RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`Application of a Quality by Design Approach to the Cell Culture
`Process of Monoclonal Antibody Production, Resulting in the
`Establishment of a Design Space
`
`HIROAKI NAGASHIMA, AKIKO WATARI, YASUHARU SHINODA, HIROSHI OKAMOTO, SHINYA TAKUMA
`
`API Process Development Department, Chugai Pharmaceutical Company, Ltd., Tokyo, Japan
`
`Received 11 August 2013; revised 9 September 2013; accepted 13 September 2013
`
`Published online 3 October 2013 in Wiley Online Library (wileyonlinelibrary.com). DOI 10.1002/jps.23744
`
`ABSTRACT: This case study describes the application of Quality by Design elements to the process of culturing Chinese hamster ovary
`cells in the production of a monoclonal antibody. All steps in the cell culture process and all process parameters in each step were
`identified by using a cause-and-effect diagram. Prospective risk assessment using failure mode and effects analysis identified the following
`four potential critical process parameters in the production culture step: initial viable cell density, culture duration, pH, and temperature.
`These parameters and lot-to-lot variability in raw material were then evaluated by process characterization utilizing a design of experiments
`approach consisting of a face-centered central composite design integrated with a full factorial design. Process characterization was
`conducted using a scaled down model that had been qualified by comparison with large-scale production data. Multivariate regression
`analysis was used to establish statistical prediction models for performance indicators and quality attributes; with these, we constructed
`contour plots and conducted Monte Carlo simulation to clarify the design space. The statistical analyses, especially for raw materials,
`identified set point values, which were most robust with respect to the lot-to-lot variability of raw materials while keeping the product
`quality within the acceptance criteria. C° 2013 Wiley Periodicals, Inc. and the American Pharmacists Association J Pharm Sci 102:4274–
`4283, 2013
`Keywords: antibody; biotechnology; cell culture; design of experiments; design space; failure mode and effects analysis; mathematical
`models; multivariate analysis; proteins; quality by design
`
`INTRODUCTION
`
`Quality by Design (QbD) is a science- and risk-based approach
`to pharmaceutical development and manufacture applicable to
`all stages of a product’s lifecycle.1,2 The QbD approach for bio-
`pharmaceutical manufacturing processes had its beginnings in
`2002 with the Current Good Manufacturing Practices for the
`21st Century initiative launched by the United States Food
`and Drug Administration,3 and the QbD concepts introduced
`there were subsequently followed by the International Con-
`ference on Harmonisation of Technical Requirements for Reg-
`istration of Pharmaceuticals for Human Use (ICH) guidance
`documents.4–7 The traditional approach to pharmaceutical de-
`velopment has been a rigid manufacturing process producing
`a product of variable quality that can be controlled by speci-
`fications; that approach is being replaced by an enhanced ap-
`proach based on QbD elements. Biopharmaceutical manufac-
`turing processes must be developed to produce a product that
`above all meets the needs of the patients while maintaining
`consistent efficacy and safety. Deep understanding of the re-
`lationships among process parameters, quality attributes, and
`
`Abbreviations used: CE-SDS, capillary electrophoresis-sodium dodecyl sulfate;
`CEX, cation exchange chromatography; CHO, Chinese hamster ovary; CPP,
`critical process parameter; CQA, critical quality attribute; DoE, design of ex-
`periments; FMEA, failure mode and effects analysis; HPLC, high-performance
`liquid chromatography; IVCD, initial viable cell density; mAb, monoclonal anti-
`body; QbD, quality by design; RPN, risk priority number; TOST, two one-sided
`t-test; VCD, viable cell density.
`Correspondence
`to: Hiroaki Nagashima (Telephone: +81-3-3968-6115;
`Fax: +81-3-3968-6259; E-mail: nagashimahra@chugai-pharm.co.jp)
`
`Journal of Pharmaceutical Sciences, Vol. 102, 4274–4283 (2013)
`C° 2013 Wiley Periodicals, Inc. and the American Pharmacists Association
`
`efficacy and safety is necessary to achieve such pharmaceutical
`development.
`The QbD approach is a more scientific, systematic, and com-
`prehensive approach that builds quality into the process in-
`stead of the traditional process of testing the quality of the
`product throughout the biopharmaceutical’s lifecycle (e.g., de-
`velopment, manufacture). It is said that pharmaceutical com-
`panies can understand their products more deeply and proceed
`with development more flexibly and robustly through a QbD
`approach, and if the development and manufacture of biophar-
`maceuticals are carried out according to QbD elements using
`appropriate tools, both pharmaceutical companies and regula-
`tory agencies will be able to reap the resulting benefits.8 For
`example, the design space for a manufacturing process is one of
`the most important components in the QbD approach. The ICH
`guidance documents define design space and working within
`the design space is not considered as a change. Some of the
`expected benefits of QbD for the biopharmaceutical industry
`thus include a reduction in the number of postapproval pro-
`cess changes and an increase in flexibility for pharmaceutical
`companies.9,10
`Interpretations of the application of QbD described in the
`guidelines differ within the industry, and concrete case studies
`of QbD as applied to biopharmaceutical drug substances are
`still limited. Some reports show original QbD approaches that
`focus on risk assessment and process characterization using de-
`sign of experiments (DoE) for the processes of cell culture11–20
`and purification,13,15,21,22 and the CMC Biotech Working Group
`published a mock case study covering all of the QbD elements
`for an antibody drug substance and drug product.23 Interpre-
`
`4274
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`Pfizer v. Genentech
`IPR2017-020(cid:20)(cid:28)
`Genentech Exhibit 2046
`
`

