Anda di halaman 1dari 10

Application of Modeling and Simulation Tools for the Evaluation of

Biocatalytic Processes: A Future Perspective


Gurkan Sin
Dept. of Chemical and Biochemical Engineering, Computer Aided Process-Product Engineering Center (CAPEC),
Technical University of Denmark, Lyngby, Denmark
John M. Woodley and Krist V. Gernaey
Dept. of Chemical and Biochemical Engineering, Center for BioProcess Engineering, Technical University of Denmark,
Lyngby, Denmark

DOI 10.1002/btpr.276
Published online September 18, 2009 in Wiley InterScience (www.interscience.wiley.com).

Modeling and simulation techniques have for some time been an important feature of bio-
catalysis research, often applied as a complement to experimental studies. In this short
review, we report on the state-of-the-art process and kinetic modeling for biocatalysis with
the aim of identifying future research needs. We have particularly focused on four aspects of
modeling: (i) the model purpose, (ii) the process model boundary, (iii) the model structure,
and (iv) the model identication procedure. First, one nds that most of the existing models
describe biocatalyst behavior in terms of enzyme selectivity, mechanism, and reaction
kinetics. More recently, work has focused on extending these models to obtain process ow-
sheet descriptions. Second, biocatalysis models remain at a relatively low level of complexity
compared with the trends observed in other engineering disciplines. Hence, there is certainly
room for additional development, i.e., detailed mixing and hydrodynamics, more process
units (e.g., biorenery). Third, biocatalysis models have been only partially subjected to for-
mal statistical analysis. In particular, uncertainty analysis is needed to ascertain reliability
of the predictions of the process model, which is necessary to make sound engineering deci-
sions (e.g., the optimal process owsheet, control strategy, etc). In summary, for modeling
studies to be more mature and successful, one needs to introduce Good Modeling Practice
and that asks for (i) a standardized and systematic guideline for model development, (ii) for-
mal identiability analysis, and (iii) uncertainty analysis. This will advance the utility of
models in biocatalysis for more rigorous application within process design, optimization,
and control strategy evaluation. V C 2009 American Institute of Chemical Engineers Biotech-

nol. Prog., 25: 15291538, 2009


Keywords: biocatalysis, kinetic model, modeling, pharmaceuticals, process development,
uncertainty analysis

Introduction of many processes.3 More recently, we have seen biocataly-


sis starting to emerge in the synthesis of lower value (bulk)
Biocatalysis, involving the use of one or more enzymes chemicals driven by the need to improve environmental
(in a soluble, whole-cell, or immobilized format) to catalyze compatibility of conventional synthetic processes, avoid the
chemical reactions has found signicant application in the use of hazardous reagents and increasingly develop synthetic
pharmaceutical sector for the synthesis of optically pure mul- routes based on sustainable (i.e., renewable) nonpetrochemi-
tichiral-centered molecules.1 Indeed, the application of bio- cal feedstocks. However, not all traditional chemical synthe-
catalysis in this sector looks set to increase as the ses are suitable candidates for implementing biocatalysis due
complexity of therapeutic molecules increases. The use of to the complexities of process integration and an unfavorable
biocatalysis enables conversion of multifunctional molecules cost/benet analysis. Assessing those cases which are most
under very mild conditions avoiding the need for multiple suitable is therefore a signicant development challenge.
protection and deprotection steps.2 In addition, the use of One set of tools to assist in this challenge is the use of mod-
mild reaction conditions, usually in an aqueous medium, also eling and simulation techniques where alternative routes and
enables a signicant improvement in the green credentials hypothetical changes to existing processes or new process
concepts can be explored in silico, before experimentation.
This enables expensive experimentation to be directed and
Additional Supporting Information may be found in the online
version of this article.
focused, thereby speeding up bioprocess development. Last
Correspondence concerning this article should be addressed to G. Sin but not least, the process models can also be used to develop
at gsi@kt.dtu.dk. and evaluate process control strategies to ensure stability and

C 2009 American Institute of Chemical Engineers


V 1529
1530 Biotechnol. Prog., 2009, Vol. 25, No. 6

mately leading to the process model, which encapsulates all


the other models and more (including unit operations other
than biocatalysis in a typical process owsheet). While natu-
rally the complexity of the model increases with the increas-
ing boundaries of the system, with proper assumptions (e.g.,
lumped vs. intrinsic kinetics) one can also have simplied
models for larger scopes. For example, a Michaelis-Menten
type kinetic reaction model is much less complex than a mo-
lecular-mechanics model of a catalyst (see below). These dif-
ferent model types will be discussed in detail in the
following section.

