Scale-Up Analysis for a CHO Cell Culture Process in Large-Scale Bioreactors

Transcription

1 ARTICLE Scale-Up Analysis for a CHO Cell Culture Process in Large-Scale Bioreactors Zizhuo Xing, Brian M. Kenty, Zheng Jian Li, Steven S. Lee Process Sciences, Biologics Manufacturing and Process Development, Worldwide Medicines Group, Bristol-Myers Squibb Company, P.O. Box 4755, Syracuse, New York ; telephone: ; fax: ; zhengjian.li@bms.com Received 2 May 2008; revision received 9 January 2009; accepted 2 February 2009 Published online 9 February 2009 in Wiley InterScience ( DOI /bit Introduction ABSTRACT: Bioprocess scale-up is a fundamental component of process development in the biotechnology industry. When scaling up a mammalian cell culture process, it is important to consider factors such as mixing time, oxygen transfer, and carbon dioxide removal. In this study, cell-free mixing studies were performed in production scale 5,000-L bioreactors to evaluate scale-up issues. Using the current bioreactor configuration, the 5,000-L bioreactor had a lower oxygen transfer coefficient, longer mixing time, and lower carbon dioxide removal rate than that was observed in bench scale 5- and 20-L bioreactors. The oxygen transfer threshold analysis indicates that the current 5,000-L configuration can only support a maximum viable cell density of cells ml 1. Moreover, experiments using a dual probe technique demonstrated that ph and dissolved oxygen gradients may exist in 5,000-L bioreactors using the current configuration. Empirical equations were developed to predict mixing time, oxygen transfer coefficient, and carbon dioxide removal rate under different mixing-related engineering parameters in the 5,000-L bioreactors. These equations indicate that increasing bottom air sparging rate is more efficient than increasing power input in improving oxygen transfer and carbon dioxide removal. Furthermore, as the liquid volume increases in a production bioreactor operated in fed-batch mode, bulk mixing becomes a challenge. The mixing studies suggest that the engineering parameters related to bulk mixing and carbon dioxide removal in the 5,000-L bioreactors may need optimizing to mitigate the risk of different performance upon process scale-up. Biotechnol. Bioeng. 2009;103: ß 2009 Wiley Periodicals, Inc. KEYWORDS: process scale-up; mixing time; oxygen transfer coefficient; dissolved carbon dioxide removal rate; mathematical model Correspondence to: Z.J. Li S.S. Lee s current address is: A-Bio Pharma Pte Ltd, 41 Science Park Road, The Gemini, Singapore Science Park II, Singapore Large-scale and high-density mammalian cell cultures have been established as a major means for manufacturing commercial products in the biopharmaceutical industry (Farid, 2007). Chinese Hamster Ovary (CHO) cells are commonly employed as a host for producing various recombinant proteins for therapeutic applications (Jayapal et al., 2007). A cell culture process is normally developed in bench-scale bioreactors and then scaled up to larger bioreactors for commercial production (Heath and Kiss, 2007). The aim of process scale-up is to produce larger quantities with equivalent productivity and product quality (Schmidt, 2005). Different agitation strategies are used to scale-up a process including specific power input ( P/V), impeller tip speed (V tip ), impeller shear rate (g), and specific impeller pumping rate (Q s ) (Chisti, 1993; Varley and Birch, 1999). Constant mixing time is not a frequently used scaling criterion, because P/V value can be prohibitively high upon scale-up and is usually overestimated (Junker, 2004). In addition, mixing time is not a readily measurable variable that is not convenient for scaling (Varley and Birch, 1999). For example, when scaling up a cell culture process from 3- to 2,500-L, constant mixing time resulted in atypical and unpractical high agitation speed (190 rpm) (Yang et al., 2007). It is not possible to maintain all variables constant simultaneously due to restriction in design and configuration of different bioreactors. The constant specific power input ( P/V) is the most often used scaling criterion in microbial fermentation (Junker, 2004) and was also used in animal cell cultures (Langheinrich and Nienow, 1999). In animal cell cultures, the commonly used scaling criteria are summarized in Table I. For the second and third criteria in Table I, mixing time is likely to be much longer for larger rather than small-scale bioreactors (Varley and Birch, 1999). Furthermore, it was found that constant tip speed led to atypical and unpractical low agitation speed (9 rpm) in 2,500-L bioreactor when a cell culture process was scaled up ß 2009 Wiley Periodicals, Inc. Biotechnology and Bioengineering, Vol. 103, No. 4, July 1,

2 Table I. Summary of scaling criteria for suspension animal cell cultures (Chisti, 1993; Ju and Chase, 1992; Varley and Birch, 1999). Scale-up criteria (1) Geometric similarity, constant k L a, and constant specific impeller pump rate (Q s ) Equations a k L a / QGS (2) Geometric similarity, constant k L a, and constant maximum shear (impeller tip speed, V tip ) P V / N3 D 5 V V tip ¼ pnd (3) Constant k L a, constant impeller tip speed, and constant Q s Q s ¼ Q G V P V / ND3 V b from 3-L bioreactor. Thus, an agitation (65 rpm) was used at the 2,500-L scale for the balance of tip speed and mixing time (Yang et al., 2007). Process scale-up from the laboratory to production scale remains a challenging task for both microbial and mammalian cell culture systems (Humphrey, 1999; Junker, 2004; Schmidt, 2005). For microbial fermentation, it has been well documented that non-ideal or uneven distribution of the liquid phase may lead to gradients in substrate, oxygen, biomass, or heat (Larsson et al., 1996). These are often cited as the reasons for decreased productivity upon scale-up and are rather common in the fermentation industry (Bylund et al., 1998; Larsson et al., 1996; Marten, 1997; Osman et al., 2001). For mammalian cell culture, hydrodynamic shear and bubble damage have traditionally been considered as the primary issues in large-scale systems due to the shear sensitivity of mammalian cells (Marks, 2003). However, with the use of surfactants such as Pluronic-F68 (Meier et al., 1999), shear sensitivity is no longer considered to be a major problem (Nienow, 2006). Studies focused