Reverse engineering gene regulatory networks from measurement with missing values
 Oyetunji E. Ogundijo^{1},
 Abdulkadir Elmas^{1} and
 Xiaodong Wang^{1}Email author
DOI: 10.1186/s1363701600558
© The Author(s) 2017
Received: 11 August 2016
Accepted: 15 December 2016
Published: 10 January 2017
Abstract
Background
Gene expression time series data are usually in the form of highdimensional arrays. Unfortunately, the data may sometimes contain missing values: for either the expression values of some genes at some time points or the entire expression values of a single time point or some sets of consecutive time points. This significantly affects the performance of many algorithms for gene expression analysis that take as an input, the complete matrix of gene expression measurement. For instance, previous works have shown that gene regulatory interactions can be estimated from the complete matrix of gene expression measurement. Yet, till date, few algorithms have been proposed for the inference of gene regulatory network from gene expression data with missing values.
Results
We describe a nonlinear dynamic stochastic model for the evolution of gene expression. The model captures the structural, dynamical, and the nonlinear natures of the underlying biomolecular systems. We present pointbased Gaussian approximation (PBGA) filters for joint state and parameter estimation of the system with onestep or twostep missing measurements. The PBGA filters use Gaussian approximation and various quadrature rules, such as the unscented transform (UT), the thirddegree cubature rule and the central difference rule for computing the related posteriors. The proposed algorithm is evaluated with satisfying results for synthetic networks, in silico networks released as a part of the DREAM project, and the real biological network, the in vivo reverse engineering and modeling assessment (IRMA) network of yeast Saccharomyces cerevisiae.
Conclusion
PBGA filters are proposed to elucidate the underlying gene regulatory network (GRN) from time series gene expression data that contain missing values. In our statespace model, we proposed a measurement model that incorporates the effect of the missing data points into the sequential algorithm. This approach produces a better inference of the model parameters and hence, more accurate prediction of the underlying GRN compared to when using the conventional Gaussian approximation (GA) filters ignoring the missing data points.
Keywords
Gene expression Missing data Bayesian inference Gaussian filters Network inference1 Introduction
Gene regulation happens to be one of the most important processes that take place in living cells [1, 2]. For instance, it includes controls over the transcription of messenger RNA (mRNA) and the eventual translation of mRNA into protein via gene regulatory networks (GRNs). A detailed network may depict various interdependencies among genes where nodes of the network represent the genes and the edges correspond to interactions among the genes [3]. The strength of these interactions represents the extent to which a gene is affected by other genes in the network. For instance, some of the genes encode specific proteins, known as the transcription factors that can bind deoxyribonucleic acid (DNA) as part of a complex or independently and regulate their rate of transcription [4, 5]. Binding of the DNA by the transcription factors may, in some occasions, include genes that encode for other transcription factors and also genes that encode proteins for other functions. Hence, this results in a complex level of interaction among the genes in the cell. Among others, understanding the complex intracellular network in a human cell may lead to the identification of diseased genes, drug targets, and biomarkers for complex diseases [6]. Thus, identifying the structure of GRNs has become a major focus in the systems approach to biology [7–10].
The generation of high throughput time series measurement of transcript levels (e.g., via microarray experiments) has become an increasingly powerful tool for investigating complex biological processes and a useful resource for GRN inference [11]. Modeling of the gene networks with gene expression data can be loosely categorized into static and dynamic models. A static approach to modeling gene expressions makes use of the following properties: correlation, statistical independence for clustering, and mutual information [12, 13]. Particularly, the clustering approach has gained significant popularity [14, 15]. On the other hand, the dynamic modeling of GRNs from time series data has also received considerable interest. For instance, Boolean network models quantize the empirical gene expression data into binary values [16] and view the network structures as constraints. Further, via the estimation of the parameters in Ssystems, a kind of nonlinear mathematical models based on power law, few authors like [17, 18] have performed the reverse engineering of GRNs. Probabilistic Boolean network models are an extension to the Boolean network models which incorporate the inherent stochasticity of gene expression and the uncertainties introduced by the measurement noise [19]. Also, dynamic Bayesian networks (DBNs) have been proposed to model the time series gene expression data [20, 21] because DBNs can model stochasticity and handle noisy/hidden variables.
