Parameter estimation in systems biology models using spline approximation
- Choujun Zhan^{1}Email author and
- Lam F Yeung^{1}Email author
https://doi.org/10.1186/1752-0509-5-14
© Zhan and Yeung; licensee BioMed Central Ltd. 2011
Received: 18 March 2010
Accepted: 24 January 2011
Published: 24 January 2011
Abstract
Background
Mathematical models for revealing the dynamics and interactions properties of biological systems play an important role in computational systems biology. The inference of model parameter values from time-course data can be considered as a "reverse engineering" process and is still one of the most challenging tasks. Many parameter estimation methods have been developed but none of these methods is effective for all cases and can overwhelm all other approaches. Instead, various methods have their advantages and disadvantages. It is worth to develop parameter estimation methods which are robust against noise, efficient in computation and flexible enough to meet different constraints.
Results
Two parameter estimation methods of combining spline theory with Linear Programming (LP) and Nonlinear Programming (NLP) are developed. These methods remove the need for ODE solvers during the identification process. Our analysis shows that the augmented cost function surfaces used in the two proposed methods are smoother; which can ease the optima searching process and hence enhance the robustness and speed of the search algorithm. Moreover, the cores of our algorithms are LP and NLP based, which are flexible and consequently additional constraints can be embedded/removed easily. Eight system biology models are used for testing the proposed approaches. Our results confirm that the proposed methods are both efficient and robust.
Conclusions
The proposed approaches have general application to identify unknown parameter values of a wide range of systems biology models.
Background
In recent years, the rapid development of sophisticated experiment tools in molecular biology allows the acquisition of high qualitative time series data which can significantly improve the ability of revealing the complex dynamics and interactions of biological systems. Profiting from the rapid technological advances, more and more researchers from different disciplines can now utilize such observation data to establish mechanism-based models which can incorporate every possible detail and functioning of biological systems [1]. One common approach is to characterize the biological system with a set of Ordinary Differential Equations (ODEs) [2–7]. Generally, there are two major aspects of building an ODE model for a biological system from experimentally measured time series: (1) to determine the structure of the system through a set of suitable ODEs with unknown parameters; (2) to determine the unknown parameters of this ODE model. The identification of these unknown parameter with fixed model structure from observations is one of the central issues of computational systems biology [8]. This type of approach can be considered as a "reverse engineering process" [9–11].
The parameter estimation problem is generally formulated as an optimization problem that minimizes an objective function which represents the fitness of the model with respect to a set of experimental data [8, 12–17]. Two major optimization approaches are commonly adopted; the gradient-based nonlinear optimization method and the evolutionary based method. Also, simulations had shown that the simulated annealing (SA) method can offer promising results [18]. In [19], many deterministic and stochastic global optimization (GO) methods for parameter estimation were further compared using a three-step pathway model with noise free data assumption; the best result was given by the Stochastic Ranking Evolution Strategy (SRES) method. It is worth mentioning that, due to its simplicity in implementation, evolutionary algorithms, such as genetic algorithm and their variants, are extensively utilized for identifying unknown parameters of systems biology models [1, 11, 20–22]. However, most of these aforementioned approaches need a numerical ODE solver to perform the numerical integration for the underlining differential equations. Studies have revealed that more than 90% of the computation time is consumed in the ODE solver during the identification process [19]. In particular, for nonlinear dynamical systems with high-parameter-dimension, one trial usually consumes tens of hours or even days [10, 20, 21, 23]. Furthermore, the convergence property is aggravated by numerical integration failure, which is a major problem in the optimization process [11]. The computational burden can be relieved by reformulating the system involving differential equations into a system of algebraic equations [12, 15, 17, 24], which can be classified as "decomposition approaches". These decomposition approaches are widely employed in the parameter estimation of S-systems [7, 25]. The reliability of the decomposed methods depends on the accuracy of the "smooth" estimated derivatives and the states of the system. In practice, these data are subject to significant observation noise. Without proper pre-processing, the estimation faces the potential of the overfitting problem and hence the estimation can deviate badly from the "true" value [26, 27]. Regularization can be considered as a mathematical pre-processing on the measured noisy data set and be used to control the trade-off between the "roughness" of the solution and the infidelity of the data [28]. Since we are dealing with a known structured bio-system, the system model itself possesses a physical inertia and can serve as physical constraints which limit the system states within a set of possible trajectories. In this paper, the over-fitting problem can be relieved by embedding the model dynamics, the mass and energy balance constraints into our constrained optimization algorithms. Owing to the nonlinearity of systems biology models, the cost function to be minimized is complex and has multiple local minima. Minimization algorithms face the high possibility of getting trapped at local optima. For these reasons, the parameter estimation problem is still a bottleneck and a challenging task of computational analysis of systems biology [1, 11]. Until now, none of the parameter estimation methods is effective in all cases and can overwhelm all the other methods. Instead, various methods have their advantages and disadvantages. Consequently, it is worthy to develop acceptably "good enough" methods within a given tolerance and time frame.