Catalyst models
The objective of devising catalyst models is to understand
Figure 1. A schematic illustration of different models applied the molecular structure of the enzyme and use this as a basis
in biocatalysis given in the order of increasing scope:
(i) intrinsic properties of biocatalyst (enzyme), (ii) to predict selectivity. Three studies have provided a review
reaction kinetics, (iii) reactor kinetics, and (iv) pro- of developments in the catalyst modeling eld.810 Kazlaus-
cess dynamics. Diagram adapted from Smith,7 kas9 classied the models according to the specic purpose
Chemical Process Design and Integration, 2005, for which they were built:
V
C Wiley.
Explanation of behavior of a known enzymesubstrate
interaction at a molecular level.
desired efciency. The development of suitable process con- Changing selectivity of an enzyme-catalyzed reaction by
trol strategies to guarantee a consistent end-product quality modication of the substrate, enzyme, or reaction conditions.
is especially relevant for the pharmaceutical sector in view Prediction of the degree of stereo-selectivity of an
of the Process Analytical Technology guideline published by enzyme-catalyzed reaction.
the FDA.4 Such objectives were reported in industrial scale
experiences as early as 1995.5,6 Now with more than a dec- To achieve these objectives, molecular models were built
ade later, in this short review, we set out to outline and based on theoretical knowledge to simulate enzymesubstrate
report the progress and challenges in process and kinetic interactions at a molecular level, e.g., calculation of free
modeling of biocatalysis. Finally, we aim to identify areas energy of enzymesubstrate interactions.810 These models
where signicant research is needed to further advance mod- are predictive and valuable. Nevertheless, the underlying
eling and its application in the eld of biocatalysis. mathematics is complex, and therefore, the computational
requirements to solve the model are signicant.10 The end
application of these models is to translate the molecular level
Application of Modeling to Biocatalytic Processes understanding into designing improved enzymes (biocata-
lysts). In this regard, the use of molecular modeling is
Mathematical modeling problems can be classied accord- viewed as a complement to screening methods and searching
ing to a number of criteria such as mathematical structure for enzymes by evolutionary approaches.11,12
(deterministic or empirical), complexity (simple or complex), In addition to the earlier classication of the models,
spatial resolution (nano-, micro-, or macro-), temporal reso- Braiuca et al.10 added chemometrics to the list of computa-
lution (steady-state or dynamic), or the intended purpose of tional methods in biocatalysis. Their review focused on the
the application (design, analysis, control, or optimization). synergistic use of molecular modeling with chemometric
For the purpose of this study, the models developed in the tools such as Principal Component Analysis (PCA) and Par-
eld of biocatalysis were classied according to their scope tial Least Squares (PLS) as well as statistical experimental
(or boundary conditions). Thus, models have been developed design by Response Surface Methodology (RSM) to optimize
in the following four areas: reaction yields. An earlier review of Hult and Holmquist8
Catalyst: Models describing catalysis at a molecular was unique in the sense that it focused rather on the method-
level leading in particular to the prediction of reaction ologies and experimental procedures followed to build and
selectivity. apply kinetic and molecular models for addressing certain
Reaction: Kinetic models describing mechanism and questions of microbial lipases, and therefore provided the
intrinsic rates based on given inputs such as concentration, underlying procedures and expertise required to set-up a ki-
pH, and temperature. netic and/or molecular model for a specic purpose (e.g.,
Reactor: Kinetic models describing observed reaction studying enzyme afnity vs. substrate binding).
kinetics in reactors incorporating mass balances as well as
hydrodynamic conditions (mixing, transport, etc).
Process: Models for the analysis of the overall perform- Reaction, reactor, and process models
ance of the process including the interactions between the The reaction, reactor, and process models all have as their
different unit operations involved in a process owsheet, objective either optimization, design or (process) analysis
e.g., enzyme reaction, crystallization, separation, granulation, and therefore prove very useful tools for engineers develop-
etc. ing, dening, and scaling-up new bioprocesses. Whilst cata-
The interaction between these models is illustrated in lyst models are concerned with the nature of the enzyme, the
Figure 1. The gure emphasizes that each model has a clear reaction (and reactor) models are mostly concerned with
boundary and focal point. Progressing from the molecular describing the observed rate of biocatalysis typically
level, each successive model builds on the previous one, ulti- obtained from initial rate type substrate-enzyme experiments.
Biotechnol. Prog., 2009, Vol. 25, No. 6 1531

Process models use the reactor model to describe the reac- (such as the one presented in Figure 2) contributes to a
tion units and couple these models with other submodels higher transparency of the model building process. More-
describing various other process units in the process ow over, it helps to move model building from an ad hoc basis
scheme such as crystallization and separation. The integrated to a more procedural method. Finally, such standardized pro-
process models ultimately aim to evaluate the interaction cedures can facilitate the assessment of the resulting model
between different process units to realistically estimate the quality and promote the transfer of knowledge regarding the
overall performance, e.g., product yield. In addition, in some model study and its results between peers in the biocatalysis
cases, it may prove advantageous to combine unit operations eld. To this end, we suggest that the modeling community
into so-called hybrid operations to enable for example in situ develops commonly agreed modeling guidelines.
product removal.13,14 In such cases, a cost-benet analysis
must be performed since under some conditions some met-
Models and their purposes
rics will be improved at the expense of others. Evaluating
such trade-offs and compromises is an important role for With respect to modeling objectives, the following could
process modeling. be distinguished from the model examples given in Table 1:
(i) kinetic analysis, in which the model (built on a reaction
mechanism, e.g., ping-pong bi-bi mechanism), is used to
Review of Process/Kinetic Models in Biocatalysis identify the rate/kinetics of the biocatalyst in question, and
(ii) process analysis, in which a validated model is used to
To structure the review, we have deliberately focused on
evaluate several process owsheets and congurations for a
the following criteria: (i) the purpose of the model, (ii) the
given biocatalyst by performing model simulations. This type
modeled process and the units considered (e.g., whole pro-
of model usage is actually predictive, i.e., the model could be
cess or a single process unit), (iii) the model structure
used outside of its calibration domain, i.e., to predict per-
(kinetics), and (iv) the model identication procedure (e.g.,
formance under different mixing/ow regimes, e.g., continu-
experimental data, parameter estimation, parameter estima-
ous vs. batch, and (iii) design, in which the model is used to
tion error (condence interval), and uncertainty of model
dimension process units for a given yield and production
predictions). Having these criteria in mind, a limited number
capacity and to evaluate the associated cost/economics.
of representative examples of reaction, reactor, and process
models developed in the eld of biocatalysis over the past Furthermore, the majority of these models have been devel-
decade have been selected to allow an in-depth review and oped and applied for the purpose of kinetic investigations (Ta-
analyze representative literature on modeling biocatalysis ble 1). This is of course valuable on its own and necessary as a
(Table 1). However, before proceeding to analyze these rst step in the development of bioprocesses. Such analysis
models in detail, it is useful to carefully evaluate the overall provides insight into underlying mechanism, rate and the limi-
procedure followed during the model building, which is dis- tations (inhibition, mass transfer, etc) of the process. For
cussed in the next section. example, Paiva et al.30 provided a comprehensive review of
the kinetics and mechanisms of lipases which concluded that
the kinetics and thermodynamics cannot be separated from
Model building procedure each other. Nevertheless, due to the complexity that arises oth-
A typical procedure followed during process model build- erwise, in considering the inventory of Gibbs free energy and
ing consists of ve steps, summarized in Figure 2. Interest- the enthalpy of the system, most research efforts focused on
ingly, most models are built having a certain objective in what is called lumped rate analysis.
mind (albeit sometimes the purpose of the modeling study is Aside from the purpose of kinetics analysis, most other
implicit rather than explicit). Having specied the objective, studies have focus on developing process models to be used
the study usually moves on to characterize the reaction for process analysis and owsheet evaluation. For example,
mechanism and other relevant process units, e.g., mass trans- the studies of Berendsen et al.25 and Chen et al.26 have per-
fer or product separation/removal. In this step, the modeling formed process analysis using model simulations and suggest
objective indeed determines the extent and the direction of the model can be used for process design purposes. Similarly
the experimental investigation required, e.g., whether or not the studies of Illanes et al.,27 Schroen et al.,22 and Willeman et
pH and temperature effects should be checked. Having col- al.,23 use the process models (previously developed based on
lected the relevant experimental data for elucidation of the experimental work) to evaluate different process congura-
reaction mechanism, the third step formulates the model tions as well as performance limitations by carrying out simu-
structure which may involve dening reaction kinetics as lations. These studies are valuable as they demonstrate how
well as other process characteristics such as the reactor process model simulations can be useful for evaluating process
model, mixing type (ideal mixing vs. 2D or 3D hydrody- performance with different reactor choices or congurations.
namic models), mass-transfer, separation model, etc. In the With respect to design, the study conducted by Mathys
fourth step, the model parameters are then estimated from et al.28 distinguishes itself from the others as it primarily
experimental data and/or adapted from relevant literature. focused on using the model for detailed process design (e.g.,
Important to mention is that in this step parameter estimation unit rating), which ultimately is used for an economic evalu-
should be supplemented with proper validation studies to as- ation. Interestingly, the model used by Mathys et al. was
certain a certain reliability of the model, reliability which is built using empirical knowledge collected from past experi-
linked to the purpose of the study. Finally, in the fth step, mental work (i.e., bioconversion rates, yields, etc).
the model is used for the purpose it was built for. This, for It is important to note that in principle process models
example, could be to simulate alternative conditions or pro- could also be used as a tool for process optimization and con-
cess ow-schemes. trol (apart from the aforementioned studies of Rohne and
Overall, it is the belief of the authors that following a Meyer5 and Brass et al.6). Indeed, examples of such model
standardized and systematic procedure during model building uses appear rather seldom in biocatalysis. The term process
1532 Biotechnol. Prog., 2009, Vol. 25, No. 6

