Soft sensor assisted dynamic bioprocess control: Efficient tools for bioprocess development

Chemical Engineering Science 96 (2013) 190–198

Contents lists available at SciVerse ScienceDirect

Chemical Engineering Science journal homepage: www.elsevier.com/locate/ces

Soft sensor assisted dynamic bioprocess control: Efficient tools for bioprocess development Patrick Sagmeister a, Patrick Wechselberger a, Mohammadhadi Jazini a, Andrea Meitz b, Timo Langemann b,c, Christoph Herwig a,n a

Institute of Biochemical Engineering, Vienna University of Technology, Austria Research Centre of Pharmaceutical Engineering GmbH, Graz, Austria c BIRD-C GmbH & Co KG, Kritzendorf, Austria b

H I G H L I G H T S








We present a soft sensor control strategy for microbial bioprocess development. A first principle soft sensor using the carbon and degree of reduction balance was used. No strain specific knowledge is needed to run the generically applicable control strategy. Specific uptake rates in microbial cultures were controlled at static and dynamic setpoints. The power of the control strategy was demonstrated on bioprocess development tasks.

art ic l e i nf o

a b s t r a c t

Article history: Received 26 October 2012 Received in revised form 12 February 2013 Accepted 22 February 2013 Available online 4 April 2013

The exploitation of dynamic process conditions increasingly finds attention for strain screening and bioprocess development of biotechnological products. Prerequisite for successful dynamic experimentation is the controlled deflection of process parameters, e.g. temperature, pH or specific substrate uptake rates, while the latter can be considered the most challenging. Here, a soft-sensor based control strategy capable of dynamic control of specific uptake rates is presented, opening new perspectives for efficient process development. A soft-sensor on the basis of a redundant equation system involving the Degree of Reduction (DoR) and Carbon (C) balance was used for the estimation of the real-time biomass concentration and used for closed-loop control of specific uptake rates using a mass balance based control approach. The power of the presented soft sensing approach for bioprocess development was demonstrated on microbial bioprocess development tasks: (a) the control of the specific substrate uptake rates in a noninduced E. coli decellerostat culture aiming at the determination of the specific acetate production as a function of the specific substrate uptake, (b) the static control of the specific substrate uptake rate in an induced P. Pastoris AOX expression system, (c) the dynamic control of specific substrate uptake rates under induced conditions in a E. coli pET expression system and (d) the simultaneous dynamic control of the specific glucose and arabinose uptake rates in an E.coli pBAD expression system. In this contribution the dynamic controllability of multiple specific substrate uptake rates using a first-principle soft sensor independent of known or dynamically changing yield coefficients in recombinant bioprocesses is demonstrated for the first time. The presented control strategy holds potential to become a key process analytical technology (PAT) tool for bioprocess development. & 2013 Elsevier Ltd. All rights reserved.

Keywords: Bioprocess Technology Process control Mathematical Modelling Dynamic Experiments Process analytical technology Soft sensors

Abbreviations: eXH20,out, water fraction in offgas [mol/mol]; Fs,in, substrate inflow _ in , mass input [g/h]; [g/h]; Fb,in, base inflow [g/h]; Fa,out, outflow air [nL/min]; M _ out , mass output [g/h]; dM, change in mass [g]; PIMS, process information M management system; qs, specific substrate uptake rate [g/g/h]; qace, specific acetate production rate [g/g/h]; ro2, rate of oxygen uptake [mol/h]; rco2, rate of carbon dioxide production [mol/h]; V, volume [l]; Vm, molar volume [l/mol]; ρ, density [g/l]; w mass, fraction of C-source in the feed [g/g]; x, biomass concentration [g/l] n Corresponding author. Tel.: +43 15 88 01166 400, mobile: +43 676 47 37 217; fax: +43 1 58801 166 980. E-mail address: [email protected] (C. Herwig). 0009-2509/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ces.2013.02.069

1. Introduction 1.1. Dynamic experimentation in bioprocess development Bioprocess development aims at the identification and understanding of interactions between process parameters and product quality and process performance attributes. Since bioprocess developing time is limited and the multivariate investigation of a

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

bioprocess is a time, cost and labour intensive task, strategies are required to accelerate process development and reduce time-to market latency of biotechnological products. One way to speed up bioprocess development is bioreactor miniaturization for the benefit of high-throughput screening and optimization of fermentation parameters (Betts and Baganz, 2006; Kumar et al., 2004; Kusterer et al., 2008; Puskeiler et al., 2005a; Puskeiler et al., 2005b; Weuster-Botz et al., 2005). However, important quantitative bioprocess variables such as adaptation times, metabolic regulations and process performance dependencies on specific rates are not yet reported to be extractable through small (millilitres) scale parallel bioprocessing. Furthermore, the limited sample size and the limited space for analytical instrumentation restrict their applicability. Dynamic experimentation, hence the controlled deflection of process states, offers the possibility to study multiple individual process parameters or multiple levels of process parameters within one fermentation run as well as adaptation behaviour. Pulse and shift experiments are among the most common dynamic methods applied in bioprocess development. Within shift experiments, the cell population is submitted to a sudden up- or downshift in process parameters aiming at the investigation of growth kinetics (Feitkenhauer et al., 2001), transient energetic regulations (Soini et al., 2005) as well as physiological regulations and strain characterization (Herwig et al., 2001; Herwig and von Stockar, 2002; Herwig and Von Stockar, 2003). Within dynamic ramp experiments, process parameters are changed slowly to allow the cell population to adapt to the changing conditions. pH ramps, ramps of the dilution rate D as well as feeding ramps in fed batch experiments were used for the investigation of cell physiology (Osman et al., 2002), metabolic strain characterization (Duboc et al., 1998) as well as the fast extraction of bioprocess parameters in real-time (Wechselberger et al., 2010). Examples for the applicability of dynamic pulse experiments include the determination of strain specific parameters of recombinant strains. The extracted strain specific information was used for the development of a feeding profile as well as strain characterization (Dietzsch et al., 2011a; Dietzsch et al., 2011b). Prerequisite for dynamic experimentation is the controlled perturbation of (physiological-) process parameters for the initiation and observation of physiological changes. Typical control parameters in focus of dynamic experimentations include T, pO2, pH, as well as feed strategy related parameters. As regards feedstrategy related process parameters, the specific substrate uptake rate qs [g/g/h], hence the amount of substrate [g] taken up by the cell population [g] in a defined time interval [h−1], can be considered the most important feed strategy related physiological parameter (Wechselberger et al., 2012a). This scaleable physiological parameter stands in direct correlation with process performance (Dietzsch et al., 2011b; Wechselberger et al., 2012a). Recent approaches in dynamic experimentation on the basis of qs approximated the desired conditions through offline OD adjustment (Dietzsch et al., 2011b), or via fixed yield correlations used in feed-forward accellero- or decellorostats (Wechselberger et al., 2010). However, the assumption of constant yield coefficients does not hold true for induced systems, hampering the controllability of the parameter qs via feed-forward approaches. This especially holds true for dynamic experimentation on qs in induced systems, where yield coefficients can change freely over time. Furthermore, the determination of yield coefficients prior to dynamic experimentation constitutes additional experimental effort that has to be avoided for the benefit of a fast and efficient process development. While a lot of attention of the scientific community has been drawn to the development of soft sensor assisted control strategies