on oxygen transfer (Lilly, 1983; Serrato et al., 2004), bulk liquid mixing (Marks, 2003), and dissolved carbon dioxide (dco 2 ) removal (Kirdar et al., 2007; Mostafa and Gu, 2003; Zhu et al., 2005) have been documented in literature as the critical engineering parameters for scaling up a mammalian cell culture process. Oxygen transfer is a critical factor in the scale-up of a cell culture process because mammalian cell culture is an aerobic process. In contrast to microbial systems, which utilize high specific power inputs, mammalian systems use relatively low power inputs because a mammalian cell has no cell wall to provide resistance to strong shear force from high agitation (Croughan et al., 1987; Marks, 2003; Varley and Birch, 1999). Therefore, agitation speed presents a key constraint for cell culture bioreactors, which may suffer from insufficient mass transfer and homogeneity due to low mixing intensity (Marks, 2003). Bulk liquid mixing is another critical factor in the scale-up of a cell culture process. Poor mixing can lead to ph and nutrient gradient in large-scale bioreactors (Bylund et al., 1998; Langheinrich and Nienow, 1999; Wayte et al., 1997). Several studies in literature have attempted to simulate these ph gradients in small-scale bioreactors. These simulated conditions resulted in reduced cell growth as well as decreased antibody production (Osman et al., 2001, 2002). The third critical factor for the scale-up of a cell culture process is the dco 2 removal. Carbon dioxide accumulation is a recurrent issue in mammalian cell culture due to the relatively low power inputs and air sparging rates that are utilized (Gray et al., 1996; Mostafa and Gu, 2003). There have been many laboratory scale studies in literature that address the impact of dco 2 level on cell growth and antibody/protein production (dezengotita et al., 1998, 2002; Gray et al., 1996; Kimura and Miller, 1996; Zhu et al., 2005). While the optimal dco 2 level varies with cell lines, relatively high dco 2 levels have an inhibitory effect on cell growth and protein production (dezengotita et al., 1998; Gray et al., 1996; Mostafa and Gu, 2003). Others have reported that high dco 2 levels may have an impact on protein quality attributes such as glycosylation (Schmelzer and Miller, 2002). In our previous publication (Zhu et al., 2005), we observed differences in cell growth, viability, and titer when scaling up the process for an antibody-fusion protein B1 production from bench to production scale. Volumetric productivity was reduced and cell viability declined more rapidly at production scale compared with bench scale. Among many variables, partial pressure of carbon dioxide ( pco 2 ) levels in the production-scale bioreactors were mmhg, much higher than the observed at bench scale (Zhu et al., 2005). The present work systematically evaluates these scale-up issues within the scope of bulk mixing, gas transfer, and carbon dioxide removal. These results will be helpful in improving equipment design and configuration to enhance cell growth and productivity. Furthermore, we report on production scale bioreactors that utilize multiple marine impeller system. This system was rarely reported in large-scale bioreactors in literature due to the limitation of cost and weight of marine impellers. Instead, multiple pitched blade impellers with good pumping characteristics were used in large-scale bioreactors when a good vertical mixing was required; for instance, a 19,000-L fermentor for viscous fungal fermentation (Pollard et al., 2007) and a 2,500-L bioreactor for animal cell culture (Yang et al., 2007). Marine impeller was occasionally used in microbial fermentation such as Kluyveromyces marxianus, of which much high agitation rate (450 rpm) was used (Yépez Silva- Santisteban and Maugeri Filho, 2005). The mixing characteristics of mild agitation for animal cell culture may be different from the microbial fermentation system using marine impellers. Therefore, the models developed in this work for bulk mixing, gas transfer, and carbon dioxide removal may apply to other large-scale bioreactors equipped with multiple marine impellers for animal cell cultures. 734 Biotechnology and Bioengineering, Vol. 103, No. 4, July 1, 2009

3 Materials and Methods Table II. Summary of the bioreactor configuration at each scale. Bioreactor Set-Up for Cell-Free Studies The cell culture process was scaled up from standard Applikon (Applikon, Inc., Foster City, CA) 5- and 20-L glass bioreactors to a stainless steel 5,000-L bioreactor (Feldmeier, Syracuse, NY). All bioreactors were equipped with marine impellers pumping axially downward. The 5-L bioreactor contained one impeller and a frit sparger with an average pore diameter of 15 mm. The 20-L bioreactor was equipped with three impellers and a frit sparger with an average pore diameter of 15 mm. The 5,000-L bioreactor had three impellers and a pipe sparger containing 1.8 mm diameter holes. The use of pipe spargers in 5,000-L bioreactors addressed the cgmp-related constraints such as clean in place (CIP) validation, foaming issue, and low efficiency of dco 2 removal with a frit sparger (Marks, 2003; Mostafa and Gu, 2003). The details of the bioreactor configurations are described in Figure 1 and Table II. A test medium was used to mimic properties of actual cell culture medium such as viscosity, buffering capacity, and bubble coalescence. The use of a test medium in this study instead of the actual cell culture medium is due to the concern of microbial contamination in our cgmp facilities (5,000-L bioreactors). Since sodium carbonate (Na 2 CO 3 )is a critical buffering reagent for ph control (Langheinrich and Nienow, 1999) and Pluronic-F68 lowers k L a significantly (Lavery and Nienow, 1987), the same concentrations of Parameters Scale 5-L 20-L 5,000-L H T (m) T (m) Max H L /T Impeller type A100 A100 A100 n N p D/T C/D DC/D N/A T 1 /D T 2 /D T 3 /D The description of each parameter is presented in Figure 1. sodium bicarbonate and Pluronic-F68 in cell culture medium were used in the test medium. The additional sodium chloride was also used in the test medium to maintain the similar viscosity and bubble coalescence as the actual cell culture medium. Thus, the test medium is composed of 6.4 g L 1 sodium chloride, 2.0 g L 1 sodium bicarbonate, and 1.0 g L 1 Pluronic-F68. It was