|

Automated dynamic fedbatch process and media optimization for high productivity cell culture process development

Automated Dynamic Fed-Batch Process andMedia Optimization for High Productivity CellCulture Process Development Franklin Lu,1 Poh Choo Toh,2 Iain Burnett,1 Feng Li,1 Terry Hudson,1Ashraf Amanullah,1 Jincai Li11Oceanside Pharma Technical Development, Genentech, Inc., 1 Antibody Way, Oceanside,California 92056; telephone: þ(86)21-50462814; fax: þ(86)21-50461000;e-mail: [email protected] Technology Institute, Singapore 138668, Singapore KEYWORDS: CHO cell culture; fed-batch; dynamic feeding; ABSTRACT: Current industry practices for large-scale capacitance; automated sampling; Nova Flex mammalian cell cultures typically employ a standard plat-form fed-batch process with fixed volume bolus feeding.
Although widely used, these processes are unable to respondto actual nutrient consumption demands from the culture, which can result in accumulation of by-products and deple-tion of certain nutrients. This work demonstrates the appli- Several different modes of bioreactor operations have cation of a fully automated cell culture control, monitoring, traditionally been used to optimize production processes and data processing system to achieve significant produc- and maximize titer for monoclonal antibody production.
tivity improvement via dynamic feeding and media optimi- Methods have varied from the simple batch and fed-batch zation. Two distinct feeding algorithms were used todynamically alter feed rates. The first method is based process to the more complex continuous culture or upon on-line capacitance measurements where cultures perfusion culture (Bibila and Robinson, 1995). Among were fed based on growth and nutrient consumption rates these approaches, fed-batch processes are the most widely estimated from integrated capacitance. The second method employed due to their ease of operation, flexibility, and is based upon automated glucose measurements obtained robustness. In these fed-batch processes, high monoclonal from the Nova Bioprofile FLEX1 autosampler where cul-tures were fed to maintain a target glucose level which in antibody yields of up to 10 g/L have been achieved through turn maintained other nutrients based on a stoichiometric feeding enhancement and optimization, which increase cell ratio. All of the calculations were done automatically productivity, maintain cell viability, and extend culture through in-house integration with a Delta V process control longevity (Huang et al., 2010; Yu et al., 2011). These system. Through both media and feed strategy optimization, strategies typically involve feeding cultures with a concen- a titer increase from the original platform titer of 5 to 6.3 g/Lwas achieved for cell line A, and a substantial titer increase of trated feed media either daily, continuously or at fixed 4 to over 9 g/L was achieved for cell line B with comparable intervals with fixed feed volumes. However, this strategy product quality. Glucose was found to be the best feed does not take into consideration the variation in growth and indicator, but not all cell lines benefited from dynamic nutrient requirements that can occur in different cell lines or feeding and optimized feed media was critical to process even with the same cell line due to biological variability improvement. Our work demonstrated that dynamic feed-ing has the ability to automatically adjust feed rates accord- which in turn can lead to nutrient depletion or accumula- ing to culture behavior, and that the advantage can be best tion of by-products in the culture.
realized during early and rapid process development stages Dynamic feeding on the other hand has been utilized as a where different cell lines or large changes in culture condi- method to adapt a process to the real time nutrient demand tions might lead to dramatically different nutrient demands.
of cells through a variety of feeding algorithms (Wlaschin Biotechnol. Bioeng. 2013;110: 191–205.
and Hu, 2006; Zhang et al., 2004). Early efforts in dynamic ß 2012 Wiley Periodicals, Inc.
feeding largely focused on maintaining the cultures at a lowglucose and/or glutamine level in order to reduce by-product accumulation, in particular lactate and ammonia, Correspondence to: J. LiReceived 5 August 2011; Revision received 22 June 2012; Accepted 26 June 2012 while also trying to balance overall nutrient demand (Gong Accepted manuscript online 5 July 2012; et al., 2006; Kuwae et al., 2005; Sauer et al., 2000). This was Article first published online 1 September 2012 in Wiley Online Library achieved by growth and nutrient consumption modeling, coupled with more frequent manual sampling, calculation ß 2012 Wiley Periodicals, Inc.
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013 and manual feeding (Xie and Wang, 1994a,b). Additionally, measurement is advantageous as an indicator for dynamic Zhou (1994) and Zhou et al. (1995) established automated fed-batch process due to its continuous on-line signal and oxygen uptake rate (OUR) measurement, and used that to does not require any multivariate chemometric analysis estimate glucose consumption via the assumed stoichio- (Junker and Wang, 2006; Kiviharju et al., 2008; Teixeira metric ratio between glucose and oxygen, and then fed et al., 2009), thus simplifying the integration into feed accordingly to maintain a low glucose level. However, algorithms. We also demonstrate the use of an automated periodic manual sampling for glucose measurement was on-line sampler to measure viable cell counts and glucose at required to adjust the glucose to oxygen ratio as it varied pre-determined intervals which could then in turn be used during the course of the culture. Other efforts included by the bioreactor control system to modulate the feed rate of measuring and monitoring single specific nutrients, for a complex stoichiometrically balanced feed. Previous example, glutamine, with an automated aseptic online methods of dynamic feeding have all required user sampling loop, and feeding glutamine accordingly to intervention to manually adjust feed rates making these maintain a low level (Chee Furng Wong et al., 2005; Lee types of experiments labor intensive to implement into et al., 2003). More recently, dynamic feeding based on the widespread development (Yu et al., 2011), but a fully manual calculation of integral viable cell concentration has automated system makes an adaptive platform feasible. We been demonstrated to produce an optimized process present two different models of control, a predictive method with titers above 10 g/L (Huang et al., 2010). In all cases, used to estimate future nutrient demand based on past better growth and productivity were demonstrated with consumption and growth, and a feedback mechanism used successful control of the by-products lactate and ammonia.
to maintain the setpoint of an indicator metabolite. Both of However, these feeding strategies are still very labor these methods use a feed that balances key components intensive involving frequent sampling, recalculation, and relative to one another through an iterative analysis of manual adjustments.
nutrient consumption. Taken together, we aim to demon- In spite of the various successful applications, dynamic strate a rapid, systematic method of cell culture process feeding as a process development platform has not been development and optimization with the capability to rapidly explored extensively. Dynamic feeding has the benefit of develop high titer processes.
