Inferring transcription factor collaborations in gene regulatory networks
- Sherine Awad^{1} and
- Jin Chen^{1, 2}Email author
https://doi.org/10.1186/1752-0509-8-S1-S1
© Awad and Chen; licensee BioMed Central Ltd. 2014
Published: 24 January 2014
Abstract
Background
Living cells are realized by complex gene expression programs that are moderated by regulatory proteins called transcription factors (TFs). The TFs control the differential expression of target genes in the context of transcriptional regulatory networks (TRNs), either individually or in groups. Deciphering the mechanisms of how the TFs control the expression of target genes is a challenging task, especially when multiple TFs collaboratively participate in the transcriptional regulation.
Results
We model the underlying regulatory interactions in terms of the directions (activation or repression) and their logical roles (necessary and/or sufficient) with a modified association rule mining approach, called mTRIM. The experiment on Yeast discovered 670 regulatory interactions, in which multiple TFs express their functions on common target genes collaboratively. The evaluation on yeast genetic interactions, TF knockouts and a synthetic dataset shows that our algorithm is significantly better than the existing ones.
Conclusions
mTRIM is a novel method to infer TF collaborations in transcriptional regulation networks. mTRIM is available at http://www.msu.edu/~jinchen/mTRIM.
Background
Considering the prohibitive costs and the tremendous number of possible combinations of higher-order gene knockouts, it is currently impossible for researchers to examine all of possible gene knockout combinations experimentally. One solution to this problem is to select only the TF groups that are most likely to bring about the phenotypic change. In order to accomplish this, we need to understand the interactions employed by multiple TFs (called regulatory interactions) to regulate their common target genes. However, this is a difficult task, because when multiple TFs simultaneously or sequentially control their target genes, a single gene responds to merged inputs, resulting in complex gene expression patterns [17, 18]. The exhaustive approach requires enumerating all TF combinations, which, given the high complexity of combinatorial, is simply impractical at the whole genome level.
In our previous research [19], a Hidden Markov model was developed to relate gene expression patterns to regulatory interactions, in order to solve a relatively simpler sub-problem that considers only two TFs. To predict regulatory interactions for all possible collaborative TFs, we propose an algorithm called "mTRIM" (m ultiple T ranscriptional R egulatory I nteraction M echanism) in this paper. By uncovering the regulatory interactions in terms of their directions (activation or repression) and corresponding logical roles (necessary and/or sufficient) from gene expression and TF-DNA binding data, mTRIM identifies TF groups that are collaboratively responsible for target gene expressions. Such inferences may provide high-quality candidate sets for further experimentally detecting the collaborative functions of gene regulations that are largely unknown [18]. Yeang and Jaakkola [3] attempted to characterize the combinatorial regulation of multiple-TF regulatory interactions using a heuristic approach to measure how well a regulatory module fits the associated binding and gene expression data with a log-likelihood function. The regulatory module's likelihood is maximized with a greedy approach by incrementally adding genes to the module and monitoring the predictions of the regulatory interactions for optimality. However, this incremental approach does not study the functions of the TFs simultaneously because of the scalability issue introduced by the greedy search. This method also uses a probability-based approach to calculate the significance of the combinatorial property of TFs, determined by the gap of likelihood scores between their model and a model built on randomized data in the entire time frame. However, as stated in [4], a TF usually functions at specific "activation time points" instead of throughout the entire time course, meaning that the identification of regulatory interaction modules should be focused on activation time-points rather than the entire time frame.
To derive dynamic regulatory networks that associate TFs with target genes at their activation time-points, an algorithm called DREM was proposed [4]. DREM integrates time-series gene expression data and protein-DNA binding data to build a global temporal map, in order to uncover transcriptional regulatory events leading to the observed temporal expression patterns and the underlying factors that control these events during a cell's response to stimuli. The method mainly works by identifying bifurcation time-points where the expression of a subset of genes diverges from the rest of the genes. The bifurcation points are then annotated with the TFs regulating these transitions, which result in a unified temporal map. The method can therefore facilitate the determination of the time when TFs are exerting their influence, and assigns genes to paths in the map based on their expression profiles and the TFs that control them. Unlike the method by Yeang and Jaakkola [3], DREM's ability to derive dynamic maps that associate TFs with the genes they regulate and their activation time-points has indeed led to better insights for the regulatory module being studied. However, DREM does not infer the logical roles of the TFs (i.e., whether a specific TF is necessary or sufficient for regulating a set of target genes). Such knowledge is extremely useful for designing high-order genetic variation experiments to understand the complex regulatory mechanisms of biological processes.
