• Open Access

Designing a fully automated multi-bioreactor plant for fast DoE optimization of pharmaceutical protein production



The identification of optimal expression conditions for state-of-the-art production of pharmaceutical proteins is a very time-consuming and expensive process. In this report a method for rapid and reproducible optimization of protein expression in an in-house designed small-scale BIOSTAT® multi-bioreactor plant is described. A newly developed BioPAT® MFCS/win Design of Experiments (DoE) module (Sartorius Stedim Systems, Germany) connects the process control system MFCS/win and the DoE software MODDE® (Umetrics AB, Sweden) and enables therefore the implementation of fully automated optimization procedures. As a proof of concept, a commercial Pichia pastoris strain KM71H has been transformed for the expression of potential malaria vaccines. This approach has allowed a doubling of intact protein secretion productivity due to the DoE optimization procedure compared to initial cultivation results. In a next step, robustness regarding the sensitivity to process parameter variability has been proven around the determined optimum. Thereby, a pharmaceutical production process that is significantly improved within seven 24-hour cultivation cycles was established. Specifically, regarding the regulatory demands pointed out in the process analytical technology (PAT) initiative of the United States Food and Drug Administration (FDA), the combination of a highly instrumented, fully automated multi-bioreactor platform with proper cultivation strategies and extended DoE software solutions opens up promising benefits and opportunities for pharmaceutical protein production.


AOX, alcohol oxidase; DoE, Design of Experiment; FDA, United States Food and Drug Administration; MLR, multiple linear regression; PAT, process analytical technology; QbD, Quality by Design; RP, reversed phase; SST, sum of squares total

1  Introduction

The application of Quality by Design (QbD) has been receiving more and more attention in the pharmaceutical community. QbD requires a thorough understanding of its manufacturing process, requiring an upfront investment in time and resources for the development of a product [1].

A basic part of QbD is to create a process design space (International Conference of Harmonisation (2009) ICH Harmonised Tripartite Guideline: Q8(R2) Pharmaceutical Development). The design space is defined by the key and the critical process parameters identified from process characterization studies. These parameters are the primary focus for in-line, on-line or at-line process analytical technology (PAT) applications [1].

Based on the ICH guidance documents, the United States Food and Drug Administration (FDA) emphasizes the need for improved on-line monitoring and control methods to maintain high product quality during manufacturing operations [2]. According to Gnoth et al. [3] the essence of the PAT initiative is primarily to monitor process variables in order to identify critical quality attributes (CQAs). Rathore et al. [4] discussed the subject more precisely and recommended the use of PAT in the aspects of process understanding, improved yield, reduction in the production cycle time, decrease in the energy consumption, cost reduction, and real-time release of the batches. Another aspect for improvement in product quality, safety, and/or efficiency in the guidance of the FDA is to improve operator safety and to reduce human errors by implementing process automation structures [5].

However, current pilot scale development is still inefficient, wasteful and time-consuming. The setup of many pilot plants does not support the integration of extended PAT and automation structures. The fact of missing PAT prevents the application of statistical tools, which means “one variable at a time” must iteratively be adopted. This leads to a quasi-optimum with lack of information and furthermore costly delays.

In this context the use of experimental design approaches becomes increasingly important. The FDA [5] points out the effectiveness of methodological experiments based on statistical principles of orthogonality, reference distribution, and randomization for identifying and studying the effect and interaction of product and process variables. Methods such as factorial design, response surface methodology, and DoE provide efficient ways to optimize cultivations using a reduced number of experiments [6].

The rising demands regarding process understanding by using analytical, statistical tools causes in widespread small scale DoE applications with Pichia pastoris [7–11].

In this view a multi-bioreactor plant was designed, which meets all the mentioned requirements (i) extended PAT for process observation/understanding, (ii) increasing automation, and (iii) effective determination of optimal expression conditions with experimental design approaches.

In previous investigations the methylotrophic yeast Pichia pastoris has been determined as suitable expression system for a high-level production with respect to the improvement of cultivation respectively purification conditions for potential malaria vaccines [12, 13]. The production process was done with the P. pastoris strain KM71H, phenotype methanol utilization pathway slow (without AOX1-gen) (MutS), which was transformed with artificial protein sequences of the Plasmodium falciparum protein apical membrane antigen 1 (PfAMA1) obtained from the Biomedical Primate Research Centre (BPRC) of The Netherlands [14].

