DK3346_C003.

fm Page 31 Wednesday, March 30, 2005 3:15 PM

3
Process Characterization

JAMES E. SEELY

CONTENTS

3.1 Introduction ................................................................... 32

3.2 Resources and Timing for Process
Characterization Studies .............................................. 33
3.3 Precharacterization Work ............................................. 33
3.3.1 Historical Data Review and
Risk Assessment ................................................. 34
3.3.2 Scale-Down Model Qualification ....................... 39
3.4 Process Characterization Studies ................................ 46
3.4.1 Impurity Clearance ............................................ 46
3.4.2 Screening Experiments ...................................... 47
3.4.3 Interactions between Key Parameters
(The Next Round of Process
Characterization Experiments) ......................... 54
3.4.4 Key and Critical Parameters............................. 60
3.4.5 Setting Acceptance Criteria for In-Process
Performance Parameters: Using Feed
Quality as a Process Input ................................ 61

31

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 32 Wednesday, March 30, 2005 3:15 PM

32 Seely

3.5 Finishing Up: Reports, Follow-Up, etc. ........................ 64

3.6 Future Challenges......................................................... 64
Acknowledgments................................................................... 66
References............................................................................... 66

3.1 INTRODUCTION
Although considered to be a significant time and resource
commitment from Process Development, process character-
Downloaded by [Universite Laval] at 04:37 17 September 2015

ization has been shown to be valuable in ensuring validation

and manufacturing success. Given the expense of producing
biopharmaceuticals at large scale, process characterization
gives an excellent return on investment over the lifetime of
a product or process. Inadequate process characterization can
result in costly lot failures and incidents, failed validation
runs, and difficult inspections [1].
The overall goal of adequate process characterization for
commercial manufacturing processes is to ensure efficient and
successful process validation and the assurance of consistent
process performance [2]. More specifically, process character-
ization provides:
• An understanding of the role of each process step, such
as an understanding of where impurities are cleared
during a particular purification step
• An understanding of the impact of process inputs
(operating parameters) on process outputs (perfor-
mance parameters) and identification of key operating
and performance parameters
• Assurance that process delivers consistent product
yields and purity within all operating ranges
• Acceptance criteria for in-process performance
parameters
In addition, although not a primary reason for doing process
characterization, these studies will frequently uncover areas
for subtle process improvements in terms of process consis-
tency, product yields, or product purity.

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 33 Wednesday, March 30, 2005 3:15 PM

Process Characterization 33

In this chapter we present an outline and some examples

for how to carry out thorough and consistent process charac-
terization. The proposed methods could provide a framework
for carrying out this work. A good portion of this chapter will
describe “precharacterization” studies. These studies are used
to help define the scope of the actual experimental character-
ization work. They also lay the foundation for the experimen-
tal studies by demonstrating the adequacy of scaled-down
process models and analytical methods. A framework and
examples for doing experimental process characterization
Downloaded by [Universite Laval] at 04:37 17 September 2015

work will also be presented.

Finally, we will discuss future directions and challenges
as our approach to process characterization evolves.

3.2 RESOURCES AND TIMING FOR PROCESS

CHARACTERIZATION STUDIES
The driver for the timing of process characterization is the
start of conformance/validation lots. Process characterization
should be completed in time such that the information gained
from these studies can be used to support operating ranges
and acceptance criteria for validation protocols. Thorough pro-
cess characterization may add as much as a year to the overall
process development time, so the completion of commercial
process development work and initiation of process charac-
terization studies should be timed with this factor in mind
[2]. Thorough process characterization requires a fully inte-
grated process characterization team (~8–12 people) including
upstream and downstream processing, analytical depart-
ments, and representatives from pilot and full-scale manufac-
turing. Resource planning from the analytical departments is
especially important, since a single characterization run may
generate several samples for analysis.

3.3 PRECHARACTERIZATION WORK

There are three key aspects to precharacterization work: (1)
historical data review and risk assessment, (2) scale-down
model qualification, and (3) analytical method qualification.

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 34 Wednesday, March 30, 2005 3:15 PM

34 Seely

3.3.1 Historical Data Review and Risk

Assessment
Retrospective review of historical data and risk assessment
analysis can be used to determine operating parameters that
need to be examined experimentally as part of process char-
acterization. Lab notebooks, technical reports, process histo-
ries, run summaries, manufacturing records, and a list of the
operating parameters and the provisional operating ranges
for each unit operation can be used by the process character-
ization team to determine knowledge gaps in the process.
Downloaded by [Universite Laval] at 04:37 17 September 2015

Information from the operating ranges tested during process

development can help identify those parameters that are most
likely to impact the process [2–4]. In particular, experimental
design studies (DOE) from the commercial process develop-
ment work can be useful for identifying key parameters or
even in determining operating ranges in certain instances,
since these experiments are carried out over a range of oper-
ating parameters and may yield information about operating
parameter interactions.
Once data mining is completed, a risk assessment anal-
ysis can be carried out on each unit operation where the
effect and likelihood of an excursion from each operating
parameter range is addressed. Hazard Analysis and Critical
Control Points (HACCP) [5], Failure Mode and Effects Anal-
ysis (FMEA) [6–9], cause-and-effect diagrams [10], and other
risk assessment tools can be used for these purposes. The
FMEA tool assigns a numerical rating to the severity of an
excursion of an operating parameter, the frequency of an
excursion, and the ability to detect the excursion before it
has an impact on the product [6–9]. The combined risk factor
(Risk Priority Number or RPN) is a multiple of these three
variables, giving a rating scale from 1 to 1000 if a 1–10
numerical rating is used [6–9]. This data is usually pre-
sented in the form of a Pareto chart [6–10], and those oper-
ating parameters below a predetermined threshold are
considered non-key and will not be examined in the charac-
terization experiments. It is a good idea to involve not only
scientists who developed the process in the FMEA exercise,

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 35 Wednesday, March 30, 2005 3:15 PM

Process Characterization 35

but also quality and plant engineers since they can bring
insight into the likelihood of certain process excursions and
the ability to detect them. There may be significant differ-
ences between the commercial and first-in-human processes,
and there may be relatively little historical data on the
commercial process. Therefore, it may be important to draw
on any development and historical data from both the first-
in-human and commercial processes.
Probably the biggest challenge in doing FMEA is coming
up with a consistent and not totally subjective risk category
Downloaded by [Universite Laval] at 04:37 17 September 2015

definition system that everyone can agree on. There are a

number of generic risk category definitions available [6–9,11].
A custom-made risk category definition system that we have
used is shown in Table 3.1. One way to better define the rating
system for FMEA is to consider the preferred operating range
for each operating parameter in manufacturing. For example,
although it may be possible to run a process at ±0.1 pH units,
operationally the process may be more robust if it can be run
at ±0.2 units. Examples of some preferred operating ranges
for different operating parameters are shown in Table 3.2. For
the FMEA exercise, we can improve the signal-to-noise ratio
of our analysis if we assume that we are considering the
severity of running the process approximately 3 times outside
the normal operating range for a given operating parameter.
For considering the operating parameter excursion frequency
and the ability to detect them, we can increase our sensitivity
by considering excursions that are just outside the tightest
controllable operating range (Table 3.2).

