate, at large scale (500-L bioreactors), the feasibility of using a Raman probe for continuous acquisition of real time data for process monitoring.
Raman spectra were collected in-line from 500-L bioreactors using a Raman spectrometer fiber optically coupled to multiple Raman immersion probes (Kaiser Optical Systems, Ann Arbor, MI). The bioreactors were sampled daily for reference method measurements (referred to as measured samples). After the daily samples, feed medium was added to the bioreactor. Post-feed nutrient, metabolite, and cell concentrations were calculated based on feed volume, measured values in the bioreactors prior to feed addition, and measured values of nutrients in the feed medium (referred to as calculated samples). Raman spectra collected immediately after the completion of feed addition were used in both the calibration and validation datasets, and were compared with the calculated concentrations. A total of 4 bioreactor runs operated under similar conditions were completed over a span of 5 months. Data from the first 3 Lots were used in the calibration dataset and data from the last lot were used to validate the models. The number of samples included in the calibration dataset was estimated based on the number of parameters being modeled.
Inoculum conditions for one of the runs used in the calibration dataset were controlled to yield a wider variation in each of the parameters. Table I summarizes the range of minimum and maximum levels included in the calibration dataset for each parameter. Comparison with the levels observed for the validation dataset shown in Figures 1 and 2 show that most of the validation data fall within the ranges found in the calibration dataset, which is expected to improve model performance. This approach for generation of the calibration dataset and model validation has been successfully applied to NIR applications (Arnold et al., 2003; Card et al., 2008; Cervera et al., 2009). Furthermore, the most immediate application of this technology would be in a manufacturing setting where similar process conditions are run repeatedly and wide variations in process performance are not generally expected. This provides further justification for the calibration and validation approaches taken here.
|Constituent||# Factors||R2||RMSECV 1 out||RMSEC||RMSECV 25% out||Calibration range|
|VCD (×106 cell/mL)||7||0.950||0.43||0.24||0.40||0.62–8.12|
|TCD (×106 cell/mL)||7||0.953||0.45||0.24||0.43||0.64–8.62|
PLS regression was applied to the training datasets as described in the Materials and Methods Section. Several different data processing techniques were evaluated, including 2nd derivatives, variance scaling, and different path length corrections. In general, it was determined that application of the gap 1st derivative with SNV path length correction performed best for these datasets (data not shown). The optimal number of factors for each model constituent was determined using PRESS plots and comparison of RMSECV and RMSEC values. After factor selection and outlier identification and removal, cross validation models were recalculated leaving approximately 25% of the training dataset out per cycle. Comparable standard errors between cross validation models generated with different numbers of files left out is an indicator of a robust model (Burman, 1989). Results for models using three lots of data in the calibration set for each constituent are summarized in Table I. As can be seen by the high coefficients of determination (R2) and low RMSECV and RMSEC, the models perform well. The average RMSECV (from leave-one-out cross validation) for all constituents was 0.39 with the highest errors coming from the ammonium model. In general, the RMSECV values obtained by leaving out one sample compared with leaving out 25% of samples per cycle were similar, indicating model stability.
The models were applied to the validation dataset and predictions were compared to measured values. The results of the model validations are summarized in Table II. The R2 for the actual versus predicted values, the average percent difference between predicted and measured values, and the estimated error for the reference methods are reported here. Three models using 1, 2, or 3 lots of data (with similar numbers of data points in each lot) in the calibration set for each parameter were evaluated using the validation dataset. The R2 increased significantly and the difference between measured and predicted values decreased significantly for most parameters as more data were included in the calibration set. Use of more than three lots of data in the calibration set did not significantly improve the accuracy of the models (data not shown). Although the R2 for the validation of the glutamate model is quite low (0.17), the average percent difference between measured and predicted values is within the error for the reference method. Examination of the glutamate trend in Figure 1 shows that the glutamate levels only vary from 2.7 to 4.0 mM through the entire run, which was only slightly larger than the error for the reference method.
|R2||Average % error||Reference method (% standard deviation)|
|1 Lot||2 Lots||3 Lots||1 Lot||2 Lots||3 Lots|
|VCD (×106 VC/mL)||0.072||0.051||0.928||219.0||168.2||14.9||10|
|TCD (×106 TC/mL)||0.079||0.121||0.927||184.9||116.7||14.5||10|
After validation of each model was completed, the models were applied to the spectra that were acquired every 2 h over the course of the cell cultures. Predictions from the validation lot are plotted along with the daily measurements and post-feed calculations that were used as the validation dataset in Figures 1 and 2.