adapting the feed strategy simultaneously with the evalua-tion of other process changes to reach an optimal process Materials and Methods with significantly fewer experiments. For instance, atraditional problem with the use of a predefined platform Cell Line and Culture Conditions during clone selection is that with a fixed feed strategy, theclone that best fits the platform will be chosen as opposed to Two recombinant CHO cell lines, cell line A and cell line B, the clone with the best potential. Alternatively, by allowing expressing two distinct IgG antibodies were maintained as the process to adapt to the growth and consumption of each continuous seed trains in a proprietary serum-free medium individual clone, the most efficient path to the optimal clone under selective pressure of either methotrexate or methio- and process can be achieved simultaneously. In addition, nine sulfoximine. The cells were passaged every 3–4 days in due to further developments in the technology of on-line shake flasks at a seeding density of 0.3E6 cells/mL. These probes, bioreactor control systems, and automated sampling seed train cultures were kept in incubators controlled at equipment it is of interest to evaluate such industrial 378C and 5% carbon dioxide, and were shaken at a speed of platform processes with a completely automated dynamic 150 rpm. The inoculum for the production cultures were fed-batch process. These advancements can supplement cultivated in 3 L bioreactors (Applikon, Foster City, CA) existing process measurements and adapt the cultures based without selective pressure with a starting densities of 0.8 to on real-time nutrient demand.
1E6 cells/mL controlled at 378C at pH 7 (Mettler Toledo, In this work we demonstrate several distinct methods of Bedford, MA) and 30% dissolved oxygen (DO; Broadley on-line measurements in order to determine a dynamic feed James, Irvine, CA) with an agitation of 275 rpm. Fed batch rate for optimal cell culture protein productivity including production cultures were seeded from inoculum cultures at capacitance and online measurements of viable cell density 1.0 to 2.0E6 cells/mL in a proprietary chemically defined and glucose. Capacitance measurement has previously been medium without selective pressure. These cultures were proven to be a useful tool in providing real-time biomass controlled at the same conditions as the inoculum cultures monitoring in fed-batch fermentations (Arnoux et al., 2005) except for the temperature setpoint, which was shifted from and cell culture (Opel et al., 2010). Capacitance measure- 37 to 358C on Day 2 or 3 of the production culture. For the ments have also been successfully employed to obtain online initial baseline studies and improved feed testing, glucose cell concentration for feed rate calculations in the fed-batch was supplemented separately from the rest of the feed fermentation of Monascus sp. for yellow pigment produc- solution as a 500 g/L solution when the glucose concentra- tion (Krairak et al., 2000). In this work we incorporate the tion decreased to below 3 g/L. For later glucose feedback use of capacitance measurements as a surrogate for cell experiments, the glucose was incorporated into the feed growth to dynamically determine feed rate through a solution. The culture duration for these fed-batch processes completely automated, predictive method. Capacitance ranged between 14 and 20 days.
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013 Cell Culture Media where R1 is the nutrient consumption rate defined inEquation (1), IVCþ1 is the predicted integral of viable cell All media for the fed-batch processes were proprietary and concentration (or surrogate such as integrated capacitance), chemically defined. The feed medium was optimized as from the current time to the next timepoint, V0 is the discussed in the Results Section. Basal media was pH and current volume of the culture, and YS is the desired osmolality adjusted before use. Feed media was formulated setpoint of the nutrient concentration. The final term, without pH and osmolality adjustment.
Y0V0
ðYSV0Þ, represents a correction factor to adjust feedvolume in the case that the predicted nutrient consumption Fed-Batch Process rate is inaccurate.
The standard fed-batch production process used a bolus For glucose-based autosampler feedback, a separate feed addition at 10% (or 5% with concentrated feed algorithm was used in order to simplify the feed method.
medium) of culture volume, every 3 days starting on Day 3.
At each sample point, an amount of feed was added to the Initial experiments continued the bolus feed until Day 9 but culture to bring the glucose concentration to a target level as later fed batch experiments with optimized feed medium described by Equation (3) continued the feeds every 3 days until the end of the culture.
Pumps were calibrated before each run through Delta V (Emerson, St. Louis, MO), and the feed source was placed on scale to monitor accuracy of additions. The dynamic ¼ target glucose ðg=LÞ
measured glucose ðg=LÞ 0 mL glucose concentration in feed ðg=LÞ fed batch algorithm employed in the experiments was setup through customized Delta V software. This algorithm enabled the culture to be fed based on one of two methods, apredictive method or a feedback-based algorithm. In the For viable cell concentration (VCC) based autosampler predictive method, the viable cell density was monitored feedback cases, the calculation was based on a pre- using a BioProfile FLEX1 (Nova Biomedical, Waltham, determined per cell consumption rate and concentration MA) automated cell counter or using the capacitance probe.
of asparagine in the feed, as described by Equation (4).
At predetermined intervals, the current measure of the celldensity (X0), either from trypan blue counts or capacitance,and the previous measure of cell density (X mL to add ¼ ½VCC  V 0  R=Freq  ½Asn to calculate a specific growth rate to predict the future celldensity or capacitance (Xþ1). The predicted future valuecould then be used to determine a future integrated VCC : cell count ðcells=mLÞ capacitance or integrated viable cell count. A nutrient V0 : culture volume ðmLÞ consumption rate (R R : asparagine consumption rateðmmol asn=cell dayÞ 1) could then be calculated using the current nutrient concentration (Y Freq : sampling frequency ð1=dayÞ 0), previous nutrient concentration (Y
Asn : asparagine feed concentrationðmmol asn=mLÞ 1), previous feed volume ( F
1), feed concentration (Yc), current tank volume (V0), and previoustank volume (V
1) to determine the total amount of feed percell per unit time required as shown by Equation (1) In Process and Offline Measurements Nutrient consumption rate Online capacitance (Fogale Biotech, Cambridge, MA)readings were taken continuously during the runs of the ¼ previous
current þ fed ¼ Y
1V
1
Y0V0 þ F
1YC predictive model to facilitate the feeding algorithm. The capacitance reading was taken directly from the probe measurement with parameters set as described in Opel et al.
Alternatively, nutrient consumption rates from previous (2010) and every 4 h these readings were used as described in experiments could be used as estimates. Once the nutrient the fed-batch process above. The offline VCC, viability, consumption rate was determined, the final feed rate could pH, ammonia, osmolality, pCO2, pO2, Caþþ, Kþ, Naþ, be calculated by Equation (2): glutamine, glutamate, glucose and lactate were all moni-tored using a BioProfile FLEX1 (Nova Biomedical).