Case Study 3.1

FMEA analysis for removal of a detergent from a protein
preparation using an ion-exchange chromatography
method is shown in Table 3.3. Scientists who developed
the process determined the severity of an excursion
approximately 2–3 times outside the preferred operating
range. Manufacturing and plant engineers determined
the frequency of excursions outside of the tightest oper-
ating range. Quality control and manufacturing provided
information about the ability to detect these excursions

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 36 Wednesday, March 30, 2005 3:15 PM

36 Seely

TABLE 3.1 Custom-Made Risk Category Rating Definitions

for FMEA
1–10
Scale Severity Occurrence Detection

10 Fails final product specs >90% >50% No way to detect

“Bad” of the time or product lost >25 times defect
or completely unrecoverable per year
9 Fails in-process performance ~30–40% Unit sampling and
parameters 100% of the 15–20 inspection;
time and final product specs times defect not
Downloaded by [Universite Laval] at 04:37 17 September 2015

>50% of the time, or over per year detected until

50% impact on step and after impact on
overall yield process
8 Fails in-process performance ~20% Unit sampling and
parameters ~75% of the 10 times inspection;
time and final product specs per year defect can be
>25% of the time, or approx. detected prior
50% impact on step yield to impacting
and over 25% impact on process
overall yield
7 Fails in-process performance ~10% All units are
parameters ~50% of the 5 times per manually
time; final product purity year inspected;
specs failed 10% of the time, defect not
or 30–40% step yield and detected until
>20% overall yield impact after impact on
process
6 May fail in-process ~5% All units
performance parameters in 2–3 times automatically
~25% of instances; may fail per year controlled;
final product specs 5% of the defect not
time, or approx. 25% step detected until
yield and >10% overall yield after impact on
impact process

(continued)

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 37 Wednesday, March 30, 2005 3:15 PM

Process Characterization 37

TABLE 3.1 Custom-Made Risk Category Rating Definitions

for FMEA (Continued)
1–10
Scale Severity Occurrence Detection

5 Runs on edge of in-process ~2% All units

performance parameters Once a automatically
and may fail these in ~10% year controlled with
instances, or ~10% impact secondary
on step yield and manual
measurable impact on inspection;
Downloaded by [Universite Laval] at 04:37 17 September 2015

overall yield (~5%) defect not

detected until
after impact on
process
4 Measurable effect on in- ~1% All units manually
process performance Once every inspected;
parameters but will not 2–3 defect detected
exceed in-process control years prior to impact
limits, or more measurable on process
effect on step yield (~5%)
3 Slightly measurable impact on ~0.5% All units
in-process quality attribute Once every automatically
parameters or slight but 5 years inspected;
measurable impact on step defect detected
yield (<3%) prior to impact
on the process
2 Measurable effect on non-key, ~0.2% All units
nonquality attribute in- Once every automatically
process performance 10 years controlled with
parameter (i.e., pool secondary
volume, peak position) manual control;
defect detected
prior to impact
on the process
1 Not noticed; no effect on Never Defect is obvious
“Good” performance and would
always be
detected prior
to starting
process

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 38 Wednesday, March 30, 2005 3:15 PM

38 Seely

TABLE 3.2 Examples of Tightest and Preferred

Operating Ranges (±) for Operating Parameters
Tightest Preferred Range for
Operating Operating FMEA Severity
Parameter Range Range and Screening

pH 0.1 0.2 0.3 or 0.4

Time 5% 10% 15–20%
Temperature 1°C 2°C 3 or 4°C
Flow rate 5% 10% 15%
Volume 2% 5% 10%
Downloaded by [Universite Laval] at 04:37 17 September 2015

OD 5% 10% 15%

Note: FMEA considers the severity of an excursion that is 2–3

times outside the preferred operating range. The test range for
initial screening experiments is ~1.5–2 times outside the preferred
operating range.

as well as, perhaps more importantly, the ability to react

to this excursion before it has product impact. For column
loading, the frequency and detection scores were quite
low and were the same whether loading was too high or
too low. However, with regard to severity, too high of a
loading was deemed to have a much greater severity since
it had the possibility of resulting in inadequate detergent
removal. This resulted in a much lower RPN score for
underloading than overloading. Likewise, high and low
flow rates have the same frequency and detection level,
but high flow rates can lead to inadequate removal of
detergent resulting in protein aggregation, giving it a
higher overall RPN score. For those parameters that do
not have a specified range (such as stop collect in this
instance), some judgment has to be made as to how much
of an excursion would have a serious impact. In the case
of this chromatography step, missing the stop collect by
a significant amount could result in detergent break-
through, again resulting in potential product aggregation
and loss of activity.
A Pareto plot of the different operating parameters versus
their respective RPN scores is shown in Figure 3.1. In
most cases with this chromatography step, those with the
highest severity impact had the highest RPN scores and

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 39 Wednesday, March 30, 2005 3:15 PM

Process Characterization 39

are the parameters that warrant further study. However,

another outcome of this exercise may be the identification
of more redundant controls in manufacturing that can
serve to either increase the ability to detect an excursion
or decrease its frequency. For example, adding a pH check
of the column eluate post-equilibration can give us extra
assurance that the column is adequately equilibrated and
decrease the frequency of equilibration errors by increas-
ing our ability to detect them. In this case, it might bring
the frequency and detection scores both down to “2” and
perhaps make it unnecessary to examine equilibration
Downloaded by [Universite Laval] at 04:37 17 September 2015

pH in our process characterization studies. Likewise, hav-

ing an additional way of checking the elution flow rate
could increase our ability to detect an excursion here and,
hence, decrease its frequency. However, the potential
severity of running at too high of a flow rate makes this
excursion something we would want to investigate as part
of our characterization studies, in any case.

3.3.2 Scale-Down Model Qualification

The development of a representative scale-down model of each
unit operation is the Achilles’ heel of good process character-
ization work. If we cannot mimic, to a reasonable degree, the
large-scale manufacturing process with our small-scale stud-
ies, any bench-scale process characterization work becomes
meaningless [2]. Some unit operations, such as homogeniza-
tion and centrifugation steps, are much more difficult to scale
down and typically have to be run in a pilot plant setting.
Other operations, such as chromatography, ultrafiltration,
and microbial fermentation operations, can usually be run at
a bench scale or smaller scales.
In general, the approach to scale-down model qualifica-
tion is to run all operating parameters at the center of the
operating range used for clinical/large-scale manufacturing.
If the process has yet to be run in clinical manufacturing, the
operating parameters should mirror those of the largest pilot
scale runs. If no appropriate large-scale data is available, data
from an earlier manufacturing process may be used; however,
the scale-down model has to give data consistent with the
large-scale runs. It is a good idea to write a protocol that

© 2005 by Taylor & Francis Group, LLC

 DK3346_C003.fm Page 40 Wednesday, March 30, 2005 3:15 PM
40
TABLE 3.3 FMEA Analysis for Removal of Detergent from a Protein Using Ion-Exchange Chromatography
(Case Study 3.1)
Downloaded by [Universite Laval] at 04:37 17 September 2015

Function Potential Potential Severity Potential Frequency Detection

Operational or Failure Effect Score Causes of Score Current Score
Parameter Requirement Mode of Failure (1–10) Failure (1–10) Controls (1–10) RPN

Ion Detergent Bad resin Insufficient 7 Poor supplier; QC 2 Monitored in 2 28

exchange retention removal failure QC; defect
lot of detected
detergent prior to use

Load volume Remove <20 CVs Loss of 3 Miscalculation; 3 MPs 3 27

detergent protein on operator error
resin

Load volume Remove >60 CVs Detergent 7 Miscalculation; 3 MPs 3 63

detergent in product operator error

Start collect Capture Too early Dilute pool 2 Poor PM; incorrect 3 Calibration; 5 30
A280 product start standardization SOP
collect