2  Material and methods

2.1  Experimental setup

The experimental setup shown schematically in Fig. 1 was extended by a newly developed BioPAT® MFCS/win DoE module (Sartorius Stedim Systems, Germany), which is linked to the DoE software MODDE® (Umetrics, Sweden) for an automatic setup of designed experiments as well as a fast and reliable data analysis. These tools support the concept of a fully automated sequential/parallel DoE cultivation strategy, which was performed in a cell-breeding BIOSTAT® Bplus and a sixfold screening BIOSTAT® Qplus reactor system [12]. The plant was set up with an industrial conformed sterile design including sterilizable transfer valves (GEMÜ GmbH & Co. KG, Germany) and quick connectors (Stäubli Tec-Systems GmbH, Germany).

Figure 1.

Set up of multi-bioreactor plant for a fully automated sequential/parallel cultivation strategy. Schematically shown are cell-breeding (B+) and screening (Q+) reactors as well as piping with valves and sterilizable quick connectors.

Using a 5-L cell-breeding bioreactor, the same inoculation conditions for six 1-L screening bioreactors were provided in a cyclic manner. The need of preparing inoculation material cyclically within a 24-h timeframe has been fulfilled by a fully automated process. This was achieved despite the requirement of switching the carbon source from glycerol for cell-breeding to methanol for induction.

Different approaches for PAT applications were implemented in terms of product quality. For simultaneous quantification of secreted recombinant proteins, the plant was equipped with an at-line reversed phase (RP)-HPLC [12] and an at-line sequential injection analysis (SIA) with Immobilized Metal Affinity Chromatography (IMAC)-Bead Injection [15]. Quality assessment and reproducibility of breeding cycles was performed via off-line measurement of cell-internal alcohol oxidase (AOX) content.

The instrumentation allows direct scale-up of the developed cultivation methods into pilot scale, as well as small scale commercial production processes.

2.2  Strain, media, and cultivation conditions

A recombinant Pichia pastoris MutS strain was kindly provided by Bart Faber from the BPRC of The Netherlands.

FM22 medium was used as batch and refresh medium. Fed-batch feeding solution contained of 630 g/L glycerol, 2 mg/L biotin, and 10 mL/L Pichia trace minerals 4 (PTM4) stock solutions. An open loop feed control for the glycerol feeding rate FR1 was realized to implement substrate-limited fed-batch operations. The methanol concentration was closed loop controlled to a given set point with pure methanol. The pH-value was controlled with 12.5% ammonium hydroxide and 0.5 M phosphoric acid during bioreactor cultivation.

The culture was air-aerated with 1.5 vvm. The dissolved oxygen tension pO2 was closed loop controlled at a set point of 25% by manipulation of agitation speed.

Investigations conducted in the current study were performed according to our previous study [12].

2.3  Measurement of potential malaria vaccine

The target protein concentration was determined by an at-line RP-HPLC method as described previously [12].

2.4  Determination of cell-specific alcohol oxidase content

One milliliter of cell suspension was centrifuged for mechanical lysis. The pellet was resuspended to a cell concentration of 100 g/L with PBS-buffer. For cell disruption 1 g of glass beads (0.4–0.6 mm, Sartorius AG, Germany) were added to 100 μL of cell suspension and 900 μL of PBS buffer in a 2-mL microtube. Cells disruption took place by using VXR basic Vibrax® (IKA-Werke GmbH & Co. KG, Germany) for 20 min at 2000 rpm. Cellular debris was removed by centrifugation (30 min, 14 000 rpm, 4°C). Stock solution for AOX assay includes 10.5 μL peroxidase (POD) (250 U/mL), 33 mg 2, 2'-azino-bis-(3-ethyl-benzthiazoline-6-sulfonic acid) (ABTS) in 20 mL 100 mM potassium phosphate buffer (pH 7.5) and 1050 μL diluted hydrogen peroxide (1 μL of 30% w/w H2O2 in 100 mL sterile filtrated DI-H2O). The reaction was started for each sample by the addition of 50 μL 1.0% v/v methanol solution to 200 μL stock solution and 50 μL sample in each well of the 96 well plate. The reaction was monitored by the microplate reader Sunrise(tm) (Tecan Group Ltd., Switzerland) at 405 nm over 4.5 min. One unit of AOX was defined as the amount of enzyme that induces the oxidation of 1 μM of ABTS per minute under the above experimental conditions. Measurements were executed in triplicates. This assay is based on refs. [16, 17].