TRIM is an HMM based model which was developed to infer the collaboration of at most two TFs that regulate the same target genes. In the HMM, the functions of a TF are hidden states. The model starts with random priors, and then is iteratively trained using EM till convergence. Since each possible function of a TF is a node in the HMM, there are four nodes (AS, AN, RS, and RN) for each TF. With the design of HMM (and the limited training data), the number of TFs TRIM can handle is limited.
The enumeration of all TF combinations is clearly a NP problem. Therefore, we focused on the most important biological problem (i.e., 2-TF combination) and therefore "hardcoded the problem in TRIM. In this paper, we solve the efficiency problem by developing an association rule mining algorithm which is capable to handle a large amount of data with high-level combinations.
In this paper, we propose a new model mTRIM for inferring regulatory interactions for multiple TFs with an EM-based Bayesian inference approach [20, 21] and a modified bottom-up association rule mining method. Experimental results evaluated with yeast genetic interactions, TF knockouts and a synthetic dataset shows that our algorithm is significantly better than the existing ones.
Methods
mTRIM is developed to efficiently infer regulatory interactions for all possible collaborative TFs in a TRN. The feasibility is achieved in two steps. First, an EM-based Bayesian inference approach is developed to identify all the significant individual TF regulatory interactions, meaning that individual TFs that can regulate the target genes independent to the existence of other TFs. For the TFs which require collaborations with other TFs to drive the target genes, or are actually non-deterministic (meaning lack of clear evidence of regulation), their p-values are insignificant. They are considered as the inputs of the second step.
Second, in order to identify the collaboration of k TFs (k ≥ 2), i.e., k-TF regulatory interaction, a bottom-up association rule mining approach is developed. While the significant TF groups are reported to the users, the insignificant ones are joined with each other to mine (k + 1)-TF regulatory interactions. It should be noted that unlike the conventional association rule mining which seeks the longest possible patterns, mTRIM outputs the shortest significant results, in that the goal of mTRIM is to discover the smallest group of TFs that can regulate the target genes, so that biological experiments with high-order genetic variations can be subsequently carried out for the understanding of the behavior of TRNs. In terms of time complexity, consider a candidate k-TF regulatory interaction $I=<{h}_{t{f}_{1}},...,{h}_{t{f}_{k}}>\Rightarrow {h}_{g}$. The algorithm computes AfnScore and p-values of all of the subsets, I - {tf_{ j }} (∀j = 1, 2, ..., k). If one of them is significant, I is immediately pruned. Hence the time complexity is O(k) for each candidate k-TFs regulatory interaction. Every merging operation requires at most k - 2 equality comparisons. In the best-case scenario, it produces a viable candidate k-TF interaction. In the worst case, the algorithm merges every pair of infrequent (k - 1)-TF candidates. Therefore, the overall cost of merging candidates is between ${\sum}_{k=2}^{\left|TF\right|}\left(k-2\right)\left|{P}_{k}\right|\phantom{\rule{0.3em}{0ex}}\mathsf{\text{and}}{\sum}_{k=2}^{\left|TF\right|}\left(k-2\right)|{P}_{k-1}{|}^{2}$, where P_{ k } is the candidate set of k-TF regulatory interactions. To improve the algorithm efficiency, a hash tree is constructed for the storage and quick access to all of the candidates. Because the maximum depth of the hash tree is k, the cost for populating the hash tree of candidates is $O\left({\sum}_{k=2}^{\left|TF\right|}k\left|{P}_{k}\right|\right)$. During candidate pruning, it is required to verify whether the k - 1 subsets of every candidate k-TF regulatory interactions are significant. Since the cost for looking up an item in a hash tree is O(k), the time complex of candidate pruning step is $O\left({\sum}_{k=2}^{w}k\left(k-2\right)\left|{P}_{k}\right|\right)$.
