AN IMPROVED PROBING CONTROLLER FOR

SUBSTRATE FEEDING IN FED-BATCH CULTURES OF

E. COLI: SIMULATIONS AND EXPERIMENTS
M. kesson
Biotecnol

P. Hagander

J. P. Axelsson

SA, Taguspark, Oeiras, Portugal

Department of Automatic Control,

Lund Institute of Technology, Lund, Sweden

Pharmacia Corp., Global Supply, Strngns, Sweden

Abstract: A strategy for feedback control of the glucose feed rate in cultivations
of E. coli is discussed. By making probing pulses in the feed rate it is possible
to detect and avoid undesirable by-product formation using a standard dissolved
oxygen sensor. The feasibility of an improved and simplified probing algorithm is
demonstrated by simulations as well as laboratory-scale experiments.
Keywords: Fermentation process, process control, extremum control, probing.

1. INTRODUCTION

Feed rate

Oxygen transfer

Escherichia coli is still one of the most important host organisms for production of recombinant proteins. Considerable efforts have been
made to improve the efficiency and to extend
the application range of E. coli-based expression
systems. One of the obstacles in attaining high
product yields and high productivity is the accumulation of the metabolic by-product acetate,
which inhibits growth [Luli and Strohl, 1990]
as well as production of recombinant protein
[Bauer et al., 1990; Bech Jensen and Carlsen,
1990; Turner et al., 1994].
Formation of acetate in E. coli cultures occurs
under anaerobic conditions but also under fully
aerobic conditions in situations with excess carbon source. These two mechanisms are often referred to as mixed-acid fermentation and overflow metabolism, respectively. The accumulation of acetate can be reduced by manipulation
of strains, media, and cultivation conditions,
and much research has been devoted to this
topic, see for instance [Lee, 1996; Riesenberg
and Guthke, 1999].

Overflow
metabolism

Overflow
metabolism

F(t)
Starvation
Time

Fig. 1. Restrictions on the carbon source feed

rate F in a fed-batch cultivation.
In fed-batch cultures, which often is the preferred mode of operation, the feed rate of the
carbon source, usually glucose, can be manipulated to restrict formation of acetate. Typical
constraints on the feed rate are illustrated in
Figure 1. In the early stage of a cultivation, the
cell density is rather low and the glucose feed
rate must remain low to avoid overfeeding and
acetate production from overflow metabolism.
As the cells grow, the feed rate can be increased
without causing production of acetate. However,
this implies an increased oxygen consumption

and eventually the maximum oxygen transfer

capacity of the reactor may be reached. This
will impose another limitation on the feed rate,
since it will lead to anaerobic conditions if the
resulting oxygen consumption exceeds the maximum oxygen transfer capacity. Finally, production of a recombinant protein may cause
metabolic changes so that overflow metabolism
again becomes a limiting factor [Qiu et al.,
1998]. When the above constraints are avoided,
acetate present in the media can also be reassimilated with a concomitant consumption of
oxygen, see for instance [Xu et al., 1999].
Consequently, by keeping the feed rate sufficiently low, it is possible to avoid accumulation
of acetate. On the other hand, choosing the feed
rate unnecessarily low will give a low growth
rate and hence a long cultivation time and low
productivity. A challenge is thus to keep a high
feed rate while avoiding overflow metabolism
and anaerobic conditions.
A number of feeding strategies have been developed to reduce or avoid acetate formation
[Lee, 1996; Riesenberg and Guthke, 1999]. Typically these are designed to avoid either overflow metabolism or anaerobic conditions and
one then has to change between different strategies during the cultivation. In some cases,
the implementation is complicated due to the
lack of cheap and reliable on-line sensors and
many methods also require considerable process
knowledge to work well. Often a key process parameter, such as the critical growth rate above
which acetate formation occurs, has to be known
a priori. This may be a disadvantage, especially
for processes for production of recombinant proteins where significant changes in the process
behavior often occur [Qiu et al., 1998; kesson
et al., 1999b; DeLisa et al., 1999].
In [kesson, 1999; kesson et al., 1999a] we
presented a probing feeding strategy that addresses the constraints from overflow metabolism
as well as oxygen transfer. Major advantages
are that it is implemented using standard sensors and that it only requires a minimum of process knowledge. Laboratory-scale experiments
under various operating conditions have shown
that the method reproducibly works well. In
[kesson and Hagander, 2000] we presented
simulations of a simplified and improved algorithm for the avoidance of overflow metabolism.
These simulations are now complemented with
experiments at two different laboratories.
2. PROCESS MODEL
We consider aerobic fed-batch processes where
E. coli bacteria are cultivated in a stirred biore-