2.5  Bioreactor instrumentation

All cultivations were performed with a 5-L BIOSTAT® Bplus bioreactor and a sixfold 1 L bioreactor system BIOSTAT® Qplus (both Sartorius Stedim Biotech, Germany). The reactors are highly instrumented (turbidity, MeOH concentration, O2 and CO2 offgas component analysis, 36 scales for accurate process balancing, and automated processing). The detailed instrumentation is described in Fricke et al. [12].

2.6  Sequential/parallel cultivation strategy for process optimization

For the optimization procedures conducted in the multi-bioreactor plant a sequential/parallel cultivation strategy was developed and implemented. The cell-breeding procedure followed a typical three step Pichia pastoris cultivation strategy with a batch, fed-batch and induction phase [18]. The process started with a cell density cXL0 of 5 g/L. In the batch phase on glycerol cS1M, the cells grew with the maximum specific growth rate μ1max of 0.23 h–1.

After an on-line batch end detection, a substrate-limited fed-batch phase on glycerol was started with an exponential increase of glycerol feed FR1 for controlling the cell-specific growth rate. The glycerol feed automatically stopped when reaching a previously defined cell-density, estimated on-line from the in-line turbidity measurement. In the induction phase the methanol concentration cS2M was controlled to a set point of 0.5 g/L for about 12 h. The resulting metabolic change of the cells allows to avoid expression delays in the subsequent production phases in the sixfold Qplus system.

In the following a defined volume of cell suspension was transferred consecutively into the six screening reactors (Q+). After a 1:1 dilution with refresh medium an initial cell density of approximately 18.5 g/L was adjusted.

The time courses of two optimization runs are shown in Fig. 2. Depicted are cell density cXL, estimated from turbidity signal, methanol concentration cS2M, pH-value, dissolved oxygen tension pO2, and target protein absorption AP1M. The experiments were only varied in settings for pH. AP1M was determined at-line with the HPLC-method, mentioned in Section 2.3.

With respect to reproducibility, nearly equal start values for cXL0 could be obtained, as demonstrated in Fig. 2. By means of global monitoring, regarding time courses of cXL, respectively, AP1M, an extended process analysis could be carried out.

Figure 2.

Time courses of two potential malaria vaccine production runs with Pichia pastoris. The culture conditions were only varied in pH-values. AP1M: at-line signal of UV absorption (▿), cS2M: methanol concentration, cXL: cell density reconstructed from turbidity signal, pO2: dissolved oxygen tension, PRDk: target protein productivity of run k.

2.7  Quality criterion for the evaluation of recombinant protein expression in Pichia pastoris

For the evaluation of the screening experiments, the mean value of target protein secretion productivity PRDk in screening step k was used as performance index yk,

equation image(1)

and was based on off-line measurements of the amount of active target protein IAP1M. Data taken at the end tkn and at the beginning tk0 of a screening run, and the amount of produced protein taken out by each sample time tkj in between, was considered. The productivity was determined for the statistically designed experiments with nearly identical initial conditions of cell density cXLk0 and target protein concentration IAP1Mk0 (Table 1).

Table 1. Data of performed cultivation runs
Run kcS2Mk (g/L)ϑLk (°C)pHk (–)cXLk0 (g/L)IAP1Mk0 (AUs)PRDk (mAUs/h)
  • a)

    *** No result.

Robustness testing
Run kpHk (–)ϑLk (°C) cXLk0 (g/L)IAP1Mk0 (AUs)PRDk (mAUs/h)
15.5525.5 17.900.35117.54
25.4026.5 19.200.33912.25
35.5525.5 19.200.34715.56
45.7024.5 19.800.32213.99
55.4024.5 18.350.29711.15
65.7026.5 18.000.30616.93
75.5525.5 19.550.27812.74
85.5525.5 19.400.27114.54

2.8  Automated DoE applications with the sequential/parallel cultivation approach

The principle of DoE is to vary simultaneously a number of “factors”, which potentially influence the outcome, respectively, response of the process. The experiments are distributed in a rectangular design space and performed at low and high levels of each “factor”, augmented with at least three additional center points. Therewith DoE provides an organized approach which requires a limited number of experiments leading to significantly faster process optimization [19].