`

`tation of these initiatives by industry has been an ongoing
`process, with some notable industry and regulatory collab-
`orations for biopharmaceuticals, and each pharmaceutical
`company should take on the challenge of this approach and
`find its own way.24,25
`This study shows our QbD approach as applied to the process
`of culturing Chinese hamster ovary (CHO) cells in the produc-
`tion of a monoclonal antibody (mAb). The aim of this QbD case
`study is to understand the cell culture process more deeply
`and to increase the sophistication of the QbD approach itself.
`The QbD approach has four important elements—identification
`of critical process parameters (CPPs), identification of critical
`quality attributes (CQAs), understanding the relationships be-
`tween CPPs and CQAs, and establishment of an appropriate
`control strategy. In this case study, we focused on quality risk
`management using failure mode and effects analysis (FMEA),
`and we conducted a multivariate study using DoE to establish
`a design space.
`Risk assessment was composed of risk identification, risk
`analysis, and risk evaluation; it was carried out for the cell
`culture process by using FMEA on the basis of established
`scientific knowledge and our experimental knowledge. FMEA
`is a common and useful risk analysis tool to find process pa-
`rameters that require further process characterization.12,13,16
`Process characterization experiments are typically conducted
`using DoE to evaluate not only the main effects of parameters
`but also their interactions.11–20 The potential CPPs that were
`identified by this risk assessment of the production culture step
`in the cell culture process were experimentally evaluated on a
`scaled down model by a multivariate study using DoE. A design
`space was established on the basis of those results and checked
`by Monte Carlo simulation. Because lot-to-lot variability of raw
`materials is said to contribute greatly to product quality, safety,
`and process performance,26–28 our multivariate study included
`an evaluation of material variability by finding the interac-
`tions between material lot and process parameters. The scaled
`down model used for process characterization was confirmed
`to be appropriate by demonstrating that its performance was
`comparable to that of a 2500 L bioreactor.
`Through this QbD case study, we were able to understand
`the cell culture process over a range of operation wider than
`that traditionally used. In this article, we present the QbD
`approach we have developed for the cell culture process of a
`mAb currently under clinical development.
`
`MATERIALS AND METHODS
`
`Cell Culture
`
`A CHO cell line expressing a glycosylated mAb was used in
`this case study. Cells were grown in suspension using a serum-
`free cell culture medium. Cells were subcultured using spinner
`flasks (Bellco Glass, Vineland, New Jersey) in a CO2 incubator
`(Thermo Scientific, Waltham, Massachusetts). Production cul-
`ture was performed in fed-batch mode with continuous feeding.
`We used 1 L bioreactors (ABLE, Tokyo, Japan) as the scaled
`down model for process characterization, and a 2500 L biore-
`actor (Hitachi, Tokyo, Japan) for the qualification of the scaled
`down model. Cell culture fluids were centrifuged and filtered to
`remove cells for analysis, and antibodies were purified from the
`fluids with a POROS A20 protein A column (Life Technologies,
`Carlsbad, California).
`
`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`4275
`
`Analytical Methods
`