reported in literature that the additive of serum-free cell culture medium to distilled water only had negligible impact on k L a value at the low agitation speed (100 rpm) (Lavery and Nienow, 1987). Therefore, the measurement from the test medium should represent that from the actual cell culture medium. Figure 1. Schematic of the bioreactor set-up for mixing experiments. The tracer for mixing time experiments was manually added as a bolus addition through the dip tube. The value of each parameter is presented in Table II. Mixing Time Measurements Mixing time is defined as the duration of time required to reach 95% homogeneity in a mixed vessel upon addition of a tracer. Many types of tracers (e.g., salt, fluorescent dye, radioactive material) can be used to measure mixing time in a vessel (Menisher et al., 2000). For a bioreactor that is used for commercial production, it is not possible to introduce foreign compounds to the system due to regulatory cgmp restrictions. For this reason, it is advantageous to utilize a material that is routinely added during the production process. Since ph is routinely controlled in bioreactors through the addition of base, the mixing time measurements in this study were performed using pulses of 4 N NaOH as the tracer, according to a previously described technique in literature (Marten et al., 1997). A schematic of the bioreactor set-up is shown in Figure 1. Two ph probes were calibrated using standard buffer solutions. One ph probe was placed at the bottom of the bioreactor and the other was located at the top of the bioreactor just below the liquid surface on the opposite side. Both probes were connected to a computer logging system to track the ph reading over time. A Bellco controller (Boston, MA) was used for the 5- and 20-L Xing et al.: Scale-Up Analysis for a CHO Cell Culture 735 Biotechnology and Bioengineering

4 bioreactors, and an ABB (Cleveland, OH) Distributed Control System (DCS) was used for the 5,000-L bioreactor. Temperature of the test medium was maintained at 37.08C. Tracer addition was performed through a tube located just above the liquid surface, diametrically opposed to the top ph probe position (Fig. 1). An addition tube was utilized to minimize the plume observed by others (Langheinrich and Nienow, 1999) and improve consistency of experiments. The volume of each addition was less than 0.5% of the liquid volume inside the bioreactor, thus the effect of tracer addition volume on mixing time was negligible. The tracer was rapidly added in a single shot and the ph excursion was recorded until the readings reached a steady state. After each trial, the ph was adjusted back to the original value using 4 N HCl. The raw data from the ph probes were normalized according to the following equation: HðtÞ ¼ phðtþ ph i ph f ph (1) i where H is referred to as homogeneity index, or normalized ph. Figure 2A is a representative normalized ph response for both top and bottom ph probes after tracer addition. In theory, as t!1, the liquid inside the bioreactor is completely homogenized, and a steady-state ph value will be reached where H is equal to 1.0. In practice, a 95% degree of mixing is commonly used to determine mixing time. Thus, a 5% H deviation from final homogenization (H ¼ 1.0) has been adopted by industry (Marten et al., 1997). Figure 2B is a magnification of Figure 2A where the normalized ph curves approach the 5% region. The average mixing time from the top and bottom ph probes was used as the overall mixing time in this study. Dynamic Determination of the Oxygen Transfer Coefficient Dynamic determination of the oxygen transfer coefficient (k L a O2 ) is based on supplying oxygen to a fluid that has been depleted of oxygen by nitrogen sparging. The rate of oxygen transfer is governed by Equation (2). To generalize the calculation for the gases used in cell culture, the equilibrium concentration of oxygen and carbon dioxide can be calculated by Equation (3). P T is the pressure at the sparger (mmhg), and y O2 and y CO2 are the molar fraction of oxygen and carbon dioxide in the inlet gas sparge flow, respectively. The Henry coefficient H O2 and H CO2 (atm l mol 1 ) under ambient pressure are adapted from the empirical equations in literature (Ducommun et al., 2000; Spérandio and Paul, 1997) d½o 2 Š dt ¼ k L a O2 ð½o 2 Š ½O 2 ŠÞ (2) ½xŠ ¼ y xðp T =760Þ H x ; x ¼ O 2 or CO 2 (3) Figure 2. Representative normalized ph (H) response curves from the top and bottom probes in a bioreactor during the mixing time experiment. A: The normalized ph response curves from the top and bottom probes in a 5,000-L production bioreactor during a mixing time test with a specific power input of 100 W m 3. B: The mixing time determination: t ¼ (t 1 þ t 2 )/2. The on-line dissolved oxygen (DO) reading is used to obtain [O 2 ], which is plotted versus time to obtain a correlation from which d[o 2 ]/dt can be calculated (Atikinson and Mavituna, 1983). d[o 2 ]/dt is plotted versus [O 2 ] which results in a straight line with a slope of k L a O2.It was assumed that the gas phase was well mixed and equal to the inlet air composition. This assumption also implies that [O 2 ](t) is constant everywhere in the bioreactor and equal to the saturation concentration with respect to air. This allows for simple solutions of unsteady state equations used to calculate k L a O2. In this study, the ranges of specific power input and bottom air sparging rate typically used for mammalian cell culture were applied (Marks, 2003; Nienow et al., 1996; Varley and Birch, 1999), described in detail in Results and Discussion Section. To begin the experiment, two DO probes were calibrated to 100% air saturation and placed in the same location as the ph probes for the mixing time study (Fig. 1). Both DO probes were connected to the computer system for trending. The medium was prepared in each bioreactor and the temperature was maintained at 37.08C. 736 Biotechnology and Bioengineering, Vol. 103, No. 4, July 1, 2009