In this DoE study, cultivations were varied in settings for pH-value, cultivation temperature ϑL and inductor concentration cS2M in the optimization procedure, respectively, pH-value, and ϑL for the investigations relating robustness of the biological production system. The pO2 was kept constant at 25% in all cultivations.

A central composite circumscribed (CCC) approach was chosen for the design of the optimization experimental setup. The use of the CCC-design allows the identification of interactions between the factors and a cubic response behavior. Table 1 shows the experimental design of the conducted optimization. In cultivations with a pH-value above six precipitations of media components were observable, therefore the upper limit for varying the pH-value was given in this case.

The goal of the optimization procedure was to approximate the determined response y,

equation image(2)

by a cubic quadratic polynomial model using multiple linear regression (MLR).

A two-level full factorial design was used for the experimental design of the robustness testing study, shown in Table 1.

The DoE procedures were conducted by combining the BioPAT® MFCS/win DoE module with the professional DoE software MODDE® [19].

Since the DoE module is linked to the process control system MFCS/win, the respective set-points could automatically be assigned to the parameter phase MFCSDOE_Factors, as shown in Fig. 3.

Figure 3.

S88 program structure of the DoE recipe. Shown are the recipe operations, a section of the SFC with different phases and the content of the MFCSDOE_Factors phase.

Shown are the ANSI/ISA-88.01 recipe structure with the operations STANDBY, INOCULATION, PRODUCTION, and TERMINATION, as well as a section of the sequential function chart used for automatic initialization of methanol and pO2 control as well as automatic adjustment of DoE parameters.

The sterilization, inoculation, cultivation, harvest, and refresh operations ran automatically by use of S88 recipes derived from the SCADA-system BioPAT® MFCS/win (Sartorius Stedim Systems, Germany). This was realized by piloting pressure valves and controlling pumps, remote-controlled via the local digital control unit DCU4.

3  Optimization of recombinant protein expression

3.1  Verification of the sequential/parallel cultivation approach

In the methylotrophic yeast Pichia pastoris, AOX is a key enzyme involved in the dissimilation of methanol [20]. AOX catalyzes the first step in the methanol utilization pathway, the oxidation of methanol to formaldehyde and hydrogen peroxide [21].

Therefore the cell-specific internal AOX content gP2/X was chosen for investigations relating changes in Pichia pastoris metabolism during long-term cyclic cell-breeding cultivations.

A start up, inoculated from shake flasks, and three cell-breeding cycles, following the strategy described before, are shown in Fig. 4, by means of cell density cXL, methanol concentration cS2M, specific cell internal AOX content gP2/X, and specific methanol uptake rate qS2/X. Each cycle consists of a glycerol (S1) batch, a substrate-limited fed batch on glycerol and a pre-induction phase on methanol (S2).

Figure 4.

Cyclically recurring cell-breeding cultivations with Pichia pastoris. Every cycle consists of a glycerol (S1) batch (u: S1 unlimited), a fed-batch phase (l: S1 limited), and a pre-induction phase on methanol (S2). cXL: cell density reconstructed from cell dry weight (○) (determined in duplicate); cS2M: methanol concentration in media phase; gP2/X: cell internal cell-specific AOX content (equation image) (determined in triplicate); qS2/X: specific methanol uptake rate (▿).

As shown in Fig. 4, gP2/X was not observed during glycerol phase in the start up cultivation. The AOX promoter is repressed by unlimited growth on glycerol, which is well known from literature [22–24]. After methanol supply, gP2/X and qS2/X began to increase rapidly. During harvest/refresh operations and after initialization of a further batch process on glycerol (S1), associated with a fully depletion of methanol (S2), the gP2/X decreased rapidly. This has been reported from Jungo et al. [20] as well.

The recurring enrichment to the same level and the subsequent decrease of gP2/X in the glycerol batch phase of gP2/X show reproducible metabolic turnovers of the Pichia pastoris cells during cyclic substrate change. Combined with final cell densities cXL, equally in each cycle, the functionality of the parallel/sequential cell-breeding approach is proven.

