B RIEFINGS IN BIOINF ORMATICS . VOL 7. NO 4. 364 ^374 doi:10.

1093/bib/bbl040
Advance Access publication November 14, 2006

Dynamic modelling and analysis

of biochemical networks:
mechanism-based models and
model-based experiments
Natal A.W. van Riel
Submitted: 28th August 2006; Received (in revised form): 4th October 2006

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

Abstract
Systems biology applies quantitative, mechanistic modelling to study genetic networks, signal transduction pathways
and metabolic networks. Mathematical models of biochemical networks can look very different. An important
reason is that the purpose and application of a model are essential for the selection of the best mathematical
framework. Fundamental aspects of selecting an appropriate modelling framework and a strategy for model building
are discussed.
Concepts and methods from system and control theory provide a sound basis for the further development of
improved and dedicated computational tools for systems biology. Identification of the network components and
rate constants that are most critical to the output behaviour of the system is one of the major problems raised
in systems biology. Current approaches and methods of parameter sensitivity analysis and parameter estimation are
reviewed. It is shown how these methods can be applied in the design of model-based experiments which iteratively
yield models that are decreasingly wrong and increasingly gain predictive power.
Keywords: systems biology; parameter sensitivity analysis; parameter estimation; optimal experiment design;
differential equations

INTRODUCTION for cells or whole organisms. Such models can be

Progress in molecular and cell biology has led to the regarded as the keystones of systems biology [2],
identification of complex biochemical networks ultimately providing scientific explanations of the
involved in the normal functioning of cells, tissues behaviour of biological systems in health and disease
and organs and also defects associated with many and assisting in the identification of therapeutic
diseases. However, most of this information is targets with the promise of molecular medicine and,
obtained in cell cultures and animal models. The on the long run, personalised medicine.
evaluation of their importance in the human disease System level understanding of complex processes
context is far from obvious [1]. The advances in is common in engineering disciplines and relies
experimental techniques provide an opportunity for heavily on mathematical models. Physiologists have
developing mechanistic mathematical models of traditionally used a holistic perspective to try to
biochemical networks, including signal transduction understand how life works by collecting quantifiable
pathways and metabolic networks. Building sound information on biological stimuli and responses.
dynamic models of biochemical networks is a key However, a systems approach is a rather new, and
step towards the development of predictive models not generally accepted, way of thinking in molecular

Natal A.W. van Riel, Biomodeling and Bioinformatics, Department of Biomedical Engineering, Eindhoven University of
Technology, PO Box 513, 5600 MB, Eindhoven, The Netherlands. Tel.: þ31-40-2475506; Fax: þ31-40-2472740;
E-mail: N.A.W.v.Riel@tue.nl
Natal van Riel is Assistant Professor of Biomodeling and Systems Biology in the Department of Biomedical Engineering at
Eindhoven University of Technology and Principal Investigator of Eindhoven Biomedical Systems Biology. His research interests
include mathematical modelling and identification of biological systems.

ß The Author 2006. Published by Oxford University Press. For Permissions, please email: journals.permissions@oxfordjournals.org
 Dynamic modelling and analysis of biochemical networks 365

and cell biology. As currently employed, the term as ‘Occam’s razor’ [12], which, as Einstein noted,
‘systems biology’ encompasses many different also implies that models should not be too simple.
approaches and models, and studies many organisms For successful application of systems biology
from bacteria to man [3]. these philosophies are complementary rather than
Despite the amount and quality of experimental mutually exclusive. The two paradigms can also be
data increase rapidly, quantitative measurements of recognised in the two principal approaches to the
many cellular components in time and space are still construction of biochemical network models [13,
relatively scarce. Data sparsity in combination with 14]. In one of his founding papers on systems
biological complexity impose severe challenges to biology, Kitano referred to a ‘bottom-up’ approach
standard engineering methodologies for modelling, as ‘trying to construct a gene regulatory network
simulation and mathematical analysis of biochemical from independent experimental data, mostly
networks and there is a great need of sound model obtained through a literature search and the rest
building methods. Predictive, mechanism-based obtained from experiments designed to provide