5 All bioreactors at three scales were run with ambient headspace pressure. The DO profile from the bottom probe was used in the k L a calculation, and the top DO probe was used to compare the DO profile difference at the top and bottom of the bioreactor. Before each run, nitrogen was sparged into the bioreactor until the DO reading was less than 10% air saturation. Once the proper agitation and sparging rate was set, the DO response was logged until the reading was at least 60%, and k L a was determined by the method described above. The range was chosen to minimize the amount of time required for each run. This process was repeated for each experimental condition. Bielefeld, Germany). The pco 2 values were measured using NOVA Bioprofile 400 (NOVA Biomedical, Waltham, MA). Models for Mixing Analysis Power input per volume (e), superficial gas velocity (v s ), and Reynolds number (Re) are calculated by Equations (4) (6). Impeller power number (N p ) in Equation (4) was provided by the vendor (Lightnin, Inc., Rochester, NY). The power inputs utilized in this study ranged from 5 to 125 W m 3 " ¼ P V ¼ N pnrn 3 D 5 V (4) Determination of the dco 2 Removal Rate The pco 2 level was measured using a YSI BioVision 8500 CO 2 monitor (YSI, Inc., Yellow Springs, OH). The probe was calibrated by sparging 10% CO 2 /90% N 2 gas standard (Airgas, Inc., Radnor, PA) through the medium until saturation at 37.08C. This value was then entered into the YSI unit. The probe was positioned in the bottom portion of the bioreactor. The 5- and 20-L bioreactors were run with ambient headspace pressure, which is standard practice for glass vessels. The 5,000-L bioreactor was operated with 5.0 psig headspace pressure to simulate normal operating conditions. In addition to the headspace pressure, the hydrostatic pressure also presents a challenge at 5,000-L scale due to the increased CO 2 solubility. The dco 2 removal rate (R dco2 ) was evaluated at different agitation speeds and air sparging rates for each scale, described in detail in the Results and Discussion Section. Before each run, pco 2 level was brought to 150 mmhg by sparging pure CO 2 into the bioreactor. Once the reading stabilized, the desired agitation and air sparging rate were set. Data were collected every minute by the internal YSI 8500 memory until the pco 2 level reached 100 mmhg. R dco2 was calculated from the slope of the pco 2 versus time profile (Mostafa and Gu, 2003). Cell Culture Condition The proprietary CHO cell line, basal medium, feeding medium, and cell culture process for an antibody-fusion protein B1 production were previously described (Xing et al., 2008; Zhu et al., 2005). Cell culture runs were seeded at a viable cell density of approximately cells ml 1. All bioreactors were operated at a starting temperature of 378C and may be changed into as low as 328C based on the temperature control strategy of the internally developed process. The ph and DO were also controlled with the internally developed process. Cell cultures were performed in fed-batch mode with daily feeding to maintain glucose at a preset concentration. Samples were taken on a daily basis right before feeding. Cell density and viability were measured using a Cedex cell counter (Innovatis AG, v s ¼ Q G S Re ¼ D2 Nr m (5) (6) Four models were developed for mixing analysis of scaleup issues. The first model predicts the oxygen transfer coefficient (k L a O2 ). Oxygen transfer is dependent on specific power input, bottom air sparging rate, as well as the top airflow rate. However, with the increase of bioreactor scale, the surface area to volume ratio decreases dramatically, thus the effect of top air is negligible when the working volume of a bioreactor is greater than 300 L (Varley and Birch, 1999). The model was adapted from Equation (7) (Nienow et al., 1996), which was originally developed for a Rushton impeller system by van t Riet (1979). The constants in this equation were estimated by log-linear regression with measured data from the present work k L a O2 ¼ A P a Q b G (7) V S The second model estimates the oxygen transfer thresholds. The thresholds are defined as the minimum k L a O2 values to support a viable cell density based on the combination of sparge flow rate and agitation speed (Marks, 2003; Mooyman, 1987). The specific oxygen uptake rate (SOUR) was measured in four 5,000-L bioreactors during cell culture run between 160 and 220 h (at approximately 160, 190, and 220 h for each run) because the peak cell density was observed in this period. For each measurement, data were collected by DCS every minute and the data for 1- h period were used for analysis. The DO and gas flow rate were controlled at constant levels (coefficient of variation is less than 5%). Since the agitation, culture volume, temperature (348C), DO, and gas flow rates were kept constant during 1-h period for a measurement, the oxygen uptake rate (SOUR X V ) was equal to the oxygen transfer rate as shown in Equation (8). Based on oxygen uptake rate and X V, SOUR was calculated. The average of 12 measurements Xing et al.: Scale-Up Analysis for a CHO Cell Culture 737 Biotechnology and Bioengineering

6 (4 runs and 3 measurements per run) was then defined as the SOUR for modeling. The oxygen transfer thresholds were determined at 3,800 L working volume, 60% DO, and 348C, because the peak cell density was observed in cell cultures under this condition. For a given viable cell density (X V ) using the known SOUR, the minimum k L a O2 value to support the SOUR X V value was calculated by Equation (8). The combination of P/V value and sparging oxygen flow rate that could provide the minimum k L a O2 value was then calculated by Equation (7). Finally, the plot of calculated P/V values and sparging oxygen flow rates was the curve of oxygen transfer thresholds for the given viable cell density literature, the k L a CO2 =k L a O2 ratio of 0.91 is used in the current study { d½co 2 Š ¼ 10 ph 10 3 RT m SCER X V dt 10 ph þ k L a CO2 ð½co 2 Š ½CO 2 ŠÞ þ K C R V dx V ¼ m dt V X V dk L a CO2 ¼ u k k L a CO2 dt d½co 2 Š ¼ u C ½CO 2 Š dt SOUR X V ¼ k L a O2 ð½o 2 Š ½O 2 ŠÞ (8) dr V dt ¼ u R R V ð11þ The third model predicts the mixing time. The model was based on Equation (9) (Nienow, 1997). According to Equation (6), Reynolds number is directly proportional to the value of D 2 N. In addition, the flow number (Fl) is correlated to the impeller pumping number, which is also a function of Reynolds number (Nienow, 1998). Therefore, Equation (9) can be rearranged as Equation (10). A loglinear regression was performed to determine the constants K 1, a 1, and b 1 using the data in the present work where t mix ¼ 5V FlND 3 Fl ¼ Q G ND 3 t mix ¼ K 1 V a 1 Re b 1 