qo

qmax
o

qa
qgcrit

qg

qac,max
Fig. 2. Relations between specific rates of glucose uptake, qg , oxygen uptake, qo , and net
acetate production, qa . Dashed lines show
potential acetate consumption and its effect
on qo when acetate is present in the media.
actor. Control loops for temperature, pH, and
dissolved oxygen ensure that suitable operating
conditions are maintained. We will now briefly
recapitulate a model used for the simulations of
the feeding strategy. For details on parameters
and metabolic expressions see [kesson, 1999].
2.1 Metabolic Relations
The cell metabolism is described by the specific
rates of growth , glucose uptake qg , oxygen
uptake qo , and net production of acetate qa . The
metabolic expressions used are largely similar
to the ones presented in [Xu et al., 1999]. Glucose provides energy and raw material for cell
growth. The specific glucose uptake, qg , is taken
to be of Monod type
qg ( G ) = qgmax G /( ks + G )
which describes a smoothly saturating glucose
uptake. Oxygen is used to metabolize the glucose, and the specific oxygen uptake qo depends
on qg .
Formation of acetate under aerobic conditions
typically occurs when qg exceeds a critical value
qgcrit . It has also been observed that qo reaches
an apparent maximum qmax
at the onset of
o
acetate formation [Andersen and von Meyenburg, 1980; Paalme et al., 1997] as illustrated
by the solid lines in Figure 2. Acetate present
in the medium may also be consumed and used
for growth. Its uptake mechanism is likewise
modeled to follow Monod kinetics
qca, pot = qac,max A/( ka + A)
However, as the consumption requires oxygen
and glucose is the preferred substrate, it is
limited both by the available oxidative capacity
and the uptake mechanism. The dashed lines
in Figure 2 outlines the maximum potential
acetate consumption and the resulting effect on
qo . The specific growth rate increases with the
glucose uptake but with a decreased yield above
qgcrit . Consumption of acetate also contributes to
the cell growth.

2.2 Bioreactor Model

Glucose
dynamics

qg

qo

Oxygen
dynamics

Op

Component-wise mass balances for the bioreactor give the following equations
dV
dt
d( V X )
dt
d( V G )
dt
d( V A)
dt
d( V Co)
dt

=F

Probing
controller

= ( G , A) V X
= FGin qg ( G ) V X
= qa ( G , A) V X


= K L a( N ) V Co Co qo ( G , A) V X

where V , X , G , A, Co are the liquid volume,

the cell concentration, the glucose concentration, the acetate concentration, and the dissolved oxygen concentration, respectively. Further, F , Gin , Co denote the feed rate, the glucose concentration in the feed, and the dissolved
oxygen concentration in equilibrium with the
oxygen in gas bubbles. The volumetric oxygen
transfer coefficient, K L a, increases with the
stirrer speed N but is also affected by factors
like viscosity, foaming, and anti-foam chemicals.
When the measurement of dissolved oxygen is
used for control, it is also important to consider
the dynamics in the dissolved oxygen probe. In
practice, most sensors measure the dissolved
oxygen tension which is related to the dissolved
oxygen concentration through Henrys law O =
H Co . The probe is here modeled as a first-order
system
dO p
Tp
+ Op = O
dt

3. A PROBING CONTROL STRATEGY

The key idea of the probing feeding controller
is to exploit the characteristic saturation in
the respiration that occurs when qg exceeds
qgcrit and acetate formation starts. The saturation can be detected by superimposing probing
pulses in F that are long enough to be seen
through the system dynamics. As long as qg is
below qgcrit , the pulses give rise to responses in
the dissolved oxygen signal O p. However, when
qg is above qgcrit , qo is saturated and no response
is seen. In this way it is possible determine if
qg is above or below qgcrit without knowledge of
the actual value [kesson et al., 1999b]. This
information is then used to adjust F to achieve
feeding at qgcrit , see Figure 3. The probing controller can be viewed as an extremum controller,
see [strm and Wittenmark, 1995], controlling
the process to the start of a saturation.