`Viable cell density (VCD) and viability were measured by using
`a Cedex automated cell counter and analyzer (F. Hoffmann-La
`Roche, Basel, Switzerland). Offline pH was measured by us-
`ing a GASTAT blood gas analyzer (Techno Medica, Kanagawa,
`Japan). Antibody titer was measured by using protein A affinity
`high-performance liquid chromatography (HPLC). Charge vari-
`ants were measured by using cation-exchange chromatography
`(CEX) HPLC. N-linked glycosylation (e.g., high-mannose) was
`measured by hydrophilic interaction chromatography HPLC
`after cutting with N-Glycosidase F (F. Hoffmann-La Roche) and
`labeling with 2-aminobenzamide (Sigma–Aldrich, St. Louis,
`Missouri). Whole antibodies (H2L2) were measured by capillary
`electrophoresis-sodium dodecyl sulfate (CE-SDS) using the PA
`800 plus Pharmaceutical Analysis System (Beckman Coulter,
`Brea, California).
`
`Risk Assessment and Process Characterization
`
`A cause-and-effect diagram was used for risk identification, and
`all process steps and process parameters in the cell culture pro-
`cess were defined. The method applied to the risk analysis and
`risk evaluation of process parameters in the cell culture process
`was based on FMEA. In this case study, two study designs gen-
`erated by DoE were integrated and used for the process char-
`acterization. The parameter IVCD, pH, and temperature were
`evaluated by face-centered central composite design, and cul-
`ture duration and material lot were evaluated by full factorial
`design. The process characterization study was performed with
`a total of 30 conditions. For the purposes of this case study, titer
`was chosen as a representative key performance indicator and
`the main peak in CEX, H2L2 in CE-SDS, and high-mannose
`were chosen as representative CQAs with which to analyze
`the impact of process parameters statistically. Multivariate re-
`gression analysis was performed for each performance indi-
`cator and quality attribute to establish mathematical models.
`Main effects and interaction effects of continuous factors on
`responses were considered statistically significant when the F
`ratio—which means the F value of the factor compared with
`the F value of error—was more than 2.0. If a model was judged
`as rational from statistical and scientific views, contour plots
`were generated and superimposed to define the design space.
`Monte Carlo simulation was performed to discover the distri-
`bution of model outputs as a function of random variation in
`the parameters and model noise. To evaluate the differences
`arising from lot-to-lot variability in raw material, the main ef-
`fect of material lot and the interaction effects between material
`lot and the other parameters were also incorporated into the
`prediction models.
`
`Qualification of Scaled Down Model
`
`Culture performance indicators and quality attributes of the
`scaled down model were compared with those of the large-scale
`production bioreactor to demonstrate the appropriateness of
`the scaled down model. We compared data from 7 runs of the
`scaled down model bioreactor, which ran during process devel-
`opment as a control with data from 6 runs of the 2500 L biore-
`actor used for the production of the clinical drug substance. The
`cells used for small-scale culture were derived from the same
`working cell bank as used for large-scale production. Five cell
`tubes were used for the small-scale runs and two cell tubes were
`used for the large-scale runs. Performance indicators, such as
`
`DOI 10.1002/jps.23744
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`

