Access the full text.
Sign up today, get DeepDyve free for 14 days.
Liqiang Chen, K. Pankiewicz (2007)
Recent development of IMP dehydrogenase inhibitors for the treatment of cancer.Current opinion in drug discovery & development, 10 4
D. Wishart, Craig Knox, Anchi Guo, S. Shrivastava, Murtaza Hassanali, P. Stothard, Zhan Chang, Jennifer Woolsey (2005)
DrugBank: a comprehensive resource for in silico drug discovery and explorationNucleic Acids Research, 34
J. Staňková, J. Shang, R. Rozen (2005)
Antisense Inhibition of Methylenetetrahydrofolate Reductase Reduces Cancer Cell Survival In vitro and Tumor Growth In vivoClinical Cancer Research, 11
T. Gordi, H. Khamis (2004)
Simple solution to a common statistical problem: interpreting multiple tests.Clinical therapeutics, 26 5
M. Hofker (2003)
Faculty Opinions recommendation of PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes.
Bar-Yehuda (2001)
1J. Graph. Alg. Appl., 5
Tianhorng Chang, C.-C. Kuo (1993)
Texture analysis and classification with tree-structured wavelet transformIEEE transactions on image processing : a publication of the IEEE Signal Processing Society, 2 4
Wojciech Rzeski, C. Ikonomidou, Lechoslaw Turski (2002)
Glutamate antagonists limit tumor growth.Proceedings of the National Academy of Sciences of the United States of America, 98 11
Hawoong Jeong, B. Tombor, Réka Albert, Z. Oltvai, A. Barabási (2000)
The large-scale organization of metabolic networksNature, 407
C. Bonferroni, C. Bonferroni (1935)
Il calcolo delle assicurazioni su gruppi di teste
Christopherson (1995)
888Biochem. Soc. Trans., 23
Chang (1993)
429IEEE Trans. Image Process., 2
R. König, G. Schramm, M. Oswald, Hanna Seitz, S. Sager, M. Zapatka, G. Reinelt, R. Eils (2006)
Discovering functional gene expression patterns in the metabolic network of Escherichia coli with wavelets transformsBMC Bioinformatics, 7
Dennis (2003)
P3Genome Biol., 4
D. Huang, Brad Sherman, R. Lempicki (2008)
Systematic and integrative analysis of large gene lists using DAVID bioinformatics resourcesNature Protocols, 4
Huang da (2009)
44Nat. Protocols, 4
Bonferroni (1935)
13
Han-Yu Chuang, Eunjung Lee, Yu-Tseung Liu, Doheon Lee, T. Ideker (2007)
Network-based classification of breast cancer metastasisMolecular Systems Biology, 3
Klimberg (1996)
418Am. J. Surg., 172
R. Bar-Yehuda, G. Even, J. Feldman, J. Naor (2001)
Computing an Optimal Orientation of a Balanced Decomposition Tree for Linear Arrangement ProblemsJ. Graph Algorithms Appl., 5
Serban Nacu, R. Critchley-Thorne, Peter Lee, S. Holmes (2007)
Gene expression network analysis and applications to immunologyBioinformatics, 23 7
Russell Jones, C. Thompson (2009)
Tumor suppressors and cell metabolism: a recipe for cancer growth.Genes & development, 23 5
V. Klimberg, J. McClellan (1996)
Claude H. Organ, Jr. Honorary Lectureship. Glutamine, cancer, and its therapy.American journal of surgery, 172 5
DeBerardinis (2007)
19345Proc. Natl Acad. Sci. USA, 104
R. Christopherson, N. Williams, S. Schoettle, E. Szabados, T. Hambley, M. Manthey (1995)
Inhibitors of dihydro-orotase, amidophosphoribosyltransferase and IMP cyclohydrolase as potential drugs.Biochemical Society transactions, 23 4
David Wise, R. Deberardinis, A. Mancuso, Nabil Sayed, Xiao-yong Zhang, H. Pfeiffer, I. Nissim, E. Daikhin, M. Yudkoff, S. McMahon, C. Thompson (2008)
Myc regulates a transcriptional program that stimulates mitochondrial glutaminolysis and leads to glutamine addictionProceedings of the National Academy of Sciences, 105
M. Schwab, F. Westermann, B. Hero, F. Berthold (2003)
Neuroblastoma: biology and molecular and chromosomal pathology.The Lancet. Oncology, 4 8
Chuang (2007)
140Mol. Syst. Biol., 3
Jeong (2000)
651Nature, 407
Chen (2007)
403Curr. Opin. Drug Discov. Dev., 10
Glynn Dennis, Brad Sherman, Douglas Hosack, Jun Yang, Wei Gao, H. Lane, R. Lempicki (2003)
DAVID: Database for Annotation, Visualization, and Integrated DiscoveryGenome Biology, 4
Jones (2009)
537Genes Dev., 23
König (2006)
119BMC Bioinformatics, 7
R. Deberardinis, A. Mancuso, E. Daikhin, I. Nissim, M. Yudkoff, S. Wehrli, C. Thompson (2007)
Beyond aerobic glycolysis: Transformed cells can engage in glutamine metabolism that exceeds the requirement for protein and nucleotide synthesisProceedings of the National Academy of Sciences, 104
V. Klimberg, J. McClellan, L. Rock (1996)
Glutamine, cancer, and its therapy*American Journal of Surgery, 172
A. Oberthuer, F. Berthold, Patrick Warnat, B. Hero, Y. Kahlert, R. Spitz, K. Ernestus, R. König, S. Haas, R. Eils, M. Schwab, B. Brors, F. Westermann, M. Fischer (2006)
Customized oligonucleotide microarray gene expression-based classification of neuroblastoma patients outperforms current clinical risk stratification.Journal of clinical oncology : official journal of the American Society of Clinical Oncology, 24 31
S. Mallat (1998)
A wavelet tour of signal processing
M. Medina (2001)
Glutamine and cancer.The Journal of nutrition, 131 9 Suppl
M. Kanehisa, M. Araki, S. Goto, M. Hattori, M. Hirakawa, M. Itoh, Toshiaki Katayama, S. Kawashima, Shujiro Okuda, T. Tokimatsu, Yoshihiro Yamanishi (2007)
KEGG for linking genomes to life and the environmentNucleic Acids Research, 36
DeBerardinis (2007)
Beyond aerobic glycolysis: transformed cells can engage in glutamine metabolism that exceeds the requirement for protein and nucleotide synthesisProc. Natl Acad. Sci. USA, 104
D. Wishart, Craig Knox, Anchi Guo, D. Cheng, S. Shrivastava, D. Tzur, Bijaya Gautam, Murtaza Hassanali (2007)
DrugBank: a knowledgebase for drugs, drug actions and drug targetsNucleic Acids Research, 36
Gordi (2004)
780Clin. Ther., 26
K. Patil, J. Nielsen (2005)
Uncovering transcriptional regulation of metabolism by using metabolic network topology.Proceedings of the National Academy of Sciences of the United States of America, 102 8
Y. Fischer, U. Böttcher, M. Eblenkamp, J. Thomas, E. Jüngling, P. Rösen, H. Kammermeier (1997)
Glucose transport and glucose transporter GLUT4 are regulated by product(s) of intermediary metabolism in cardiomyocytes.The Biochemical journal, 321 ( Pt 3)
Qun Wang, S. Diskin, E. Rappaport, E. Attiyeh, Y. Mossé, Daniel Shue, E. Seiser, Jayanti Jagannathan, S. Shusterman, M. Bansal, Deepa Khazi, C. Winter, Erin Okawa, G. Grant, A. Cnaan, Huaqing Zhao, N. Cheung, W. Gerald, W. London, K. Matthay, G. Brodeur, J. Maris (2006)