191

for bioprocess production processes, the development of physiological control strategies for bioprocess development tasks was not in focus so far. However, a significant improvement of bioprocess development (speed up of development times, higher titres, increased process understanding) through better controlled and therefore better investigated bioprocesses can be expected from the establishment of soft-sensor assisted control strategies, especially in combination with dynamic experimentation. This aim stands in accordance with the quality by design initiative where the development of process analytical (PAT) tools capable of “designing, analyzing and controlling” is highly encouraged. This contribution aims at presenting the first soft-sensor assisted control strategy for bioprocess development, complying with the demands of (dynamic) process development in respect to (a) independence of unknown or variable yield coefficients, (b) dynamic closed loop control of multiple specific substrate uptake rates and (c) the sole use of simple on-line analyzers that do not violate the sterile barrier of the reactor. 1.2. Soft-sensors: process analytical technology (PAT) tools for process monitoring and control Following regulatory initiatives (FDA, 2004), Process Analytical Technology (PAT) emerged as a key tool for science and risk based process development and manufacturing along Quality by Design (QbD) principles. Following the regulatory definition, PAT aims at “…designing, analyzing, and controlling manufacturing through timely measurements […] with the goal of ensuring final product quality” (FDA, 2004). It should be highlighted, that PAT is not limited to timely (online) measurements, but covers the whole framework how to efficiently generate and apply process knowledge, as outlined by several excellent recent contributions (Glassey et al., 2011; Mandenius et al., 2009; Rathore et al., 2010; Rathore and Winkle, 2009; Undey et al., 2011). This contribution presents a PAT bioprocess control tool applicable in process development for the fast generation of process knowledge through advanced physiological bioprocess control. The highly complex and non-linear nature of bioprocesses poses great difficulties to bioprocess monitoring and bioprocess control approaches. Next to substrate concentration, the biomass dry cell weight concentration can serve as the basis of bioprocess control approaches (Caramihai et al., 2007; Karakuzu et al., 2006; Komives and Parker, 2003). This important process variable can be estimated via correlation to an on-line accessible signal, e.g. the broth turbidity, broth fluorescence or permittivity using different hard-type sensors, as reviewed elsewhere (Sonnleitner et al., 1992). However, every available hard-type device for on-line biomass measurement has its drawbacks as reviewed recently (Kiviharju et al., 2008). Alternatively to hard-type sensors, the biomass concentration can be estimated by soft sensors. Soft-Sensors are process analytical devices that grant access to important non-measured process variables by mathematical processing of readily available process data. They typically come with reduced costs compared to hard type sensors and do not violate the sterility barrier of the system. A recent review on soft sensors in biotechnology is given by Luttmann et al. (Luttmann et al., 2012). Presupposing the correct estimation of process states (e.g. the biomass concentration), they pose great opportunities to be embedded within physiological bioprocess control strategies. The soft-sensor estimation relies on a process model, which can be derived from first principle relationships or statistical regressions derived from multivariate data analysis, e.g. from training data sets. Adaptive observers for the estimation of process states (Mailleret et al., 2004) can be used and were demonstrated to be especially powerful for bioprocess application in the

192

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

formulation of the Kalman filter (Dubach and Märkl, 1992; Ghoul et al., 1991; San and Stephanopoulos, 1984; Wilson et al., 1998). Furthermore, soft sensors based on artificial neural networks (ANNs) were successfully applied in bioprocesses (Gadkar et al., 2005; Liu et al., 2011; Nayak and Gomes, 2009). A recent review on the performance of soft-sensors for the applications on bioprocesses is given by de Assis and Filho (de Assis and Filho, 2000).

investigation of recombinant bioprocesses are given: (a) the investigation of acetate production and uptake in E. coli via a dynamic decellerostat culture (b) and (c) the control of static and ramp set points of the specific substrate uptake rate qs in induced a non-induced conditions for different systems (P. Pastoris and E. coli) and (d) the independent control of glucose and inducer uptake rates as necessary in the development of transcription controlled bioprocesses.