1 D ð9þ (10) The fourth model predicts the dco 2 removal rate (R CO2 ) in the fed-batch CHO cell culture process for the production of antibody-fusion protein B1. The model is based on Equation (11), which was a modification of the equation reported for CHO cell cultures using batch mode (Mostafa and Gu, 2003). In this equation, the equilibrium constant K C is M (Mostafa and Gu, 2003), T m is the temperature of the bioreactor (K), R is the ideal gas constant (8.134 J mol 1 K 1 ). Since the unit of time (t) is hour, the volume ratio of the gas phase to the liquid phase (R V ) can be measured by volume of gas (L) sparged in 1-h period per liquid volume of cultures (L). The k L a CO2 value was calculated from k L a O2 value based on the k L a CO2 =k L a O2 ratio of 0.91 (Spérandio and Paul, 1997), which is derived from the ratio of liquid-phase diffusivities of CO 2 and O 2 (Royce and Thornhill, 1991). The measured k L a values for CO 2 and O 2 were reported in agreement with the k L a CO2 =k L a O2 ratio of at low agitation speed (25 and 53 rpm) in a mineral culture medium (Spérandio and Paul, 1997). The equivalent k L a CO2 =k L a O2 ratio (0.89) was also used in hybridoma cell cultures (Bonarius et al., 1995; Frahm et al., 2002). Thus, consistent with the practice in Since k L a CO2, [CO 2 ], and R V are determined by a gas flow rate that depends on time, the dynamic changes of these three variables can be obtained. The rates u ¼ [m V u k u C u R ] corresponding to variables Y ¼½X V k L a CO2 ½CO 2 Š R V Š were calculated by Equation (12) and used in Equation (11) for modeling u i ðtþ ¼ where y i ðn þ 1Þ y i ðnþ 0:5ðy i ðn þ 1Þþy i ðnþþðtðn þ 1Þ tðnþþ y 1 ¼ X V ; y 2 ¼ k L a CO2 ; y 3 ¼½CO 2 Š ; y 4 ¼ R V u 1 ¼ m V ; u 2 ¼ u k ; u 3 ¼ u C ; u 4 ¼ u R (12) Two modifications were made for the dco 2 removal model in this study compared to the model in literature (Mostafa and Gu, 2003). First, because of daily feeding for the fed-batch mode, Equation (11) was applied to each time period between two successive feed additions, which were denoted as time points t(n) and t(n þ 1). Second, the value at t(n) was used to predict the [CO 2 ] value at t(n þ 1) by Equation (11), which was numerically solved by the ODE23 function of MATLAB version 7.5 (The MathWork, Inc., Natick, MA). The second component of the d[co 2 ]/dt term in Equation (11) is the dco 2 removal rate presented by Equation (13) R dco2 10 ph ¼ 10 ph k L a CO2 ð½co 2 Š ½CO 2 ŠÞ (13) þ K C Results and Discussion Oxygen Transfer The k L a values were measured under normal operational ranges of engineering parameters for the CHO cell culture process in the 5,000-L bioreactors. These ranges are represented by three levels of culture volume (3,000, 3,800, and 4,400 L), four levels of superficial gas velocity 738 Biotechnology and Bioengineering, Vol. 103, No. 4, July 1, 2009

7 (0.14, 0.71, 1.43, and ms 1 ), and three levels of agitation (30, 50, and 80 rpm). A complete factorial design was applied for k L a measurement, leading to 36 experimental conditions. The measured k L a values at the maximum allowed superficial gas velocity for the CHO cell culture ( ms 1 ) are presented in Figure 3, which are between and min 1. The k L a values in a cellfree study mimicking bench-scale CHO cell culture conditions are min 1 for 5-L bioreactors and min 1 for 20-L bioreactors. The maximum achieved k L a value in the 5,000-L bioreactor (0.057 min 1 ) was less than 50% of the maximum value observed at 5- and 20-L bioreactors. One significant difference in the configuration of the three bioreactors is that a frit sparger was used in the 5- and 20-L bioreactors, whereas a pipe sparger was used in the 5,000-L bioreactor. A frit sparger can produce very fine air bubbles in the order of microns, which have a high surface area for oxygen transfer. The air bubbles produced by a pipe sparger are in the size of millimeters (Marks, 2003; Mostafa and Gu, 2003), and therefore have much less surface area for oxygen transfer. However, due to cgmp-related constraints such as CIP validation, foaming issue, and low efficiency of dco 2 removal, frit spargers are rarely used beyond pilot-scale bioreactors (Marks, 2003; Mostafa and Gu, 2003). To investigate the scale-up issues that caused the reduction of cell culture performance previously reported (Zhu et al., 2005), the bioreactor configurations for the original process development using frit spargers at bench scales (5- and 20-L bioreactors) were kept in the current study. For a model to predict k L a value for the 5,000-L bioreactors, the constants in Equation (7) were estimated under the 36 conditions described above. The measured and the predicted k L a values are presented in Figure 3, where the P/V values are calculated from agitation speed and culture volume by Equation (4). Based on the data from all 36 conditions, r 2 value of the fitting model was 0.96, indicating a good agreement between the predicted and measured values. These constants are incorporated into Equation (14). The predicted k L a values were then used to measure SOUR value in four 5,000-L bioreactors during cell culture run between 160 and 220 h. The average SOUR value of 12 samples was 0.32 pmol cell 1 h 1. This value is within the range of reported SOUR values for CHO cell cultures between 0.25 and 0.29 pmol cell 1 h 1 (Ducommun et al., 2000) and 0.35 pmol cell 1 h 1 (Mostafa and Gu, 2003). This result also indicates that Equation (14) is applicable in the CHO cell culture process k L a O2 ¼ 0:075 P 0:47 Q 0:80 G (14) V S As described in the Materials and Methods Section, Equation (8) defines the relationship between a cell density and its oxygen demand that is supported by the bioreactor s k L a. Furthermore, the k L a value can be calculated from specific power input and oxygen flow rate by Equation (14). Therefore, for a target cell density, the plot of specific power input against the sparge oxygen flow rate was defined as an oxygen transfer threshold curve (Marks, 2003; Mooyman, 1987). The peak cell density was observed at 348C in our cell culture process. Since temperature affects oxygen solubility in liquid and hence impacts the driven force for oxygen transfer ([O 2 ] [O 2 ]) based on Equation (2), we tested the value of ([O 2 ] [O 2 ]) at 60% DO at 34 and 378C. The result showed that the value at 348C was only 4% higher