The statespace approach, an extension of the DBNs, is a popular technique to model the GRNs [22, 23], where the hidden state of the network can be estimated by Gaussian approximation (GA) filters. The conventional Kalman filter, being optimal for a linear Gaussian system [24], requires some modifications to be able to cope with the nonlinearity of the activation function that regulates the gene activity profile. For instance, the extended Kalman filter (EKF) uses the firstorder terms of the Taylor’s series expansion [25] to linearize the nonlinear functions in the model. The EKF only calculates the posterior densities accurately to the first order with all higher moments truncated. A different paradigm of the GA filtering approach is the pointbased filtering technique, which involves numerically integrating nonlinear functions by using a set of deterministic points. This approach lowers the computational complexity when compared to the Monte Carlo numerical integration which relies on randomly generated points, since it requires much less number of points with the same accuracy.
However, in reality, gene expression time series data may not contain sufficient quantity of data in the appropriate format for the inference of GRNs because of the missing data points [26]. For example, in microarray measurement of gene expression, errors such as insufficient resolution and image corruption or simply due to dust or scratches on the slide of a microarray chip may occur in the experimental process which lead to corruption or absence of some expression measurements. In the engineering literature, similar problems are inherent in networked control systems (NCS) and sensor networks where packet dropouts and time delays are an unavoidable phenomenon during data transmission [27]. Classical methods fail to solve the filtering and estimation problems for such cases with delays and missing data and cannot accurately infer the underlying network structure.
In this paper, we present a class of GA filters for inferring GRN from data with missing measurement values, which can be modeled in the same unifying framework as in the case of state estimation from onestep or twostep randomly delayed measurements [28]. A general framework is presented through augmenting the state variables and with Gaussian assumptions on the posterior state and missing measurement. To make GRN inference from measurements that contain missing data, we describe the network by a nonlinear model and a measurement model that incorporates the missing data. The inferred parameter set can be used to identify the underlying regulatory network structure.
In the literature, several pointbased Gaussian approximation (PBGA) filters have been used for solving the GRN inference problem from DNA microarray gene expression data and genomewide knockout fitness data [29, 30]; however, there is no solution that outperforms all other counterparts. Thus, one has to pick the filter balancing the estimation performance, implementation complexity, and filter stability. Prominent among the PBGA filters are the cubature Kalman filter (CKF) that makes use of the thirddegree cubature rule [31], the unscented Kalman filter (UKF) that makes use of the unscented transformation [30, 32], and the central difference Kalman filter (CDKF) that makes use of the difference rule.
The remainder of this paper is organized as follows. In Section 2, we describe the system model and problem formulation. In Section 3, we describe the corresponding GA filter. In Section 4, we investigate the performance of the proposed algorithm on a synthetic network and a diverse set of in silico networks released as a part of the DREAM project, from which observations can be made for benchmarking purposes [33, 34]. In addition, we present results on a real data obtained from the IRMA network of yeast Saccaromyces cerevisiae [35]. Finally, Section 5 concludes the paper.
 1.
\(\mathcal {N}(\mathrm {x};\mu, \Sigma)\) denotes the Gaussian probability density function with mean μ and covariance Σ.
 2.
\(\mathbb {E}_{g}\lbrace \cdot \vert \mu, \Sigma \rbrace \) denotes the Gaussian integral with respect to \(\mathcal {N}(\mathrm {x};\mu, \Sigma)\).
 3.
\(\hat {\mathrm {x}}\) represents the estimate of variable x, \(\tilde {\mathrm {x}}= \mathrm {x}\hat {\mathrm {x}}\) is the estimation error, and \(\mathbb {E}[\!\cdot ]\) denotes the expectation operation.
 4.
X ^{−1} and X ^{ T } represent the inverse and transpose of matrix X, respectively, and I _{ n } denotes the ndimensional identity matrix.
2 Methods
2.1 Problem formulation and system model
with μ being a parameter to be identified and I _{0i } being the external bias on the ith gene. The noise vector e _{ k }=[e _{ k,1},e _{ k,2},…,e _{ k,N }]^{ T } is Gausssian distributed with zero mean and covariance matrix \(\mathbf {Q}^{'}_{k}\), for k=1,…,K.
where f(·) is the nonlinear function associated with (2) and (4); w_{ k } = [ e _{ k,1},…,e _{ k,N },0,…,0] is the augmented noise vector with covariance matrix \(\mathbf {Q}_{k} = \text {diag} ([\mathbf {\!Q}_{k}^{'}~~\mathbf {0}_{2N + 2N^{2}}])\), where 0 _{ m } denotes an m×m allzero matrix.