1.3. Soft sensors for bioprocess development tasks In our point of view, the task of soft-sensing in process control and monitoring in process development differs conceptually from soft-sensor applications in manufacturing. In the latter case huge amounts of process data are typically available that can be used to construct first-principle based or statistical process models. In process development no or very little prior data on the strain or process is available and soft-sensing acts as a tool for the development of strain- and process knowledge. This conceptual difference discards artificial neural networks (ANNs) and multivariate soft-sensing approaches for the investigation of unexplored systems within process development. In contrast, elemental balancing approaches for bioprocesses (Jobé et al., 2003) are attractive for the design of a process development soft sensing strategy: no training data sets are required, simple easily measurable input parameters and variables are used, they can be checked for gross errors via reconciliation procedures (van der Heijden et al., 1994) and multiple stoichiometries can be identified in real-time (Herwig et al., 2001). In contrast to data driven soft sensors such as ANNs, firstprinciple soft sensors do not require training data sets but rely on first principle relationships such as mechanistic models or elemental balances. A control strategy for the control of the specific growth rate in CHO cell processes was reported recently (Aehle et al., 2011). The respective approach demands prior knowledge on yield coefficients and presupposes that the yield coefficient is not changing over time. However, dynamically changing yield coefficients are frequently observed in induced cultures (Sagmeister et al., 2012, Jenzsch et al., 2006), restricting the applicability of fixed yield approaches for bioprocess development tasks. Elemental balancing first-principle soft-sensor approaches were used successfully for control of bioprocesses in respect to the avoidance of overflow metabolites (Jobé et al., 2003). Using a statistical test, Jobé et al. identified the metabolic state by testing different metabolic stoichiometries for their significance, a concept first reported by Herwig et al. (2001). Using this information, a controller acts on the feed exponent of an exponential feeding ramp. Although overflow metabolites can be successfully avoided and the metabolic state effectively controlled, the approach reported by Jobé et al. does not constitute an effective control of neither specific substrate nor specific growth rates. Soft-sensors in extended Kalman filter formulation were applied for the dynamic control of the specific growth rate m, where changes in yield coefficients were successfully handled (Jenzsch et al., 2006). However, this approach demanded at-line biomass concentration determination via optical density measurement to be made available to the sensor. In this contribution we present what is to our knowledge the first soft-sensor based control strategy capable of effective (dynamic) control of specific substrate uptake rates that is independent of training data sets, elaborate strain specific information, fixed or known yield coefficients or offline measurements. Therefore, the first soft sensor control strategy that qualifies as process development tool is presented. To demonstrate the applicability of the presented soft sensor control strategy, applications aiming at the fast design and

2. Materials and methods 2.1. Strains E. coli C41 (F– ompT hsdSB (rB−mB−) gal dcm (DE3); obtained from Lucigene, Middleton, WI, USA) carrying the plasmid pBMPPET or pBK-BMP encoding for recombinant human bone morphogenetic protein 2 (rhBMP-2) was used. pBMPPET originates from pET24c(+) (Novagen, Merck KGaA, Darmstadt, Germany) by cloning rhBMP-2 under the control of the IPTG inducible T7 promotor. pBK-BMP originates from pBAD24 (provided by BIRDC, Vienna, Austria); rhBMP-2 was cloned under the control of the L-arabinose inducible PBAD and the ampicillin resistance cassette was changed to kanamycin. E. coli C41 strains have an intact L-arabinose metabolism and therefore L-arabinose is used by E. coli C41 (pBK-BMP) as inducer for recombinant protein production and as a second carbon source. Furthermore recombinant KM71H Pichia pastoris was used. This strain contains a gene-sequence encoding horseradish peroxidase under the control of AOX promoter. 2.2. Media A defined minimal medium according to De Lisa et al. (DeLisa et al., 1999) supplemented with kanamycin with glucose as main carbon source (batch medium glucose concentration: 20 g/l; fedbatch medium glucose concentration 400 g/l) was used for E. coli processes. IPTG was added as a pulse (1 mM final concentration). The addition of L-arabinose was executed as a second independent feed with a feed concentration of 100 g/l to allow the decoupling of L-arabinose and glucose feed rate. The detailed description of media used for cultivation of P. pastoris is available elsewhere (Dietzsch et al., 2011a). 2.3. Bioreactor setup Dynamic experiments were carried out on two Techfors-S bioreactors (Infors, Bottmingen, Switzerland) holding 10 L and 20 L working volume. Base- and substrate feeding was quantified gravimetrically: base- and feed solutions were placed on balances (Sartorius, Göttingen, Germany). A peristaltic analogue pump (Lamda, Baar, Switzerland) as well as a Techfors-S integrated analogue pump assembled with silicon tubing was used for the addition of main substrate (glucose) and inducer (arabinose) feeds. Base addition (NH4OH) was achieved via a Techfors-S peristaltic pump. The ports on the top plate of the reactor were used for a dissolved oxygen sensor (Hamilton, Reno, USA), pH probe (Hamilton, Reno, USA), pressure sensor (Keller, Winterthur, Switzerland), a septum and an overpressure valve (Infors, Bottmingen, Switzerland). CO2 and O2 in the off-gas stream was measured by a gas analyzer (Müller Systems AG, Egg, Switzerland) following non dispersive infra-red (CO2) and paramagnetic (O2) principle. Signals were recorded by the process management system Lucullus PIMS (Secure Cell, Switzerland). P. pastoris cultivation was carried out in a Labfors fermenter with 3 L working volume (Infors, Bottmingen, Switzerland) with equivalent instrumentation as described above.

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

2.4. Fermentation parameters At first, a pre-culture was grown in a 1000 ml baffled Erlenmeyer flask containing 70 ml of pre-culture to an OD of 1–2 and used for the inoculation of the bioreactor (20 g/l batch medium). At the time point when the CO2 off-gas signal started to decrease, the biomass was estimated using an OD correlation (x ¼0.3  OD) and used as a starting value for the soft-sensor assisted control strategy. No further off-line data was delivered as input to the softsensor. A detailed description of the applied feeding and control profiles is given in the results section. The detailed description of the cultivation of P. pastoris is described elsewhere (Dietzsch et al., 2011a). Dissolved oxygen levels (DO2) were kept over 40% (100% were set before inoculation at 35 1C, 0.3 bar gauge, pH 7.2). The pH was kept constant by adding 12.5% NH4OH, which also served as nitrogen source. 2.5. Offline analytical methods 2.5.1. Biomass dry weight concentration After centrifugation (RZB 5171, 10 min, 4 1C) of 2 ml of the cell suspension in pre-weighted glass tubes the pellets were washed twice using distilled water and dried at 105 1C for 72 h. The biomass dry weight concentration was determined in duplicate. 2.5.2. Metabolite concentrations Supernatant for metabolites and residual substrate concentration determination were sampled using a ceramic-type 0.2 mm filtration device (IBA, Heiligenstadt, Germany). Acetate and arabinose concentrations in the supernatant were determined via HPLC (Supelcogel C-610, Sigma Aldrich, St. Louis, USA), applying an isocratic gradient (0.5 ml/min) with 0.1% H3PO4 as eluent. 2.6. Calculation of rates, estimation of rates and reconciliation procedure The applied approach is based on the real-time calculation of metabolic reaction rates. On the basis of the calculated metabolic reaction rates an over determined equation system using the Degree of Reduction (DoR) as well as the carbon balance is established. This allows the estimation of unknown rates as well as the consistency check of the observed system. A detailed description of rate calculation, estimation of non-measured rates and reconciliation as adapted from (van der Heijden et al., 1994) is given elsewhere (Jobé et al., 2003, Wechselberger et al., 2012b).