`

`4276
`
`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`Figure 1. A part of the cause-and-effect diagram identifying the process steps and associated process parameters in the cell culture process.
`All identified process parameters were evaluated for further risk assessment. CQA: critical quality attribute, KPI: key performance indicator.
`
`cell growth and titer, and quality attributes, such as CEX vari-
`ants and whole antibody, were compared between the scales.
`Comparability between the scaled down model and the large-
`scale bioreactor was judged by using the mean ± 3SD of each.
`
`RESULTS AND DISCUSSION
`
`Risk Assessment for Process Parameters
`
`Quality risk assessment comprises risk identification, risk
`analysis, and risk evaluation. A cause-and-effect diagram—also
`called a fishbone diagram or Ishikawa diagram—is commonly
`used in many fields for risk identification.21,29 All process steps
`and all process parameters in the cell culture process were iden-
`tified by cause-and-effect diagram. Part of this cause-and-effect
`diagram is shown in Figure 1, indicating that the inoculum
`train step has 11 parameters, the production medium reconsti-
`tution step has 10 parameters, the feed medium reconstitution
`
`step has 12 parameters, and the production culture step has 16
`parameters. In total, the cell culture process was found to have
`14 steps and 176 process parameters (data not shown).
`Risk analysis and evaluation should be the qualitative or
`quantitative process of linking the likelihood of occurrence and
`severity of harm and the ability to detect the harm.5 Although
`many risk analysis tools are known, and no one tool or set
`of tools is applicable to every situation, FMEA is one of the
`most common and useful tools for risk analysis in pharmaceu-
`tical process development.12,13,16 The risks of all of the process
`parameters in the 14 cell culture steps were analyzed and eval-
`uated using FMEA. Each process parameter was assessed in
`terms of three factors: severity of impact, probability of occur-
`rence, and likelihood of detection (Table 1). Severity of impact
`was ranked on a five-level scale with a score of 1–11; a score
`of 11 indicates that the parameter has already been found to
`have a deleterious impact on CQAs or key performance indica-
`tors. Probability of occurrence was ranked on a four-level scale
`
`Table 1. Definitions Used for FMEA
`
`Score
`
`Definition
`
`1
`2
`4
`
`7
`11
`
`1
`2
`3
`5
`
`1
`3
`5
`
`(A)
`It is known that the parameter does not affect CQAs and KPIs because we already have data.
`The parameter is not expected to affect CQAs and KPIs based on experience and prior knowledge (Other cell lines’ data show no effect).
`The parameter might affect CQAs and KPIs based on experience, prior knowledge, and literature (Some other cell lines’ data show
`effects).
`The parameter is expected to affect CQAs and KPIs based on experience and prior knowledge (Other cell lines’ data show effects).
`It is known that the parameter affects CQAs and KPIs because we already have data.
`
`(B)
`Never in the foreseeable future.
`Never, but might happen.
`Rare (less than once per 20 batches, below 5%)
`Sometimes (more than once per 20 batches, over 5%)
`
`(C)
`Change can be detected and dealt with immediately.
`Change can be detected and dealt with within a day.
`Change cannot be detected or dealt with.
`
`(A) Severity of impact, (B) probability of occurrence, and (C) likelihood of detection.
`CQA, critical quality attribute; KPI, key performance indicator.
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`DOI 10.1002/jps.23744
`
`