Table 1. Examples of Reaction, Reactor, and Process Models Applied to Biocatalytic Process
Model Identication
Purpose Process Model Structure Parameters Variables Data Ref.
Kinetic Hydrolysis of adipyl-7-ADCA to (i) Thermodynamic based 7 3 14 equilibrium; 15
analysis 7-ADCA using immobilized equilibrium constant/yield 3 initial rate; 12
glutaryl acylase calculation, (ii) Two-step temperature exps
reversible equilibrium reaction
kinetics plus Arrhenius
temperature correction.
Kinetic Synthesis of cephalexin using Two-step reversible equilibrium 10 Unclear 15 synthesis & 16
analysis phenylglycine amid and reaction kinetics plus Arrhenius hydrolysis exps
7-ADCA catalyxzed by temperature correction.
AssemblaseVimmobilized
R

penicillin G acylase
Kinetic Catalyze the conversion of Reversible bi-uni model for reaction 10 6 11 progress curves; 17
analysis benzaldehyde into (R)- kinetics plus mass transfer mass transfer; 11
mandelonitrile using Prunus kinetics between aqueous and validation exps
amygdalus hydroxynitrile organic phases plus mass
lyase balances.
Kinetic Acylation of glucose by lauric Simplied reversible ping-pong 3 4 3 initial rate; 1 pro- 18
analysis acid in 2-methyl 2-butanol by bi-bi kinetic model. gress curve; 1
Candida antarctica lipase equilibrium exps
Kinetic Whole-cell biocatalysis of Bioconversion kinetics 23 7 4 fed-batch at 2 L & 19
analysis pseudocumene to DMB- (biocatalysis), cell growth, mass 30 L
aldehyde by recombinant transfer and mass balances for
E. coli containing XMO aqueous and solvent phases.
Kinetic Esterication of propionic acid Ping-pong bi-bi mechanism with 4 3 initial rate exps at 20
analysis and 1-butanol in a water/ alcohol inhibition (competitive). micro and bench
n-decane two-phase system reactors
using lipase type B from
Candida antarctica
Kinetic/ Lipase catalyzed (R/S)-1- Reversible ping-pong bi-bi kinetics 11 6 6 initial rates; 1 21
process methoxy-2-propanol with extended with inhibition (Haldane enantiometric
analysis ethyl acetate type) coupled with a diffusion excess exps
model.
Process Enzymatic hydrolysis of The model of Schroen et al.15 10 3 1 validation exp 22
analysis adipyl-7-ADCA using extended with
immobilized glutaryl acylase pseudo-rst order rate of product
degradation.
Process Synthesis of (R)-Mandelonitrile The model of Willeman et al.17 10 6 23
analysis in an aqueous-organic
biphasic system
Process Baeyer-Villiger oxidation Regime analysis using linear 3 3 Two pilot-scale exps 24
analysis catalyzed by recombinant relations constrained with (50 L)
E. coli inequalities.
Process Lipase catalyzed (R/S)-1- 2-D model (mass and energy 11 6 Validation with enan- 25
analysis methoxy-2-propanol with balances) coupled to the kinetic tiometric excess &
vinyl acetate and diffusion model of Berendsen process perform-
et al.21 ance data
Process TK mediated carboncarbon Ping-pong bi-bi mechanism with 7 3 Literature 26
analysis bond formation using substrate inhibition integrated
hydroxypyruvate as ketol with empirical process variables.
donor
Process Enzymatic isomerization of An integrated process model Unclear Unclear No info 27
design glucose to fructose by consisting of mass balances,
immobilized glucose enzymatic action rates (reversible
isomerase Michaelis-Menten), enzyme
decay and mass transfer
limitations (due to immobilized
nature of enzyme).
Process Oxidation of n-octane to Steady-state mass and energy No info No info No info 28
design 1-octanol by Pseudomonas balances set-up for each process
oleovorans in fed-batch and unit in the process ow-diagram
continuous fermentation (19 total) that includes storage of
processes downstream raw materials to fermentation,
processing and product downstream processing and
purication (whole-cell product recovery. Aspen
biocatalysis) software.
Process Enzymatic production of A process model integrating enzyme 21 4 29
design hydrolyzedisomerized whey kinetics with competitive
permeate syrup (HIPS) using inhibition, negligible mass-trans-
immobilized lactase and fer limitation, enzyme inactiva-
glucose isomerase tion and a steady-state packed
bed reactor model.
Biotechnol. Prog., 2009, Vol. 25, No. 6 1533