Concepts
A TRN can be represented as a directed graph in which each node is a TF or a gene, and each edge pointing from a TF to a gene represents a regulation relationship between them. In many organisms, in-depth transcriptome analysis has revealed the modular architecture of gene expression [22]. A regulatory module is a self-consistent regulatory unit R(TF, G, I) representing a set of co-expressed genes G = {g_{1}, g_{2}, ..., g_{ n }} regulated in concert by a group of TFs in TF = {tf_{1}, tf_{2}, ..., tf_{ m }} that govern the target genes' behaviors via regulatory interaction I [5]. An example of the regulatory module is shown in Figure 1b.
A regulatory interaction $I=<{{h}_{tf}}_{{}_{1}},\phantom{\rule{2.77695pt}{0ex}}\dots ,\phantom{\rule{2.77695pt}{0ex}}{h}_{t{f}_{i}},\phantom{\rule{2.77695pt}{0ex}}\dots ,\phantom{\rule{2.77695pt}{0ex}}{h}_{t{f}_{m}}>\Rightarrow {h}_{g}$ (which is the final output of mTRIM) is defined as a set of TFs {tf_{1}, ..., tf_{ m }} co-regulating a set of genes {g_{1}, ..., g_{ n }}, where ${h}_{t{f}_{i}}$ is the behavior of TF i; h_{ g } is the behavior of all the target genes in R, and h_{ x } ∈ {↑, ↓, -}, meaning up-express, down-express and no change respectively. For example, if tf_{1} ↑ and tf_{2} ↓ always cause the target genes g_{1} and g_{2} to be up-regulated, the regulatory interaction is <tf_{1} ↑, tf_{2} ↓> ⇒ g ↑. For individual regulatory interactions, I ∈ {AN, AS, RN, RS, ANS, RNS}. In this work, we assume that a regulatory interaction is consistent in the context of transcriptional control as long as the experimental conditions are unchanged. Note that binaries gene expression values are used in mTRIM, since TF activity is not always proportional to its mRNA abundance [23].
mTRIM Step 1. Inferring individual regulatory interactions
To solve a relatively easier problem of inferring the regulatory interactions for each individual TF and to prepare input for multi-TF regulatory interaction inference, an EM-based Bayesian inference algorithm has been developed [20, 21].
To define the probabilities in Eq. 2 and Eq. 3, we followed the definitions in [20]. Eq 2 represents the prior probability of the interaction model I_{ m }, and Eq 3 represents the probability of gene expression correlation between TFs and targets given the interaction model I_{ m }. In the Bayesian model, the training dataset is a matrix that contains gene expression levels of TFs and their targets, from which Γ(I_{ m }) is estimated using Eq 4. And then, the likelihood is calculated using Eq 3. The prior probabilities are randomly assigned initially. In each iteration, the posterior probabilities and the frequency of I_{ m } are updated. The iteration will continue till the posterior probabilities converge.
where Pri(I_{ m }) is the prior probability of regulatory interaction I_{ m } (defined in Eq 2) and the likelihood Lk(tf_{ m }, R_{ k }, I_{ m }) is defined in Eq 3.
where fre(I_{ m }) is the frequency of regulatory interaction I_{ m } in all of the regulatory modules, |R| is the number of the regulatory modules, |TF| is the number of TFs, and I_{ m } ∈ {AS, RS, AN, RN}.
An expectation-maximization (EM) algorithm is adopted to maximize the posterior probabilities Pos(tf_{ m }, R_{ k }, I_{ m }). The EM model is initialized with each TF assigned a random regulatory interaction. In the expectation step, we compute the likelihood of each TF to be a specific interaction using Eq 3. Consequently, the posterior probabilities of interactions for every TF is updated with Eq 1. As a result, each TF is assigned with the regulatory interaction with the highest posterior probability. In the maximization step, we maximize the scoring function $S\left({R}_{k}\right)={\sum}_{m=1}^{\left|TF\right|}{\sum}_{n=1}^{\left|G\right|}\Gamma \left({I}_{m}\right)$ for each regulatory module R_{ k }, which measures how the interaction of each TF in R_{ k } matches the target gene expression changes. Note that in the iteration the priors are updated but the likelihoods are constant.
Finally, in order to determine whether I_{ m } is "necessary and sufficient" (ANS and RNS) or "no decision", the following strategy is adopted: if none of the posterior probabilities are significant, the output is "no decision"; if the probabilities of both N and S states are significant, and there is no significant difference between them, the output is ANS or RNS depending on the target gene expression direction; otherwise the output is the regulatory interaction with the highest posterior probability.