For practical purpose, some essential issues should be taken into account when developing a parameter estimation method: first, the method must be "efficient" enough that a trial can be completed within a reasonable computation time; second, for biological systems, the observation data is often corrupted by high level of noise, which complicates the objective function surface and introduces unwanted additional local minima in the search space [29]. Hence, the approach should be robust subject to noise; third, it needs to be flexible enough for adding/removing physical constraints, such as model dynamics, the mass and energy balance constraints. Furthermore, the representative cost function should have less local minima so as to ease the optimization algorithm in converging to the global minima. In this paper, two parameter estimation methods of combining spline theory [28] with Linear Programming (LP) and Nonlinear Programming (NLP) are developed, respectively. These methods remove the need for an ODE solver. Our analysis exhibits that the cost function surfaces of the two proposed methods are smooth.
Moreover, the cores of our algorithms are LP and NLP based, which are very flexible and hence additional constraints can be embeded/removed easily. Eight systems biology models were used to test the proposed algorithms. Experimental results show that the proposed methods are both efficient and robust (see additional file 1 for details).
This paper is organized as follows: The preliminary problem formulation is given and the bottleneck of the problem is highlighted in the next section. Then, two parameter estimation methods surmounting those bottlenecks are presented. In section 3, two trials are given, a simple enzyme kinetic model and the mammalian G1/S transition network model, in order to illustrate the robustness and the effectiveness of these two proposed methods (more models and trial results are given in additional file 1 and 2). Finally, conclusions and discussions are given in section 4.
Methods
Parameter estimation problem of systems biology models
where x ϵ R^{ n } is the system's state vector (for example the concentrations of a process), θ ϵ R^{ k } is the system's parameter vector (for instance, the reaction rates), u(t) ϵ R^{ p } is system's input, y ϵ R^{ m } denotes the measured data subject to a Gaussian white noise η(t) ~ N(0, σ^{2}), and x_{0} is the initial state. f(·) is a set of nonlinear transition functions describing the dynamical properties of a biological system. Here, g(·) represents a measurement function. If all the states can be measured, the observer g(·) becomes an identity matrix. Otherwise, g(·) usually is a rectangular zero-one matrix with corresponding rows deleted (represent the immeasurable states) from the identity matrix I_{ n } .
P_{0} minimizes a cost function that measures the fitness of the model with respect to a given set of experiment data subjecting to a set of constraints, where $\widehat{\theta}\in {R}^{k}$ is the set of parameters to be estimated, ||·|| _{ l } denotes the l-norm with l > 0, ${\widehat{x}}_{0}$ is the estimated initial condition, $\widehat{x}\in {R}^{k}$ is the estimated system states ($\widehat{x}({t}_{j}|\widehat{\theta})$ represents the estimated variable at time t_{ j } with parameter $\widehat{\theta}$ and initial condition ${\widehat{x}}_{0}$), w_{ ij } are the weighting coefficients, $\widehat{y}$ is the estimated measured data. In some applications, additional constraints are introduced to impose special structural properties of a given system; they can be implemented in the form of the equality and inequality constraints C_{ eq } and C_{ ineq } (for instance the system performance and the mass balance constraints). Finally, θ_{L} and θ_{U} are simple structural constraints such as the parameter's upper/lower bounds (they can be part of the C_{ ineq } ).