Integrative genomics identifies distinct molecular classes of neuroblastoma and shows that multiple genes are targeted by regional alterations in DNA copy number.Cancer research, 66 12
Kanehisa (2008)
D480Nucleic Acids Res., 36
Fischer (1997)
629Biochem. J., 321
W. Huber, A. Heydebreck, H. Sültmann, A. Poustka, M. Vingron (2002)
Variance stabilization applied to microarray data calibration and to the quantification of differential expressionBioinformatics, 18 Suppl 1
Huber (2002)
S96Bioinformatics, 18
F. Westermann, D. Muth, A. Benner, T. Bauer, K. Henrich, A. Oberthuer, B. Brors, T. Beissbarth, J. Vandesompele, F. Pattyn, B. Hero, R. König, M. Fischer, M. Schwab (2008)
Distinct transcriptional MYCN/c-MYC activities are associated with spontaneous regression or malignant progression in neuroblastomasGenome Biology, 9
Vol. 26 no. 9 2010, pages 1225–1231 BIOINFORMATICS ORIGINAL PAPER doi:10.1093/bioinformatics/btq113 Systems biology Advance Access publication March 24, 2010 PathWave: discovering patterns of differentially regulated enzymes in metabolic pathways 1,2 3,4 1 1,2 Gunnar Schramm , Stefan Wiesberg , Nicolle Diessl , Anna-Lena Kranz , 5 3,4 3 5 Vitalia Sagulenko , Marcus Oswald , Gerhard Reinelt , Frank Westermann , 1,2,∗ 1,2,∗ Roland Eils and Rainer König Department of Bioinformatics and Functional Genomics, Institute of Pharmacy and Molecular Biotechnology, and Bioquant, University of Heidelberg, Im Neuenheimer Feld 267, Department of Theoretical Bioinformatics, German 3 4 Cancer Research Center (DKFZ), Im Neuenheimer Feld 280, Institute of Computer Science, Interdisciplinary Center for Scientific Computing, University of Heidelberg and Department of Tumor Genetics, German Cancer Research Center (DKFZ), Im Neuenheimer Feld 280, 69120 Heidelberg, Germany Associate Editor: Alfonso Valencia ABSTRACT (Chen and Pankiewicz, 2007). Besides this, the introduction of experimental high-throughput methods in functional genomics Motivation: Gene expression profiling by microarrays or transcript such as gene expression profiling using microarrays and genome- sequencing enables observing the pathogenic function of tumors on wide sequencing has evoked the challenging task to observe and a mesoscopic level. understand the pathogenic function of tumors on a mesoscopic Results: We investigated neuroblastoma tumors that clinically level. However, the large volume of information generated in these exhibit a very heterogeneous course ranging from rapid growth with experiments must be funneled into manageable and functionally fatal outcome to spontaneous regression and detected regulatory sensible partitions to select components that well describe the tumor oncogenetic shifts in their metabolic networks. In contrast to pathologies. Intelligent embedding of the expression data into the common enrichment tests, we took network topology into account underlying topology of the metabolic network may enable the by applying adjusted wavelet transforms on an elaborated and detection of tumor-specific metabolism for directed and specific new 2D grid representation of curated pathway maps from the targeting of tumor cells. Kyoto Enzyclopedia of Genes and Genomes. The aggressive form Neuroblastoma is the most common solid, extracranial tumor of the tumors showed regulatory shifts for purine and pyrimidine of early childhood, mainly affecting children at the age of about biosynthesis as well as folate-mediated metabolism of the one- 1 year. It is derived from primitive cells of the sympathetic nervous carbon pool in respect to increased nucleotide production. We system. In many patients, neuroblastoma is metastatic at the time spotted an oncogentic regulatory switch in glutamate metabolism of diagnosis and undergoes rapid progression with fatal outcome. for which we provided experimental validation, being the first steps Alternatively, neuroblastomas, especially in infants younger than towards new possible drug therapy. The pattern recognition method 1 year at diagnosis can regress spontaneously, and the tumor can we used complements normal enrichment tests to detect such differentiate into benign ganglioneuroma in older infants (Schwab functionally related regulation patterns. et al., 2003). Detailed diagnosis and appropriate adjustment of Availability and Implementation: PathWave is implemented in a therapy requires the support of investigations at the molecular level. package for R (www.r-project.org) version 2.6.0 or higher. It is freely DNA microarray technology improved the prediction of patient available from http://www.ichip.de/software/pathwave.html outcome in comparison to established risk markers (Oberthuer et al., Contact: r.koenig@dkfz.de; r.eils@dkfz.de 2006). Supplementary information: Supplementary data are available at We wanted to track how aggressive neuroblastomas have Bioinformatics online. specifically regulated their metabolism to optimize oncogenetic Received on November 24, 2009; revised on February 17, 2010; fitness, and to elucidate ways to severely perturb this process. accepted on March 13, 2010 An analysis method was required that discovers pathways in the metabolic network, especially showing significant shifts (similar and contrasting) in regulation. When analyzing data on a metabolic 1 INTRODUCTION network, enzymes can be represented by their corresponding Cancer cells exhibit a dramatically disturbed metabolism to satisfy genes. Transcriptional data, and the topological information derived their high bioenergetic demands for cell proliferation (Jones and from the metabolic network, was connected by calculating Z- Thompson, 2009). Accordingly, a long-standing strategy for cancer scores of highly correlated sub-networks (Patil and Nielsen, 2005). treatment is to attack basic tumor metabolism. Mainly, these Chuang et al. (2007) improved classification of breast cancers treatments are rather unspecific and hinder nucleotide biosynthesis with expression patterns of small subnets of a signal transduction network. Common gene expression levels of neighboring nodes in a To whom correspondence should be addressed. metabolic network were calculated by averaging over all neighbors © The Author 2010. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org 1225 [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1225 1225–1231 G.Schramm et al. of a gene and revealed several interesting regulated pathways 99 pathways with 1826 different reactions. Each pathway was represented by its adjacency-matrix. An entry at row a and column b was set to one if for the human immune system (Nacu et al., 2007). However, there existed a metabolite that was produced by reaction a and consumed these approaches were not developed to detect highly contrasting by reaction b or vice versa. The sizes of the symmetric adjacency-matrices expression of neighboring genes that undergo a switch-like shift of were between 2 × 2 and 92 × 92 reactions. regulation in a tumor cell. Especially, these switches can be highly relevant to identify potential drug targets that specifically attack the 2.2 Ordering the 2D pathway representation with the tumor at sites of flux-redirections with which the tumor established parasitic advantages. grid arrangement method Wavelet transforms have been commonly applied in information To apply our feature extraction method we required a 2D arrangement technology and image processing to track congeneric and contrasting of the metabolic network. We calculated an embedding of the metabolic signatures (Chang and Kuo, 1993; Mallat, 1998). However, applying networks for every KEGG pathway into a 2D, regular square lattice grid. To this powerful technology to analyze cellular networks is challenging. preserve neighborhood characteristics of the reactions, we were looking for embeddings in which adjacent nodes of the network were placed onto the While the underlying topology of an ordinary image is a simple grid as close to each other as possible. As a measure of distance in the lattice, lattice grid, cellular networks exhibit a rather complex scale-free we used the Manhattan distance, i.e. for any two grid points u = (i , j ) and 1 1 architecture (Jeong et al., 2000). In our initial approach, we mapped v = (i ,j ) the distance was given by d =|i −i |+| j −j |. We wanted to 2 2 1 2 1 2 the gene expression data onto (lattice grid-like) adjacency matrices determine an optimal neighborhood in which the total edge length of the of networks and applied Haar wavelet transforms onto dense subsets graph on the lattice was minimized while conserving the network topology. of the metabolic network. Features of the wavelet transforms that This resulted in an NP-hard combinatorial optimization problem. We stated could significantly separate samples from different treatments were this problem as an integral linear program (IP; see Nemhauser and Wolsey, used to extract metabolic pathways showing the most significant 1999 for an introduction to integer programming). We formulated the IP by gene expression patterns (König et al., 2006). introducing 3D binary variables x for every node v and every grid point vij In the present study, we substantially improved this technology (i, j) stating whether or not node v has to be placed on grid point (i, j). For each pair of nodes (u, v), we calculated their distance d . For a given by (i) using cell-physiologically well-defined and curated pathways uv lattice grid g, the undirected network graph G =(V, E) with node set V, edge from the Kyoto Enzyclopedia of Genes and Genomes (KEGG) set E and adjacency matrix M, the most basic IP was given by finding an (Kanehisa et al., 2008) which extensively simplified interpretation optimum for of the results, (ii) developing a new and elaborated metric to arrange min M(a,b) ·d (1) ab the order of enzyme representations as a lattice grid-like architecture a,b∈V ,a<b x,d of single pathways taking network priorities from curators of KEGG with the constraints into account, (iii) implementing a one-step frameshift concept for x =1, ∀ v ∈ V (2) vij (i,j)∈g wavelet transforms to overcome their rigidity, and (iv) we made the software freely available and easy applicable by a package x ≤1, ∀(i,j) ∈ g (3) vij v∈V for R (www.r-project.org). When applying this method to the A +B neuroblastoma tumors, the most significant expression patterns were A −B detected in purine and pyrimidine biosynthesis and folate-mediated d ≥ , ∀(a,b) ∈ V ×V , a < b (4) ab ⎪ −A +B one-carbon metabolism. These pathways can account for increased −A −B nucleotide production for proliferation. We spotted a significant where switch-like regulation pattern in glutamate metabolism that hints A := i ·x − i ·x , (5) aij bij towards de-regulated neurotransmitter production and glutamine (i, j)∈g (i, j)∈g uptake from the bloodstream. A potential drug target was proposed B := j ·x − j ·x aij bij for this pathway and experimentally validated by drug-treatment of (i, j)∈g (i, j)∈g several neuroblastoma cell lines. and x ≥0, x ∈Z, ∀ v ∈ V ,(i,j) ∈ g. (6) vij vij Constraints (2) and (3) guaranteed that all nodes were placed exactly once 2 METHODS and that each grid point could be used at most once. Constraints (4) ensured For simplification, we explained the method with a toy example of a that the distance of node a and b is given by |A|+|B|, where A and B are small synthetic pathway and simulated expression patterns (Supplementary computed by Equation (5) as A = i −i and B = j −j . All variables were a b a b Material). For all multiple testing corrections in this study, we used the enforced to values 0 or 1 by constraint (6). The problem was solved by method of Bonferroni, 1935 (Gordi and Khamis, 2004). P-values for the CPLEX 8.1 (ILOG, Gentilly, France) for 99 lattice grids (representing 99 pathways of all gene set enrichment tests and our method were corrected for KEGG-maps) with an average optimality of 96% for embeddings on square multiple testing by this method. grids of side length |V | +1, rounded up to the next integer. 