Andrzej Kraslawski and Ilkka Turunen (Editors) Proceedings of the 23rd European Symposium on

Computer Aided Process Engineering ESCAPE 23, June 9-12, 2013, Lappeenranta, Finland
2013 Elsevier B.V. All rights reserved. 115

A model based approach to an adaptive design

space in chromatography
Edward Closea,b, Daniel G Bracewellb and Eva Srensena
a.CPSE, Department of Chemical Engineering, University College London, Torrington Place,
London, WC1E 7JE, UK. e-mail: e.sorensen@ucl.ac.uk
b. ACBE, Department of Biochemical Engineering, University College London, Torrington Place,
London, WC1E 7JE, UK.

Abstract
We demonstrate an approach to construct an adaptive design space in the face of
process variability, for an industrial hydrophobic interaction chromatography (HIC)
with resin lot-to-lot variability. The step has a complex mixture of impurities in the feed
stream and a multi component product. The step must deliver a specific distribution of
product forms in the elution peak whilst maintaining product recovery and impurity
removal. In our approach, a mechanistic model is used which gives a good
representation of the system, and has been validated experimentally. The model is used
to quickly and efficiently explore the impact of process parameters on process
performance, utilizing stochastic simulation to generate probabilistic design spaces for
different resin lots. The results indicate that significant increases in process robustness
can be made by adapting the design space based on the resin lot in use, rather than
fixing the design space for all resin lots. An adaptive design space enables operation
further away from high risk regions, increases the size of operating regions and
improves flexibility to variations in process inputs.

Keywords: Process systems modelling, quality by design, QbD, chromatography.

1. Introduction
Despite several decades of effort to improve R&D efficiency and performance, the
procedure for bringing a new biopharmaceutical product to market remains an
extraordinarily expensive, time-consuming, and risky proposition (Lightfoot and
Moscariello, 2004). Chromatographic separations are the workhorse of therapeutic
protein purification (Kelley, 2007), but the design and development of these separations
is a challenging task both technologically and economically.
The US Food and Drugs Administration (FDA) are now encouraging
applications based on the Quality by Design (QbD) paradigm (US Food and Drug
Administration, 2006). Key to the QbD approach is a thorough understanding of process
inputs and their impact on performance, the relationship between the process and the
products critical quality attributes (CQA), and the association between the CQAs and
a products clinical properties (Jiang et al., 2010). The expected benefit of the approach
is an increase in the assurance of product quality, and in turn, the FDA will allow
manufacturers greater flexibility to operate with lower regulatory burden, enabling
continuous process improvement, as well as greater robustness.
Central to QbD is the concept of a design space, which is defined by the
International Conference on Harmonisation (ICH) guidance document as the
multidimensional combination and interaction of input variables and process
parameters that have been demonstrated to provide an assurance of product quality
116 E. Close et al.

(US Food and Drug Administration, 2006). Working within a regulatory approved
design space is not deemed to be a change from normal operating conditions, and this
therefore offers the attractive possibility that a chromatographic process can be
optimized for different process scenarios and objective functions whilst still remaining
within the regulatory approved design space.
Process characterization to map the impact of process parameters on process
performance and product CQAs is an essential part of this QbD strategy (Jiang et al.,
2010). A model based approach supported by appropriate experimentation has clear
advantages over a purely experimental approach to process characterization, as an
accurate mathematical model can enable quick and efficient investigation and screening
of different design and operating alternatives, with minimal experimental effort. A good
representation should guarantee the selection of the optimal operating region of the
design space, and enhances process understanding and knowledge in view of the QbD
framework.
This paper considers a commercial hydrophobic interaction chromatographic
(HIC) step with a complex industrial feed stream and a multi-component protein
product. The step must deliver a specific distribution of product forms in the elution
peak, i.e. not just a single component at a given concentration, whilst maintaining
product recovery and removal of impurities, regardless of variations in the feed stream
composition; which is a challenging set of objectives. The chromatography is
particularly challenging because the resin has been found to exhibit lot to-lot-variability
which significantly limits the achievable design space.
In this work, we demonstrate the novel use of a mechanistic model and
stochastic simulation methods to generate probabilistic design spaces for an industrial
chromatographic purification process. We then use the probabilistic design spaces to
explore the possibility of adapting the chromatography design space based on the resin
lot in use, rather than fixing the design space for all resin lots. We develop a model
which simulates the separation of the multicomponent protein product during the load,
wash and elution of the chromatographic cycle. The model is able to accurately predict
the distribution of product forms in the elution peak and the product recovery. We
combine the mechanistic model with stochastic simulation methods to generate
probabilistic design spaces which show the probability of meeting the process
objectives based on known variability in inlet variables. The demonstration illustrates
how models can be a key component in the successful implementation of a Quality by
Design (QbD) processing strategy.

