PETALS: Proteomic Evaluation and Topological Analysis of a mutated Locus' Signaling
© Bebek et al; licensee BioMed Central Ltd. 2010
Received: 1 July 2010
Accepted: 13 December 2010
Published: 13 December 2010
Colon cancer is driven by mutations in a number of genes, the most notorious of which is Apc. Though much of Apc's signaling has been mechanistically identified over the years, it is not always clear which functions or interactions are operative in a particular tumor. This is confounded by the presence of mutations in a number of other putative cancer driver (CAN) genes, which often synergize with mutations in Apc.
Computational methods are, thus, required to predict which pathways are likely to be operative when a particular mutation in Apc is observed.
We developed a pipeline, PETALS, to predict and test likely signaling pathways connecting Apc to other CAN-genes, where the interaction network originating at Apc is defined as a "blossom," with each Apc-CAN-gene subnetwork referred to as a "petal." Known and predicted protein interactions are used to identify an Apc blossom with 24 petals. Then, using a novel measure of bimodality, the coexpression of each petal is evaluated against proteomic (2 D differential In Gel Electrophoresis, 2D-DIGE) measurements from the Apc1638N+/-mouse to test the network-based hypotheses.
The predicted pathways linking Apc and Hapln1 exhibited the highest amount of bimodal coexpression with the proteomic targets, prioritizing the Apc-Hapln1 petal over other CAN-gene pairs and suggesting that this petal may be involved in regulating the observed proteome-level effects. These results not only demonstrate how functional 'omics data can be employed to test in silico predictions of CAN-gene pathways, but also reveal an approach to integrate models of upstream genetic interference with measured, downstream effects.
It is clear that sporadic colorectal cancer - as well as other cancers - is largely the product of acquired somatic mutations . Though many of these mutations are functionally relevant to the tumor ("driver" genes), the most well-studied cancer driver gene remains Apc (adenomatous polyposis coli), thought to be the first hit in the majority of nonhereditary colon cancers . While Apc is commonly known as an antagonist to β-catenin and WNT signaling, a growing body of evidence points to the importance of Apc in a variety of other cellular contexts - from microtubule polymerization  to cell migration . Apc also plays important roles in chromosome segregation and stability, localizing to spindles, kinetochores, and centrosomes in mitosis [5, 6]. The myriad aspects of Apc signaling may not be relevant in all cellular contexts, however, as signaling depends upon the background gene expression program and, in cancer biology, is often the result of multiple mutations. In fact, mouse models mutated at two driver genes simultaneously have shown a synergistic (i.e. non-additive) increase in tumor burden, such as in Pten-Apc , Kras-Tgfb , and Apc-Trp53  double mutants. Such genetic synergy suggests that the pathways emanating from the two genes intersect downstream, supporting the idea that only a subset of all possible pathways are involved in a tumor harboring a mutation in Apc. We hypothesize that these mutations have distinct synergistic effects on the cancer phenotype, such that the activities of these networks are greatly associated with the measured downstream changes in the proteome of the intestine. We argue that these measured molecular changes can be leveraged to elucidate which pathways are most relevant to the disease model at hand.
Results and Discussion
In this paper, we present a method to capture the likely signaling pathways of a cancer driver gene, focusing on the signaling related to Apc as an example. The initial set of pathway predictions are mined from protein-protein interaction networks, coupled to mRNA coexpression data and Gene Ontology association rules. We refer to this data-mining process as the Blossom Algorithm (Figure 1 top), as it produces a network connecting a driver gene (e.g. Apc) to a set of putative signaling partners, referred to as the Apc blossom (Figure 1 center). The Apc blossom is then pruned using biological evidence (microarray and proteomic data) to identify a candidate petal, or subnetwork, most likely to be involved in Apc signaling (Figure 1 bottom).
The Apc Blossom: A PETALS Network
A blossom can be constructed for a wide variety of genes, with the stipulation that corresponding microarray data is available. In our case study of Apc, we employ mRNA expression data from intestinal tumors harvested from ApcMin/+mice. As multiple mutations are present in these samples, coexpression measurements calculated for this dataset are representative of the tumor microenvironment; as such, both Apc signaling, as well as additional CAN-gene signaling, are likely to be active simultaneously. While the presence of these multiple, active pathways increases the signal associated with cross-talk within in each petal, it does not allow us to determine which pathways are most strongly associated with Apc signaling alone. To answer this question, as outlined in the next section, we used mice with a particular heterozygous mutation in Apc - 1638N - that results in a mild intestinal cancer phenotype , thereby minimizing the noise arising from the many pathways activated in a full-blown tumor. Since we are interested in assessing the systems-level effects of such mutations, we focus on measuring the downstream effects of these genes via 'omic experiments.