2.1 Assembling the metabolic pathways 2.3 Network motifs constrain the optimization problem Pathways were defined according to curated pathway maps of the KEGG This basic model was enhanced by a number of graph dependent, additional database (version from February 4, 2009) (Kanehisa et al., 2008). Each constraints on the distance variables. They provided lower bounds for the metabolic pathway was established by defining neighbors of reactions distance sums of well-known sub-graph motifs. That is, for an edge induced using the information from KEGG (ftp://ftp.genome.jp/pub/kegg/xml/kgml/ sub-graph G ⊂ G with a least objective function contribution of lb(G ), the metabolic/organisms/hsa/). Two reactions were neighbors if a metabolite following inequality can be added or dynamically separated: existed that was the product of one and the substrate of the other. We defined reactions as the nodes and metabolites as the edges between them. d ≥ lb G (7) uv (u,v)∈E G Pathways without any connected reaction were discarded. This resulted in [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1226 1225–1231 PathWave: discovering regulation patterns in metabolic pathways very poor prognosis if the oncogene MYCN is genetically amplified, i.e. abundant in high copy number (Schwab et al., 2003). We normalized the data with the variance normalization method (Huber et al., 2002). The raw and normalized data are deposited at ArrayExpress (http:// www.ebi.ac.uk/arrayexpress; experiment accession number E-TABM-38). The expression data of each dataset was mapped onto the corresponding reactions of the transcribed enzymes using the gene–protein information from KEGG (ftp://ftp.genome.jp/pub/kegg/xml/kgml/metabolic/organisms/hsa). Mean values were taken if a reaction was catalyzed by a complex of proteins. The expression values of each reaction were z-transformed to facilitate combinations of the values that were needed for the wavelet transforms. Expression data were available for 1103 enzymatic reactions extracted from KEGG. The expression data of all samples were mapped onto the optimally ordered grid representations of all KEGG pathways, respectively. This resulted in 84 different patterns (of stage 1 and stage 4 patients) for each KEGG pathway. 2.5 Pattern recognition on lattice grids with Haar Fig. 1. Workflow of data integration, analysis and experimental validation. wavelet transforms (A) Workflow of the method: each pathway of the metabolic network We wanted to explore every possible expression pattern of neighboring was represented on optimally arranged 2D grids, gene expression data genes and groups of genes within a KEGG pathway that showed significant were mapped onto these grids for every patient, features were generated differences between samples of different conditions. For this, we performed of combined expression values of neighboring reactions in the grid, the a Haar wavelet transform for each optimized grid representation of the discriminative power for each feature was statistically estimated, pathways pathways. The wavelet transformed expression values were statistically and patterns of significance were given out for functional interpretation and tested to identify pathways with a discriminative pattern between tumors experimental validation. (B) Specific network motifs constrained the optimal of favorable and unfavorable outcome. Such a Haar wavelet transform can arrangement of the metabolic network leading to (C) an optimal arrangement be regarded as systematically applying low pass and high pass filters from of the network on a lattice-grid. (D) Gene expression data were mapped fine grain to coarse grain resolutions. It is therefore well suited to not only onto the optimally arranged grids and features were generated allowing the detect commonly regulated pathways (low-pass filter), but also to elucidate identification of discriminative patterns. switch-like behaviors within the pathways (high-pass filter), for more details see Mallat (1998). To overcome rigidity of wavelet transforms, we covered any possible combination of neighboring reactions by shifting the frame for We considered the sub-graph motifs of star graphs, cliques consisting of up to applying the transforms: we conducted the Haar wavelet transform on the 10 vertices and odd cycles (2k+1-cycles). Moreover, a certain class of trees original grid, on the grid without the first row and column, respectively, and with maximum vertex degree (T) ≤ 4 (Fig. 1B) decreased computation finally without the two, first row and column. If the number of rows (columns) time and enhanced separation ability. became odd after deletion of the first row (column) then the last row (column) These motifs included six representatives of typical sub-graph constraint was also removed. By this, we avoided an unnecessary weighting of the last classes: the 2k+1-cycles for k = 1 (Fig. 1B.1) and k = 2 (Fig. 1B.2), a star and isolated rows (columns). This procedure was carried out for all KEGG graph with eight vertices (Fig. 1B.3), a 5-clique (Fig. 1B.4) and neighborhood pathways of every sample. The results of the transforms were stored as the star graphs with 12 (Fig. 1B.5) and 15 (Fig. 1B.6) vertices. As the graphs corresponding features for every sample. were embedded optimally, the numbers gave the total edge lengths of the embeddings as well as the right-hand sides lb(G ). Furthermore, calculation time was reduced by symmetry-breaking constraints eliminating all but a 2.6 Significance tests for the pathways few representative embeddings from each equivalence class of symmetrical Pathways were ranked according to the statistical significance of an assigned embeddings. For this, grid symmetries due to translation, rotation and score. T-tests were applied on all features F of pathway π returning reflection of the embeddings were considered as well as vertex subsets whose P-values P . The pathway-score S was derived by inner permutations did not change the value of the objective function. S =max log P (8) 10 i 2.4 Assembling the gene expression data and mapping onto the network To estimate the statistical significance of pathway-score S , we randomly Our metabolic analysis was performed with gene expression data from an sampled the samples (patients, drawing without replacement) n times earlier study in which we supported clinical diagnoses of neuroblastoma (n = 10 000 for the neuroblastoma study). Assuming an extreme value tumors (Oberthuer et al., 2006). The gene expression profiling was performed distribution (Gumbel distribution), for each pathway a curve was fitted to the for 251 patients diagnosed between 1989 and 2004, in duplicate as dye-swap distribution of the pathway-scores of