- Proceedings
- Open Access
A CoD-based stationary control policy for intervening in large gene regulatory networks
- Noushin Ghaffari^{1}Email author,
- Ivan Ivanov^{2},
- Xiaoning Qian^{3} and
- Edward R Dougherty^{1, 4}
https://doi.org/10.1186/1471-2105-12-S10-S10
© Ghaffari et al; licensee BioMed Central Ltd. 2011
Published: 18 October 2011
Abstract
Background
One of the most important goals of the mathematical modeling of gene regulatory networks is to alter their behavior toward desirable phenotypes. Therapeutic techniques are derived for intervention in terms of stationary control policies. In large networks, it becomes computationally burdensome to derive an optimal control policy. To overcome this problem, greedy intervention approaches based on the concept of the Mean First Passage Time or the steady-state probability mass of the network states were previously proposed. Another possible approach is to use reduction mappings to compress the network and develop control policies on its reduced version. However, such mappings lead to loss of information and require an induction step when designing the control policy for the original network.
Results
In this paper, we propose a novel solution, CoD-CP, for designing intervention policies for large Boolean networks. The new method utilizes the Coefficient of Determination (CoD) and the Steady-State Distribution (SSD) of the model. The main advantage of CoD-CP in comparison with the previously proposed methods is that it does not require any compression of the original model, and thus can be directly designed on large networks. The simulation studies on small synthetic networks shows that CoD-CP performs comparable to previously proposed greedy policies that were induced from the compressed versions of the networks. Furthermore, on a large 17-gene gastrointestinal cancer network, CoD-CP outperforms other two available greedy techniques, which is precisely the kind of case for which CoD-CP has been developed. Finally, our experiments show that CoD-CP is robust with respect to the attractor structure of the model.
Conclusions
The newly proposed CoD-CP provides an attractive alternative for intervening large networks where other available greedy methods require size reduction on the network and an extra induction step before designing a control policy.
Keywords
Introduction
A key purpose of modeling gene regulation via gene regulatory networks (GRNs) is to derive strategies to shift long-run cell behavior towards desirable phenotypes. To date, the majority of the research regarding intervention in GRNs has been carried out in the context of probabilistic Boolean networks (PBNs) [1]. Assuming random gene perturbation in a PBN, the associated Markov chain is ergodic, and thus it possesses a steady-state distribution (SSD), and (from a theoretical standpoint) one can always change the long-run behavior using an optimal control policy derived via dynamic programming [2, 3]. In practice, however, the computational requirements of dynamic programming limit this approach to small networks [4, 5]. As an alternative to such optimal intervention, greedy control approaches using mean-first-passage time (MFPT-CP algorithm) or the steady-state distribution directly (SSD-CP algorithm) have been proposed (CP denoting control policy) [6, 7]; nonetheless, these algorithms have their own computational issues owing to their need to use the state transition matrix (STM) of the Markov chain. To overcome the computational problems associated with the design of control policies for larger PBNs, previous studies have proposed reduction mappings that either delete genes [8] or states [9]. Deletion of network components compresses large networks, but at the cost of information loss. Furthermore, reduction mappings themselves can be computationally demanding [8, 9].
The control approach taken in this paper circumvents many of the computational impediments of previous methods by basing its intervention strategy directly on inter-predictability among genes. Referring to a gene that characterizes a particular phenotype as a Target (T) gene and a gene used to alter the long-run behavior of the network by controlling the expression of T as a Control (C) gene, the method proposed herein relies on the predictive power of a small group of genes, which includes the control gene, and designs a stationary control policy that alters the steady-state distribution of the model. The algorithm is designed for the specific class of networks where there is a path from the control to the target gene – an assumption having a natural interpretation in terms of the biochemical regulatory pathways present in cells. Our method simplifies the procedure of designing the stationary control policy and eliminates the need to have a complete knowledge about the STM. Most importantly, the new algorithm can be used to design stationary control policy directly on large networks without deleting any genes/states. It only requires knowledge about the SSD of the network which can be estimated without inferring the STM. The coefficient of determination (CoD) is used for measuring the power of gene interactions [10]. Thus, our new algorithm is optimized for and performs especially well on network models that are inferred from data using CoD-based approaches, e.g. the well-known seed-growing algorithm [11]. The proposed algorithm, called CoD-CP because the CoD is the main tool, uses the marginal probabilities of the individual genes obtained from the steady-state distribution of the network to calculate the CoDs.