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

mathematical models of biochemical networks information about specific aspects of the network
require quantitative information on reaction rates of the interest’. A ‘top-down’ approach ‘tries to
and molecular concentrations. For most processes, make use of data gathered using a high-throughput
these parameters are not directly accessible in vivo. DNA micro-array and other measurement technol-
Existing biochemical data usually originate from ogies to infer genetic network structures’ [15], which
different experimental settings, cell types and states of is also referred to as ‘reverse engineering’ [16–19].
cells and can therefore not be automatically trusted A substantial number of methods have been devel-
and used for quantitative models. Furthermore, oped to try to solve these complex inverse problems
cellular processes are described on different levels of [20–24]. However, the focus of this review is on
information quality, ranging from mechanistically mechanism-based models of biochemical networks.
well-understood interactions to purely hypothetical It should be noted that in several papers on
interactions described in qualitative terms like systems biology the ‘top-down’ and ‘bottom-up’
activation or inhibition. In cases where the mechan- terms have been used with a different meaning. In
istic details are unclear, it is necessary to fill in the drug discovery ‘bottom-up’ refers to going from
gaps by postulating simple mechanisms without molecular biology techniques applied to in vitro
having any kinetic parameters available [4]. If the systems up to clinical trials [1, 3]. This interpretation
dynamic behaviour of a system is highly dependent of bottom-up and top-down is more closely
on the value of (some of) the parameters, then related to their meaning in computational physiol-
accurate and reliable quantification of the parameters ogy, such as applied in the Physiome project,
is essential for the development of predictive models reflecting the multi-scale hierarchy going from
and therefore parameter estimation is an important organ systems (top) to the genome (bottom,
research topic in systems biology [5–8]. Figure 1) [11, 25].
Whether an inductive or a parsimonious,
hypothesis-driven approach is most suited, and
BOTTOM-UP VERSUS TOP-DOWN which mathematical framework to use is essentially
Approaches to mathematical and computational dependent on the purpose of the model. Therefore,
modelling in biology can be thought of as falling any modelling effort should start with defining this
into two philosophical categories [9–11]. The first purpose. In brief, the following reasons could
includes large-scale modelling approaches for warrant the use of mathematical models [10, 26]:
which it is assumed that important biological (1) to organise disparate information into a coherent,
features emerge from their simulation and analysis self-consistent whole, (2) to think (and calculate)
(inductive science). Complementary to this approach logically about what components and interactions are
is hypothesis-driven research, traditional in among important in a complex system, (3) to simulate,
other physics, which aims for the simplest possible predict, and optimise procedures, experiments and
model that captures the key features of a system therapies, (4) to disprove hypotheses and to define
consistent with the level of available experimental improved hypotheses and (5) to understand the
data. This fundamental principle of generating essential features of a system. Disproof of hypotheses
parsimonious hypotheses and models is referred to that are formalised as mathematical models is based
366 van Riel

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

Figure 1: Integrative physiology and (molecular) systems biology. ‘Omics’ focuses on the identification and global
measurement of molecular components and reverse engineering of the connection network. Mechanistic modelling
attempts to form integrative (across scales) models of human physiology and disease. One could argue that the term
‘systems biology’ is currently inappropriately limited to the molecular scale and needs to be associated with all spatial
and temporal scales [11].

on the comparison of model-predicted and experi- Several review, opinion and tutorial papers use
mentally measured variables. Thus working toward rather simple and abstracted examples of signalling
goal 4 requires developing models and designing pathways to illustrate how mathematical modelling
experiments in tandem, ensuring that sets of and analysis can help in understanding concepts
modelled and measured variables can be matched like emergence, feedforward and feedback loops,
to each other and that experiments are optimally stability, robustness (and fragility), adaptation and
designed to identify model unknowns [10, 11, 27]. homeostasis. This might leave the impression that
Over the past years, many papers have reviewed systems biology per se aims for simplification and
approaches to the development of computational functional descriptions, while loosing sight of the
models of biochemical networks and described details and complexity of actual reaction networks.
the insights gained from models [4, 28–37]. Some Many efforts pursue realistic large-scale complex
cover an extensive overview of many different models, but simplified models show to be
mathematical frameworks, including Boolean and valuable in terms of understanding the essential
Bayesian networks and stochastic models [38], and/or qualitative behaviour of biological systems
others are restricted to a specific class of models, [10, 34, 40]. A focus on system behaviour of the
such as so-called hybrid systems, which combine network does not imply that known molecular
continuous differential equations and discrete details are always to be neglected [41]. Nevertheless,
events [39]. Like most reviews, we focus on conti- a course-grained level of description is often the
nuous, deterministic descriptions of the dynamics by only practical approach if molecular details of parts
applying non-linear ordinary differential equation of the biochemical network lack [4, 42].
(ODE) models of biochemical pathways. The
so-called state-space formulation is most convenient
to describe and implement an input–output system MODEL BUILDING CYCLE
as a set of first-order differential equations for Arguably, systems theory provides the most formal
a vector of state variables and an algebraic output and most general (not application domain specific)
equation. definition of the cycle of building models of
 Dynamic modelling and analysis of biochemical networks 367

dynamic systems. The typical model building cycle Despite such endpoints will often be static, this
as considered in the area of system identification helps to verify the consistency of the model and is
[5, 43–45] starts from a goal definition (purpose important to gain confidence in the predictions.
of the model) and some a priori knowledge Next, system analysis tools, such as parameter
(i.e. preliminary data and initial hypotheses) on the sensitivity analysis, can be used to design experiments
basis of which a model framework is chosen and to generate optimally informative data for parameter
a model structure is proposed. From the available estimation or to validate model predictions. Optimal
data, parameter estimation is then performed, leading experimental design (OED) includes how to perturb
to a first working model. This initial model must be the system, which components to measure and
validated with new experiments, which, in most at which time points to sample (model-based
cases, will reveal a number of deficiencies. Thus, experiments). To develop predictive models of
a new model structure and/or a new experimental realistically large biochemical networks dozens of
design must be planned and the process is repeated components (proteins, metabolites) will have to be