For the NLP-P_{0} , the direct optimization methods, such as Newton type methods and many GO methods, require solving the nonlinear dynamic model (1) for $\widehat{x}$ in order to compute the cost function. The common method to estimate $\widehat{\dot{x}}({t}_{i})$ and $\widehat{x}({t}_{i})$ is using ODE solvers, which perform the numerical integration with $\widehat{\theta}$ fixed at each iteration [19]. During the process of identification, the integration has to be executed thousands, even millions of times. That is the main reason more than 90% of the time is consumed in the ODE solver [24] and the computation time spent on the P_{0} can be hours even days [10, 20]. Moreover, P_{0} is a nonlinear optimization problem subjecting to a set of linear and non-linear differential equation constraints. Hence, P_{0} is often multimodal (non-convex) and has many local minima. In a high-noise environment, the situation becomes more difficult. Consequently, P_{0} requires further manipulation in order to reduce the complexity so as to relieve the computation burden and also to avoid being trapped in local minima.
where ${\dot{b}}_{j}(t)={[{\dot{b}}_{0,{n}_{d}}(t),{\dot{b}}_{0,{n}_{d}}(t),\cdots ,{\dot{b}}_{{L}_{j}-{n}_{d}-2,{n}_{d}}(t)]}^{T}$ is the set of the derivatives of the basis functions. There are various types of splines suitable for this application, such as cubic spline, B-spline, uniform spline, nonuniform spline and interpolating spline. For more detail information about spline approximation theory, please refer to chapter (IX, XI, and XIV) in [28]. As B-spline is simple in formation and efficient for computation, it is adopted here. Our extensive tests have shown that uniform B-spline basis with $\frac{N}{3}\le {L}_{i}\le \frac{N}{2}$ produces good results. Hence, in this paper, unless otherwise indicated, the uniform B-spline basis with ${L}_{i}\approx \frac{N}{3}$ was used in the parameter identification process.
Next, two techniques based on spline for parameter estimation will be proposed: one is based on linear programming (LP) which is very efficient and can cover many special structured systems and the other one is based on NLP which is flexible and can cater for general system structures.
The LP Approach
where Φ(x) ϵ R^{n ×k}is a matrix and its elements are a function of the state x. Systems with structure (8) covers a large set of systems biology models, such as enzyme kinetic pathway model, RKIP pathway model, Iκ B-NF-κ B model TNFα-Mediated NF-κ B-signaling pathway model, irreversible inhibition of HIV proteinase model, Laub and Loomis model [2–4, 30]. In addition, these types of models are usually subject to the mass balance constraints which can be incorporated into the LP easily (It is demonstrated in the results section via the Enzyme kinetic model).
${A}_{eq}\xb7\widehat{x}={b}_{eq}$ stands for the equality constraints ${C}_{eq}(\widehat{x},\widehat{\theta})=0$. A_{ eq } is a constant matrix, b_{ eq } is a constant vector. It is found empirically that m = 2 and 0 ≤ λ ≤ 0.05 produce relatively good results. Hence, the parameters m = 2, and λ = {0, 0.01, 0.03} corresponding to a noise level {0%, 5%, 10%}, respectively, were used in this study. Then, the "smooth" estimated state $\widehat{x}$ can be generated by the B-spline approximation (5).
It is a Linear Programming (LP) problem with variable {α, θ}, which is a convex problem with a wealth of fast and efficient routines available [31, 32].
Combine spline theory and NLP
Note that the constraint-(i) of (2) has been replaced by constraints (i)-(iii) of P_{3}. Then, NLP-P_{0}-(2) with differential-algebraic constraints turns into NLP-P_{3}-(14) with only algebraic equation constraints. Hence, P_{3} does not require ODE solvers, which eases the computation burden (as shown in Examples). In contrast to the the decomposition methods [12, 15], which divide the estimation of the system states (and its derivative) and the parameter estimation into two separate steps, P_{3} computes the estimated states (and its derivative) and parameter values at the same time. Note that constraint (iii) of P_{3} governs the estimated state (and its derivative) so as to ensure these estimates belong to the trajectory $\widehat{x}(t|\widehat{\theta})$, which is a solution of system (1). Thus, the system model itself serves as a filter performing regularization. Hence, the overfitting problem can be relieved (see additional file 2 for details).