Fig. 3. The probing control strategy. By making

probing pulses in F it is possible to determine if qg is above or below qgcrit from the
response in O p.
3.1 An Improved Feedback Algorithm
The previous implementation of the algorithm
[kesson, 1999; kesson et al., 1999a] made
use of both up and down pulses in a switching scheme depending on the pulse responses.
In [kesson and Hagander, 2000], the down
pulses were omitted while maintaining, and
even improving, performance in simulation experiments. The simplification also makes implementation and analysis easier.
At each cycle of the algorithm, a pulse is given
to get information. When qg < qgcrit the process model suggests a region where the amplitude of the pulse responses y are proportional
to the distance qgcrit qg . Within this proportional band the feed rate is adjusted using a
proportional-integral (PI) algorithm

F ( k) = ( y( k) yr ) +

+ i

k
X

( y( j ) yr )

j =0

F
O Osp

The set-point yr enhances convergence and

makes it possible to keep qg slightly below qgcrit .
If the amplitude of the pulse response y is
judged too small, that is below a threshold level
Oreac , it is concluded that qg exceeds qgcrit and
F is decreased with a fixed amount chosen to
be equal to the pulse size Fpulse. An example
is shown in Figure 4 where also some of the
algorithm parameters are defined.

3.2 Safety Nets and Tuning Rules

To ensure that O p starts from the same level,
Osp, at each pulse, it is regulated by manipulation of the stirrer speed. During the feed pulses,
however, the stirrer speed is fixed in order not
to interfere with the detection algorithm. If disturbances cause O p to deviate from the set-point
at the time for a pulse, a safety net delays the
pulse to avoid erroneous interpretations of the
pulse response.

qg [g/(gh)]

Fpulse

A [g/l]

F (k)

0.5

0
0

Tpulse

Tcontrol

1.5

2.5

3.5

0.5

1.5

2.5

3.5

0.5

1.5

2.5

3.5

0.5

1.5

2.5

3.5

40

O p [ %]

35
30
25

Osp

Op

0.5

20
0

Oreac

0.06

Fig. 4. Example of a cycle in the algorithm. During the time Tcontrol between two pulses, O p
is regulated by a controller manipulating
the stirrer speed.
Derivation of tuning rules and a stability
analysis for the algorithm were performed in
[kesson, 1999; kesson and Hagander, 2000]
based on a linearized process model. The only
required information is an upper bound for the
lumped time constant Tmax in the relation between F and O p together with the noise level
in O p. These quantities are mainly equipment
related and do not depend critically on a particular strain or product. The former reference also
describes how to avoid anaerobic conditions due
to limitations in the reactor oxygen transfer.

3.3 Simulation Example

A simulation of the improved algorithm is
shown in Figure 5 where the initial feed rate
after a batch phase gives a specific glucose uptake qg below qgcrit . The controller increases F
until qgcrit is reached and qg is then kept approximately constant as desired. After 2 hours
the value of qgcrit is decreased, and as can be
seen the algorithm is able to detect and adjust
the feed accordingly. The undershoot is due to
reconsumption of the acetate produced when
qg temporarily exceeds qgcrit . The accompanying oxygen consumption gives qo = qmax
even
o
though qg < qgcrit and hence F is decreased as
no oxygen responses can be seen. This example
illustrates the ability to achieve feeding at qgcrit
without a priori information and in spite of
changes that may occur in the process.