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

iteratively until the validation step is considered measured in time. To use parameter estimation
satisfactory. techniques, quantitative and repeatedly measured
An underlying assumption of this cycle is that the data on cell responses to physiological stimuli and to
system inputs can be easily and freely perturbed by pharmaceutical agents are needed [3]. Compared
the experimenter to reveal a rich ensemble of to the high-throughput experiments in functional
dynamic responses and that the responses of many genomics these are generally medium or low
of the system variables can be measured accurately. throughput assays. To validate the model, it is used
Furthermore, it is assumed that this cycle can be to formulate hypotheses about the system behaviour
iterated several times without significant extra ‘costs’. under different conditions than used to estimate the
These assumptions do not hold for biological parameters. Subsequently, these experiments are
systems. For systems biology a somewhat different, performed and the measured responses are compared
more specific workflow is emerging as a practical quantitatively with the model predictions. The
approach towards mechanism-based models of iterative cycle of prediction and experimental
biochemical networks. After the purpose of the validation progressively strengthens the predictive
model has been defined, the next step is to define the power of the model. Sensitivity analysis, parameter
connection map of the network. For all the compo- estimation and optimal experiment design will
nents and their interactions parameter information be discussed in the remainder of this review.
needs to be collected. This involves (basal) concen-
trations for each component and kinetic rate
constants for interactions and enzymatic reactions. PARAMETER SENSITIVITY
Some of this information is available as tabulated ANALYSIS
numerical values in the literature. However, in many Mechanism-based models allow to predict the
cases, the (in vitro) quantitative information has to be behaviour of the specified system over time and to
manually extracted from the primary literature. track its dynamics for each set of fixed system
In cases where the mechanistic details are unclear, parameters. However, all of the parameters, includ-
it is necessary to fill in the gaps with simple ing rate constants and initial component concentra-
mechanism [1, 4, 42]. Before starting any biological tions, must be experimentally measured or inferred
experiment, it is most efficient to adapt the model to specify the model. Even for those models with
until it is consistent with most of the heterologous experimentally determined parameters, it is still
datasets (i.e. using data from different tissues or uncertain whether the particular set of parameters
species). Based on model simulations that are closely approximates the corresponding biological
discrepant for (some of) the data, the model can system, because (some of) the kinetic parameters are
already be iteratively improved without performing usually taken or estimated from measurements
own ‘wet’ experiments. Preferably, the literature reported by different laboratories using different
also provides quantitative physiological ‘endpoints’ in vitro models and conditions. It is therefore
that have been measured in an intact system important not only to study the dynamical properties
(in vivo, in situ, ex vivo) and can be compared to the governed by the particular kinetic parameters,
corresponding endpoints as predicted by the model. but also to further investigate the effects of their
368 van Riel

perturbations on the overall system [46]. The between the output for the reference parameter
identification of the network components and rate values and the perturbed system output [45, 46, 59],
constants that are most critical to the output the area under the curve of the output [8], the
behaviour of the system is one of the major problems amplitude and period of oscillation [56, 57], the
raised in systems biology [47]. Besides analysis of the deviation from the steady-state values [51] and
effects of parameter uncertainty on model outcome, the output value after a specific amount of time,
other applications of parameter sensitivity analysis which could be an endpoint that is clinically used
are: (1) analysing robustness [48, 49], (2) determining to assess drug efficacy [1, 50].
modularity of biochemical networks and model Local sensitivity analysis allows only one param-
reduction [42], (3) identification of control points eter to vary for each simulation and deals with only
for, among others, drug target selection [1], (4) small perturbations of the reference model, i.e. local
discrimination of phenotypes for biomarker selection sensitivity analysis pertains to a particular point in
[50] and (5) experiment design [5, 42, 45, 51, 52]. the parameter space. However, rate constants and

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

Robustness allows a system to maintain its functions concentrations of diverse molecules may vary
despite external and internal perturbations [53] and extensively in an interactive manner among different
is one of the system properties of biochemical cellular environments. For these reasons it is more
networks that has been widely addressed using appropriate to explore, in a probabilistic context,
computational approaches [49]. With parameter possibilities of non-linear effects from simultaneous
sensitivity analysis it can be investigated if a system parameter variations of arbitrary magnitudes by
is robust in the sense that it is capable to operate means of a global sensitivity analysis [46]. Only
reliably when its physical and biochemical param- recently global parameter sensitivity analysis (GPSA)
eters vary within their expected ranges, or that its methods have been started to be applied to
outcome is insensitive to the precise value of the biochemical network models [42, 45, 46, 59].
parameters when these are not (accurately) known. GPSA has the advantage that it provides a more
Conversely, parameters and components that are comprehensive analysis. However, varying all possi-
very sensitive to parametric variations can introduce ble combinations of parameters for a wide range of
fragility into the system. Sensitivity analysis assists in possible values per parameters rapidly becomes
identifying the critical steps in the network that computationally infeasible for realistically large
could drive disease development and these control biochemical networks. Therefore, all GPSA methods
points provide potential therapeutic targets. If, by apply a Monte Carlo strategy using sampling to
changing parameters (including basal component generate random sets of parameter values for
concentrations), the model can be made to repro- simulations [59]. A range of the parameter distribu-
duce different characteristic disease phenotypes and tions can often be obtained from the available
this outcome is sensitive and specific for a certain literature or guided by the experiences of the
parameter value, then this parameter might be a researchers. Latin hypercube sampling [60] is an
novel biomarker. efficient method to sample random parameter
Local parameter sensitivity analysis (LPSA) has vectors while guaranteeing that individual parameter
been used as an in silico investigation method to ranges are evenly covered [46]. Zhang and Rundell
identify critical parameters in dynamic models of [59] embedded GPSA in a random-search parameter
signal transduction [8, 54–56] and metabolic path- identification routine. The sensitivity index is
ways [57]. The sensitivity coefficient is defined as: calculated for each parameter set generated by the
Genetic Algorithm during parameter optimisation.
MðyÞ=MðyÞ
SM

¼ ð1Þ The fit of the output of the system with the

=

perturbed parameters to the output for the reference
where
represents the parameter that is varied and parameter values has commonly been used as
M(y) a characteristic of the response of the system sensitivity criterion. A sum of squared errors
output y. M is the change in M due to the change between the perturbed and reference output is

in
. This relative sensitivity coefficient is similar used as objective function to evaluate the fit [42,
to the control coefficients of metabolic control 45, 46, 59].
analysis [58]. Typical response characteristics that Bentele et al. [42] calculated global sensitivities
have been applied are: sum of squared differences from a weighted average of local sensitivities.
 Dynamic modelling and analysis of biochemical networks 369