Illustrative example of time-series gene expression data for the genes in Figure 1b.
t _{0} | t _{1} | t _{2} | t _{3} | t _{4} | t _{5} | t _{6} | t _{7} | t _{8} | t _{9} | t _{10} | t _{11} | |
---|---|---|---|---|---|---|---|---|---|---|---|---|
tf _{1} | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ |
tf _{2} | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ | ↑ |
tf _{3} | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ |
g _{1} | ↑ | ↑ | ↑ | ↑ | ↑ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ |
g _{2} | ↑ | ↑ | ↑ | ↑ | ↑ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ |
Illustrative example of regulatory interaction identification on the TRN in Figure 1b.
Regulatory Interaction | AfnScore | p-value | |
---|---|---|---|
I _{0} | <tf_{1} ↓>⇒ g ↓ | - | 0.05 |
I _{1} | <tf_{1} ↑, tf_{2} ↑>⇒ g ↑ | 0.347 | 0.06 |
I _{2} | <tf_{2} ↑, tf_{3} ↓>⇒ g ↑ | 0.173 | 0.09 |
I _{3} | <tf_{1} ↑, tf_{3} ↓>⇒ g ↑ | 0.347 | 0.06 |
I _{4} | <tf_{1} ↑, tf_{2} ↑, tf_{3} ↓>⇒ g ↑ | 0.347 | 0.04 |
mTRIM Step 2. Mining multiple-TF regulatory interactions
where P(x) is the number of times that x appears in the given time series gene expression dataset divided by the product of the total number of time points and the total number of target genes. The p-value of each candidate regulatory interaction is computed by considering the distribution of AfnScore for the regulatory interactions with the same number of TFs. Only the candidate interactions with p-values smaller than 0.05 are reported to the user. Specifically, if all the TFs in I are up-regulated, the TFs are "sufficient"; if they are all down-regulated, the TFs are "necessary"; otherwise, each TF acts differently to drive the target genes to the same direction.
For an illustrative example, there are 40 possible multiple-TF regulatory interactions in the regulatory module shown in Figure 1b. Using the time-series gene expression data in Table 1 all the 2-TF regulatory interaction candidates are screened and their p-values are computed (see Table 2 row 2-4). Since none of the 2-TF regulatory interaction candidates is significant, a 3-TF interaction I_{4} =<tf_{1} ↑, tf_{2} ↑, tf_{3} ↓>=>g ↑ is generated by merging I_{2} and I_{3}. The AfnScore of I_{4} is ((10/24) * (10/24))/(12/24)) = 0.347 and its p-value is 0.04 (see Table 2 row 5). Based on I_{0} and I_{4}, we conclude that the target genes g_{1} and g_{2} are induced by the up-expression of tf_{1} and tf_{2} and the down-expression of tf_{3}, and the same target genes are repressed by the down-expression of tf_{1}.
Experimental results
mTRIM was applied on two independently-constructed yeast transcriptional regulatory networks (the Harbison dataset [15] and the Reimand dataset [12]) to identify regulatory interactions. For performance comparison, DREM v3.0 [17] and TRIM [19] were both applied on the same datasets. We did not compare mTRIM with Yeang's method [3] because the latter's objective is to build a reliable TRN instead of predicting regulatory interactions. We evaluated these methods systematically with three independent sources: single TF knockouts [16] for individual regulatory interactions, genetic interactions (GI) [25] for 2-TF regulatory interactions and synthetic data for high-order regulatory interactions.
The number and type of the regulatory interactions for individual TFs predicted by mTRIM.
Dataset | Necessary | Sufficient | Necessary & Sufficient | No Decision | |||
---|---|---|---|---|---|---|---|
Activator | Repressor | Activator | Repressor | Activator | Repressor | ||
Harbison | 194 | 184 | 118 | 162 | 29 | 69 | 4414 |
Reimand | 22 | 43 | 42 | 32 | 7 | 18 | 1543 |
Number of the multiple-TF regulatory interactions identified by mTRIM.