`

`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`4277
`
`Table 2. Results of Failure Mode and Effects Analysis for the Production Culture Step
`
`Process Parameter
`
`Control Space
`
`Characterization Range
`
`IVCD (×105 cells/mL)
`Duration (day)
`pH (day 4–14)
`Feed start timing (h)
`pH control start timing (h)
`Cell age (day)
`pH (upper, day 0–3)
`Minimum glucose concentration (g/L)
`Feed rate (%)
`Dilution rate (%)
`Inner pressure (MPa)
`Temperature (◦C)
`Agitation power (rpm)
`Viability (%)
`Working volume (%)
`Dissolved oxygen (%)
`
`2.03–2.79
`13.4–13.8
`6.70–6.88
`72.00–72.02
`72.00–72.32
`–
`7.16–7.28
`–
`24.6–28.6
`7.2–8.1
`0.013–0.017
`36.7–37.2
`18.0–18.1
`98.8–99.8
`−0.4–0.1
`28.8–70.9
`
`1.20–2.80
`13.0–14.0
`6.65–6.95
`60.00–96.00
`60.00–96.00
`130
`<7.40
`>0.2
`20.0–40.0
`6.0–9.0
`0.010–0.020
`36.5–37.5
`17.0–19.0
`>80.0
`−10.0–10.0
`>10.0
`
`S
`
`7
`7
`7
`4
`4
`7
`4
`4
`4
`2
`2
`7
`2
`4
`2
`2
`
`O
`
`D
`
`RPN
`
`2
`2
`2
`2
`2
`1
`2
`2
`2
`2
`2
`2
`2
`2
`2
`2
`
`5
`5
`3
`5
`5
`5
`3
`3
`3
`5
`5
`1
`3
`1
`1
`1
`
`70
`70
`42
`40
`40
`35
`24
`24
`24
`20
`20
`14
`12
`8
`4
`4
`
`IVCD, initial viable cell density; S, severity of impact score; O, probability of occurrence score; D, likelihood of detection score; RPN, risk priority number.
`
`with a score of 1–5; a score of 5 indicates that the parame-
`ter has the highest probability of deviating from the control
`space. Likelihood of detection was ranked on a three-level scale
`with a score of 1–5; a score of 5 indicates that the parameter
`cannot be detected or cannot be dealt with even if detected.
`Considering that the manufacturing process is operated within
`a certain control space, characterization ranges should be wider
`than the control space. Characterization ranges were defined
`as “ coefficient × control space” and were used throughout the
`risk assessment and process characterization as ranges within
`which the set point values of process parameters were changed.
`A coefficient of 3 was basically used to calculate the character-
`ization ranges; however, when the resulting ranges were so
`wide that product quality or process performance had to devi-
`ate beyond preferable ranges, a coefficient of less than 3 was
`used. In the risk assessment, the severity of impact for each
`process parameter was evaluated within the characterization
`range, whereas probability of occurrence and likelihood of de-
`tection for each process parameter were evaluated at the edges
`of the control spaces. The risk analysis was conducted on the
`basis of information taken from the literature, experimental
`and manufacturing data on this and other similar processes
`during developments, and general scientific principles.
`Table 2 summarizes the risk analysis results for the produc-
`tion culture step. Although in the risk analysis, the severity
`of impact scores of process parameters were defined with re-
`spect to each quality attribute and performance indicator, the
`highest score is taken as the severity of impact score. Each
`risk priority number (RPN) was calculated by multiplying the
`scores for severity of impact, probability of occurrence, and like-
`lihood of detection, and the resulting scores were used to judge
`which process parameter should be subsequently evaluated in
`the process characterization. The RPN threshold was deter-
`mined by experts in each function and, based on an acceptable
`level of risk at this stage, was set at 40. IVCD, culture duration,
`and pH, each of which had an RPN of over 40, were defined as
`potential CPPs and were evaluated in the process characteri-
`zation. From among all the process steps, FMEA identified no
`potential CPPs other than those for the production culture step
`(data not shown).
`
`Parameters, which had an RPN below the threshold, were
`carefully reviewed by the risk assessment team. Although tem-
`perature in the production culture step was not defined as a
`potential CPP by FMEA, this parameter was evaluated in the
`process characterization for two reasons. One reason was that
`the control space and the characterization range for tempera-
`ture were very narrow because temperature in the large-scale
`bioreactor had been strictly controlled and the risk posed by
`variations in temperature might have been undervalued in the
`FMEA. The other reason was the concern that there are poten-
`tial interaction effects between temperature and other process
`parameters. For these two reasons, temperature was also de-
`fined as a potential CPP. Material lot—medium lot and other
`materials lot in medium reconstitution step—was also defined
`independently of the FMEA as a factor that should be evalu-
`ated in the production culture step because lot-to-lot variation
`in raw material has the potential risk of leading to variability in
`quality attributes and performance indicators. Material lot for
`the production culture was evaluated to statistically analyze
`its variability. Material lot for the subculture was not included
`because the purpose of that step is only to passage and expand
`cells for production culture and the impact on quality attributes
`and performance indicators was considered to be limited. In to-
`tal, four process parameters and material lot were evaluated in
`process characterization.
`
`Process Characterization
`
`The risk assessment stage of our QbD case study identified
`four potential CPPs (IVCD, culture duration, pH, and tem-
`perature) and material lot in the production culture step for
`experimental evaluation in the process characterization stage.
`The actual impact of these potential CPPs and material lot
`on quality attributes and performance indicators was evalu-
`ated by using a DoE-based multivariate study. Because the
`four continuous factors (temperature, pH, IVCD, and duration)
`and the one nominal factor (material lot) cannot be comprehen-
`sively evaluated by a single experimental design, a three-level
`face-centered central composite design to evaluate IVCD, pH,
`and temperature was integrated with a two-level full factorial
`design to evaluate culture duration and material lot to create
`
`DOI 10.1002/jps.23744
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`