than that at 378C. Thus, Equation (14) based on the measurements at 378C could be used in modeling. For a cell culture in 5,000-L bioreactor at set points of 60% DO and 348C, oxygen transfer threshold curves for the target cell densities of 6, 7, and cells ml 1 are presented as examples in Figure 4. When the maximum allowed specific power input and oxygen sparge flow rate for the specific configuration of 5,000-L bioreactors were plotted in Figure 4, it falls onto the oxygen transfer threshold curve for cells ml 1. This result indicates that the current bioreactor configuration of 5,000-L bioreactors could only support a maximum cell density of cells ml 1. Equation (14) provides guidance for the optimization of agitation and gas flow rate in our 5,000-L bioreactors equipped with multiple marine impellers. In comparison with the equation reported in literature for the Rushton system (Nienow et al., 1996; van t Riet, 1979), the a value is Figure 3. Comparison of the predicted (black bars) and the measured oxygen mass transfer coefficient values (white bars) in 5,000-L bioreactors. The predicted value was calculated by Equation (14). Xing et al.: Scale-Up Analysis for a CHO Cell Culture 739 Biotechnology and Bioengineering

8 Figure 4. Oxygen transfer threshold curves at three target viable cell densities in 5,000-L bioreactors with 3,800-L volume culture at 348C. The area below the threshold curve is where respiration becomes growth limiting. Figure 5. Model predicted (lines) and measured (symbol) values of mixing times in 5,000-L bioreactors. The liquid levels at 3,000-L (square), 3,800-L (diamond), and 4,400-L (triangle) were evaluated at airflow rates of 10 SLPM (solid symbols) and 15 SLPM (open symbols), respectively. similar (0.47 vs. 0.4 in literature) but the b value is higher (0.8 vs. 0.5 in literature) in Equation (14). Because the b value is larger than the a value, it is concluded that the k L a O2 value is more dependent on superficial gas velocity than the specific power input in the 5,000-L bioreactor system equipped with multiple marine impellers. Thus, to improve the oxygen transfer for a bioreactor with fixed cross-sectional area (S), increasing gas flow rate (Q G ) is more effective than increasing specific power input. As mentioned above, the use of a frit sparger in place of pipe sparger in a 5,000-L bioreactor is the more effective way to increase oxygen transfer. However, due to the cgmp-related constraints such as CIP validation, foaming issue, and low efficiency of dco 2 removal, a frit sparger is rarely used beyond pilot scale (Marks, 2003; Mostafa and Gu, 2003). Consistent with this industrial practice, it is not recommended to use frit spargers in 5,000-L bioreactors. Bulk Mixing Time Mixing times were measured under normal operational ranges of engineering parameters in 5,000-L bioreactors. A complete factorial design was used for mixing time measurement, which includes three levels of agitation rate (30, 50, and 80 rpm) and three levels of culture volume (3,000, 3,800, and 4,400 L). For each condition, we measured mixing time at four different air sparging rates (0, 5, 10, and 15 SLPM). Since no significant difference in mixing time was observed under any condition (the coefficient of variation is less than 20%), the average mixing time from four sparging rates for each condition was used for analysis. This observation is in agreement with the generally thought that gas flow rate does not have a significant influence on the bulk mixing in a bioreactor as long as flooding does not occur at the agitator (Vasconcelos et al., 2000). As shown in Figure 5, the mixing times at the maximum allowed agitation in the 5,000-L bioreactors with the specific configuration were >100 s at high liquid levels (>3,800 L). The mixing times in a cell-free study mimicking bench-scale CHO cell culture conditions were 3 8 s for 5-L bioreactors and s for 20-L bioreactors. These results indicate that longer mixing time may be an issue in 5,000-L bioreactors, because poor mixing usually leads to ph, DO, and nutrient gradients in a large-scale bioreactor (Bylund et al., 1998; Langheinrich and Nienow, 1999; Marks, 2003; Wayte et al., 1997). To investigate if DO gradients may be present due to poor mixing in 5,000-L bioreactors, dual DO probes were utilized to measure the difference between the top and bottom DO probe readings. An equivalent specific power input (100 Wm 3 ), bottom air sparging rate ( vvm), and liquid aspect ratio were used for both 20- and 5,000-L bioreactors as shown in Figure 6. The top and bottom DO profiles in the 20-L bioreactor were nearly identical (Fig. 6A). However, in 5,000-L bioreactors, the bottom DO probe responds right away while the top probe did not respond for over 6 min (Fig. 6B). These results suggest that there may be significant oxygen concentration gradients in large-scale bioreactors, which were not observed in small-scale systems. The mixing time in the 5,000-L bioreactor was in the range of s whereas the top probe did not respond for over 6 min (Fig. 6). So the mixing time is not a direct factor causing the DO gradient. Instead, less efficient mixing (indicated by long mixing time) resulted in significantly lower k L a at the top than at the bottom of the bioreactor, which created the DO gradient. The liquid height may influence the localized k L a at the bottom and upper portions in a bioreactor. Unlike 20-L bioreactors, of which the liquid height is negligible, the 3,800-L liquid in 5,000-L bioreactors added 0.16 atm hydrostatic pressure at bottom DO probe compared to the top DO probe. This led to an increase of oxygen solubility around the bottom probe. This DO gradient has serious implications for the cell culture environment. For mammalian cells, there is an 740 Biotechnology and Bioengineering, Vol. 103, No. 4, July 1, 2009