193

Subsequently, the change in reactor volume can be calculated by taking into account the broth density: ! dM dV dt ¼ ð3Þ dt ρν 2.8. On-line computation Soft-sensors as described in Section 3.1 were implemented using the Sim-Fit tool of the Lucullus PIMS (Secure Cell, Schlieren, Switzerland). The respective tools were called in a time interval of 300 s.

3. Results and discussion 3.1. Implementation of the soft-sensor assisted control methodology Following a cumulative, rate based estimation approach in matrix formulation as described in sections 2.6–2.8, a soft-sensor based control strategy aiming at bioprocess monitoring and dynamic bioprocess control was implemented in the PIMS. The general concept relies on the interplay of different modules, as illustrated in Fig. 1. Constants such as biomass and substrate stoichiometry, feed concentration and densities were determined in advance and made available to the individual devices “Volume Calculation”, “Soft Sensor”, as well as the “Feed Rate Setpoint Calculator”. Since the soft-sensor strategy aims at applications in bioprocess control and the exact stoichiometry of the biomass is in the most cases not known, the tools were run using a uniform C-molar biomass stoichiometry (CH1.8N0.2O0.56, ash 3%) determined for an E. coli K12 strain. For the P. Pastoris KM71H following composition was used: CH1.8N0.16O0.56, ash 10%. Using a mass balancing approach outlined in section 2.7, the volume calculation device estimates the broth volume and delivers the obtained estimation to the Soft Sensor and the Feed Rate Setpoint Calculator. The soft sensor performs biomass estimation following a cumulative estimation approach outlined in sections 2.6–2.8 using real-time process data and constants. Output from Soft Sensor is the biomass concentration, which is delivered to the Feed Rate Setpoint Calculator. Furthermore, the pre-set constant

2.7. Volume calculation The volume can be calculated by means of a mass balance and subsequent division through the broth density. The mass balance was set up as follows: _ in −M _ out ¼ dM M dt

ð1Þ

The input term considers mass input through substrate inflow and base inflow as well as oxygen fixation through the microorganisms. The output term considers water stripping, sampling, outflow terms in the continuous process case as well as mass loss through carbon dioxide production. Hence, the change in broth mass can be calculated as follows: dM ¼ F S;in þ F b;in þ r O2  32 dt
 F a;out  exH2 O ;out  60  18 − rCO2  44 þ Vm

ð2Þ

Fig. 1. Constants (C-molar stoichiometries, feed concentrations, feed densities) as well as inputs from the process (Off-gas data, gas and liquid flows) are used for the estimation of the biomass concentration (Soft Sensor) as well as the estimation of the volume (Volume Calculation). The obtained real-time process data is delivered to Feed Rate Setpoint Calculators, that provides a feed rate setpoints for different feeds reflecting the current process state using a substrate balance approach assuming no substrate accumulates. Execution of the feed rate setpoint is done via a simple PI flow controller.

194

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

mass fraction of C-source in the feed (g C-source/g feed) as well as a freely selectable set-point for the specific substrate uptake rate qs needs to be made available for the Feed Rate Setpoint Calculator. This set-point can be subjected to any function (ramps, static setpoints etc.) as demanded from the experimenter. Following a simple substrate balancing approach, a feed rate set-point is calculated that is delivered to a PI controller.

3.2. Dynamic control of a decellerostat culture aiming at the investigation of specific acetate production as a function of the specific substrate uptake rate Acetate as overflow metabolite in E. coli has reported negative effects on recombinant protein production as well as on RNA, lipid and protein synthesis (Lee, 1996) and can therefore be considered critical for recombinant processes (Koh et al., 1992). In fed-batch processes, acetate formation can be avoided by limiting and controlling the specific substrate uptake rate through the fedbatch feeding profile. However, this demands the strain specific knowledge of the onset of acetate production, i.e., which is maximum specific substrate uptake rate qs without the formation of acetate. Metabolic limitation like the onset of acetate production can be determined in a fast and efficient way by dynamic ramp experimentation, where the culture is submitted to a slow change in the specific substrate uptake rate qs or the specific growth rate, as applied elsewhere (Dietzsch et al., 2011a, Wechselberger et al., 2010). However, feed-forward control approaches as typically applied demand knowledge on known yield coefficients and the

assumption of constant yield coefficients, which is the information that is not available in the early stages of process development. Here, we apply the described soft-sensing strategy for the dynamic control (decellerostat) of the specific substrate uptake rate qs in a non-induced E. coli C41 strain, aiming at investigation of the onset of acetate production. Following a batch phase (data not shown), the culture was submitted to a linear decline in qs set point ranging from −1 to 0 [g/g/h] within 6 h (Fig. 2B). The dynamic control of the specific substrate uptake rate qs was performed using the presented soft-sensor assisted control strategy. From a dynamic experimental point of view the problem of finding the acetate onset can be either tackled by running accellero- or decellerostat fed-batch experiments. However, decellerostat experiments exploit the technical reactor capabilities (OTR, cooling) in a more efficient way, since higher specific substrate uptake rates are run in the beginning of the experiment, where the biomass concentrations are still low. This allows the full investigation of the physiologically feasible qs range, starting from the maximum specific substrate uptake rate (as observed in the batch phase). The real-time biomass concentration (Fig. 2A) was estimated via the biomass estimation device as described in section 3.1. The feed rate set point (Fig. 2B) was calculated in real-time using the estimated biomass concentration, the estimated broth volume as well as the constant feed concentration (400 g/l) and feed density (1100 g/l). Offline samples were taken every 30 min for the determination of biomass dry cell weight as well as acetate (Fig. 2A) and used for the calculation of the specific acetate production rate qace as well as the specific substrate uptake rate

Fig. 2. Estimated biomass concentrations and measured biomass dry cell weight concentrations and acetate concentrations are given in Fig. 2A. The setpoint of the specific substrate uptake rate (dashed grey line) as well as the specific substrate uptake rate calculated on the basis of biomass dry cell weight data (grey line) as well as the feed rate is given in Fig. 2B. Fig. 2C shows the specific substrate uptake rates (qs) as well as the specific acetate formation rates (qace) as calculated from offline data in the respective experiment. Fig. 2D shows a plot for the determination of the strain specific function specific acetate production (qace) as a function of the specific substrate uptake rate (qs).