the permutated samples. The P-value experiments on Agilent oligonucleotide microarrays (www.agilent.com) for each pathway resulted from this fitting-curve and was corrected for with 10 163 neuroblastoma-specific probes. The age of the patients was multiple testing (Bonferroni, 1935; Gordi and Khamis, 2004). To focus on between 0 and 296 months (median age: 15 months). For our study, we the most relevant features, only pathways with more than five significantly, compared stage 1 patients without MYCN amplification, 65 in total, to differentially regulated reactions and genes were further investigated 19 stage 4 patients with MYCN amplification. According to the International (P ≤ 0.01, not corrected for multiple testing). For this, two reactions that Neuroblastoma Staging System (INSS), stage 1 tumors are localized and consisted of exactly the same genes were counted as one reaction. Pathways confined to the area of origin. Stage 4 tumors disseminate to distant with significant features were further functionally characterized by analyzing lymph nodes, bone marrow, bone, liver or other organs and have a the literature (Section 3 and Supplementary Material). [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1227 1225–1231 G.Schramm et al. Table 1. Identified significant differentially regulated pathways with more 2.7 Discovering local patterns than five differentially regulated KEGG reactions All features were statistically tested with t-tests to separate between favorable and unfavorable tumors and corrected for multiple testing. Statistically Rank Pathway P-value Score size significant features contained those sub-graphs of the metabolic network that showed differentially regulated patterns. Such regions of interest could 1 Purine metabolism <1E-16 1 directly be accessed by reconstructing the regions that were represented by 2 Glutamate metabolism <1E-16 1 the significant features and were given out by PathWave (expression patterns 3 Glycolysis/Gluconeogenesis 1.1E-14 1 in Supplementary Figs S4–S13 and Table S2). 4 Pyrimidine metabolism 5.5E-14 2 5 One carbon pool by folate 3.2E-13 1 2.8 Cell culture, reagents, treatments and data analysis 6 Phosphatidylinositol signaling 9.1E-12 2 for the experimental D-cycloserine study system 7 Pyruvate metabolism 1.3E-11 3 We used four human neuroblastoma cell lines (SK-N-SH, SH-EP, IMR32 and 8 Valine, leucine and isoleucine 6.9E-11 1 Kelly) for our analyses. They were maintained in DMEM medium (Lonza, degradation Verviers, Belgium) supplemented with l-glutamine and 10% fetal calf serum ◦ 9 Lysine degradation 1.5E-10 3 (FCS) at 37 Cin5%CO atmosphere. d-cycloserine was purchased from 10 Glycine, serine and threonine 2.1E-10 1 Sigma (Munich, Germany). For cytotoxicity assays approximately 1E+5 metabolism cells were seeded onto 24-well plates in DMEM without l-glutamine and 11 Urea cycle and metabolism of amino 2.5E-10 1 supplemented with 5% FCS. d-cycloserine was added to the cells 18 h after groups seeding at a final concentration of 5 mM. The concentration of d-cycloserine 12 Inositol phosphate metabolism 1.3E-9 1 was one order of magnitude below the reported toxic concentration in rats 13 Fatty acid metabolism 1.7E-9 3 [see DrugBank (Wishart et al., 2006, 2008)]. Total amount of cells in culture 14 Folate biosynthesis 1.8E-9 2 was determined with crystal violet (Serva) after 20 min (0 h), 24 h, 48 h 15 Glutathione metabolism 2.3E-9 1 and 72 h. The cells were fixed with 3.7% formaldehyde and stained with 16 N-Glycan biosynthesis 6.2E-8 2 0.5% crystal violet. After several washings with DPBS (Lonza), crystal 17 Starch and sucrose metabolism 1.7E-7 1 violet was re-solubilized in a buffer containing 0.1 M sodium citrate, pH 4.2 18 Glycerophospholipid metabolism 3.9E-7 1 and 50% ethanol. The absorbance of crystal violet was measured at 580 nm 19 Glycosphingolipid 5.3E-7 1 and taken as a measure for cell proliferation. Measurements were done in a biosynthesis—neo-lactoseries FLUOstar OPTIMA plate reader (BMG Labtech, Offenburg, Germany). All 20 Sphingolipid metabolism 1.4E-6 2 experiments were done in quadruplicates. 21 Tyrosine metabolism 6.6E-6 2 For each time point, raw values were subtracted by the mean background 22 Aminoacyl-tRNA biosynthesis 1.3E-5 2 intensity coming from four blank wells treated similar to the test wells. 23 Fatty acid biosynthesis 4.1E-5 1 To obtain replication levels, for each cell line and replicate, the intensities were divided by their intensities at time point 0 h, respectively. Mean values Size of the most significant pattern with which the score was calculated, 1 = 1st wavelet and standard deviations were calculated for each time point (after 0 h) using with 2 × 2 pixel on the grid; 2 = 2nd wavelet with 4 × 4 pixel, 3 = 3rd wavelet with the data from each cell line and replicate. The difference of the distributions 8 × 8 pixel. was tested with a Student’s t-test. 3 RESULTS to focus on the most relevant features. This procedure yielded significant features from 23 different pathways (Table 1). All 3.1 Combining regulation patterns of enzymes reactions of these features are shown in the Supplementary Material We compared gene expression profiles of 19 aggressive (Supplementary Table S2) together with the regulation patterns in neuroblastomas having unfavorable prognosis (stage 4 according the pathways from KEGG for the first 10 pathways (Supplementary to the international neuroblastoma staging system and amplification Figs S4–S13). The Supplementary Material also includes a summary of the MYCN oncogene) with 65 tumors having favorable prognosis of information from the literature regarding the oncogenic relevance (stage 1, no MYCN amplification). Our aim was to compare the of pathways not discussed in the main text. We denoted an enzyme regulation of metabolism of these two tumor entities. The workflow as being up- or down-regulated in the aggressive tumors only if its was as follows (Fig. 1): metabolic pathways were extracted from differential regulation was significant (P ≤ 0.01, not multiple testing the KEGG database and a lattice grid-like representation was corrected). constructed for each pathway. Enzymes were optimally arranged We included the enzyme commission (EC) numbers for all on the grid to preserve the relevant pathway topology. Gene enzymes to a convenient tracking in the pathway maps. We used expression data were mapped onto the corresponding enzymes in the the same abbreviations for metabolites as in KEGG. optimally arranged grid. Neighboring enzymes were grouped and their gene expression values combined using wavelet transforms. 