Figure 2. Typical steps involved in building a process model: arrows indicate feedback loops (iterations between different process
modeling steps).

Figure 3. Model complexity vs. model publication year: no signicant increase in time for models developed in biocatalysis (left); ex-
ponential increase of complexity of activated sludge models (environmental engineering) that follows Moores law (right:
Reproduced with permission from Gujer, Water Sci Technol, 2006, 53, 111119).

optimization means that a certain objective function, e.g., et al.,25 on the other hand, elaborated on a detailed descrip-
product yield or enzyme use, is iteratively optimized via opti- tion of the mixing in the reactor by means of a 2D model,
mization algorithms subject to a number of constraints (as in which was demonstrated to capture the substrate gradient
generic optimization problems). Likewise process control prole in uidized bed reactor. In summary, most mechanis-
refers to studies where models are used to develop and/or tic models describe phenomena ranging from kinetics (rate)
evaluate control algorithms for achieving and maintaining an to mass/energy conservation to mass transfer limitations and
optimal performance, again dened by a certain objective mixing/hydrodynamics. However, it has not become standard
function. These two application objectives could indeed be a practice to take them all into account simultaneously when
further opportunity to exploit the process models. building process models.

Models and their structure Models and their complexity


Most models can be classied as mechanisticas the The complexity of dynamic models such as the ones
models are built assuming a certain reaction mechanism. developed for biocatalysis relates to the structure of the
When the focus of the model is on reactor design, mass bal- model. A simple indicator of model complexity can be
ances are also included to describe/compare various reactor inferred from the number of parameters, the number of state
congurations (batch or fed-batch vs. continuous). Moreover, variables, and the spatial resolution (i.e., 1D vs. 2D) it
mass transfer limitations are considered particularly when employs. According to Gujer,32 a pragmatic way to calculate
building models for immobilized enzyme systems29 as well model complexity can be done as follows: C npro  nvar
as biphasic systems (e.g., between organic and aqueous  ndim, where C refers to complexity, npro refers to the num-
phases when organic solvents are used in the biocatalytic ber of process steps described (e.g., enzymatic reaction,
process17,19,31). The effects of mixing and hydrodynamics in mass-transfer, etc), nvar refers to the number of state varia-
reactors have typically been neglected by assuming ideally bles, and ndim refers to the spatial resolution, e.g., 1D vs. 2D
mixed conditions (homogenous concentration throughout the description of the hydrodynamics in a reactor. In this way,
reactor), which were described using the continuously stirred the complexity of the models reported in Table 1 was calcu-
tank reactor (CSTR) concept. The study of Berendsen lated and plotted in Figure 3-left. The plot conveys a pattern
1534 Biotechnol. Prog., 2009, Vol. 25, No. 6