`

`4278
`
`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`Figure 2. Multivariate regression analysis for main peak in CEX. (a) Distribution of main peaks in CEX. (b) Correlation between actual values
`and predicted values from the multivariate regression model. (c) Residual analysis for the multivariate regression model. (d) Significant factors
`affecting main peak in CEX. IVCD, initial viable cell density.
`
`a single multivariate experimental design. All of the main ef-
`fects of the five experimental factors and all of the interaction
`effects between the factors could be estimated independently in
`this experimental design. Because the ranges evaluated during
`process characterization are typically wider than the operating
`ranges encountered during manufacturing, we used the char-
`acterization ranges as the experimental ranges during process
`characterization (The characterization range of temperature
`was reset as 36.0–38.0). A total of 30 fed-batch production cul-
`ture runs were conducted using 1 L bioreactors, and the titer,
`main peak in CEX, H2L2 in CE-SDS, and high-mannose were
`used as potential key performance indicators or potential CQAs
`to statistically analyze the impact of process parameters. The
`acceptance criteria for the main peak in CEX and for H2L2 in
`CE-SDS were set using the specifications for manufacturing,
`and the acceptance criterion for high-mannose was set using
`the mean + 3SD of data obtained during process development.
`The acceptance criterion for titer was not set because titer does
`not affect the efficacy or safety of the mAb in patients.
`As a representative example of the procedure used, the
`statistical analysis using JMP software (SAS Institute, Cary,
`North Carolina) for the main peak in CEX is described in detail
`below, although the same analysis was performed for each qual-
`ity attribute and performance indicator. First, the distribution
`of CEX main peaks was monitored by histogram to confirm that
`the data were normally distributed (Fig. 2a). In cases where the
`data were not normally distributed (e.g., viability), a variable
`
`transformation such as logit transformation or Box–Cox trans-
`formation was applied before performing statistical analysis
`that assumes the analyzed data are normally distributed. The
`main effects, interaction effects, and quadratic effects with an
`F ratio of over 2.0 were considered to be significant and were
`incorporated into a multivariate regression model. The actual
`values and the predicted values by the model were shown in a
`scatter plot (Fig. 2b).
`As an indication of how well the mathematical model fitted
`the experimental data, rather than using R2, which increases
`mechanically as more factors are incorporated into the model,
`we used adjusted R2 that takes into consideration the number
`of factors incorporated. A lack of fit is also a useful indication to
`confirm whether sufficient factors including interaction effects
`and quadratic effects are incorporated into the model or not.
`The error measured for these exact replicates is called pure
`error and is the portion of the sample error that cannot be ex-
`plained or predicted by the mathematical model. A lack of fit
`can be significantly greater than pure error if sufficient factors
`including interaction effects and quadratic effects are not in-
`corporated into the model. The adjusted R2 and the P value of
`lack of fit of the mathematical model were 0.914 and 0.302, re-
`spectively, indicating the appropriateness of the mathematical
`model. Residuals of mathematical models should be randomly
`distributed around zero; a scatter plot of residuals by predicted
`values indicated that no notable tendency can be observed
`(Fig. 2c). The statistical analysis showed that the main peak
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`DOI 10.1002/jps.23744
`
`