Plucking Petals: Testing the Bimodality of Coexpression
To understand how a mutation affects information flow in a tumor, one must consider both the proximal and distal signaling effects. Proximally, a mutation in a gene may result in a truncated protein product that affects physical interactions, or it may result in a hyperphosphorylated and active state. These small, upstream effects are then amplified and result in distal changes in signaling, affecting mRNA and protein levels of tens to hundreds of seemingly unrelated nodes. While the field of cell signaling is adept at dissecting the proximal effects of a mutation - mechanistically mapping out perturbed pathways - it has not yet developed the tools to fully understand the distal effects and, more importantly, their connection with more proximal signaling. Indeed, currently available commercial software for network analysis can only associate these distal effects amongst themselves, with no regard to the upstream causative mutation. In this study, we present a method by which the distal effects measured in two 'omics experiments - microarray and proteomics - can be simultaneously leveraged to test network-based hypotheses. After testing the hypotheses (petals) against proteomic evidence, the refined petal subnetworks we present not only reveal the relationship between upstream genetic interference and its downstream effects at the proteomics level, but also allow us to prioritize other cancer-driver genes that are likely to act cooperatively with Apc to drive tumorigenesis. This new approach - linking in silico predictions with experimental measurements - provides a way forward in mining context-specific pathways that may prove to be useful in identifying pathways active in individual cancer patients.
The Blossom Algorithm
The Apc blossom is built using the Blossom algorithm, based on the PathFinder architecture . A recent study compared various frameworks developed for detecting signaling networks , and the PathFinder architecture had the best recall rate compared to other available methods, whereas all methods described had a similar precision rate.
In the Blossom algorithm, networks (e.g. pathways) connecting proteins of interest are built by integrating and mining multiple datasets. First, the network of publicly available interactions [20, 21] (over 80K interactions) is filtered to remove less reliable interactions, i.e. likely false positives, and, then, new interactions are added to enrich the network to account for missing interactions, i.e. false negatives. To remove false postives, a logistic regression model that incorporates (i) the number of times a PPI is observed, (ii) coexpression measurements for the corresponding genes, (iii) the protein's small world clustering coefficient, and (iv) the protein subcellular localization data of interacting partners .
Coexpression values (Pearson's correlation coefficient) are calculated from mRNA expression profiles of the laser-capture microdissected epithelium from the ApcMin/+mouse (series GSE422 ), providing coregulatory information specific to our tissue and organism of interest. The logistic regression model that predicts the validity of interactions is trained on positive (1000 PPIs from the MIPS database ) and negative training data sets (1000 randomly selected PPIs not in MIPS, assuming that most interactions are unreliable or irrelevant [11, 25]). Repeating these trials 100 times, an optimized cut off point for the probability of a true interaction is set, and a network of reliable interactions is formed (~ 30K PPIs).
Finally, false negative interactions are inferred using sequence homology relationships, as it has been shown that similar sequences share similar interaction partners in the same organism [26–29]. An interaction edge is inferred among two proteins if no record of interaction exists, and there exists at least one interaction between the protein families of these two proteins (since sequences sharing similar domains share similar interaction partners [30, 31]) (Pfam release 23.0 used ).
These steps resulted in a filtered network with predicted edges within which we searched for pathways linking Apc and CAN-genes. GO biological process annotations  are used to generate functional association rules from know pathways [24, 33–35] as outlined in . Association rules are tuples representing a noteworthy relationship, in this case functional relationships, between two interacting proteins. For each protein, leaf terms on the GO term graph are used. In addition, the average absolute coexpression is calculated for each path, and paths are then filtered according to a set threshold (γ = 0.6). These rules and parameters are used to evaluate candidate paths for possible occurrences of these rules. The p-value, p ϕ , for a path, ϕ, is calculated with the null hypothesis being that every simple path connecting two proteins has a number of association rules associated with these interactions, but the average number of rules on these paths are uniform across various paths. Significant paths, i.e. p ϕ <p threshold , are merged into a subnetwork, thus representing a petal in the blossom. An empty set is returned when there is not a significant path.