about the complexity of the biocatalysis models which sequences, biocatalytic kinetic models can normally be
actually shows statistically insignicant increase over the expressed by a single rate equation. The second distinguish-
past years. This trend is in sharp contrast with typical obser- ing feature is the sensitivity to operational conditions. This
vations in other engineering disciplines. For example, the means that, for example, the kinetic dependence of pH and
evolution of the complexity of dynamic models in environ- temperature is always required and may result in extra pa-
mental engineering displays an exponential increase over rameters in the expression. Finally, the third and most signi-
time (Figure 3-right taken from Gujer32). In fact, the increase cant difference is the presence of substrate and product
in the complexity of wastewater treatment plant models inhibition of enzyme activity at high concentrations. This is a
reported by Gujer closely follows Moores Law, which stipu- vital element of bioconversion reaction rate expressions and
lates that computational power (i.e., density of transistors per makes enzymatic reaction kinetic models more difcult to
unit area of integrated circuits) doubles every 2 years (Figure determine since they require estimation of many parameters
3-right). From this comparative analysis, one can argue that in a single differential equation. This automatically leads to
the complexity of biocatalysis models is not computationally the question of identiability, i.e., whether or not the model
limited. Hence, it would be technically feasible to further parameters can reliably be estimated from the available meas-
expand the biocatalysis models by, for example, increasing urements. This issue with its implications on the model appli-
the level of detail and the number of processes, phenomena, cations in biocatalysis is discussed at length below.
and spatial resolution currently described. Of course, next to
technical possibilities lies an important question: is a more
complex model needed? If so, what kind of motivation or Identiability of biocatalysis models
purpose would justify such a research undertaking? An an- Good Modeling Practice (GMoP) in general requires ana-
swer to this question is formulated below. lyzing the model identication results at least considering
the following criteria: (i) parameter estimation quality: e.g.,
95% condence interval and parameter correlation, (ii)
Model identication uncertainty of model predictions (i.e., uncertainty of the
With respect to the model identication procedure, one model outputs), and (iii) model corroboration (validation) on
observes that most often the nonlinear least-squares method an independent data set. Such formal statistical analysis,
is used to estimate a number of parameters based on an avail- described by Saltelli et al.,33 is deemed necessary to ascer-
able number of measurements. Most of these mechanistic tain scientic quality, which helps build credible and rigor-
models (particularly, those developed for the purpose of ously defendable models.3335 Different sensitivity analysis
kinetics analysis), contain three or four variables (substrate, methods are available, each with their own strengths and
product, enzyme, or cosubstrate concentrations) and four pa- weaknesses.36,37 For example, local methods such as differ-
rameters (Michaelis-Menten kinetic parameters, inhibition ential analysis provide in-depth insight into the sensitivity
coefcient(s), temperature, or pH correction parameters). Fur- function of a parameter. However, the results are conditional
thermore, the identication of these kinetic models is done to the point in parameter-space where the analysis is per-
using typical progress-time curve measurements with sub- formed, and therefore, the results cannot be extrapolated.
strate-enzyme tests in batch and also enantiomeric excess The Monte-Carlo-based regression methods (such as Standar-
curves. On the other hand, the process models developed dized Regression Coefcient) overcome this issue of local
with the aim of describing conversion performance in various sensitivity and provide good approximation of the global
reactor congurations naturally contain more parameters, sensitivity measure of a parameter. However, the associated
e.g., 23 in the case of the whole-cell biocatalysis study of computational demand will be high. For further discussion,
Buhler et al.19 These process models can also explain more the reader is referred to specic literature in the eld.33,36,37
variables, e.g., upto seven variables in the case of Buhler That said, one observes that the models developed so far
et al.19 The higher the number of model parameters, however, within biocatalysis have only partially been exposed to for-
the more measurementsor better put, the more information- mal statistical analysis. For example, among the models
rich measurementsare needed for their identication. listed in Table 1, rather few studies have performed a check
Hence, studies dealing with process models require measure- on the resulting quality of parameter estimates.15,16,18,20
ments of substrate consumption and product formation in Noteworthy is that some of these studies reported signi-
reactors.21 On the other hand, if the model is too complex to cantly large 95% condence intervals (CI) on some of the
identify its parameters from the available measurements, then estimated parameters. The 95% condence interval is an in-
one needs to consider simplifying the model structure as has dicator of the estimation error, and hence the higher its value
been demonstrated in the study of Flores and Halling.18 the worse the quality of the estimated parameter. Such a
large CI means that these parameters are unidentiable and
hence their values may not be reliable.34,38,39 Faced with
A Critical Discussion of Process/Kinetic such an issue (actually commonly encountered in the eld of
Models in Biocatalysis engineering), GMoP requires a proper diagnosis of the
Compared with chemical catalytic reactions, enzymatic unidentiability to overcome the problem (e.g., experimental
reactions have three typical features. The rst feature is the design or model-structure modication).34,40 It should be
reaction selectivity. Most chemical catalytic reactions are not mentioned that the study of Berendsen et al.21 recognized
selective (i.e., they will catalyze similar reactions involving the importance of the identiability issue during model
many different kinds of substrates or reactants). Although a building but did not provide a formal treatment of the issue,
few enzymes have been characterized that are not very selec- e.g., no parameter estimation quality was reported.
tive, the vast majority will catalyze only one reaction involv- Another important anecdote related to the unidentiability
ing only certain substrates. Therefore, unlike the use of issue can be found in the whole-cell biocatalysis study of
differential equations to describe chemical catalytic reaction Buhler et al.,19 which reports that different parameter values
Biotechnol. Prog., 2009, Vol. 25, No. 6 1535

Figure 4. Illustrating model parameter estimation errors and model prediction uncertainty traditional model identication where
model ts and mean parameter values are reported (left) vs. formal identiability analysis in which parameter estimation
errors are reported together with the model prediction uncertainty (right).

had to be used when describing process performance data ments. For example, in case one could provide evidence
obtained from batch runs at different scales (2 L vs. 30 L). following a formal identiability analysis41,42that the bio-
The authors attributed this issue to poor understanding of catalysis models are unidentiable, then this means the esti-
external inuences on the cell metabolism. This and other mated-parameter values may not be interpreted as unique
process related hypotheses (e.g., mass-transfer may have (the values themselves become arbitrary while the model t
been rate limiting at the larger reactor scale) could indeed be remains equally good41,42). In such a case, one needs to ana-
true. However, from a model identication point of view, lyze the implications of parameter identiability on model
these changes made to the parameters simply mean that the predictions which calls for uncertainty analysis.39,40
model is not valid (i.e., the model fails to describe a data set
independent from the one used for the parameter estimation).
On the other hand, the study of Berendsen et al.21 shows Prediction uncertainty of a biocatalysis model
convincingly good validation results for their model when To introduce the concept of prediction uncertainty in mod-
confronted with an independent data set obtained in a mini- eling biocatalysis, we have chosen a simple and known
plant. Again, the model validation results were not statisti- example from the eld of enzyme kinetics (a Michaelis-
cally quantied but based on expert-judgment of the quality Menten model, i.e., v Vmax SK S
m
). Our purpose is to illus-
of the ts. On the other hand, the study of Flores and Hal- trate the interactions between available experimental data, pa-
ling18 did indeed consider a statistical criterionmean rameter estimation error (e.g., 95% condence interval), and
squared error to be specicwhen comparing alternative prediction uncertainty of the model. Ultimately, we aim to
models in which simplications to the originally complex make the case for the importance of quantifying the predic-
model were introduced either by (a) replacing some of the tion uncertainty of biocatalysis models, particularly, when the
model parameters by dening a lumped (combined) parame- model predictions will be used as a basis for making engi-
ter or (b) removing the parameter from the model in case it neering judgments such as process design and optimization.
was found insensitive. While this study is valuable in its For this simple example, we perform a simulation exercise
own right to suggest a parsimonious model (from an original using the standard Michaelis-Menten model. The main moti-
model containing nine parameters down to ve), neverthe- vation for using simulated data is that true values of Michae-
less, it would have been valuable to perform a formal sensi- lis-Menten parameters are often (if not always) unknown for
tivity analysis to quantify the signicance of parameters on experimental data and hence limiting the discussion for the
the model predictions. purpose of this study. Simulation data (initial rates vs. sub-
With respect to the uncertainty aspect of model predic- strate concentration) were then generated using realistic val-
tions, one of the only studies known to us to consider this ues of Vmax and Km (15 lM/min and 25 lM, respectively)
issue was that of Mathys et al.28 who provided an expert- with a random measurement error (standard deviation equal
judgment-based uncertainty range (albeit no details were dis- to 5% Vmax) and are presented in Figure 4-left (shown with
closed). Needless to say this requires a formal treatment circles). Parameter identication (nonlinear least-squares
using a statistical method such as linear error propagation or method) using the simulation data provided a rather good t
a Monte-Carlo procedure.36,37,39 This uncertainty issue to the data (visually speaking). Despite the fact usually
becomes particularly important when the model is to be used model ts well, the parameter estimates can be seen to devi-
to support engineering decisions, e.g., extrapolating models ate from the true values. In fact, a closer look at the con-
outside the conditions of the experimental data used for pa- dence intervals shows, however, that the true values (see
rameter identication. Process design is an example to this above) are in fact included within the reported condence
end (see below for an illustration of identiability vs. interval, which is rather large (around 100% of the estimated
uncertainty). value for Km). Using the linear error method for uncertainty
Overall, a rigorous analysis of identiability is lacking for analysis37 (the method used for this exercise is provided as
biocatalysis models. Such a study is needed to theoretically Supporting Information), the parameter estimation errors
establish whether biocatalysis models can in principle be were propagated to quantify the error in the model predic-
identied reliably from the typical set of available measure- tionsshown as dashed lines on Figure 4-right. These
1536 Biotechnol. Prog., 2009, Vol. 25, No. 6