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

qs (Fig. 2C). An offset between the qs calculated from the offline biomass dry-cell weight determination and the qs set-point was detected, converging in within to small qs values (Fig. 2B). The specific substrate uptake rate qs has negative values, since it is defined as uptake rate. The calculation of qs/qace from offline data followed by linear regression allows the extraction of the strain specific function qace f(qs) (Fig. 2D). The qs threshold for acetate production can be read from the interception of this function with the x-axis, as indicated by the arrows (Fig. 2D). Furthermore, from the strain specific function depicted in Fig. 2D, the bioprocess developer can obtain the information how long it takes for acetate (accumulated in the batch phase) to be consumed within the fed-batch at a specific substrate uptake rate qs. In the presented configuration, no metabolite formation is considered, resulting in deviation in the experimental cases where the culture was meaningful shifted to oxido-reductive metabolism (Fig. 2A and B), resulting in an error of approximately 10–20% on the controlled specific rate. The introduction of a third balance (e.g. the N-balance) would result in a more precise estimation, since the formation of metabolites can be considered.

3.3. Static control of the specific substrate uptake rate in an induced P. pastoris AOX expression system Because of the dependency of the productivity of horseradish peroxidase produced by the recombinant P. pastoris on substrate uptake rate, the control of a static qs set point is of great interest (Dietzsch et al., 2011a). Here it is demonstrated, that the control of

195

a static qs set point can be achieved without knowledge on known yield coefficients and the assumption of constant yield coefficients. A batch phase and consequently a fed batch phase was conducted to achieve a high cell density culture ready for induction (data not shown). Consequently, the presented control strategy was used for the control of the specific methanol uptake rate in the induction phase (Fig. 3A and Fig. 3B). The feed setpoint profile and the corresponding actual flow rate as well as offline qs calculated by using offline biomass and the corresponding setpoint are depicted in Fig. 3A. Biomass estimation by the soft sensor and corresponding offline values are depicted in Fig. 3B. The desired control of the specific substrate uptake rate was achieved without knowledge of the yield coefficient with a maximum deviation of 10%.

3.4. Dynamic control of the specific substrate uptake rate in an induced E. coli C41 (pBMPPET) expression system The control of dynamically changing setpoints of the specific substrate uptake rate qs is considered of practical interest for bioprocess development, especially for the design of feeding strategies for recombinant processes (Dietzsch et al., 2011). However, dynamically changing yield coefficients are frequently observed in induced conditions (Sagmeister et al., 2012, Jenzsch et al., 2006), restricting the applicability of feed-forward control strategies as well as control strategies based on total carbon dioxide evolution or total oxygen consumption. Here, the applicability of the presented soft-sensor control strategy for the dynamic control of the specific substrate uptake

Fig. 3. The setpoint of the specific substrate uptake rate as well as the specific substrate uptake rate calculated on the basis of biomass dry cell weight data as well as the feed rate is depicted in Fig. 3A. The estimated biomass concentrations and measured biomass dry cell well concentrations are shown in Fig. 3B.

Fig. 4. The biomass estimation in an E. coli C41 strain carrying the plasmid pBMPPET is shown in Fig. 4A. The time point of induction with IPTG (1 mM) is highlighted with an arrow (A and B). Fig. 4B shows the dynamic control of the specific substrate uptake rate qs. The qs setpoint is given in dashed grey lines. qs calculated on the basis of offline biomass dry cell weight data is given in full grey lines. The feed rate of glucose is plotted as full black line.

196

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

Fig. 5. Independent control of arabinose and glucose fluxes is depicted in Fig. 5A. The flow rate of glucose and arabinose are given as dashed grey line (glucose) and full grey line (arabinose). Fig. 5B shows the independent control of specific glucose and arabinose fluxes. The setpoint of the specific glucose uptake rate (qs glucose) is given as dashed grey line; the actual glucose uptake rate based on offline biomass dry cell weight measurements is given as full grey line. The setpoint of the specific arabinose uptake rate (qs arabinose) is given as dashed black line, the actual arabinose uptake rate based on offline biomass dry cell weight measurements is given as full black line.

rate qs on an induced process with dynamically changing yield coefficients due to the effect of recombinant protein production is demonstrated. Aiming at the demonstration of dynamic control of specific substrate uptake rate (qs) static setpoints and dynamic ramps under induced and non-induced conditions, the described softsensor based strategy was applied on a recombinant E. coli C41 strain carrying the plasmid pBMPPET. Following a batch phase (data not shown), a fed-batch with a qs controlled at −0.4 g/g/h was started. The biomass estimation that serves as the basis for the qs control is plotted in Fig. 4A, the feed profile as a consequence of the qs control strategy is plotted in Fig. 4B. Following 3 h at qs of −0.4 g/g/h, the culture was induced via addition of IPTG to a final concentration of 1 mM, as indicated by an arrow in Fig. 4. The control was continued at a qs level of −0.4 g/g/h until t ¼13.5 h and then subjected to a linear qs increase from -0.4 g/g/h to −0.2 g/g/h for 1 h. Thereafter, the control was continued at −0.2 g/g/h. The biomass yield coefficient estimated by the soft sensor was found to decrease in course of the process (Fig. 4A, grey line). For the validation of the qs control, the specific substrate uptake rate was calculated on the basis of biomass offline data (Fig. 4B).