Dataset | 2-TF | 3-TF | 4-TF | 5-TF | 6-TF |
---|---|---|---|---|---|
Harbison | 350 | 61 | 82 | 43 | 10 |
Reimand | 95 | 15 | 7 | 7 | 0 |
Data preparation
Yeast ChlP-chip binding data [15] was downloaded from http://younglab.wi.mit.edu/regulatory_code, and a p-value cutoff of 0.001 was applied (the same threshold used in [4]) to obtain the Harbison dataset. It contains 169 TFs, 2,864 target genes and 6,253 TF-DNA bindings. Next we applied the same statistical approach as in [12] to filter the union of the yeast ChlP-chip binding data [26] and the binding-site predictions [27, 28] to generate the Reimand dataset with 2,230 TF-DNA binding relationships between 268 TFs and 1,509 target genes. To obtain the regulatory modules in the TRNs, all the target genes were clustered based on their gene expression values with Cluster 3.0 (specifically, k-means), which uses Pearson correlation coefficient for gene similarity metric [29], resulting in 50 clusters. The clusters are then evaluated with Gene Ontology enrichment analysis using Bingo [30], and unenriched clusters are discarded. To construct regulatory modules from the clustering results, the target genes that are regulated by the same TFs were partitioned if they are not in the same cluster. Finally, 2,172 and 1,031 regulatory modules were obtained in the Harbison and Reimand networks respectively. The distribution of genes and regulatory modules (Figure SI and Table S2 in Additional file 1) reveal that many genes are bound by multiple TFs.
To identify the individual and collaborative regulatory interactions in the above datasets, three widely used time-series microarray datasets (alpha, CDC28 and elu) from yeast cell cycle studies were collected [31] as training data. These datasets contain 49 time points in total. In these experiments, yeast cells were first synchronized to the same cell cycle stage, released from synchronization, and then the total RNA samples were taken at even intervals for a period of time (Table SI in Additional file 1). In order to decide whether a gene is significantly up or down regulated, a gene expression change cutoff of 0.35 was applied (the same threshold used in [19]).
To evaluate the individual regulatory relations, single-TF knockout microarray data were collected [16], and a p-value cut-off of 0.05 (as used in [16]) was applied to determine whether a gene is significantly affected by a TF knockout. To evaluate the 2-TF regulatory interactions, we downloaded the SGA genetic interaction dataset [25], which is composed of 1,711 queries crossed to 3,885 array strains. Of the 1,711 queries, 1,377 are deletion mutants of non-essential genes and 334 are essential gene alleles. The SGA dataset contains 762,146 genetic interactions. Two genes are genetically interacted if mutations in both of them produce a phenotype that is significantly different to each mutation's individual effects. In a 2-TF regulatory interaction, if TFs collaboratively regulate the same target genes, the down-regulation of both TFs should have a significantly different phenotype as the down regulation of each individual TF. Therefore, such TF pairs should have a significant p-value in the GI dataset. To evaluate the high-order multiple-TF regulatory interactions, a synthetic binding network were built, which contains 11 TFs, 17 target genes and 58 regulation/binding relationships. The network also contains two feed forward loops. Corresponding time-series gene expression data containing 500 time-points were randomly generated with 10% or 40% noise rate.
Evaluation 1. Single TF knock-outs
We used the single TF knockout microarray data to evaluate the performance of mTRIM on individual TF regulatory interaction predictions in terms of the identification of "necessary" TFs (i.e., if the expression values of the target genes are significantly changed when the TF is knocked out). For the Harbison dataset, the prediction precision of mTRIM is 94.44%, higher than the results of TRIM (82.50%). Using the Reimand dataset, mTRIM has a precision of 91.94%, significantly higher than the results of TRIM (61.54%). DREM is not compared since it does not predict "necessary" TFs.
Evaluation 2. Genetic interaction
Evaluation 3. Synthetic transcriptional regulatory networks
Case studies
Conclusion
Revealing the mechanisms of the transcriptional regulatory programs in TRNs is essential for understanding the complex control by which genes are expressed in living cells. The inference of collaborative protein-DNA functions helps paving the critical path for new drug development. In this work, we identify the regulatory interactions between TFs and target genes with mTRIM, an integration of an EM-based Bayesian inference and a new association rule mining approach built on a set of basic constraints that relate gene expression patterns to regulatory interactions. mTRIM is not limited by the number of TFs. The experimental results show that mTRIM is clearly better than the existing algorithms. Since it is difficult to obtain the ground truth for algorithm performance evaluation on real data, we generated two sets of synthetic data and used them to validate the results of our algorithm. In our future work, we will use third-party biological evidences including multiple TF knockouts, metabolic pathways, protein-protein interactions, etc., for biological validation. In our future work, we would like to extend this work by including extra data in addition to wild-type gene expression datasets. For example, since miRNA can degrade the genes induced by certain TFs [32], we will consider miRNA-target bindings and miRNA expressions, aiming to understand how miRNAs and TFs collaborate to regulate target gene expressions.