The developed multi-bioreactor plant met the requirement of reproducibility in cell-breeding and ensures stable initial conditions in the screening cultivations, which is achieved by implementing complex automation structures. By using linked DoE tools and extended PAT, investigations regarding process optimization in pharmaceutical protein production become more manageable and reliable.

3.2  Optimization of malaria vaccine expression using the multi-bioreactor plant

The measured product concentrations were used to calculate the related performance index in Eq. (1). Data, shown in Table 1, was used to generate the coefficients a0 to a123 by fitting the model (Eq. (2)) to the measurements with MLR.

The goodness of fit R2 of 0.98 shows that only 2% of the total variation is not explainable by the model. The Prediction Error Sum of Squares (PRESS) value of 0.59 (AUs/h)2 related to the Sum of Squares Total (SST) value of 3.96 (AUs/h)2 indicates a low predicted variation. The goodness of prediction Q2, calculated from PRESS and SST, of 0.85 is close to the determination coefficient R2.

The reproducibility RP,

equation image(3)

could be calculated to 0.92 with the Sum of Squares Center Points (SSCP), the SST, the number of experiments ntot and the number of independent Center Point experiments nCP.

High significance of the model can be claimed after consideration of ANalysis Of Variance (ANOVA) for a significance level of 95%. The p-value for regression (<0.001) supports the analysis. The model shows no lack of fit with a plof-value of 0.76.

Figure 5 illustrates the results of the optimization via a response surface plot by fitting the model to the experimental data.

Figure 5.

Response surface plot of the DoE optimization results as a function of cultivation temperature ϑL and pH-value at a methanol concentration cS2M of 1.0 g/L. Model data was fitted to the data determined experimental by using MLR. A p-value for regression of < 0.001 claims high model significance. The model shows no lack of fit.

The conclusive result indicates optimal productivity PRD for a methanol concentration cS2Mopt of 1.0 g/L, a pHopt of 5.55 and a cultivation temperature ϑLopt of 25.8°C.

As known from literature, optimum protein production varies in accordance to the used Pichia pastoris strain, especially relating to the geno-, respectively, phenotype of organism (e.g. [25–27]) and to the secreted foreign protein, which are directly influenced by cultivation conditions (e.g. [28–30]).

Gasser et al. [27] reported that cultivation temperature strongly impacts the regulation of specific genes. Many important cellular processes, including the central carbon metabolism, stress response, and protein folding are affected by changing the growth temperature [31]. However, Cos et al. [32] observed less protein expression above 32°C with Pichia pastoris, which matches the results obtained in this report. The resulting optimum temperature ϑLopt is specific for the expression of the potential malaria vaccine protein but commonly used in Pichia pastoris cultivations as well.

The ability of Pichia pastoris to grow across a relatively broad pH range [33] is well known. In contrast, the recombinant protein stability is closely coupled to the pH-value, controlled in cultivations. Different investigations in this field relating to optimal protein production have shown a wide range in adjusted pH-values [34, 35].

Compared to our results, pH-values have been fixed around 5.5 to reduce protease effects, reported in several works [11, 36].

The stability of the target protein was tested in culture supernatant at different temperatures and pH-values (data not shown). The investigations have shown a significant increase of protease effects at temperatures ≥ 27°C. The degradation of the product increases with lowering the pH-value from 6.0 to 4.8.

In summary, the optimal target protein productivity seems to be a sensitive steady state between accelerating cellular processes and decreasing protease effects in the culture supernatant.

3.3  Contemplation of system sensitivity

In the robustness testing procedure the sensitivity around a response variable optimum was investigated by varying the factor levels in a small range.

In this context ϑL was varied by 1°C and the pH-value by 0.15 pH-units around the center point with the factor settings ϑL = 25.8°C and pH = 5.55. A full factorial design in two levels was applied as regression model. The methanol concentration was kept constant in the optimum at 1 g/L. The statistical insignificance of small methanol concentration changes has been proven in pretests (data not shown). The experimental set up, the start values for cXL, respectively IAP1M, as well as the calculated results are shown in Table 1.

The productivity PRD varies around the mean of the center points within a given tolerance limit of ± 2 SDs.