Autosamplers were setup and cleaned as described in Feed volume ¼ predicted consumption ðmmolÞ concentration in feed ðmmol=mLÞ where Derfus et al. (2009). Amino acid analysis and vitamin predicted consumption ¼ consumption rate analyses (by high performance liquid chromatography,HPLC), as well as trace element analysis (by inductively  predicted IVC
ðcurrent nutrient concentration coupled plasma mass spectrometry, ICP-MS) were also
desired nutrient concentrationÞ performed after each experiment. Titer was measured ¼ ðR1IVCþ1Þ
ðY0V0
ðYSV0ÞÞ through HPLC with Protein A columns.
Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering Antibody Product Quality Analysis defined feed media and to evaluate the impact of feedtiming. The cultures were fed with the same feed medium, The product quality attributes of antibodies, including level either by bolus feeding (10% of working volume on Day 2, 5, of aggregation, acidic and basic variants, and various and 8), daily feeding (adjusted daily based on VCC), or glycoforms were measured and compared. The various continuous feeding (1.8 mL/L/h starting at Day 2). All cases product quality analysis methods were previously described were designed to deliver similar total feed volumes (Fig. 1F) (Li et al., 2012).
and resulted in similar growth and titer profiles (Fig. 1A andB). This initial data suggested that when the total amount ofnutrients is controlled, a variety of feed timings and volumescould still be used to produce comparable results. While Errors, Uncertainties, and Reproducibility of Results growth and titer were not impacted by feed timing, amino Most data presented here are from representative singlet acid time course profiles revealed that all three cultures were conditions as the overall effort was an iterative process so exhausted or limited in several key nutrients including each experiment varied slightly. Error bars presented are cysteine, asparagine, and tyrosine during the late stage of the based on error of analytical measurements only and are run (Fig. 1C–E) which may have led to slowed product described in each corresponding figure. We feel confident in accumulation during the late stage of the culture.
the results and general applicability of the method, as severalsets of iterative conditions were studied. We saw similarimprovements in two cell lines (Fig. 9) with three different Feed Media Design and Optimization on-line inputs (capacitance, glucose, and OUR) and theperformance trend was consistent throughout the optimi- The platform feed medium was redesigned and refined after zation process (Table I).
several experiments according to the flow chart shown inFigure 2 to obtain a stoichiometrically balanced feedmedium. For this iterative method, several experiments were required in order to assess the effectiveness of thefeeding method for maintaining consistent nutrientconsumption rates since adjustment of the feed could Baseline Data (Platform Feed Formulation, potentially change the relative consumption rates. Daily amino acid concentrations were used to calculate daily In order to provide baseline data, cell lines A and B were consumption rates that were then averaged to determine evaluated in a standard fed-batch process as described in the an overall consumption rate. The relative ratios for all Materials and Methods Section. These cell lines were chosen components of the feed medium were then normalized to an due to their high titer and high nutrient consumption rates indicator metabolite. When cultures were fed based on the in order to maximize the potential improvements of indicator metabolite, other components were supplemented dynamic feeding. The results of cell line A are shown in in stoichiometric amounts to avoid exhaustion or over Figure 1. Nutrient data was also collected in the initial accumulation. In these studies, asparagine was selected baseline experiments in order to redesign the chemically as one benchmark metabolite because of its role as one Evolution of feed media.
Feed concentration of unlisted metabolites Zinc sulfate (mg/L) Ferric citrate (mg/L) Other amino acids 1.4 leucine, lysine, serine Glycine, threonine, valine, (others unchanged) isoleucine, leucine, serine 1.8–2.5 of CDF1 (others 1.5 of CDF1) Osmolality (mOsm/kg) Titer (cell line B)a 4.1  0.2 (3 feeds to Day 9) 6.0  0.3 (3 feeds to Day 9) 7.0  0.4 (5 feeds to Day 14) 8.6  0.4; 9.1  0.5 aTiter ranges (t) are based on analytical error.
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013

Figure 1. Three cultures of cell line A, designed to receive about the same volume of a chemically defined feed formulation (CDF1) were evaluated with either bolus feeding(10% WV on Day 2, 5, 8), daily feeding (adjusted daily based on VCC), or continuous feeding (1.8 mL/L/h starting at Day 2). A: VCC; (B) titer profiles of baseline fed-batch runs weresimilar independent of feeding method; (C) asparagine; (D) tyrosine; (E) cystine amino acid profiles indicating nutrient limitation in later part of culture; and (F) total feed volumeprofile. Error bars are based on analytical accuracy of measurements: VCC  5%, titer  5%, amino acids  5%, and feed volume (by weight)  5 g.
of the key energy sources in the tricarboxylic acid (TCA) The evolution of the feed media is described in Table I. As cycle and its relatively consistent consumption rate shown in Figure 1, cultures with the initial platform feed throughout our baseline study cultures. In addition, showed depletion of several amino acids. Therefore, the asparagine is regarded as a key amino acid for product original platform feed was adjusted with higher levels of synthesis, as depletion of asparagine had been shown to asparagine, tyrosine, leucine, lysine, and serine to create lead to decreases in cell specific productivity (unpublished chemically defined feed #1 (CDF1). In addition, zinc and observations). Other metabolites such as glucose were iron levels were also increased in CDF1. With CDF1, the also used as an indicator due to its key role as a primary culture was able to maintain appropriate levels of most of energy source.
the amino acids, for example, asparagine, tyrosine, serine, Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering

and isoleucine. However, the total feed volume increasedabout 40% compared to the control culture (data notshown), resulting in significant dilution and thus limitingincrease of volumetric productivity. The majority of the feedcomponents were then concentrated by 2 in order tominimize dilution effects, and relative ratios were furtheradjusted (chemically defined feed #2, CDF2). CDF2 wastested for cell line B with asparagine as indicator metabolite,and average asparagine consumption rate from platformbolus feed culture was used as pre-determined nutrientconsumption rate. However, significant accumulation ofasparagine as well as serine and isoleucine was observed, asactual asparagine consumption rate on cell line B culturewas lower than predicted. The overall concentration of thefeed medium was then lowered slightly to 1.5 of CDF1 andratios of several components were further refined forthe final version, chemical defined feed #3 (CDF3), toaccommodate the changes in consumption rates. This Figure 2. Iterative method of a complex stoichiometrically balanced feed.
resulted in balanced nutrient profiles, as shown in Figure 3of asparagine, tyrosine, serine, and isoleucine profiles.