where z_{ k } is the output data from the experiments at time k, h(x_{ k })=g _{ k } and \(\mathrm {v}_{k} \in \mathbb {R}^{N}\) is Gaussian distributed noise with zero mean and covariance matrix \(\mathbf {R}_{k} \in \mathbb {R}^{N\times N}\).
Finally, (5)–(8) describe the dynamic model we propose for inferring GRNs with onestep or twostep missing measurements.
 1.The onestep posterior predictive PDF of the state x_{ k } conditioned on Y_{ k−1} is Gaussian, i.e.,$$ p(\mathrm{x}_{k}\mathrm{Y}_{k1}) = \mathcal{N}(\mathrm{x}_{k};\hat{\mathrm{x}}_{kk1},\mathrm{P}_{kk1}^{\mathrm{x}\mathrm{x}}), $$(10)where$${} \hat{\mathrm{x}}_{kk1} = \mathbb{E}\left[ \mathrm{x}_{k}\mathrm{Y}_{k1}\right], ~~ \mathrm{P}_{kk1}^{\mathrm{x}\mathrm{x}} = \mathbb{E}\left[ \tilde{\mathrm{x}}_{kk1}\tilde{\mathrm{x}}_{kk1}^{T}\mathrm{Y}_{k1}\right]. $$(11)
 2.The onestep posterior predictive PDF of y_{ k } conditioned on Y_{ k−1} is Gaussian, i.e.,$$ p(\mathrm{y}_{k}\mathrm{Y}_{k1}) = \mathcal{N}(\mathrm{y}_{k};\hat{\mathrm{y}}_{kk1},\mathrm{P}_{kk1}^{\mathrm{y}\mathrm{y}}), $$(12)where$${} \hat{\mathrm{y}}_{kk1} = \mathbb{E}\left[ \mathrm{y}_{k}\mathrm{Y}_{k1}\right],~~ \mathrm{P}_{kk1}^{\mathrm{y}\mathrm{y}} = \mathbb{E}\left[ \tilde{\mathrm{y}}_{kk1}\tilde{\mathrm{y}}_{kk1}^{T}\mathrm{Y}_{k1}\right]. $$(13)
2.2 Gaussian approximation filters with missing measurements
As with the general GA filtering, the filtering procedure consists of the state update and measurement update.
2.2.1 State update
For the detailed derivations, see Additional file 1.
2.2.2 Measurement update
The delayed/missing measurement statistics \(\hat {\mathrm {z}}_{kdk1}\), \(\mathrm {P}_{kdk1}^{\mathrm {z}\mathrm {z}}\), and \(\mathrm {P}_{k,kdk1}^{\mathfrak {X}\mathrm {z}}\) are defined as follows.
The filtering estimate \(\hat {\mathrm {x}}_{kk}\) and covariance Pkkxx of the system state are obtained from \(\hat {\mathfrak {X}}_{kk}\) and \(\mathrm {P}^{\mathfrak {X}\mathfrak {X}}_{kk}\) respectively. (See Additional file 1 for derivations).
However, the Gaussian weighted integrals in (22) and (25)–(27) contain nonlinear functions which render the analytical calculation infeasible and the algorithm becomes intractable. To deal with this, we employ the pointbased numerical integration techniques, which is presented in Additional file 1.
3 Results
We assess the proposed algorithm using both synthetic data and real data. Gold standards or the groundtruths are provided for both categories of data and the inferred networks are “benchmarked" against the gold standards. Benchmarking is done by counting the number of links correctly predicted by the algorithm (true positives, TP), the number of incorrectly predicted links (false positives, FP), the number of true links missed in the inferred network (false negatives, FN), and the number of correctly identified nonexisting links (true negatives, TN). Thus, the following performance metrics will be defined accordingly: true positive rate or recall also known as the sensitivity (TPR = TP/(TP+FN)), positive predictive value or precision (PPV = TP/(TP+FP)), and false positive rate (FPR = FP/(FP+TN), where specificity = 1FPR). All the metrics are computed for different thresholds and the area under the receiver operating characteristic (AUROC) curve and the area under the precisionrecall (AUPR) curve are estimated to illustrate the overall inference performance of the algorithms. As the inference result comprises of the estimates of both the linear and nonlinear regulatory coefficients among the genes, if at least one of the regulatory coefficients between any two genes is recovered, the link is designated as TP.
In addition, y_{1}=z_{1}; at time k=2 the measurement output can be missing by onestep; and at any time instant k≥3 it can be missing by onestep or twostep. With the prior knowledge of the number of missing data points to be replaced in the experimental output, an estimate of the value of q, the success probability of the Bernoulli variable ς _{ k } can be made. Specifically, if the number of missing data points is less than 20% of the total number of data points, a q value chosen in the interval [0.05, 0.2] is a good choice. In our experiments, q=0.1, so that the probability that z_{ k } is used in the estimation is \({p^{0}_{k}} = 0.9\), the probability that z_{ k−1} is used in the estimation \({p^{1}_{k}} = 0.09\), and the probability that z_{ k−2} is used in the estimation is \({p^{2}_{k}} = 0.01\). In the remainder of this paper, we denote the datasets that have no missing values as the complete measurements (CM) and we demote the datasets with missing but replaced data points as the missing measurements (MM). The MM is created in the following manner: at time k, if z_{ k } is missing and z_{ k−1} is available, we replace z_{ k } with z_{ k−1}; otherwise, we replace z_{ k } with z_{ k−2}, as there can be maximum of two consecutive missing data points in the measurement.
3.1 Synthetic network
Network parameters for the synthetic network
Edge  LCC and NCC  UKF with CM  UKF with MM  UKFMM with MM 