For a non-linear system, the Lagrangian of (2), $L(\widehat{\theta},{\widehat{X}}_{0})$, is an implicit function of $\{\widehat{\theta},{\widehat{x}}_{0}\}$[31]. However, many traditional optimization algorithms require the derivative $\partial L/\partial \widehat{\theta}$ during the optimization process. As $L(\widehat{\theta},{\widehat{x}}_{0})$ is an implicit function of $\widehat{\theta}$, $\partial L/\partial \widehat{\theta}$ cannot be obtained directly, but has to be computed via approximation methods [16], which makes the algorithm unreliable. For P_{3}, the Lagrangian $L(\widehat{\theta},p)$ now consists of simple algebraic constraints. Thus, $\partial L/\partial \widehat{\theta}$ and ∂L/∂p are explicit functions of $\widehat{\theta}$ and p. In conclusion, many of the aforementioned difficulties can be reduced. P_{3} can be solved by a number of optimization approaches; either via evolution type algorithms, such as genetic algorithm (GA), simulated annealing (SA) and etc, or via traditional NLP algorithms, such as sequential quadratic programming(SQP), sequential penalty function, the trust region approach and etc [33, 34].
Results
Two biological system models, a simple enzyme kinetic model and the mammalian G1/S transition network model, are chosen as benchmarks for evaluating the performance of P_{2} and P_{3} respectively.
Enzyme kinetic model
where ${A}_{eq}=\left[\begin{array}{cccc}0& 1& 1& 0\\ 1& 0& 1& 1\end{array}\right]$ and ${b}_{eq}=\left[\begin{array}{c}{x}_{2}({t}_{0})+{x}_{3}({t}_{0})\\ {x}_{1}({t}_{0})+{x}_{3}({t}_{0})+{x}_{4}({t}_{0})\end{array}\right]$.
where ${\widehat{x}}_{i}({t}_{j})$ is the estimated time-course at time t_{ j } of a state variable x_{ i } , and x_{ i } (t_{ j } ) represents the "true" time-courses without noise at time t_{ j } . Note that smaller RSE J reflects better estimation. In order to obtain a statistical result on the quality of the estimation, 5,000 trials were performed. At each trial an estimated $\widehat{\theta}$ is computed using P_{2}. Then, the mean estimation and standard deviation were deduced. The computation was very efficient and only took a few seconds for one estimation trial.
Statistical results of parameter estimation of enzyme kinetic model.
Nominal Value | Mean estimation ± standard deviation | |||
---|---|---|---|---|
Noise level: 0% | Noise level: 5% | Noise level: 10% | ||
k _{1} | 0.18 | 0.1796 ± 33.5081e-6 | 0.1794 ± 0.0008 | 0.1808 ± 0.0025 |
k _{2} | 0.20 | 0.1993 ± 108.4891e-6 | 0.1968 ± 0.0031 | 0.1963 ± 0.0106 |
k _{3} | 0.23 | 0.2300 ± 1.5946e-6 | 0.2326 ± 0.0001 | 0.2345 ± 0.0004 |
J | 6.8620e-8 ± 7.7337e-8 | 1.0996e-4 ± 9.3270e-5 | 3.0328e-4 ± 3.2160e-4 |
The mammalian G1/S transition network model
Definition of Variables for G1/S Transition Model.
Symbol | x _{1} | x _{2} | x _{3} | x _{4} | x _{5} | x _{6} | x _{7} | x _{8} | x _{9} |
---|---|---|---|---|---|---|---|---|---|
Acronym | pRB | E2F1 | CycD _{ i } | CycD _{ a } | AP - 1 | pRB _{ p } | pRB _{ p } p | CycE _{ i } | CycE _{ a } |
Results of parameter estimation of the mammalian G1/S transition network model.