`

`Table 3. Design Space for the Production Culture Step
`
`Potential CPP
`
`IVCD (×105 cells/mL)
`pH
`Culture duration (day)
`Temperature (◦C)
`
`Range
`
`1.2–2.8
`6.65–6.90
`13–14
`36–38
`
`The range of pH in the design space is narrower than the characterization
`range because of the deviation of high-mannose from the desired criterion.
`CPP, critical process parameter; IVCD, initial viable cell density.
`
`in CEX was most significantly influenced by the main effects of
`temperature, pH, IVCD, and culture duration; by the interac-
`tion effects between temperature and pH and between temper-
`ature and duration; and by the quadratic effects of temperature
`and pH (Fig. 2d). Similar results were obtained for the other
`responses (data not shown), and the prediction models were
`used for the further analysis.
`The four statistical prediction models (for main peak in CEX,
`titer, H2L2 in CE-SDS, and high-mannose) were then depicted
`as contour plots to clarify the multivariate acceptable range in
`the production culture step (Fig. 3). In this case study, spec-
`ifications or mean + 3SD from the process development data
`were used as acceptance criteria. Deviation beyond the accep-
`tance criteria is shown in the contour plots as shaded areas.
`Although no deviation was observed for the main peak in CEX
`and H2L2 in CE-SDS, high-mannose levels were higher than
`the acceptable criteria in some ranges. These results indicate
`that the ranges of some parameters should be narrower than
`the characterization range to establish an appropriate design
`space for the production culture step. One possible constraint
`is to narrow the range of pH from 6.65–6.95 to 6.65–6.90 be-
`
`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`4279
`
`cause pH had the biggest impact on high-mannose and this
`constraint could completely remove all shaded areas from the
`contour plots. Table 3 shows an example design space for the
`production culture step in the cell culture process.
`Another description of a design space can be achieved by ap-
`plying constraints to the prediction models, as has been shown
`in other case studies.12,13,16,18,20,21 In such a case, the area within
`the criteria with values that predict high-mannose can be de-
`fined as the design space. Although this design space will be
`wider than the design space shown in Table 3, the description
`is complicated and difficult to understand intuitively. If con-
`straints are not critical for the process, such as in the current
`case study, a simpler description of the design space might be
`better.
`The distributions of model outputs with random variation in
`the parameters and model noise were evaluated by Monte Carlo
`simulation. The set point value and the perturbative range
`observed in the large-scale production were used as the mean
`and SD of each parameter, respectively. The relative sums of
`mean error of the models were used as the SD of response, and
`a total of 5000 runs were evaluated in the simulation. Figure 4a
`shows the statistical parameters used for the simulation, and
`the results of the simulations are shown in Figures 4b–4e as
`histograms. The minimum values of main peak in CEX, titer,
`and H2L2 in CE-SDS, and the maximum value of high-mannose
`were respectively 79.69%, 94.02%, 87.03%, and 7.30%, which
`were all within the acceptable criteria. These results indicate
`that the production culture step in the cell culture process was
`robust even considering possible noise such as mathematical
`model noise and perturbative noise in large-scale production.
`To promote better understanding of the effect of raw ma-
`terials in the cell culture process, further statistical analysis
`was performed. The main effect of material lot and the interac-
`tion effects between material lot and the other four parameters
`
`Figure 3. Understanding the design space in terms of contour plots. Internal x- and y-axes are temperature and pH, respectively. External
`x- and y-axes are IVCD and culture duration, respectively. Deviation beyond the acceptance criteria is shown as shaded areas. The dashed line
`shows an example of a constraint to establish a design space. IVCD, initial viable cell density.
`
`DOI 10.1002/jps.23744
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`

`

`4280
`
`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`Identifying the distribution of model outputs using Monte Carlo simulation. (a) Statistical parameters used for the simulation.
`Figure 4.
`The set point value and perturbative range observed in large-scale production were used as the mean and SD of each parameter, respectively.
`The relative sums of mean error of models were used as SD of response. Results of 5000 runs in the simulation for (b) main peak in CEX, (c) titer,
`(d) H2L2 in CE-SDS, and (e) high-mannose. IVCD, initial viable cell density.
`
`Figure 5. Robustness analysis for the variability of material lot using the profiles of interaction between material lot and parameters for main
`peak in CEX, titer, H2L2 in CE-SDS, and high-mannose. Each line in the interaction profile represents each material lot. IVCD, initial viable
`cell density.
`
`Nagashima et al., JOURNAL OF PHARMACEUTICAL SCIENCES 102:4274–4283, 2013
`
`DOI 10.1002/jps.23744
`
`

`

`RESEARCH ARTICLE – Pharmaceutical Biotechnology
`
`4281
`
`Figure 6. Qualification of scaled down model. (a) Mean ± 3SD of viable cell density from 2500 L scale production runs (n = 6) are shown
`as dashed lines. Viable cell densities from the scaled down model runs (n = 7) are shown as solid lines. (b) Main peak in CEX, (c) titer, and
`(d) high-mannose were compared between the scales. Results are expressed as mean ± 3SD. VCD, viable cell density.
`
`were incorporated into the four statistical prediction models
`to evaluate differences in variability of material lot. Figure 5
`shows the profiles of interaction between material lot and
`parameters for main peak in CEX, titer, H2L2 in CE-SDS, and
`high-mannose. For each parameter, the set point value that
`results in the smallest difference in response between the dif-
`ferent material lots means that the set point is more robust
`with respect to the variability of the material lot as compared
`with other set points. The effect of temperature on H2L2, pH on
`H2L2, IVCD on H2L2, and IVCD on high-mannose were each af-
`fected by material lot. Although a lower temperature is pr