In addition, vitamin and trace metals profiles were alsomonitored and adjustments were made throughout the feedmedium optimization (data not shown). The titer results in Figure 3. Amino acid profiles of (A) asparagine, (B) serine, (C) tyrosine, and (D) isoleucine during dynamic feeding with evolved feed medias CDF1, CDF2, and CDF3 with bothcell line A and B. Error bars are based on analytical accuracy of measurements: amino acids  5%. GTR (greater than measurement range of 10 mM).
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013 Table I showed the performance improvement trends for cell that of the typical PID control. Two primary advancements line B throughout the optimization process.
were key to these developments: the integration of on-lineprocess monitoring which was able to provide realtime analysis of key variables (i.e., capacitance) and theautomation of at-line sampling that is typically done by Design of Dynamic Feeding algorithm hand by researchers using Nova Bioprofile Automated Two different types of dynamic feeding algorithms were Sampler system. In collaboration with Nova Biomedical, we employed in this work. The predictive method utilized an were able to build a custom OPC module to read additional online indicator of cell growth and calculated the expected parameters that could be measured on an adaptable growth rate and cell density in order to estimate future nutrient demand. The alternative approach was a feedback- Figure 4 depicts the set up of the automated fed-batch based approach in which the historical nutrient consump- process. Initially, a signal (capacitance, OD, OUR, cell tion is used to determine amount of feed to maintain an count, or glucose) was collected in order to assess the indicator metabolite at a target. With the automated current state of the bioreactor. Once the current state of the sampling method, the sampling interval is set frequent bioreactor had been assessed, previous data could also be enough (6 h) to avoid dramatic changes in cell mass and used to predict future nutrient demand. In the case of the target concentration is set high enough to avoid capacitance, the specific growth rate was coupled with a exhaustion between feeding intervals. In either method, predetermined nutrient consumption rate to account for dynamic feeding is not being used in a nutrient limited changes in cell mass before the next feed point. Finally, after mode to alter metabolism, but solely as a means to each pre-determined interval, the feed pump was activated determine how much and how often to feed, which may to supply the required volume of feed.
avoid potential changes in product quality due to altered In order to account for noise and aberrant readings, metabolism (Chee Furng Wong et al., 2005). In addition, correction factors were incorporated into both algorithms.
the dynamic feed was restricted to small bolus feeds that For the predictive methods an additional term were built in occurred several times a day. Consequently, nutrient to the algorithms to handle over or underfeeding (the composition was never fully static, but was controlled second term of Equation 2).For the feedback method, the within desired ranges to optimize productivity.
response to fluctuation is built into the feedback control so One of the key advancements of this work was integration that if at any one data point, an excess of feed is added, the of our feeding algorithms into our control system. Delta V is feed volume added will be reduced or eliminated upon a process control platform that can be easily scaled for large the next measurement. Neither of these algorithms are able bioreactor networks. In this work, we were able to build to immediately discard an errant data point, which is a custom modules within the control system that linked to weakness of the system, as an incorrect measurement could novel inputs to build an additional layer of control beyond lead to a large spike of unneeded feed. While not Figure 4. Design of automated dynamic feeding system.
Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering

implemented for these experiments, a criteria could be productivity profiles (Fig. 5A and B). Ammonium increased implemented that would trigger a re-calibration and more rapidly in late phase of the dynamic feeding culture recalculation of feed addition if, for instance, the calculated (Fig. 5C), potentially due to higher levels of asparagine consumption rate changed more drastically than expected.
available, which may have resulted in higher rate of These sort of error checks would be simple for an automated deamidation to aspartatic acid. Asparagine concentration method to incorporate once bi-directional communication was well maintained throughout the cultures, although there with analytical instruments is enabled. Currently, this is were more day-to-day changes for the bolus feeding case possible with the Nova Flex and was in development at the compared to the dynamic feeding case (Fig. 5D). Other time of this work.
nutrient and metabolism profiles were also similar betweenthe two cultures (data not shown). Titer reached 6.1 and6.3 g/L for the bolus and dynamic feeding cases, respectively, Capacitance Based Predictive Method Using Optimized which represents about 25% increase when compared to the Feed Medium: Cell Line A platform bolus feeding cultures shown in Figure 1. Theimprovement in titer appeared to be primarily due to The automated system described above was first attempted optimization of feed formulation and cell line A was not using the predictive model with online capacitance sensitive to the feeding frequency using the optimized feed measurement. In addition, the feed medium formulation was optimized from the original platform feed medium afterseveral iterative experiments. Figure 5 shows the resultsusing the optimized feed medium CDF3 for cell line A. The Capacitance Based Predictive Method Using Optimized dynamic feed was based on capacitance measurement and Feed Medium: Cell Line B was added every 4 h. A bolus feed culture was also performedfor comparison, where a feed volume of 5% initial culture The capacitance-based dynamic feeding method was also volume was delivered every 3 days until Day 15. Overall, the applied to cell line B with the same optimized feed medium dynamic feeding case and the manual bolus feeding case had CDF3. Cell line B had marked difference on growth and very similar performance, as shown by the growth and nutrient consumption rates compared to cell line A (data Figure 5. Two cultures of cell line A were tested with optimized feed medium (CDF3) using either the standard bolus feeding method (5% every 72 h) or using the dynamicfeeding method with capacitance as on-line feedback signal (fed every 4 h). Comparison of (A) VCC, (B) titer, (C) ammonium, and (D) asparagine profiles. Error bars are based onanalytical accuracy of measurements: VCC  5%, titer  5%, ammonium  5%, and amino acids  5%.