The most important advantage of the proposed CoD-CP is that it can be designed on networks with many genes, and without any compression of the model. All of the previously proposed methods for working with large GRNs, e.g. CoD-Reduce[8] or state reduction [9], require ‘deletion’ of network components to achieve a compressed model, which allows for the design of the control policy. An induction step is then required in order to induce those control policies back to the original networks. In this paper, we propose a new approach, which designs control policies directly on the original network and requires neither reduction/compression nor induction.
We performed a series of simulation studies to validate CoD-CP performance. Our experiments show that in small networks, where it is possible to derive the currently available greedy MFPT-CP [6] and SSD-CP [7] policies, CoD-CP achieves a similar performance. Most importantly, when the size of the network is large and MFPT-CP or SSD-CP cannot be designed directly on the original model, CoD-CP is easily constructed and applied to the network without any reduction mappings and induction of the control policy from the reduced network back to the original model. Section describes our simulations results. When the network is large, a reduction step is needed before designing the MFPT-CP or SSD-CP. In these cases, CoD-CP can be designed directly on the large networks and performs better than the induced MFPT-CP and SSD-CP on average for networks with singleton attractors only or models where cyclic attractors are allowed. We examined CoD-CP performance for two different perturbation probabilities and the results show consistent patterns. Furthermore, we examined the performance of the three algorithms on a 17-gene gastrointestinal cancer network derived from microarray data. CoD-CP designed on that model network outperforms the stationary MFPT-CP and SSD-CP policies induced from the reduced versions of the 17-gene model. Thus, our new approach provides an attractive alternative to the methods that require network reduction and an extra induction step before designing a control policy.
Background
Boolean networks
Coefficient of determination (CoD)
[10]. The CoD can be used to measure the strength of the connection between a target gene and its predictors and has been used since the early days of DNA microarray analysis to characterize the nonlinear multivariate interactions between genes [14]. More recently, CoD was used to characterize canalizing genes [15] and contextual genomic regulation [16]. We have restricted ourselves to the Boolean case, thereby arriving at the preceding representations of ε_{ opt }(Y,X) and ε_{0}(Y); however, the basic definition for CoD_{ X }(Y) is not so restricted [10].
MFPT control policy (MFPT-CP)
Optimal intervention is usually formulated as an optimal stochastic control problem [4]. We focus on intervention via a single control gene c, and stationary control policies µ_{ c } : S → {0,1} based on c. The values 0/1 are interpreted as off/on for the application of the control: 1 meaning that the current value of c is flipped, and 0 meaning that no control is applied.
where e denotes the vector of dimension 2^{ n }^{– 1} with all of its co-ordinates equal to 1.
To understand the intuition behind the MFPT-CP algorithm it is important to notice that, because the control gene c is different from the target gene, every state s belongs to the same class of states, D or U, as its flipped state . With this in mind, if a desirable state s reaches U on average faster than , it is reasonable to apply control and start the next network transition from its flipped state . Thus, the design of the stationary MFPT-CP is based on the differences and . The MFPT-CP algorithm uses a tuning parameter γ > 0, and these differences are compared to the value of γ, which is related to the cost of applying control. For example, γ is set to a larger value when the ratio of the cost of control to the cost of the undesirable states is higher, the intent being to apply the control less frequently [6].
The MFPT concept could be used in two different ways to design the intervention strategy. The first approach is called “model-dependent” and needs the state transition matrix of the Markov Chain. The time-course measurements can be used to estimate the transition probabilities for all states. Then the STM is used to find the K_{ U } and K_{ D } vectors to design the control policy. In the second approach, called “model-free,” the MFPTs are directly estimated from the time-course data and the inference of the STM is skipped. In this paper we focus on the model-dependent MFPT-CP.
SSD control policy (SSD-CP)
where β^{ T } = b^{ T }Z and e(k) is the elementary vector with a 1 in the k th position and 0s elsewhere [17–19]. To define the SSD-CP policy let be the flipped state (with respect to control gene c) corresponding to state s (as with MFPT-CP). Let π_{ U } be the original steady-state mass of the undesirable states and let and denote the steady-state masses of the undesirable states resulting from altering the original state transition matrix by changing the starting state for the next transition from s to and from to s, respectively. The SSD-CP policy is defined on pairs of states, s and , in the following manner: if both and are larger than π_{ U }, then control is applied to neither; otherwise, if , then control is applied to s, and if , then control is applied to .