Start collect Capture Too late Lower yield 5 Poor PM; incorrect 3 Calibration; 5 75
A280 product start standardization SOP
collect

Seely
© 2005 by Taylor & Francis Group, LLC
 DK3346_C003.fm Page 41 Wednesday, March 30, 2005 3:15 PM
Process Characterization
© 2005 by Taylor & Francis Group, LLC

Stop collect Capture Too early Lower yield 5 Poor PM; incorrect 3 Calibration; 5 75
A280 product stop standardization SOP
w/o collect
Downloaded by [Universite Laval] at 04:37 17 September 2015

detergent

Stop collect Capture Too late Detergent 7 Poor PM; incorrect 3 Calibration; 5 105
A280 product stop in product standardization SOP
w/o collect
detergent

Flow rate Remove Low flow Longer 3 Pump failure; 4 SOP; MP; 3 36
detergent rate process power outage; calibration
time, operator error;
possible incorrect
loss of calibration
product
on resin

Flow rate Remove High flow Inadequate 6 Pump failure; 4 SOP; MP; 3 72
detergent rate detergent power outage; calibration
removal operator error;
incorrect
calibration

41
© 2005 by Taylor & Francis Group, LLC
DK3346_C003.fm Page 42 Wednesday, March 30, 2005 3:15 PM

42 Seely

Low load volume

Ion exchange lot

Early start collect

Low flow rate

High load volume

High flow rate

Downloaded by [Universite Laval] at 04:37 17 September 2015

Early stop collect

Late start collect

Late stop collect

0 20 40 60 80 100 120
RPN Score

Figure 3.1 Pareto chart of RPN scores from detergent removal

step (Case Study 3.1).

describes the scale-down model, how it is controlled, and

acceptance criteria for each unit operation in terms of perfor-
mance parameters. Generally, a numerical indication that the
small-scale process is acceptable, such as a t-test tolerance
interval or some other statistical means, should be used, or
there may be a simple requirement that the mean of the data
from the scale-down run be within the historical range of the
large-scale data. Doing replicate runs can improve the level
of statistical rigor for the scale-down model, since it is good
to know something about the variability that occurs with the
unit operation at small scale [3]. This can help in determining
the number of replicates that are needed for the actual char-
acterization studies.
The following chapter in this book provides points to
consider and strategies for scaling down a number of different
process steps, including chromatography, chemical modifica-
tion reactions, ultrafiltration, microfiltration, and several

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 43 Wednesday, March 30, 2005 3:15 PM

Process Characterization 43

other unit operations. Some points to consider for scaling

down microbial fermentation and cell culture processes are
described in the following section.

[Link] Scaling Down Fermentation and Cell

Culture
Comprehensive characterization of microbial fermentation
work requires many experiments and is usually carried out
at the 10-liter scale with some supporting experiments run
at larger scales. In our experience, outputs from 10-liter
Downloaded by [Universite Laval] at 04:37 17 September 2015

microbial fermentation experiments have been fairly repre-

sentative of what we have seen at manufacturing scale. How-
ever, scaling down mammalian cell culture has proven to be
more difficult. Although we have been able to determine the
impact that a change in operating parameters might have on
a cell culture process, we frequently have no good way of
predicting the magnitude of that response from bench-scale
data. Even so, process characterization experiments can still
give insights into which operating parameters are the most
important to control for cell culture processes.
Some of the points to consider when scaling down fer-
mentation and cell culture processes include the following:
• Use the most current manufacturing procedures for
scaled-down process.
• Use released GMP materials whenever possible. This
includes master or working cell bank vials.
• Sterilization times of media and feed should match
manufacturing scale. In some instances, it may be
important to extend the heating time to make sure
there is no effect of the sterilization on the media.
Make sure media mass change from pre- to postster-
ilization is the same at both scales.
• Use same size and shape of shaker flask (baffled or
unbaffled), as well as incubator conditions (shaker
speed, throw, etc.).
• Maintain constant inoculum ratios for seed and pro-
duction fermentation between scales.

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 44 Wednesday, March 30, 2005 3:15 PM

44 Seely

• Operational parameters (pH, temperature, back pres-

sure, dissolved oxygen [DO] set point, etc.) should be
identical to those specified in the manufacturing pro-
cess. Ensure that the DO calibrations are compatible.
• Total airflow should be scaled down on the scale factor
to ensure similar sweeping of CO2 from the fermenta-
tion. The overall oxygen control strategy should be
equivalent to that used at scale (i.e., order of cascade,
back pressures employed, etc.). A maximum agitation
for the small scale should be selected that mimics as
Downloaded by [Universite Laval] at 04:37 17 September 2015

closely as possible power input limitations at scale.

This might be accomplished with theoretical calcula-
tion or through equipment design.
• Addition order for all the ingredients should be the
same at both scales. Also, make sure the hold times
for all ingredients are within the same historical
ranges.
• Antifoam strategy or total amounts of antifoam should
match manufacturing scale. Antifoam usage typically
increases with scale due to higher superficial gas veloc-
ities.
• Both scales should have the same feed rates and step
times.
• Make sure you have good calibration of all DO probes,
pH meters, spectrophotometers, etc., and that they
match between scales.
Some of the key performance parameters to monitor include
the following:
• Growth curves and final OD
• Growth rates
• Titer
• % solids
• Feed/acid-base usage
• Nutrient profiles
• Times (total seed fermentation time, main fermenta-
tion time, etc.)
• Genetic stability

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 45 Wednesday, March 30, 2005 3:15 PM

Process Characterization 45

• Cell viability
• Product quality
All of these are good to monitor as a part of the scale-
down work, but the ones that are deemed critical may be on
a case-by-case basis. Certainly, titer and product quality are
two important performance parameters that should be con-
sidered. In addition, cell viability and percent solids could
impact subsequent cell processing and purification steps.
Therefore, it may be necessary to process one or two more
steps downstream to determine the impact of these parame-
Downloaded by [Universite Laval] at 04:37 17 September 2015

ters on subsequent processing steps or product quality.

[Link] Analytical Methods Qualification

The analytical methods for process characterization studies
should be robust and representative of what will be used for
the commercial process. Ideally, the methods would be vali-
dated; however, for a product in early phase III clinical stage
this may not always be possible. In these instances, the
method should be qualified and developed to the point where
there is a high degree of confidence that it can be validated
at some point in the future. It is important, therefore, to allow
enough time for method qualification prior to process charac-
terization work. This can require anywhere from 3–12 months
depending on the assay, the complexity of the protein and
protein matrices, etc.
Points to consider for method qualification depend on
what the assay is used for. For all assays, critical variables
and nominal target values should be defined. For product
quantification assays, the sample handling and preparation,
particularly for in-process samples, should be established.
There should be minimum interference from matrix compo-
nents and adequate resolution. The linearity and range of the
analysis should be determined, as well as the limit of quan-
tification, if applicable. Reproducibility should be established
based on a statistical equivalence between labs, and the rel-
ative standard deviation should be less than or equal to 5%
for chromatographic methods, if possible. Any critical assay
variables should be identified and the nominal target value

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 46 Wednesday, March 30, 2005 3:15 PM

46 Seely

defined. For product immunoassays, there should be minimal

matrix interference and the antibody specificity should be
confirmed. Linearity, precision, and ranges should be estab-
lished on quantitative immunoassays for both product and
process-related impurities (i.e., host cell proteins). For iden-
tification immunoassays, linearity and precision are not
required, but the limit of detection and limit of quantitation
should be confirmed.
As mentioned earlier, the analytical group supporting
process characterization, whether in a process development
Downloaded by [Universite Laval] at 04:37 17 September 2015

department or in quality control, will be one of the hardest

hit from a resource standpoint for characterization studies. A
single purification run may produce as many as four or five
samples for analysis. The analytical group supporting the
characterization effort must be staffed appropriately so the
sample turnaround does not become too much of a rate-lim-
iting step for the completion of process characterization.