3.5. Simultaneous dynamic control of the specific substrate and inducer uptake rates in a E. coli C41 (pBK-BMP) expression system aiming at the investigation of mixed feed metabolic capabilities “Mixed feed” refers to the use of multiple carbon sources in recombinant bioprocesses aiming at increased process performance (Arnau et al., 2011; Hellwig et al., 2001; Jungo et al., 2007). Recently, a novel dynamic experimental strategy based on the decoupling of the individual feed streams was reported (Zalai et al., 2012). Design of mixed feed systems demands the investigation of simultaneous substrate utilization capabilities, which is not straight forward due to physiological constrains such as the onset of catabolite repression. Here, the applicability of the presented soft-sensor assisted control strategy in the field of mixed feed bioprocess design is illustrated. Aiming at the investigation of mixed feed capabilities of an E. coli C41 strain carrying the plasmid pBK-BMP, the culture was subjected to two independent C sources (glucose as main C-source, arabinose as second C-source and inducer for recombinant protein production) following the qs control strategy described in Section 3.1. The goal of the dynamic experiment was to investigate the capability of L-arabinose metabolization in a range of 0 to 0.3 g/g/h specific uptake of L-arabinose in the presence of an uptake of D-glucose of -0.2 g/g/h.

Following a batch phase, the process was subjected to an arabinose pulse to allow the culture to adapt to arabinose metabolism (data not shown). Thereafter, the qs glucose setpoint was kept constant at a level of −0.2 g/g/h and the qs arabinose was subjected to a linear ramp starting from -0.05 g/g/h to −0.3 g/g/h (Fig. 5B). No L-arabinose was found to accumulate, demonstrating that the culture is able to metabolize L-arabinose in a flux range from −0.05 g/g/h to −0.3 g/g/h in the presence of specific glucose uptake of −0.2 g/g/h. Obtained specific substrate uptake rates were calculated on the basis of biomass dry cell weight measurements (Fig. 5A). 3.6. Remarks on error propagation in the described soft-sensor control strategy The presented soft-sensor assisted control strategy for the control of the specific substrate uptake rate qs strongly depends upon a correct estimation of the biomass concentration by the soft-sensor. The estimation of the biomass concentration is influenced by (a) the initial biomass concentration delivered to the soft sensor (typically estimated via OD correlation or a fixed yield relationship after the batch-phase) and (b) the error on the estimation of the biomass formation rate. The signal to noise ratio of real-time estimated rates and its enhancements through reconciliation procedures is discussed elsewhere (Wechselberger et al., 2013).

4. Conclusions and outlook 4.1. Application of the strategy within bioprocess development Core elements of process analytical technology (PAT) is the capability to “…design, analyze and control… (FDA, 2004)” the manufacturing process. The presented soft-sensor assisted control strategy especially qualifies as a full PAT tool for process development, since it is independent from prior strain specific knowledge (e.g. known yield coefficients) and can be applied to the full definition of PAT; hence to design, analyze and control microbial fermentation processes. 4.1.1. Applications for process development Feed strategy development for recombinant fed-batch processes is a highly challenging task due to numerous possible feeding trajectories (static feed, linear increasing feed, exponential feed etc.) at various levels (Ramalingam et al., 2007; Wong et al., 1998). Recently Wechselberger et al. showed that the impact of typical

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198 Table 1 Applications of the presented soft-sensor assisted control strategy. Applications in process development Soft sensor controlled dynamic ramp experiments in fed-batch mode. Soft sensor controlled dynamic shift experiments in fed-batch mode. Control of defined trajectories of the specific substrate- and inducer uptake rates for feed strategy development. Applications in manufacturing Static control of specific rates independent of changing yield coefficients in induced and non-induced bioprocesses. Control of defined trajectories of specific substrate and inducer uptake rates independent of changing yield coefficients.

feeding profile related process parameters on recombinant protein production can be fully explained by the specific substrate uptake rate qs (Wechselberger et al., 2012a). Wechselberger et al. suggested to include the specific substrate uptake rate as a factor in Design of Experiments (DoEs) instead of feed profile related process parameters. The demonstrated static controllability of the specific substrate uptake rate qs is a prerequisite for their inclusion as factors in statistical experimental designs (DoEs), which can lead to increased experimental efficiency due to a more adequate choice of experimental factors (Wechselberger et al., 2012a). Furthermore, the soft-sensor assisted control strategy presented in this contribution allows the investigation of the impact of controlled dynamic trajectories of the specific substrate uptake rate qs on recombinant protein production. Applications are summarized in Table 1. 4.1.2. Application for analyzing and controlling manufacturing processes Following the described strategy, yield coefficients and specific rates are readily available on-line (Section 3.4, and already demonstrated in detail by Jobé et al., 2003). This can serve as the basis of process monitoring during manufacturing and process development. The extracted information is directly linked to the cellular metabolism and can be exploited for multivariate data analysis for the investigation of the physiological impact of process parameters on the biological system (Sagmeister et al., 2012) as well as physiological monitoring during manufacturing. The realtime availability of physiological information can possibly contribute to real-time release decisions (EMA, 2010). In case large data sets are available, soft sensor control approaches on the basis of data driven models such as ANNs, PCR or PLS as reviewed by Kadlec et al. (Kadlec et al., 2009) can probably lead to a more precise state estimation and a more precise process control. Their application and significance for biotechnological processes was included in a recent review by Luttmann et al. (Luttmann et al., 2012). However, elaborate data sets are not always available and therefore control strategies on the basis of first-principles (as presented here) can be of great value for manufacturing approaches. Applications are summarized in Table 1. Summarizing, the presented physiological control strategy for bioprocess development proved applicable for a variety of (dynamic-) microbial bioprocess development tasks and therefore holds potential to become a key tool in bioprocess development. Furthermore, it can possibly find applications within manufacturing where elaborate data sets do not exist.

Acknowledgements This project was supported by FFG, Land Steiermark and SFG. Strains and plasmids were gratefully provided by BIRD-C GmbH & Co KG, Kritzendorf and Morphoplant GmbH, Bochum.