Overall, the model predictions based on the Raman spectra matched measured and calculated values for glutamine, glutamate, glucose, lactate, and ammonium throughout the course of the run as shown in Figure 1. The Raman models accurately predicted increases in nutrient levels and decreases in metabolite levels after feeding. The average difference between the measured and predicted values for glutamine (Fig. 1a) was 30%, which was close to the 22% relative standard deviation (RSD) for that measurement. RSD for each at-line constituent reference method were determined experimentally, which is described below in the Materials and Methods Section. In general, Raman predictions appear to be closer to the reference method for glutamine concentrations >1 mM. This likely is due to lower accuracy of both the reference method and model predictions at lower glutamine levels. For glutamate (Fig. 1b), the average difference was 12% compared to the reference method RSD of 17%. Furthermore, the increases in both the glutamine and glutamate concentrations after feed additions, as well as the gradual decline until the next feed addition, were accurately predicted by the models. The 15% average percent difference for glucose was greater than the reference method standard deviation of 4%; however, the predictions shown in Figure 1c followed the overall expected trend and accurately predicted the large increase in glucose concentration at 250 h from a bolus addition of a concentrated glucose solution. The model predictions matched the measured values for lactate concentrations well—13% average difference compared to a 10% RSD, respectively—and the model accurately predicted decreases in lactate concentrations (Fig. 1d) following dilutions by the daily additions of feed medium. Ammonium predictions (Fig. 1e) differed from measured values by an average of 11% during the run, which was larger than the 4% RSD of the reference method. This larger difference is primarily a result of ammonium predictions being 20–30% below measured values during the first 4 days of the culture. Including wider ranges in the calibration dataset would likely improve the robustness of the model, however, the performance demonstrated here was sufficient to show utility for this technique. In a manufacturing setting, the calibration dataset can be easily updated as additional samples are available to continuously improve the models.
The at-line measured and in-line predicted data for VCD, TCD, and culture viability are shown in Figure 2. Cell densities were included as individual model constituents and predicted culture viabilities were calculated from these values. The VCD and TCD estimates matched measured data within expected errors with only slight deviations for both parameters between days 0 and 2 when cell densities were very low. As a result, the calculated estimates for culture viability matched the measured data, and the overall trend followed observed behavior. These data support recent reports that Raman spectroscopy can be used to distinguish between live and dead cells, as previously reported using Raman micro-spectroscopy (reviewed in Notingher, 2007). The methods described here have also been applied to a second mammalian cell culture fed-batch process and resulted in similar agreement between the measured and predicted values for those constituents reported here, including VCD and TCD (data not shown).
Overall, the data presented here demonstrate the feasibility of using Raman spectroscopy for monitoring mammalian cell culture processes. To the authors' knowledge, this is the first report applying Raman technology using an in-line probe for real time measurement of nutrients, metabolites, and cell densities in large scale mammalian cell culture bioreactors. Real time measurement of such parameters can provide immediate feedback on process performance to ensure consistent manufacture of biologics using PAT and QbD principles. This study lays the foundation for future studies to expand the capabilities for both monitoring and control of mammalian cell bioreactors using Raman spectroscopy. Based on a review of the literature, it seems likely that the number and types of components analyzed in bioreactors using Raman spectroscopy can significantly expand (Fagnano and Fini, 1993; Lopes et al., 2004; Mourant et al., 2003; Short et al., 2005; Tuma, 2005). Implementation of real-time monitoring in both manufacturing and development settings will lead to an increased understanding of the interaction between cells and their environments leading to more efficient design and control of processes. Ultimately this will facilitate streamlined regulatory approvals of innovative and cutting edge processing strategies consistent with the FDA's PAT and Quality by Design (QbD) initiatives.