4. EXPERIMENTS
The experiment shown in Figure 6 was performed using a probing controller of PI-type.
After the feed start, the controller increases the
feed rate in an almost exponential fashion, as
expected. Recombinant protein production is induced at 6.7 h, and shortly after the feed rate is

F [l/h]

0.04
0.02
0
0
1200
1000

N [rpm]

800
600
400
0

Time [h]

Fig. 5. Simulation with PI controller ( =

0.96, i = 0.64, yr = 3) where qgcrit is
changed after two hours. The dotted lines
indicate the critical glucose uptake, qgcrit ,
and the reaction level Oreac . The probing
controller achieves feeding at qgcrit without
prior information of the value and in spite
of the change.
decreased when the amplitude of the dissolved
oxygen responses are reduced. The off-line analysis shows that concentrations of glucose and
acetate are kept low throughout the cultivation,
in spite of disturbances from antifoam additions
around 8.6 and 9.1 h. This experiment shows
the feasibility of the probing technique and also
illustrates how qgcrit can change significantly after induction of recombinant protein production.
Figure 7 shows an experiment with another
E. coli-based process, now using a probing controller of proportional type. This experiment
was performed at very low cell densities. As a
consequence the oxygen consumption is lower
and therefore the setpoint in dissolved oxygen
had to be at a higher level. Again, the feed rate
is increased rapidly until recombinant protein
production is induced after 19.3 h, whereafter it
decreases slowly after having touched the maximum pump speed. Unfortunately, no acetate
measurements are available but glucose test
sticks with a detection limit of 0.5 g/l showed
that no glucose accumulation occurred.
A comparison with a cultivation run in parallel
under identical conditions using a exponential
feeding profile is shown in Figure 8. Interestingly, the cultivation with the fixed feeding pro-

0.06

F [l/h]

100

0.04

F [ %]

50

0.02
0

15

16

17

18

19

20

21

22

23

21

22

23

40

Induction

OD

30

20

10

O p [ %]

5
0

15

16

17

18

19

20

Time [h]
20

X [g/l] 10

Fig. 8. Comparison between the probing feeding strategy (solid line and circles) and a
fixed exponential feeding profile (dashed
line and crosses). At the last two samples
glucose accumulated for the fixed feeding
profile even though overall growth rate was
lower than for probing feeding strategy.

Induction

2G

[g/l]

0.5

A
0

Time [h]

Fig. 6. Fed-batch part of an experiment using a

probing controller with PI adjustments.
100

F [ %]

50

15

16

17

18

19

20

21

22

23

15

16

17

18

19

20

21

22

23

21

22

23

60

O p [ %]

50

40

Induction

OD

10
5
0

15

16

17

18

19

20

Time [h]

Fig. 7. Fed-batch part of an experiment with

a probing controller using proportional adjustments.
file showed glucose accumulation in the two
samples after 22 h even though the overall
growth rate was slower. This clearly shows the
advantage of using a feeding strategy based on
feedback control.

5. CONCLUSIONS
In cultivations of E. coli it is important to avoid
accumulation of the by-product acetate, which
is produced when the specific glucose uptake qg

exceeds a critical value qgcrit . We have discussed

an approach for feedback control of glucose feeding based on a superimposed probing signal
in the feed rate. It can be implemented with
a regular dissolved oxygen sensor and enables
control of qg at or below qgcrit without prior
knowledge and in spite of changes in the process. The feasibility of an improved and simplified probing algorithm was demonstrated by
simulations as well as laboratory experiments.
The key idea in the probing algorithm is
to exploit a characteristic saturation in the
metabolism of E. coli. By-product formation
linked to a limitation in the metabolism is, however, a characteristic E. coli shares with many
other organisms. A natural step is therefore to,
where appropriate, investigate the applicability
of the probing technique also in these cases.
Preliminary experiments showing good results
have been performed with Saccharomyces cerevisiae [Fredriksson, 2000] and further investigations are currently made with other microbial hosts. Another interesting extension is to
perform experiments in a larger scale.
Acknowledgments
This work was financially supported by Pharmacia, the Swedish National Board for Industrial and Technical Development (project 1N1197-09517), and the Alf kerman foundation.
The opportunity to perform experiments at
Pharmacia, Stockholm, Sweden, and the Department of Biotechnology, Lund University,
Lund, Sweden, is gratefully acknowledged. In
particular, the authors would like to thank Finn
Duns, Maher Abou Hachem, Olle Holst, and
Veronica Martins for assistance and valuable
discussions.