197

References Aehle, M., Kuprijanov, A., Schaepe, S., Simutis, R., Lübbert, A., 2011. Increasing batch-to-batch reproducibility of CHO cultures by robust open-loop control. Cytotechnology 63, 41–47. Arnau, C., Casas, C., Valero, F., 2011. The effect of glycerol mixed substrate on the heterologous production of a Rhizopus oryzae lipase in Pichia pastoris system. Biochem. Eng. J. 57, 30–37. De Assis, A.J., Filho, R.M., 2000. Soft sensors development for on-line bioreactor state estimation. Comput. Chem. Eng. 24, 1099–1103. Betts, J.I., Baganz, F., 2006. Miniature bioreactors: current practices and future opportunities. Microb. Cell Fact. 5, 21. Caramihai, M., Chirvase, A., Fonteix, C., Marc, I., Fournier, F., Misleanu, R., Ungureanu, C., 2007. Optimal fed-batch bioprocess control. An advanced approach. Comput. Aided Chem. Eng. 24, 787–792. DeLisa, M.P., Li, J., Rao, G., Weigand, W.A., Bentley, W.E., 1999. Monitoring GFPoperon fusion protein expression during high cell density cultivation of Escherichia coli using an on-line optical sensor. Biotechnol. Bioeng. 65, 54–64. Dietzsch, C., Spadiut, O., Herwig, C., 2011a. A fast approach to determine a fed batch feeding profile for recombinant Pichia pastoris strains. Microb. Cell Fact. 10, 85. Dietzsch, C., Spadiut, O., Herwig, C., 2011b. A dynamic method based on the specific substrate uptake rate to set up a feeding strategy for Pichia pastoris. Microb. Cell Fact. 10, 14. Dubach, A.C., Märkl, H., 1992. Application of an extended kalman filter method for monitoring high density cultivation of Escherichia coli. J. Fermentation and Bioengineering 73, 396–402. Duboc, P., von Stockar, U., Villadsen, J., 1998. Simple generic model for dynamic experiments withSaccharomyces cerevisiae in continuous culture: decoupling between anabolism and catabolism. Biotechnol. Bioeng. 60, 180–189. EMA, 2010. Guideline on real time release testing (formerly guideline on parametric release). Retrieved October 2012 from 〈http://www.emea.europa.eu/docs/en_GB/ document_library/Scientific_guideline/2010/03/WC500075028.pdf〉. FDA 2004. PAT—A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance. Retrieved October 2012 from http://www. fda.gov/downloads/Drugs/…/Guidances/ucm070305.pdf. Feitkenhauer, H., Schnicke, S., Müller, R., Märkl, H., 2001. Determination of the kinetic parameters of the phenol-degrading thermophile Bacillus themoleovorans sp. A2. Appl. Microbiol. Biotechnol. 57, 744–750. Gadkar, K.G., Mehra, S., Gomes, J., 2005. On-line adaptation of neural networks for bioprocess control. Comput. Chem. Eng. 29, 1047–1057. Ghoul, M., Dardenne, M., Fonteix, C., Marc, A., 1991. Extended Kalman filtering technique for the on-line control of OKT3 hybridoma cultures. Biotechnol. Tech. 5, 367–370. Glassey, J., Gernaey, K.V., Clemens, C., Schulz, T.W., Oliveira, R., Striedner, G., Mandenius, C.-F., 2011. Process analytical technology (PAT) for biopharmaceuticals. Biotechnol. J. 6, 369–377. van der Heijden, R.T.J.M., Heijnen, J.J., Hellinga, C., Romein, B., K.C.A.M., Luyben, 1994. Linear constraint relations in biochemical reaction systems: I. Classification of the calculability and the balanceability of conversion rates. Biotechnol. Bioeng. 43, 3–10. Hellwig, S., Emde, F., Raven, N.P., Henke, M., van Der Logt, P., Fischer, R., 2001. Analysis of single-chain antibody production in Pichia pastoris using on-line methanol control in fed-batch and mixed-feed fermentations. Biotechnol. Bioeng. 74, 344–352. Herwig, C., Marison, I., von Stockar, U., 2001. On-line stoichiometry and identification of metabolic state under dynamic process conditions. Biotechnol. Bioeng. 75, 345–354. Herwig, C., von Stockar, U., 2002. A small metabolic flux model to identify transient metabolic regulations in Saccharomyces cerevisiae. Bioprocess and Biosyst. Eng. 24, 395–403. Herwig, C., Von Stockar, U., 2003. Quantitative comparison of transient growth of Saccharomyces cerevisiae, Saccharomyces kluyveri, and Kluyveromyces lactis. Biotechnol. Bioeng. 81, 837–847. Jenzsch, M., Simutis, R., Luebbert, A., 2006. Generic model control of the specific growth rate in recombinant Escherichia coli cultivations. J. Biotechnol. 122, 483–493. Jobé, A.M., Herwig, C., Surzyn, M., Walker, B., Marison, I., von Stockar, U., 2003. Generally applicable fed‐batch culture concept based on the detection of metabolic state by on‐line balancing. Biotechnol. Bioeng. 82, 627–639. Jungo, C., Marison, I., von Stockar, U., 2007. Mixed feeds of glycerol and methanol can improve the performance of Pichia pastoris cultures: a quantitative study based on concentration gradients in transient continuous cultures. J. Biotechnol. 128, 824–837. Kadlec, P., Gabrys, B., Strandt, S., 2009. Data-driven soft sensors in the process industry. Comput. Chem. Eng. 33, 795–814. Karakuzu, C., Türker, M., Öztürk, S., 2006. Modelling, on-line state estimation and fuzzy control of production scale fed-batch baker's yeast fermentation. Control Eng. Pract. 14, 959–974. Kiviharju, K., Salonen, K., Moilanen, U., Eerikäinen, T., 2008. Biomass measurement online: the performance of in situ measurements and software sensors. J. Indust. Microbiol. Biotechnol. 35, 657–665. Koh, B.T., Nakashimada, U., Pfeiffer, M., Yap, M.G.S., 1992. Comparison of acetate inhibition on growth of host and recombinant E. coli K12 strains. Biotechnol. Lett. 14, 1115–1118. Komives, C., Parker, R.S., 2003. Bioreactor state estimation and control. Current Opin. Biotechnol. 14, 468–474.