9 Figure 6. Comparison of DO readings from the top and bottom probes in a bioreactor. The liquid aspect ratios in the 5,000- and 20-L bioreactors were similar, and equivalent specific power inputs and air sparging rates (vvm) were used. A: DO probe responses in a 20-L bioreactor. B: DO probe responses in the 5,000-L bioreactor. optimal DO range for cell growth and antibody/protein production (Heidemann et al., 1998; Jan et al., 1997). In the case of the DO probe being installed in the bottom of the bioreactor, the oxygen concentration in the upper portion of the bioreactor is likely to be lower than the bottom portion, especially at high volumes (Fig. 6B). If the DO gradient inside a bioreactor is severe enough, the upper portion of the bioreactor may be outside of the optimal range. As the cells circulate through the tank, they are constantly circulated in and out of this low DO zone, so cell growth and protein production may be negatively impacted. In our previous publication (Zhu et al., 2005), we observed that volumetric productivity was reduced and cell viability declined more rapidly at production scale in comparison with bench-scale performance. Further study would be required to determine if such gradients in DO are truly influencing cellular performance at the 5,000-L scale. To investigate if ph gradients may be present due to long mixing time in 5,000-L bioreactors, dual ph probes were utilized to measure the ph overshoot. As shown in Figure 2A, when tracer was added to the liquid surface, the top ph probe response was characterized by a rapid peak followed by a steady decline to the final ph value, while the bottom ph probe response was a steady increase to the final ph value. The ph overshoot is defined as the difference between the peak of the top normalized ph curve and the corresponding bottom probe value at that time point and may be used to characterize the transient ph gradient. As shown in Table III, the ph overshoot in the 5,000-L bioreactors at the maximum allowed liquid volume (4,400 L) and specific power input (21 W m 3 ) for the CHO cell culture process is 1.67 normalized units. The ph overshoot values in a cell-free study mimicking bench-scale CHO cell culture conditions are 0.08 normalized units for 5-L bioreactors and 0.81 normalized units for 20-L bioreactors. The ph overshoot value in 5,000-L bioreactors is more than 20- and 2-folds higher than that was observed in 20- and 5-L bioreactors, even though the bench bioreactors were run at lower specific power input (5 W m 3 ). These results suggest that there are significant ph gradients in 5,000-L bioreactors but were not observed in the bench-scale systems. This ph gradient has serious implications for the cell culture environment. In the CHO cell culture process, due to the production of CO 2 and lactate by cells, a concentrated alkali is added above the culture surface to maintain the ph at a set point. As the cells circulate through the tank, they may be exposed to regions close to the base addition area that are outside of the optimal ph range (Langheinrich and Nienow, 1999). To predict mixing time in 5,000-L bioreactors, we first examined Equation (15) described in literature (Nienow, 1998) using the nine conditions described above. This empirical equation was developed from a multiple radial flow Rushton turbine system (Cooke et al., 1988), which was tested in literature with different types of dual radial flow turbines and dual axial Lightnin A-315 impellers. The mixing time observed with the dual Rushton impellers was reported to be very close to the predicted value, but the dual A-315 mixing time was 50% of what Equation (15) predicted (Otomo, 1996). Prior to testing in the 5,000-L bioreactors, Equation (15) was evaluated in the 20-L bioreactors. The measured mixing times were 2 30% higher than the predicted mixing times, indicating that Table III. Comparison of ph overshoot among 5-, 20-, and 5,000-L bioreactors. a Bioreactor Volume (L) P/V (W m 3 ) ph overshoot (normalized) 5 L (1 impeller) L (3 impellers) ,000 L (3 impellers) 4, a Each value is the average standard deviation with n ¼ 3. Xing et al.: Scale-Up Analysis for a CHO Cell Culture 741 Biotechnology and Bioengineering

10 Equation (15) yielded acceptable results in the 20-L bioreactor. However, when Equation (15) was evaluated in the 5,000-L bioreactor, the measured values were 20 70% lower than the predicted values, indicating that an alternative model may be needed to predict mixing time for this system equipped with three marine impellers t mix ¼ 3: =3 D 2:43 (15) N N p H L We then evaluated an alternative equation (Eq. (10)) in the 5,000-L bioreactors under the nine conditions described above. The constants of Equation (10) were estimated as shown in Equation (16). Figure 5 indicates that the values predicted by the model are in agreement with the measured values, where the r 2 value of the fitting model was 0.97 t mix ¼ 13; 546V 2:398 Re 0:7455 D 1 (16) Equation (16) provides guidance on how to improve the bulk mixing inside the 5,000-L bioreactor. Since the absolute value of the exponential constant for volume is threefold higher than the value for Re number (2.398 vs ), it can be concluded that mixing time value is more dependent on the culture volume than the agitation rate. Therefore, an effort that reduces the overall feeding volume may be required to improve mixing inside the 5,000-L bioreactor. Furthermore, alkali addition into impeller zone should be used for ph control to break ph gradients in 5,000-L bioreactors as suggested in literature (Langheinrich and Nienow, 1999). Carbon Dioxide Removal The dco 2 removal rates were measured under normal operational ranges of engineering parameters in 5,000-L bioreactors. It has been reported in literature that increasing the top airflow rate is an effective way to enhance dco 2 removal (Marks, 2003; Mitchell-Logean and Murhammer, 1997; Mostafa and Gu, 2003). However, we found that increasing the headspace airflow rate by as much as 200% did not have any significant effect on the dco 2 removal rate in the 5,000-L bioreactor. Therefore, dco 2 removal rates were measured at different specific power inputs and sparge gas flow rates in the 5,000-L bioreactors. The dco 2 removal rates at 3,800-L liquid volume are presented in Figure 7. At a fixed power input, the dco 2 removal rate was positively correlated to sparge gas flow rate at three power inputs. At a fixed sparge gas flow rate, the dco 2 removal rate did not respond to an increase of specific power input between 5.2 and 24 W m 3 ; however, it became positively correlated to specific power inputs between 24 and 100 W m 3. Figure 7 also shows that the dco 2 removal rate is approximately 0.53 mmhg min 1 at a sparge gas flow rate of 15 SLPM and a specific power input of 24 W m 3, which are Figure 7. Carbon dioxide removal