2. Experimental Material and Methods

2.1 Hydrophobic Interaction Chromatography

Laboratory scale hydrophobic interaction chromatography experiments were conducted
to obtain validation data for the chromatographic model. Two different lots of Butyl
Sepharose 4B fast flow hydrophobic interaction resin were obtained from GE
Healthcare (Uppsala, Sweden). All preparative scale laboratory experiments were
carried out using an KTA FPLC chromatography system from GE Healthcare
(Uppsala, Sweden). Laboratory columns were 1.1 cm in diameter and 7.4 cm in bed
height. The columns were first equilibrated with equilibration buffer to the desired pH
and NaCl concentration. Isocratic experiments were then conducted. The elution peak
from a preceding capture chromatography unit was applied to the columns, followed by
a 10 column volume (CV) wash step using the equilibration buffer, before elution buffer
 A model based approach to an adaptive design space in chromatography 117

was applied and the product peak collected. The load material was fully representative
of a commercial process, and contained the typical impurity profile of a CHO
expression system. Fractions were collected every 0.5CV during the load and wash
phases and were subsequently analysed on an in house cation exchange HPLC assay
in order to determine the percentage of each component in each fraction

3. Mechanistic model development

In this work, we developed an equilibrium dispersive model coupled with a competitive
Langmuir isotherm (Guiochon et al., 1994), coded within gPROMS, an equation based
modelling and optimization software package (Process Systems Enterprise, 2012). Our
objective was to develop a model with sufficient accuracy in order to simulate the
process, meeting the aims and objectives, whilst minimising experimental and
computational effort. Following this philosophy, in the model we neglect the impurities
present in the load material and only consider the 3 product forms that make up the final
product. This is justified, and was confirmed experimentally, as the main contaminants
in the load material have a much weaker affinity with the resin than the product, and
therefore have minimal effect on the binding of these components. Parameter estimation
conducted within gPROMS was used to estimate the competitive Langmuir isotherm
adsorption parameters using data from experimental batch adsorption studies. The
apparent axial dispersion coefficient and total column porosity were calculated using the
retention time of an unretained small molecule and the number of theoretical plates of
the column, both determined from pulse injection experiments.

100 200
Resin A
75
UV Trace (mAU)

Predicted Experiment
Percentage

product product
50 100
distribution distribution
79:21 81:19
25

0 0
0 5000 10000 15000
Time (s)
100 200
Resin B
75 Predicted Experiment
UV Trace (mAU)

product product
Percentage

50 100 distribution distribution

91:9 90:10
25

0 0
0 5000 10000 15000
Time (s)

Figure 1. Model validation. Primary Y axis: Product form percentages. Lines: simulation predictions,
symbols: experiment results. Secondary Y axis: UV trace.
 118 E. Close et al.

We validated the model performance by comparing model predictions with

experimentally determined percentages of product forms in the elution peak, and in
samples taken every column volume (CV) during runs. This was an important and
industrially relevant test to determine the model accuracy, since achieving a specific
percentage distribution of the product isoforms in the elution pool is a key objective of
this particular chromatography. During all validation studies the flow rate, wash length,
bed height and column volumes were kept constant throughout. The component/product
distribution in the load material, total load concentration and load challenge were
varied. The values were chosen to give a good representation of the product
distributions experienced in normal load material, and the load challenges used during
day to day operation of the chromatography. An example of the good agreement
between model predictions and experimental results are shown in Figure 1.

4. Stochastic simulations
We used stochastic simulations to generate probabilistic design spaces for a given resin
lot A and lot B. The design spaces show the probability of meeting the chromatography
objective (a specific distribution of product forms in the elution peak) as a function of
manipulated variables (wash length and mass challenge) given the variability
experienced during normal process operation in the inlet variables (inlet concentration
of the multiple product forms). In the following we set out our methodology.

4.1 Accounting for variability in inlet variables

For each product form, instead of assigning a literal value to the inlet concentration in
the simulation, we assign a stochastic value, where the batch-to-batch variability is
accounted for by specifying the average inlet concentration and standard deviation
experienced during day-to-day operation of the HIC process. When simulating the HIC
using the mechanistic model, a built in function within gPROMS is used that returns a
random value sampled from a normal distribution generated using the specified average
and standard deviation. Each time a simulation is run, a different value is picked. An
example normal distribution of the inlet concentration of a product form is shown in
Figure 2A.