The methods of Cho et al. [45], Zi et al. [46] studies, sufficient components of a biochemical
and Zhang and Rundell [59] classify the sampled network were measured in time such that the
parameter sets as either acceptable or unaccept- parameters of a simplified model could be uniquely
able given a threshold value of the objective and accurately estimated given the data of inputs and
function. The distributions of the parameter values outputs (the model was shown to be ‘identifiable’)
associated with acceptable and unacceptable cases are [6, 8]. Swameye et al. [8] estimated five unknown
statistically evaluated to rank the sensitivities. parameters in a model with four state variables. The
LPSA and GPSA have been compared by model of van Riel and Sontag contained five state
applying the methods to the same model. The variables and seven unknown model parameters [6].
resulting rankings of the parameters were signifi- The parameters of both models were estimated by
cantly different [46, 59]. Zi et al. [46] applied LPSA applying the maximum likelihood estimator (MLE)
and their GPSA method to a model of the IFN- and the accuracy of the estimates was assessed by
induced JAK-STAT signalling pathway, a stress calculating the variances from the Fisher Information

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

response pathway related to the immune system, Matrix (FIM). MLE has been widely used for
containing 32 state variables and 51 parameters. The parameter estimation problems [5–7, 43, 52,
response of phosphorylated STAT1 dimers in the 65, 66]. It is assumed that the measured data have
nucleus (output) to a step function in IFN- (input) been obtained from a stochastic process, because of
was used as the characteristic response for MPSA. the presence of unmodelled dynamics and measure-
Besides LPSA, Zhang and Rundell [59] also ment noise. The measurement error is assumed to be
compared five GPSA methods: the method of additive with a Gaussian distribution with zero mean
Bentele et al. [42], their own novel method, the and known variance. This implicates that the
partial rank correlation coefficient (PRCC) method estimated parameters are also stochastic quantities
[61], Sobol’s method [62] and the Fourier amplitude with a Gaussian distribution. It can be shown that an
sensitivity test (FAST) [63]. The latter three have unbiased parameter estimate (the expected mean
been used in engineering applications [64]. The value) can be obtained by minimising a weighted
methods were applied to a model of a T-cell sum of squared model residuals, which is similar to
receptor-activated Erk-MAPK signalling pathway the cost function used in MPSA. The covariance
(24 state variables and 69 parameters of which 49 matrix of parameter estimates has the inverse of the
were included in the sensitivity analysis). The FIM as lower bound (the so-called Cramér–Rao
sensitivities as predicted by the ‘traditional’ GPSA bound [44]). The FIM quantifies the accuracy of the
methods (PRCC, Sobol’s method and FAST) were estimates. If the weighting matrix in the cost
highly similar. Like LPSA, the method of weighted function is selected as the inverse of the data
average of local sensitivities [42] produced a very covariance matrix, then the parameter covariance
different sensitivity pattern. The results of the Zhang matrix is equal to (FIM)
1. The parameter estimates
and Rundell method were in between the local have a minimum variance and the diagonal elements
approaches and the existing GPSA methods. of (FIM)
1 are approximations of this variance.
The method of Zi et al. [46] was not included in The non-linear least squares criterion in combi-
this comparison. nation with gradient-based optimisation methods is
available in many software tools and computational
languages [67]. One difficulty with parameter
PARAMETER ESTIMATION FROM estimation in non-linear differential–algebraic equa-
PARTIAL MEASUREMENTS tion models is the existence of multiple local minima
The number of assessable parameters and therefore in the cost function. The traditional gradient-based
the maximum size of biochemical network models optimisation methods often fail to arrive at the global
have been very limited due to the large amount of optimal solutions. Banga and coworkers [68]
experimental data required for high-dimensional have compared several advanced deterministic and
parameter estimation problems. The space of possible stochastic global optimisation methods for parameter
sets of parameter values grows exponentially with the identification. The random search methods include
number of unknown parameters, which severely methods such as Genetic Algorithms and Simulated
impairs the search for the globally optimal parameter Annealing. The stochastic strategies appeared to
values (the ‘curse of dimensionality’ [42]). In a few be the most efficient and reliable to locate the
370 van Riel

parameter region containing the global solution, but of components while minimising the use of reaction
global optimality is not guaranteed. Furthermore, all fluxes. This requirement is analogous to a classic
global optimisation methods suffer from the large problem in automatic control, namely, the state
computational burden required, even for moderately regulator problem (SRP), which penalises deviations
sized problems. (As case study, 36 parameters of a of the state variables as well as large control action.
non-linear biochemical model were estimated using The SRP estimator calculates an optimal set of state
a PC Pentium-III, which took 40 h.) In a recent trajectories via a constrained convex programming
contribution, Banga and co-workers [5] presented problem in which the available measurements are the
a new optimisation methodology reducing compu- constraints. It is important to note that the reaction
tation time by one order of magnitude by means kinetics and therefore the unknown parameters are
of a hybrid method combining global and local not used in the SRP algorithm.
optimisation methods which increases efficiency
while guaranteeing robustness.