Formally, let G(V, E) denote the PPI network gathered from publicly available interactions. Also, let G' and G'' be networks built on the same set of nodes, V, using the procedures described above, where false positive interactions, F, are removed, E' = (E - F), to obtain G'(V, E'), and a set of additional interactions, H, are imputed (based on sequence information) to obtain E'' = E' ∪ H forming G''(V, E'').
The objective of the proposed Blossom framework is to find a petal for a given protein c a ∈ V (in our case, Apc) and each protein c i in the candidate set of proteins C ⊂ V (CAN-genes). To reduce the search space, Blossom employs a network diameter heuristic. Namely, for each node pair (c a and c i ), let d i denote the shortest path between c a and c i in G(V, E) (PPI network without inferred edges). For each c i ∈ C, we then search G'(V, E') for every path that connects c a to c i with path length smaller than d i that connect c a and c i . This guarantees at least one path for consideration if the two nodes are connected.
The paths on the network are discovered using all paths depth first search (AllPathsDFS), where every path connecting c a and c i that is less than d i , Φ i , is identified. In the final step of the algorithm, these paths are compared against the null distribution for significance. For the shortest path calculation, a single-source shortest path solution is used (e.g. Dijkstra's algorithm). The Blossom algorithm's run time is the same as the all-paths depth first search: where .
Input: c a , C, G(V, E), G''(V, E''), p threshold , γ
foreach c i ∈ C do
d i = ShortestPathDistance(G(V, E), c a , c i );
if d i = = ∞ then
d i = ShortestPathDistance(G''(V, E'), c a , c i );
Φ i = AllPathsDFS(G''(V, E''), c i , c a , d i );
forall the ϕ ∈ Φ i do
if r(ϕ) ≤ γ and p ϕ <p threshold then
Algorithm 1: The Blossom algorithm that returns the blossom network for protein c a .
Plucking Petals: Testing Bimodality of Coexpression
For a particular petal, a single node perturbation (e.g. a mutation at Apc) within the petal itself will perturb pathways that are expected to associate with the given petal more strongly than others, assuming that the network predictions were accurate. To identify the best petal in the Apc blossom, we employed a mouse mutant, Apc1638N+/-, representing a perturbation at the stamen. The transcript and protein levels of Apc itself have been verified in previous studies ; in this study, we were interested in distilling the myriad downstream effects into a coherent set of candidate pathways. As proteins are the ultimate mediators of function, targets from proteomic experiments - such as label-free, or, in our case, 2 D DIGE - represent an ideal dataset for assessing the downstream effects of such perturbations. However, proteomic technologies often sample the most abundant quartile of proteins , while cancer network predictions - such as those in the Apc blossom - often focus on low-abundance signaling proteins.
In order to make inferences about identified petals, a relational map must be used to connect the proteomic targets to the petal of interest. Coexpression networks are currently the most informative and accessible mapping available, as proteins correlated at the mRNA-level are hypothesized to be coregulated.
Thus, for a hypothesized petal, P, mRNA coexpression (Pearson's correlation coefficient) was calculated between the nodes, i ∈ P , and the 2D-DIGE targets, d ∈ D (where D ⊂ S and S is the set of all genes on the array) measured in the Apc1638N+/-mouse intestinal epithelium. The 2D-DIGE targets' Mascot DAT files are available through the Proteomics Identifications Database (accession number 10638) .
Apc1638N+/-microarray data is available through the Gene Expression Omnibus (GSE19338) . Two fractions, representing crypts and villi, were available with four samples in each group (eight samples each, wild-type and Apc1638N+/-). Though the mild phenotype of the Apc1638N+/-mouse appears to result in a low signal - in stark contrast to that observed from ApcMin/+mice - many molecular changes are still measurable, as evidenced by the 'omic experiments. The proteins identified within each fraction were pooled to arrive at a set of 31 2D-DIGE targets shown on the periphery of Figure 4 (see  for detailed methods). Robust Multiarray Averaging was used to normalize mRNA expression measurements, and differential expression was calculated between the eight mutant samples versus the eight wild-type samples. For coexpression, the wild-type and Apc1638N+/-microarray data were normalized by dChip  to avoid artificially inflating coexpression values .