Declarations
Acknowledgements
This project has been funded by the Egyptian Government GM 845.
Declarations
The publication costs for this article were funded by the corresponding author's institution.
This article has been published as part of BMC Systems Biology Volume 8 Supplement 1, 2014: Selected articles from the Twelfth Asia Pacific Bioinformatics Conference (APBC 2014): Systems Biology. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcsystbiol/supplements/8/S1.
Authors’ Affiliations
References
- Qiu P: Recent advances in computational promoter analysis in understanding the transcriptional regulatory network. Biochem Bioph Res Co. 2003, 309: 495-501. 10.1016/j.bbrc.2003.08.052.View ArticleGoogle Scholar
- Maienschein-Cline M, Zhou J, White K, Sciammas R, Dinner A: Discovering Transcription Factor Regulatory Targets Using Gene Expression and Binding Data. Bioinformatics. 2011, 28: 206-213.PubMed CentralView ArticlePubMedGoogle Scholar
- Yeang H, Jaakkola T: Modeling the combinatorial functions of multiple transcription factors. J Comput Biol. 2006, 13: 463-480. 10.1089/cmb.2006.13.463.View ArticlePubMedGoogle Scholar
- Ernst J, Vainas O, Harbison C, Simon I, Bax-Joseph Z: Reconstructing dynamic regulatory maps. Mol Syst Biol. 2007, 3 (74): 1-13.Google Scholar
- Segal E, Shapira M, Regev A, Pe'er D, Botstein D, Koller D, Friedman N: Module networks: identifying regulatory modules and their condition-specific regulators from gene expression data. Nat Genet. 2003, 34: 166-167. 10.1038/ng1165.View ArticlePubMedGoogle Scholar
- Hagen D, McCaffrey G, Sprague G: Pheromone Response Elements Are Necessary and Sufficient for Basal and Pheromone-Induced Transcription of the FUS1 Gene of Saccharomyces cerevisiae. Method Mol Cell Biol. 1991, 11 (6): 2952-2961.View ArticleGoogle Scholar
- Babur O, Demir E, Gonen M, Sander C, Dogrusoz U: Discovering modulators of gene expression. Nucleic Acids Res. 2010, 38: 5648-5656. 10.1093/nar/gkq287.PubMed CentralView ArticlePubMedGoogle Scholar
- Park P: ChlPseq: advantages and challenges of a maturing technology. Nat Rev Genet. 2009, 10 (10): 669-680. 10.1038/nrg2641.PubMed CentralView ArticlePubMedGoogle Scholar
- Deplancke B, Dupuy D, Vidal M, Walhout A: A gateway-compatible yeast one-hybrid system. Genome Res. 2004, 14 (10b): 2093-2101. 10.1101/gr.2445504.PubMed CentralView ArticlePubMedGoogle Scholar
- Deplancke B, Mukhopadhyay A, Ao W, et al: A gene-centered C. elegans protein-DNA interaction network. Cell. 2006, 125: 1193-1205. 10.1016/j.cell.2006.04.038.View ArticlePubMedGoogle Scholar
- Ren B, Robert F, Wyrick J, et al: Genome-wide location and function of DNA binding proteins. Science. 2000, 290: 2306-2309. 10.1126/science.290.5500.2306.View ArticlePubMedGoogle Scholar
- Reimand J, Vaquerizas J, Todd A, Vilo J, Luscombe N: Comprehensive reanalysis of transcription factor knockout expression data in Saccharomyces cerevisiae reveals many new targets. Nucleic Acids Res. 2010, 38: 4768-4777. 10.1093/nar/gkq232.PubMed CentralView ArticlePubMedGoogle Scholar
- Hoth S, Morgante M, Sanchez J, et al: Genome-wide gene expression profiling in Arabidopsis thaliana reveals new targets of abscisic acid and largely impaired gene regulation in the abil-1 mutant. J Cell Sci. 2006, 115: 4891-4900.View ArticleGoogle Scholar
- Tong A, Boone C: Synthetic genetic array analysis in Saccharomyces cerevisiae. Meth Mol Biol. 2006, 313: 171-191.Google Scholar