20 0.2
A B
Inlet concentration (mg/ml)

15

10 0.1
P

0 0.0
0.0 0.1 0.2 0 250 500 750 1000
Inlet concentration (mg/ml) Simulation number

Figure 2. A Normal distribution of inlet concentration of example product form. Mean: 0.1089. Standard
deviation: 0.0239. B. Example of randomly selected inlet concentrations during stochastic simulation.
 A model based approach to an adaptive design space in chromatography 119

Product form percentage in elution peak

50
A B
0.10

25

P
0.05

0 0.00
0 250 500 750 1000 0 25 50 75 100
Simulation number Percentage
Figure 3. A. Product form percentage in elution peak over 1000 simulations. B. Probability density function
of product form percentage in peak for mass challenge: 2 mg/ml, wash length 5 column volumes. Highlighted
area: where HIC objective is met, Probability = 0.58.

4.2 Conducting multiple simulations at unique design space points

At each unique mass challenge and wash length combination, we conducted multiple
simulations (>5000) in order to determine how variability in the inlet concentration
affects the distribution of product forms in the elution peak. In each individual
simulation the inlet concentration was randomly assigned as described previously
(section 4.1), as illustrated in Figure 2B. The resulting product distribution in the elution
peak is recorded for each individual simulation (Figure 3A), and used to generate a
probability density function (Figure 3B). The total area under the curve is equal to one,
therefore, the area under the curve where the process objective is met is the probability
that the objective will be met at that mass challenge and wash length, i.e. the probability
that the product distribution in the elution peak is acceptable.

4.2 Combining results from multiple design space points

In order to generate the probabilistic design space, we conducted multiple simulations
as described previously (section 4.2), at different mass challenge and wash length
combinations. For each unique combination we generated the probability density
function, and calculated the probability that the process objective will be met. By
combining this data in one graph, we were able to generate the probabilistic design
space. The probabilistic design spaces for resin lots A and B are shown in Figure 4 and
5.
1.00 1.00
A B
0.75 0.75

0.50 0.50
P

Mass Challenge Mass challenge

0.25 1.5 mg/ml
2 mg/ml
0.25 1.5 mg/ml
2 mg/ml
2.5 mg/ml 2.5 mg/ml
3 mg/ml 3 mg/ml
0.00 0.00
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Wash length (Column Volumes) Wash length (Column Volumes)

Figure 4. Probability of meeting HIC objecives as a function of washlenth and mass challenge. A. Resin lot A
with high binding affinity B. Resin lot B with low binding affinity.
 120 E. Close et al.

3.0 3.0
A B
Mass challenge (mg/ml)

Mass challenge (mg/ml)

2.5 .40 0.60 2.5 0.80 0.60 0.40

2.0 0.80 2.0 0.80

1.5 1.5
1 2 3 4 5 1 2 3 4 5
Wash Length (Column Volumes) Wash Length (Column Volumes)
Figure 5. Probabilistic design spaces showing probability that HIC objectives will be me as a fuinction of
washlength and mass challenge. A. Resin lot A with high binding affinity for product. B. Resin lot B with low
binding affinit for product.

5. Results and discussion

Two distinct approaches can be taken to account for resin lot variability. In the first
approach, the design space must be fixed for all resin lots. If we assume that the design
space is any combination of mass challenge and wash length that will meet the process
objectives 80 percent of the time, then the achievable design space is where both the
low and high resin lot probability is above 0.8. In the second, where the design space is
adapted depending on which resin lot is in use, i.e. an adaptive design space, the
achievable design space for each resin lot is where the individual resins probability of
meeting the process objective is above 0.8. The probabilistic design spaces generated in
this work indicate that significant increases in process robustness can be made by
adapting the design space based on the resin lot in use, rather than fixing the design
space for all resin lots. An adaptive design space enables operation further away from
high risk regions, increases the size of operating regions and improves flexibility to
variations in process inputs.

6. Concluding remarks
We have proposed an approach using mechanistic chromatography models combined
with stochastic simulation which can quickly and efficiently investigate different design
and operating alternatives to find an optimum design space with minimum impact of
process related variability. The approach was demonstrated using an industrial HIC step
with a complex feed and product distribution. The approach enables significant savings
in terms of time and cost when compared with a traditional, purely experimental,
approach.

References
C. Jiang, L. Flansburg, S. Ghose, P. Jorjorian, and A. A Shukla, 2010. Biotech Bioeng, 107, 985.
G. Guiochon, S. Goldshan-Shirazi, A. Katti. 1994. Fundamentals of preperative and non-linear
chromatography. Boston:Academic Press.
B. Kelley. Biotechnol Prog. 2007. 23: 995-1008.
E. Lightfoot, J. Moscariello. 2004. Biotech Bioeng. 87: 259-273.
A. Rathore. 2009. Trends Biotechnol, 27(9), 546553.
US Food and Drug Administration (FDA). 2006. Q8 pharmaceutical development.
Process Systems Enterprise, 2012, gPROMS, www.psenterprise.com/gproms.