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

Recently, three different approaches to overcome Feature-based constraints
the curses of data scarcity and high dimensionality of In contrast to the scarce quantitative time-series data,
the parameter optimisation problem in large-scale, biology is gifted with enormous amounts of
mechanistic models of biochemical networks were qualitative observations and semi-quantitative infor-
reported: (1) infer the state trajectories of compo- mation. Recently, several publications have tried to
nents that have not been measured and subsequently make use of qualitative features of network beha-
use these profiles as additional data [52, 69], viour for the automatic quantification of model
(2) include well-known biological features (semi- parameters [6, 70, 71]. It is not trivial to translate
quantitative information) of the behaviour of the qualitative features into a mathematical function that
system in the parameter estimation algorithm [6, 70, can be included in the numerical optimisation
71] and (3) make use of the modular topology of algorithm for parameter estimation. The most
many biochemical networks [6, 42]. applied approach is to heuristically construct an
objective function by summing penalty functions for
State regulator approach each of the feature constraints [70, 71]. The relative
Generally, it is not possible to measure all time- importance of the different terms is rather arbitrary
varying components in a genetic, signalling or and is determined by introducing a weighting factor
metabolic network. Doyle III and coworkers for each term, yielding a weighted optimisation
[52, 69] have decoupled the problem of parameter criterion. Amonlirdviman et al. [71] used the patterns
estimation from partial measurements into two parts. of the hairs on the Drosophila wings as qualitative
First, the available measurements are used to estimate features. Locke et al. [70] combined qualitative
the profiles of all unmeasured concentrations and features of mRNA profiles extracted from different
fluxes. In the second part the reconstructed state publications on the circadian clock circuit in the
trajectories and fluxes are used as additional ‘data’ plant Arabidopsis thaliana.
to estimate the model parameters. In system and Van Riel and coworkers [6, 74] developed an
control theory a well-known solution to estimate the approach which combines data fitting and feature
behaviour of unmeasured components given partial constraints in a multi-objective estimation problem.
measurements of other system variables is the The data fit objective function was augmented with
Kalman filter. However, a Kalman filter requires an a penalty function prohibiting parameter values that
accurate model of the system and the development would violate the a priori information. The a priori
of such models is exactly the aim of systems biology. information consisted of the steady-state fluxes,
Doyle III and coworkers used an extension of calculated by flux analysis, in a metabolic network
dynamic flux balance analysis (dFBA, [72, 73]) of nitrogen metabolism in the yeast Saccharomyces
to estimate unmeasured protein and reaction rate cerevisiae. These flux values were subsequently used
trajectories. The premise of dFBA is that cellular as constraints when a dynamic model of the same
processes have evolved regulatory structures that network was fitted to time-series data of extracellular
optimally use cellular resources. This translates and intracellular metabolites after perturbing
into the postulate that the network flows are the steady-state cell culture by injecting a nitrogen
managed to minimise internal accumulation substrate [74].
 Dynamic modelling and analysis of biochemical networks 371

Modular topology in pharmacokinetic compartment modelling to

As indicated before, parameter sensitivity analysis can estimate subject specific physiological parameters
be used to analyse the modularity of biochemical (e.g. quantification of insulin sensitivity [75]). In
networks. If the network has a modular structure the systems biology the approach is particularly useful in
distributions of the sensitivities calculated with GPSA validation experiments, because model parameters
show distinct and narrow peaks indicating that the can be fitted to experimental data generated by a
sensitivities of the system are highly robust to large reference cell type (‘wild-type’) and then testing this
variations in most of the parameter values [42]. model on data generated by a variation (‘mutant’).
Because of this modular structure of biochemical
networks the inverse problem to estimate model
parameters from experimental data can be split in a OPTIMAL EXPERIMENTAL DESIGN
hierarchically nested set of parameter estimation For a systems biology approach and especially
problems [42]. Clusters can be identified that contain parameter estimation, the dynamics of cellular inter-

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

a subset of molecules whose concentrations depend action networks need to be measured by taking
on a subset of parameters only, indicating that these comprehensive, high-resolution quantitative mea-
parameters can be estimated locally. Global param- surements of the intracellular status, such as the
eters, those that belong to more than one cluster, concentrations, interactions, modifications and loca-
are estimated by optimising all clusters. Bentele et al. lisations of molecules, and of cellular structures in
[42] could reduce the dimensionality of the time and space under various conditions [76]. The
optimisation problem from 58 unknown parameters experiments to generate such data are complex and
to 18 in a model of CD95-induced apoptosis. expensive, as a consequence of which the time-series
Van Riel and Sontag [6] developed a different available are usually rather short, with few if any
approach to utilise the modular structure of replicates. Almost certainly, not all variables one
biochemical networks to reduce the complexity would like to include in a model can be measured.
of parameter estimation [6]. If the profiles of some of Furthermore, the robustness of many biomolecular
the proteins or metabolites can be (accurately) systems implies that arbitrary perturbations may result
measured in time, these signals can be used as in a minimal response in many of its molecular
inputs during simulation. These measured inputs components. Thus, if the experiments are not
emerge from the total system and cannot be appropriately designed, it is very likely that little
independently defined in an experiment; therefore information of value can be recovered from the
Van Riel and Sontag have called these as ‘dependent experimental measurements for identifying the
inputs’. The concept of dependent inputs allows model parameters. OED is therefore an important
the ‘opening’ of some of the connections of the research problem in systems biology [7, 51, 77].
pathways of interest with the rest of the cell and Identification of the most sensitive parameters can
its environment (Figure 2). The model can be guide the design of model-based experiments to
restricted to a smaller part of the network and the provide data that are required to estimate those
unmodelled dynamics are replaced by the measured parameters [42, 45]. However, system identification
dependent inputs. This approach of closed-loop theory offers a more rigorous approach to OED.
system identification has been previously applied According to the theory of MLE, the FIM is a
measure of the information content of the data and
its inverse provides an estimate of the accuracy of the
parameter estimates. Therefore, OED can be
formulated as an optimisation problem, in which a
scalar function of the FIM is maximised to provide
Figure 2: The concept of the ‘dependent inputs’ makes
‘rich’ experimental data to increase the accuracy of
use of the modular structure of biochemical networks
and can be used to reduce the complexity of the model.
the parameter estimates [5, 52, 77]. The shape of the
Input u can be experimentally manipulated and output y input perturbation is one of the important aspects of
is measured. Signal v results from the interaction of an optimal design. In general, complex perturbations
module A with module B. If v is also measured it can be generate more informative data than simple ones
treated as a dependent input to simulate a model of A, [78]. OED can also be used to optimise the sampling
without requiring a model of B. times [7]. Restrictions with respect to the number
372 van Riel