Figure 5. Prediction uncertainty of the model when extrapolated outside of the calibration conditions (left); improving prediction
quality (decreasing uncertainty) by additional measurements (initial rate experiments) (right).

condence intervals indicate the degree of uncertainty sur- may also involve confronting the simulation results with the
rounding each prediction. This uncertainty stems from both valuable process knowledge thereby keeping the model in
measurement error as well as model structure error. The lat- check and balance.
ter relates to the error due to the fact that the Michaelis- Another way of dealing with such uncertainty is to elimi-
Menten kinetics may not be the true model but an approxi- nate the sources contributing to the uncertainty. For example,
mate one to describe this particular set of data. This eventu- one can invest in further but focused experimental efforts to
ally contributes to the observed discrepancy between model improve the model prediction quality, i.e., decrease uncer-
predictions and experimental data. In the given example, the tainty to acceptable levels. This can be done formally fol-
condence intervals on the model predictions become larger, lowing DoE or based on a more traditional approach in
especially, at higher substrate concentration. This simple which empirical knowledge is used. For example, one could
example illustrates how parameter estimation errors actually collect data at higher substrate concentrations. It is com-
translate as model prediction errors. monly known that for proper estimation of Km, one need to
Taking into account, this prediction uncertainty will help have experimental data corresponding to substrate concentra-
discriminate between alternative models, as demonstrated in tions that are much higher than Km.43 In the given example,
the study of Flores and Halling.18 However, what is more we therefore assumed that initial rate measurements were
relevant, is that when the kinetic values derived from experi- also performed at 75 lM (Figure 5-right). This one addi-
ments are used in process models, their corresponding pa- tional data point (coming from one initial rate experiment)
rameter estimation errors will also be propagated as errors in led to a ve times decrease in the variance of the predicted
the model simulations. This is illustrated with the following catalysis rate at a substrate concentration of 500 lM (a
example in which the model (using the kinetic parameters variance of 4.5 lM/min compared with the variance of
that were identied earlier) is extrapolated beyond the exper- 21.1 lM/min previously (Figure 5-left)). This result provides
imental region to predict the enzymatic rate, i.e., beyond a a rigorous explanation for the well-known observation that
substrate concentration of 30 lM and upto a relatively high- measurements should additionally be made above Km value
substrate concentration of 500 lM. The results are shown in to allow adequate prediction quality.
Figure 5-left and it is clear that the model prediction error
(or the uncertainty) increases with the increasing substrate
Conclusions and Future Outlook for Process/Kinetic
concentration for this specic example. For example, the
Model Applications in Biocatalysis
variance of the predicted catalysis rate (r2v ) increases
from 0.3 lM/min at a substrate concentration of 30 lM to The application of mathematical models in enzyme tech-
21.1 lM/min at a substrate concentration of 500 lM. This nology has been strongly advocated over the past 20 years.
indicates that as one moves away from the model building In our opinion, modeling will play an increasingly important
region (i.e., the data set used for model identication), the role in the biocatalysis eld in the future. The idea behind
uncertainty in the predictions becomes very pronounced. this review article was to document the status of modeling in
From this simple exercise, one can contemplate the conse- the biocatalysis eld, and based on that status to nd out
quences when these parameter estimates are used in a pro- what is necessary to improve future modeling efforts.
cess model intended for process evaluation, e.g., for Most of the existing models describe biocatalyst behavior
developing process design concepts or control strategies. It is in terms of enzyme selectivity, mechanism, and reaction
clear that the quality of the parameter estimates has an kinetics. These models have generally been subject to con-
impact on the quality of the simulations and hence their reli- siderable simplication but nevertheless describe enzyme
ability. However, this does not mean the model cannot be action under dilute conditions and have been well reviewed
used for this purpose. On the contrary, it just means that the in the literature. More recently, work has focused on the
model predictions should be used with an open awareness of integration of these models with mass balances over reactors
the limitations (i.e., the uncertainty) surrounding model pre- to compare alternative operating schemes (batch, fed-batch)
dictions. Once the uncertainty is quantied, one can go on and hydrodynamics (plug-ow, well-mixed). In addition,
with evaluating different alternatives, but this time interpret- there are isolated cases where product recovery has been
ing the simulations with due care. This careful interpretation integrated into the process model.
Biotechnol. Prog., 2009, Vol. 25, No. 6 1537