Parameters | Nominal value | Estimated parameters | |||
---|---|---|---|---|---|
Noise level 0% | Noise level: 2.5% | Noise level: 5% | Noise level: 10% | ||
k _{1} | 1 | 0.9957 | 1.0150 | 1.1105 | 1.6037 |
k _{2} | 1.6 | 1.5989 | 1.4138 | 1.5187 | 1.0315 |
k _{3} | 0.05 | 0.0500 | 0.0528 | 0.0392 | 0.0381 |
k _{16} | 0.4 | 0.4002 | 0.4440 | 0.3959 | 0.9331 |
k _{34} | 0.04 | 0.0400 | 0.0414 | 0.0337 | 0.0215 |
k _{43} | 0.01 | 0.0100 | 0.0142 | 0.0090 | 1.45e-10 |
k _{61} | 0.3 | 0.2985 | 0.3432 | 0.2847 | 0.8185 |
k _{67} | 0.7 | 0.6999 | 0.4535 | 1.3974 | 1.3108 |
k _{76} | 0.1 | 0.0999 | 0.0457 | 0.2446 | 0.1845 |
k _{23} | 0.3 | 0.1219 | 0.4134 | 0.6132 | 0.5579 |
k _{25} | 0.9 | 0.1785 | 0.7063 | 0.8291 | 0.7874 |
k _{28} | 0.06 | 0.0601 | 0.0669 | 0.0222 | 0.0198 |
k _{39} | 0.07 | 0.0700 | 0.0549 | 0.0520 | 0.0334 |
k _{96} | 0.01 | 0.0100 | 0.0441 | 0.0002 | 4.55e-14 |
a | 0.04 | 0.0400 | 0.1257 | 0.1260 | 0.1265 |
J _{11} | 0.5 | 0.4992 | 0.5612 | 0.4252 | 0.6523 |
J _{12} | 5 | 5.0025 | 4.8940 | 4.6892 | 5.4021 |
J _{15} | 0.001 | 0.0051 | 0.0011 | 0.0010 | 0.0011 |
J _{18} | 0.6 | 0.5990 | 0.7253 | 0.8014 | 1.1290 |
J _{61} | 5 | 5.2581 | 4.1474 | 6.4585 | 7.2003 |
J _{62} | 8 | 8.0088 | 29.734 | 39.403 | 41.408 |
J _{65} | 6 | 5.9222 | 8.7804 | 9.3474 | 7.8076 |
J _{68} | 7 | 6.9916 | 31.979 | 25.125 | 36.795 |
J _{13} | 0.002 | 0.0050 | 0.0013 | 0.0016 | 2.61e-14 |
J _{63} | 2 | 1.9740 | 1.4726 | 0.4203 | 19.871 |
K _{m1} | 0.5 | 0.4905 | 0.5267 | 0.5601 | 0.0410 |
K _{m2} | 4 | 3.9985 | 4.0482 | 4.1061 | 3.8495 |
K _{m4} | 0.3 | 0.2999 | 0.2838 | 0.2735 | 0.2338 |
K _{m9} | 0.005 | 0.0054 | 3.69e-5 | 2.03e-5 | 3.88e-6 |
K _{ p } | 0.05 | 0.0499 | 0.0452 | 0.0496 | 0.0311 |
ϕ _{1} | 0.005 | 0.0044 | 0.0057 | 0.0041 | 0.0073 |
ϕ _{2} | 0.1 | 0.0999 | 0.0920 | 0.0983 | 0.0693 |
ϕ _{3} | 0.023 | 0.0230 | 0.0261 | 0.0164 | 0.0152 |
ϕ _{4} | 0.03 | 0.0300 | 0.0279 | 0.0253 | 0.0218 |
ϕ _{5} | 0.01 | 0.0100 | 0.0098 | 0.0101 | 0.0101 |
ϕ _{6} | 0.06 | 0.0606 | 0.0627 | 0.0608 | 0.1518 |
ϕ _{7} | 0.04 | 0.0401 | 0.0436 | 0.0404 | 0.0788 |
ϕ _{8} | 0.06 | 0.0600 | 0.1546 | 0.0024 | 0.0260 |
ϕ _{9} | 0.05 | 0.0500 | 0.0025 | 0.0439 | 0.0276 |
J | 7.5399e-6e | 0.0005 | 0.0009 | 0.0025 |
Trials were performed using Matlab-7. The main reason to use Matlab is that it is a convenient environment to visualize all the information arising from the optimization runs of the solver, evaluate new algorithms and modify existing algorithms. In contrast to the convenience, it is worth mentioning that Matlab programs usually are one order of magnitude (10 times or more) slower than equivalent compiled Fortan or C codes [19]. This is the major drawbacks of carrying programs out with Matlab. However, even in this situation, the performance of the proposed methods is acceptable.
For fair comparison, we also used the SRES algorithm to solve the same parameter estimation problem in the same searching region, but using NLP-P_{0} with differential algebraic constraints as cost function. In this condition, after running 1 day, the algorithm failed to produce a set of parameters that can produce reasonable simulation result. We further reduced the searching region to [0, 3θ ] and used noise free data, but the estimation result was still not good and the RSE J is larger than 10.
Furthermore, P_{3} only involves algebraic equations as objective function and constraints. These properties make the NLP-P_{3} easier to solve.