3.4 PROCESS CHARACTERIZATION STUDIES

3.4.1 Impurity Clearance
Much of this data may be available prior to the process char-
acterization studies. An understanding of what each process
step delivers in terms of yield, impurity clearance, or in-
process pool quality should be an outcome of this work. In the
case of cell culture, this may be a titer or some qualitative
assessment of the product (such as the degree of glycosylation
or sialylation). For a chromatography step, it would be not
only what impurities are cleared during the step, but at what
point are they cleared, i.e., in the wash step, before the product
elution, after the product elution, during the regeneration
step, etc. For a diafiltration step, one might examine conduc-
tivity or pH after different turnover volumes. The outcome of
this work would be the identification of key performance
parameters for each process step, which can help in the
design of further characterization experiments. How these
performance parameters are affected by excursions from the

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 47 Wednesday, March 30, 2005 3:15 PM

Process Characterization 47

operating ranges will be the objective of the next set of char-

acterization studies.

Case Study 3.2

This study was designed to track impurity clearance
across a cation exchange capture/purification step for a
recombinant protein made in E. coli. The impurities
tracked included DNA, endotoxin, E. coli proteins (ECPs),
product charge variants (measured by cation-exchange
HPLC), and oxidized methionine (as measured by
Downloaded by [Universite Laval] at 04:37 17 September 2015

reversed-phase HPLC). The clearance of these impurities

was tracked by collecting fractions starting with the load
and going through the stop collect. (In some instances,
we would also collect eluate from the regeneration step;
however, all impurities had cleared prior to this step in
this case.) Clearance of these impurities relative to the
product peak is shown in Figure 3.2A–D. Endotoxin clear-
ance was virtually identical to the DNA clearance and is
not shown. From this data, we can provide a rationale for
pool criteria. In addition, these studies can be used for
addressing certain nonconformances, such as inadequate
wash volumes, early pool collections, etc. This information
is also used to identify the quality indicating performance
parameters that should be monitored for this process
step, which would include all of the impurities tested in
this example.

3.4.2 Screening Experiments

Screening experiments are designed to eliminate the less crit-
ical parameters from further, more rigorous process charac-
terization work. Experimental design (DOE) approaches
can be used for these studies to easily screen a number of
operating parameters [12,13], and there are a number of DOE
software packages that can aid in the design and interpreta-
tion of these experiments [14,15]. Fractional factorial, Plack-
ett-Burman, and D-Optimal DOE designs can be used for
screening [2,12,13]. An example of a simple Resolution III
fractional-factorial design for a fermentation process is
shown in Table 3.4. In this design, only nine experiments are

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 48 Wednesday, March 30, 2005 3:15 PM

48 Seely

24 1100
1000
20 900

Nucleic Acids, µg / mL
800
mg/mL Product

16
700

(diamonds)
(squares)

600
12
500
400
8
Downloaded by [Universite Laval] at 04:37 17 September 2015

300

4 200
100
0 0
1 3 5 7 9 11 13 15
CVs from the Load Start

Figure 3.2 (A) Impurity clearance study (Case Study 3.2) exam-
ining the removal of DNA (A), E. coli proteins (B), reversed-phase
impurities (C), and cation-exchange impurities (D).

24.00 16000
20.00
ng ECP/mg Product
mg / mL Product

12000
(diamonds)

16.00
(squares)

12.00 8000
8.00
4000
4.00
0.00 0
1
3
5
7
9
11
13
15

CVs from the Load Start

Figure 3.2 (B)

required to screen six operating parameters. Results can be

analyzed using Pareto charts or regression analysis [12,13].
While these studies do not allow us to see interactions
between parameters, they enable us to determine the main

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 49 Wednesday, March 30, 2005 3:15 PM

Process Characterization 49

24 100

99
20
98

97
16
mg/mL Product

% RP Purity
96

(diamonds)
(squares)

12 95

94
8
93

92
Downloaded by [Universite Laval] at 04:37 17 September 2015

4
91

0 90
1 3 5 7 9 11 13 15
CVs from the Load Start

Figure 3.2 (C)

24 100

98
20
96

94
16
mg/mL Product

% CEX Purity

92
(diamonds)
(squares)

12 90

88
8
86

84
4
82

0 80
1 3 5 7 9 11 13 15
CVs from the Load Start

Figure 3.2 (D)

effects and identify what parameters have the greatest effect

on key performance parameters. In some cases, it may be
simpler to do “one-off ” studies where only a single operating
parameter is tested and all other parameters are held at their

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 50 Wednesday, March 30, 2005 3:15 PM

50 Seely

TABLE 3.4 Resolution III Fractional-Factorial Design for a

Fermentation Process
Induction
Start Feed Inducer Feed Rate Time
Run pH Temp OD Volume (kg/h) (h)

1 6.7 27 17 9.6 97 17
2 6.7 27 23 10.4 72 13
3 6.7 33 17 10.4 72 17
4 6.7 33 23 9.6 97 13
5 7.3 27 17 10.4 97 13
Downloaded by [Universite Laval] at 04:37 17 September 2015

6 7.3 27 23 9.6 72 17
7 7.3 33 17 9.6 72 13
8 7.3 33 23 10.4 97 17
9 7 30 20 10 84 15

center point. While this approach can require more experi-

ments, it can give results that are more easily interpreted.
Typically, we assume that the performance response over
the operating parameter range tested will be linear; therefore,
only two-level (a high and a low value) studies are used for
these experiments. In addition, the two levels we prefer to test
are at approximately 1.5–2 times the preferred operating range
in manufacturing [2]. Of course, the “preferred” operating
range may be different for different unit operations and with
different processes. (Examples of these ranges for several com-
monly used operating parameters are shown in Table 3.2.) This
approach can enable us to get a clear indication of the effect
of an operating parameter on unit operation performance. The
range is wide enough to see an effect should one exist, yet not
so wide as to make a performance failure inevitable. Because
of this, it can give information on process robustness, i.e., the
ability of the process to run outside the prescribed operating
range. This can be useful for closing out manufacturing inci-
dents and excursions from set operating ranges. In addition,
information from these studies can be used to tell manufac-
turing which parameters need tighter control or narrower
ranges and which require less attention. In some instances
they may be able to run at a more preferred operating range,
and in others they may have to tighten the range (Table 3.3).

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 51 Wednesday, March 30, 2005 3:15 PM

Process Characterization 51

Finally, these studies will clearly separate key from non-key

parameters. If an operating parameter tested over this range
has no significant effect on process performance, we may des-
ignate it a non-key parameter. However, even if excursions
from these operating parameters have no product impact, we
still may monitor them to ensure consistent process control.
Those parameters that have a significant, measurable effect
are identified as key parameters that should be tested in the
next set of characterization experiments.
For some operations, particularly purification steps
Downloaded by [Universite Laval] at 04:37 17 September 2015

where several contaminants are removed, there will be more

than one performance parameter, and some will be more
important than others. For example, an operating parameter
that has a large effect on pool volume and a minimal effect
on product purity would be considered to be of less importance
than a parameter where the reverse was true. In these cases,
the impact of the operating parameters will have to be
weighed relative to their effect on the different performance
parameters.