One important conclusion is that biocatalysis models process optimization, and process control strategy evalua-
remain at a relatively low level of complexity, at least com- tions. Modeling should, for example, allow the comparison
pared with the trends in the development of model complex- of several bioprocess development alternatives by simulation,
ity observed in other engineering disciplines. As a thus saving experimental time and effort, while uncertainty
consequence, we conclude that there certainly is room in the analysis should allow the model user to have an idea about
biocatalysis eld for additional development at the model the degree of condence that can be placed on the results of
structure level. This is in fact very fortunate, since future the model study.
modeling challenges will indeed require more detailedand
thus increasingly complexprocess models to enable evalua-
Acknowledgments
tion of the suitable conditions under which enzyme-based ca-
talysis can replace or provide an alternative to chemical The research work of Dr. Gurkan Sin is nanced by the Dan-
synthesis. For example, many proponents of bioreneries ish Research Council for Technology and Production Sciences
suggest that renewable resources can be used to make plat- (FTP project No. 274-07-0339).
form chemicals as well as chemical building blocks based on
enzyme and heterogeneous catalysis.44 To evaluate the alter-
Literature Cited
native technologies for a given case, detailed process models
will be required. The use of life cycle analysis to aid the de- 1. Patel RN, editor. Stereoselective Biocatalysis. NY-Basel: Marcel
Dekker; 2000.
cision-making process is another example where detailed
2. Pollard DJ, Woodley JM. Biocatalysis for pharmaceutical inter-
mass balances will be required. Last but not least, another mediates: the future is now. Trends Biotechnol. 2007;25:6673.
example where detailed modeling (such as ux-balance anal- 3. Woodley JM. New opportunities for biocatalysis: making phar-
ysis) can be valuable is the development of bioprocesses maceutical processes greener. Trends Biotechnol. 2008;26:321
relying on whole-cell biocatalysis, in which modeling can 327.
help understanding the underlying constraints of the metabo- 4. FDA. Guidance for industry: PAT-A framework for innovative
pharmaceutical manufacturing & quality assurance, Available
lism45,46 and thereby help in devising optimal strategies for at: http://www.fda.gov/cder/guidance/6419fnl.pdf, 2004.
engineering cells as powerful biocatalysts.47 5. Rohner M, Meyer HP. Applications of modelling for bioprocess
On the basis of the analysis, we have undertaken of the design and control in industrial production. Bioprocess Biosyst
scientic literature we believe there are four areas which Eng. 1995;13:6978
require signicant development in the future: 6. Brass JM, Hoeks FWJMM, Rohner M. Application of modeling
techniques for the improvement of industrial processes. J Bio-
1. The majority of process models are steady-state and technol. 1997;59:6372.
extension to dynamic models will be required. 7. Smith R. Chemical Process Design and Integration. Chichester,
2. All the process models to date have been used for deci- England: Wiley; 2005.
sion-making based on a unique result for one set of parame- 8. Hult K, Holmquist M. Kinetics, molecular modeling, and syn-
thetic applications with microbial lipases. Methods Enzymol.
ters. There is a need to introduce uncertainty analysis to give
1997;286:386405.
condence in decision making and to direct experimental/ 9. Kazlauskas RJ. Molecular modeling and biocatalysis: explana-
parameter estimation efforts. tions, predictions, limitations and opportunities. Curr Opin
3. Assumptions to date are based on ideal conditions in Chem Biol. 2000;4:8188.
reactors. However, in the future, many large reactors (due to 10. Braiuca P, Ebert C, Basso A, Linda P, Gardossi L. Computa-
dilute aqueous solutions) will be used and the nonidealities tional methods to rationalize experimental strategies in biocatal-
ysis. Trends Biotechnol. 2006;24:419425.
will need to be modeled to facilitate scale-up. The develop- 11. Turner NJ. Directed evolution of enzymes for applied biocataly-
ment of suitable computational uid dynamics computer sis. Trends Biotechnol. 2003;21:474478.
codes might be required for such reactors and later will need 12. Kuchner O, Arnold FH. Directed evolution of enzyme catalyst.
to be integrated with existing models. Trends Biotechnol. 1997;15:523530.
4. Attempts to validate models will need to be enhanced. 13. Lye GJ, Woodley JM. Application of in situ product-removal
This is required not only to provide feedback to improve the techniques to biocatalytic processes. Trends Biotechnol. 1999;
17:395402.
models and assumptions made but also to ensure that con- 14. Woodley JM, Bisschops M, Straathof AJJ, Ottens M. Future
dence is built to promote the wider application of such directions for in-situ product removal (ISPR). J Chem Technol
models. Biotechnol. 2008;83:121123.
In summary, for modeling studies to be more mature and 15. Schroen CGPH, VandeWiel S, Kroon PJ, DeVroom E, Janssen
successful in this eld, we feel that there is need for a AEM, Tramper J. Equilibrium position, kinetics and reactor
concepts for the adipyl-7-ADCA hydrolysis process. Biotechnol
clearly dened modeling protocol, outlining the different Bioeng. 2000A;70:654661.
steps to be taken in a modeling study. Important elements in 16. Schroen CGPH, Nierstrasz VA, Moody HM, Hoogschagen MJ,
such a protocol should be as follows: Kroon PJ, Bosma R, Beeftink HH, Janssen AEM, Tramper J.
1. Development of a standardized and systematic guide- Modeling of the enzymatic kinetic synthesis of Cephalexin
line for model development and also model documentation inuence of substrate concentration. Biotechnol Bioeng. 2000B;
73:171178.
(an example of which has been proposed herein). The latter 17. Willeman WF, Gerrits PJ, Hanefeld U, Brussee J, Straathof AJJ,
will facilitate model exchange and thus knowledge transfer. van der Gen A, Heijnen JJ. Development of a process model to
2. Introduction of formal statistical analysis of identiabil- describe the synthesis of (R)-mandelonitrile by Prunus amygda-
ity of biocatalysis models as a default step in the modeling lus hydroxynitrile lyase in an aqueous-organic biphasic reactor.
procedure. Biotechnol Bioeng. 2001;77:239247.
3. Use of uncertainty/sensitivity analysis of the predictions 18. Flores MV, Halling PJ. Full model for reversible kinetics
of lipasecatalyzed sugar-ester synthesis in 2-methyl-2-butanol.
of biocatalysis models.
Biotechnol Bioeng. 2002;78:794800.
In our opinion, adoption of GMoP will indeed lead to 19. Buhler B, Straathof AJJ, Witholt B, Schmid A. Analysis of
more rigorous application of models within process design, two-liquid-phase multistep biooxidation based on a process
1538 Biotechnol. Prog., 2009, Vol. 25, No. 6