- Harbison C , B G, Lee T, et al: Transcriptional regulatory code of a eukaryotic genome. Nature. 2004, 431: 99-104. 10.1038/nature02800.PubMed CentralView ArticlePubMedGoogle Scholar
- Hu Z, Killion P, Iyer V: Genetic reconstruction of a functional transcriptional regulatory network. Nat Genet. 2007, 39: 683-687. 10.1038/ng2012.View ArticlePubMedGoogle Scholar
- Bar-Joseph Z, Gerber G, Lee T, et al: Computational discovery of gene modules and regulatory networks. Nat Biotechnol. 2003, 21: 1337-1342. 10.1038/nbt890.View ArticlePubMedGoogle Scholar
- Balaji S, Babu M, Iyer M, Luscombe M, Aravind L: Comprehensive analysis of combinatorial regulation using the transcriptional regulatory network of yeast. Mol Biol. 2001, 360: 213-227.View ArticleGoogle Scholar
- Awad S, Panchy N, Ng S, Chen J: Inferring the regulatory interaction types of transcription factors in transcriptional regulatory networks. J Bioinfo Comp Bio. 2012, 10 (5): 1250012-10.1142/S0219720012500126.View ArticleGoogle Scholar
- Duda R, Hart P, Stork D: Pattern Classification. 2001, John Wiley and Sonss, 2Google Scholar
- Thorne T, Stumpf M: Inference of temporally varying Bayesian Networks. Bioinformatics. 2012, 28 (24): 3298-3305. 10.1093/bioinformatics/bts614.PubMed CentralView ArticlePubMedGoogle Scholar
- Ihmels J, Friedlander G, et al: Revealing modular organization in the yeast transcriptional network. Nat Genet. 2002, 31: 370-377.PubMedGoogle Scholar
- Gygi S, Rochon Y, Franza B, Aebersold R: Correlation between protein and mRNA abundance in yeast. Molecular and cellular biology. 1999, 19 (3): 1720-1730.PubMed CentralView ArticlePubMedGoogle Scholar
- Agrawal R, Srikant R: Fast algorithms for mining association rules. Proc of VLDB. 1994, 487-499.Google Scholar
- Costanzo M, Baryshnikova A, Bellay J, et al: The Genetic Landscape of a Cell. Science. 2010, 327: 425-431. 10.1126/science.1180823.View ArticlePubMedGoogle Scholar
- Lee T, Rinaldi N, Robert F, et al: Transcriptional Regulatory Networks in Saccharomyces cerevisiae. Science. 2002, 298: 799-804. 10.1126/science.1075090.View ArticlePubMedGoogle Scholar
- Erb I, Nimwegen E: Statistical features of yeast's transcriptional regulatory code. Proc of ISCB. 2006, 1: 111-118.Google Scholar
- Maclsaac K, Wang T, Gordon B, Gifford D, Stormo G, Fraenkel E: An improved map of conserved regulatory sites for Saccharomyces cerevisiae. BMC Bioinformatics. 2006, 7: 113-10.1186/1471-2105-7-113.View ArticleGoogle Scholar
- Eisen M, Spellman P, Brown P, Botstein D: Cluster analysis and display of genome-wide expression patterns. P Natl Acad Sci USA. 1998, 95: 14863-14868. 10.1073/pnas.95.25.14863.View ArticleGoogle Scholar
- Maere S, Heymans K, Kuiper M: BiNGO a Cytoscape plugin to assess overrepresentation of Gene Ontology categories in Biological Networks. Bioninformatics. 2005, 21: 3448-3449. 10.1093/bioinformatics/bti551.View ArticleGoogle Scholar
- Spellman P, Sherlock G, Zhang M, et al: Comprehensive Identification of Cell Cycle-regulated Genes of the Yeast Saccharomyces cerevisiae by Microarray Hybridization. Mol Biol Cell. 1998, 9: 3273-3297. 10.1091/mbc.9.12.3273.PubMed CentralView ArticlePubMedGoogle Scholar
- Joung J, Hwang K, et al: Discovery of microRNA-mRNA modules via population-based probabilistic learning. Bioinformatics. 2007, 23: 1141-1147. 10.1093/bioinformatics/btm045.View 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. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.