3.2 A cellular switch in the glutamate metabolism These transforms yielded combined expression values (‘features’) by applying low-pass filters to detect similar expression changes and A significant pattern was identified in glutamate metabolism. high-pass filters to detect contrasting regulation patterns. We tested It consisted of six reactions (P < 1E-16, Fig. 2 and Supplementary the performance of all non-trivial features (10 377) to separate the Fig. S5). Glutamine is produced by the host, and is highly two tumor entities by a t-test. Pathways were ranked according to abundant in the blood. It is consumed by the parasitic tumor their significance. Only pathways with more than five significantly, cells to provide an ammonium supply (Medina, 2001). It has differentially regulated reactions and genes were further investigated previously been shown that tumors depress systemic glutamine [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1228 1225–1231 PathWave: discovering regulation patterns in metabolic pathways Fig. 3. Proliferation of neuroblastoma cells treated with d-cycloserine. Proliferation is shown for cells treated with 5 mM d-cycloserine (red) and Fig. 2. Regulation pattern in glutamate metabolism. In the aggressive untreated control cultures (blue). Error bars are the standard deviations (1σ) tumors, the ASCT2 glutamine transporter, amido phosphoribosyltransferase for the corresponding time point and treatment. The y-axis denotes replication (PPAT) and one amino transferase (GOT) were significantly up-regulated levels in accordance to absorbance of a dye that stained fixed cells (crystal to take up glutamine from the bloodstream and metabolize it for purine violet), normalized to 0 h (1× proliferation). In comparison to non-treated biosynthesis, amino acids biosynthesis and anapleurosis of the TCA cycle. cells, the proliferation of the treated cells decreased significantly at 72 h Ammonium detoxification by glutaminase (GLS) was significantly down- (P = 1.3E-5). There was no significant difference in proliferation at 24 h regulated. Probes for the amino transferase GPT were not included in the and 48 h. microarrays. levels in cancer patients (Klimberg and McClellan, 1996). MYCN can compensate for c-MYC activity and that MYCN/c-MYC Glutamine is catabolized into ribosylamine-5-phosphate by signaling is more active in more aggressive neuroblastoma subtypes amidophosphoribosyl transferase to synthesize purines. Amido- (Westermann et al., 2008). phosphoribosyl transferase has been proposed as a drug target to treat cancer (Christopherson et al., 1995). It may be combined 3.3 The aggressive tumors employ increased nucleotide with glutamate antagonists. Glutamate has previously been biosynthesis shown to be important for tumor growth, since inhibition using glutamate antagonists led to reduced proliferation (Rzeski Highly significant regulation patterns were detected for the pathways of purine and pyrimidine biosynthesis as well as folate-mediated et al., 2002). We observed a significant regulation pattern one-carbon metabolism. The pathways ranked at positions 1, 4 and for glutamine metabolism in the investigated neuroblastomas, 5(P ≤ 1E-16, 5.5E-14 and 3.2E-13, respectively). These pathways which is in line with these observations from the literature were mainly up-regulated to enable enforced nucleotide biosynthesis (Fig. 2). More specifically, ATP-dependent glutaminase (GLS, for increased cell cycle activity of aggressive tumors. All enzymes EC 3.5.1.2) was down-regulated, reducing glutamine catabolism involved in the biosynthetic pathway for the purines ATP and to ammonium and glutamate. Glutamine flux was redirected GTP were up-regulated (Supplementary Fig. S4). We defined the into purine biosynthesis via down-regulation of ATP-dependent pathway for ATP and GTP biosynthesis as starting with ribose- GLS and up-regulation of amidophosphoribosyl transferase (PPAT, 5-phosphate originating from the pentose phosphate pathway and EC 2.4.2.14). Amidophosphoribosyl transferase also produces glutamate metabolism using, among others, AICAR, IMP, XMP, glutamate that is used as a building block for synthesizing further GMP, GDP, ADP. Phosphatases (EC 3.1.3.5, EC 3.6.1.5, EC non-essential amino acids. Glutamate was shown to be used for TCA cycle anapleurosis in glioblastomas (DeBerardinis et al., 3.6.1.6) that degrade compounds for purine production were down- 2007). To identify an anapleurotic tendency in neuroblastoma, we regulated. We also detected a significant regulation pattern for TTP experimentally investigated the proliferation of neuroblastoma cell and CTP biosynthesis in pyrimidine metabolism (Supplementary lines after treatment with d-cycloserine. d-cycloserine compromises Fig. S7). Enzymes were up-regulated for RNA synthesis (RNA the action of two aminotransferases in the glutamate metabolism polymerase, EC 2.7.7.6), the final biosynthetic step of UTP and (GOT, EC 2.6.1.1 and GPT, EC 2.6.1.1) (Fischer et al., 1997) that CTP (nucleotide phosphate kinase, EC 2.7.4.6), dUPD production are necessary for this TCA anapleurosis. Growth was significantly from dUMP (thymidylate kinase, EC 2.7.4.9) and the conversions reduced after 72 h in neuroblastoma cells treated with 5 mM of CDP to dCDP and dUMP to dTMP (ribonucleotide reductase, d-cycloserine (P = 1.3E-5, Fig. 3). Interestingly, Wise and co- EC 1.17.4.1 and thymidylate synthetase, EC 2.1.1.45). Enzymes workers (Wise et al., 2008) have recently shown that c-MYC reversing pyrimidine anabolism were also down-regulated (EC regulated the stimulation of just such a glutaminolysis program, 3.1.3.5, EC 3.6.1.5, EC 3.6.1.6 and EC 3.5.4.5), similar to the in which glutamine is used for TCA anapleurosis, in glioma regulatory patterns for purine metabolism. cells. It is likely that this oncogenetic regulation scheme is A significant regulation pattern was found in folate-mediated metabolism