Two step design of control policy: reduction followed by induction
The derivation of the optimal or greedy control policies becomes infeasible as the number of genes in the GRN increases. As a solution, deleting the genes is proposed by methods outlined in [8]. The idea is to delete genes sequentially until the size of the network is small enough for designing the control policy. Because the dimension of the control policy designed on the reduced network is not compatible with the original network, it is necessary to induce the control policy from the reduced network to the original one. The best candidate gene for deletion is selected by an algorithm that measures strength of gene-connectivity using the CoD. Genes not predicting any other genes or being predicted by any other genes are called constant genes and are the first choice for deletion. If there are not any constant genes, then the gene that has minimum CoD for predicting the target gene is selected as the best candidate, d, for deletion. After selecting d, a reduction mapping is used to define the transition rules for states in the reduced network [20]. The design of the reduction mapping is based on the notion of a selection policy[8]. A selection policy ν^{ d } corresponding to the deleted gene d is a 2^{ n } dimensional vector, ν^{ d } ∈ {0, 1}^{2}^{ n }, indexed by the states of S and having components equal to 1 at exactly one of the positions corresponding to each pair , s ∈ S . For each gene d there are 2^{2 – n} different selection policies.
Since finding the optimal selection policy is computationally impossible in large GRNs, an heuristic approach is proposed by [8]: if either state s or is an attractor, then the attractor state is chosen to determine the function structure, but if neither is an attractor, then the transitions of the state possessing larger steady-state probability mass are kept as transitions for the reduced state.
for any z_{1}, …, z_{ n–m } ∈ {0,1}.
Proposed methodology
This section describes our new algorithm, CoD-CP. The algorithm takes advantage of the predictive power of triplets of genes that include the control gene to predict the expression of the target gene with a small estimated error. To achieve the best performance of the algorithm, it is necessary to have a direct connection or a path from the control gene to the target gene in the regulatory network. The algorithm uses the CoD to measure that predictive power and to design a control policy.
CoD-CP is a greedy technique for designing a stationary control policy. The target gene defines the phenotype and divides states into two mutually disjoint sets, D (desirable) and U (undesirable). The gene with the most predictive power over the target gene T among the genes connected with a path to T is used as the control gene C. The goal of the algorithm is to increase the total probability mass of desirable states in the long-run by controlling C.
MAXCPD Table: the first three columns represent the binary combinations of the three MAXCOD genes. The last two columns are filled by summing up the SSD probabilities of states in each corresponding block.
MAXCOD | T | ||||
---|---|---|---|---|---|
C | Predictor 1 | Predictor 2 | 0 | 1 | |
row 1 | 0 | 0 | 0 | P _{10} | P _{11} |
row 2 | 0 | 0 | 1 | P _{20} | P _{21} |
row 3 | 0 | 1 | 0 | P _{30} | P _{31} |
row 4 | 0 | 1 | 1 | P _{40} | P _{41} |
row 5 | 1 | 0 | 0 | P _{50} | P _{51} |
row 6 | 1 | 0 | 1 | P _{60} | P _{61} |
row 7 | 1 | 1 | 0 | P _{70} | P _{71} |
row 8 | 1 | 1 | 1 | P _{80} | P _{81} |
Example 1 , part a: This example explains the entries of the MAXCPD table using a 7-gene network with 128 states. Without loss of generality, assume that x_{1} and x_{2} are the T and C genes, respectively, and x_{1} = 0 defines desirable states. After examining all the triples, MAXCOD is found to be {x_{2}, x_{3}, x_{4}}, which has maximum CoD for predicting x_{1}. The first three columns of the MAXCPD table contain 8 binary combinations of x_{2}, x_{3} and x_{4}, as table 1 shows. The last two columns of the table contain the summation of the SSD probabilities of the states with common value for MAXCOD genes. The only difference in columns four and five is the value of the T gene. The size of each block of states is 2^{ n }^{– 4} = 2^{3} = 8. The first block is Block(1) = {0000000, 0000001, 0000010, 0000011, 0000100, 0000101, 0000110, 0000111}, where all have {x_{2}, x_{3}, x_{4}} = 000 and x_{1} = 0. The second block is Block(2) = {1000000, 1000001, 1000010, 1000011, 1000100, 1000101, 1000110, 1000111}, where {x_{2}, x_{3}, x_{4}} = 000 and x_{1} = 1. Each entry of the forth and fifth columns of the CPD table are represented by P_{ ij }, where i ∈ {1, …, 8} represents a row and j ∈ {0,1} is the T value. Each P_{ ij } is the summation of the SSD probabilities of the states in a block. For columns four and five of the first row (i = 1), we have to sum up all the SSD probabilities for the states in Block(1) to find P_{10}. The summation of the SSD probabilities of Block(2) forms P_{11}. The rest of the P_{ ij }s are calculated similarly.