Case Study 3.3

This study involves a reversed-phase column that is used
to remove several product-related variants as well as host
cell proteins from a recombinant glycoprotein. Risk
assessment analysis determined that there were six
potential key parameters that could affect the perfor-
mance of the chromatography step: pH, bed height, col-
umn load factor, temperature, resin type (elutriated
versus nonelutriated), and flow rate. A near-resolution IV
fractional factorial design was set up to screen these oper-
ating parameters with regard to eight different process
performance parameters, which included % yield, product
variants 1, 2, and 3, host cell proteins, pool volume, reten-
tion time, and peak asymmetry (Table 3.5). (This exper-
iment was a bit unusual in that the ranges tested were
somewhat wider than what we usually test.) The relative
impact of each operating parameter on each performance
parameter was determined using Pareto plots, and the
data is summarized in Table 3.6. The importance of the
performance parameter was multiplied by the impact of

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 52 Wednesday, March 30, 2005 3:15 PM

52 Seely

the operating parameter (size of the effect) to give the

rating for a given operating parameter/performance
parameter pair (e.g., temperature/yield had a relative
effect of 6 × 2 = 12). The total score for each operating
parameter is the sum of the effect on all of the perfor-
mance parameters (e.g., total score for pH was 21.0, tem-
perature was 28.5, etc.). Another factor was added to
account for the “ease of control” for each operating param-
eter to give the final adjusted score (“ADJUSTED
SCORE”). From this analysis, it is evident that load factor
has the largest overall impact on the process; pH, tem-
Downloaded by [Universite Laval] at 04:37 17 September 2015

perature, and bed height have some effect; and flow rate
and resin type have a minimal role. Therefore, for our
next set of experiments, we would only want to focus on
load rate, pH, temperature, and bed height.

TABLE 3.5 Fractional-Factorial Screening Study to

Examine Operating Parameters from a Reversed-Phase
Column (Case Study 3.3)
Run Protein Resin Bed Flow
No. pH Temperature Load Type* Height Rate

1 6 4 4 NE 3 50
2 6 4 4 E 15 100
3 6 4 20 NE 15 100
4 6 4 20 E 3 50
5 7.5 4 4 NE 3 100
6 6.4 7 8 E 8.5 75.5
7 7.5 4 4 E 15 50
8 7.5 4 20 NE 15 50
9 7.5 4 20 E 3 100
10 6 22 4 NE 15 50
11 6 22 4 E 3 100
12 6.4 7 8 E 8.5 75.5
13 6 22 20 NE 3 100
14 6 22 20 E 15 50
15 7.5 22 4 NE 15 100
16 7.5 22 4 E 3 50
17 7.5 22 20 NE 3 50
18 7.5 22 20 E 15 100
19 6.4 7 8 E 8.5 75.5

*E = elutriated, NE = non-elutriated.

© 2005 by Taylor & Francis Group, LLC

 DK3346_C003.fm Page 53 Wednesday, March 30, 2005 3:15 PM
Process Characterization
TABLE 3.6 Impact of Operating Parameters on Different Performance Parameters from a Reversed-Phase
Column (Case Study 3.3)*
Downloaded by [Universite Laval] at 04:37 17 September 2015

Relative Size of Effect (Pareto Plot) Total Effect (Size of Effect X Importance)
Bed Flow Resin Bed Flow Resin
Output pH Temperature Load Height Rate Type pH Temperature Load Height Rate Type

Product purity 1 1 1.5 6.0 6.0 9.0

(column pool)
Yield 2 1 12.0 6.0
Product variant 2 1 0.5 2 1.2 5.0 2.5 10.0 6.0 0.0 0.0
Product variant 1 1 1 2 1 3.0 3.0 6.0 3.0 0.0 0.0
Pool volume 1 2 1 1 3.0 6.0 3.0 3.0
Host cell protein 1 2.4 2.0 0.0 4.8 0.0 0.0 0.0
Peak position 1 2 1 2.0 4.0 2.0
Peak asymmetry 1 1 1.5 1.0 1.0 1.5

TOTAL SCORE 21.0 28.5 27.8 24.5 6.0 3.0

Ease of control 1.5 1.5 3.0 2.0 1.0 1.0
(1 = easiest)
ADJUSTED SCORE 32 43 83 49 6 3
(score × control factor)
RANK 4 3 1 2 5 6
Include in next DOE? Yes Yes Yes Yes No No

* Relative importance of output parameters: product purity = 6, yield = 6, product variant 2 = 5, product variant 1 = 3, pool volume =
3, host cell protein = 2, peak position = 2, peak asymmetry = 1.

53
© 2005 by Taylor & Francis Group, LLC
DK3346_C003.fm Page 54 Wednesday, March 30, 2005 3:15 PM

54 Seely

3.4.3 Interactions between Key Parameters (The

Next Round of Process Characterization
Experiments)
From the screening experiments, we know that the operating
parameters being tested at this juncture have some effect on
process performance. Therefore, we typically test these
parameters only to the edge of their normal or preferred
operating ranges. As in the screening experiments, a DOE
approach may be used. Depending on the number of variables
to be tested, a full-factorial, fractional-factorial, or other
Downloaded by [Universite Laval] at 04:37 17 September 2015

design could be used [12–15]. In general, however, we will

want to use a design where the effect of any suspected inter-
actions can be determined, which will necessarily mean a
higher-resolution (IV or V) experimental design [12–15].
Although we typically assume the performance response to
be linear over the operating ranges tested, some judgment
has to be made as to whether or not this is a valid assumption.
In those cases where nonlinearity is suspected, multilevel
experimental designs should be used. In addition, we gener-
ally look for no more than two-factor interactions, since inter-
actions with more than two variables are quite rare and would
require extensive studies to detect. One of the outcomes of
these experiments may be the identification of other process
weak spots, due to interactions or additive effects between
parameters that were not observed in the initial screening
experiments. In some cases, certain operating ranges may
have to be readjusted. The ultimate deliverable from this set
of experiments is to provide assurance that the process pro-
vides consistent yields and product quality attributes within
the confines of all combinations of the operating limits.

Case Study 3.4

In order to more accurately characterize the behavior of
the chromatography experiment from Case Study 3.3 and
to determine if there were interactions between key
parameters that could result in process failure, we under-
took a second, more rigorous DOE with the four key
inputs identified from the DOE screening studies. An

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 55 Wednesday, March 30, 2005 3:15 PM

Process Characterization 55

TABLE 3.7 A 20-Run, Three-Level D-Optimal Design Study

(Case Study 3.4)
Run No. pH Temperature Protein Load Rate Bed Height

1 5.8 3 5 4.5
2 7.0 3 5 10
3 5.8 3 5 15
4 6.4 3 10 15
5 7.0 3 15 4.5
6 6.4 3 15 4.5
7 5.8 3 15 10
Downloaded by [Universite Laval] at 04:37 17 September 2015