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013

not shown), and thus the pre-determined feeding rate had to and physiological state in addition to biomass (Opel et al., be adjusted for each cell line. Specifically, the asparagine feed 2010). Titer was also dramatically improved with the rate was reduced by about 30% to account for lower specific optimized feed formulation. Combined with dynamic asparagine consumption rate. For the control case, manual feeding, cultures were able to reach titers up to 8.6 g/L, a bolus feeding was done every 3 days at 5% initial culture two-fold improvement over the baseline process (Fig. 6C).
volume using the same CDF3 feed medium. The results are Cell specific productivity was similar during the early stage shown in Figure 6. In addition, results from the original of culture, and the capacitance based predictive method platform bolus feed culture are also depicted in Figure 6 for maintained productivity at higher levels and for a longer period than bolus feeds (Fig. 6D). The bolus feed case with Cell growth was significantly higher in the dynamic the same optimized feed was also improved from the feeding case than both the improved bolus feed case and the baseline, reaching 7 g/L titer, indicating that the rebalancing initial platform bolus feed case (Fig. 6A). Peak density was of the feed as well as higher inoculation density were also reached around Day 8 in both improved bolus feed and significant contributions to the increase in titer. However, dynamic feeding cases. Osmolality was maintained below even in the optimized bolus feed case shown here, rapid 400 mOsm/kg at the end of the culture, indicating absence of changes of over 50 mOsm/kg still occurred after each significant metabolite accumulation (Fig. 6B). Capacitance bolus feed, potentially negatively impacting growth and profiles tracked closely with offline VCC values until peak productivity relative to the dynamic feeding case. In VCC was reached. Capacitance then continued to increase addition, the total feed volume was not controlled as in for several days before beginning to decrease around Day 11.
the baseline studies. The dynamic feed culture actually used The divergence between capacitance and VCC profiles had approximately 50% more total feed volume (Fig. 6E) over been reported previously, as capacitance depends on cell size the entire culture duration compared to the bolus feed Figure 6. Two cultures of cell line B were tested with CDF3 using either the standard bolus feeding method (5% every 72 h) or using the dynamic feeding method withcapacitance as on-line feedback signal (fed every 4 h). Both cultures also incorporated a higher inoculation density of 2e
6 cells/mL to increase build up of cell mass. The platformbolus feed process (with initial platform feed medium and lower inoculation density) is shown for reference: (A) VCC (solid symbols) and capacitance (open symbols) profiles, steepdrops in capacitance profile in the optimized bolus case were due to volume changes due to feeds; (B) osmolality profiles, (C) titer, (D) average specific productivity, (E) total feedvolumes, and (F) asparagine concentration profiles. Error bars are based on analytical accuracy of measurements: VCC  5%, osmolality  5 mOsm, titer  5%, specificproductivity  7% (based on propagation of error), feed volume (by weight)  5 g, and amino acids  5%.
Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering

culture, although no amino acids dropped below detectablelevels (0.2 mM) in either case. While the lack of nutrientexhaustion demonstrated the successful redesign of the feedformulation using the iterative method (Fig. 6F), it alsoindicated that feeding the additional volume in the dynamicfeeding case led to greater nutrient consumption andincreased antibody production. These differences under-score the possibilities that different feed algorithms canresult in alternative feed volumes that could result in evenhigher titers. Consequently, simply searching for nutrientexhaustion might not be sufficient to determine when aprocess has reached an optimal feed regime.
A key aspect of process improvement is product quality.
Differences in product quality have been observed inprevious applications of dynamic feeding that was focusedon altering cellular metabolism to improve processproductivity (Chee Furng Wong et al., 2005). However,in our case our feeding algorithms are not designed tooperate in nutrient limited mode, but to more efficientlypredict the demand of the culture and adjust feed accordingto shifts in cell density and cell physiology. Consequently, inour cases only minor differences in key product qualityattributes were observed that were all within acceptableranges for typical molecules (Fig. 7).
Autosampler Based Feedback Control Method UsingOptimized Feed Medium: Cell Line B As an alternative to a predictive model of nutrient demandused in previous studies, we explored a direct feedbackmechanism where feed rates were altered dynamically tomaintain a preset target of an indicator metabolite. From thelimited number of metabolites directly measured by theBioProfile FLEX instrument, glucose was chosen as a keyindicator of metabolism. Glucose has the advantage of beinga direct indicator of nutrient consumption that inherently Figure 7. Product quality comparison between dynamic feeding and bolus incorporates changes in cell growth and volumetric feeding cases. A: %Main peak of charged variants as measured by iCIEF at Day12, 14, and 16 using the capacitance based predictive method for the two test cell lines.
consumption rate. From the capacitance based predictive B: %Monomer analysis of product by size exclusion chromatography, (C) %GO glycan dynamic feeding study, it was observed that there was a analysis by capillary electrophoresis. Error bars are based on analytical accuracy of relatively constant ratio between glucose consumption rate measurements: monomer  0.2%, charge variants  1%, and glycoforms  1%.
and the asparagine consumption rate throughout theculture. This suggested that glucose could be a surrogateindicator for the previously asparagine balanced feed this method, additional components could be modified or formulation (CDF3) to dynamically feed the cultures.
added easily to a chemically defined feed as long as they are Feed rate in glucose feedback control is determined by balanced relative to other nutrients based on their Equation (3) described in the Materials and Methods and by consumption rates.
in-process glucose measurements. The glucose feedback Figure 8 shows the results of the autosampler based method benefits from incorporating multiple sources of dynamic feed back control study. In addition to feeding variability such as changes in cell mass and metabolism into based on the measured glucose concentration, VCC was also a single direct input that could be used to determine the feed used as a feeding indicator in this same study. For the glucose-based feeding, the target glucose level was set at 5 g/L, The feed medium used in this study had one change from which was chosen based on previous capacitance based CDF3, which was used in the experiment described in feeding study. For the VCC-based feeding, feed rate was Figure 6. Glucose, which had previously been added calculated automatically according to Equation (4) in the separately as a second feed, was added to the feed medium Materials and Methods Section. In addition, a fourth case and balanced stoichiometrically with other metabolites. The from a separate experiment, labeled as ‘‘Manual adjusted rest of the CDF3 composition remained unchanged. With feed, every 6 h'' was also included for comparison, where the Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013

Figure 8. Auto sampler based dynamic feeding results of cell line B. Cultures were inoculated at initial VCC of 2.0e
6/mL and feeding with CDF3 was initiated on Day 2. For theglucose and VCC-based feeding, cultures were sampled every 6 h via the autosampler, and fed automatically based on Equations (3) and (4) after each sampling. A fourth case froma separate experiment, labeled as ‘‘manual adjusted feed, every 6 h'' was also included for comparison, where the feed medium was also delivered every 6 h, but no auto samplerwas used and feed volume was preset once a day based on off-line glucose measurement. The daily adjustment on feed volume was done to target culture glucose level of between4 and 6 g/L. The bolus feed case was fed every 3 days at 6.7% initial culture volume. A: VCC and (B) titer profiles indicating cell mass and productivity improved with glucosefeedback. C: Glucose and (D) asparagine profiles illustrating direct and indirect metabolite control through dynamic feeding. E: lactate and (F) osmolality profiles. Error bars arebased on analytical accuracy of measurements: VCC  5%, titer  5%, glucose  0.2 g/L, asparagine  5%, lactate  5%, and osmolality  5 mOsm.