In the PBN setting, control of the network is achieved by toggling the value of the control gene. The derivation of a stationary control policy µ ∈ {0, 1}2n, means defining control actions for each state s ∈ {StateSpace}. If the control action for the state s is set to 1, it means that the network should transition from its flipped with respect to . Otherwise the network transitions as specified by its STM. The CoD-CP algorithm finds the MAXCPD table in order to specify the control actions. It uses the total probabilities P_{ ij } to define the control actions. Algorithm 1 details all the steps of CoD-CP. In the binary representation of each state s, we find the values of MAXCOD genes. The decimal conversion of the values of MAXCOD genes determines the row of the MAXCPD table corresponding to state s. Then, the total probabilities P_{ ij } are used to find D(.), as described by algorithm 1, where D(.) defines the difference between the total probability of a block of states to be desirable from that of being undesirable in the long run. Using this difference we can define the control actions: if , then flip the value of C in to start the transition from s; otherwise, flip the value of C in s and start the next transition of the Markov chain from . If , then we can select one of them uniformly randomly. Example 1, part b illustrates how control actions are assigned to the states.
Performance comparison
In this section we compare the performances of CoD-CP, SSD-CP, and MFPT-CP, first with respect to run time and then to shift of the steady-state distribution.
Run-time comparison
The dynamics of a GRN and its associated Markov chain are determined by its state transition matrix. The STM provides the full knowledge about the states and their transitions in the network; however, inferring the STM is difficult, especially when available data about the network are limited or the size of the network is large. The main advantage of the CoD-CP algorithm is that it can be directly designed on large networks without inferring the STM and only needs an estimation of the SSD of the Markov chain. This section provides a comparison of CoD-CP with MFPT-CP [6] and SSD-CP [7].
For comparing the performance of the three algorithms one needs to keep in mind their important characteristics. The CoD-CP algorithm needs the SSD to design the control policy. In cases when the SSD is known, one can directly proceed to the CoD calculations and design the control policy for the network. When the SSD is not known, it can be calculated using equation (1) or can be estimated by methods described in [22]. The model-dependent version of the MFPT algorithm requires an extra step to infer the STM. It then uses matrix inversion to find the mean-first-passage-time vectors K_{ D } and K_{ U }, this step having the same time complexity as finding the SSD. The model-free version of MFPT-CP requires time-course measurements to estimate the necessary mean-first-passage time vectors. In such a case the algorithm can skip the inference of the STM, and the complexity of estimating MFPT vectors is constant with respect to the number of genes. However, the availability of time-course data is very limited in practice. The other available greedy approach, SSD-CP also requires the SSD and STM of the network. Moreover, the SSD-CP algorithm needs to find the perturbed SSD for each state, which increases the time spent for designing the control policy.
As described in the section , CoD-CP uses the MAXCPD table to design the control policy, which divides the state space into blocks of size 2^{ n }^{– 4}. These blocks are used to assign the same control actions to all of the states in a given block and the complement control action for the block of flipped states. This significantly reduces the complexity of the control policy design and leads to shorter run times.
Steady-state performance
This section provides simulation experiments to demonstrate the performance of the CoD-CP algorithm with respect to its main goal, to shift undesirable steady-state mass to desirable steady-state mass. In the first part, the algorithm is applied to randomly generated networks. In the second part, we demonstrate CoD-CP on a real-world-derived gastrointestinal cancer network with 17 genes, which can be considered large, given that even with binary quantization, the dimension of its Boolean network STM is 2^{17} × 2^{17}.