8 7.0 3 15 15
9 7.0 7 5 4.5
10 5.8 7 5 15
11 5.8 7 10 4.5
12 6.4 7 15 10
13 6.4 11 5 4.5
14 5.8 11 5 10
15 7.0 11 5 15
16 7.0 11 10 10
17 5.8 11 15 4.5
18 7.0 11 15 4.5
19 7.0 11 15 15
20 5.8 11 15 15

experimental design with the fewest number of experi-

ments that would provide a model capable of discerning
nonlinear behavior and two-factor interactions was a 20-
run D-optimal design with each of the four inputs set at
three levels: low, intermediate, and high (Table 3.7). The
input settings (high and low) were narrowed compared to
those in the first study since we only wanted to test to
the edge of the proposed operating ranges in this set of
experiments. Column performance was monitored by
seven output parameters: step yield, product purity, prod-
uct variants 1 and 2, an additional assay for product
variant 3, host cell proteins, and pool volume.
The results for the study are show in Table 3.8. Even
though effects, some of them significant, were observed
on most performance parameters, all (with the exception
of host cell proteins) fell within the acceptable range.
There were several instances where host cell proteins

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 56 Wednesday, March 30, 2005 3:15 PM

56 Seely

TABLE 3.8 Results from D-Optimal Experimental Design Study

for Reversed-Phase Column (Case Study 3.4)
Pool Product Product Product Host
Run Volume Product Variant Variant Variant Cell
No. Yield CVs Purity 1 2 3 Protein

1 38.6 2.2 90.5 0.57 0.81 0.5 2077

2 38.7 1.8 93.2 0.67 0.68 0.7 817
3 39.5 1.6 91.1 0.34 N/D* 0.3 1524
4 39.8 2.1 91.1 0.57 0.66 0.5 2596
5 43.2 5.3 90.1 1.29 0.93 1.3 10835
Downloaded by [Universite Laval] at 04:37 17 September 2015

6 36.7 3.9 90.5 1.03 0.85 1.1 4599

7 39.0 2.7 90.1 0.72 0.80 0.8 3658
8 41.3 3.1 91.1 1.04 0.82 1.0 3357
9 41.9 2.4 91.3 0.86 0.78 0.9 265
10 37.0 1.5 91.8 0.27 0.57 0.3 938
11 38.2 2.6 91.0 0.72 0.78 0.8 256
12 36.2 2.7 91.8 0.92 0.80 0.9 5163
13 38.6 2.0 91.7 0.31 0.80 0.5 39
14 33.4 1.5 92.9 0.35 0.67 0.3 137
15 39.4 1.7 92.9 0.42 0.81 0.4 68
16 39.4 2.4 91.7 0.73 0.75 0.9 443
17 35.3 3.0 91.0 0.80 0.77 0.9 4138
18 38.9 4.5 91.8 1.14 0.80 1.4 4861
19 40.4 3.0 91.0 0.83 0.93 1.0 2506
20 33.1 2.2 91.6 0.31 0.86 0.4 4662
ctrl 1^ 39.1 2.0 91.8 0.36 0.84 0.5 2653
ctrl 2^ 37.4 2.2 91.7 0.53 0.80 0.5 3659
ctrl 3^ 36.8 2.2 92.9 0.49 0.85 0.6 N/D*

exceeded the upper acceptance limit (~4000 ppm). Based

on the Pareto plot for host cell protein (Figure 3.3), it is
apparent that the main factor affecting host cell protein
content in the product pool is the column load factor. In
order to ensure that we are consistently below the histor-
ical upper limit for host cell proteins, we set the load
factor at no greater than 10. With this new limit on load
factor, we now have a process that will deliver acceptable
yields and quality attributes within the confines of all
operating ranges. The Pareto plots also gave some hints
as to how we might improve step yields (increasing pH
and decreasing temperature) or increase product purity
(also by decreasing load).

© 2005 by Taylor & Francis Group, LLC

 DK3346_C003.fm Page 57 Wednesday, March 30, 2005 3:15 PM
Process Characterization
© 2005 by Taylor & Francis Group, LLC

Effects Pareto for Product Purity Effects Pareto for Product Variant 1
Effects Pareto for Yield 1 0.5
4

Product Variant 1
Product Purity
0.5 B C
2 A
Downloaded by [Universite Laval] at 04:37 17 September 2015

0
Yield

0
0
-0.5
-2 -0.5
-1

Load

pH ^2

Load * pH
pH * T

bed ht * T

bed ht
pH

bed ht * Load

T
T ^2
-4 -1.5
pH pH* T bed ht^2 bedht T T bed ht pH bed ht^2 pH* T Load

Effects Pareto for Product Variant 2 Effects Pareto for Product Variant 3 Effects Pareto for Host Cell Protein
0.4 0.15 6000
Product Variant 2

Product Variant 3

Host Cell Protein

D E 4000 F
0.2 0.1

2000
0 0.05
0
-0.2 0
-2000
-0.4 -0.05
-4000
Load
*T
bed ht
Load
bed ht *
pH

bed ht ^2

bed ht
Load pH pH ^2 Load *pHT ^2 T bed ht Load pH bed ht T bed ht *
pH

2
Effects Pareto for Pool Volume
Pool Volume

1
G

-1
Load

pH

Load * pH

bed ht^2

bed ht * T
T
pH
bed ht *
Load
bed ht *

bed ht

Figure 3.3 Pareto plots from D-optimal study of reversed-phase column (Case Study 3.4).

57
© 2005 by Taylor & Francis Group, LLC
DK3346_C003.fm Page 58 Wednesday, March 30, 2005 3:15 PM

58 Seely

[Link] Combining Variables

In some instances, it may be possible to combine variables in
such a way as to more quickly explore the “operating space”
of a given unit operation [2,12]. This will greatly cut down on
the amount of experimentation required. Sometimes, how-
ever, it may give you results that are difficult to interpret if
there is a process failure and it only assumes that there are
additive effects between operating parameters. We typically
use this approach when we have some degree of confidence
in our operating ranges and only need confirmatory informa-
Downloaded by [Universite Laval] at 04:37 17 September 2015

tion that the process will perform properly over these ranges.
One example would be to combine pH and conductivity as
inputs to an ion-exchange chromatography step. An even more
extreme example is shown in the following case study.

Case Study 3.5

This study was from a cation-exchange chromatography
step using a linear salt gradient. From an initial screen-
ing study, seven operating parameters were determined
to have a significant effect on the performance of a cation-
exchange step: the low salt buffer conductivity and pH,
the high salt conductivity and pH, the feed conductivity
and pH, and the temperature. These operating param-
eters were combined to the edges of their operating range
in such a way as to obtain the very earliest or latest
retention times, as well as the largest or smallest pool
volumes (Table 3.9). One way of looking at this is as a
two-factor full-factorial design, with retention time and
pool volume as inputs or combined variables and with
step yields, quality attributes, etc. as outputs. Table 3.10
shows the effect of pool volume as an input variable. There
is a slight effect on yield (but nothing outside of what had
been observed historically) and no effect on quality
attributes over the range of possible pool volumes from
this process. Table 3.11 shows the results of an experi-
ment where retention time is set up as an input variable.
At the earliest retention times, the yield drops dramati-
cally. Also, for both the late and early retention times,
there is a slight increase in E. coli proteins. When the
column feed pH and conductivity are kept at their center

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 59 Wednesday, March 30, 2005 3:15 PM

Process Characterization 59

TABLE 3.9 Combination of Operating Parameters for a Cation-

Exchange Step to Obtain Earliest and Latest Retention Times and
Minimum and Maximum Pool Volumes (Case Study 3.5)
Earliest Latest Minimum Maximum
Operating Retention Retention Pool Pool
Parameter Time Time Volume Volume Control

Equil pH Low High High Low Center

Equil conductivity High Low Low High Center
Elution pH Low High Low High Center
Elution conductivity High Low High Low Center
Downloaded by [Universite Laval] at 04:37 17 September 2015