(1,1)  0.5 (0.4)  — (0.5880)  — —  0.7313 — 
(3,1)  0.5 (0.4)  0.3837 (0.4391)  0.7357 (0.5043)  0.4079 (0.4223) 
(3,2)  0.5 (0.4)  0.7380 (0.4390)  — —  0.7380 (0.4192) 
(3,5)  0.5 (0.4)  0.6098 (0.4391)  0.6623 (0.5043)  0.8285 (0.4354) 
(4,3)  0.5 (0.4)  0.7257 (0.3059)  0.3953 (0.2123)  0.7256 (0.3235) 
(5,2)  0.5 (0.4)  — (0.3837)  — —  — (0.3235) 
(5,4)  0.5 (0.4)  0.6677 (0.3839)  0.5813 (0.3706)  0.7850 (0.3464) 
(4,4)  — —  — —  — (0.8417)  — — 
(5,1)  — —  — —  — (0.5916)  — — 
(5,5)  — —  — —  — (0.3705)  — — 
(1,5)  — —  — —  — —  0.5722 — 
Average TPR and PPV for the synthetic network (standard deviations are shown in parentheses)
UKF with CM  UKF with MM  UKFMM with MM  

TPR  1.00  0.48 (0.035)  0.91 (0.017) 
PPV  1.00  0.53 (0.028)  0.86 (0.013) 
3.2 DREAM4 in silico gene regulatory networks
In order to assess the performance of GRN inference algorithms, several in silico gene networks have been produced as the benchmarking data sets, specifically, the DREAM in silico gene networks [39–41]. We made use of the 10gene networks by the DREAM4 challenge to test the efficacy of the proposed algorithm. All networks and data were generated with version 2.0 of GeneNetWeaver (GNW) [42]. In total, there are five separate networks, each with 10 genes, whose topologies were extracted from the known GRNs in Escherichia coli and Saccharomyces cerevisiae. The time series measurements were generated using parametrized stochastic differential equations (SDEs), with observations uniformly sampled (21 time points, single replicate) under five different perturbations, for a total of 105 observations per gene. The inference is performed by using all the perturbations. Selfinteraction/autoregulatory edges were not expected in the predictions and were subsequently removed. Since the number of possible edges in an Ngene network without autoregulatory interactions is N(N−1), the length of a complete list of predictions is 90 edges for a network of size 10 [33, 34].
AUROC and AUPR curves for the DREAM4 networks
UKF with CM  UKF with MM  GP4GRN with CM  GP4GRN with MM  UKFMM with MM  