Synthetic networks
Where and are the total probability masses of the desirable states after applying control and before applying any control, respectively, a larger λ being desirable. In real-world situations the target (T) and control (C) genes are often pre-selected by the biologists/clinicians, the basis for choice being that a phenotypically related target is to be up- or down-regulated and the control gene is known to be related to the target. However, in our simulation studies, where knowledge about T and C does not exist, we have designed a procedure to identify reasonable target and control genes. The objective of the procedure is to select a (C, T) pair such that there is a direct connection, or path, from C to T, which would be a natural constraint in applications. The strength of connection between C and T is measured by the CoD. The selected pair is called CoD-strongly-connected pair. To select this pair, we consider all two-gene combinations such that each gene in a given pair is treated as both the candidate target and candidate control gene, and the CoD of the candidate C for predicting candidate T is calculated. The pair with the maximum CoD of C candidate for predicting candidate T is picked. Then the algorithm checks if there is a path from C to the T. If such a path exists, then the (C, T) pair is chosen. If no path exists, then the pair is discarded and the next highest CoD pair is considered as the candidate (C,T) pair. For checking the existence of a path, we use the breadth-first-search (BFS) algorithm [25]. For more information please refer to the supplemental document (Additional file 1).
CoD-CP performance for p ∈ {0.1, 0.01} and singleton or cyclic attractors, averaged for 100 BN_{ p }s with 7, 8, 9 and 10 gene networks.
Network Size | ||||||
---|---|---|---|---|---|---|
p | C-T Pair | Attractors | 7 | 8 | 9 | 10 |
0.01 | Connected | Singleton | 0.442813327 | 0.43722164 | 0.343500124 | 0.26431826 |
0.01 | Connected | Cyclic | 0.438436274 | 0.309464685 | 0.288786759 | 0.204716543 |
0.01 | Random | Singleton | 0.117863711 | 0.148585675 | 0.160006514 | 0.102484183 |
0.01 | Random | Cyclic | 0.069442199 | 0.068265255 | 0.116798596 | 0.06892594 |
0.1 | Connected | Singleton | 0.322014658 | 0.256250706 | 0.205475527 | 0.162102211 |
0.1 | Connected | Cyclic | 0.315366842 | 0.237692755 | 0.191486782 | 0.132842645 |
0.1 | Random | Singleton | 0.072929047 | 0.069787742 | 0.069085431 | 0.043901727 |
0.1 | Random | Cyclic | 0.078443655 | 0.045092185 | 0.057163771 | 0.044105186 |
The first point to recognize is that using CoD-strongly-connected target-control pairs is more realistic because in practice one would control a target with gene that is strongly connected to it via prediction and the CoD is a measure of prediction. On the other hand, one could hardly expect to achieve as good results by randomly selecting targets and controls. We see this contrast reflected in the SSD shifts in Table 2. In addition, we see that using CoD-strongly-connected target-control pairs results in decreasing SSD shift for increasing network size, whereas this trend is replaced by sporadic behavior for randomly selected target-control pairs. Finally, we note the better performance for p = 0. 01 than for p = 0. 1. This reflects the more random network behavior for higher perturbation probability because the control algorithm utilizes the predictive structure in the network (as measured by the CoD) and this structure is less determinative when perturbations are more likely. In this regard we note that both MFPT-CP and SSD-CP also perform better for p = 0. 01 than for p = 0. 1, in both their non-induced and induced modes.
Gastrointestinal cancer network
The 17-gene network has a 2^{17} × 2^{17} state transition matrix. The generation and manipulation of the STM needed for the design of the MFPT-CP and SSD-CP is a hard computational problem, thus, reduction and induction are necessary steps for obtaining the two control policies. We use an estimation of the SSD because for such a large network it is infeasible to derive it analytically. The approximation method proposed in [22] is used to estimate the SSD of the network. Since CoD-CP can use the estimated SSD of the network, it can be used for directly designing the stationary control policy on the 17-gene network. The estimation procedure uses the Kolmogorov-Smirnov test to decide if the network has reached its steady-state.
Comparing SSD shift before and after applying control policies, p = 0.1. The CoD-CP designed on the 17-gene Gastrointestinal cancer network. For MFPT-CP and SSD-CP, The 17-gene network is reduced to 10 genes, the control policies designed for it and then these policies induced back and applied on the original 17-gene network.