Small ion capacity Low High Center Center Center

Temperature High Low Center Center Center
Feed pH Low High Center Center Center
Feed Conductivity High Low Center Center Center

TABLE 3.10 Effect of Cation-Exchange Pool Volume on Product

Purity and Yields (Case Study 3.5)
Pool % %
Pool Product Volume % RP CEX E. Coli
Volumes Runs Concentration (ml) Yield Purity Purity Proteins

Low 1 20.8 195.6 102.2 95.3 95.1 3776

(1.47 CV)
2 20.8 194.6 105.7 95.2 95.0 3878
(1.47 CV)
Center 1 14.8 257.3 95.7 94.4 95.9 3526
(1.94 CV)
2 14.6 259.3 97.0 94.4 95.2 3833
(1.95 CV)
High 1 11.2 313.1 90.4 96.2 95.2 3144
(2.36 CV)
2 11.7 300.1 89.8 94.7 96.3 3037
(2.26 CV)

points, the yield effect is greatly reduced and there is a

slight reduction in the amount of E. coli proteins (Table
3.12). We concluded from these experiments that to
improve the robustness of this step we needed to lower
the center point of the column feed pH by 0.1 units (main-
taining the operating range at ±0.1 unit).

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 60 Wednesday, March 30, 2005 3:15 PM

60 Seely

TABLE 3.11 Effect of Cation-Exchange Retention Time on Product

Purity and Yield (Case Study 3.5)
Product
Retention Concentration % % % E. Coli
Times Runs (mg/ml) Yield RP Purity CEX Purity Proteins

Early 1 7.7 37.9 92.9 94.8 4682

2 7.0 36.2 94.5 95.2 5195
Center 1 12.7 87.0 94.8 95.7 4068
2 12.7 88.2 95.5 95.6 3387
Late 1 14.1 102.7 93.6 94.4 4289
Downloaded by [Universite Laval] at 04:37 17 September 2015

2 13.5 102.0 93.6 94.1 5085

TABLE 3.12 Effect of Cation-Exchange Retention Time on

Product Yields and Purity when Column Feed pH and
Conductivity Are Run at Their Center Points
Retention % % % E. Coli
Times Runs Yield RP Purity CEX Purity Proteins

Early 1 84.1 95.1 95.8 4590

2 83.5 94.8 96.5 3788
Center 1 94.2 94.2 96.3 3229
2 91.0 94.2 95.8 3520
Late 1 98.6 95.0 95.6 4050
2 98.4 95.5 95.4 4519

The advantage of this study design is that with no more

than six experiments we are able to explore the edges of
the operating space for this unit operation. The disadvan-
tage is that any nonlinear responses or significant inter-
actions between variables could be lost. As with any
design, one has to determine the risk-benefit of combining
variables or using a lower-powered study.

3.4.4 Key and Critical Parameters

Operating parameters that have a significant effect on the
performance of a unit operation are considered key param-
eters. More specifically, this would include those parameters

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 61 Wednesday, March 30, 2005 3:15 PM

Process Characterization 61

from the initial screening experiments that have a significant,

measurable effect on process performance parameters (par-
ticularly product quality attributes, impurities, or step yields).
These parameters should be included as the key operating
parameters for full-scale validation runs, although, in some
instances, there may be additional parameters included in the
full-scale validation runs as key parameters as well.
In addition to key parameters, there may be a subset of
“critical” operating parameters that have an even greater
effect on product quality. The FDA definition clarifies the
Downloaded by [Universite Laval] at 04:37 17 September 2015

definition of critical parameters as “process parameters that

must be controlled within established operating ranges to
ensure that the API or intermediate will meet specifications
for quality and purity” [16]. Most parameters, even most key
parameters, can be run slightly outside their prescribed oper-
ating range without resulting in a failed product or product
intermediate specification. However, there may be a handful
of operating parameters for a given process for which an
excursion outside the prescribed range would result in such
a failure. Generally, for a robust process there should be very
few of these. Those that are identified as critical should be
highly characterized in terms of defining their edge of failure
and identification of any interactions with other operating
parameters that could result in process or product failures.

3.4.5 Setting Acceptance Criteria for In-Process

Performance Parameters: Using Feed
Quality as a Process Input
A number of factors need to be considered for setting accep-
tance criteria for in-process performance parameters. Setting
the criteria too stringently can lead to unnecessary validation
failures. Setting the criteria too loosely may result in product
that may ultimately fail the final product release specifications.
Statistical analysis of pilot, clinical, and commercial-
scale manufacturing data is one of the most common ways of
setting acceptance criteria on key performance parameters.
Frequently, however, the data sets may not be representative
(due to process differences between pilot and large scale) or

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 62 Wednesday, March 30, 2005 3:15 PM

62 Seely

there may be a limited number of manufacturing runs. In

these cases, basing acceptance criteria on a historical range
or a statistical evaluation of the data, such as three standard
deviations or tolerance intervals [16], may not give the appro-
priate acceptable ranges for process performance parameters.
Therefore, it is important to use process characterization data
to determine the process capability and redundancy so that
acceptance criteria are based on what the process can actually
deliver.
In most process characterization studies, representative
Downloaded by [Universite Laval] at 04:37 17 September 2015

feed material should be used, whether this is material from

a seed fermentor feeding a production fermentor, a column
feed, or a feed material going into an ultrafiltration step.
However, to really test the “top-to-bottom” robustness of the
process, the effect of feed quality on each unit operation
should be tested. This will give us an understanding of the
downstream sensitivity to upstream process excursions. For
fermentation, this might include different seed fermentor cell
densities. For cell harvesting, it might include different fer-
mentation media OD or viscosity. In the case of chromatog-
raphy steps, the purity of product from the previous step or
load factor should be considered (although we usually run all
process characterization experiments at the upper end of the
loading range). All other operating parameters (pH, temper-
ature, etc.) are run at the center of their respective ranges,
because the likelihood of having both an operating parameter
excursion and a feed quality excursion is remote.
These experiments can be used to set performance
parameter acceptance criteria for each unit operation. One
way to address this is to run a unit operation under conditions
where it fails to perform adequately. The pool from the failed
unit operation is processed further downstream to see if sub-
sequent process steps can make up for the poor performance
from the failed unit operation and allow for the product to
stay within specifications. These experiments give informa-
tion on process redundancy with regard to different key per-
formance parameters [2].
An example of this is shown in Figure 3.4. In this
instance, column 1 is run in such a way that a product-related

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 63 Wednesday, March 30, 2005 3:15 PM

Process Characterization 63

10
Column 2
9 Column 1
8
7
6
% Impurity 5
4
3
Downloaded by [Universite Laval] at 04:37 17 September 2015

2
1
0
Worst Case Historical Mean
Impurity Range (high)

Figure 3.4 Impurity clearance showing process redundancy

between columns 1 and 2.

variant was present at values higher than the historical

range. When this column 1 pool is processed through column
2, however, the amount of this variant falls within the histor-
ical range for the column 2 pool. Therefore, rather than setting
the acceptance criteria for the column 1 pool based on histor-
ical ranges for the product-related variant, a wider acceptance
criteria can be set due to the process redundancy between
columns 1 and 2. There may be other instances where a pool
has to be processed through more than one additional step
downstream to determine the process redundancy, but the
principle is the same. This method of setting acceptance cri-
teria is scientifically based and can give a more realistic indi-
cation of what the process can actually deliver than statistical
analysis of historical data (standard deviation, tolerance
intervals, etc. [17]), particularly since at the time the process
characterization is carried out there may be very little if any
historical data to set a statistically valid acceptance criteria.
Setting the acceptance criteria for process validation studies
based on what the process can actually deliver will result in
fewer validation failures due to acceptance criteria being set

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 64 Wednesday, March 30, 2005 3:15 PM

64 Seely

incorrectly. Of course, as the process matures and more lots

are run, the acceptance criteria can more easily be set based
on statistical analysis.