Where E(·) indicates the expectation. Thus, we see that β P is the difference between the second moments of the two distributions (or the difference of their variances, if both distributions are centered at zero).
With β rand being the bimodality for a randomly selected set of candidate 2D-DIGE targets; 10000 such sets (of cardinality equal to that of P) were generated. Then, the null hypothesis is that the coexpression pattern between the network-petal and the proteomic targets is random, and the p-value is the probability of attaining at least a value of |β P | via stochastic generation of 2D-DIGE targets.
Authors would like to mention support from National Institutes of Health grants R25T-CA094186, P30-CA043703 and UL1-RR024989. We are grateful to Dr. Mehmet Koyutürk for critically reviewing this manuscript and for his insightful advice.
- Sjöblom T, Jones S, Wood LD, Parsons DW, Lin J, Barber TD, Velculescu VE: The consensus coding sequences of human breast and colorectal cancers. Science 2006, 314(5797):268–74.View ArticlePubMedGoogle Scholar
- Powell SM, Zilz N, Beazer-Barclay Y, Bryan TM, Hamilton SR, Thibodeau SN, Vogelstein B, Kinzler KW: APC mutations occur early during colorectal tumorigenesis. Nature 1992, 359(6392):235–7. 10.1038/359235a0View ArticlePubMedGoogle Scholar
- Nakamura M, Zhou XZ, Lu KP: Critical role for the EB1 and APC interaction in the regulation of microtubule polymerization. Curr Biol 2001, 11(13):1062–7. 10.1016/S0960-9822(01)00297-4View ArticlePubMedGoogle Scholar
- Kawasaki Y, Senda T, Ishidate T, Koyama R, Morishita T, Iwayama Y, Higuchi O, Akiyama T: Asef, a link between the tumor suppressor APC and G-protein signaling. Science 2000, 289(5482):1194–7. 10.1126/science.289.5482.1194View ArticlePubMedGoogle Scholar
- Fodde R, Smits R, Clevers H: APC, signal transduction and genetic instability in colorectal cancer. Nat Rev Cancer 2001, 1: 55–67. 10.1038/35094067View ArticlePubMedGoogle Scholar
- Kaplan KB, Burds AA, Swedlow JR, Bekir SS, Sorger PK, Näthke IS: A role for the Adenomatous Polyposis Coli protein in chromosome segregation. Nat Cell Biol 2001, 3(4):429–32. 10.1038/35070123View ArticlePubMedGoogle Scholar
- Marsh V, Winton DJ, Williams GT, Dubois N, Trumpp A, Sansom OJ, Clarke AR: Epithelial Pten is dispensable for intestinal homeostasis but suppresses adenoma development and progression after Apc mutation. Nat Genet 2008, 40(12):1436–44. 10.1038/ng.256View ArticlePubMedGoogle Scholar
- Trobridge P, Knoblaugh S, Washington MK, Munoz NM, Tsuchiya KD, Rojas A, Grady WM: TGF-beta receptor inactivation and mutant Kras induce intestinal neoplasms in mice via a beta-catenin-independent pathway. Gastroenterology 2009, 136(5):1680–8.e7. 10.1053/j.gastro.2009.01.066View ArticlePubMedPubMed CentralGoogle Scholar
- Halberg RB, Chen X, Amos-Landgraf JM, White A, Rasmussen K, Clipson L, Dove WF: The pleiotropic phenotype of Apc mutations in the mouse: allele specificity and effects of the genetic background. Genetics 2008, 180: 601–9. 10.1534/genetics.108.091967View ArticlePubMedPubMed CentralGoogle Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Sherlock G: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 2000, 25: 25–29. 10.1038/75556View ArticlePubMedPubMed CentralGoogle Scholar