Total Probability mass of D states, before control | 0.499871 |
---|---|
Total Probability mass of D states, after original CoD-CP | 0.726651 |
Total Probability mass of D states, after induced MFPT-CP | 0.547329 |
Total Probability mass of D states, after induced SSD-CP | 0.548864 |
Comparing SSD shift before and after applying control policies, p=0.01. The CoD-CP designed on the 17-gene Gastrointestinal cancer network. For MFPT-CP and SSD-CP, The 17-gene network is reduced to 10 genes, the control policies designed for it and then these policies induced back and applied on the original 17-gene network.
Total Probability mass of D states, before control | 0.497549 |
---|---|
Total Probability mass of D states, after original CoD-CP | 0.988466 |
Total Probability mass of D states, after induced MFPT-CP | 0.788339 |
Total Probability mass of D states, after induced SSD-CP | 0.747307 |
Conclusions
In this paper we propose a new algorithm, CoD-CP, for designing a greedy stationary control policy that beneficially alters the dynamics of large gene regulatory networks. The proposed algorithm needs minimum knowledge about the structure of the model and only uses the steady-state distribution of the associated Markov chain. This is particularly important for large networks, where it is computationally prohibitive to use the previously proposed optimal or greedy approaches for designing stationary control policies. The CoD-CP algorithm uses CoD computations based on the steady-state distribution for measuring the strength of connection between the target gene and its candidate predictor genes. CoD-CP is particularly designed for the class of network models where there is a path between the target and control genes, a condition that is reasonable in practical applications. The control action for each state of the network is defined based on the values of the strongest predictor set for the target gene. Simulations demonstrate that CoD-CP outperforms the induced versions of the MFPT-CP and SSD-CP algorithms relative to shifting the steady-state distribution of the network toward more desirable states when there is a significant amount of reduction, a requirement for large networks.
Authors contributions
NG proposed the main idea, developed the algorithm, designed and performed the simulations, and prepared the manuscript. II collaborated on the design of the algorithm and simulations, interpretation of the results and manuscript preparation. XQ provided insights on the interpretation of the algorithm and results and helped on the manuscript. ERD conceived the study, participated in the analysis and interpretation of the results, and helped draft the manuscript. All authors read and approved the final manuscript.
Declarations
Acknowledgements
This article has been published as part of BMC Bioinformatics Volume 12 Supplement 10, 2011: Proceedings of the Eighth Annual MCBIOS Conference. Computational Biology and Bioinformatics for a New Decade. The full contents of the supplement are available online at http://www.biomedcentral.com/1471-2105/12?issue=S10.
Authors’ Affiliations
References
- Shmulevich I, Dougherty ER: Probabilistic Boolean Networks: The Modeling and Control of Gene Regulatory Networks. New York: SIAM Press; 2010.View ArticleGoogle Scholar
- Pal R, Datta A, Dougherty E: Optimal infinite-horizon control for Probabilistic Boolean Networks. IEEE Trans. on Signal Processing 2006, 54(6):2375–2387.View ArticleGoogle Scholar
- Datta A, Pal R, Choudhary A, Dougherty E: Control Approaches for probabilistic gene regulatory networks. IEEE Signal Processing Mag 2007, 24: 54–63.View ArticleGoogle Scholar
- Bertsekas D: Dynamic Programming and Optimal Control. 2005.Google Scholar
- Akutsu T, Hayashida M, Ching WK, Ng MK: Control of Boolean networks: Hardness results and algorithms for the tree structured networks. Journal of Theoretical Biology 2007, 244(4):670–679. 10.1016/j.jtbi.2006.09.023View ArticlePubMedGoogle Scholar
- Vahedi G, Faryabi B, Chamberland JF, Data A, Dougherty E: Intervention in gene regulatory networks via a stationary Mean-first-Passage-Time control policy. IEEE Trans. of Biomed. Eng 2008, 55(10):2319–2331.View ArticleGoogle Scholar