feed medium was delivered in the same frequency as with the nutrient accumulation (Fig. 8C, D, and F). This may be due glucose based feeding, that is, every 6 h. But no auto sampler to the fact that VCC is an indirect indicator of nutrient was used and feed volume was adjusted only once a day consumption and the total cell mass may not be an accurate based on off-line glucose measurement so that the same indicator of changing metabolic activity. As a result, the feed volume would be used four times before the next continual overfeeding and resulting osmolality increases adjustment. The daily adjustment on feed volume was done may have led to more rapid reduction in VCC (Fig. 8A) and to target culture glucose level of between 4 and 6 g/L. The reduced productivity. The bolus feeding case (fed every design can be interpreted as manual dynamic feeding.
3 days at 6.7% initial culture volume) did not perform as The highest titer of 9.1 g/L was achieved using the glucose well in this study, with culture viability dropping to <60% feedback method (Fig. 8B). In the auto sampler based VCC by Day 12 and titer reaching 4.1 g/L. Compared to the bolus feedback case, glucose, osmolality, and asparagine levels feeding case in Figure 5 where titer reached 7 g/L, a key were increasing late in the culture, indicating potential difference between the two cultures was the increased feed Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering

Figure 9. A: Ratio of feed concentrations and consumption rates of arginine, glycine, threonine, and tyrosine relative to feed concentration and consumption rates ofasparagine in feed versus consumption rates in various stages of culture for auto sampler based glucose feedback culture: (black) indicates ratio in feed medium, (white) indicatesconsumption rate across entire run, (light gray) consumption rate from Day 0 to Day 3, (gray) consumption rate from Day 3 to Day 11 and (dark gray) consumption rates from Day 12to Day 19 based on amino acid analysis of cell culture supernatants by HPLC. B: Ratio of glucose feed concentrations and consumption rates to asparagine feed concentrations andconsumption rates for auto sampler based glucose feedback case. Error bars are based on analytical accuracy of measurements and propogation of error: amino acids ratio  7%,glucose ratio  6.7% (based on glucose error of 0.2 g/L and average glucose measurement of 4.5 g/L).
volume from 5% to 6.7% initial culture volume per feed, with the objective of increasing total feed volume to matchthe dynamic feeding case. However, the bolus feed case Most industrial mammalian cell culture processes still use might have suffered from apparent high osmolality and high manual bolus feeding methods for nutrient feeding (Kelley, nutrient concentration shock accompanied by each feed, as 2009). This is in contrast to an industrial microbial each bolus feed added about 100 mOsmo to the culture fermentation process where automatic, dynamic feeding (Fig. 8F). The fourth case with the daily manual adjusted often is necessary to avoid nutrient over-accumulation and feed rate performed quite well, reaching titer of 8.2 g/L, waste product formation. Examples include widely used which was only slightly lower than the 9.1 g/L reached by the dissolved oxygen (DO-stat) or pH (pH-stat) based feeding glucose-based dynamic feeding case and comparable with strategies for E. coli fermentation (Chen et al., 1997; Kim the 8.6 g/L reached by the capacitance-based dynamic et al., 2004). The difference in practice is largely due to the feeding case in Figure 6.
fast metabolism of microbial cells in comparison to When looking at the metabolite and nutrient profiles, the mammalian cells and hence the perception that mammalian glucose feedback case and the daily manual adjusted case cultures do not need to be fed so frequently. In this work, tracked very closely in glucose, lactate, asparagine, and part of our objective was to understand whether given the osmolality (Fig. 8C–F). In contrast, the VCC-based dynamic same feed medium, the feeding strategy and frequency feeding case and the bolus feed case both had higher glucose, difference would lead to different culture performance.
asparagine, and osmolality levels.
Little has been reported on such comparison studies. In their Even with glucose as a direct indicator of metabolism, the work, Po¨rtner et al. (2004) did compare the various control relative consumption rates of other components varied and feed strategies for a hybridoma cell line expressing a during the culture, and in some cases the overall Mab. However, the studies with the various feed strategies consumption rates also differed slightly than predicted were not conducted in parallel; rather they were done over from previous studies (Fig. 9A and B). For instance, the several years, when specific Mab productivity varied average consumptions rate (white bar) of arginine and significantly. The feed medium used for the studies also glucose matched well with the ratios predicted in feed varied slightly. Thus strict comparison on productivity formulation (black bars), but other components such as across the different methods was not possible. Many other tyrosine and glycine showed slight differences from expected reports compared performance between batch culture and relative consumption. In addition, as expected, the relative optimized fed-batch culture, but again not among fed-batch consumption rates also varied throughout the length of the culture modes with the same feed medium (Xie and Wang, culture from relatively minor changes of glucose to more 1994a,b, 1996; Xie et al., 1997; Yu et al., 2011; Zhou et al., dramatic shifts of components such as glycine.
1995, 1997). To our knowledge, our work is the first that Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013 Summary of comparisons between preset bolus feed and automatic dynamic feed cultures.
Bolus (daily adjusted) 1.91 (1.18 on Day 12) 6.5 (5.4 on Day 12) 640 (395 on Day 12) 1.64 (1.01 on Day 12) 9.1 0.5 (6.3  0.3 Day 12) Bolus (daily adjusted) 530 (335 on Day 12) 1.36 (0.86 on Day 12) 8.2  0.4 (5.4  0.3 Day 12) aAdjusted based on 1.45 L initial culture volume, 5 g.
bTiter ranges (5%) are based on analytical error.
cNutrient depletion seen on all three cultures.
dBolus feed rate changed from 5% to 6.7%. High nutrient accumulation, high osmolality seen on the culture with bolus, every 3-day feeding, which resulted in early termination of bolus culture.
eThe glucose concentration was adjusted to 133 g/L (739 mM) in CDF3 for this experiment to combine a previously separate glucose and nutrient feed into one solution.
offers direct comparison between manual bolus feeding and noted, however, that for manual bolus feeding to work automatic dynamic feeding with industrially relevant and optimally, appropriate feed dose, and feeding frequency still high producing cell lines. We offered four sets of data need to be established for different cell lines and processes.
