B I O P R O C E S S TECHNICAL

Using Design of Experiments

To Assess Escherichia coli
Fermentation Robustness
Arun Tholudur, Todd Sorensen, Xiaoyue Zhu, Scot Shepard

A
s initial bench-scale processes data that was thus available, we
progress to full-scale identified parameters that were likely
commercial manufacturing, it to affect process performance. Finally,
is important to identify critical we performed statistically designed
parameters and assess their robustness. experiments to assess the impact of
If manufacturing deviations occur, those process parameters. As a result,
researchers with process robustness we were able to establish a data set to
data in hand have critical information support broader parameter ranges for
with which to evaluate the effects of routine manufacturing.
those deviations. With this in mind,
we recently conducted a statistical FERMENTATION PROCESS AND FACTORS
study to assess fermentation process A two-stage seed-expansion procedure
robustness in the manufacture of an provided 100-mL of culture that we
active pharmaceutical ingredient. used to inoculate 15-L production
Design of experiments (DOE) is a vessels containing 9.1 L of complex
structured and efficient methodology fermentation medium with glycerol as
for planning experiments such that the primary carbon source. During the
statistically valid relationships between fermentation, we maintained
factors affecting a process and its temperature at 37 °C, pH at 7.10, and
outputs can be established. DOE was dissolved oxygen (DO) at 15% of air
a key element in our study because it is saturation. Agitation and airflow rates WWW.PHOTOS.COM

ideally suited to establishing critical remained constant throughout

parameters and their acceptable ranges. fermentation while pure oxygen was Factor A: culture-medium
To begin the study, we first scaled sparged to maintain DO at the desired temperature
down an Escherichia coli fed-batch set point. The pH of the medium was Factor B: fermentation-medium pH
fermentation process that was already maintained at its set point by means of Factor C: DO concentration
in full-scale operation producing a base (NaOH) addition. Factor D: induction cell density
recombinant protein in the form of Following an initial batch growth (OD600) — cell density at which feeds
inclusion bodies. Next, using historical phase, when optical density (OD600) initiate
reached 10 ± 1 absorbance unit (AU), Factor E: yeast extract/glycerol
two constant-rate feed solutions began: feedrate — carbon/nitrogen mixture
PRODUCT FOCUS: FERMENTATION PRODUCTS
one solution a yeast extract/glycerol Factor F: lactose feedrate — induces
PROCESS FOCUS: PRODUCTION (YEG) feed and the other a lactose product formation
feed. Adding the lactose solution We evaluated the following three
WHO SHOULD READ: PROCESS initiated product expression. After 4.5 responses.
DEVELOPMENT, MANUFACTURING, AND hours, the feeds were stopped and the Response 1: harvest cell density
ANALYTICAL STATISTICIANS cells were harvested. (OD600) — acceptable harvest range
We studied process robustness by compatible with downstream
KEYWORDS: ROBUSTNESS, OPTIMIZATION,
focusing on the six factors below. The processing capacity
DESIGN OF EXPERIMENTS, FED-BATCH, E. COLI
first three specifically affect cellular Response 2: product band purity —
LEVEL: INTERMEDIATE growth rates and product formation. desired product protein amount

2 BioProcess International OCTOBER 2005

Figure 1: DOE design table. Six factors and three responses appear in their standard run order to summarize the experimental design. This matrix is the
initial step in identifying critical process parameters and ranges. To protect proprietary data, actual responses are omitted here.
��� ��� ����� �������� �������� �������� �������� �������� �������� ���������� ���������� ����������

� � � �� ��� � � ��� ���

� �� � �� ��� � � ��� ���
� �� � �� ��� � � ��� ���
� �� � �� ��� � � ��� ���
� �� � �� ��� �� � ��� ���
� �� � �� ��� �� � ��� ���
� � � �� ��� �� � ��� ���
� � � �� ��� �� � ��� ���
� �� � �� ��� � �� ��� ���
�� �� � �� ��� � �� ��� ���
�� � � �� ��� � �� ��� ���
�� � � �� ��� � �� ��� ���
�� � � �� ��� �� �� ��� ���
�� � � �� ��� �� �� ��� ���
�� �� � �� ��� �� �� ��� ���
�� �� � �� ��� �� �� ��� ���
�� � � �� ��� �� �� ��� ���
�� � � �� ��� �� �� ��� ���
�� �� � �� ��� �� �� ��� ���
�� �� � �� ��� �� �� ��� ���

relative to other cellular proteins being We applied a ¼-fractional factorial Figure 2: Half-normal probability plot.
made (Ideally, batch-to-batch variation design that allowed us to estimate Induction OD (D) and the temperature-
should remain relatively low.) main effects unambiguously. That induction OD interaction (AD) significantly
affect harvest OD. Although temperature, by
Response 3: inclusion bodies (IB) proved to be a good compromise itself, is not a significant parameter, it is
— actual product concentration between number of experiments and included in the model to preserve hierarchy.
data quality, allowing us to draw ����������������
EXPERIMENT DESIGN reasonable conclusions. The high and
��
Factorial designs are a type of DOE low levels we selected for this study
�������������������������

that allow researchers to efficiently represented the extremes of normal ��

��
�
explore potentially interacting effects operating ranges, which accentuated all ��
��
of multiple factors on process outputs. effects while we tested 16 unique ��
��
�
Fractional factorial designs are a subset factor combinations. Those 16 ��
of full factorial designs (using fewer combinations ran in four blocks of five ��
��
runs) that can be useful when experiments, with a center point ��
attempting to establish process included in each block. Thus, 20 �
robustness — in which the desired fermentations were performed in all.
outcome is a lack of effect of various The entire design matrix appears as �����������������������������������������������������������������
�����
factors on process outputs. Figure 1. To protect proprietary data,