of samples that can be taken or the type of input References

perturbations that can be applied can be taken into
1. Rullmann JAC, Struemper H, Defranoux NA, etal. Systems
account. biology for battling rheumatoid arthritis: application of the
entelos physiolab platform, Syst Biol, IEE Proc 2005;152:
256–62.
2. Wolkenhauer O. Systems biology: the reincarnation of
CONCLUSION systems theory applied in biology? Brief Bioinform 2001;2:
258–70.
Systems biology is arguably one of the most
3. Butcher EC, Berg EL, Kunkel EJ. Systems biology in drug
influential ideas in life science research today, discovery. Nat Biotech 2004;22:1253–9.
promising to yield a more global and in-depth 4. Neves SR, Iyengar R. Modeling of signaling networks.
understanding of biological systems, furnishing Bioessays 2002;24:1110–7.
insights into the root causes of complex diseases, 5. Rodriguez-Fernandez M, Mendes P, Banga JR. A hybrid
contributing to the drug discovery process and facil- approach for efficient and robust parameter estimation in

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

biochemical pathways. Biosystems 2006;83:248–65.
itating a variety of aspects of health care including
6. van Riel NAW, Sontag ED. Parameter estimation in
treatment evaluation and predictive medicine [79]. models combining signal transduction and metabolic
Precise estimates of parameters are not always pathways: the dependent input approach. IEE Proc Syst Biol
needed to understand certain qualitative features of 2006;153:263–74.
the dynamic behaviour of biochemical networks. 7. Kutalik Z, Cho KH, Wolkenhauer O. Optimal sampling
time selection for parameter estimation in dynamic pathway
Parameter sensitivity analysis is an important tool to modeling. Biosystems 2004;75:43–55.
analyse which parameters do need to be accurately 8. Swameye I, Muller TG, Timmer J, et al. Identification of
known to develop predictive models. nucleocytoplasmic cycling as a remote sensor in cellular
signaling by databased modeling. Proc Natl Acad Sci 2003;
100:1028–33.
Key Points 9. Kell DB, Oliver SG. Here is the evidence, now what is
 Inductive research, common to omics approaches, and holistic the hypothesis? The complementary roles of inductive and
and hypothesis-driven research, typical of physiology and physics, hypothesis-driven science in the post-genomic era. Bioessays
should be applied in parallel for successful systems biology. 2004;26:99–105.
 Mechanism-based dynamic models provide a platform to inte- 10. Beard DA, Bassingthwaighte JB, Greene AS.
grate heterologous data into a self-consistent description that Computational modeling of physiological systems. Physiol
can be used for hypothesis-driven research and the design of Genom 2005;23:1–3.
model-based experiments.
 System and control theory provides many valuable concepts and 11. Hunter P, Nielsen P. A strategy for integrative computa-
methods for systems biology. Concepts from control theory are tional physiology. Physiology 2005;20:316–25.
successful in explaining how biochemical networks can be very 12. Domingos P. The role of Occam’s razor in knowledge
stable in the face of unwanted perturbations, while simulta- discovery. Data Mining Knowl Disc 1999;3:409–25.
neously responding in a highly sensitive and specific manner to 13. Palsson B. In silico biology through ‘‘omics’’. Nat Biotech
desired stimuli. System identification theory provides methods 2002;20:649–50.
for parameter sensitivity analysis, parameter estimation and
the design of optimally informative experiments, which itera-
14. Stark J, Callard R, Hubank M. From the top down: towards
tively yield models that are decreasingly wrong and increasingly
a predictive biology of signalling networks. Trends Biotechnol
gain predictive power. The specific requirements of systems
2003;21:290–3.
biology also stimulate new developments in system theory [80]. 15. Kitano H. Perspectives on systems biology. New Gener
 Local methods of parameter sensitivity analysis and parameter Comput 2000;18:199–216.
estimation are well-established and have been successfully 16. Tegner J, Yeung MKS, Hasty J, et al. Reverse engineering
applied. More advanced methods for parameter sensitivity analy- gene networks: Integrating genetic perturbations with
sis based on Monte Carlo simulations and global optimisation dynamical modeling. Proc Natl Acad Sci 2003;100:5944–9.
methods based on random search algorithms for parameter esti-
17. Sachs K, Perez O, Pe’er D, et al. Causal protein-signaling
mation problems and optimal experimental design receive
networks derived from multiparameter single-cell data.
increasing attention, but these areas need further development.
Science 2005;308:523–9.
Currently the methods are tailored towards specific problems
to reduce the computational burden. 18. Workman CT, Mak HC, McCuine S, et al. A systems
approach to mapping DNA damage response pathways.
Science 2006;312:1054–9.
19. Brazhnik P, de la Fuente A, Mendes P. Gene networks:
how to put the function in genomics. Trends Biotechnol 2002;
20:467–72.
Acknowledgements 20. Cho KH, Choo SM, Wellstead P, etal. A unified framework
The author wishes to thank Mark Musters for his helpful for unraveling the functional interaction structure
comments and Ad Damen and Paul van den Bosch for useful of a biomolecular network based on stimulus-response
discussions. experimental data. FEBS Lett 2005;579:4520–8.
 Dynamic modelling and analysis of biochemical networks 373