- Bebek G, Yang J: PathFinder: mining signal transduction pathway segments from protein-protein interaction networks. BMC Bioinformatics 2007, 8: 335. 10.1186/1471-2105-8-335View ArticlePubMedPubMed CentralGoogle Scholar
- Wood LD, Parsons DW, Jones S, Lin J, Sj¨oblom T, Leary RJ, Vogelstein B: The genomic landscapes of human breast and colorectal cancers. Science 2007, 318(5853):1108–13. 10.1126/science.1145720View ArticlePubMedGoogle Scholar
- Gygi SP, Rochon Y, Franza BR, Aebersold R: Correlation between protein and mRNA abundance in yeast. Mol Cell Biol 1999, 19(3):1720–30.View ArticlePubMedPubMed CentralGoogle Scholar
- Pascal LE, True LD, Campbell DS, Deutsch EW, Risk M, Coleman IM, Eichner LJ, Nelson PS, Liu AY: Correlation of mRNA and protein levels: cell type-specific gene expression of cluster designation antigens in the prostate. BMC Genomics 2008, 9: 246. 10.1186/1471-2164-9-246View ArticlePubMedPubMed CentralGoogle Scholar
- Horvath S, Zhang B, Carlson M, Lu KV, Zhu S, Felciano RM, Mischel PS: Analysis of oncogenic signaling networks in glioblastoma identifies ASPM as a molecular target. Proc Natl Acad Sci USA 2006, 103(46):17402–7. 10.1073/pnas.0608396103View ArticlePubMedPubMed CentralGoogle Scholar
- Stumpf MPH, Thorne T, de Silva E, Stewart R, An HJ, Lappe M, Wiuf C: Estimating the size of the human interactome. Proc Natl Acad Sci USA 2008, 105(19):6959–64. 10.1073/pnas.0708078105View ArticlePubMedPubMed CentralGoogle Scholar
- Patel VN, Bebek G, Mariadason JM, Wang D, Augenlicht LH, Chance MR: Prediction and testing of biological networks underlying intestinal cancer. PLoS One 2010., 5(9): 10.1371/journal.pone.0012497
- Yang WC, Mathew J, Velcich A, Edelmann W, Kucherlapati R, Lipkin M, Yang K, Augenlicht LH: Targeted Inactivation of the p21WAF1/cip1 Gene Enhances Apc-initiated Tumor Formation and the Tumor-promoting Activity of a Western-Style High-Risk Diet by Altering Cell Maturation in the Intestinal Mucosa. Cancer Research 2001, 61(2):565–569. [http://cancerres.aacrjournals.org/cgi/content/abstract/61/2/565]PubMedGoogle Scholar
- Zhao XM, Wang RS, Chen L, Aihara K: Uncovering signal transduction networks from high-throughput data by integer linear programming. Nucleic Acids Res 2008, 36(9):e48. 10.1093/nar/gkn145View ArticlePubMedPubMed CentralGoogle Scholar
- Prasad TSK, Kandasamy K, Pandey A: Human Protein Reference Database and Human Proteinpedia as discovery tools for systems biology. Methods Mol Biol 2009, 577: 67–79. full_textView ArticlePubMedGoogle Scholar
- Stark C, Breitkreutz BJ, Reguly T, Boucher L, Breitkreutz A, Tyers M: BioGRID: a general repository for interaction datasets. Nuc Ac Res 2006, 34(suppl 1):D535–539. 10.1093/nar/gkj109View ArticleGoogle Scholar
- Sprenger J, Lynn Fink J, Karunaratne S, Hanson K, Hamilton NA, Teasdale RD: LOCATE: a mammalian protein subcellular localization database. Nucleic Acids Res 2008, (36 Database):D230–3.