With six factors, several fractional all response columns in the figure have
two-level factorial designs are possible. been left blank.
A one-half fractional factorial requires We used a DOE software package
32 experiments, whereas one-quarter (Design-Expert software from Stat-
and one-eighth designs using six Ease, Inc., www.statease.com) to
factors require 16 and 8 experiments, calculate all effects that could be Aliasing (a
respectively. With higher fractional estimated from the data, and then we confounding of
factorial designs, fewer experiments are assessed their significance. Subsequent effects that makes
required; however, estimation of effects software analysis found our
becomes complicated because of fermentation process to be robust with
results ambiguous)
aliasing (a confounding of effects that respect to purity percentage of is the PRICE we pay
prevents us from unambiguously inclusion bodies and protein for running a
estimating some effects and/or concentration. However, induction OD reduced number of
interactions), which is the price we pay and temperature affected harvest OD,
for running a reduced number of as depicted in Figure 2. Equally
experiments.
experiments. important, an interaction occurred

OCTOBER 2005 BioProcess International 3

Figure 3: Interaction graph. When cells are
induced at the low level, harvest OD is temperature. (Because the induction robust. As well as developing processes
insensitive to temperature. However, wide OD–temperature interaction term is “from scratch,” we are significantly
variations in harvest OD, with respect to integral to our model, the parent enhancing existing processes for
temperature, occur when cells are induced at
higher ODs.
temperature term is included to transfer to full-production facilities.
�����������������
preserve hierarchy even though it does DOE is helping us select and
��� not significantly affect harvest OD on develop serum-free and protein-free
its own.) media formulations in addition to
�����������������������

��� High Figure 3 shows the induction OD– nutrient feed strategies that optimize
Indu
ction
temperature interaction effect on cell growth and protein productivity.
harvest OD. Harvest OD is relatively DOE is also being used for process-
��
insensitive to temperature at low robustness pilot studies in preparation
induction ODs, but the graph reveals for process validation and routine
Low Induction
��
the temperature effect to be more commercial manufacturing. Finally, as
dramatic at high induction ODs. demonstrated in this article, DOE also
�� Harvest OD insensitivity to can be used for further characterization
������������������������������������������������������������������������� temperature is also revealed in Figure of process ranges even as routine
��������������
4, which is a three-dimensional plot of manufacturing continues.
between induction OD and harvest OD as a function of induction
OD and temperature. The plot FOR FURTHER READING?
Figure 4: Harvest OD 3-D plot. Harvest OD
appears as a function of induction OD and indicates insensitivity of harvest OD to 1 Anderson MJ, Whitcomb PJ. DOE
Simplified: Practical Tools for Effective
temperature. This graphic analysis indicates temperature excursions.
Experimentation. Productivity Press: New York,
harvest OD’s lack of sensitivity to specific Figure 5 is an overlay plot generated NY, 2000.
parameters. by setting ranges on input and output 2 Montgomery DC. Design and Analysis of
variables in the DOE software.
Inc.: Hoboken, NJ, 2004. 
Experiments, 6th Edition. John Wiley & Sons,
��� Because our manufacturing process has
�� a specification of 50–100 AU for
harvest OD, we sought factor
����������

��
�� combinations resulting in harvest OD
Corresponding author Arun Tholudur is a
��
specification violations. The overlay senior scientist in the upstream
plot indicates that within the ranges of development and production group at
induction OD and temperature Diosynth Biotechnology, 3000 Weston
�� studied, harvest OD is always >50 AU. Parkway, Cary, NC 27513; 1-919-388-5721,
��
��
��
�� However, at high induction ODs and fax 1-919-678-0366; arun.
��� ��
��
��� ��
��
���
� low temperatures, harvest OD can tholudur@diosynth-rtp.com; www.
��
��
� ���
�� � ��
�� exceed 100 AU. diosynthbiotechnology.com. Todd
� ����
Although current manufacturing Sorensen and Xiaoyue Zhu are scientists,
specifications state that temperature and Scot Shepard is director of the
should be maintained at 37 ± 1 °C, our upstream process development group at
Figure 5: Overlay plot. Shaded area in upper Diosynth Biotechnology. The authors’
left shows temperature and induction OD study indicated that the process is
primary focus is on developing and
combinations resulting in harvests greater robust with respect to temperatures implementing scalable microbial
than 100 AU. Because specifications are set at over 37 ± 2 °C for a wide range of fermentation and animal cell culture
50–100 AU, any batch resulting in harvest induction ODs. As a result, broader processes for manufacturing active
OD > 100 AU could be rejected. This plot
acceptable operating ranges have been pharmaceutical ingredients. Their company
reveals that temperature and induction OD
ranges are almost entirely within range except established for temperature. Although is a contract manufacturer with process
for the shaded region. manufacturing specifications were not development and manufacturing
������������ changed, data from this study can be operations in the United States and Europe
��
��������������� used in technical evaluations and to experienced in biopharmaceutical
close out any deviations that may occur production using a wide range of cell types
�� during routine manufacturing. including mammalian, hybridoma, and
���������������

insect cells, E. coli, and Pichia pastoris.

�� THE DOE HORIZON

We have developed several DOE-
characterized processes that are in use
��
at our company for producing clinical
and commercial recombinant protein
�
��������������������������������������������������������������������������������
products. Our goal is to develop
�������������� cGMP production processes that are
economically viable, scalable, and

4 BioProcess International OCTOBER 2005