N1  [0.63] [0.42]  [0.44(0.024)][0.24(0.020)]  [0.66] [0.42]  [0.42(0.027)][0.29(0.021)]  [0.61(0.015)][0.42(0.008)] 
N2  [0.67] [0.49]  [0.48(0.018)][0.26(0.017)]  [0.69] [0.44]  [0.44(0.015)][0.28(0.018)]  [0.64(0.013)][0.44(0.011)] 
N3  [0.72] [0.50]  [0.45(0.020)][0.30(0.012)]  [0.70] [0.47]  [0.50(0.022)][0.33(0.016)]  [0.72(0.021)][0.53(0.012)] 
N4  [0.75] [0.52]  [0.56(0.019)][0.28(0.011)]  [0.62] [0.35]  [0.36(0.031)][0.25(0.027)]  [0.72(0.009)][0.50(0.010)] 
N5  [0.81] [0.44]  [0.53(0.021)][0.26(0.019)]  [0.86] [0.65]  [0.55(0.022)][0.40(0.019)]  [0.80(0.012)][0.42(0.014)] 
We also compared our algorithm against a relevant computational method designed for the GRN network inference, i.e., [43], which is based on the use of Bayesian analysis with ordinary differential equations (ODEs) and nonparametric Gaussian process, an algorithm referred to as GP4GRN. The inference result of GP4GRN with CM is shown in the fourth column in Table 3. Similarly, we tested GP4GRN with the MM where we ran 1000 experiments. At each run, we created the MM by randomly deleting up to five data points with maximum of two consecutive data points and replaced the deleted data points accordingly. The averages and standard deviations of AUROC and AUPR are obtained and the corresponding results are summarized in the fifth column in Table 3. We conclude that the GP4GRN method has comparable performance to the UKF in all data sets, and similarly it is outperformed by the proposed UKFMM algorithm under missing data conditions.
3.3 Saccharomyces cerevisiae IRMA network
Saccharomyces cerevisiae GAL network in yeast is one of the most prominent model systems due to its importance for the studies of eukaryotic regulation and relatively selfcontained nature [44–47]. A synthetic GRN that contains 5 genes has previously been constructed in the budding yeast [35]. In the well studied network, popularly referred to as in vivo reverse engineering and modeling assessment (IRMA) network, each of the genes regulate at least one other gene in the network. Expression within the network is activated in the presence of galactose and then switched to glucose to obtain the switchoff data which represents the expressive samples at 21 time points. The switchon data consists of 16 sample points and is obtained by growing the cells in a glucose medium and then changing to galactose.
AUROC and AUPR curve for the yeast networks
UKF with CM  UKF with MM  GP4GRN with CM  GP4GRN with MM  Proposed UKFMM  