21. de la Fuente A, Brazhnik P, Mendes P. Linking the genes: 44. Walter E, Pronzato L. Qualitative and quantitative experi-
inferring quantitative gene networks from microarray data. ment design for phenomenological models – a survey.
Trends Genet 2002;18:395–8. Automatica 1990;26:195–213.
22. Kholodenko BN, Kiyatkin A, Bruggeman FJ, et al. 45. Cho K, Shin S, Kolch W, et al. Experimental design in
Untangling the wires: a strategy to trace functional systems biology, based on parameter sensitivity analysis
interactions in signaling and gene networks. Proc Natl using a Monte-Carlo method: a case study for the TNF-
Acad Sci 2002;99:12841–6. mediated NF-B signal transduction pathway. Simulation
23. Rangel C, Angus J, Ghahramani Z, et al. Modeling T-cell 2003;79:726–39.
activation using gene expression profiling and state-space 46. Zi Z, Cho KH, Sung MH, et al. In silico identification
models. Bioinformatics 2004;20:1361–72. of the key components and steps in IFN-
24. Sontag E, Kiyatkin A, Kholodenko BN. Inferring dynamic induced JAK-STAT signaling pathway. FEBS Lett 2005;
architecture of cellular networks using time series of gene 579:1101–8.
expression, protein and metabolite data. Bioinformatics 2004; 47. Kitano H. Systems biology: a brief overview. Science 2002;
20:1877–86. 295:1662–4.
25. Noble D. Modeling the heart – from genes to cells to the 48. von Dassow G, Meir E, Munro EM, et al. The segment

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

whole organ. Science 2002;295:1678–82. polarity network is a robust developmental module. Nature
26. Bailey JE. Mathematical modeling and analysis in biochem- 2000;406:188–92.
ical engineering: past accomplishments and future 49. El Samad H, Kurata H, Doyle JC, et al. Surviving heat
opportunities. Biotechnol Prog 1998;14:8–20. shock: control strategies for robustness and performance.
27. Endy D, Brent R. Modelling cellular behaviour. Nature Proc Natl Acad Sci 2005;102:2736–41.
2001;409:391–5. 50. de Pillis LG, Radunskaya AE, Wiseman CL. A validated
28. McAdams HH, Arkin A. Simulation of prokaryotic mathematical model of cell-mediated immune response to
genetic circuits. Annu Rev Biophys Biomol Struct 1998;27: tumor growth. Cancer Res 2005;65:7950–8.
199–224. 51. Feng Xj, Rabitz H. Optimal identification of biochemical
29. Asthagiri AR, Lauffenburger DA. Bioengineering models of reaction networks. BiophysJ 2004;86:1270–81.
cell signaling. Annu Rev Biomed Eng 2000;2:31–53. 52. Gadkar KG, Gunawan R, Doyle FJ, III. Iterative approach
30. Gombert AK, Nielsen J. Mathematical modelling of to model identification of biological networks. BMC
metabolism. Curr Opin Biotechnol 2000;11:180–6. Bioinform 2005;6:155.
31. Smolen P, Baxter DA, Byrne JH. Modeling transcriptional 53. Kitano H. Biological robustness. Nat Rev Genet 2004;5:
control in gene networks – methods, recent results, and 826–37.
future directions,. Bull Math Biol 2000;62:247–92. 54. Schoeberl B, Eichler-Jonsson C, Gilles ED, et al.
32. Hasty J, McMillen D, Isaacs F, et al. Computational studies Computational modeling of the dynamics of the MAP
of gene regulatory networks: in numero molecular biology. kinase cascade activated by surface and internalized EGF
Nat Rev Genet 2001;2:268–79. receptors. Nat Biotech 2002;20:370–5.
33. Bolouri H, Davidson EH. Modeling transcriptional 55. Lee E, Salic A, Krüger R, et al. The roles of APC and axin
regulatory networks. Bioessays 2002;24:1118–29. derived from experimental and theoretical analysis of the
34. Tyson JJ, Chen KC, Novak B. Sniffers, buzzers, toggles and Wnt pathway. PLoS Biol 2003;1:116–32.
blinkers: dynamics of regulatory and signaling pathways in 56. Ihekwaba AEC, Broomhead DS, Grimley RL, et al.
the cell. Cur Opin Cell Biol 2003;15:221–31. Sensitivity analysis of parameters controlling oscillatory
35. Eungdamrong NJ, Iyengar R. Modeling cell signaling signalling in the NF-B pathway: the roles of IKK and
networks,. Biol Cell 2004;96:355–62. IB. IEE Proc Syst Biol 2004;1:93–103.
36. Stelling J. Mathematical models in microbial systems 57. van Stiphout RGPM, Verhoog PJ, van Riel NAW, et al.
biology. Curr Opin Microbiol 2004;7:513–8. Computational model of excitable cell indicates ATP free
energy dynamics in response to calcium oscillations are
37. Araujo RP, Liotta LA. A control theoretic paradigm for cell undampened by cytosolic ATP buffers. IEE Proc Syst Biol
signaling networks: a simple complexity for a sensitive 2006;153:405–8.
robustness. Curr Opin Chem Biol 2006;10:81–7.
58. Westerhoff HV, Palsson BO. The evolution of
38. de Jong H. Modeling and simulation of genetic regulatory
molecular biology into systems biology. Nat Biotech 2004;
systems: a literature review. J Comput Biol 2002;9:67–103.
22:1249–52.
39. Alur R, Belta C, Kumar V, et al. Modeling and analyzing
59. Zhang Y, Rundell A. Comparative study of
biomolecular networks. Comput Sci Eng 2002;4:20–31.
parameter sensitivity analyses of the TCR-activated
40. Brandman O, Ferrell JE, Jr, Li R, et al. Interlinked fast and erk-MAPK signalling pathway. IEE Proc Syst Biol 2006;153:
slow positive feedback loops drive reliable cell decisions. 201–11.
Science 2005;310:496–8.
60. McKay MD, Beckman RJ, Conover WJ. A comparison of
41. Bornholdt S. Systems biology: less is more in modeling large three methods for selecting values of input variables in the
genetic networks. Science 2005;310:449–51. analysis of output from a computer code. Technometrics 1979;
42. Bentele M, Lavrik I, Ulrich M, etal. Mathematical modeling 21:239–45.
reveals threshold mechanism in CD95-induced apoptosis. 61. Blower SM, Dowlatabadi H. Sensitivity and uncertainty
J Cell Biol 2004;166:839–51. analysis of complex models of disease transmission: an HIV
43. Ljung L. System Identification Theory for the User, 2nd edn. model, as an example. Int Stat Rev/Revue Internationale de
Upper Saddle River: NJ, USA, 1999. Statistique 1994;62:229–43.
374 van Riel