6. REFERENCES
kesson, M. (1999): Probing control of glucose feeding in Escherichia coli cultivations.
PhD Thesis ISRN LUTFD2/TFRT--1057-SE. Department of Automatic Control, Lund
Institute of Technology, Sweden.
kesson, M. and P. Hagander (2000): A simplified probing controller for glucose feeding
in Escherichia coli cultivations. In Proceedings of the 39th IEEE Conference on Decision and Control, Dec 12-15, 2000, Sydney,
Australia., vol. 1, pp. 45204525.
kesson, M., P. Hagander, and J. P. Axelsson
(1999a): A probing feeding strategy for Escherichia coli cultures. Biotechnology Techniques, 13, pp. 523528.
kesson, M., E. Nordberg Karlsson, P. Hagander, J. P. Axelsson, and A. Tocaj (1999b):
On-line detection of acetate formation in Escherichia coli cultures using dissolved oxygen responses to feed transients. Biotechnology and Bioengineering, 64, pp. 590598.
Andersen, K. and K. von Meyenburg (1980):
Are growth rates of Escherichia coli in
batch cultures limited by respiration? Journal of Bacteriology, 144, pp. 114123.
strm, K. and B. Wittenmark (1995): Adaptive
Control, 2nd edition. Addison-Wesley, Reading, MA.
Bauer, K., A. Ben-Bassat, M. Dawson, V. de la
Puente, and J. Neway (1990): Improved expression of human interleukin-2 in high-celldensity fermentor cultures of Escherichia
coli K-12 by a phosphotransacetylase mutant. Applied and Environmental Microbiology, 56, pp. 12961302.
Bech Jensen, E. and S. Carlsen (1990): Production of recombinant human growth hormone
in Escherichia coli: expression of different
precursors and physiological effects of glucose, acetate and salts. Biotechnology and
Bioengineering, 36, pp. 111.
DeLisa, M., J. Li, G. Rao, W. Weigand, and
W. Bentley (1999): Monitoring GFP-operon
fusion protein expression durig high cell
density cultivation of Escherichia coli using
an on-line optical sensor. Biotechnology and
Bioengineering, 65, pp. 5464.
Fredriksson, J. (2000): Probing control of
glucose feeding in cultivation of Saccharomyces cerevisiae. Master thesis ISRN
LUTFD2/TFRT--5660--SE. Department of
Automatic Control, Lund Institute of Technology, Sweden.
Lee, S. Y. (1996): High cell-density culture of
Escherichia coli. Trends in Biotechnology,
14, pp. 98105.
Luli, G. W. and W. R. Strohl (1990): Comparison of growth, acetate production, and

acetate inhibition of Escherichia coli strains

in batch and fed-batch fermentations. Applied and Environmental Microbiology, 56,
pp. 10041011.
Paalme, T., R. Elken, A. Kahru, K. Vanatalu,
and R. Vilu (1997): The growth rate control
in Escherichia coli at near to maximum
growth rates: the A-stat approach. Antonie
van Leeuwenhook, 71, pp. 217230.
Qiu, J., J. R. Swartz, and G. Georgiou (1998):
Expression of active human tissue-type
plasminogen activator in Escherichia coli.
Applied and Environmental Microbiology,
64, pp. 48914896.
Riesenberg, D. and R. Guthke (1999): Highcell-density cultivation of microorganisms.
Applied Microbiology and Biotechnology, 51,
pp. 422430.
Turner, C., M. Gregory, and M. Turner (1994):
A study of the effect of specific growth
rate and acetate on recombinant protein
production of E. coli JM 107. Biotechnology
Letters, 16, pp. 891896.
Xu, B., M. Jahic, and S.-O. Enfors (1999):
Modeling of overflow metabolism in batch
and fed-batch cultures of Escherichia coli.
Biotechnology Progress, 15, pp. 8190.