Extreme Scale-Down Approaches for Rapid Chromatography Column Design and Scale-Up During Bioprocess Development - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 131

is not necessarily suitable in every situation, and other design-of-experiments
methodologies may be more suitable in those cases. The simplex method is
essentially suited to early-stage process development when trying to identify a
feasible operating point that merits further characterisation in later studies. Sub-
sequent work (e.g. upon scale-up in later development) is perhaps best conducted
using more conventional factorial design models such as two-level factorials or
response surface methodologies to provide the type of robust design space char-
acterisation that ends up supporting a regulatory filing. Table 1 summarises the
circumstances in which one would use the simplex method versus an alternative.
5.2.2 Genetic Algorithm-Based Methods
There are other ways of structuring an experimental campaign, such as genetic
algorithms, which are suitable for complex problems involving numerous variables
and which are robust in the presence of experimental noise. Susanto et al. [ 30 ]
proposed a way of performing chromatographic process optimisation at very small
scales in a resource- and time-efficient manner by using closed-loop (i.e. no
manual intervention) and semi-closed-loop strategies that integrated HTS robotics
together with genetic algorithms. The procedure carries out process exploration,
data analysis and optimization iteratively and was illustrated through two IEX case
studies. The first was a closed-loop case looking at the capture of lysozyme from a
solution also containing cytochrome and focussed upon how varying the equili-
bration buffer pH and NaCl concentration affected capacity and selectivity. The
study was conducted in a robotically controlled batch plate to enable prediction of
a 1-mL column. Resin slurries were aliquoted using the Atoll ResiQuot device to
provide a 7.7 lL settled resin volume per well. Closed-loop optimisation was used
to drive the search, and in spite of analytical noise, successive iterations enabled
the search to converge to a global optimum within 4 h (96 experiments). After this,
optimal and sub-optimal conditions identified by HTS were verified in a 1-mL bed.
The second study involved an interesting approach to predict the optimal elu-
tion gradient shape. The method involved deconvoluting the A280 'total' protein
absorbance signal into individual component peaks, alongside the use of mass
balance and prior user knowledge, to distinguish the position and size of individual
peaks for a three-protein mixture and so assess peak resolution. The study used
robotically operated 200-lL miniature columns to look at the optimisation of the
shape of a multilinear elution gradient consisting of two different linear slopes in
order to maximise peak resolution for a mixture containing ribonuclease, cyto-
chrome and lysozyme. Eight columns were run simultaneously using different
elution conditions, and a 96-well UV-transparent plate was used to collect frac-
tions by moving the plate underneath the outlet of the column row. The two
successive linear slopes were converted into a succession of discrete step gradients
that incremented the eluate salt concentration appropriately. The elution gradient
was optimized by using a genetic algorithm, but instead of using a fully automatic
closed loop, this second example involved a semi-closed case that demonstrated