rates at 3,800-L liquid level in 5,000-L bioreactors. The specific power input at 5.2 (cross), 24 (solid diamond), 100 W m 3 (open triangle) are tested. the maximum allowed parameters for the 5,000-L CHO cell culture process. The dco 2 removal rate in a cell-free study mimicking bench-scale CHO cell culture conditions was mmhg min 1 for 5-L bioreactors and mmhg min 1 for 20-L bioreactors. These results indicate that the dco 2 removal rate in a 5,000-L bioreactor is less than 50% of that observed in the bench-scale bioreactors. To improve dco 2 removal, the current bioreactor configuration of 5,000-L bioreactors may need modifying to allow higher sparge gas flow rate and/or higher specific power input. The dco 2 removal rates measured in the cell-free study are in agreement with pco 2 profiles observed in the cell culture process. Figure 8 displays the average pco 2 profiles for multiple CHO cell culture runs in the 5-, 20-, and 5,000- L bioreactors from historical B1 protein production data. Equivalent bottom air sparging rates and headspace air flow rates (vvm) were utilized for these runs. While the viable cell Figure 8. Comparison of average carbon dioxide profiles among 5-, 20-, and 5,000-L bioreactors during a cell culture process for the production of an antibodyfusion protein B1. The carbon dioxide level is normalized to the initial pco 2 value at 5-L scale (in ratio). The cell culture age is normalized to the harvest time at 5-L scale (in ratio). 742 Biotechnology and Bioengineering, Vol. 103, No. 4, July 1, 2009

11 densities were comparable at each scale, the average final pco 2 level in the 5,000-L bioreactor was significantly higher than the levels observed in both 5- and 20-L bioreactors. These results suggest that the higher pco 2 level in cell culture is a result of the lower dco 2 removal rate in the 5,000-L bioreactors. A theoretical model was developed to predict the dco 2 removal rate (Equation 11). Three fed-batch CHO cell culture runs for an antibody-fusion protein B1 production at 5,000-L scale were used to calculate CO 2 concentrations by Equation (11). To be consistent with the practice of others who evaluated a CO 2 model in the cell growth phase (Mostafa and Gu, 2003), we evaluated our model at a culture duration of h. This time period is appropriate because external CO 2 is sparged for ph control prior to 140 h and the viable cell density is declined significantly after 240 h in our process. Both external CO 2 addition and low cell activity invalidated the analysis. A summary of pco 2 and R dco2 values predicted by Equations (11) and (13) is presented in Table IV. The results indicate that eight out of nine predicted pco 2 values fit well with the measured values, of which the differences were less than 11%. However, the predicted pco 2 value was 36% higher than the measured value in the second cell culture run at 184 h. This result might be caused by an error of cell density measurement. While the viability continuously decreased from 94% to 85% between 184 and 231 h, measured viable cell density showed a different trend. It first decreased between 184 and 209 h and then increased between 209 and 231 h. Except for this overestimated [CO 2 ] value, our model fits well with the experimental data. In comparison with the model used for batch culture in literature, which is only a function of time (Mostafa and Gu, 2003), our model reflected the dynamic change of viable cell density, gas flow rate, and CO 2 concentration during the fed-batch process. The predicted dco 2 removal rates based on Equations (11) and (13) are in agreement with the results from the cellfree study. Between 160 and 240 h, the specific power input, liquid level, and total gas sparging rate are approximately 24 W m 3, 3,800-L, and SLPM, respectively. The cellfree study under the condition with the same values of above parameters yielded the dco 2 removal rate of mmhg min 1 based on Figure 7. As shown in Table IV, simulation for the three cell culture runs returned results ( mmhg min 1 ) that were comparable to the cellfree study. Overall, the dco 2 removal rate was lower, and pco 2 level was higher for the cell culture process in the 5,000-L bioreactors compared to the bench-scale bioreactors, indicating that the dco 2 removal may be one of scale-up issues that led to the reduction of cell culture performance at the 5,000-L scale previously reported (Zhu et al., 2005). Equations (11) and (13) were successfully used to predict CO 2 concentrations and dco 2 removal rate in fed-batch cell culture, which provide an alternative to the traditional methods that either used cell-free condition or based on batch culture mode (Mostafa and Gu, 2003). Since [CO 2 ], ph, and k L a CO2 impact the dco 2 removal rate, under the assumption of constant ratio of k L a CO2 =k L a O2 (0.91) reported in literature (Spérandio and Paul, 1997), the factors that affect k L a O2 also impact the dco 2 removal rate. Based on Equation (14), the k L a O2 value is more dependent on superficial gas velocity than the specific power input in the 5,000-L bioreactor, because the b value is almost twofold higher than the a value. The current bioreactor configuration with lower airflow rate may be not high enough for efficient dco 2 removal. The further study on the optimal gas transfer parameters should be conducted to improve dco 2 removal in 5,000-L bioreactors. Conclusion When scaling up a mammalian cell culture process, characteristics such as the oxygen transfer coefficient, mixing time, and dco 2 removal rate will often change. In this study, a mixing analysis was conducted to identify scaleup issues that could impact cell culture performance at 5,000-L scale using the current bioreactor configuration. Table IV. Comparison of the predicted and the measured dco 2 values using cell culture data of three manufacturing runs. pco 2 (mmhg) a R dco2 (mmhg min 1 ) b Run # Cell age (h) X v (10 6 cell ml 1 ) % viability Measured Predicted Difference c a The predicted pco 2 value is based on Equation (11). b The predicted R CO2 value is the average value at time t(n) and t(n 1), which are calculated by Equation (13). c The value of % fitting is calculated by the following equation: % difference ¼ 100 (predicted measured)/measured. Xing et al.: Scale-Up Analysis for a CHO Cell Culture 743 Biotechnology and Bioengineering