- Qian X, Ivanov I, Ghaffari N, Dougherty ER: Intervention in Gene Regulatory Networks via Greedy Control Policies Based on Long-Run Behavior. BMC Systems Biology 2009.Google Scholar
- Ghaffari N, Ivanov I, Quian X, Dougherty E: A CoD-based reduction algorithm preserving the effects of stationary control policies for Boolean networks. Bioinformatics 2010, 26: 1556–1563. 10.1093/bioinformatics/btq225View ArticlePubMedGoogle Scholar
- Qian X, Ghaffari N, Ivanov I, Dougherty ER: State reduction for network intervention with probabilistic Boolean networks. Bioinformatics 2010, 26(24):3098–3104. 10.1093/bioinformatics/btq575PubMed CentralView ArticlePubMedGoogle Scholar
- Dougherty E, Kim S, Chen Y: Coeffcient of Determination in Nonlinear Signal Processing. Signal Processing 2000, 80: 2219–2235. 10.1016/S0165-1684(00)00079-7View ArticleGoogle Scholar
- Hashimoto R, Kim S, Shmulevich I, Zhang W, Bittner M, Dougherty E: Growing genetic regulatory networks from seed genes. Bioinformatics 2004, 20(8):124–1247.View ArticleGoogle Scholar
- Shmulevich I, Dougherty E, Kim S, Zhang W: Probabilistic Boolean Networks: a rule-based uncertainty model for gene regulatory networks. Bioinformatics 2002, 18(2):261–274. 10.1093/bioinformatics/18.2.261View ArticlePubMedGoogle Scholar
- Norris J: Markov Chains. Cambridge university Press; 1998.Google Scholar
- Kim S, Dougherty ER, Bittner M, Chen Y, Sivakumar K, Meltzer P, Trent JM: A General Framework for the Analysis of Multivariate Gene Interaction via Expression Arrays. Biomedical Optic 2000, 5(4):411–424. 10.1117/1.1289142View ArticleGoogle Scholar
- Martins D, Braga-Neto U, Hashimoto R, Bittner M, Dougherty ER: A General Framework for the Analysis of Multivariate Gene Interaction via Expression Arrays. IEEE Journal of Selected Topics in Signal Processing 2008, 2(3):424–439.View ArticleGoogle Scholar
- Dougherty E, Brun M, Trent J, Bittner ML: A Conditioning-Based Model of Contextual Regulation. IEEE/ACM Transactions on Computational Biology and Bioinformatics 2009, 6(2):310–320.View ArticlePubMedGoogle Scholar
- Qian X, Dougherty ER: Effect of Function Perturbation on the Steady-State Distribution of Genetic Regulatory Networks: Optimal Structural Intervention. IEEE Transactions on Signal Processing 2008, 56(10):4966–4975.View ArticleGoogle Scholar
- Schweitzer P: Perturbation theory and finite Markov chains. J. Applied Probability 1968, 5: 401–413. 10.2307/3212261View ArticleGoogle Scholar
- Hunter J: Stationary distributions and mean first passage times of perturbed Markov chains. Linear Algebra and its Applications 2005, 410: 217–243.View ArticleGoogle Scholar
- Ivanov I, Pal R, Dougherty E: Dynamics preserving size reduction mappings for probabilistic Boolean networks. IEEE Transactions on Signal Processing 2007, 55: 2310–2322.View ArticleGoogle Scholar
- Kauffman S: The Origins of Order: Self-Organization and Selection in Evolution. New York: Oxford University Press; 1993.Google Scholar
- Kim S, Li H, Dougherty E, Chao N, Chen Y, Bittner M, Suh E: Can Markov Chain Models Mimic Biological Regulation. Biological Systems 2002, 10(4):447–458. 10.1142/S021833900200069XView ArticleGoogle Scholar
- Pal R, Ivanov I, Datta A, Dougherty E: Generating Boolean Networks with a Prescribed Attractor Structure. Bioinformatics 2005, 54(21):4021–4025.View ArticleGoogle Scholar
- Pal R, Datta A, Dougherty E: Robust Intervention in Probabilistic Boolean Networks. IEEE Trans. on Signal Processing 2008, 56(3):1280–1294.View ArticleGoogle Scholar
- Cormen T, Leiserson C, Rivest RL, Stein C: Introduction to Algorithms. 2001.Google Scholar
- Price ND, Trent J, El-Naggar AK, Cogdell D, Taylor E, Hunt KK, Pollock RE, Hood L, Shmulevich I, Zhang W: Highly accurate two-gene classifier for differentiating gastrointestinal stromal tumors and leiomyosarcomas. Proceedings of the National Academy of Sciences 2007, 104(9):3414–3419. 10.1073/pnas.0611373104View ArticleGoogle Scholar
- R Development Core Team: R: A Language and Environment for Statistical Computing.R Foundation for Statistical Computing, Vienna, Austria; 2008. [http://www.R-project.org]
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.