Discussion and Conclusion
In this paper, two parameter estimation methods based on spline theory are proposed. One aims at a narrower class of systems which is linear in parameters; however, it can cover many commonly found biological systems. The benefit is that the estimation problem can be transformed in an LP sub-algorithm which are fast and robust. Additional linear constraints can be embedded relative easily. For general systems, the problem is solved by an NLP with algebraic constraints, which is more computationally demanding.
A simple enzyme kinetic model and the mammalian G1/S transition network model were used as benchmarks to evaluate the performance of the two proposed methods. We illustrate the usefulness with more examples in additional files 1 but these do not remotely cover all the conditions.
During the simulation of the mammalian G1/S transition network model, we found that the estimated parameter set Φ _{ A } ≡ {k_{1}, k_{2}, k_{ p }, J_{11}, J_{12}, K_{m 1}, K_{m 2}, ϕ_{1}, ϕ_{2}} were well within the respective nominal values. While the set Φ_{ B }≡ {J_{61}, J_{62}, J_{63}, J_{65}, J_{68}} were far from their nominal values. However, the time-series produced by the estimated model were very similar to the original data. This phenomena reveals that some parameter values are insensitive in the searching region. Interestingly, we find that the "sensitive" or "easily identified" parameters set Φ_{ A }are also the parameters of the double-activator-inhibitor module of the antagonistic players E2F/DP and pRB, which makes up the core unit of the G1/S transition model [5]. This phenomenon may imply that the parameter values of the core module are sensitive and easy to identify. In contrast, the parameters set Φ_{ B }seems to be insensitive, which may reflect that pRG_{ p } (x_{6}) is not a key element of the total system. However, to identify which parameter values or variables are important, a sensitivity analysis is needed [38], which is another important topic in systems biology and deserves a more detailed study. This sensitivity analysis is a pre-process for isolating those states and parameters which are sensitive in order to reduce the dimension of the system model and to improve the numerical stability for the core estimation problem.
- 1.
High quality experiment data is essential for identifying accurate biology systems. When the experiment data is corrupted with high level noise, it needs more experimental data. If an insufficient amount of time-series data is given as observed profiles, the high degree-freedom of systems biology models ensures that many candidate solutions will be found.
- 2.
Perform a sensitivity analysis and identifiability analysis before the identification phase [36–38].
- 3.
For systems models with insensitive or non-identifiable parameters, the search may lead to a solution where some parameters can have large deviation, but still produce satisfactory system responses. This problem can be partly relieved by introducing auxiliary information (additional constraints such as shrinking the searching region) of the model into the algorithm. However, it remains difficult to be solved completely by improving parameter estimation strategy. It indicates that researchers should focus on predictions rather than on accurately estimating every parameter.
Although the proposed algorithms are fast and robust, there is certainly room for improvement: for method 1, it is not general enough to catch every case; for method 2, the price for the simplicity and generality is at the expansion of the optimization variable dimension. Under high noise condition, method 2 is still not robust enough. At the moment, the testing is based on Matlab which is much slower than native codes produced by C, Fortran, etc, however the conversion is straight forward. Currently, many high-speed computation engines are available that make use of parallelism, for instance multi-cluster engines, array-processing engines etc. Hence, one possible way is developed algorithm on these high-speed computation engines environment. Another possible way is developing hybrid algorithms to incorporate elements from evolution algorithms such as GA, SA and PSO. In this paper, we have considered the parameter estimation problem with known structure. However, it is easy to expand our method to structure identification by introducing an additional penalty term to the objective function [39].
Declarations
Acknowledgements
The work is supported by CERG Grant 9041393 (CityU 123808). We would like to express our appreciation for Dr. Robin S. Bradbeer and Dr. K. T. Ko for their fruitful suggestions.