- Paoni NF, Feldman MW, Gutierrez LS, Ploplis VA, Castellino FJ: Transcriptional profiling of the transition from normal intestinal epithelia to adenomas and carcinomas in the APCMin/+ mouse. Physiol Genomics 2003, 15(3):228–35.View ArticlePubMedGoogle Scholar
- Mewes HW, Heumann K, Kaps A, Mayer K, Pfeiffer F, Stocker S, Frishman D: MIPS: a database for genomes and protein sequences. Nuc Ac Res 1999, 27: 44–48. [http://mips.gsf.de] 10.1093/nar/27.1.44View ArticleGoogle Scholar
- Suthram S, Shlomi T, Ruppin E, Sharan R, Ideker T: A direct comparison of protein interaction confidence assignment schemes. BMC Bioinformatics 2006, 7: 360. 10.1186/1471-2105-7-360View ArticlePubMedPubMed CentralGoogle Scholar
- Deane CM, Salwinski L, Xenarios I, Eisenberg D: Protein interactions: two methods for assessment of the reliability of high throughput observations. Mol Cell Proteomics 2002, 1(5):349–356. 10.1074/mcp.M100037-MCP200View ArticlePubMedGoogle Scholar
- Obenauer JC, Cantley LC, Yaffe MB: Scansite 2.0: Proteome-wide prediction of cell signaling interactions using short sequence motifs. Nucleic Acids Res 2003, 31(13):3635–41. 10.1093/nar/gkg584View ArticlePubMedPubMed CentralGoogle Scholar
- Sharan R, Suthram S, Kelley RM, Kuhn T, McCuine S, Uetz P, Sittler T, Karp RM, Ideker T: Conserved patterns of protein interaction in multiple species. Proc Natl Acad Sci USA 2005, 102(6):1974–1979. 10.1073/pnas.0409522102View ArticlePubMedPubMed CentralGoogle Scholar
- Bebek G, Berenbrink P, Cooper C, Friedetzky T, Nadeau J, Sahinalp SC: Improved Duplication Models for Proteome Network Evolution. RECOMB 2005 Ws on Regulatory Genomics LNBI 4023: 119–137.
- Yeang CH, Haussler D: Detecting coevolution in and among protein domains. PLoS Comput Biol 2007, 3(11):e211. 10.1371/journal.pcbi.0030211View ArticlePubMedPubMed CentralGoogle Scholar
- Itzhaki Z, Akiva E, Altuvia Y, Margalit H: Evolutionary conservation of domain-domain interactions. Genome Biol 2006, 7(12):R125. 10.1186/gb-2006-7-12-r125View ArticlePubMedPubMed CentralGoogle Scholar
- Bateman A, Coin L, Durbin R, Finn RD, Hollich V, Griffiths-Jones S, Eddy SR: The Pfam protein families database. Nuc Ac Res 2004, 32: D138–141. [http://www.sanger.ac.uk/Software/Pfam] 10.1093/nar/gkh121View ArticleGoogle Scholar
- Kanehisa M, Goto S: KEGG: Kyoto Encyclopedia of Genes and Genomes. Nuc Ac Res 2000, 28: 27–30. 10.1093/nar/28.1.27View ArticleGoogle Scholar
- Campagne F, Neves S, Chang CW, Skrabanek L, Ram PT, Iyengar R, Weinstein H: Quantitative information management for the biochemical computation of cellular networks. Sci STKE 2004., 2004(248): 10.1126/stke.2482004pl11
- Gough NR, Adler EM, Ray LB: Focus Issue: Cell Signaling-Making New Connections. Sci STKE 2004, 2004(261):12. 10.1126/2612004eg12Google Scholar
- Chang J, Chance MR, Nicholas C, Ahmed N, Guilmeau S, Flandez M, Wang D, Byun DS, Nasser S, Albanese JM, Corner GA, Heerdt BG, Wilson AJ, Augenlicht LH, Mariadason JM: Proteomic changes during intestinal cell maturation in vivo. J Proteomics 2008, 71(5):530–46. 10.1016/j.jprot.2008.08.003View ArticlePubMedPubMed CentralGoogle Scholar
- Vizcaíno JA, Côté R, Reisinger F, Foster JM, Mueller M, Rameseder J, Hermjakob H, Martens L: A guide to the Proteomics Identifications Database proteomics data repository. Proteomics 2009, 9(18):4276–83.View ArticlePubMedPubMed CentralGoogle Scholar
- Barrett T, Troup DB, Wilhite SE, Ledoux P, Rudnev D, Evangelista C, Kim IF, Soboleva A, Tomashevsky M, Marshall KA, Phillippy KH, Sherman PM, Muertter RN, Edgar R: NCBI GEO: archive for high-throughput functional genomic data. Nuc Ac Res 2009, (37 Database):D885–90. 10.1093/nar/gkn764
- Li C, Wong WH: Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection. Proc Natl Acad Sci USA 2001, 98: 31–6. 10.1073/pnas.011404098View ArticlePubMedPubMed CentralGoogle Scholar
- Harr B, Schlötterer C: Comparison of algorithms for the analysis of Affymetrix microarray data as evaluated by co-expression of genes in known operons. Nucleic Acids Res 2006, 34(2):e8. 10.1093/nar/gnj010View ArticlePubMedPubMed CentralGoogle Scholar
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.