198

P. Sagmeister et al. / Chemical Engineering Science 96 (2013) 190–198

Kumar, S., Wittmann, C., Heinzle, E., 2004. Review: minibioreactors. Biotechnol. Lett. 26, 1–10. Kusterer, A., Krause, C., Kaufmann, K., Arnold, M., Weuster-Botz, D., 2008. Fully automated single-use stirred-tank bioreactors for parallel microbial cultivations. Bioprocess. Biosyst. Eng. 31, 207–215. Lee, S., 1996. High cell-density culture of Escherichia coli. Trends Biotechnol. 14, 98–105. Liu, G., Yu, S., Mei, C., Ding, Y., 2011. A novel soft sensor model based on artificial neural network in the fermentation process. Afr. J. Biotechnol. 10, 19780–19787. Luttmann, R., Bracewell, D.G., Cornelissen, G., Gernaey, K.V., Glassey, J., Hass, V.C., Kaiser, C., Preusse, C., Striedner, G., Mandenius, C.-F., 2012. Soft sensors in bioprocessing: a status report and recommendations. J. Biotechnol. 7, 1040–1048. Mandenius, C.-F., Graumann, K., Schultz, T.W., Premstaller, A., Olsson, I.-M., Petiot, E., Clemens, C., Welin, M., 2009. Quality-by-design for biotechnology-related pharmaceuticals. Biotechnol. J. 4, 600–609. Mailleret, L., Bernard, O., Steyer, J.-P., 2004. Nonlinear adaptive control for bioreactors with unknown kinetics. Automatica 40, 1379–1385. Nayak, R., Gomes, J., 2009. Sequential adaptive networks: an ensemble of neural networks for feed forward control of L-methionine production. Chem. Eng. Sci. 64, 2401–2412. Osman, J.J., Birch, J., Varley, J., 2002. The response of GS-NS0 myeloma cells to single and multiple pH perturbations. Biotechnol. Bioeng. 79, 398–407. Puskeiler, R., Kaufmann, K., Weuster-Botz, D., 2005a. Development, parallelization, and automation of a gas-inducing milliliter-scale bioreactor for highthroughput bioprocess design (HTBD). Biotechnol. Bioeng. 89, 512–523. Puskeiler, R., Kusterer, A., John, G.T., Weuster-Botz, D., 2005b. Miniature bioreactors for automated high-throughput bioprocess design (HTBD): reproducibility of parallel fed-batch cultivations with Escherichia coli. Biotechnol. Appl. Biochem. 42, 227–235. Ramalingam, S., Gautam, P., Mukherjee, K.J., Jayaraman, G., 2007. Effects of postinduction feed strategies on secretory production of recombinant streptokinase in Escherichia coli. Biochem. Eng. J. 33, 34–41. Rathore, A.S., Bhambure, R., Ghare, V., 2010. Process analytical technology (PAT) for biopharmaceutical products. Anal. Bioanal. Chem. 398, 137–154. Rathore, A.S., Winkle, H., 2009. Quality by design for biopharmaceuticals. Nat. Biotechnol. 27, 26–34.

Sagmeister, P., Wechselberger, P., Herwig, C., 2012. Information processing: rate based investigation of physiological changes along design space development. PDA J. Pharm. Sci. Technol. 66, 526–541. San, K., Stephanopoulos, G., 1984. Studies on on‐line bioreactor identification. II. Numerical and experimental results. Biotechnol. Bioeng. 26, 1189–1197. Soini, J., Falschlehner, C., Mayer, C., Böhm, D., Weinel, S., Panula, J., Vasala, A., Neubauer, P., 2005. Transient increase of ATP as a response to temperature upshift in Escherichia coli. Microb. Cell Fact. 4, 9. Sonnleitner, B., Locher, G., Fiechter, A., 1992. Biomass determination. J. Biotechnol. 25, 5–22. Undey, C., Low, D., Menezes, J.M.C.D., 2011. Pat applied in biopharmaceutical process development and manufacturing: an enabling tool for quality-bydesign 1st ed. Crc. Pr. Inc., 327. Wechselberger, P., Sagmeister, P., Engelking, H., Schmidt, T., Wenger, J., Herwig, C., 2012a. Efficient feeding profile optimization for recombinant protein production using physiological information. Bioprocess Biosyst. Eng. 35, 1637–1649. Wechselberger, P., Herwig, C., 2013. Model-based analysis on the extractability of information from data in dynamic fed-batch experiments. Biotechnol. Prog. 29, 285–296. Wechselberger, P., Sagmeister, P., Herwig, C., 2012b. Real-time estimation of biomass and specific growth rate in physiologically variable recombinant fedbatch processes. Bioprocess Biosyst. Eng., Nov 23, (Epub ahead of print). Wechselberger, P., Seifert, A., Herwig, C., 2010. PAT method to gather bioprocess parameters in real-time using simple input variables and first principle relationships. Chem. Eng. Sci. 65, 5734–5746. Weuster-Botz, D., Puskeiler, R., Kusterer, A., Kaufmann, K., John, G.T., Arnold, M., 2005. Methods and milliliter scale devices for high-throughput bioprocess design. Bioprocess Biosyst. Eng. 28, 109–119. Wilson, D.I., Agarwal, M., Rippin, D.W.T., 1998. Experiences implementing the extended Kalman filter on an industrial batch reactor. Comput. Chem. Eng. 22, 1653–1672. Wong, H.H., Kim, Y.C., Lee, S.Y., Chang, H.N., 1998. Effect of post-induction nutrient feeding strategies on the production of bioadhesive protein in Escherichia coli. Biotechnol. Bioeng. 60, 271–276. Zalai, D., Dietzsch, C., Herwig, C., Spadiut, O., 2012. A dynamic fed batch strategy for a Pichia pastoris mixed feed system to increase process understanding. Biotechnol. Prog. 28, 878–886.