3.5 FINISHING UP: REPORTS, FOLLOW-UP, ETC.

Process characterization reports should be written for each
unit operation. These reports include the results from the
clearance studies, the screening and interactions studies, and
the feed quality studies. Key operating parameters and their
Downloaded by [Universite Laval] at 04:37 17 September 2015

respective ranges are identified, as are acceptance criteria for

all key in-process performance parameters. A rationale for
why certain parameters were identified as non-key is included
as well. Data from these reports will be used to support the
validation studies, and the reports should be completed prior
to writing the validation protocols.
After the characterization work is completed, it may be
valuable to go back and repeat the FMEA exercise. The sever-
ity factor for each operating parameter will be known at this
point and this could, in some instances, dramatically change
the outcome of the risk priority number. In this way, plant
engineers can best devote their time to those unit operations
and operating parameters that require the greatest control
and detection.

3.6 FUTURE CHALLENGES

Process characterization requires a significant commitment
of time and resources, but the payoff in terms of better process
understanding, improved success rate in manufacturing, and
avoidance of costly regulatory delays makes it a very worth-
while investment. The approaches described in this chapter
can provide information used for setting operating ranges and
performance parameter acceptance criteria and can give an
indication of overall process robustness.
There are a number of challenges that we continue to
face in process development and in other departments
involved in carrying out process characterization studies.
Some of these include the following:

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 65 Wednesday, March 30, 2005 3:15 PM

Process Characterization 65

• Developing appropriate scale-down models for cell cul-

ture, centrifugation, and other difficult-to-scale
upstream steps. Finding a representative scale that
will still be amenable to the high run number generally
required for thorough process characterization work
can be a real challenge. For cell culture, one approach
may be to use small-scale bioreactors to identify key
operating parameters and then to use larger, more
representative bioreactors for confirming ranges or
studying interactions between key parameters. Even
Downloaded by [Universite Laval] at 04:37 17 September 2015

so, this entails a great deal of work and may still give
information that is of marginal value.
• Use of appropriate analytical methods and analytical
method turnaround time. Having “mature” analytical
methods available in time for starting process charac-
terization studies can be a real problem. The availabil-
ity of more generic analytical method platforms so that
methods can be more easily qualified would help
ensure that process characterization studies are
started at the appropriate time. In the absence of this,
analytical resources will have to be spent “at risk”
earlier in the product development cycle in order to
ensure timely qualification of analytical methods. Also,
since sample analysis can be a major bottleneck for
completing process characterization work, develop-
ment of more rapid and automated methods can help
in sample turnaround time and in planning of subse-
quent process characterization experiments.
• Appropriate resources for process characterization.
Many companies are still dialing in the appropriate
resource requirements for process characterization. It
is important to tailor the process characterization
requirements for each product to key business drivers
in order to have the most efficient use of resources. All
products will have certain requirements with regard
to regulatory commitments, such as providing data
that the process will provide consistent yields and
product quality attributes within the normal operating
ranges. For products with less intensive run rates or

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 66 Wednesday, March 30, 2005 3:15 PM

66 Seely

fewer cost-of-goods issues, broader yield ranges or

higher process excursion rates may be acceptable. In
these instances, it may be possible to reduce the scope
of some of the characterization work by reducing the
number of variables to examine or by testing only to
the edge of the normal operating range for certain
parameters. Another problem encountered is how to
deal with products that have accelerated development
timelines that may not allow time for thorough charac-
terization studies. In these cases, a more “bare-bones”
Downloaded by [Universite Laval] at 04:37 17 September 2015

approach to characterization might be used prior to

conformance runs, with more thorough studies being
carried out later (but still in time for the BLA filing).
It is hoped that we will be able to better address these and
other challenges as the strategies for doing process character-
ization continue to evolve in the biopharmaceutical industry.

ACKNOWLEDGMENTS
The author wishes to acknowledge Steve Rausch, David
Dripps, Carl Richey, and David Smiley for input and discus-
sion on this manuscript.

REFERENCES
1. Bobrowicz, G., The compliance costs of hasty process develop-
ment, BioPharm, 12, 35–38, 1999.
2. Seely, J. and Seely, R., A rational, step-wise approach to process
characterization, BioPharm Int., 16, 24–34, 2003.
3. Gardner, A. and Smith, T., Identification and establishment of
operating ranges of critical process variables, in Pharmaceutical
Process Validation, Sofer, G. and Zabriskie, D., Eds., Marcel
Dekker, New York, 1999, pp. 61–76.
4. Rathore, A.S., Johnson, G.V., Buckley, J.J., Boyle, D.M., and
Gustafson, M.E., Process characterization of the chromatogra-
phy steps in the purification process of a recombinant Escher-
ichia coli–expressed protein, Biotechnol. Appl. Biochem., 37,
51–61, 2003.

© 2005 by Taylor & Francis Group, LLC

DK3346_C003.fm Page 67 Wednesday, March 30, 2005 3:15 PM

Process Characterization 67

5. Armbruster, A. and Feldsien, T., Applying HACCP to pharma-

ceutical process validation, BioPharm, 13, 170–178, 2000.
6. Kieffer, R. et al., Applications of failure mode and effects anal-
ysis in the pharmaceutical industry, Pharm. Tech. Europe, Sept.,
36–49, 1997.
7. Nobel, P., Reduction of risk and the evaluation of quality assur-
ance, PDA J. Pharm. Sci. Technol., 55, 235–239, 2001.
8. Sahni, A., Using failure mode and effects analysis to improve
manufacturing processes, Med. Device Diagn. Ind., July, 47–51,
Downloaded by [Universite Laval] at 04:37 17 September 2015

1993.
9. McDermott, R. et al., The Basics of FMEA, Productivity, Inc.,
Portland, OR, 1996.
10. Burr, J.T., SPC Tools for Everyone, ASQC Quality Press,
Milwaukee, 1993.
11. Rath and Strong Consultants, Six Sigma Pocket Guide, Division
of Aon Worldwide, Lexington, MA, 2001, pp. 26–31.
12. Kelley, B., Establishing process robustness using designed
experiments, in Pharmaceutical Process Validation, Sofer, G.
and Zabriskie, D., Eds., Marcel Dekker, New York, 1999, pp.
29–60.
13. Haaland, P., Experimental Design in Biotechnology, Marcel Dek-
ker, New York, 1989.
14. Juran, J.M. and Godfry, A.B., Juran’s Quality Handbook, 5th
ed., McGraw-Hill, New York, 1999, pp. 47.1–47.77.
15. Montgomery, D.C., Design and Analysis of Experiments, 5th ed.,
John Wiley & Sons, New York, 2001.
16. FDA Guidance for Industry: Manufacturing, Processing or
Holding of Active Pharmaceutical Ingredients, August 1996.
17. Seely, R., Munyakazi, L., and Haury, J., Statistical tools for
setting in process acceptance criteria, BioPharm, 14, 28–34,
2001.

© 2005 by Taylor & Francis Group, LLC