62. Sobol IM. Global sensitivity indices for nonlinear mathe- 72. Mahadevan R, Edwards JS, Doyle FJ, III. Dynamic flux
matical models and their Monte Carlo estimates. Math balance analysis of diauxic growth in escherichia coli.
Comput Simul 2001;55:271–80. BiophysJ 2002;83:1331–40.
63. Saltelli A, Tarantola S, Chan KPS. A quantitative model- 73. van Riel NAW, Giuseppin MLF, Verrips CT. Dynamic
independent method for global sensitivity analysis of model optimal control of homeostasis: an integrative system
output. Technometrics 1999;41:39–56. approach for modeling of the central nitrogen metabolism
64. Frey CH, Patil SR. Identification and review of sensitivity in saccharomyces cerevisiae. Metabolic Eng 2000;2:49–68.
analysis methods. Risk Anal 2002;22:553–78. 74. van Riel NAW, Giuseppin MLF, ter Schure EG, et al.
65. Hidalgo ME, Ayesa E. Numerical and graphical description A structured, minimal parameter model of the central
nitrogen metabolism in saccharomyces cerevisiae: the
of the information matrix in calibration experiments for
prediction of the behaviour of mutants,. J Theor Biol 1998;
state-space models. Water Res 2001;35:3206–14.
191:397–414.
66. Li X, Feng D, Wong KP. A general algorithm for optimal
75. Bergman RN, Ider YZ, Bowden CR, et al. Quantitative
sampling schedule design in nuclear medicine imaging.
estimation of insulin sensitivity. Am J Physiol Endocrinol
Comput Meth Progr Biomed 2001;65:45–59.
Metab 1979;236:E667–77.
67. Mendes P, Kell D. Non-linear optimization of biochemical
76. Kitano H. Computational cellular dynamics: a network-

Downloaded from https://academic.oup.com/bib/article/7/4/364/184297 by guest on 03 March 2023

pathways: applications to metabolic engineering and param- physics integral. Nat Rev Mol Cell Biol 2006;7:163.
eter estimation. Bioinformatics 1998;14:869–83.
77. Faller D, Klingmuller U, Timmer J. Simulation methods for
68. Moles CG, Mendes P, Banga JR. Parameter estimation in optimal experimental design in systems biology. Simulation
biochemical pathways: a comparison of global optimization 2003;79:717–25.
methods. Genome Res 2003;13:2467–74.
78. Zak DE, Gonye GE, Schwaber JS, etal. Importance of input
69. Gadkar KG, Varner J, Doyle FJ, III Model identification of perturbations and stochastic gene expression in the reverse
signal transduction networks from data using a state engineering of genetic regulatory networks: insights from an
regulator problem. Syst Biol, IEE Proc 2005;2:17–30. identifiability analysis of an in silico network. Genome Res
70. Locke JCW, Millar AJ, Turner MS. Modelling genetic 2003;13:2396–405.
networks with noisy and varied experimental data: the 79. Weston A, Hood L. Systems biology, proteomics,
circadian clock in Arabidopsis thaliana. J Theor Biol 2005; and the future of health care: toward predictive, preventa-
234:383–93. tive, and personalized medicine. J Proteome Res 2004;3:
71. Amonlirdviman K, Khare NA, Tree DRP, et al. 179–96.
Mathematical modeling of planar cell polarity to understand 80. Sontag ED. Some new directions in control theory inspired
domineering nonautonomy. Science 2005;307:423–6. by systems biology. Syst Biol 2004;1:9–18.