(Figs. 1, 5, 6, and 8) across two cell lines for the direct This is especially true for highly concentrated feed medium, comparison. Table II summarized the studies by highlight- for example, CDF3, where feeding in every 3 days at 5% ing the feeding strategy, frequency and total feed volume.
initial volume seemed not enough for cell line B, and feeding Each set of the studies used exactly the same feed medium, in every 3 days at 6.7% initial volume resulted in large with the only difference being feeding strategy and osmolality and nutrient shock.
frequency. The studies shown in Figures 1 and 5 show In essence, the key advantage of dynamic feeding, when basically identical performance between manual bolus coupled with feedback control, is the ability to automatically feeding and dynamic feeding for cell line A, whereas studies adjust feed rates according to culture behavior, aka, feed by on Figures 6 and 8 show higher titer with the dynamic demand, thus avoiding either under-feeding or over-feeding feeding case than the bolus feeding case for cell line B.
of the cultures. The advantage can be best realized during Consequently, improvements from dynamic feeding will early and rapid process development stages where different vary based on relative sensitivity of cell line, which may be a cell lines or large changes in culture conditions might lead to combination of sensitivity to large shifts in osmolality from dramatically different nutrient demands, making it impos- bolus feeds and sensitivity to variations in metabolites that sible to determine feed rates beforehand. Conversely, once a may alter nutrient uptake and metabolism. For some cell process is well established, a fixed rate, predefined bolus lines, such as cell line A, this difference may be negligible.
feeding method could very well deliver similar performance.
An interesting observation from this work was that the The actual dynamic feed profiles (Figs. 1F and 6E) of this relative performance among the cases appear to be work appear to be relatively constant, where a continuous correlated with total feed volume, with the total feed feed could also potentially replicate results while simplifying volume approximately the same between the feeding implementation. However, these simplifications require first methods for Figure 1 studies and Figure 5 studies, performing dynamic feeding experiment to determine the respectively, whereas for Figure 6 studies, feed volume appropriate feed rate. In addition, even in well established was about 1.5 higher with the dynamic feeding case. This manufacturing processes, the bolus strategy does not take suggests that it is possible to achieve similar culture into consideration the variation in growth and nutrient performance when similar volumes of feed are delivered via requirements that can occur due to variations in parameters different feeding methods. Consequently, with further like inoculation density, raw materials or cell age, which in adjustments a manual bolus feed could potentially achieve turn can lead to nutrient depletion, or accumulation of by- the same titer as a dynamic feed with cell line B. However, products in the culture.
two key points need to be considered a priori. The total In our efforts, we evaluated both the capacitance-based bolus feed volume would need to be predetermined, and the predictive method and the auto sampler-based feedback frequency of bolus feeding needs to be optimized for control method. Similar performance was achieved between different feed medium and different processes. It should be the two methods. The predictive method requires accurate Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering modeling of culture growth rates or nutrient consumption The authors are grateful for the valuable contributions and technical rates (Dowd et al., 2001a,b; Kurokawa et al., 1994; Xie and discussions by Srikanth Chary, Gayle Derfus, Veronica Carvalhal, Wang, 1994a,b). However, since growth rates and nutrient Jason Goodrick, and Robert Kiss.
consumption rates vary during the course of any givenculture, which have impact on specific substrate uptakerates, an on-line or off-line measurement is needed to either automatically or manually adjust feed rates. In our work, thecapacitance signal was successfully used as a surrogate of Arnoux AS, Preziosi-Belloy L, Esteban G, Teissier P, Ghommidh C. 2005.
culture growth, and constant re-calculation of growth rates Lactic acid bacteria biomass monitoring in highly conductive media bypermittivity measurements. Biotechnol Lett 27(20):1551–1557.
was done automatically to allow for change of feed rates.
Bibila TA, Robinson DK. 1995. In pursuit of the optimal fed-batch process The asparagine consumption rate, used as an indicator for monoclonal antibody production. Biotechnol Prog 11(1):1–13.
metabolite for feed rate calculation, was initially derived Chee Furng Wong D, Tin Kam Wong K, Tang Goh L, Kiat Heng C, Gek Sim from historical data and then adjusted throughout evolution Yap M. 2005. Impact of dynamic online fed-batch strategies on of the feed medium formulations. On the other hand, the metabolism, productivity and N-glycosylation quality in CHO cellcultures. Biotechnol Bioeng 89(2):164–177.
feedback-based method does not rely on growth or nutrient Chen W, Graham C, Ciccarelli RB. 1997. Automated fed-batch fermenta- consumption rate models. Instead, it relies on more frequent tion with feed-back controls based on dissolved oxygen (DO) and measurement of the indicator metabolite, for example, pH for production of DNA vaccines. J Ind Microbiol Biotechnol 18: glucose, and the subsequent feeding that is designed to bring the indicator metabolite to a preset control level (Chee Derfus GE, Abramzon D, Tung M, Chang D, Kiss R, Amanullah A.
2009. Cell culture monitoring via an auto-sampler and an integrated Furng Wong et al., 2005; Lee et al., 2003; Zhou et al., 1995, multi-functional off-line analyzer. Biotechnol Prog 26(1):284– 1997). This strategy has been successfully used to control low levels of glucose and glutamine, but for this work glucose Dowd JE, Kwok KE, Piret JM. 2001a. Glucose-based optimization of CHO- was maintained at around 4–5 g/L in order to keep the cell perfusion cultures. Biotechnol Bioeng 75:252–256.
culture from either over-feeding or under-feeding. Glucose Dowd JE, Kwok KE, Piret JM. 2001b. Predictive modeling and loose-loop control for perfusion bioreactors. Biochem Eng J 9:1–9.
as a feed indicator is simple, requires the least prior Gong X, Li D, Li X, Fang Q, Han X, Wu Y, Yang S, Shen BQ. 2006. Fed- knowledge and is adaptable to both changes in cell density batch culture optimization of a growth-associated hybridoma cell and metabolism.
line in chemically defined protein-free media. Cytotechnology 52(1): It is also important to note that the overall improvements on productivity, for example, from 4 g/L of the baseline Huang YM, Hu W, Rustandi E, Chang K, Yusuf-Makagiansar H, Ryll T.