AUROC curve  0.70  0.42  0.76  0.49  0.68 
AUPR curve  0.46  0.34  0.57  0.38  0.46 
4 Discussion
This work presents a novel algorithm for GRN inference from timeseries gene expression data with onestep or twostep missing measurements. Gene regulation is assumed to follow a nonlinear state evolution model described in (1). The parameters of the model, which are assumed to be the regulatory coefficients between the genes, are estimated with a modified unscented Kalman filtering algorithm. We considered the experimental scenarios that lead to total loss of expression values for all genes at a particular time point or few successive time points which may significantly diminish the performance of GRN inference algorithms.
In the proposed algorithm, the state vector which is the gene expression at each time point in (1) is concatenated with the model parameters and an augmented state vector in (3) is defined for the joint estimation of gene expression values and system parameters. We consider the possibility that each real measurement is randomly missing and the estimation is made from the available measurements. The use of the UKF, an instance of the PBGA filters, for the state and parameter estimation renders the algorithm computationally efficient and capable of working offline or online (when all the measurements are readily available, or they become available successively, respectively). The proposed algorithm is tested on both synthetic and real biological data to evaluate the efficacy of the predictions. From the series of results obtained for both synthetic data and the real biological data, we conclude that the gene network structure can be inferred from time series data with missing values.
In this paper, we have applied the proposed algorithm to the time series data generated from the DNA microarray because to our best of knowledge, DNA microarray is still of interest in transcriptome profiling due to its reduced cost and widespread use as compared to the RNAseq. In addition, it has been shown that there is there is high correlation between the gene expression profiles generated between the DNA microarray and RNAseq [48, 49]. Hence, the proposed method can easily be extended to time series gene expression data from RNAseq.
In general, this work addresses the possibility of having onestep or twostep missing expression values by considering them as the delayed observations of the full set of genes. Future work will focus on the inference of the structure of a (potentially larger) network by incorporating a general sstep missing values for sconsecutive time points, which may address more complex missing data scenarios.
5 Conclusions
Time series gene expression data be modeled with statespace model and the model parameters can be estimated using different GA filters. Unfortunately, there are situations which result in loss of expression values for all genes at a particular time point or few successive time points. In this case, conventional filtering approach fails to correctly estimate the model parameters, which are used to elucidate the underlying GRN. We have proposed PBGA filters that treat the missing measurement values as a set of delayed measurements and demonstrated that the modified filter can estimate the model parameters, with missing measurements, as accurate as the conventional filter with no missing measurements.
Abbreviations
 AUPR:

Area under precisionrecall
 AUROC:

Area under the receiver operating characteristic
 CKF:

Cubature Kalman filter
 CDKF:

Central difference Kalman filter
 CM:

Complete measurements
 DBN:

Dynamic Bayesian network
 DNA:

Deoxyribonucleic acid
 DREAM:

Dialogue on reverse engineering assessment and methods
 EKF:

Extended Kalman filter
 FN:

False negative
 FP:

False positive
 FPR:

False positive rate
 GA:

Gaussian approximation
 GRN:

Gene regulatory network
 GWN:

GeneNetWeaver
 IRMA:

Invivo reverseengineering and modeling assessment
 LCC:

Linear connection coefficient
 MM:

Missing measurements
 NCC:

Nonlinear connection coefficient
 NCS:

Networked control systems
 ODE:

Ordinary differential equation
 PBGA:

Pointbased Gaussian approximation
 PDF:

Probability density function
 PPV:

Positive predictive value
 RNA:

Ribonucleic acid
 SDE:

Stochastic differential equation
 TN:

True negative
 TP:

True positive
 TPR:

True positive rate
 UKF:

Unscented Kalman filter
 UKFMM:

Unscented Kalman filter with onestep or twostep missing measurements
 UT:

Unscented transform
Declarations
Acknowledgements
Not applicable.
Availability of data and materials
DREAM4 in silico challenge datasets analyzed during the current study are available for download at http://gnw.sourceforge.net/dreamchallenge.html. IRMA in vivo dataset analyzed during the current study is available for download at http://www.cell.com/supplemental/S00928674(09)001561.
Authors’ contributions
OEO developed the idea, performed the implementation of the algorithm, ran most of the experiments, performed the data analysis and interpretation, and heavily involved in the writing of the manuscript. AE performed some of the simulation experiments, was involved with data analysis and interpretation, and was a major contributor in writing the manuscript. XW conceived the idea, contributed significantly to its development, supervised its implementation, and performed a thorough revision of the manuscript. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Consent for publication
Not applicable.
Ethics approval and consent to participate
Not applicable.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 EI Palmero, SGP de Campos, M Campos, NC Souza, IDC Guerreiro, AL Carvalho, MMC Marques, Mechanisms and role of microRNA deregulation in cancer onset and progression. Genet. Mol. Biol.34(3), 363–370 (2011).View ArticleGoogle Scholar
 D Baek, J Villén, C Shin, FD Camargo, SP Gygi, DP Bartel, The impact of microRNAs on protein output. Nature. 455(7209), 64–71 (2008).View ArticleGoogle Scholar
 J Jin, K He, X Tang, Z Li, L Lv, Y Zhao, J Luo, G Gao, An arabidopsis transcriptional regulatory map reveals distinct functional and evolutionary features of novel transcription factors. Mol. Biol. Evol.32(7), 1767–1773 (2015).View ArticleGoogle Scholar
 VB Teif, K Rippe, Statistical–mechanical lattice models for protein–DNA binding in chromatin. J. Phys. Condensed Matter. 22(41), 414105 (2010).View ArticleGoogle Scholar
 RE Moellering, M Cornejo, TN Davis, C Del Bianco, JC Aster, SC Blacklow, AL Kung, DG Gilliland, GL Verdine, JE Bradner, Direct inhibition of the notch transcription factor complex. Nature. 462(7270), 182–188 (2009).View ArticleGoogle Scholar
 AL Barabási, N Gulbahce, J Loscalzo, Network medicine: a networkbased approach to human disease. Nat. Rev. Genet.12(1), 56–68 (2011).View ArticleGoogle Scholar
 LE Chai, SK Loh, ST Low, MS Mohamad, S Deris, Z Zakaria, A review on the computational approaches for gene regulatory network construction. Comput. Biol. Med.48:, 55–65 (2014).View ArticleGoogle Scholar
 F EmmertStreib, M Dehmer, B HaibeKains, Gene regulatory networks and their applications: understanding biological and medical problems in terms of networks. Front. Cell Dev. Biol.2:, 38 (2014).View ArticleGoogle Scholar
 N Vijesh, SK Chakrabarti, J Sreekumar, et al., Modeling of gene regulatory networks: a review. J. Biomed. Sci. Eng.6(02), 223 (2013).View ArticleGoogle Scholar
 M Hecker, S Lambeck, S Toepfer, E Van Someren, R Guthke, Gene regulatory network inference: data integration in dynamic models? A review. Biosystems. 96(1), 86–103 (2009).View ArticleGoogle Scholar
 CE Gagna, WC Lambert, Novel multistranded, alternative, plasmid and helical transitional dna and RNA microarrays: implications for therapeutics. Pharmacogenomics. 10(5), 895–914 (2009).View ArticleGoogle Scholar
 W Zhao, E Serpedin, ER Dougherty, Inferring connectivity of genetic regulatory networks using informationtheoretic criteria. Comput. Biol. Bioinform. IEEE/ACM Trans.5(2), 262–274 (2008).View ArticleGoogle Scholar
 J Dougherty, I Tabus, J Astola, Inference of gene regulatory networks based on a universal minimum description length. EURASIP J. Bioinforma. Syst. Biol.2008(1), 1 (2008).View ArticleGoogle Scholar
 B Godsey, Improved inference of gene regulatory networks through integrated bayesian clustering and dynamic modeling of timecourse expression data. PloS ONE. 8(7), 68358 (2013).View ArticleGoogle Scholar
 CD Giurcarneanu, I Tabus, J Astola, J Ollila, M Vihinen, Fast iterative gene clustering based on information theoretic criteria for selecting the cluster structure. J. Comput. Biol.11(4), 660–682 (2004).View ArticleGoogle Scholar
 S Kauffman, C Peterson, B Samuelsson, C Troein, Random boolean network models and the yeast transcriptional network. Proc. Natl. Acad. Sci.100(25), 14796–14799 (2003).View ArticleGoogle Scholar
 X Yang, JE Dent, C Nardini, An ssystem parameter estimation method (SPEM) for biological networks. J. Comput. Biol.19(2), 175–187 (2012).MathSciNetView ArticleGoogle Scholar
 OR Gonzalez, C Küper, K Jung, PC Naval, E Mendoza, Parameter estimation using simulated annealing for ssystem models of biochemical networks. Bioinformatics. 23(4), 480–486 (2007).View ArticleGoogle Scholar
 I Shmulevich, ER Dougherty, W Zhang, From boolean to probabilistic boolean networks as models of genetic regulatory networks. Proc. IEEE. 90(11), 1778–1792 (2002).View ArticleGoogle Scholar
 Y Huang, J Wang, J Zhang, M Sanchez, Y Wang, Bayesian inference of genetic regulatory networks from time series microarray data using dynamic Bayesian networks. J. Multimedia. 2(3), 46–56 (2007).View ArticleGoogle Scholar