of the one-carbon pool. Ten reactions were up-regulated, transferable to neuroblastoma, for which we have shown that [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1229 1225–1231 G.Schramm et al. specifically, EC 1.5.1.3 (2), EC 2.1.2.1, EC 2.1.2.2 (2), EC all high ranking pathways (ranks 1–6) of the first analysis showed 2.1.2.3, EC 2.1.1.45, EC 3.5.4.9, EC 6.3.4.3, and EC 1.5.1.15 up again in this dataset confirming our initial results. (Supplementary Fig. S8 and Table S2). This pathway serves to carry and activate single carbons for purine and pyrimidine biosynthesis, 4 DISCUSSION AND CONCLUSION utilizing pyrimidine thymidylate synthase (EC 2.1.1.45), purine GAR formyltransferase (EC 2.1.2.2) and AICAR formyltransferase We mapped gene expression data from neuroblastomas having (EC 2.1.2.3). Folate derivatives are also needed for methionine two very distinct clinical courses onto the human metabolic synthesis, which is essential for cancer cell survival (Stankova et al., network. We revealed interesting insights into tumor cell regulation 2005). To synthesize methionine, 5,10-methylene tetrahydrofolate when applying our novel method PathWave. The aggressive is processed by methylene tetrahydrofolate reductase (EC tumors showed significantly up-regulated pathways for purine 1.5.1.20) into 5-methyl tetrahydrofolate from which methionine and pyrimidine synthesis. This was expected, as tumors need synthase synthesizes methionine. In fact, inhibition of methylene these building blocks to maintain quick mitotic cycles. We tetrahydrofolate reductase has been shown to reduce tumor growth observed an interesting regulatory switch in glutamate metabolism: by depleting the cellular methionine pool (Stankova et al., the energy consuming ammonium elimination from glutamine 2005). Interestingly, methylene tetrahydrofolate reductase itself was was down-regulated, while amidophosphorybosyl transferase was not significantly regulated in the pattern we identified. Instead, up-regulated. This may have redirected toxic ammonium from thymidylate synthase (EC 2.1.1.45), was up-regulated, and may degradation into nucleotide anabolism in the aggressive tumors. have taken over producing 5,10-methylene tetrahydrofolate in these Additionally, neuroblastoma tumor cells may utilize systemic tumor cells. Based on the regulatory patterns identified by PathWave, glutamine for anapleurosis of the TCA cycle, similar to what we propose that inhibiting thymidylate synthetase in combination has been recently reported for glioblastoma cells (DeBerardinis with methylene tetrahydrofolate reductase will reduce tumor growth et al., 2007). We provided experimental evidence for this regulatory more effectively. switch in tumor cell glutamate metabolism by targeting the amino transferases at the necessary metabolic junction, which resulted in reduced proliferation of neuroblastoma cell lines. Folate- 3.4 Comparison to established methods mediated one-carbon metabolism was also differentially regulated Although PathWave was not designed as an enrichment test but in the aggressive neuroblastomas. We suggest targeting thymidylate rather to point to regulatory patterns in pathways and regions synthetase in combination with methylene tetrahydrofolate therein, we were interested in the results from an established reductase in neuroblastoma cell lines to assess the relevance gene set enrichment method. Therefore, we applied the Gene of these results for therapeutic intervention. These interesting Set Enrichment Analysis (GSEA; Mootha et al., 2003) on the regulation switches could be found by our pattern recognition expression data of all pathways that we also had analyzed with method as our implementation of wavelet transforms systematically PathWave. GSEA revealed three significantly enriched pathways, tracked co-regulated and anti-co-regulated neighboring nodes in i.e. pyrimidine metabolism, purine metabolism and polyunsaturated the network. Our network representation by lattice grids simplified fatty acid biosynthesis (P ≤1E-16, ≤1E-16, 9.9E-4, respectively, their topology as it avoided hub-like structures by including only corrected for multi-testing, results for all pathways are in the compounds that were selected to be relevant for the corresponding Supplementary Table S4). Furthermore, we used DAVID which is pathway by KEGG curators. The grid arrangement method placed another commonly used gene set enrichment test (Dennis et al., players in pathways on a 2D map while conserving the direct 2003; Huang da et al., 2009). Although DAVID revealed various neighborhoods of the players. This method is new and has potential enriched KEGG Pathways (see Supplementary Table S5a and S5b) for other applications. It is a generalization of the one-dimensional only three metabolic pathways were identified (pyrimidine and linear arrangement optimization problem (Bar-Yehuda et al., 2001) purine metabolism, and glycolysis/gluconeogenesis). Hence, GSEA to two dimensions. The model can be extended to higher dimensions, and DAVID were capable to identify enriched pathways, but showed but becomes more difficult to solve with each additional dimension. less sensitivity in comparison to PathWave The newest version of our branch-and-cut algorithm tests whether a given set of values d , (u,v) ∈ E exhibits a feasible embedding. uv PathWave enabled to focus on pathways with distinct regulated 3.5 Applying PathWave to another neuroblastoma patterns in the network and pointed to sections in these pathways dataset at which a switch-like regulation may have occurred. However, To verify our findings, we analyzed a further independent expression manual inspection and interpretation of the regulation of these dataset of neuroblastomas (Wang et al., 2006) consisting of primary sub-graphs is still necessary to derive their relevance within the tumors from 27 stage 1 patients without MYCN amplification and functional context. When using such a global scanning device, lack 20 stage 4 patients with MYCN amplification. Raw expression of specificity and sensitivity must still be tackled. For about one- data were downloaded from NCBI, normalized and analyzed third of the enzymes in KEGG, we could not assign any probe from with PathWave. The significance