Authors’ Affiliations
References
- Chang WC, Li CW, Chen BS: Quantitative inference of dynamic regulatory pathways via microarray data. BMC Bioinformatics. 2005, 6: 1-19. 10.1186/1471-2105-6-44View ArticleGoogle Scholar
- Cho KH, Shin SY, Kim HW, Wolkenhauer O, McFerran B, Kolch W: Mathematical modeling of the influence of RKIP on the ERK signaling pathway. Lecture Notes in Computer Science. 2003, 2602: 127-131. full_text. full_text full_textView ArticleGoogle Scholar
- Cho KH, Shin SY, Lee HW, Wolkenhauer O: Investigations into the analysis and modeling of the TNFα -Mediated NF-κ B-Signaling pathway. Genome Res. 2003, 13 (11): 2413-2422. 10.1101/gr.1195703PubMed CentralView ArticlePubMedGoogle Scholar
- Hoffmann A, Levchenko A, Scott ML, Baltimore D: The IkB-NF-kB Signalling module: temporal control and selective gene activation. Science. 2002, 298 (5596): 1241-1245. 10.1126/science.1071914View ArticlePubMedGoogle Scholar
- Swat M, Kel A, Herzel H: Bifurcation analysis of the regulatory modules of the mammalian G1/S transition. Bioinformatics. 2004, 20 (10): 1506-1511. 10.1093/bioinformatics/bth110View ArticlePubMedGoogle Scholar
- Tyson JJ, Chen K, Novak B: Network dynamics and cell physiology. Nature Reviews Molecular Cell Biology. 2001, 2: 908-916. 10.1038/35103078View ArticlePubMedGoogle Scholar
- Voit EO: Computational analysis of biochemical systems, a practical guide for biochemists and molecular biologists. 2000, Cambridge: Cambridge University Press,Google Scholar
- Goel G, Chou IC, Voit EO: System estimation from metabolic time-series data. Bioinformatics. 2008, 24 (21): 2505-2511. 10.1093/bioinformatics/btn470PubMed CentralView ArticlePubMedGoogle Scholar
- Ashyraliyev M, Fomekong-Nanfack Y, Kaandorp JA, Blom JG: Systems biology: parameter estimation for biochemical models. FEBS J. 2009, 276 (4): 886-902. 10.1111/j.1742-4658.2008.06844.xView ArticlePubMedGoogle Scholar
- Kikuchi S, Tominaga D, Arita M, Takahashi K, Tomita M: Dynamic modeling of genetic networks using genetic algorithm and S-system. Bioinformatics. 2003, 19 (5): 643-650. 10.1093/bioinformatics/btg027View ArticlePubMedGoogle Scholar
- Tsai KY, Wang FS: Evolutionary optimization with data, collocation for reverse engineering of biological networks. Bioinformatics. 2005, 21 (7): 1180-1188. 10.1093/bioinformatics/bti099View ArticlePubMedGoogle Scholar
- Chou IC, Martens H, Voit EO: Parameter estimation in biochemical systems models with alternating regression. Theor Biol Med Model. 2006, 19: 3-25.Google Scholar
- Gadkar KG, Gunawan R, Doyle F: Iterative approach to model identification of biological networks. BMC Bioinformatics. 2005, 6: 155- 10.1186/1471-2105-6-155PubMed CentralView ArticlePubMedGoogle Scholar
- Gonzalez OR, Küer C, Jung K, Naval PCJ, Mendoza E: Parameter estimation using Simulated Annealing for S-system models of biochemical networks. Bioinformatics. 2007, 23 (4): 480-486. 10.1093/bioinformatics/btl522.View ArticlePubMedGoogle Scholar
- Lall R, Voit EO: Parameter estimation in modulated, unbranched reaction chains within biochemical systems. Comput Biol Chem. 2005, 29 (5): 309-318. 10.1016/j.compbiolchem.2005.08.001View ArticlePubMedGoogle Scholar
- Peifer M, Timmer J: Parameter estimation in ordinary differential equations for biochemical processes using the method of multiple shooting. IET Syst Biol. 2007, 1 (2): 78-88. 10.1049/iet-syb:20060067View ArticlePubMedGoogle Scholar
- Veflingstad SR, Almeida J, Voit EO: Priming nonlinear searches for pathway identification. Theor Biol Med Model. 2004, 1: 8- 10.1186/1742-4682-1-8PubMed CentralView ArticlePubMedGoogle Scholar
- Mendes P, Kell D: Non-linear optimization of biochemical pathways: application to metabolic engineering and parameter estimation. Bioinformatics. 1998, 14 (10): 869-883. 10.1093/bioinformatics/14.10.869View ArticlePubMedGoogle Scholar
- Moles CG, Mendes P, Banga JR: Parameter estimation in biochemical pathways: a comparison of global optimization methods. Genome Res. 2003, 13 (11): 2467-2474. 10.1101/gr.1262503PubMed CentralView ArticlePubMedGoogle Scholar