 TF Liu, WK Sung, A Mittal, Model gene network by semifixed Bayesian network. Expert Syst. Appl.30(1), 42–49 (2006).View ArticleGoogle Scholar
 J Angus, M Beal, J Li, C Rangel, D Wild, in Learning and Inference in Computational Systems Biology, ed. by ND Lawrence, M Girolami, M Rattray, and G Sanguinetti. Inferring transcriptional networks using prior biological knowledge and constrained statespace models (MIT PressCambridge, 2010), pp. 117–152.Google Scholar
 O Hirose, R Yoshida, S Imoto, R Yamaguchi, T Higuchi, DS CharnockJones, S Miyano, et al., Statistical inference of transcriptional modulebased gene networks from time course gene expression profiles by using state space models. Bioinformatics. 24(7), 932–942 (2008).View ArticleGoogle Scholar
 RE Kalman, A new approach to linear filtering and prediction problems. J. Basic Eng.82(1), 35–45 (1960).View ArticleGoogle Scholar
 Z Wang, X Liu, Y Liu, J Liang, V Vinciotti, An extended Kalman filtering approach to modeling nonlinear dynamic gene regulatory networks via short gene expression time series. IEEE/ACM Trans. Comput. Biol. Bioinformatics (TCBB). 6(3), 410–419 (2009).View ArticleGoogle Scholar
 MSB Sehgal, I Gondal, LS Dooley, R Coppel, How to improve postgenomic knowledge discovery using imputation. EURASIP J. Bioinforma. Syst. Biol.2009(1), 1 (2009).View ArticleGoogle Scholar
 S Sun, L Xie, W Xiao, N Xiao, Optimal filtering for systems with multiple packet dropouts. Circ. Syst. II Express Briefs IEEE Trans.55(7), 695–699 (2008).View ArticleGoogle Scholar
 M Sahebsara, T Chen, SL Shah, Optimal filtering with random sensor delay, multiple packet dropout and uncertain observations. Int. J. Control.80(2), 292–301 (2007).MathSciNetView ArticleMATHGoogle Scholar
 A Noor, E Serpedin, M Nounou, H Nounou, Reverse engineering sparse gene regulatory networks using cubature Kalman filter and compressed sensing. Adv. Bioinformatics. 2013:, 205763 (2013).Google Scholar
 L Wang, X Wang, AP Arkin, MS Samoilov, Inference of gene regulatory networks from genomewide knockout fitness data. Bioinformatics. 29(3), 338–346 (2013).View ArticleGoogle Scholar
 I Arasaratnam, S Haykin, Cubature Kalman filters. IEEE Trans. Autom. Control. 54(6), 1254–1269 (2009).MathSciNetView ArticleMATHGoogle Scholar
 EA Wan, R Van Der Merwe, in Adaptive Systems for Signal Processing, Communications, and Control Symposium 2000. ASSPCC. The IEEE 2000. The unscented Kalman filter for nonlinear estimation (IEEEAlberta, Canada, 2000), pp. 153–158.Google Scholar
 RJ Prill, D Marbach, J SaezRodriguez, PK Sorger, LG Alexopoulos, X Xue, ND Clarke, G AltanBonnet, G Stolovitzky, Towards a rigorous assessment of systems biology models: the DREAM3 challenges. PLoS ONE. 5(2), 9202 (2010).View ArticleGoogle Scholar
 D Marbach, RJ Prill, T Schaffter, C Mattiussi, D Floreano, G Stolovitzky, Revealing strengths and weaknesses of methods for gene network inference. Proc. Natl. Acad. Sci.107(14), 6286–6291 (2010).View ArticleGoogle Scholar
 I Cantone, L Marucci, F Iorio, MA Ricci, V Belcastro, M Bansal, S Santini, M Di Bernardo, D Di Bernardo, MP Cosma, A yeast synthetic network for in vivo assessment of reverseengineering and modeling approaches. Cell. 137(1), 172–181 (2009).View ArticleGoogle Scholar
 L Chen, K Aihara, Chaos and asymptotical stability in discretetime neural networks. Physica D Nonlinear Phenomena. 104(3), 286–325 (1997).MathSciNetView ArticleMATHGoogle Scholar
 A HermosoCarazo, J LinaresPérez, Unscented filtering algorithm using twostep randomly delayed observations in nonlinear systems. Appl. Math. Model.33(9), 3705–3717 (2009).MathSciNetView ArticleMATHGoogle Scholar
 K Ito, K Xiong, Gaussian filters for nonlinear filtering problems. IEEE Trans. Autom. Control. 45(5), 910–927 (2000).MathSciNetView ArticleMATHGoogle Scholar
 D Marbach, T Schaffter, C Mattiussi, D Floreano, Generating realistic in silico gene networks for performance assessment of reverse engineering methods. J. Comput. Biol.16(2), 229–239 (2009).View ArticleGoogle Scholar
 G Stolovitzky, RJ Prill, A Califano, Lessons from the DREAM2 challenges. Ann. N. Y. Acad. Sci.1158(1), 159–195 (2009).View ArticleGoogle Scholar
 G Stolovitzky, D Monroe, A Califano, Dialogue on reverseengineering assessment and methods. Ann. N. Y. Acad. Sci.1115(1), 1–22 (2007).View ArticleGoogle Scholar
 T Schaffter, D Marbach, D Floreano, Genenetweaver: in silico benchmark generation and performance profiling of network inference methods. Bioinformatics. 27(16), 2263–2270 (2011).View ArticleGoogle Scholar
 T Äijö, H Lähdesmäki, Learning gene regulatory networks from gene expression measurements using nonparametric molecular kinetics. Bioinformatics. 25(22), 2937–2944 (2009).View ArticleGoogle Scholar
 O Egriboz, F Jiang, JE Hopper, Rapid gal gene switch of Saccharomyces cerevisiae depends on nuclear gal3, not nucleocytoplasmic trafficking of GAL3 and GAL80. Genetics. 189(3), 825–836 (2011).View ArticleGoogle Scholar
 M Johnston, JS Flick, T Pexton, Multiple mechanisms provide rapid and stringent glucose repression of GAL gene expression in Saccharomyces cerevisiae. Mol. Cell. Biol.14(6), 3834–3841 (1994).View ArticleGoogle Scholar
 D Lohr, P Venkov, J Zlatanova, Transcriptional regulation in the yeast gal gene family: a complex genetic network. FASEB J.9(9), 777–787 (1995).Google Scholar
 S Ostergaard, L Olsson, J Nielsen, Metabolic engineering of Saccharomyces cerevisiae. Microbiol. Mol. Biol. Rev.64(1), 34–50 (2000).View ArticleGoogle Scholar
 S Zhao, WP FungLeung, A Bittner, K Ngo, X Liu, Comparison of RNAseq and microarray in transcriptome profiling of activated t cells. PLoS ONE. 9(1), 78644 (2014).View ArticleGoogle Scholar
 S Kogenaru, Q Yan, Y Guo, N Wang, RNAseq and microarray complement each other in transcriptome profiling. BMC Genomics. 13(1), 1 (2012).View ArticleGoogle Scholar