threshold was again set to our microarray chips. Especially, for small networks this may have P = 0.01. As the fraction of significantly regulated reactions (P ≤ led to inappropriate overestimation of single nodes. For this reason, 0.01) was lower in this dataset (29%) compared to our first we discarded pathways with to few reactions from our analyses. study (43%), we focused on pathways containing four or more In general, we detected cellular switches in the metabolism of the differentially regulated reactions. PathWave revealed 20 pathways tumor under study. The presented analysis technique is capable of (Supplementary Table S3) with significant regulation patterns, 15 out reducing the relevant pathways to those having significant patterns of which were also found in the first dataset we studied. Specifically, of functionally connected proteins. [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1230 1225–1231 PathWave: discovering regulation patterns in metabolic pathways ACKNOWLEDGEMENTS Jones,R.G. and Thompson,C.B. (2009) Tumor suppressors and cell metabolism: a recipe for cancer growth. Genes Dev., 23, 537–548. We thank Kathy Astrahantseff for stylistic corrections. Kanehisa,M. et al. (2008) KEGG for linking genomes to life and the environment. Nucleic Acids Res., 36, D480–D484. Funding: BMBF-FORSYS Consortium Viroquant (#0313923); the Klimberg,V.S. and McClellan,J.L. (1996) Claude H. Organ, Jr. Honorary Lectureship. Helmholtz Alliance on Systems Biology and the Nationales Genom- Glutamine, cancer, and its therapy. Am. J. Surg., 172, 418–424. Forschungs-Netz (NGFN+) for the neuroblastoma project, ENGINE König,R. et al. (2006) Discovering functional gene expression patterns in the metabolic network of Escherichia coli with wavelets transforms. BMC Bioinformatics, 7, 119. (#01GS0898). Mallat,S. (1998) A Wavelet Tour of Signal Processing. Academic Press, New York. Conflict of Interest: none declared. Medina,M.A. (2001) Glutamine and cancer. J. Nutrition, 131, 2539S–2542S; discussion 2550S–2531S. Mootha,V.K. et al. (2003) PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat. Genet., REFERENCES 34, 267–273. Bar-Yehuda,R. et al. (2001) Computing an optimal orientation of a balanced Nacu,S. et al. (2007) Gene expression network analysis, and applications to decomposition tree for linear arrangement problems. J. Graph. Alg. Appl., 5, 1–27. immunology. Bioinformatics, 23, 850–858. Bonferroni,C.E. (1935) Il calcolo delle assicurazioni su gruppi di teste. In, Studi in Nemhauser,G. and Wolsey,L. (1999) Integer and Combinatorial Optimization. John Onore del Professore Salvatore Ortu Carboni. Rome, 13-60. Wiley & Sons, New York. Chang,T. and Kuo,C.-C. (1993) Texture analysis and classification with tree-structured Oberthuer,A. et al. (2006) Customized oligonucleotide microarray gene expression- wavelet transform. IEEE Trans. Image Process., 2, 429–441. based classification of neuroblastoma patients outperforms current clinical risk Chen,L. and Pankiewicz,K.W. (2007) Recent development of IMP dehydrogenase stratification. J. Clin. Oncol., 24, 5070–5078. inhibitors for the treatment of cancer. Curr. Opin. Drug Discov. Dev., 10, 403–412. Patil,K.R. and Nielsen,J. (2005) Uncovering transcriptional regulation of metabolism Christopherson,R.I. et al. (1995) Inhibitors of dihydro-orotase, amidophos- by using metabolic network topology. Proc. Natl Acad. Sci. USA, 102, 2685–2689. phoribosyltransferase and IMP cyclohydrolase as potential drugs. Biochem. Soc. Rzeski,W. et al. (2002) Glutamate antagonists limit tumor growth. Biochem. Trans., 23, 888–893. Pharmacol., 64, 1195–1200. Chuang,H.Y. et al. (2007) Network-based classification of breast cancer metastasis. Schwab,M. et al. (2003) Neuroblastoma: biology and molecular and chromosomal Mol. Syst. Biol., 3, 140. pathology. Lancet Oncol., 4, 472–480. DeBerardinis,R.J. et al. (2007) Beyond aerobic glycolysis: transformed cells can engage Stankova,J. et al. (2005) Antisense inhibition of methylenetetrahydrofolate reductase in glutamine metabolism that exceeds the requirement for protein and nucleotide reduces cancer cell survival in vitro and tumor growth in vivo. Clin. Cancer Res., synthesis. Proc. Natl Acad. Sci. USA, 104, 19345–19350. 11, 2047–2052. Dennis,G. Jr. et al. (2003) DAVID: database for annotation, visualization, and integrated Wang,Q. et al. (2006) Integrative genomics identifies distinct molecular classes of discovery. Genome Biol., 4,P3. neuroblastoma and shows that multiple genes are targeted by regional alterations Fischer,Y. et al. (1997) Glucose transport and glucose transporter GLUT4 are regulated in DNA copy number. Cancer Res., 66, 6050–6062. by product(s) of intermediary metabolism in cardiomyocytes. Biochem. J., 321 Westermann,F. et al. (2008) Distinct transcriptional MYCN/c-MYC activities are (Pt 3), 629–638. associated with spontaneous regression or malignant progression in neuroblastomas. Gordi,T. and Khamis,H. (2004) Simple solution to a common statistical problem: Genome Biol., 9, R150. interpreting multiple tests. Clin. Ther., 26, 780–786. Wise,D.R. et al. (2008) Myc regulates a transcriptional program that stimulates Huang da,W. et al. (2009) Systematic and integrative analysis of large gene lists using mitochondrial glutaminolysis and leads to glutamine addiction. Proc. Natl Acad. DAVID bioinformatics resources. Nat. Protocols, 4, 44–57. Sci. USA, 105, 18782–18787. Huber,W. et al. (2002) Variance stabilization applied to microarray data calibration Wishart,D.S. et al. (2008) DrugBank: a knowledgebase for drugs, drug actions and drug and to the quantification of differential expression. Bioinformatics, 18 (Suppl. 1), targets. Nucleic Acids Res., 36, D901–D906. S96–S104. Wishart,D.S. et al. (2006) DrugBank: a comprehensive resource for in silico drug Jeong,H. et al. (2000) The large-scale organization of metabolic networks. Nature, 407, discovery and exploration. Nucleic Acids Res., 34, D668–D672. 651–654. [14:00 18/4/2010 Bioinformatics-btq113.tex] Page: 1231 1225–1231
Bioinformatics – Oxford University Press
Published: Mar 24, 2010
You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
Access the full text.
Sign up today, get DeepDyve free for 14 days.
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.