model: indications for biological energy shortage. Org Process 35. Saltelli A. The critique of modeling and sensitivity analysis in
Res Dev. 2006;10:628643. the scientic discourse. An overview of good practices.
20. Swarts JW, Vossenberg P, Meerman MH, Janssen AEM, Boom EUR 22487 EN: EUR Scientic and Technical Research series:
RM. Comparison of two-phase lipase-catalyzed esterication on Luxembourg; 2006.
micro and bench scale. Biotechnol Bioeng. 2008;99:855861. 36. Helton JC, Davis FJ. Latin hypercube sampling and the propa-
21. Berendsen WR, Lapin A, Reuss MA. Kinetic study of lipase- gation of uncertainty in analyses of complex systems. Reliab
catalyzed reversible kinetic resolution involving verication at Eng Syst Saf. 2003;81:2369.
miniplant-scale. Biotechnol Bioeng. 2006;95:883892. 37. Sin G, Eliasson AL, Gernaey KV. Good modelling practice
22. Schroen CGPH, Nierstrasz VA, Bosma R, Dijkstra ZJ, Vande- (GMoP) for PAT applications: propagation of input uncertainty
Sandt EJAX, Beeftink HH, Tramper J. Process design for enzy- and sensitivity analysis. Biotechnol Prog. 2009; DOI:10.1002/
matic adipyl-7-ADCA hydrolysis. Biotechnol Prog. 2002;18: btpr.166.
745751. 38. Holmberg A. On the practical identiability of microbial growth
23. Willeman WF, Straathof AJJ, Heijnen JJ. Comparison of a models incorporating Michaelis-Menten type nonlinearities.
batch, fed-batch and continuously operated stirred-tank reactor Math Biosci. 1982;62:2343.
for the enzymatic synthesis of (R)-mandelonitrile by using a 39. Omlin M, Reichert PA. Comparison of techniques for the esti-
process model. Bioprocess Biosyst Eng. 2002;24:281287. mation of model prediction uncertainty. Ecol Modell. 1999;
24. Law HEM, Baldwin CVF, Chen BH, Woodley JM. Process lim- 115:4559.
itations in a whole-cell catalysed oxidation: sensitivity analysis. 40. Dochain D, Vanrolleghem PA. Dynamical Modelling and Esti-
Chem Eng Sci. 2006;61:66466652. mation in Wastewater Treatment Processes. London UK: IWA
25. Berendsen WR, Lapin A, Reuss MA. Non-isothermal lipase-cata- Publishing; 2001.
lyzed kinetic resolution in a packed bed reactor: modeling, simu- 41. Brun R, Kuhni M, Siegrist H, Gujer W, Reichert P. Practical
lation and miniplant studies. Chem Eng Sci. 2007;62:23752385. identiability of ASM2d parameters-systematic selection and
26. Chen BH, Baganz F, Woodley JM. Modelling and optimisation tuning of parameter subsets. Water Res. 2002;36:41134127.
of a transketolase-mediated carboncarbon bond formation reac- 42. Sin G, Odman P, Petersen N, Lantz AE, Gernaey KV. Matrix
tion. Chem Eng Sci. 2007;62:31783184. notation for efcient development of rst-principles models
27. Illanes A, Zunaiga M, Contreras S, Guerrero A. Reactor design within a PAT framework: integrated modeling of antibiotic pro-
for the enzymatic isomerization of glucose to fructose. Biopro- duction with Streptomyces coelicolor. Biotechnol Bioeng. 2008;
cess Eng. 1992;7:199204 101:153171.
28. Mathys RG, Schmid A, Witholt B. Integrated two-liquid phase 43. Chen BH, Hibbert EG, Dalby PA, Woodley JM. A new
bioconversion and product-recovery processes for the oxidation approach to bioconversion of reaction kinetic parameter identi-
of alkanes: process design and economic evaluation. Biotechnol cation. AIChE J. 2008;54:21552163.
Bioeng. 1999;64:459477. 44. Hayes DJ. An examination of biorening processes, catalysts
29. Illanes A, Wilson L, Raiman L. Design of immobilized enzyme and challenges. Catal Today 2009;145:138151.
reactors for the continuous production of fructose syrup from 45. Chin JW, Khankal R, Monroe CA, Maranas CD, Cirino PC.
whey permeate. Bioprocess Eng. 1999;21:509515. Analysis of NADPH supply during xylitol production by
30. Paiva AL, Balcao VM, Malcata FX. Kinetics and mechanisms engineered Escherichia coli. Biotechnol Bioeng. 2009;102:209
of reactions catalyzed by immobilized lipases. Enzyme Microb 220.
Technol. 2000;27:187204. 46. Buhler B, Park JB, Blank LM, Schmid A. NADH availability
31. Woodley JM, Cunnah PJ, Lilly MD. Stirred tank two-liquid limits asymmetric biocatalytic epoxidation in a growing
phase biocatalytic reactor studies: kinetics, evaluation and mod- recombinant Escherichia coli strain. Appl Environ Microbiol.
eling of substrate mass transfer. Biocatalysis. 1991;5:112. 2008;74:14361446.
32. Gujer W. Activated sludge modelling: past, present and future. 47. Blank LM, Ebert BE, Buhler B, Schmid A. Metabolic capacity
Water Sci Technol. 2006;53:111119. estimation of Escherichia coli as a platform for redox biocataly-
33. Saltelli A, Ratto M, Andres T, Campolongo F, Cariboni J, Gate- sis: constraint-based modeling and experimental verication.
lli D, Saisana M, Tarantola S. Global Sensitivity Analysis: The Biotechnol Bioeng. 2008;100:10501065.
Primer. West Sussex, England: Wiley; 2008.
34. Seber G, Wild C. Nonlinear Regression. NY: Wiley; 1989. Manuscript received Dec. 18, 2008, and revision received Apr. 8, 2009.

Anda mungkin juga menyukai