The calculated Q2-value of 0.13 for the chosen model indicates an extremely weak relationship between the factors and the response.

The ANOVA of the regression model demonstrates the insignificance of the robustness test model.

The low predictive power of the selected model and the high p-value verify robustness of the response secretion productivity for small variations in factor settings around the response variable optimum.

4  Concluding remarks

In this report a fast optimization of malaria vaccine expression with Pichia pastoris via Design of Experiments in a multi-bioreactor plant is presented. The whole DoE procedure comprised 38 experiments, consisting of screening [12], optimization, and robustness testing, and was conducted in 7 sequential/parallel approaches.

The developed multi-bioreactor plant equipped with sterilizable piping, transfer valves and quick connectors is based on industrial oriented sterile design and ensures constant conditions for investigations regarding protein expression.

Several implemented automation structures enable fully automated operations, e.g. piping sterilization, cultivation, and refresh, by using BioPAT® MFCS/win recipes complying with the ANSI/ISA-88.01 standard.

In addition, the implementation of experimental DoE designs into the BioPAT® MFCS/win recipe structure, and furthermore the statistical evaluation of cultivation data has been easily conducted by combining the new BioPAT® MFCS/win DoE module with the DoE software MODDE®.

An optimum in product expression was found in two cultivation procedures, screening for significant parameters/search space [12] and optimization of parameter setting in at least five 24 h cultivation cycles via DoE applications.

Robustness in consistent productivity was observed by varying the factors in a small range. In respect of ICH-regulations the ability of the process to tolerate changes in the process without negative impact on quality was demonstrated.

Regarding the incentives, pointed out in PAT initiative of the FDA [5], the highly instrumented multi-bioreactor plant fulfills the requirements in (i) process observation and understanding perfectly. In the sense of ICH's guideline Q8R2, where process supervision is understood as continuous process verification, improvements by monitoring the target protein and controlling the critical parameters could be highlighted.

Automation structures (ii) for easier process handling were implemented and maintained reproducible research results. Hence, the conducted fully automated cultivation process reduces human errors and improves the plausibility and reliability in process optimization.

The assembled statistical tools for DoE (iii) improved plant automation and biological process understanding in terms of interaction effects of the investigated process parameters. An effective determination of optimum expression conditions was observed by using an experimental design approach. At last, the multidimensional combination and interaction of process parameters has been demonstrated to provide assurance of constant productivity.

In summary the described multi-bioreactor plant enables fast process developments. Product quality is proven if operated within the design space. The small-scale system already allowed seamless scale-up into lab scale (15 L) and into pilot scale (40 L) and could be an interesting tool for pharmaceutical protein production stakeholder to improve manufacturing scale process.

5  Nomenclature

AP1MUV absorption of target protein P1 in media phase M (AU)
cIKmass concentration of component I in subsystem K (g/L)
FKflow rate in or out of subsystem K (L/h)
IAP1Mintegral of UV absorption of target protein P1 in media phase M (AUs)
mimass of component i (g)
NStstirrer agitation speed (rpm)
pO2dissolved oxygen tension (%)
qI/Xcell-specific reaction rate of component I (g/(gh))
tcultivation time (h)
VKvolume of subsystem K (L)
ϑKtemperature in subsystem K (°C)
μcell-specific biomass reaction rate (g/(gh))
C, CO2carbon dioxide
Ggaseous phase
jsample index
k0first sample in screening run k
knfinal sample in screening run k
Lliquid (media and cell) phase
Mmedia phase
MCiMedia component i
O, O2oxygen
P1target protein
R1glycerol reservoir
R2methanol reservoir
R3refresh medium reservoir
S1substrate glycerol
S2substrate methanol
T1titration (acid)
T2titration (base)
wset point
Xbio dry mass
Zbio wet mass


The authors thank the BMBF – German Federal Ministry of Education and Research for financial support, Project “Malariavakzine” – FKZ 1756X09, as well as Umetrics AB, Sartorius Stedim Systems GmbH, HAMILTON Bonaduz AG, and Stäubli Tec-Systems GmbH for providing soft- and hardware components. The malaria vaccine candidate described here were developed under auspices of the European Malaria Vaccine Development Association, grant number LSHP-CT-2007-037506.

The authors declare no conflict of interest.