2010. Maximizing productivity of CHO cell-based fed-batch culture platform process to 9.1 g/L for the dynamic feeding process, using chemically defined media conditions and typical manufacturing is only possible with the optimized feed medium. The equipment. Biotechnol Prog 26(5):1400–1410.
automatic dynamic feeding method, however, made it Junker BH, Wang HY. 2006. Bioprocess monitoring and computer control: possible to evaluate the various feed medium formulations key roots of the current PAT initiative. Biotechnol Bioeng 95(2):226– without the need to establish a priori how these different Kelley B. 2009. Industrialization of mAb production technology: The feed media should be administered in terms of dose and bioprocessing industry at a crossroads. MAbs 1(5):443–452.
frequency. Therefore, the combined approach of dynamic Kim BS, Lee SC, Lee SY, Chang YK, Chang HN. 2004. High cell density fed- feeding and iterated feed medium optimization has the clear batch cultivation of Escherichia coli using exponential feeding com- synergistic benefit. Automation is also a main feature in our bined with pH-stat. Bioprocess Biosyst Eng 26:147–150.
studies, as many earlier efforts on dynamic feeding had to Kiviharju K, Salonen K, Moilanen U, Eerikainen T. 2008. Biomass mea- surement online: The performance of in situ measurements and rely on manual adjustments based on off-line measurements software sensors. J Ind Microbiol Biotechnol 35(7):657–665.
(Xie and Wang, 1994a,b, 1996; Xie et al., 1997; Yu et al., Krairak S, Yamamura K, Irie R, Nakajima M, Shimizu H, Chim-Anage P, 2011; Zhou et al., 1995, 1997a,b). Advances in automation Yongsmith B, Shioya S. 2000. Maximizing yellow pigment production have enabled more advanced feedback control mechanisms in fed-batch culture of Monascus sp. J Biosci Bioeng 90(4):363–367.
without a concomitant increases in labor. A fully integrated Kurokawa H, Park YS, Iijima S, Kobayashi T. 1994. Growth characteristics in fed-batch culture of hybridoma cells with control of glucose and sampling, data acquisition, and control system has the glutamine concentrations. Biotechnol Bioeng 44:95–103.
capability of responding to a wide variety of inputs.
Kuwae S, Ohda T, Tamashima H, Miki H, Kobayashi K. 2005. Development Alternative methods could also be developed for subgroups of a fed-batch culture process for enhanced production of recombinant of metabolites that could be broken up into separate feeds human antithrombin by Chinese hamster ovary cells. J Biosci Bioeng that vary in predetermined ratios to one another. An Lee YY, Yap MG, Hu WS, Wong KT. 2003. Low-glutamine fed-batch expansion in the number of available inputs could further cultures of 293-HEK serum-free suspension cells for adenovirus pro- improve feeding algorithms to develop a control system that duction. Biotechnol Prog 19(2):501–509.
is dependent on inherent cell metabolism. An evolution of Li J, Wong CL, Vijayasankaran N, Hudson T, Amanullah A. 2012. Feeding this feeding algorithm could define a platform process that lactate for CHO cell culture processes: Impact on culture metabolism adapts to natural variations in cell lines in order to ensure and performance. Biotechnol Bioeng 109:1173–1186.
Opel CF, Li J, Amanullah A. 2010. Quantitative modeling of viable cell optimal productivity and quality is reached in the shortest density, cell size, intracellular conductivity, and membrane capacitance time possible, translating technological advancements into in batch and fed-batch CHO processes using dielectric spectroscopy.
high throughput process development.
Biotechnol Prog 26(4):1187–1199.
Biotechnology and Bioengineering, Vol. 110, No. 1, January, 2013 Po¨rtner R, Schwabe JO, Frahm B. 2004. Evaluation of selected control Xie L, Nyberg G, Gu X, Li H, Mo¨llborn F, Wang DIC. 1997. Gamma- strategies for fed-batch cultures of a hybridoma cell line. Biotechnol interferon production and quality in stoichiometric fed-batch cultures Appl Biochem 40:47–55.
of Chinese hamster ovary (CHO) cells under serum-free conditions.
Sauer PW, Burky JE, Wesson MC, Sternard HD, Qu L. 2000. A high- Biotechnol Bioeng 56:577–582.
yielding, generic fed-batch cell culture process for production of Yu M, Hu Z, Pacis E, Vijayasankaran N, Shen A, Li F. 2011. Understanding recombinant antibodies. Biotechnol Bioeng 67(5):585–597.
the intracellular effect of enhanced nutrient feeding toward high Teixeira AP, Oliveira R, Alves PM, Carrondo MJ. 2009. Advances in on-line titer antibody production process. Biotechnol Bioeng 108(5):1078– monitoring and control of mammalian cell cultures: Supporting the PAT initiative. Biotechnol Adv 27(6):726–732.
Zhang L, Shen H, Zhang Y. 2004. Fed-batch culture of hybridoma cells in Wlaschin KF, Hu WS. 2006. Fedbatch culture and dynamic nutrient serum-free medium using an optimized feeding strategy. J Chemi Tech feeding. Adv Biochem Eng Biotechnol 101:43–74.
Xie L, Wang DI. 1994a. Fed-batch cultivation of animal cells using different Zhou W, Hu W-S. 1994. On-line characterization of a hybridoma cell medium design concepts and feeding strategies. Biotechnol Bioeng culture process. Biotechnol Bioeng 44:170–177.
Zhou W, Rehm J, Hu WS. 1995. High viable cell concentration fed-batch Xie L, Wang DI. 1994b. Applications of improved stoichiometric model in cultures of hybridoma cells through on-line nutrient feeding. Biotech- medium design and fed-batch cultivation of animal cells in bioreactor.
nol Bioeng 46(6):579–587.
Zhou W, Chen CC, Buckland B, Aunins J. 1997. Fed-batch culture of Xie L, Wang DI. 1996. High cell density and high monoclonal antibody recombinant NS0 myeloma cells with high monoclonal antibody production through medium design and rational control in a bioreac- production. Biotechnol Bioeng 55(5):783–792.
tor. Biotechnol Bioeng 51:725–729.
Lu et al.: Automated Dynamic Fed-Batch Process and Media Optimization Biotechnology and Bioengineering

Source: http://www.novabio.us/press/NovaBio_FLEX_CellCulture.pdf