- Cho DY, Cho KH, Zhang BT: Identification of biochemical networks by S-tree based genetic programming. Bioinformatics. 2006, 22 (13): 1631-1640. 10.1093/bioinformatics/btl122View ArticlePubMedGoogle Scholar
- Koh G, Teong HF, Cléent MV, Hsu D, Thiagarajan PS: A decompositional approach to parameter estimation in pathway modeling: a case study of the Akt and MAPK pathways and their crosstalk. Bioinformatics. 2006, 22 (24): 271-280. 10.1093/bioinformatics/btl264.View ArticleGoogle Scholar
- Kimura S, Ide K, Kashihara A, Kano M, Hatakeyama M, Masui R, Nakagawa N, Yokoyama S, Kuramitsu S, A K: Inference of S-system models of genetic networks using a cooperative coevolutionary algorithm. Bioinformatics. 2005, 21 (7): 1154-1163. 10.1093/bioinformatics/bti071View ArticlePubMedGoogle Scholar
- Polisetty PK, Voit EO, Gatzke EP: Identification of metabolic system parameters using global optimization methods. Theor Biol Med Model. 2006, 27 (3): 4-10.1186/1742-4682-3-4.View ArticleGoogle Scholar
- Voit EO, Almeida J: Decoupling dynamical systems for pathway identification from metabolic profiles. Bioinformatics. 2004, 20 (11): 1670-1681. 10.1093/bioinformatics/bth140View ArticlePubMedGoogle Scholar
- Savageau MA: Biochemical Systems Analysis: a study of Function and Design in Molecular Biology. 1976, Reading, MA: Addison-Wesley, Reading,Google Scholar
- Craven P, Wahba G: Smoothing noisy data with spline functions estimating the correct degree of smoothing by the method of generalized cross validation. Numer Math. 1979, 31: 377-403. 10.1007/BF01404567.View ArticleGoogle Scholar
- Tetko IV, Livingstone DJ, Luik AI: Neural network studies. 1. Comparison of overfitting and overtraining. J Chem Inf Comput Sci. 1995, 35: 826-833.View ArticleGoogle Scholar
- Boor D: A practical guide to splines. 1987, New York: Springer,Google Scholar
- Voss HU, Timmer J, Kurths J: Nonlinear dynamical system identification from uncertain and indirect measurements. International Journal of Bifurcation and Chaos. 2004, 14 (6): 1905-1933. 10.1142/S0218127404010345.View ArticleGoogle Scholar
- Kuzmic P: Program DYNAFIT for the analysis of enzyme kinetic data: application to HIV proteinase. Ana Biochem. 1996, 237: 260-273. 10.1006/abio.1996.0238.View ArticleGoogle Scholar
- Boyd SP, Vandenberghe L: Convex Optimization. 2003, Cambridge, U.K.: Cambridge University Press,Google Scholar
- Winston WL: Operations research: applications and algorithms. 1994, Belmont, CA: Duxbury Press,Google Scholar
- Fletcher RE: Numerical experiments with an exact penalty function method. Nonlinear Programming. 1981, 4: 99-129.Google Scholar
- Powell MJD: A Fast Algorithm for Nonlinearly Constrained Optimization Calculations. Numerical Analysis. 1978, 27: 630-Google Scholar
- Runarsson TP, Yao X: Stochastic ranking for constrained evolutionary optimization. IEEE Trans Evol Comput. 2000, 4: 284-294. 10.1109/4235.873238.View ArticleGoogle Scholar
- Raue A, Kreutz C, Maiwald T, J B, Schilling M, Klingmuller U, Timmer J: Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics. 2009, 25 (15): 1923-1929. 10.1093/bioinformatics/btp358View ArticlePubMedGoogle Scholar
- Hengl S, Kreutz C, Timmer J, Maiwald T: Data-based identifiability analysis of non-linear dynamical models. Bioinformatics. 2007, 23 (19): 2612-2618. 10.1093/bioinformatics/btm382View ArticlePubMedGoogle Scholar
- Yue H, Brown M, Knowles J, Wang H, Broomhead DS, Kell DB: Insights into the behaviour of systems biology models from dynamic sensitivity and identifiability analysis: a case study of an NF-kB signalling pathway. Mol Biosyst. 2006, 2 (12): 640-649. 10.1039/b609442bView ArticlePubMedGoogle Scholar
- Liu PK, Wang FS: Inference of biochemical network models in S-system using multiobjective optimization approach. Bioinformatics. 2008, 24 (8): 1085-1092. 10.1093/bioinformatics/btn075View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.